1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2002, Linus Torvalds.
7 * Contains all the functions related to writing back and waiting
8 * upon dirty inodes against superblocks, and writing back dirty
9 * pages against inodes. ie: data writeback. Writeout of the
10 * inode itself is not handled here.
12 * 10Apr2002 Andrew Morton
13 * Split out of fs/inode.c
14 * Additions for address_space-based writeback
17 #include <linux/kernel.h>
18 #include <linux/export.h>
19 #include <linux/spinlock.h>
20 #include <linux/slab.h>
21 #include <linux/sched.h>
24 #include <linux/pagemap.h>
25 #include <linux/kthread.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include <linux/device.h>
31 #include <linux/memcontrol.h>
35 * 4MB minimal write chunk size
37 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
40 * Passed into wb_writeback(), essentially a subset of writeback_control
42 struct wb_writeback_work {
44 struct super_block *sb;
45 enum writeback_sync_modes sync_mode;
46 unsigned int tagged_writepages:1;
47 unsigned int for_kupdate:1;
48 unsigned int range_cyclic:1;
49 unsigned int for_background:1;
50 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
51 unsigned int auto_free:1; /* free on completion */
52 enum wb_reason reason; /* why was writeback initiated? */
54 struct list_head list; /* pending work list */
55 struct wb_completion *done; /* set if the caller waits */
59 * If an inode is constantly having its pages dirtied, but then the
60 * updates stop dirtytime_expire_interval seconds in the past, it's
61 * possible for the worst case time between when an inode has its
62 * timestamps updated and when they finally get written out to be two
63 * dirtytime_expire_intervals. We set the default to 12 hours (in
64 * seconds), which means most of the time inodes will have their
65 * timestamps written to disk after 12 hours, but in the worst case a
66 * few inodes might not their timestamps updated for 24 hours.
68 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
70 static inline struct inode *wb_inode(struct list_head *head)
72 return list_entry(head, struct inode, i_io_list);
76 * Include the creation of the trace points after defining the
77 * wb_writeback_work structure and inline functions so that the definition
78 * remains local to this file.
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/writeback.h>
83 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
85 static bool wb_io_lists_populated(struct bdi_writeback *wb)
87 if (wb_has_dirty_io(wb)) {
90 set_bit(WB_has_dirty_io, &wb->state);
91 WARN_ON_ONCE(!wb->avg_write_bandwidth);
92 atomic_long_add(wb->avg_write_bandwidth,
93 &wb->bdi->tot_write_bandwidth);
98 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
100 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
101 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
102 clear_bit(WB_has_dirty_io, &wb->state);
103 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104 &wb->bdi->tot_write_bandwidth) < 0);
109 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
110 * @inode: inode to be moved
111 * @wb: target bdi_writeback
112 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
114 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
115 * Returns %true if @inode is the first occupant of the !dirty_time IO
116 * lists; otherwise, %false.
118 static bool inode_io_list_move_locked(struct inode *inode,
119 struct bdi_writeback *wb,
120 struct list_head *head)
122 assert_spin_locked(&wb->list_lock);
124 list_move(&inode->i_io_list, head);
126 /* dirty_time doesn't count as dirty_io until expiration */
127 if (head != &wb->b_dirty_time)
128 return wb_io_lists_populated(wb);
130 wb_io_lists_depopulated(wb);
134 static void wb_wakeup(struct bdi_writeback *wb)
136 spin_lock_bh(&wb->work_lock);
137 if (test_bit(WB_registered, &wb->state))
138 mod_delayed_work(bdi_wq, &wb->dwork, 0);
139 spin_unlock_bh(&wb->work_lock);
142 static void finish_writeback_work(struct bdi_writeback *wb,
143 struct wb_writeback_work *work)
145 struct wb_completion *done = work->done;
150 wait_queue_head_t *waitq = done->waitq;
152 /* @done can't be accessed after the following dec */
153 if (atomic_dec_and_test(&done->cnt))
158 static void wb_queue_work(struct bdi_writeback *wb,
159 struct wb_writeback_work *work)
161 trace_writeback_queue(wb, work);
164 atomic_inc(&work->done->cnt);
166 spin_lock_bh(&wb->work_lock);
168 if (test_bit(WB_registered, &wb->state)) {
169 list_add_tail(&work->list, &wb->work_list);
170 mod_delayed_work(bdi_wq, &wb->dwork, 0);
172 finish_writeback_work(wb, work);
174 spin_unlock_bh(&wb->work_lock);
178 * wb_wait_for_completion - wait for completion of bdi_writeback_works
179 * @done: target wb_completion
181 * Wait for one or more work items issued to @bdi with their ->done field
182 * set to @done, which should have been initialized with
183 * DEFINE_WB_COMPLETION(). This function returns after all such work items
184 * are completed. Work items which are waited upon aren't freed
185 * automatically on completion.
187 void wb_wait_for_completion(struct wb_completion *done)
189 atomic_dec(&done->cnt); /* put down the initial count */
190 wait_event(*done->waitq, !atomic_read(&done->cnt));
193 #ifdef CONFIG_CGROUP_WRITEBACK
196 * Parameters for foreign inode detection, see wbc_detach_inode() to see
199 * These paramters are inherently heuristical as the detection target
200 * itself is fuzzy. All we want to do is detaching an inode from the
201 * current owner if it's being written to by some other cgroups too much.
203 * The current cgroup writeback is built on the assumption that multiple
204 * cgroups writing to the same inode concurrently is very rare and a mode
205 * of operation which isn't well supported. As such, the goal is not
206 * taking too long when a different cgroup takes over an inode while
207 * avoiding too aggressive flip-flops from occasional foreign writes.
209 * We record, very roughly, 2s worth of IO time history and if more than
210 * half of that is foreign, trigger the switch. The recording is quantized
211 * to 16 slots. To avoid tiny writes from swinging the decision too much,
212 * writes smaller than 1/8 of avg size are ignored.
214 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
215 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
216 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
217 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
219 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
220 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
221 /* each slot's duration is 2s / 16 */
222 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
223 /* if foreign slots >= 8, switch */
224 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
225 /* one round can affect upto 5 slots */
226 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
229 * Maximum inodes per isw. A specific value has been chosen to make
230 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
232 #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
233 / sizeof(struct inode *))
235 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
236 static struct workqueue_struct *isw_wq;
238 void __inode_attach_wb(struct inode *inode, struct page *page)
240 struct backing_dev_info *bdi = inode_to_bdi(inode);
241 struct bdi_writeback *wb = NULL;
243 if (inode_cgwb_enabled(inode)) {
244 struct cgroup_subsys_state *memcg_css;
247 memcg_css = mem_cgroup_css_from_page(page);
248 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
250 /* must pin memcg_css, see wb_get_create() */
251 memcg_css = task_get_css(current, memory_cgrp_id);
252 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
261 * There may be multiple instances of this function racing to
262 * update the same inode. Use cmpxchg() to tell the winner.
264 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
267 EXPORT_SYMBOL_GPL(__inode_attach_wb);
270 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
271 * @inode: inode of interest with i_lock held
272 * @wb: target bdi_writeback
274 * Remove the inode from wb's io lists and if necessarily put onto b_attached
275 * list. Only inodes attached to cgwb's are kept on this list.
277 static void inode_cgwb_move_to_attached(struct inode *inode,
278 struct bdi_writeback *wb)
280 assert_spin_locked(&wb->list_lock);
281 assert_spin_locked(&inode->i_lock);
283 inode->i_state &= ~I_SYNC_QUEUED;
284 if (wb != &wb->bdi->wb)
285 list_move(&inode->i_io_list, &wb->b_attached);
287 list_del_init(&inode->i_io_list);
288 wb_io_lists_depopulated(wb);
292 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
293 * @inode: inode of interest with i_lock held
295 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
296 * held on entry and is released on return. The returned wb is guaranteed
297 * to stay @inode's associated wb until its list_lock is released.
299 static struct bdi_writeback *
300 locked_inode_to_wb_and_lock_list(struct inode *inode)
301 __releases(&inode->i_lock)
302 __acquires(&wb->list_lock)
305 struct bdi_writeback *wb = inode_to_wb(inode);
308 * inode_to_wb() association is protected by both
309 * @inode->i_lock and @wb->list_lock but list_lock nests
310 * outside i_lock. Drop i_lock and verify that the
311 * association hasn't changed after acquiring list_lock.
