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);
123 assert_spin_locked(&inode->i_lock);
125 list_move(&inode->i_io_list, head);
127 /* dirty_time doesn't count as dirty_io until expiration */
128 if (head != &wb->b_dirty_time)
129 return wb_io_lists_populated(wb);
131 wb_io_lists_depopulated(wb);
135 static void wb_wakeup(struct bdi_writeback *wb)
137 spin_lock_irq(&wb->work_lock);
138 if (test_bit(WB_registered, &wb->state))
139 mod_delayed_work(bdi_wq, &wb->dwork, 0);
140 spin_unlock_irq(&wb->work_lock);
143 static void finish_writeback_work(struct bdi_writeback *wb,
144 struct wb_writeback_work *work)
146 struct wb_completion *done = work->done;
151 wait_queue_head_t *waitq = done->waitq;
153 /* @done can't be accessed after the following dec */
154 if (atomic_dec_and_test(&done->cnt))
159 static void wb_queue_work(struct bdi_writeback *wb,
160 struct wb_writeback_work *work)
162 trace_writeback_queue(wb, work);
165 atomic_inc(&work->done->cnt);
167 spin_lock_irq(&wb->work_lock);
169 if (test_bit(WB_registered, &wb->state)) {
170 list_add_tail(&work->list, &wb->work_list);
171 mod_delayed_work(bdi_wq, &wb->dwork, 0);
173 finish_writeback_work(wb, work);
175 spin_unlock_irq(&wb->work_lock);
179 * wb_wait_for_completion - wait for completion of bdi_writeback_works
180 * @done: target wb_completion
182 * Wait for one or more work items issued to @bdi with their ->done field
183 * set to @done, which should have been initialized with
184 * DEFINE_WB_COMPLETION(). This function returns after all such work items
185 * are completed. Work items which are waited upon aren't freed
186 * automatically on completion.
188 void wb_wait_for_completion(struct wb_completion *done)
190 atomic_dec(&done->cnt); /* put down the initial count */
191 wait_event(*done->waitq, !atomic_read(&done->cnt));
194 #ifdef CONFIG_CGROUP_WRITEBACK
197 * Parameters for foreign inode detection, see wbc_detach_inode() to see
200 * These paramters are inherently heuristical as the detection target
201 * itself is fuzzy. All we want to do is detaching an inode from the
202 * current owner if it's being written to by some other cgroups too much.
204 * The current cgroup writeback is built on the assumption that multiple
205 * cgroups writing to the same inode concurrently is very rare and a mode
206 * of operation which isn't well supported. As such, the goal is not
207 * taking too long when a different cgroup takes over an inode while
208 * avoiding too aggressive flip-flops from occasional foreign writes.
210 * We record, very roughly, 2s worth of IO time history and if more than
211 * half of that is foreign, trigger the switch. The recording is quantized
212 * to 16 slots. To avoid tiny writes from swinging the decision too much,
213 * writes smaller than 1/8 of avg size are ignored.
215 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
216 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
217 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
218 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
220 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
221 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
222 /* each slot's duration is 2s / 16 */
223 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
224 /* if foreign slots >= 8, switch */
225 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
226 /* one round can affect upto 5 slots */
227 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
230 * Maximum inodes per isw. A specific value has been chosen to make
231 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
233 #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
234 / sizeof(struct inode *))
236 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
237 static struct workqueue_struct *isw_wq;
239 void __inode_attach_wb(struct inode *inode, struct page *page)
241 struct backing_dev_info *bdi = inode_to_bdi(inode);
242 struct bdi_writeback *wb = NULL;
244 if (inode_cgwb_enabled(inode)) {
245 struct cgroup_subsys_state *memcg_css;
248 memcg_css = mem_cgroup_css_from_page(page);
249 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
251 /* must pin memcg_css, see wb_get_create() */
252 memcg_css = task_get_css(current, memory_cgrp_id);
253 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
262 * There may be multiple instances of this function racing to
263 * update the same inode. Use cmpxchg() to tell the winner.
265 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
268 EXPORT_SYMBOL_GPL(__inode_attach_wb);
271 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
272 * @inode: inode of interest with i_lock held
273 * @wb: target bdi_writeback
275 * Remove the inode from wb's io lists and if necessarily put onto b_attached
276 * list. Only inodes attached to cgwb's are kept on this list.
278 static void inode_cgwb_move_to_attached(struct inode *inode,
279 struct bdi_writeback *wb)
281 assert_spin_locked(&wb->list_lock);
282 assert_spin_locked(&inode->i_lock);
284 inode->i_state &= ~I_SYNC_QUEUED;
285 if (wb != &wb->bdi->wb)
286 list_move(&inode->i_io_list, &wb->b_attached);
288 list_del_init(&inode->i_io_list);
289 wb_io_lists_depopulated(wb);
293 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
294 * @inode: inode of interest with i_lock held
296 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
297 * held on entry and is released on return. The returned wb is guaranteed
298 * to stay @inode's associated wb until its list_lock is released.
300 static struct bdi_writeback *
301 locked_inode_to_wb_and_lock_list(struct inode *inode)
302 __releases(&inode->i_lock)
303 __acquires(&wb->list_lock)
306 struct bdi_writeback *wb = inode_to_wb(inode);
309 * inode_to_wb() association is protected by both
310 * @inode->i_lock and @wb->list_lock but list_lock nests
311 * outside i_lock. Drop i_lock and verify that the
312 * association hasn't changed after acquiring list_lock.
315 spin_unlock(&inode->i_lock);
316 spin_lock(&wb->list_lock);
318 /* i_wb may have changed inbetween, can't use inode_to_wb() */
319 if (likely(wb == inode->i_wb)) {
320 wb_put(wb); /* @inode already has ref */
324 spin_unlock(&wb->list_lock);
327 spin_lock(&inode->i_lock);
332 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
333 * @inode: inode of interest
335 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
338 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
339 __acquires(&wb->list_lock)
341 spin_lock(&inode->i_lock);
342 return locked_inode_to_wb_and_lock_list(inode);
345 struct inode_switch_wbs_context {
346 struct rcu_work work;
349 * Multiple inodes can be switched at once. The switching procedure
350 * consists of two parts, separated by a RCU grace period. To make
351 * sure that the second part is executed for each inode gone through
352 * the first part, all inode pointers are placed into a NULL-terminated
353 * array embedded into struct inode_switch_wbs_context. Otherwise
354 * an inode could be left in a non-consistent state.
356 struct bdi_writeback *new_wb;
357 struct inode *inodes[];
360 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
362 down_write(&bdi->wb_switch_rwsem);
365 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
367 up_write(&bdi->wb_switch_rwsem);
370 static bool inode_do_switch_wbs(struct inode *inode,
371 struct bdi_writeback *old_wb,
372 struct bdi_writeback *new_wb)
374 struct address_space *mapping = inode->i_mapping;
375 XA_STATE(xas, &mapping->i_pages, 0);
377 bool switched = false;
379 spin_lock(&inode->i_lock);
380 xa_lock_irq(&mapping->i_pages);
383 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
384 * path owns the inode and we shouldn't modify ->i_io_list.
386 if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
389 trace_inode_switch_wbs(inode, old_wb, new_wb);
392 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
393 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
394 * pages actually under writeback.
396 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
397 if (PageDirty(page)) {
398 dec_wb_stat(old_wb, WB_RECLAIMABLE);
399 inc_wb_stat(new_wb, WB_RECLAIMABLE);
404 xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
405 WARN_ON_ONCE(!PageWriteback(page));
406 dec_wb_stat(old_wb, WB_WRITEBACK);
407 inc_wb_stat(new_wb, WB_WRITEBACK);
410 if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
411 atomic_dec(&old_wb->writeback_inodes);
412 atomic_inc(&new_wb->writeback_inodes);
418 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
419 * the specific list @inode was on is ignored and the @inode is put on
420 * ->b_dirty which is always correct including from ->b_dirty_time.
421 * The transfer preserves @inode->dirtied_when ordering. If the @inode
422 * was clean, it means it was on the b_attached list, so move it onto
423 * the b_attached list of @new_wb.
425 if (!list_empty(&inode->i_io_list)) {
426 inode->i_wb = new_wb;
428 if (inode->i_state & I_DIRTY_ALL) {
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,
436 pos->i_io_list.prev);
438 inode_cgwb_move_to_attached(inode, new_wb);
441 inode->i_wb = new_wb;
444 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
445 inode->i_wb_frn_winner = 0;
446 inode->i_wb_frn_avg_time = 0;
447 inode->i_wb_frn_history = 0;
451 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
452 * ensures that the new wb is visible if they see !I_WB_SWITCH.