314 spin_unlock(&inode->i_lock);
315 spin_lock(&wb->list_lock);
317 /* i_wb may have changed inbetween, can't use inode_to_wb() */
318 if (likely(wb == inode->i_wb)) {
319 wb_put(wb); /* @inode already has ref */
323 spin_unlock(&wb->list_lock);
326 spin_lock(&inode->i_lock);
331 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
332 * @inode: inode of interest
334 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
337 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
338 __acquires(&wb->list_lock)
340 spin_lock(&inode->i_lock);
341 return locked_inode_to_wb_and_lock_list(inode);
344 struct inode_switch_wbs_context {
345 struct rcu_work work;
348 * Multiple inodes can be switched at once. The switching procedure
349 * consists of two parts, separated by a RCU grace period. To make
350 * sure that the second part is executed for each inode gone through
351 * the first part, all inode pointers are placed into a NULL-terminated
352 * array embedded into struct inode_switch_wbs_context. Otherwise
353 * an inode could be left in a non-consistent state.
355 struct bdi_writeback *new_wb;
356 struct inode *inodes[];
359 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
361 down_write(&bdi->wb_switch_rwsem);
364 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
366 up_write(&bdi->wb_switch_rwsem);
369 static bool inode_do_switch_wbs(struct inode *inode,
370 struct bdi_writeback *old_wb,
371 struct bdi_writeback *new_wb)
373 struct address_space *mapping = inode->i_mapping;
374 XA_STATE(xas, &mapping->i_pages, 0);
376 bool switched = false;
378 spin_lock(&inode->i_lock);
379 xa_lock_irq(&mapping->i_pages);
382 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
383 * path owns the inode and we shouldn't modify ->i_io_list.
385 if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
388 trace_inode_switch_wbs(inode, old_wb, new_wb);
391 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
392 * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
393 * folios actually under writeback.
395 xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
396 if (folio_test_dirty(folio)) {
397 long nr = folio_nr_pages(folio);
398 wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr);
399 wb_stat_mod(new_wb, WB_RECLAIMABLE, nr);
404 xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
405 long nr = folio_nr_pages(folio);
406 WARN_ON_ONCE(!folio_test_writeback(folio));
407 wb_stat_mod(old_wb, WB_WRITEBACK, -nr);
408 wb_stat_mod(new_wb, WB_WRITEBACK, nr);
411 if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
412 atomic_dec(&old_wb->writeback_inodes);
413 atomic_inc(&new_wb->writeback_inodes);
419 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
420 * the specific list @inode was on is ignored and the @inode is put on
421 * ->b_dirty which is always correct including from ->b_dirty_time.
422 * The transfer preserves @inode->dirtied_when ordering. If the @inode
423 * was clean, it means it was on the b_attached list, so move it onto
424 * the b_attached list of @new_wb.
426 if (!list_empty(&inode->i_io_list)) {
427 inode->i_wb = new_wb;
429 if (inode->i_state & I_DIRTY_ALL) {
432 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
433 if (time_after_eq(inode->dirtied_when,
436 inode_io_list_move_locked(inode, new_wb,
437 pos->i_io_list.prev);
439 inode_cgwb_move_to_attached(inode, new_wb);
442 inode->i_wb = new_wb;
445 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
446 inode->i_wb_frn_winner = 0;
447 inode->i_wb_frn_avg_time = 0;
448 inode->i_wb_frn_history = 0;
452 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
453 * ensures that the new wb is visible if they see !I_WB_SWITCH.
455 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
457 xa_unlock_irq(&mapping->i_pages);
458 spin_unlock(&inode->i_lock);
463 static void inode_switch_wbs_work_fn(struct work_struct *work)
465 struct inode_switch_wbs_context *isw =
466 container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
467 struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
468 struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
469 struct bdi_writeback *new_wb = isw->new_wb;
470 unsigned long nr_switched = 0;
471 struct inode **inodep;
474 * If @inode switches cgwb membership while sync_inodes_sb() is
475 * being issued, sync_inodes_sb() might miss it. Synchronize.
477 down_read(&bdi->wb_switch_rwsem);
480 * By the time control reaches here, RCU grace period has passed
481 * since I_WB_SWITCH assertion and all wb stat update transactions
482 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
483 * synchronizing against the i_pages lock.
485 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
486 * gives us exclusion against all wb related operations on @inode
487 * including IO list manipulations and stat updates.
489 if (old_wb < new_wb) {
490 spin_lock(&old_wb->list_lock);
491 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
493 spin_lock(&new_wb->list_lock);
494 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
497 for (inodep = isw->inodes; *inodep; inodep++) {
498 WARN_ON_ONCE((*inodep)->i_wb != old_wb);
499 if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
503 spin_unlock(&new_wb->list_lock);
504 spin_unlock(&old_wb->list_lock);
506 up_read(&bdi->wb_switch_rwsem);
510 wb_put_many(old_wb, nr_switched);
513 for (inodep = isw->inodes; *inodep; inodep++)
517 atomic_dec(&isw_nr_in_flight);
520 static bool inode_prepare_wbs_switch(struct inode *inode,
521 struct bdi_writeback *new_wb)
524 * Paired with smp_mb() in cgroup_writeback_umount().
525 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
526 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
527 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
534 /* while holding I_WB_SWITCH, no one else can update the association */
535 spin_lock(&inode->i_lock);
536 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
537 inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
538 inode_to_wb(inode) == new_wb) {
539 spin_unlock(&inode->i_lock);
542 inode->i_state |= I_WB_SWITCH;
544 spin_unlock(&inode->i_lock);
550 * inode_switch_wbs - change the wb association of an inode
551 * @inode: target inode
552 * @new_wb_id: ID of the new wb
554 * Switch @inode's wb association to the wb identified by @new_wb_id. The
555 * switching is performed asynchronously and may fail silently.
557 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
559 struct backing_dev_info *bdi = inode_to_bdi(inode);
560 struct cgroup_subsys_state *memcg_css;
561 struct inode_switch_wbs_context *isw;
563 /* noop if seems to be already in progress */
564 if (inode->i_state & I_WB_SWITCH)
567 /* avoid queueing a new switch if too many are already in flight */
568 if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
571 isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC);
575 atomic_inc(&isw_nr_in_flight);
577 /* find and pin the new wb */
579 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
580 if (memcg_css && !css_tryget(memcg_css))
586 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
591 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
594 isw->inodes[0] = inode;
597 * In addition to synchronizing among switchers, I_WB_SWITCH tells
598 * the RCU protected stat update paths to grab the i_page
599 * lock so that stat transfer can synchronize against them.
600 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
602 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
603 queue_rcu_work(isw_wq, &isw->work);
607 atomic_dec(&isw_nr_in_flight);
614 * cleanup_offline_cgwb - detach associated inodes
617 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
618 * to eventually release the dying @wb. Returns %true if not all inodes were
619 * switched and the function has to be restarted.
621 bool cleanup_offline_cgwb(struct bdi_writeback *wb)
623 struct cgroup_subsys_state *memcg_css;
624 struct inode_switch_wbs_context *isw;
627 bool restart = false;
629 isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW),
634 atomic_inc(&isw_nr_in_flight);
636 for (memcg_css = wb->memcg_css->parent; memcg_css;
637 memcg_css = memcg_css->parent) {
638 isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
642 if (unlikely(!isw->new_wb))
643 isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
646 spin_lock(&wb->list_lock);
647 list_for_each_entry(inode, &wb->b_attached, i_io_list) {
648 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
651 isw->inodes[nr++] = inode;
653 if (nr >= WB_MAX_INODES_PER_ISW - 1) {
658 spin_unlock(&wb->list_lock);
660 /* no attached inodes? bail out */
662 atomic_dec(&isw_nr_in_flight);
669 * In addition to synchronizing among switchers, I_WB_SWITCH tells
670 * the RCU protected stat update paths to grab the i_page
671 * lock so that stat transfer can synchronize against them.
672 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
674 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
675 queue_rcu_work(isw_wq, &isw->work);
681 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
682 * @wbc: writeback_control of interest
683 * @inode: target inode
685 * @inode is locked and about to be written back under the control of @wbc.
686 * Record @inode's writeback context into @wbc and unlock the i_lock. On
687 * writeback completion, wbc_detach_inode() should be called. This is used
688 * to track the cgroup writeback context.
690 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
693 if (!inode_cgwb_enabled(inode)) {
694 spin_unlock(&inode->i_lock);
698 wbc->wb = inode_to_wb(inode);
701 wbc->wb_id = wbc->wb->memcg_css->id;
702 wbc->wb_lcand_id = inode->i_wb_frn_winner;
703 wbc->wb_tcand_id = 0;
705 wbc->wb_lcand_bytes = 0;
706 wbc->wb_tcand_bytes = 0;
709 spin_unlock(&inode->i_lock);
712 * A dying wb indicates that either the blkcg associated with the
713 * memcg changed or the associated memcg is dying. In the first
714 * case, a replacement wb should already be available and we should
715 * refresh the wb immediately. In the second case, trying to
716 * refresh will keep failing.
718 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
719 inode_switch_wbs(inode, wbc->wb_id);
721 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
724 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
725 * @wbc: writeback_control of the just finished writeback
727 * To be called after a writeback attempt of an inode finishes and undoes
728 * wbc_attach_and_unlock_inode(). Can be called under any context.
730 * As concurrent write sharing of an inode is expected to be very rare and
731 * memcg only tracks page ownership on first-use basis severely confining
732 * the usefulness of such sharing, cgroup writeback tracks ownership
733 * per-inode. While the support for concurrent write sharing of an inode
734 * is deemed unnecessary, an inode being written to by different cgroups at
735 * different points in time is a lot more common, and, more importantly,
736 * charging only by first-use can too readily lead to grossly incorrect
737 * behaviors (single foreign page can lead to gigabytes of writeback to be
738 * incorrectly attributed).
740 * To resolve this issue, cgroup writeback detects the majority dirtier of
741 * an inode and transfers the ownership to it. To avoid unnnecessary
742 * oscillation, the detection mechanism keeps track of history and gives
743 * out the switch verdict only if the foreign usage pattern is stable over
744 * a certain amount of time and/or writeback attempts.
746 * On each writeback attempt, @wbc tries to detect the majority writer
747 * using Boyer-Moore majority vote algorithm. In addition to the byte
748 * count from the majority voting, it also counts the bytes written for the
749 * current wb and the last round's winner wb (max of last round's current
750 * wb, the winner from two rounds ago, and the last round's majority
751 * candidate). Keeping track of the historical winner helps the algorithm
752 * to semi-reliably detect the most active writer even when it's not the
755 * Once the winner of the round is determined, whether the winner is
756 * foreign or not and how much IO time the round consumed is recorded in
757 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
758 * over a certain threshold, the switch verdict is given.
760 void wbc_detach_inode(struct writeback_control *wbc)
762 struct bdi_writeback *wb = wbc->wb;
763 struct inode *inode = wbc->inode;
764 unsigned long avg_time, max_bytes, max_time;
771 history = inode->i_wb_frn_history;
772 avg_time = inode->i_wb_frn_avg_time;
774 /* pick the winner of this round */
775 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
776 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
778 max_bytes = wbc->wb_bytes;
779 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
780 max_id = wbc->wb_lcand_id;
781 max_bytes = wbc->wb_lcand_bytes;
783 max_id = wbc->wb_tcand_id;
784 max_bytes = wbc->wb_tcand_bytes;
788 * Calculate the amount of IO time the winner consumed and fold it
789 * into the running average kept per inode. If the consumed IO
790 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
791 * deciding whether to switch or not. This is to prevent one-off
792 * small dirtiers from skewing the verdict.
794 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
795 wb->avg_write_bandwidth);
797 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
798 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
800 avg_time = max_time; /* immediate catch up on first run */
802 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
806 * The switch verdict is reached if foreign wb's consume
807 * more than a certain proportion of IO time in a
808 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
809 * history mask where each bit represents one sixteenth of
810 * the period. Determine the number of slots to shift into
811 * history from @max_time.
813 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
814 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
816 if (wbc->wb_id != max_id)
817 history |= (1U << slots) - 1;
820 trace_inode_foreign_history(inode, wbc, history);
823 * Switch if the current wb isn't the consistent winner.
824 * If there are multiple closely competing dirtiers, the
825 * inode may switch across them repeatedly over time, which
826 * is okay. The main goal is avoiding keeping an inode on
827 * the wrong wb for an extended period of time.
829 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
830 inode_switch_wbs(inode, max_id);
834 * Multiple instances of this function may race to update the
835 * following fields but we don't mind occassional inaccuracies.
837 inode->i_wb_frn_winner = max_id;
838 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
839 inode->i_wb_frn_history = history;
844 EXPORT_SYMBOL_GPL(wbc_detach_inode);
847 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
848 * @wbc: writeback_control of the writeback in progress
849 * @page: page being written out
850 * @bytes: number of bytes being written out
852 * @bytes from @page are about to written out during the writeback
853 * controlled by @wbc. Keep the book for foreign inode detection. See
854 * wbc_detach_inode().
856 void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
859 struct cgroup_subsys_state *css;
863 * pageout() path doesn't attach @wbc to the inode being written
864 * out. This is intentional as we don't want the function to block
865 * behind a slow cgroup. Ultimately, we want pageout() to kick off
866 * regular writeback instead of writing things out itself.
868 if (!wbc->wb || wbc->no_cgroup_owner)
871 css = mem_cgroup_css_from_page(page);
872 /* dead cgroups shouldn't contribute to inode ownership arbitration */
873 if (!(css->flags & CSS_ONLINE))
878 if (id == wbc->wb_id) {
879 wbc->wb_bytes += bytes;
883 if (id == wbc->wb_lcand_id)
884 wbc->wb_lcand_bytes += bytes;
886 /* Boyer-Moore majority vote algorithm */
887 if (!wbc->wb_tcand_bytes)
888 wbc->wb_tcand_id = id;
889 if (id == wbc->wb_tcand_id)
890 wbc->wb_tcand_bytes += bytes;
892 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
894 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
897 * inode_congested - test whether an inode is congested
898 * @inode: inode to test for congestion (may be NULL)
899 * @cong_bits: mask of WB_[a]sync_congested bits to test
901 * Tests whether @inode is congested. @cong_bits is the mask of congestion
902 * bits to test and the return value is the mask of set bits.
904 * If cgroup writeback is enabled for @inode, the congestion state is
905 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
906 * associated with @inode is congested; otherwise, the root wb's congestion
909 * @inode is allowed to be NULL as this function is often called on
910 * mapping->host which is NULL for the swapper space.
912 int inode_congested(struct inode *inode, int cong_bits)
915 * Once set, ->i_wb never becomes NULL while the inode is alive.
916 * Start transaction iff ->i_wb is visible.
918 if (inode && inode_to_wb_is_valid(inode)) {
919 struct bdi_writeback *wb;
920 struct wb_lock_cookie lock_cookie = {};
923 wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
924 congested = wb_congested(wb, cong_bits);
925 unlocked_inode_to_wb_end(inode, &lock_cookie);
929 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
931 EXPORT_SYMBOL_GPL(inode_congested);
934 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
935 * @wb: target bdi_writeback to split @nr_pages to
936 * @nr_pages: number of pages to write for the whole bdi
938 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
939 * relation to the total write bandwidth of all wb's w/ dirty inodes on
942 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
944 unsigned long this_bw = wb->avg_write_bandwidth;
945 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
947 if (nr_pages == LONG_MAX)
951 * This may be called on clean wb's and proportional distribution
952 * may not make sense, just use the original @nr_pages in those
953 * cases. In general, we wanna err on the side of writing more.
955 if (!tot_bw || this_bw >= tot_bw)
958 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
962 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
963 * @bdi: target backing_dev_info
964 * @base_work: wb_writeback_work to issue
965 * @skip_if_busy: skip wb's which already have writeback in progress
967 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
968 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
969 * distributed to the busy wbs according to each wb's proportion in the
970 * total active write bandwidth of @bdi.
972 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
973 struct wb_writeback_work *base_work,
976 struct bdi_writeback *last_wb = NULL;
977 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
978 struct bdi_writeback, bdi_node);
983 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
984 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
985 struct wb_writeback_work fallback_work;
986 struct wb_writeback_work *work;
994 /* SYNC_ALL writes out I_DIRTY_TIME too */
995 if (!wb_has_dirty_io(wb) &&
996 (base_work->sync_mode == WB_SYNC_NONE ||
997 list_empty(&wb->b_dirty_time)))
999 if (skip_if_busy && writeback_in_progress(wb))
1002 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
1004 work = kmalloc(sizeof(*work), GFP_ATOMIC);
1007 work->nr_pages = nr_pages;
1008 work->auto_free = 1;
1009 wb_queue_work(wb, work);
1013 /* alloc failed, execute synchronously using on-stack fallback */
1014 work = &fallback_work;
1016 work->nr_pages = nr_pages;
1017 work->auto_free = 0;
1018 work->done = &fallback_work_done;
1020 wb_queue_work(wb, work);
1023 * Pin @wb so that it stays on @bdi->wb_list. This allows
1024 * continuing iteration from @wb after dropping and
1025 * regrabbing rcu read lock.
1031 wb_wait_for_completion(&fallback_work_done);
1041 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1042 * @bdi_id: target bdi id
1043 * @memcg_id: target memcg css id
1044 * @reason: reason why some writeback work initiated
1045 * @done: target wb_completion
1047 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1048 * with the specified parameters.
1050 int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
1051 enum wb_reason reason, struct wb_completion *done)
1053 struct backing_dev_info *bdi;
1054 struct cgroup_subsys_state *memcg_css;
1055 struct bdi_writeback *wb;
1056 struct wb_writeback_work *work;
1057 unsigned long dirty;
1060 /* lookup bdi and memcg */
1061 bdi = bdi_get_by_id(bdi_id);
1066 memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1067 if (memcg_css && !css_tryget(memcg_css))
1076 * And find the associated wb. If the wb isn't there already
1077 * there's nothing to flush, don't create one.