454 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
456 xa_unlock_irq(&mapping->i_pages);
457 spin_unlock(&inode->i_lock);
462 static void inode_switch_wbs_work_fn(struct work_struct *work)
464 struct inode_switch_wbs_context *isw =
465 container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
466 struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
467 struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
468 struct bdi_writeback *new_wb = isw->new_wb;
469 unsigned long nr_switched = 0;
470 struct inode **inodep;
473 * If @inode switches cgwb membership while sync_inodes_sb() is
474 * being issued, sync_inodes_sb() might miss it. Synchronize.
476 down_read(&bdi->wb_switch_rwsem);
479 * By the time control reaches here, RCU grace period has passed
480 * since I_WB_SWITCH assertion and all wb stat update transactions
481 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
482 * synchronizing against the i_pages lock.
484 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
485 * gives us exclusion against all wb related operations on @inode
486 * including IO list manipulations and stat updates.
488 if (old_wb < new_wb) {
489 spin_lock(&old_wb->list_lock);
490 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
492 spin_lock(&new_wb->list_lock);
493 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
496 for (inodep = isw->inodes; *inodep; inodep++) {
497 WARN_ON_ONCE((*inodep)->i_wb != old_wb);
498 if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
502 spin_unlock(&new_wb->list_lock);
503 spin_unlock(&old_wb->list_lock);
505 up_read(&bdi->wb_switch_rwsem);
509 wb_put_many(old_wb, nr_switched);
512 for (inodep = isw->inodes; *inodep; inodep++)
516 atomic_dec(&isw_nr_in_flight);
519 static bool inode_prepare_wbs_switch(struct inode *inode,
520 struct bdi_writeback *new_wb)
523 * Paired with smp_mb() in cgroup_writeback_umount().
524 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
525 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
526 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
533 /* while holding I_WB_SWITCH, no one else can update the association */
534 spin_lock(&inode->i_lock);
535 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
536 inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
537 inode_to_wb(inode) == new_wb) {
538 spin_unlock(&inode->i_lock);
541 inode->i_state |= I_WB_SWITCH;
543 spin_unlock(&inode->i_lock);
549 * inode_switch_wbs - change the wb association of an inode
550 * @inode: target inode
551 * @new_wb_id: ID of the new wb
553 * Switch @inode's wb association to the wb identified by @new_wb_id. The
554 * switching is performed asynchronously and may fail silently.
556 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
558 struct backing_dev_info *bdi = inode_to_bdi(inode);
559 struct cgroup_subsys_state *memcg_css;
560 struct inode_switch_wbs_context *isw;
562 /* noop if seems to be already in progress */
563 if (inode->i_state & I_WB_SWITCH)
566 /* avoid queueing a new switch if too many are already in flight */
567 if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
570 isw = kzalloc(sizeof(*isw) + 2 * sizeof(struct inode *), GFP_ATOMIC);
574 atomic_inc(&isw_nr_in_flight);
576 /* find and pin the new wb */
578 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
579 if (memcg_css && !css_tryget(memcg_css))
585 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
590 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
593 isw->inodes[0] = inode;
596 * In addition to synchronizing among switchers, I_WB_SWITCH tells
597 * the RCU protected stat update paths to grab the i_page
598 * lock so that stat transfer can synchronize against them.
599 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
601 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
602 queue_rcu_work(isw_wq, &isw->work);
606 atomic_dec(&isw_nr_in_flight);
613 * cleanup_offline_cgwb - detach associated inodes
616 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
617 * to eventually release the dying @wb. Returns %true if not all inodes were
618 * switched and the function has to be restarted.
620 bool cleanup_offline_cgwb(struct bdi_writeback *wb)
622 struct cgroup_subsys_state *memcg_css;
623 struct inode_switch_wbs_context *isw;
626 bool restart = false;
628 isw = kzalloc(sizeof(*isw) + WB_MAX_INODES_PER_ISW *
629 sizeof(struct inode *), GFP_KERNEL);
633 atomic_inc(&isw_nr_in_flight);
635 for (memcg_css = wb->memcg_css->parent; memcg_css;
636 memcg_css = memcg_css->parent) {
637 isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
641 if (unlikely(!isw->new_wb))
642 isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
645 spin_lock(&wb->list_lock);
646 list_for_each_entry(inode, &wb->b_attached, i_io_list) {
647 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
650 isw->inodes[nr++] = inode;
652 if (nr >= WB_MAX_INODES_PER_ISW - 1) {
657 spin_unlock(&wb->list_lock);
659 /* no attached inodes? bail out */
661 atomic_dec(&isw_nr_in_flight);
668 * In addition to synchronizing among switchers, I_WB_SWITCH tells
669 * the RCU protected stat update paths to grab the i_page
670 * lock so that stat transfer can synchronize against them.
671 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
673 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
674 queue_rcu_work(isw_wq, &isw->work);
680 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
681 * @wbc: writeback_control of interest
682 * @inode: target inode
684 * @inode is locked and about to be written back under the control of @wbc.
685 * Record @inode's writeback context into @wbc and unlock the i_lock. On
686 * writeback completion, wbc_detach_inode() should be called. This is used
687 * to track the cgroup writeback context.
689 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
692 if (!inode_cgwb_enabled(inode)) {
693 spin_unlock(&inode->i_lock);
697 wbc->wb = inode_to_wb(inode);
700 wbc->wb_id = wbc->wb->memcg_css->id;
701 wbc->wb_lcand_id = inode->i_wb_frn_winner;
702 wbc->wb_tcand_id = 0;
704 wbc->wb_lcand_bytes = 0;
705 wbc->wb_tcand_bytes = 0;
708 spin_unlock(&inode->i_lock);
711 * A dying wb indicates that either the blkcg associated with the
712 * memcg changed or the associated memcg is dying. In the first
713 * case, a replacement wb should already be available and we should
714 * refresh the wb immediately. In the second case, trying to
715 * refresh will keep failing.
717 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
718 inode_switch_wbs(inode, wbc->wb_id);
720 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
723 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
724 * @wbc: writeback_control of the just finished writeback
726 * To be called after a writeback attempt of an inode finishes and undoes
727 * wbc_attach_and_unlock_inode(). Can be called under any context.
729 * As concurrent write sharing of an inode is expected to be very rare and
730 * memcg only tracks page ownership on first-use basis severely confining
731 * the usefulness of such sharing, cgroup writeback tracks ownership
732 * per-inode. While the support for concurrent write sharing of an inode
733 * is deemed unnecessary, an inode being written to by different cgroups at
734 * different points in time is a lot more common, and, more importantly,
735 * charging only by first-use can too readily lead to grossly incorrect
736 * behaviors (single foreign page can lead to gigabytes of writeback to be
737 * incorrectly attributed).
739 * To resolve this issue, cgroup writeback detects the majority dirtier of
740 * an inode and transfers the ownership to it. To avoid unnnecessary
741 * oscillation, the detection mechanism keeps track of history and gives
742 * out the switch verdict only if the foreign usage pattern is stable over
743 * a certain amount of time and/or writeback attempts.
745 * On each writeback attempt, @wbc tries to detect the majority writer
746 * using Boyer-Moore majority vote algorithm. In addition to the byte
747 * count from the majority voting, it also counts the bytes written for the
748 * current wb and the last round's winner wb (max of last round's current
749 * wb, the winner from two rounds ago, and the last round's majority
750 * candidate). Keeping track of the historical winner helps the algorithm
751 * to semi-reliably detect the most active writer even when it's not the
754 * Once the winner of the round is determined, whether the winner is
755 * foreign or not and how much IO time the round consumed is recorded in
756 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
757 * over a certain threshold, the switch verdict is given.
759 void wbc_detach_inode(struct writeback_control *wbc)
761 struct bdi_writeback *wb = wbc->wb;
762 struct inode *inode = wbc->inode;
763 unsigned long avg_time, max_bytes, max_time;
770 history = inode->i_wb_frn_history;
771 avg_time = inode->i_wb_frn_avg_time;
773 /* pick the winner of this round */
774 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
775 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
777 max_bytes = wbc->wb_bytes;
778 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
779 max_id = wbc->wb_lcand_id;
780 max_bytes = wbc->wb_lcand_bytes;
782 max_id = wbc->wb_tcand_id;
783 max_bytes = wbc->wb_tcand_bytes;
787 * Calculate the amount of IO time the winner consumed and fold it
788 * into the running average kept per inode. If the consumed IO
789 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
790 * deciding whether to switch or not. This is to prevent one-off
791 * small dirtiers from skewing the verdict.