1079 wb = wb_get_lookup(bdi, memcg_css);
1086 * The caller is attempting to write out most of
1087 * the currently dirty pages. Let's take the current dirty page
1088 * count and inflate it by 25% which should be large enough to
1089 * flush out most dirty pages while avoiding getting livelocked by
1090 * concurrent dirtiers.
1092 * BTW the memcg stats are flushed periodically and this is best-effort
1093 * estimation, so some potential error is ok.
1095 dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1096 dirty = dirty * 10 / 8;
1098 /* issue the writeback work */
1099 work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1101 work->nr_pages = dirty;
1102 work->sync_mode = WB_SYNC_NONE;
1103 work->range_cyclic = 1;
1104 work->reason = reason;
1106 work->auto_free = 1;
1107 wb_queue_work(wb, work);
1122 * cgroup_writeback_umount - flush inode wb switches for umount
1124 * This function is called when a super_block is about to be destroyed and
1125 * flushes in-flight inode wb switches. An inode wb switch goes through
1126 * RCU and then workqueue, so the two need to be flushed in order to ensure
1127 * that all previously scheduled switches are finished. As wb switches are
1128 * rare occurrences and synchronize_rcu() can take a while, perform
1129 * flushing iff wb switches are in flight.
1131 void cgroup_writeback_umount(void)
1134 * SB_ACTIVE should be reliably cleared before checking
1135 * isw_nr_in_flight, see generic_shutdown_super().
1139 if (atomic_read(&isw_nr_in_flight)) {
1141 * Use rcu_barrier() to wait for all pending callbacks to
1142 * ensure that all in-flight wb switches are in the workqueue.
1145 flush_workqueue(isw_wq);
1149 static int __init cgroup_writeback_init(void)
1151 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1156 fs_initcall(cgroup_writeback_init);
1158 #else /* CONFIG_CGROUP_WRITEBACK */
1160 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1161 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1163 static void inode_cgwb_move_to_attached(struct inode *inode,
1164 struct bdi_writeback *wb)
1166 assert_spin_locked(&wb->list_lock);
1167 assert_spin_locked(&inode->i_lock);
1169 inode->i_state &= ~I_SYNC_QUEUED;
1170 list_del_init(&inode->i_io_list);
1171 wb_io_lists_depopulated(wb);
1174 static struct bdi_writeback *
1175 locked_inode_to_wb_and_lock_list(struct inode *inode)
1176 __releases(&inode->i_lock)
1177 __acquires(&wb->list_lock)
1179 struct bdi_writeback *wb = inode_to_wb(inode);
1181 spin_unlock(&inode->i_lock);
1182 spin_lock(&wb->list_lock);
1186 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1187 __acquires(&wb->list_lock)
1189 struct bdi_writeback *wb = inode_to_wb(inode);
1191 spin_lock(&wb->list_lock);
1195 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1200 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1201 struct wb_writeback_work *base_work,
1206 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1207 base_work->auto_free = 0;
1208 wb_queue_work(&bdi->wb, base_work);
1212 #endif /* CONFIG_CGROUP_WRITEBACK */
1215 * Add in the number of potentially dirty inodes, because each inode
1216 * write can dirty pagecache in the underlying blockdev.
1218 static unsigned long get_nr_dirty_pages(void)
1220 return global_node_page_state(NR_FILE_DIRTY) +
1221 get_nr_dirty_inodes();
1224 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1226 if (!wb_has_dirty_io(wb))
1230 * All callers of this function want to start writeback of all
1231 * dirty pages. Places like vmscan can call this at a very
1232 * high frequency, causing pointless allocations of tons of
1233 * work items and keeping the flusher threads busy retrieving
1234 * that work. Ensure that we only allow one of them pending and
1235 * inflight at the time.
1237 if (test_bit(WB_start_all, &wb->state) ||
1238 test_and_set_bit(WB_start_all, &wb->state))
1241 wb->start_all_reason = reason;
1246 * wb_start_background_writeback - start background writeback
1247 * @wb: bdi_writback to write from
1250 * This makes sure WB_SYNC_NONE background writeback happens. When
1251 * this function returns, it is only guaranteed that for given wb
1252 * some IO is happening if we are over background dirty threshold.
1253 * Caller need not hold sb s_umount semaphore.
1255 void wb_start_background_writeback(struct bdi_writeback *wb)
1258 * We just wake up the flusher thread. It will perform background
1259 * writeback as soon as there is no other work to do.
1261 trace_writeback_wake_background(wb);
1266 * Remove the inode from the writeback list it is on.
1268 void inode_io_list_del(struct inode *inode)
1270 struct bdi_writeback *wb;
1272 wb = inode_to_wb_and_lock_list(inode);
1273 spin_lock(&inode->i_lock);
1275 inode->i_state &= ~I_SYNC_QUEUED;
1276 list_del_init(&inode->i_io_list);
1277 wb_io_lists_depopulated(wb);
1279 spin_unlock(&inode->i_lock);
1280 spin_unlock(&wb->list_lock);
1282 EXPORT_SYMBOL(inode_io_list_del);
1285 * mark an inode as under writeback on the sb
1287 void sb_mark_inode_writeback(struct inode *inode)
1289 struct super_block *sb = inode->i_sb;
1290 unsigned long flags;
1292 if (list_empty(&inode->i_wb_list)) {
1293 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1294 if (list_empty(&inode->i_wb_list)) {
1295 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1296 trace_sb_mark_inode_writeback(inode);
1298 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1303 * clear an inode as under writeback on the sb
1305 void sb_clear_inode_writeback(struct inode *inode)
1307 struct super_block *sb = inode->i_sb;
1308 unsigned long flags;
1310 if (!list_empty(&inode->i_wb_list)) {
1311 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1312 if (!list_empty(&inode->i_wb_list)) {
1313 list_del_init(&inode->i_wb_list);
1314 trace_sb_clear_inode_writeback(inode);
1316 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1321 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1322 * furthest end of its superblock's dirty-inode list.
1324 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1325 * already the most-recently-dirtied inode on the b_dirty list. If that is
1326 * the case then the inode must have been redirtied while it was being written
1327 * out and we don't reset its dirtied_when.
1329 static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1331 assert_spin_locked(&inode->i_lock);
1333 if (!list_empty(&wb->b_dirty)) {
1336 tail = wb_inode(wb->b_dirty.next);
1337 if (time_before(inode->dirtied_when, tail->dirtied_when))
1338 inode->dirtied_when = jiffies;
1340 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1341 inode->i_state &= ~I_SYNC_QUEUED;
1344 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1346 spin_lock(&inode->i_lock);
1347 redirty_tail_locked(inode, wb);
1348 spin_unlock(&inode->i_lock);
1352 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1354 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1356 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1359 static void inode_sync_complete(struct inode *inode)
1361 inode->i_state &= ~I_SYNC;
1362 /* If inode is clean an unused, put it into LRU now... */
1363 inode_add_lru(inode);
1364 /* Waiters must see I_SYNC cleared before being woken up */
1366 wake_up_bit(&inode->i_state, __I_SYNC);
1369 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1371 bool ret = time_after(inode->dirtied_when, t);
1372 #ifndef CONFIG_64BIT
1374 * For inodes being constantly redirtied, dirtied_when can get stuck.
1375 * It _appears_ to be in the future, but is actually in distant past.
1376 * This test is necessary to prevent such wrapped-around relative times
1377 * from permanently stopping the whole bdi writeback.
1379 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1384 #define EXPIRE_DIRTY_ATIME 0x0001
1387 * Move expired (dirtied before dirtied_before) dirty inodes from
1388 * @delaying_queue to @dispatch_queue.
1390 static int move_expired_inodes(struct list_head *delaying_queue,
1391 struct list_head *dispatch_queue,
1392 unsigned long dirtied_before)
1395 struct list_head *pos, *node;
1396 struct super_block *sb = NULL;
1397 struct inode *inode;
1401 while (!list_empty(delaying_queue)) {
1402 inode = wb_inode(delaying_queue->prev);
1403 if (inode_dirtied_after(inode, dirtied_before))
1405 list_move(&inode->i_io_list, &tmp);
1407 spin_lock(&inode->i_lock);
1408 inode->i_state |= I_SYNC_QUEUED;
1409 spin_unlock(&inode->i_lock);
1410 if (sb_is_blkdev_sb(inode->i_sb))
1412 if (sb && sb != inode->i_sb)
1417 /* just one sb in list, splice to dispatch_queue and we're done */
1419 list_splice(&tmp, dispatch_queue);
1423 /* Move inodes from one superblock together */
1424 while (!list_empty(&tmp)) {
1425 sb = wb_inode(tmp.prev)->i_sb;
1426 list_for_each_prev_safe(pos, node, &tmp) {
1427 inode = wb_inode(pos);
1428 if (inode->i_sb == sb)
1429 list_move(&inode->i_io_list, dispatch_queue);
1437 * Queue all expired dirty inodes for io, eldest first.
1439 * newly dirtied b_dirty b_io b_more_io
1440 * =============> gf edc BA
1442 * newly dirtied b_dirty b_io b_more_io
1443 * =============> g fBAedc
1445 * +--> dequeue for IO
1447 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1448 unsigned long dirtied_before)
1451 unsigned long time_expire_jif = dirtied_before;
1453 assert_spin_locked(&wb->list_lock);
1454 list_splice_init(&wb->b_more_io, &wb->b_io);
1455 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1456 if (!work->for_sync)
1457 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1458 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1461 wb_io_lists_populated(wb);
1462 trace_writeback_queue_io(wb, work, dirtied_before, moved);
1465 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1469 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1470 trace_writeback_write_inode_start(inode, wbc);
1471 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1472 trace_writeback_write_inode(inode, wbc);
1479 * Wait for writeback on an inode to complete. Called with i_lock held.
1480 * Caller must make sure inode cannot go away when we drop i_lock.
1482 static void __inode_wait_for_writeback(struct inode *inode)
1483 __releases(inode->i_lock)
1484 __acquires(inode->i_lock)
1486 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1487 wait_queue_head_t *wqh;
1489 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1490 while (inode->i_state & I_SYNC) {
1491 spin_unlock(&inode->i_lock);
1492 __wait_on_bit(wqh, &wq, bit_wait,
1493 TASK_UNINTERRUPTIBLE);
1494 spin_lock(&inode->i_lock);
1499 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1501 void inode_wait_for_writeback(struct inode *inode)
1503 spin_lock(&inode->i_lock);
1504 __inode_wait_for_writeback(inode);
1505 spin_unlock(&inode->i_lock);
1509 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1510 * held and drops it. It is aimed for callers not holding any inode reference
1511 * so once i_lock is dropped, inode can go away.
1513 static void inode_sleep_on_writeback(struct inode *inode)
1514 __releases(inode->i_lock)
1517 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1520 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1521 sleep = inode->i_state & I_SYNC;
1522 spin_unlock(&inode->i_lock);
1525 finish_wait(wqh, &wait);
1529 * Find proper writeback list for the inode depending on its current state and
1530 * possibly also change of its state while we were doing writeback. Here we
1531 * handle things such as livelock prevention or fairness of writeback among
1532 * inodes. This function can be called only by flusher thread - noone else
1533 * processes all inodes in writeback lists and requeueing inodes behind flusher
1534 * thread's back can have unexpected consequences.
1536 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1537 struct writeback_control *wbc)
1539 if (inode->i_state & I_FREEING)
1543 * Sync livelock prevention. Each inode is tagged and synced in one
1544 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1545 * the dirty time to prevent enqueue and sync it again.
1547 if ((inode->i_state & I_DIRTY) &&
1548 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1549 inode->dirtied_when = jiffies;
1551 if (wbc->pages_skipped) {
1553 * writeback is not making progress due to locked
1554 * buffers. Skip this inode for now.
1556 redirty_tail_locked(inode, wb);
1560 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1562 * We didn't write back all the pages. nfs_writepages()
1563 * sometimes bales out without doing anything.
1565 if (wbc->nr_to_write <= 0) {
1566 /* Slice used up. Queue for next turn. */
1567 requeue_io(inode, wb);
1570 * Writeback blocked by something other than
1571 * congestion. Delay the inode for some time to
1572 * avoid spinning on the CPU (100% iowait)
1573 * retrying writeback of the dirty page/inode
1574 * that cannot be performed immediately.
1576 redirty_tail_locked(inode, wb);
1578 } else if (inode->i_state & I_DIRTY) {
1580 * Filesystems can dirty the inode during writeback operations,
1581 * such as delayed allocation during submission or metadata
1582 * updates after data IO completion.
1584 redirty_tail_locked(inode, wb);
1585 } else if (inode->i_state & I_DIRTY_TIME) {
1586 inode->dirtied_when = jiffies;
1587 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1588 inode->i_state &= ~I_SYNC_QUEUED;
1590 /* The inode is clean. Remove from writeback lists. */
1591 inode_cgwb_move_to_attached(inode, wb);
1596 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1597 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1599 * This doesn't remove the inode from the writeback list it is on, except
1600 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1601 * expiration. The caller is otherwise responsible for writeback list handling.
1603 * The caller is also responsible for setting the I_SYNC flag beforehand and
1604 * calling inode_sync_complete() to clear it afterwards.
1607 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1609 struct address_space *mapping = inode->i_mapping;
1610 long nr_to_write = wbc->nr_to_write;
1614 WARN_ON(!(inode->i_state & I_SYNC));
1616 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1618 ret = do_writepages(mapping, wbc);
1621 * Make sure to wait on the data before writing out the metadata.
1622 * This is important for filesystems that modify metadata on data
1623 * I/O completion. We don't do it for sync(2) writeback because it has a
1624 * separate, external IO completion path and ->sync_fs for guaranteeing
1625 * inode metadata is written back correctly.
1627 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1628 int err = filemap_fdatawait(mapping);
1634 * If the inode has dirty timestamps and we need to write them, call
1635 * mark_inode_dirty_sync() to notify the filesystem about it and to
1636 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1638 if ((inode->i_state & I_DIRTY_TIME) &&
1639 (wbc->sync_mode == WB_SYNC_ALL ||
1640 time_after(jiffies, inode->dirtied_time_when +
1641 dirtytime_expire_interval * HZ))) {
1642 trace_writeback_lazytime(inode);
1643 mark_inode_dirty_sync(inode);
1647 * Get and clear the dirty flags from i_state. This needs to be done
1648 * after calling writepages because some filesystems may redirty the
1649 * inode during writepages due to delalloc. It also needs to be done
1650 * after handling timestamp expiration, as that may dirty the inode too.
1652 spin_lock(&inode->i_lock);
1653 dirty = inode->i_state & I_DIRTY;
1654 inode->i_state &= ~dirty;
1657 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1658 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1659 * either they see the I_DIRTY bits cleared or we see the dirtied
1662 * I_DIRTY_PAGES is always cleared together above even if @mapping
1663 * still has dirty pages. The flag is reinstated after smp_mb() if
1664 * necessary. This guarantees that either __mark_inode_dirty()
1665 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1669 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1670 inode->i_state |= I_DIRTY_PAGES;
1671 else if (unlikely(inode->i_state & I_PINNING_FSCACHE_WB)) {
1672 if (!(inode->i_state & I_DIRTY_PAGES)) {
1673 inode->i_state &= ~I_PINNING_FSCACHE_WB;
1674 wbc->unpinned_fscache_wb = true;
1675 dirty |= I_PINNING_FSCACHE_WB; /* Cause write_inode */
1679 spin_unlock(&inode->i_lock);
1681 /* Don't write the inode if only I_DIRTY_PAGES was set */
1682 if (dirty & ~I_DIRTY_PAGES) {
1683 int err = write_inode(inode, wbc);
1687 wbc->unpinned_fscache_wb = false;
1688 trace_writeback_single_inode(inode, wbc, nr_to_write);
1693 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1694 * the regular batched writeback done by the flusher threads in
1695 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1696 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1698 * To prevent the inode from going away, either the caller must have a reference
1699 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1701 static int writeback_single_inode(struct inode *inode,
1702 struct writeback_control *wbc)
1704 struct bdi_writeback *wb;
1707 spin_lock(&inode->i_lock);
1708 if (!atomic_read(&inode->i_count))
1709 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1711 WARN_ON(inode->i_state & I_WILL_FREE);
1713 if (inode->i_state & I_SYNC) {
1715 * Writeback is already running on the inode. For WB_SYNC_NONE,
1716 * that's enough and we can just return. For WB_SYNC_ALL, we
1717 * must wait for the existing writeback to complete, then do
1718 * writeback again if there's anything left.
1720 if (wbc->sync_mode != WB_SYNC_ALL)
1722 __inode_wait_for_writeback(inode);
1724 WARN_ON(inode->i_state & I_SYNC);
1726 * If the inode is already fully clean, then there's nothing to do.
1728 * For data-integrity syncs we also need to check whether any pages are
1729 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1730 * there are any such pages, we'll need to wait for them.
1732 if (!(inode->i_state & I_DIRTY_ALL) &&
1733 (wbc->sync_mode != WB_SYNC_ALL ||
1734 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1736 inode->i_state |= I_SYNC;
1737 wbc_attach_and_unlock_inode(wbc, inode);
1739 ret = __writeback_single_inode(inode, wbc);
1741 wbc_detach_inode(wbc);
1743 wb = inode_to_wb_and_lock_list(inode);
1744 spin_lock(&inode->i_lock);
1746 * If the inode is now fully clean, then it can be safely removed from
1747 * its writeback list (if any). Otherwise the flusher threads are
1748 * responsible for the writeback lists.