793 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
794 wb->avg_write_bandwidth);
796 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
797 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
799 avg_time = max_time; /* immediate catch up on first run */
801 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
805 * The switch verdict is reached if foreign wb's consume
806 * more than a certain proportion of IO time in a
807 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
808 * history mask where each bit represents one sixteenth of
809 * the period. Determine the number of slots to shift into
810 * history from @max_time.
812 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
813 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
815 if (wbc->wb_id != max_id)
816 history |= (1U << slots) - 1;
819 trace_inode_foreign_history(inode, wbc, history);
822 * Switch if the current wb isn't the consistent winner.
823 * If there are multiple closely competing dirtiers, the
824 * inode may switch across them repeatedly over time, which
825 * is okay. The main goal is avoiding keeping an inode on
826 * the wrong wb for an extended period of time.
828 if (hweight16(history) > WB_FRN_HIST_THR_SLOTS)
829 inode_switch_wbs(inode, max_id);
833 * Multiple instances of this function may race to update the
834 * following fields but we don't mind occassional inaccuracies.
836 inode->i_wb_frn_winner = max_id;
837 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
838 inode->i_wb_frn_history = history;
843 EXPORT_SYMBOL_GPL(wbc_detach_inode);
846 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
847 * @wbc: writeback_control of the writeback in progress
848 * @page: page being written out
849 * @bytes: number of bytes being written out
851 * @bytes from @page are about to written out during the writeback
852 * controlled by @wbc. Keep the book for foreign inode detection. See
853 * wbc_detach_inode().
855 void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
858 struct cgroup_subsys_state *css;
862 * pageout() path doesn't attach @wbc to the inode being written
863 * out. This is intentional as we don't want the function to block
864 * behind a slow cgroup. Ultimately, we want pageout() to kick off
865 * regular writeback instead of writing things out itself.
867 if (!wbc->wb || wbc->no_cgroup_owner)
870 css = mem_cgroup_css_from_page(page);
871 /* dead cgroups shouldn't contribute to inode ownership arbitration */
872 if (!(css->flags & CSS_ONLINE))
877 if (id == wbc->wb_id) {
878 wbc->wb_bytes += bytes;
882 if (id == wbc->wb_lcand_id)
883 wbc->wb_lcand_bytes += bytes;
885 /* Boyer-Moore majority vote algorithm */
886 if (!wbc->wb_tcand_bytes)
887 wbc->wb_tcand_id = id;
888 if (id == wbc->wb_tcand_id)
889 wbc->wb_tcand_bytes += bytes;
891 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
893 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
896 * inode_congested - test whether an inode is congested
897 * @inode: inode to test for congestion (may be NULL)
898 * @cong_bits: mask of WB_[a]sync_congested bits to test
900 * Tests whether @inode is congested. @cong_bits is the mask of congestion
901 * bits to test and the return value is the mask of set bits.
903 * If cgroup writeback is enabled for @inode, the congestion state is
904 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
905 * associated with @inode is congested; otherwise, the root wb's congestion
908 * @inode is allowed to be NULL as this function is often called on
909 * mapping->host which is NULL for the swapper space.
911 int inode_congested(struct inode *inode, int cong_bits)
914 * Once set, ->i_wb never becomes NULL while the inode is alive.
915 * Start transaction iff ->i_wb is visible.
917 if (inode && inode_to_wb_is_valid(inode)) {
918 struct bdi_writeback *wb;
919 struct wb_lock_cookie lock_cookie = {};
922 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
923 congested = wb_congested(wb, cong_bits);
924 unlocked_inode_to_wb_end(inode, &lock_cookie);
928 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
930 EXPORT_SYMBOL_GPL(inode_congested);
933 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
934 * @wb: target bdi_writeback to split @nr_pages to
935 * @nr_pages: number of pages to write for the whole bdi
937 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
938 * relation to the total write bandwidth of all wb's w/ dirty inodes on
941 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
943 unsigned long this_bw = wb->avg_write_bandwidth;
944 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
946 if (nr_pages == LONG_MAX)
950 * This may be called on clean wb's and proportional distribution
951 * may not make sense, just use the original @nr_pages in those
952 * cases. In general, we wanna err on the side of writing more.
954 if (!tot_bw || this_bw >= tot_bw)
957 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
961 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
962 * @bdi: target backing_dev_info
963 * @base_work: wb_writeback_work to issue
964 * @skip_if_busy: skip wb's which already have writeback in progress
966 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
967 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
968 * distributed to the busy wbs according to each wb's proportion in the
969 * total active write bandwidth of @bdi.
971 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
972 struct wb_writeback_work *base_work,
975 struct bdi_writeback *last_wb = NULL;
976 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
977 struct bdi_writeback, bdi_node);
982 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
983 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
984 struct wb_writeback_work fallback_work;
985 struct wb_writeback_work *work;
993 /* SYNC_ALL writes out I_DIRTY_TIME too */
994 if (!wb_has_dirty_io(wb) &&
995 (base_work->sync_mode == WB_SYNC_NONE ||
996 list_empty(&wb->b_dirty_time)))
998 if (skip_if_busy && writeback_in_progress(wb))
1001 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
1003 work = kmalloc(sizeof(*work), GFP_ATOMIC);
1006 work->nr_pages = nr_pages;
1007 work->auto_free = 1;
1008 wb_queue_work(wb, work);
1013 * If wb_tryget fails, the wb has been shutdown, skip it.
1015 * Pin @wb so that it stays on @bdi->wb_list. This allows
1016 * continuing iteration from @wb after dropping and
1017 * regrabbing rcu read lock.
1022 /* alloc failed, execute synchronously using on-stack fallback */
1023 work = &fallback_work;
1025 work->nr_pages = nr_pages;
1026 work->auto_free = 0;
1027 work->done = &fallback_work_done;
1029 wb_queue_work(wb, work);
1033 wb_wait_for_completion(&fallback_work_done);
1043 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1044 * @bdi_id: target bdi id
1045 * @memcg_id: target memcg css id
1046 * @reason: reason why some writeback work initiated
1047 * @done: target wb_completion
1049 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1050 * with the specified parameters.
1052 int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
1053 enum wb_reason reason, struct wb_completion *done)
1055 struct backing_dev_info *bdi;
1056 struct cgroup_subsys_state *memcg_css;
1057 struct bdi_writeback *wb;
1058 struct wb_writeback_work *work;
1059 unsigned long dirty;
1062 /* lookup bdi and memcg */
1063 bdi = bdi_get_by_id(bdi_id);
1068 memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1069 if (memcg_css && !css_tryget(memcg_css))
1078 * And find the associated wb. If the wb isn't there already
1079 * there's nothing to flush, don't create one.
1081 wb = wb_get_lookup(bdi, memcg_css);
1088 * The caller is attempting to write out most of
1089 * the currently dirty pages. Let's take the current dirty page
1090 * count and inflate it by 25% which should be large enough to
1091 * flush out most dirty pages while avoiding getting livelocked by
1092 * concurrent dirtiers.
1094 * BTW the memcg stats are flushed periodically and this is best-effort
1095 * estimation, so some potential error is ok.
1097 dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1098 dirty = dirty * 10 / 8;
1100 /* issue the writeback work */
1101 work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1103 work->nr_pages = dirty;
1104 work->sync_mode = WB_SYNC_NONE;
1105 work->range_cyclic = 1;
1106 work->reason = reason;
1108 work->auto_free = 1;
1109 wb_queue_work(wb, work);
1124 * cgroup_writeback_umount - flush inode wb switches for umount
1126 * This function is called when a super_block is about to be destroyed and
1127 * flushes in-flight inode wb switches. An inode wb switch goes through
1128 * RCU and then workqueue, so the two need to be flushed in order to ensure
1129 * that all previously scheduled switches are finished. As wb switches are
1130 * rare occurrences and synchronize_rcu() can take a while, perform
1131 * flushing iff wb switches are in flight.
1133 void cgroup_writeback_umount(void)
1136 * SB_ACTIVE should be reliably cleared before checking
1137 * isw_nr_in_flight, see generic_shutdown_super().
1141 if (atomic_read(&isw_nr_in_flight)) {
1143 * Use rcu_barrier() to wait for all pending callbacks to
1144 * ensure that all in-flight wb switches are in the workqueue.