1750 if (!(inode->i_state & I_DIRTY_ALL))
1751 inode_cgwb_move_to_attached(inode, wb);
1752 spin_unlock(&wb->list_lock);
1753 inode_sync_complete(inode);
1755 spin_unlock(&inode->i_lock);
1759 static long writeback_chunk_size(struct bdi_writeback *wb,
1760 struct wb_writeback_work *work)
1765 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1766 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1767 * here avoids calling into writeback_inodes_wb() more than once.
1769 * The intended call sequence for WB_SYNC_ALL writeback is:
1772 * writeback_sb_inodes() <== called only once
1773 * write_cache_pages() <== called once for each inode
1774 * (quickly) tag currently dirty pages
1775 * (maybe slowly) sync all tagged pages
1777 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1780 pages = min(wb->avg_write_bandwidth / 2,
1781 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1782 pages = min(pages, work->nr_pages);
1783 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1784 MIN_WRITEBACK_PAGES);
1791 * Write a portion of b_io inodes which belong to @sb.
1793 * Return the number of pages and/or inodes written.
1795 * NOTE! This is called with wb->list_lock held, and will
1796 * unlock and relock that for each inode it ends up doing
1799 static long writeback_sb_inodes(struct super_block *sb,
1800 struct bdi_writeback *wb,
1801 struct wb_writeback_work *work)
1803 struct writeback_control wbc = {
1804 .sync_mode = work->sync_mode,
1805 .tagged_writepages = work->tagged_writepages,
1806 .for_kupdate = work->for_kupdate,
1807 .for_background = work->for_background,
1808 .for_sync = work->for_sync,
1809 .range_cyclic = work->range_cyclic,
1811 .range_end = LLONG_MAX,
1813 unsigned long start_time = jiffies;
1815 long wrote = 0; /* count both pages and inodes */
1817 while (!list_empty(&wb->b_io)) {
1818 struct inode *inode = wb_inode(wb->b_io.prev);
1819 struct bdi_writeback *tmp_wb;
1821 if (inode->i_sb != sb) {
1824 * We only want to write back data for this
1825 * superblock, move all inodes not belonging
1826 * to it back onto the dirty list.
1828 redirty_tail(inode, wb);
1833 * The inode belongs to a different superblock.
1834 * Bounce back to the caller to unpin this and
1835 * pin the next superblock.
1841 * Don't bother with new inodes or inodes being freed, first
1842 * kind does not need periodic writeout yet, and for the latter
1843 * kind writeout is handled by the freer.
1845 spin_lock(&inode->i_lock);
1846 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1847 redirty_tail_locked(inode, wb);
1848 spin_unlock(&inode->i_lock);
1851 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1853 * If this inode is locked for writeback and we are not
1854 * doing writeback-for-data-integrity, move it to
1855 * b_more_io so that writeback can proceed with the
1856 * other inodes on s_io.
1858 * We'll have another go at writing back this inode
1859 * when we completed a full scan of b_io.
1861 spin_unlock(&inode->i_lock);
1862 requeue_io(inode, wb);
1863 trace_writeback_sb_inodes_requeue(inode);
1866 spin_unlock(&wb->list_lock);
1869 * We already requeued the inode if it had I_SYNC set and we
1870 * are doing WB_SYNC_NONE writeback. So this catches only the
1873 if (inode->i_state & I_SYNC) {
1874 /* Wait for I_SYNC. This function drops i_lock... */
1875 inode_sleep_on_writeback(inode);
1876 /* Inode may be gone, start again */
1877 spin_lock(&wb->list_lock);
1880 inode->i_state |= I_SYNC;
1881 wbc_attach_and_unlock_inode(&wbc, inode);
1883 write_chunk = writeback_chunk_size(wb, work);
1884 wbc.nr_to_write = write_chunk;
1885 wbc.pages_skipped = 0;
1888 * We use I_SYNC to pin the inode in memory. While it is set
1889 * evict_inode() will wait so the inode cannot be freed.
1891 __writeback_single_inode(inode, &wbc);
1893 wbc_detach_inode(&wbc);
1894 work->nr_pages -= write_chunk - wbc.nr_to_write;
1895 wrote += write_chunk - wbc.nr_to_write;
1897 if (need_resched()) {
1899 * We're trying to balance between building up a nice
1900 * long list of IOs to improve our merge rate, and
1901 * getting those IOs out quickly for anyone throttling
1902 * in balance_dirty_pages(). cond_resched() doesn't
1903 * unplug, so get our IOs out the door before we
1907 blk_flush_plug(current->plug, false);
1912 * Requeue @inode if still dirty. Be careful as @inode may
1913 * have been switched to another wb in the meantime.
1915 tmp_wb = inode_to_wb_and_lock_list(inode);
1916 spin_lock(&inode->i_lock);
1917 if (!(inode->i_state & I_DIRTY_ALL))
1919 requeue_inode(inode, tmp_wb, &wbc);
1920 inode_sync_complete(inode);
1921 spin_unlock(&inode->i_lock);
1923 if (unlikely(tmp_wb != wb)) {
1924 spin_unlock(&tmp_wb->list_lock);
1925 spin_lock(&wb->list_lock);
1929 * bail out to wb_writeback() often enough to check
1930 * background threshold and other termination conditions.
1933 if (time_is_before_jiffies(start_time + HZ / 10UL))
1935 if (work->nr_pages <= 0)
1942 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1943 struct wb_writeback_work *work)
1945 unsigned long start_time = jiffies;
1948 while (!list_empty(&wb->b_io)) {
1949 struct inode *inode = wb_inode(wb->b_io.prev);
1950 struct super_block *sb = inode->i_sb;
1952 if (!trylock_super(sb)) {
1954 * trylock_super() may fail consistently due to
1955 * s_umount being grabbed by someone else. Don't use
1956 * requeue_io() to avoid busy retrying the inode/sb.
1958 redirty_tail(inode, wb);
1961 wrote += writeback_sb_inodes(sb, wb, work);
1962 up_read(&sb->s_umount);
1964 /* refer to the same tests at the end of writeback_sb_inodes */
1966 if (time_is_before_jiffies(start_time + HZ / 10UL))
1968 if (work->nr_pages <= 0)
1972 /* Leave any unwritten inodes on b_io */
1976 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1977 enum wb_reason reason)
1979 struct wb_writeback_work work = {
1980 .nr_pages = nr_pages,
1981 .sync_mode = WB_SYNC_NONE,
1985 struct blk_plug plug;
1987 blk_start_plug(&plug);
1988 spin_lock(&wb->list_lock);
1989 if (list_empty(&wb->b_io))
1990 queue_io(wb, &work, jiffies);
1991 __writeback_inodes_wb(wb, &work);
1992 spin_unlock(&wb->list_lock);
1993 blk_finish_plug(&plug);
1995 return nr_pages - work.nr_pages;
1999 * Explicit flushing or periodic writeback of "old" data.
2001 * Define "old": the first time one of an inode's pages is dirtied, we mark the
2002 * dirtying-time in the inode's address_space. So this periodic writeback code
2003 * just walks the superblock inode list, writing back any inodes which are
2004 * older than a specific point in time.
2006 * Try to run once per dirty_writeback_interval. But if a writeback event
2007 * takes longer than a dirty_writeback_interval interval, then leave a
2010 * dirtied_before takes precedence over nr_to_write. So we'll only write back
2011 * all dirty pages if they are all attached to "old" mappings.
2013 static long wb_writeback(struct bdi_writeback *wb,
2014 struct wb_writeback_work *work)
2016 long nr_pages = work->nr_pages;
2017 unsigned long dirtied_before = jiffies;
2018 struct inode *inode;
2020 struct blk_plug plug;
2022 blk_start_plug(&plug);
2023 spin_lock(&wb->list_lock);
2026 * Stop writeback when nr_pages has been consumed
2028 if (work->nr_pages <= 0)
2032 * Background writeout and kupdate-style writeback may
2033 * run forever. Stop them if there is other work to do
2034 * so that e.g. sync can proceed. They'll be restarted
2035 * after the other works are all done.
2037 if ((work->for_background || work->for_kupdate) &&
2038 !list_empty(&wb->work_list))
2042 * For background writeout, stop when we are below the
2043 * background dirty threshold
2045 if (work->for_background && !wb_over_bg_thresh(wb))
2049 * Kupdate and background works are special and we want to
2050 * include all inodes that need writing. Livelock avoidance is
2051 * handled by these works yielding to any other work so we are
2054 if (work->for_kupdate) {
2055 dirtied_before = jiffies -
2056 msecs_to_jiffies(dirty_expire_interval * 10);
2057 } else if (work->for_background)
2058 dirtied_before = jiffies;
2060 trace_writeback_start(wb, work);
2061 if (list_empty(&wb->b_io))
2062 queue_io(wb, work, dirtied_before);
2064 progress = writeback_sb_inodes(work->sb, wb, work);
2066 progress = __writeback_inodes_wb(wb, work);
2067 trace_writeback_written(wb, work);
2070 * Did we write something? Try for more
2072 * Dirty inodes are moved to b_io for writeback in batches.
2073 * The completion of the current batch does not necessarily
2074 * mean the overall work is done. So we keep looping as long
2075 * as made some progress on cleaning pages or inodes.
2080 * No more inodes for IO, bail
2082 if (list_empty(&wb->b_more_io))
2085 * Nothing written. Wait for some inode to
2086 * become available for writeback. Otherwise
2087 * we'll just busyloop.
2089 trace_writeback_wait(wb, work);
2090 inode = wb_inode(wb->b_more_io.prev);
2091 spin_lock(&inode->i_lock);
2092 spin_unlock(&wb->list_lock);
2093 /* This function drops i_lock... */
2094 inode_sleep_on_writeback(inode);
2095 spin_lock(&wb->list_lock);
2097 spin_unlock(&wb->list_lock);
2098 blk_finish_plug(&plug);
2100 return nr_pages - work->nr_pages;
2104 * Return the next wb_writeback_work struct that hasn't been processed yet.
2106 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2108 struct wb_writeback_work *work = NULL;
2110 spin_lock_bh(&wb->work_lock);
2111 if (!list_empty(&wb->work_list)) {
2112 work = list_entry(wb->work_list.next,
2113 struct wb_writeback_work, list);
2114 list_del_init(&work->list);
2116 spin_unlock_bh(&wb->work_lock);
2120 static long wb_check_background_flush(struct bdi_writeback *wb)
2122 if (wb_over_bg_thresh(wb)) {
2124 struct wb_writeback_work work = {
2125 .nr_pages = LONG_MAX,
2126 .sync_mode = WB_SYNC_NONE,
2127 .for_background = 1,
2129 .reason = WB_REASON_BACKGROUND,
2132 return wb_writeback(wb, &work);
2138 static long wb_check_old_data_flush(struct bdi_writeback *wb)
2140 unsigned long expired;
2144 * When set to zero, disable periodic writeback
2146 if (!dirty_writeback_interval)
2149 expired = wb->last_old_flush +
2150 msecs_to_jiffies(dirty_writeback_interval * 10);
2151 if (time_before(jiffies, expired))
2154 wb->last_old_flush = jiffies;
2155 nr_pages = get_nr_dirty_pages();
2158 struct wb_writeback_work work = {
2159 .nr_pages = nr_pages,
2160 .sync_mode = WB_SYNC_NONE,
2163 .reason = WB_REASON_PERIODIC,
2166 return wb_writeback(wb, &work);
2172 static long wb_check_start_all(struct bdi_writeback *wb)
2176 if (!test_bit(WB_start_all, &wb->state))
2179 nr_pages = get_nr_dirty_pages();
2181 struct wb_writeback_work work = {
2182 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2183 .sync_mode = WB_SYNC_NONE,
2185 .reason = wb->start_all_reason,
2188 nr_pages = wb_writeback(wb, &work);
2191 clear_bit(WB_start_all, &wb->state);
2197 * Retrieve work items and do the writeback they describe
2199 static long wb_do_writeback(struct bdi_writeback *wb)
2201 struct wb_writeback_work *work;
2204 set_bit(WB_writeback_running, &wb->state);
2205 while ((work = get_next_work_item(wb)) != NULL) {
2206 trace_writeback_exec(wb, work);
2207 wrote += wb_writeback(wb, work);
2208 finish_writeback_work(wb, work);
2212 * Check for a flush-everything request
2214 wrote += wb_check_start_all(wb);
2217 * Check for periodic writeback, kupdated() style
2219 wrote += wb_check_old_data_flush(wb);
2220 wrote += wb_check_background_flush(wb);
2221 clear_bit(WB_writeback_running, &wb->state);
2227 * Handle writeback of dirty data for the device backed by this bdi. Also
2228 * reschedules periodically and does kupdated style flushing.
2230 void wb_workfn(struct work_struct *work)
2232 struct bdi_writeback *wb = container_of(to_delayed_work(work),
2233 struct bdi_writeback, dwork);
2236 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2237 current->flags |= PF_SWAPWRITE;
2239 if (likely(!current_is_workqueue_rescuer() ||
2240 !test_bit(WB_registered, &wb->state))) {
2242 * The normal path. Keep writing back @wb until its
2243 * work_list is empty. Note that this path is also taken
2244 * if @wb is shutting down even when we're running off the
2245 * rescuer as work_list needs to be drained.
2248 pages_written = wb_do_writeback(wb);
2249 trace_writeback_pages_written(pages_written);
2250 } while (!list_empty(&wb->work_list));
2253 * bdi_wq can't get enough workers and we're running off
2254 * the emergency worker. Don't hog it. Hopefully, 1024 is
2255 * enough for efficient IO.
2257 pages_written = writeback_inodes_wb(wb, 1024,
2258 WB_REASON_FORKER_THREAD);
2259 trace_writeback_pages_written(pages_written);
2262 if (!list_empty(&wb->work_list))
2264 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2265 wb_wakeup_delayed(wb);
2267 current->flags &= ~PF_SWAPWRITE;
2271 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2272 * write back the whole world.
2274 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2275 enum wb_reason reason)
2277 struct bdi_writeback *wb;
2279 if (!bdi_has_dirty_io(bdi))
2282 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2283 wb_start_writeback(wb, reason);
2286 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2287 enum wb_reason reason)
2290 __wakeup_flusher_threads_bdi(bdi, reason);
2295 * Wakeup the flusher threads to start writeback of all currently dirty pages
2297 void wakeup_flusher_threads(enum wb_reason reason)
2299 struct backing_dev_info *bdi;
2302 * If we are expecting writeback progress we must submit plugged IO.
2304 if (blk_needs_flush_plug(current))
2305 blk_flush_plug(current->plug, true);
2308 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2309 __wakeup_flusher_threads_bdi(bdi, reason);
2314 * Wake up bdi's periodically to make sure dirtytime inodes gets
2315 * written back periodically. We deliberately do *not* check the
2316 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2317 * kernel to be constantly waking up once there are any dirtytime
2318 * inodes on the system. So instead we define a separate delayed work
2319 * function which gets called much more rarely. (By default, only
2320 * once every 12 hours.)
2322 * If there is any other write activity going on in the file system,
2323 * this function won't be necessary. But if the only thing that has
2324 * happened on the file system is a dirtytime inode caused by an atime
2325 * update, we need this infrastructure below to make sure that inode
2326 * eventually gets pushed out to disk.
2328 static void wakeup_dirtytime_writeback(struct work_struct *w);
2329 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2331 static void wakeup_dirtytime_writeback(struct work_struct *w)
2333 struct backing_dev_info *bdi;
2336 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2337 struct bdi_writeback *wb;
2339 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2340 if (!list_empty(&wb->b_dirty_time))
2344 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2347 static int __init start_dirtytime_writeback(void)
2349 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2352 __initcall(start_dirtytime_writeback);
2354 int dirtytime_interval_handler(struct ctl_table *table, int write,
2355 void *buffer, size_t *lenp, loff_t *ppos)
2359 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2360 if (ret == 0 && write)
2361 mod_delayed_work(system_wq, &dirtytime_work, 0);
2366 * __mark_inode_dirty - internal function to mark an inode dirty
2368 * @inode: inode to mark
2369 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2370 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2371 * with I_DIRTY_PAGES.
2373 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2374 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2376 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2377 * instead of calling this directly.
2379 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2380 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2381 * even if they are later hashed, as they will have been marked dirty already.
2383 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2385 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2386 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2387 * the kernel-internal blockdev inode represents the dirtying time of the
2388 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2389 * page->mapping->host, so the page-dirtying time is recorded in the internal
2392 void __mark_inode_dirty(struct inode *inode, int flags)
2394 struct super_block *sb = inode->i_sb;
2397 trace_writeback_mark_inode_dirty(inode, flags);
2399 if (flags & I_DIRTY_INODE) {
2401 * Notify the filesystem about the inode being dirtied, so that
2402 * (if needed) it can update on-disk fields and journal the
2403 * inode. This is only needed when the inode itself is being
2404 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2405 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2407 trace_writeback_dirty_inode_start(inode, flags);
2408 if (sb->s_op->dirty_inode)
2409 sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE);
2410 trace_writeback_dirty_inode(inode, flags);
2412 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2413 flags &= ~I_DIRTY_TIME;
2416 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2417 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2418 * in one call to __mark_inode_dirty().)