1147 flush_workqueue(isw_wq);
1151 static int __init cgroup_writeback_init(void)
1153 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1158 fs_initcall(cgroup_writeback_init);
1160 #else /* CONFIG_CGROUP_WRITEBACK */
1162 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1163 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1165 static void inode_cgwb_move_to_attached(struct inode *inode,
1166 struct bdi_writeback *wb)
1168 assert_spin_locked(&wb->list_lock);
1169 assert_spin_locked(&inode->i_lock);
1171 inode->i_state &= ~I_SYNC_QUEUED;
1172 list_del_init(&inode->i_io_list);
1173 wb_io_lists_depopulated(wb);
1176 static struct bdi_writeback *
1177 locked_inode_to_wb_and_lock_list(struct inode *inode)
1178 __releases(&inode->i_lock)
1179 __acquires(&wb->list_lock)
1181 struct bdi_writeback *wb = inode_to_wb(inode);
1183 spin_unlock(&inode->i_lock);
1184 spin_lock(&wb->list_lock);
1188 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1189 __acquires(&wb->list_lock)
1191 struct bdi_writeback *wb = inode_to_wb(inode);
1193 spin_lock(&wb->list_lock);
1197 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1202 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1203 struct wb_writeback_work *base_work,
1208 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1209 base_work->auto_free = 0;
1210 wb_queue_work(&bdi->wb, base_work);
1214 #endif /* CONFIG_CGROUP_WRITEBACK */
1217 * Add in the number of potentially dirty inodes, because each inode
1218 * write can dirty pagecache in the underlying blockdev.
1220 static unsigned long get_nr_dirty_pages(void)
1222 return global_node_page_state(NR_FILE_DIRTY) +
1223 get_nr_dirty_inodes();
1226 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1228 if (!wb_has_dirty_io(wb))
1232 * All callers of this function want to start writeback of all
1233 * dirty pages. Places like vmscan can call this at a very
1234 * high frequency, causing pointless allocations of tons of
1235 * work items and keeping the flusher threads busy retrieving
1236 * that work. Ensure that we only allow one of them pending and
1237 * inflight at the time.
1239 if (test_bit(WB_start_all, &wb->state) ||
1240 test_and_set_bit(WB_start_all, &wb->state))
1243 wb->start_all_reason = reason;
1248 * wb_start_background_writeback - start background writeback
1249 * @wb: bdi_writback to write from
1252 * This makes sure WB_SYNC_NONE background writeback happens. When
1253 * this function returns, it is only guaranteed that for given wb
1254 * some IO is happening if we are over background dirty threshold.
1255 * Caller need not hold sb s_umount semaphore.
1257 void wb_start_background_writeback(struct bdi_writeback *wb)
1260 * We just wake up the flusher thread. It will perform background
1261 * writeback as soon as there is no other work to do.
1263 trace_writeback_wake_background(wb);
1268 * Remove the inode from the writeback list it is on.
1270 void inode_io_list_del(struct inode *inode)
1272 struct bdi_writeback *wb;
1274 wb = inode_to_wb_and_lock_list(inode);
1275 spin_lock(&inode->i_lock);
1277 inode->i_state &= ~I_SYNC_QUEUED;
1278 list_del_init(&inode->i_io_list);
1279 wb_io_lists_depopulated(wb);
1281 spin_unlock(&inode->i_lock);
1282 spin_unlock(&wb->list_lock);
1284 EXPORT_SYMBOL(inode_io_list_del);
1287 * mark an inode as under writeback on the sb
1289 void sb_mark_inode_writeback(struct inode *inode)
1291 struct super_block *sb = inode->i_sb;
1292 unsigned long flags;
1294 if (list_empty(&inode->i_wb_list)) {
1295 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1296 if (list_empty(&inode->i_wb_list)) {
1297 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1298 trace_sb_mark_inode_writeback(inode);
1300 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1305 * clear an inode as under writeback on the sb
1307 void sb_clear_inode_writeback(struct inode *inode)
1309 struct super_block *sb = inode->i_sb;
1310 unsigned long flags;
1312 if (!list_empty(&inode->i_wb_list)) {
1313 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1314 if (!list_empty(&inode->i_wb_list)) {
1315 list_del_init(&inode->i_wb_list);
1316 trace_sb_clear_inode_writeback(inode);
1318 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1323 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1324 * furthest end of its superblock's dirty-inode list.
1326 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1327 * already the most-recently-dirtied inode on the b_dirty list. If that is
1328 * the case then the inode must have been redirtied while it was being written
1329 * out and we don't reset its dirtied_when.
1331 static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1333 assert_spin_locked(&inode->i_lock);
1335 if (!list_empty(&wb->b_dirty)) {
1338 tail = wb_inode(wb->b_dirty.next);
1339 if (time_before(inode->dirtied_when, tail->dirtied_when))
1340 inode->dirtied_when = jiffies;
1342 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1343 inode->i_state &= ~I_SYNC_QUEUED;
1346 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1348 spin_lock(&inode->i_lock);
1349 redirty_tail_locked(inode, wb);
1350 spin_unlock(&inode->i_lock);
1354 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1356 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1358 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1361 static void inode_sync_complete(struct inode *inode)
1363 inode->i_state &= ~I_SYNC;
1364 /* If inode is clean an unused, put it into LRU now... */
1365 inode_add_lru(inode);
1366 /* Waiters must see I_SYNC cleared before being woken up */
1368 wake_up_bit(&inode->i_state, __I_SYNC);
1371 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1373 bool ret = time_after(inode->dirtied_when, t);
1374 #ifndef CONFIG_64BIT
1376 * For inodes being constantly redirtied, dirtied_when can get stuck.
1377 * It _appears_ to be in the future, but is actually in distant past.
1378 * This test is necessary to prevent such wrapped-around relative times
1379 * from permanently stopping the whole bdi writeback.
1381 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1386 #define EXPIRE_DIRTY_ATIME 0x0001
1389 * Move expired (dirtied before dirtied_before) dirty inodes from
1390 * @delaying_queue to @dispatch_queue.
1392 static int move_expired_inodes(struct list_head *delaying_queue,
1393 struct list_head *dispatch_queue,
1394 unsigned long dirtied_before)
1397 struct list_head *pos, *node;
1398 struct super_block *sb = NULL;
1399 struct inode *inode;
1403 while (!list_empty(delaying_queue)) {
1404 inode = wb_inode(delaying_queue->prev);
1405 if (inode_dirtied_after(inode, dirtied_before))
1407 spin_lock(&inode->i_lock);
1408 list_move(&inode->i_io_list, &tmp);
1410 inode->i_state |= I_SYNC_QUEUED;
1411 spin_unlock(&inode->i_lock);
1412 if (sb_is_blkdev_sb(inode->i_sb))
1414 if (sb && sb != inode->i_sb)
1419 /* just one sb in list, splice to dispatch_queue and we're done */
1421 list_splice(&tmp, dispatch_queue);
1426 * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1427 * we don't take inode->i_lock here because it is just a pointless overhead.
1428 * Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1429 * fully under our control.
1431 while (!list_empty(&tmp)) {
1432 sb = wb_inode(tmp.prev)->i_sb;
1433 list_for_each_prev_safe(pos, node, &tmp) {
1434 inode = wb_inode(pos);
1435 if (inode->i_sb == sb)
1436 list_move(&inode->i_io_list, dispatch_queue);
1444 * Queue all expired dirty inodes for io, eldest first.
1446 * newly dirtied b_dirty b_io b_more_io
1447 * =============> gf edc BA
1449 * newly dirtied b_dirty b_io b_more_io
1450 * =============> g fBAedc
1452 * +--> dequeue for IO
1454 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1455 unsigned long dirtied_before)
1458 unsigned long time_expire_jif = dirtied_before;
1460 assert_spin_locked(&wb->list_lock);
1461 list_splice_init(&wb->b_more_io, &wb->b_io);
1462 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1463 if (!work->for_sync)
1464 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1465 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1468 wb_io_lists_populated(wb);
1469 trace_writeback_queue_io(wb, work, dirtied_before, moved);
1472 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1476 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1477 trace_writeback_write_inode_start(inode, wbc);
1478 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1479 trace_writeback_write_inode(inode, wbc);
1486 * Wait for writeback on an inode to complete. Called with i_lock held.
1487 * Caller must make sure inode cannot go away when we drop i_lock.
1489 static void __inode_wait_for_writeback(struct inode *inode)
1490 __releases(inode->i_lock)
1491 __acquires(inode->i_lock)
1493 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1494 wait_queue_head_t *wqh;
1496 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1497 while (inode->i_state & I_SYNC) {
1498 spin_unlock(&inode->i_lock);
1499 __wait_on_bit(wqh, &wq, bit_wait,
1500 TASK_UNINTERRUPTIBLE);
1501 spin_lock(&inode->i_lock);
1506 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1508 void inode_wait_for_writeback(struct inode *inode)
1510 spin_lock(&inode->i_lock);
1511 __inode_wait_for_writeback(inode);
1512 spin_unlock(&inode->i_lock);
1516 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1517 * held and drops it. It is aimed for callers not holding any inode reference
1518 * so once i_lock is dropped, inode can go away.