2420 dirtytime = flags & I_DIRTY_TIME;
2421 WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2425 * Paired with smp_mb() in __writeback_single_inode() for the
2426 * following lockless i_state test. See there for details.
2430 if (((inode->i_state & flags) == flags) ||
2431 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2434 spin_lock(&inode->i_lock);
2435 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2436 goto out_unlock_inode;
2437 if ((inode->i_state & flags) != flags) {
2438 const int was_dirty = inode->i_state & I_DIRTY;
2440 inode_attach_wb(inode, NULL);
2442 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2443 if (flags & I_DIRTY_INODE)
2444 inode->i_state &= ~I_DIRTY_TIME;
2445 inode->i_state |= flags;
2448 * If the inode is queued for writeback by flush worker, just
2449 * update its dirty state. Once the flush worker is done with
2450 * the inode it will place it on the appropriate superblock
2451 * list, based upon its state.
2453 if (inode->i_state & I_SYNC_QUEUED)
2454 goto out_unlock_inode;
2457 * Only add valid (hashed) inodes to the superblock's
2458 * dirty list. Add blockdev inodes as well.
2460 if (!S_ISBLK(inode->i_mode)) {
2461 if (inode_unhashed(inode))
2462 goto out_unlock_inode;
2464 if (inode->i_state & I_FREEING)
2465 goto out_unlock_inode;
2468 * If the inode was already on b_dirty/b_io/b_more_io, don't
2469 * reposition it (that would break b_dirty time-ordering).
2472 struct bdi_writeback *wb;
2473 struct list_head *dirty_list;
2474 bool wakeup_bdi = false;
2476 wb = locked_inode_to_wb_and_lock_list(inode);
2478 inode->dirtied_when = jiffies;
2480 inode->dirtied_time_when = jiffies;
2482 if (inode->i_state & I_DIRTY)
2483 dirty_list = &wb->b_dirty;
2485 dirty_list = &wb->b_dirty_time;
2487 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2490 spin_unlock(&wb->list_lock);
2491 trace_writeback_dirty_inode_enqueue(inode);
2494 * If this is the first dirty inode for this bdi,
2495 * we have to wake-up the corresponding bdi thread
2496 * to make sure background write-back happens
2500 (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2501 wb_wakeup_delayed(wb);
2506 spin_unlock(&inode->i_lock);
2508 EXPORT_SYMBOL(__mark_inode_dirty);
2511 * The @s_sync_lock is used to serialise concurrent sync operations
2512 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2513 * Concurrent callers will block on the s_sync_lock rather than doing contending
2514 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2515 * has been issued up to the time this function is enter is guaranteed to be
2516 * completed by the time we have gained the lock and waited for all IO that is
2517 * in progress regardless of the order callers are granted the lock.
2519 static void wait_sb_inodes(struct super_block *sb)
2521 LIST_HEAD(sync_list);
2524 * We need to be protected against the filesystem going from
2525 * r/o to r/w or vice versa.
2527 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2529 mutex_lock(&sb->s_sync_lock);
2532 * Splice the writeback list onto a temporary list to avoid waiting on
2533 * inodes that have started writeback after this point.
2535 * Use rcu_read_lock() to keep the inodes around until we have a
2536 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2537 * the local list because inodes can be dropped from either by writeback
2541 spin_lock_irq(&sb->s_inode_wblist_lock);
2542 list_splice_init(&sb->s_inodes_wb, &sync_list);
2545 * Data integrity sync. Must wait for all pages under writeback, because
2546 * there may have been pages dirtied before our sync call, but which had
2547 * writeout started before we write it out. In which case, the inode
2548 * may not be on the dirty list, but we still have to wait for that
2551 while (!list_empty(&sync_list)) {
2552 struct inode *inode = list_first_entry(&sync_list, struct inode,
2554 struct address_space *mapping = inode->i_mapping;
2557 * Move each inode back to the wb list before we drop the lock
2558 * to preserve consistency between i_wb_list and the mapping
2559 * writeback tag. Writeback completion is responsible to remove
2560 * the inode from either list once the writeback tag is cleared.
2562 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2565 * The mapping can appear untagged while still on-list since we
2566 * do not have the mapping lock. Skip it here, wb completion
2569 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2572 spin_unlock_irq(&sb->s_inode_wblist_lock);
2574 spin_lock(&inode->i_lock);
2575 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2576 spin_unlock(&inode->i_lock);
2578 spin_lock_irq(&sb->s_inode_wblist_lock);
2582 spin_unlock(&inode->i_lock);
2586 * We keep the error status of individual mapping so that
2587 * applications can catch the writeback error using fsync(2).
2588 * See filemap_fdatawait_keep_errors() for details.
2590 filemap_fdatawait_keep_errors(mapping);
2597 spin_lock_irq(&sb->s_inode_wblist_lock);
2599 spin_unlock_irq(&sb->s_inode_wblist_lock);
2601 mutex_unlock(&sb->s_sync_lock);
2604 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2605 enum wb_reason reason, bool skip_if_busy)
2607 struct backing_dev_info *bdi = sb->s_bdi;
2608 DEFINE_WB_COMPLETION(done, bdi);
2609 struct wb_writeback_work work = {
2611 .sync_mode = WB_SYNC_NONE,
2612 .tagged_writepages = 1,
2618 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2620 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2622 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2623 wb_wait_for_completion(&done);
2627 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2628 * @sb: the superblock
2629 * @nr: the number of pages to write
2630 * @reason: reason why some writeback work initiated
2632 * Start writeback on some inodes on this super_block. No guarantees are made
2633 * on how many (if any) will be written, and this function does not wait
2634 * for IO completion of submitted IO.
2636 void writeback_inodes_sb_nr(struct super_block *sb,
2638 enum wb_reason reason)
2640 __writeback_inodes_sb_nr(sb, nr, reason, false);
2642 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2645 * writeback_inodes_sb - writeback dirty inodes from given super_block
2646 * @sb: the superblock
2647 * @reason: reason why some writeback work was initiated
2649 * Start writeback on some inodes on this super_block. No guarantees are made
2650 * on how many (if any) will be written, and this function does not wait
2651 * for IO completion of submitted IO.
2653 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2655 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2657 EXPORT_SYMBOL(writeback_inodes_sb);
2660 * try_to_writeback_inodes_sb - try to start writeback if none underway
2661 * @sb: the superblock
2662 * @reason: reason why some writeback work was initiated
2664 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2666 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2668 if (!down_read_trylock(&sb->s_umount))
2671 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2672 up_read(&sb->s_umount);
2674 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2677 * sync_inodes_sb - sync sb inode pages
2678 * @sb: the superblock
2680 * This function writes and waits on any dirty inode belonging to this
2683 void sync_inodes_sb(struct super_block *sb)
2685 struct backing_dev_info *bdi = sb->s_bdi;
2686 DEFINE_WB_COMPLETION(done, bdi);
2687 struct wb_writeback_work work = {
2689 .sync_mode = WB_SYNC_ALL,
2690 .nr_pages = LONG_MAX,
2693 .reason = WB_REASON_SYNC,
2698 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2699 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2700 * bdi_has_dirty() need to be written out too.
2702 if (bdi == &noop_backing_dev_info)
2704 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2706 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2707 bdi_down_write_wb_switch_rwsem(bdi);
2708 bdi_split_work_to_wbs(bdi, &work, false);
2709 wb_wait_for_completion(&done);
2710 bdi_up_write_wb_switch_rwsem(bdi);
2714 EXPORT_SYMBOL(sync_inodes_sb);
2717 * write_inode_now - write an inode to disk
2718 * @inode: inode to write to disk
2719 * @sync: whether the write should be synchronous or not
2721 * This function commits an inode to disk immediately if it is dirty. This is
2722 * primarily needed by knfsd.
2724 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2726 int write_inode_now(struct inode *inode, int sync)
2728 struct writeback_control wbc = {
2729 .nr_to_write = LONG_MAX,
2730 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2732 .range_end = LLONG_MAX,
2735 if (!mapping_can_writeback(inode->i_mapping))
2736 wbc.nr_to_write = 0;
2739 return writeback_single_inode(inode, &wbc);
2741 EXPORT_SYMBOL(write_inode_now);
2744 * sync_inode_metadata - write an inode to disk
2745 * @inode: the inode to sync
2746 * @wait: wait for I/O to complete.
2748 * Write an inode to disk and adjust its dirty state after completion.
2750 * Note: only writes the actual inode, no associated data or other metadata.
2752 int sync_inode_metadata(struct inode *inode, int wait)
2754 struct writeback_control wbc = {
2755 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2756 .nr_to_write = 0, /* metadata-only */
2759 return writeback_single_inode(inode, &wbc);
2761 EXPORT_SYMBOL(sync_inode_metadata);