1520 static void inode_sleep_on_writeback(struct inode *inode)
1521 __releases(inode->i_lock)
1524 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1527 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1528 sleep = inode->i_state & I_SYNC;
1529 spin_unlock(&inode->i_lock);
1532 finish_wait(wqh, &wait);
1536 * Find proper writeback list for the inode depending on its current state and
1537 * possibly also change of its state while we were doing writeback. Here we
1538 * handle things such as livelock prevention or fairness of writeback among
1539 * inodes. This function can be called only by flusher thread - noone else
1540 * processes all inodes in writeback lists and requeueing inodes behind flusher
1541 * thread's back can have unexpected consequences.
1543 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1544 struct writeback_control *wbc)
1546 if (inode->i_state & I_FREEING)
1550 * Sync livelock prevention. Each inode is tagged and synced in one
1551 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1552 * the dirty time to prevent enqueue and sync it again.
1554 if ((inode->i_state & I_DIRTY) &&
1555 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1556 inode->dirtied_when = jiffies;
1558 if (wbc->pages_skipped) {
1560 * Writeback is not making progress due to locked buffers.
1561 * Skip this inode for now. Although having skipped pages
1562 * is odd for clean inodes, it can happen for some
1563 * filesystems so handle that gracefully.
1565 if (inode->i_state & I_DIRTY_ALL)
1566 redirty_tail_locked(inode, wb);
1568 inode_cgwb_move_to_attached(inode, wb);
1572 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1574 * We didn't write back all the pages. nfs_writepages()
1575 * sometimes bales out without doing anything.
1577 if (wbc->nr_to_write <= 0) {
1578 /* Slice used up. Queue for next turn. */
1579 requeue_io(inode, wb);
1582 * Writeback blocked by something other than
1583 * congestion. Delay the inode for some time to
1584 * avoid spinning on the CPU (100% iowait)
1585 * retrying writeback of the dirty page/inode
1586 * that cannot be performed immediately.
1588 redirty_tail_locked(inode, wb);
1590 } else if (inode->i_state & I_DIRTY) {
1592 * Filesystems can dirty the inode during writeback operations,
1593 * such as delayed allocation during submission or metadata
1594 * updates after data IO completion.
1596 redirty_tail_locked(inode, wb);
1597 } else if (inode->i_state & I_DIRTY_TIME) {
1598 inode->dirtied_when = jiffies;
1599 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1600 inode->i_state &= ~I_SYNC_QUEUED;
1602 /* The inode is clean. Remove from writeback lists. */
1603 inode_cgwb_move_to_attached(inode, wb);
1608 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1609 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1611 * This doesn't remove the inode from the writeback list it is on, except
1612 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1613 * expiration. The caller is otherwise responsible for writeback list handling.
1615 * The caller is also responsible for setting the I_SYNC flag beforehand and
1616 * calling inode_sync_complete() to clear it afterwards.
1619 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1621 struct address_space *mapping = inode->i_mapping;
1622 long nr_to_write = wbc->nr_to_write;
1626 WARN_ON(!(inode->i_state & I_SYNC));
1628 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1630 ret = do_writepages(mapping, wbc);
1633 * Make sure to wait on the data before writing out the metadata.
1634 * This is important for filesystems that modify metadata on data
1635 * I/O completion. We don't do it for sync(2) writeback because it has a
1636 * separate, external IO completion path and ->sync_fs for guaranteeing
1637 * inode metadata is written back correctly.
1639 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1640 int err = filemap_fdatawait(mapping);
1646 * If the inode has dirty timestamps and we need to write them, call
1647 * mark_inode_dirty_sync() to notify the filesystem about it and to
1648 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1650 if ((inode->i_state & I_DIRTY_TIME) &&
1651 (wbc->sync_mode == WB_SYNC_ALL ||
1652 time_after(jiffies, inode->dirtied_time_when +
1653 dirtytime_expire_interval * HZ))) {
1654 trace_writeback_lazytime(inode);
1655 mark_inode_dirty_sync(inode);
1659 * Get and clear the dirty flags from i_state. This needs to be done
1660 * after calling writepages because some filesystems may redirty the
1661 * inode during writepages due to delalloc. It also needs to be done
1662 * after handling timestamp expiration, as that may dirty the inode too.
1664 spin_lock(&inode->i_lock);
1665 dirty = inode->i_state & I_DIRTY;
1666 inode->i_state &= ~dirty;
1669 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1670 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1671 * either they see the I_DIRTY bits cleared or we see the dirtied
1674 * I_DIRTY_PAGES is always cleared together above even if @mapping
1675 * still has dirty pages. The flag is reinstated after smp_mb() if
1676 * necessary. This guarantees that either __mark_inode_dirty()
1677 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1681 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1682 inode->i_state |= I_DIRTY_PAGES;
1684 spin_unlock(&inode->i_lock);
1686 /* Don't write the inode if only I_DIRTY_PAGES was set */
1687 if (dirty & ~I_DIRTY_PAGES) {
1688 int err = write_inode(inode, wbc);
1692 trace_writeback_single_inode(inode, wbc, nr_to_write);
1697 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1698 * the regular batched writeback done by the flusher threads in
1699 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1700 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1702 * To prevent the inode from going away, either the caller must have a reference
1703 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1705 static int writeback_single_inode(struct inode *inode,
1706 struct writeback_control *wbc)
1708 struct bdi_writeback *wb;
1711 spin_lock(&inode->i_lock);
1712 if (!atomic_read(&inode->i_count))
1713 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1715 WARN_ON(inode->i_state & I_WILL_FREE);
1717 if (inode->i_state & I_SYNC) {
1719 * Writeback is already running on the inode. For WB_SYNC_NONE,
1720 * that's enough and we can just return. For WB_SYNC_ALL, we
1721 * must wait for the existing writeback to complete, then do
1722 * writeback again if there's anything left.
1724 if (wbc->sync_mode != WB_SYNC_ALL)
1726 __inode_wait_for_writeback(inode);
1728 WARN_ON(inode->i_state & I_SYNC);
1730 * If the inode is already fully clean, then there's nothing to do.
1732 * For data-integrity syncs we also need to check whether any pages are
1733 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1734 * there are any such pages, we'll need to wait for them.
1736 if (!(inode->i_state & I_DIRTY_ALL) &&
1737 (wbc->sync_mode != WB_SYNC_ALL ||
1738 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1740 inode->i_state |= I_SYNC;
1741 wbc_attach_and_unlock_inode(wbc, inode);
1743 ret = __writeback_single_inode(inode, wbc);
1745 wbc_detach_inode(wbc);
1747 wb = inode_to_wb_and_lock_list(inode);
1748 spin_lock(&inode->i_lock);
1750 * If the inode is freeing, its i_io_list shoudn't be updated
1751 * as it can be finally deleted at this moment.
1753 if (!(inode->i_state & I_FREEING)) {
1755 * If the inode is now fully clean, then it can be safely
1756 * removed from its writeback list (if any). Otherwise the
1757 * flusher threads are responsible for the writeback lists.
1759 if (!(inode->i_state & I_DIRTY_ALL))
1760 inode_cgwb_move_to_attached(inode, wb);
1761 else if (!(inode->i_state & I_SYNC_QUEUED)) {
1762 if ((inode->i_state & I_DIRTY))
1763 redirty_tail_locked(inode, wb);
1764 else if (inode->i_state & I_DIRTY_TIME) {
1765 inode->dirtied_when = jiffies;
1766 inode_io_list_move_locked(inode,
1773 spin_unlock(&wb->list_lock);
1774 inode_sync_complete(inode);
1776 spin_unlock(&inode->i_lock);
1780 static long writeback_chunk_size(struct bdi_writeback *wb,
1781 struct wb_writeback_work *work)
1786 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1787 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1788 * here avoids calling into writeback_inodes_wb() more than once.
1790 * The intended call sequence for WB_SYNC_ALL writeback is:
1793 * writeback_sb_inodes() <== called only once
1794 * write_cache_pages() <== called once for each inode
1795 * (quickly) tag currently dirty pages
1796 * (maybe slowly) sync all tagged pages
1798 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1801 pages = min(wb->avg_write_bandwidth / 2,
1802 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1803 pages = min(pages, work->nr_pages);
1804 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1805 MIN_WRITEBACK_PAGES);
1812 * Write a portion of b_io inodes which belong to @sb.
1814 * Return the number of pages and/or inodes written.
1816 * NOTE! This is called with wb->list_lock held, and will
1817 * unlock and relock that for each inode it ends up doing
1820 static long writeback_sb_inodes(struct super_block *sb,
1821 struct bdi_writeback *wb,
1822 struct wb_writeback_work *work)
1824 struct writeback_control wbc = {
1825 .sync_mode = work->sync_mode,
1826 .tagged_writepages = work->tagged_writepages,
1827 .for_kupdate = work->for_kupdate,
1828 .for_background = work->for_background,
1829 .for_sync = work->for_sync,
1830 .range_cyclic = work->range_cyclic,
1832 .range_end = LLONG_MAX,
1834 unsigned long start_time = jiffies;
1836 long total_wrote = 0; /* count both pages and inodes */
1838 while (!list_empty(&wb->b_io)) {
1839 struct inode *inode = wb_inode(wb->b_io.prev);
1840 struct bdi_writeback *tmp_wb;
1843 if (inode->i_sb != sb) {
1846 * We only want to write back data for this
1847 * superblock, move all inodes not belonging
1848 * to it back onto the dirty list.
1850 redirty_tail(inode, wb);
1855 * The inode belongs to a different superblock.
1856 * Bounce back to the caller to unpin this and
1857 * pin the next superblock.
1863 * Don't bother with new inodes or inodes being freed, first
1864 * kind does not need periodic writeout yet, and for the latter
1865 * kind writeout is handled by the freer.
1867 spin_lock(&inode->i_lock);
1868 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1869 redirty_tail_locked(inode, wb);
1870 spin_unlock(&inode->i_lock);
1873 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1875 * If this inode is locked for writeback and we are not
1876 * doing writeback-for-data-integrity, move it to
1877 * b_more_io so that writeback can proceed with the
1878 * other inodes on s_io.
1880 * We'll have another go at writing back this inode
1881 * when we completed a full scan of b_io.
1883 requeue_io(inode, wb);
1884 spin_unlock(&inode->i_lock);
1885 trace_writeback_sb_inodes_requeue(inode);
1888 spin_unlock(&wb->list_lock);
1891 * We already requeued the inode if it had I_SYNC set and we
1892 * are doing WB_SYNC_NONE writeback. So this catches only the
1895 if (inode->i_state & I_SYNC) {
1896 /* Wait for I_SYNC. This function drops i_lock... */
1897 inode_sleep_on_writeback(inode);
1898 /* Inode may be gone, start again */
1899 spin_lock(&wb->list_lock);
1902 inode->i_state |= I_SYNC;
1903 wbc_attach_and_unlock_inode(&wbc, inode);
1905 write_chunk = writeback_chunk_size(wb, work);
1906 wbc.nr_to_write = write_chunk;
1907 wbc.pages_skipped = 0;
1910 * We use I_SYNC to pin the inode in memory. While it is set
1911 * evict_inode() will wait so the inode cannot be freed.
1913 __writeback_single_inode(inode, &wbc);
1915 wbc_detach_inode(&wbc);
1916 work->nr_pages -= write_chunk - wbc.nr_to_write;
1917 wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
1918 wrote = wrote < 0 ? 0 : wrote;
1919 total_wrote += wrote;
1921 if (need_resched()) {
1923 * We're trying to balance between building up a nice
1924 * long list of IOs to improve our merge rate, and
1925 * getting those IOs out quickly for anyone throttling
1926 * in balance_dirty_pages(). cond_resched() doesn't
1927 * unplug, so get our IOs out the door before we
1930 blk_flush_plug(current);
1935 * Requeue @inode if still dirty. Be careful as @inode may
1936 * have been switched to another wb in the meantime.
1938 tmp_wb = inode_to_wb_and_lock_list(inode);
1939 spin_lock(&inode->i_lock);
1940 if (!(inode->i_state & I_DIRTY_ALL))
1942 requeue_inode(inode, tmp_wb, &wbc);
1943 inode_sync_complete(inode);
1944 spin_unlock(&inode->i_lock);
1946 if (unlikely(tmp_wb != wb)) {
1947 spin_unlock(&tmp_wb->list_lock);
1948 spin_lock(&wb->list_lock);
1952 * bail out to wb_writeback() often enough to check
1953 * background threshold and other termination conditions.
1956 if (time_is_before_jiffies(start_time + HZ / 10UL))
1958 if (work->nr_pages <= 0)
1965 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1966 struct wb_writeback_work *work)
1968 unsigned long start_time = jiffies;
1971 while (!list_empty(&wb->b_io)) {
1972 struct inode *inode = wb_inode(wb->b_io.prev);
1973 struct super_block *sb = inode->i_sb;
1975 if (!trylock_super(sb)) {
1977 * trylock_super() may fail consistently due to
1978 * s_umount being grabbed by someone else. Don't use
1979 * requeue_io() to avoid busy retrying the inode/sb.
1981 redirty_tail(inode, wb);
1984 wrote += writeback_sb_inodes(sb, wb, work);
1985 up_read(&sb->s_umount);
1987 /* refer to the same tests at the end of writeback_sb_inodes */
1989 if (time_is_before_jiffies(start_time + HZ / 10UL))
1991 if (work->nr_pages <= 0)
1995 /* Leave any unwritten inodes on b_io */
1999 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
2000 enum wb_reason reason)
2002 struct wb_writeback_work work = {
2003 .nr_pages = nr_pages,
2004 .sync_mode = WB_SYNC_NONE,
2008 struct blk_plug plug;
2010 blk_start_plug(&plug);
2011 spin_lock(&wb->list_lock);
2012 if (list_empty(&wb->b_io))
2013 queue_io(wb, &work, jiffies);
2014 __writeback_inodes_wb(wb, &work);
2015 spin_unlock(&wb->list_lock);
2016 blk_finish_plug(&plug);
2018 return nr_pages - work.nr_pages;
2022 * Explicit flushing or periodic writeback of "old" data.
2024 * Define "old": the first time one of an inode's pages is dirtied, we mark the
2025 * dirtying-time in the inode's address_space. So this periodic writeback code
2026 * just walks the superblock inode list, writing back any inodes which are
2027 * older than a specific point in time.
2029 * Try to run once per dirty_writeback_interval. But if a writeback event
2030 * takes longer than a dirty_writeback_interval interval, then leave a
2033 * dirtied_before takes precedence over nr_to_write. So we'll only write back
2034 * all dirty pages if they are all attached to "old" mappings.
2036 static long wb_writeback(struct bdi_writeback *wb,
2037 struct wb_writeback_work *work)
2039 long nr_pages = work->nr_pages;
2040 unsigned long dirtied_before = jiffies;
2041 struct inode *inode;
2043 struct blk_plug plug;
2045 blk_start_plug(&plug);
2046 spin_lock(&wb->list_lock);
2049 * Stop writeback when nr_pages has been consumed
2051 if (work->nr_pages <= 0)
2055 * Background writeout and kupdate-style writeback may
2056 * run forever. Stop them if there is other work to do
2057 * so that e.g. sync can proceed. They'll be restarted
2058 * after the other works are all done.
2060 if ((work->for_background || work->for_kupdate) &&
2061 !list_empty(&wb->work_list))
2065 * For background writeout, stop when we are below the
2066 * background dirty threshold
2068 if (work->for_background && !wb_over_bg_thresh(wb))
2072 * Kupdate and background works are special and we want to
2073 * include all inodes that need writing. Livelock avoidance is
2074 * handled by these works yielding to any other work so we are
2077 if (work->for_kupdate) {
2078 dirtied_before = jiffies -
2079 msecs_to_jiffies(dirty_expire_interval * 10);
2080 } else if (work->for_background)
2081 dirtied_before = jiffies;
2083 trace_writeback_start(wb, work);
2084 if (list_empty(&wb->b_io))
2085 queue_io(wb, work, dirtied_before);
2087 progress = writeback_sb_inodes(work->sb, wb, work);
2089 progress = __writeback_inodes_wb(wb, work);
2090 trace_writeback_written(wb, work);
2093 * Did we write something? Try for more
2095 * Dirty inodes are moved to b_io for writeback in batches.
2096 * The completion of the current batch does not necessarily
2097 * mean the overall work is done. So we keep looping as long
2098 * as made some progress on cleaning pages or inodes.
2103 * No more inodes for IO, bail
2105 if (list_empty(&wb->b_more_io))
2108 * Nothing written. Wait for some inode to
2109 * become available for writeback. Otherwise
2110 * we'll just busyloop.
2112 trace_writeback_wait(wb, work);
2113 inode = wb_inode(wb->b_more_io.prev);
2114 spin_lock(&inode->i_lock);
2115 spin_unlock(&wb->list_lock);
2116 /* This function drops i_lock... */
2117 inode_sleep_on_writeback(inode);
2118 spin_lock(&wb->list_lock);
2120 spin_unlock(&wb->list_lock);
2121 blk_finish_plug(&plug);
2123 return nr_pages - work->nr_pages;
2127 * Return the next wb_writeback_work struct that hasn't been processed yet.
2129 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2131 struct wb_writeback_work *work = NULL;
2133 spin_lock_irq(&wb->work_lock);
2134 if (!list_empty(&wb->work_list)) {
2135 work = list_entry(wb->work_list.next,
2136 struct wb_writeback_work, list);
2137 list_del_init(&work->list);
2139 spin_unlock_irq(&wb->work_lock);
2143 static long wb_check_background_flush(struct bdi_writeback *wb)
2145 if (wb_over_bg_thresh(wb)) {
2147 struct wb_writeback_work work = {
2148 .nr_pages = LONG_MAX,
2149 .sync_mode = WB_SYNC_NONE,
2150 .for_background = 1,
2152 .reason = WB_REASON_BACKGROUND,
2155 return wb_writeback(wb, &work);
2161 static long wb_check_old_data_flush(struct bdi_writeback *wb)
2163 unsigned long expired;
2167 * When set to zero, disable periodic writeback
2169 if (!dirty_writeback_interval)
2172 expired = wb->last_old_flush +
2173 msecs_to_jiffies(dirty_writeback_interval * 10);
2174 if (time_before(jiffies, expired))
2177 wb->last_old_flush = jiffies;
2178 nr_pages = get_nr_dirty_pages();
2181 struct wb_writeback_work work = {
2182 .nr_pages = nr_pages,
2183 .sync_mode = WB_SYNC_NONE,
2186 .reason = WB_REASON_PERIODIC,
2189 return wb_writeback(wb, &work);
2195 static long wb_check_start_all(struct bdi_writeback *wb)
2199 if (!test_bit(WB_start_all, &wb->state))
2202 nr_pages = get_nr_dirty_pages();
2204 struct wb_writeback_work work = {
2205 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2206 .sync_mode = WB_SYNC_NONE,
2208 .reason = wb->start_all_reason,
2211 nr_pages = wb_writeback(wb, &work);
2214 clear_bit(WB_start_all, &wb->state);
2220 * Retrieve work items and do the writeback they describe
2222 static long wb_do_writeback(struct bdi_writeback *wb)
2224 struct wb_writeback_work *work;
2227 set_bit(WB_writeback_running, &wb->state);
2228 while ((work = get_next_work_item(wb)) != NULL) {
2229 trace_writeback_exec(wb, work);
2230 wrote += wb_writeback(wb, work);
2231 finish_writeback_work(wb, work);
2235 * Check for a flush-everything request
2237 wrote += wb_check_start_all(wb);
2240 * Check for periodic writeback, kupdated() style
2242 wrote += wb_check_old_data_flush(wb);
2243 wrote += wb_check_background_flush(wb);
2244 clear_bit(WB_writeback_running, &wb->state);
2250 * Handle writeback of dirty data for the device backed by this bdi. Also
2251 * reschedules periodically and does kupdated style flushing.
2253 void wb_workfn(struct work_struct *work)
2255 struct bdi_writeback *wb = container_of(to_delayed_work(work),
2256 struct bdi_writeback, dwork);
2259 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2260 current->flags |= PF_SWAPWRITE;
2262 if (likely(!current_is_workqueue_rescuer() ||
2263 !test_bit(WB_registered, &wb->state))) {
2265 * The normal path. Keep writing back @wb until its
2266 * work_list is empty. Note that this path is also taken
2267 * if @wb is shutting down even when we're running off the
2268 * rescuer as work_list needs to be drained.
2271 pages_written = wb_do_writeback(wb);
2272 trace_writeback_pages_written(pages_written);
2273 } while (!list_empty(&wb->work_list));
2276 * bdi_wq can't get enough workers and we're running off
2277 * the emergency worker. Don't hog it. Hopefully, 1024 is
2278 * enough for efficient IO.
2280 pages_written = writeback_inodes_wb(wb, 1024,
2281 WB_REASON_FORKER_THREAD);
2282 trace_writeback_pages_written(pages_written);
2285 if (!list_empty(&wb->work_list))
2287 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2288 wb_wakeup_delayed(wb);
2290 current->flags &= ~PF_SWAPWRITE;
2294 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2295 * write back the whole world.
2297 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2298 enum wb_reason reason)
2300 struct bdi_writeback *wb;
2302 if (!bdi_has_dirty_io(bdi))
2305 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2306 wb_start_writeback(wb, reason);
2309 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2310 enum wb_reason reason)
2313 __wakeup_flusher_threads_bdi(bdi, reason);
2318 * Wakeup the flusher threads to start writeback of all currently dirty pages
2320 void wakeup_flusher_threads(enum wb_reason reason)
2322 struct backing_dev_info *bdi;
2325 * If we are expecting writeback progress we must submit plugged IO.
2327 if (blk_needs_flush_plug(current))
2328 blk_schedule_flush_plug(current);
2331 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2332 __wakeup_flusher_threads_bdi(bdi, reason);
2337 * Wake up bdi's periodically to make sure dirtytime inodes gets
2338 * written back periodically. We deliberately do *not* check the
2339 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2340 * kernel to be constantly waking up once there are any dirtytime
2341 * inodes on the system. So instead we define a separate delayed work
2342 * function which gets called much more rarely. (By default, only
2343 * once every 12 hours.)
2345 * If there is any other write activity going on in the file system,
2346 * this function won't be necessary. But if the only thing that has
2347 * happened on the file system is a dirtytime inode caused by an atime
2348 * update, we need this infrastructure below to make sure that inode
2349 * eventually gets pushed out to disk.
2351 static void wakeup_dirtytime_writeback(struct work_struct *w);
2352 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2354 static void wakeup_dirtytime_writeback(struct work_struct *w)
2356 struct backing_dev_info *bdi;
2359 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2360 struct bdi_writeback *wb;
2362 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2363 if (!list_empty(&wb->b_dirty_time))
2367 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2370 static int __init start_dirtytime_writeback(void)
2372 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2375 __initcall(start_dirtytime_writeback);
2377 int dirtytime_interval_handler(struct ctl_table *table, int write,
2378 void *buffer, size_t *lenp, loff_t *ppos)
2382 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2383 if (ret == 0 && write)
2384 mod_delayed_work(system_wq, &dirtytime_work, 0);
2389 * __mark_inode_dirty - internal function to mark an inode dirty
2391 * @inode: inode to mark
2392 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2393 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2394 * with I_DIRTY_PAGES.
2396 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2397 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2399 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2400 * instead of calling this directly.
2402 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2403 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2404 * even if they are later hashed, as they will have been marked dirty already.
2406 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2408 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2409 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2410 * the kernel-internal blockdev inode represents the dirtying time of the
2411 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2412 * page->mapping->host, so the page-dirtying time is recorded in the internal
2415 void __mark_inode_dirty(struct inode *inode, int flags)
2417 struct super_block *sb = inode->i_sb;
2419 struct bdi_writeback *wb = NULL;
2421 trace_writeback_mark_inode_dirty(inode, flags);
2423 if (flags & I_DIRTY_INODE) {
2425 * Inode timestamp update will piggback on this dirtying.
2426 * We tell ->dirty_inode callback that timestamps need to
2427 * be updated by setting I_DIRTY_TIME in flags.
2429 if (inode->i_state & I_DIRTY_TIME) {
2430 spin_lock(&inode->i_lock);
2431 if (inode->i_state & I_DIRTY_TIME) {
2432 inode->i_state &= ~I_DIRTY_TIME;
2433 flags |= I_DIRTY_TIME;
2435 spin_unlock(&inode->i_lock);
2439 * Notify the filesystem about the inode being dirtied, so that
2440 * (if needed) it can update on-disk fields and journal the
2441 * inode. This is only needed when the inode itself is being
2442 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2443 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2445 trace_writeback_dirty_inode_start(inode, flags);
2446 if (sb->s_op->dirty_inode)
2447 sb->s_op->dirty_inode(inode,
2448 flags & (I_DIRTY_INODE | I_DIRTY_TIME));
2449 trace_writeback_dirty_inode(inode, flags);
2451 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2452 flags &= ~I_DIRTY_TIME;
2455 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2456 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2457 * in one call to __mark_inode_dirty().)
2459 dirtytime = flags & I_DIRTY_TIME;
2460 WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2464 * Paired with smp_mb() in __writeback_single_inode() for the
2465 * following lockless i_state test. See there for details.
2469 if ((inode->i_state & flags) == flags)
2472 spin_lock(&inode->i_lock);
2473 if ((inode->i_state & flags) != flags) {
2474 const int was_dirty = inode->i_state & I_DIRTY;
2476 inode_attach_wb(inode, NULL);
2478 inode->i_state |= flags;
2481 * Grab inode's wb early because it requires dropping i_lock and we
2482 * need to make sure following checks happen atomically with dirty
2483 * list handling so that we don't move inodes under flush worker's
2487 wb = locked_inode_to_wb_and_lock_list(inode);
2488 spin_lock(&inode->i_lock);
2492 * If the inode is queued for writeback by flush worker, just
2493 * update its dirty state. Once the flush worker is done with
2494 * the inode it will place it on the appropriate superblock
2495 * list, based upon its state.
2497 if (inode->i_state & I_SYNC_QUEUED)
2501 * Only add valid (hashed) inodes to the superblock's
2502 * dirty list. Add blockdev inodes as well.
2504 if (!S_ISBLK(inode->i_mode)) {
2505 if (inode_unhashed(inode))
2508 if (inode->i_state & I_FREEING)
2512 * If the inode was already on b_dirty/b_io/b_more_io, don't
2513 * reposition it (that would break b_dirty time-ordering).
2516 struct list_head *dirty_list;
2517 bool wakeup_bdi = false;
2519 inode->dirtied_when = jiffies;
2521 inode->dirtied_time_when = jiffies;
2523 if (inode->i_state & I_DIRTY)
2524 dirty_list = &wb->b_dirty;
2526 dirty_list = &wb->b_dirty_time;
2528 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2531 spin_unlock(&wb->list_lock);
2532 spin_unlock(&inode->i_lock);
2533 trace_writeback_dirty_inode_enqueue(inode);
2536 * If this is the first dirty inode for this bdi,
2537 * we have to wake-up the corresponding bdi thread
2538 * to make sure background write-back happens
2542 (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2543 wb_wakeup_delayed(wb);
2549 spin_unlock(&wb->list_lock);
2550 spin_unlock(&inode->i_lock);
2552 EXPORT_SYMBOL(__mark_inode_dirty);
2555 * The @s_sync_lock is used to serialise concurrent sync operations
2556 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2557 * Concurrent callers will block on the s_sync_lock rather than doing contending
2558 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2559 * has been issued up to the time this function is enter is guaranteed to be
2560 * completed by the time we have gained the lock and waited for all IO that is
2561 * in progress regardless of the order callers are granted the lock.
2563 static void wait_sb_inodes(struct super_block *sb)
2565 LIST_HEAD(sync_list);
2568 * We need to be protected against the filesystem going from
2569 * r/o to r/w or vice versa.
2571 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2573 mutex_lock(&sb->s_sync_lock);
2576 * Splice the writeback list onto a temporary list to avoid waiting on
2577 * inodes that have started writeback after this point.
2579 * Use rcu_read_lock() to keep the inodes around until we have a
2580 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2581 * the local list because inodes can be dropped from either by writeback
2585 spin_lock_irq(&sb->s_inode_wblist_lock);
2586 list_splice_init(&sb->s_inodes_wb, &sync_list);
2589 * Data integrity sync. Must wait for all pages under writeback, because
2590 * there may have been pages dirtied before our sync call, but which had
2591 * writeout started before we write it out. In which case, the inode
2592 * may not be on the dirty list, but we still have to wait for that
2595 while (!list_empty(&sync_list)) {
2596 struct inode *inode = list_first_entry(&sync_list, struct inode,
2598 struct address_space *mapping = inode->i_mapping;
2601 * Move each inode back to the wb list before we drop the lock
2602 * to preserve consistency between i_wb_list and the mapping
2603 * writeback tag. Writeback completion is responsible to remove
2604 * the inode from either list once the writeback tag is cleared.
2606 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2609 * The mapping can appear untagged while still on-list since we
2610 * do not have the mapping lock. Skip it here, wb completion
2613 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2616 spin_unlock_irq(&sb->s_inode_wblist_lock);
2618 spin_lock(&inode->i_lock);
2619 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2620 spin_unlock(&inode->i_lock);
2622 spin_lock_irq(&sb->s_inode_wblist_lock);
2626 spin_unlock(&inode->i_lock);
2630 * We keep the error status of individual mapping so that
2631 * applications can catch the writeback error using fsync(2).
2632 * See filemap_fdatawait_keep_errors() for details.
2634 filemap_fdatawait_keep_errors(mapping);
2641 spin_lock_irq(&sb->s_inode_wblist_lock);
2643 spin_unlock_irq(&sb->s_inode_wblist_lock);
2645 mutex_unlock(&sb->s_sync_lock);
2648 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2649 enum wb_reason reason, bool skip_if_busy)
2651 struct backing_dev_info *bdi = sb->s_bdi;
2652 DEFINE_WB_COMPLETION(done, bdi);
2653 struct wb_writeback_work work = {
2655 .sync_mode = WB_SYNC_NONE,
2656 .tagged_writepages = 1,
2662 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2664 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2666 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2667 wb_wait_for_completion(&done);
2671 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2672 * @sb: the superblock
2673 * @nr: the number of pages to write
2674 * @reason: reason why some writeback work initiated
2676 * Start writeback on some inodes on this super_block. No guarantees are made
2677 * on how many (if any) will be written, and this function does not wait
2678 * for IO completion of submitted IO.
2680 void writeback_inodes_sb_nr(struct super_block *sb,
2682 enum wb_reason reason)
2684 __writeback_inodes_sb_nr(sb, nr, reason, false);
2686 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2689 * writeback_inodes_sb - writeback dirty inodes from given super_block
2690 * @sb: the superblock
2691 * @reason: reason why some writeback work was initiated
2693 * Start writeback on some inodes on this super_block. No guarantees are made
2694 * on how many (if any) will be written, and this function does not wait
2695 * for IO completion of submitted IO.
2697 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2699 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2701 EXPORT_SYMBOL(writeback_inodes_sb);
2704 * try_to_writeback_inodes_sb - try to start writeback if none underway
2705 * @sb: the superblock
2706 * @reason: reason why some writeback work was initiated
2708 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2710 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2712 if (!down_read_trylock(&sb->s_umount))
2715 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2716 up_read(&sb->s_umount);
2718 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2721 * sync_inodes_sb - sync sb inode pages
2722 * @sb: the superblock
2724 * This function writes and waits on any dirty inode belonging to this
2727 void sync_inodes_sb(struct super_block *sb)
2729 struct backing_dev_info *bdi = sb->s_bdi;
2730 DEFINE_WB_COMPLETION(done, bdi);
2731 struct wb_writeback_work work = {
2733 .sync_mode = WB_SYNC_ALL,
2734 .nr_pages = LONG_MAX,
2737 .reason = WB_REASON_SYNC,
2742 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2743 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2744 * bdi_has_dirty() need to be written out too.
2746 if (bdi == &noop_backing_dev_info)
2748 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2750 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2751 bdi_down_write_wb_switch_rwsem(bdi);
2752 bdi_split_work_to_wbs(bdi, &work, false);
2753 wb_wait_for_completion(&done);
2754 bdi_up_write_wb_switch_rwsem(bdi);
2758 EXPORT_SYMBOL(sync_inodes_sb);
2761 * write_inode_now - write an inode to disk
2762 * @inode: inode to write to disk
2763 * @sync: whether the write should be synchronous or not
2765 * This function commits an inode to disk immediately if it is dirty. This is
2766 * primarily needed by knfsd.
2768 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2770 int write_inode_now(struct inode *inode, int sync)
2772 struct writeback_control wbc = {
2773 .nr_to_write = LONG_MAX,
2774 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2776 .range_end = LLONG_MAX,
2779 if (!mapping_can_writeback(inode->i_mapping))
2780 wbc.nr_to_write = 0;
2783 return writeback_single_inode(inode, &wbc);
2785 EXPORT_SYMBOL(write_inode_now);
2788 * sync_inode_metadata - write an inode to disk
2789 * @inode: the inode to sync
2790 * @wait: wait for I/O to complete.
2792 * Write an inode to disk and adjust its dirty state after completion.
2794 * Note: only writes the actual inode, no associated data or other metadata.
2796 int sync_inode_metadata(struct inode *inode, int wait)
2798 struct writeback_control wbc = {
2799 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2800 .nr_to_write = 0, /* metadata-only */
2803 return writeback_single_inode(inode, &wbc);
2805 EXPORT_SYMBOL(sync_inode_metadata);