4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
29 #include <linux/device.h>
30 #include <linux/memcontrol.h>
34 * 4MB minimal write chunk size
36 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
38 struct wb_completion {
43 * Passed into wb_writeback(), essentially a subset of writeback_control
45 struct wb_writeback_work {
47 struct super_block *sb;
48 unsigned long *older_than_this;
49 enum writeback_sync_modes sync_mode;
50 unsigned int tagged_writepages:1;
51 unsigned int for_kupdate:1;
52 unsigned int range_cyclic:1;
53 unsigned int for_background:1;
54 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
55 unsigned int auto_free:1; /* free on completion */
56 unsigned int single_wait:1;
57 unsigned int single_done:1;
58 enum wb_reason reason; /* why was writeback initiated? */
60 struct list_head list; /* pending work list */
61 struct wb_completion *done; /* set if the caller waits */
65 * If one wants to wait for one or more wb_writeback_works, each work's
66 * ->done should be set to a wb_completion defined using the following
67 * macro. Once all work items are issued with wb_queue_work(), the caller
68 * can wait for the completion of all using wb_wait_for_completion(). Work
69 * items which are waited upon aren't freed automatically on completion.
71 #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
72 struct wb_completion cmpl = { \
73 .cnt = ATOMIC_INIT(1), \
78 * If an inode is constantly having its pages dirtied, but then the
79 * updates stop dirtytime_expire_interval seconds in the past, it's
80 * possible for the worst case time between when an inode has its
81 * timestamps updated and when they finally get written out to be two
82 * dirtytime_expire_intervals. We set the default to 12 hours (in
83 * seconds), which means most of the time inodes will have their
84 * timestamps written to disk after 12 hours, but in the worst case a
85 * few inodes might not their timestamps updated for 24 hours.
87 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
89 static inline struct inode *wb_inode(struct list_head *head)
91 return list_entry(head, struct inode, i_wb_list);
95 * Include the creation of the trace points after defining the
96 * wb_writeback_work structure and inline functions so that the definition
97 * remains local to this file.
99 #define CREATE_TRACE_POINTS
100 #include <trace/events/writeback.h>
102 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
104 static bool wb_io_lists_populated(struct bdi_writeback *wb)
106 if (wb_has_dirty_io(wb)) {
109 set_bit(WB_has_dirty_io, &wb->state);
110 WARN_ON_ONCE(!wb->avg_write_bandwidth);
111 atomic_long_add(wb->avg_write_bandwidth,
112 &wb->bdi->tot_write_bandwidth);
117 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
119 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
120 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
121 clear_bit(WB_has_dirty_io, &wb->state);
122 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
123 &wb->bdi->tot_write_bandwidth) < 0);
128 * inode_wb_list_move_locked - move an inode onto a bdi_writeback IO list
129 * @inode: inode to be moved
130 * @wb: target bdi_writeback
131 * @head: one of @wb->b_{dirty|io|more_io}
133 * Move @inode->i_wb_list to @list of @wb and set %WB_has_dirty_io.
134 * Returns %true if @inode is the first occupant of the !dirty_time IO
135 * lists; otherwise, %false.
137 static bool inode_wb_list_move_locked(struct inode *inode,
138 struct bdi_writeback *wb,
139 struct list_head *head)
141 assert_spin_locked(&wb->list_lock);
143 list_move(&inode->i_wb_list, head);
145 /* dirty_time doesn't count as dirty_io until expiration */
146 if (head != &wb->b_dirty_time)
147 return wb_io_lists_populated(wb);
149 wb_io_lists_depopulated(wb);
154 * inode_wb_list_del_locked - remove an inode from its bdi_writeback IO list
155 * @inode: inode to be removed
156 * @wb: bdi_writeback @inode is being removed from
158 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
159 * clear %WB_has_dirty_io if all are empty afterwards.
161 static void inode_wb_list_del_locked(struct inode *inode,
162 struct bdi_writeback *wb)
164 assert_spin_locked(&wb->list_lock);
166 list_del_init(&inode->i_wb_list);
167 wb_io_lists_depopulated(wb);
170 static void wb_wakeup(struct bdi_writeback *wb)
172 spin_lock_bh(&wb->work_lock);
173 if (test_bit(WB_registered, &wb->state))
174 mod_delayed_work(bdi_wq, &wb->dwork, 0);
175 spin_unlock_bh(&wb->work_lock);
178 static void wb_queue_work(struct bdi_writeback *wb,
179 struct wb_writeback_work *work)
181 trace_writeback_queue(wb->bdi, work);
183 spin_lock_bh(&wb->work_lock);
184 if (!test_bit(WB_registered, &wb->state)) {
185 if (work->single_wait)
186 work->single_done = 1;
190 atomic_inc(&work->done->cnt);
191 list_add_tail(&work->list, &wb->work_list);
192 mod_delayed_work(bdi_wq, &wb->dwork, 0);
194 spin_unlock_bh(&wb->work_lock);
198 * wb_wait_for_completion - wait for completion of bdi_writeback_works
199 * @bdi: bdi work items were issued to
200 * @done: target wb_completion
202 * Wait for one or more work items issued to @bdi with their ->done field
203 * set to @done, which should have been defined with
204 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
205 * work items are completed. Work items which are waited upon aren't freed
206 * automatically on completion.
208 static void wb_wait_for_completion(struct backing_dev_info *bdi,
209 struct wb_completion *done)
211 atomic_dec(&done->cnt); /* put down the initial count */
212 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
215 #ifdef CONFIG_CGROUP_WRITEBACK
217 /* parameters for foreign inode detection, see wb_detach_inode() */
218 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
219 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
220 #define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
221 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
223 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
224 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
225 /* each slot's duration is 2s / 16 */
226 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
227 /* if foreign slots >= 8, switch */
228 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
229 /* one round can affect upto 5 slots */
231 void __inode_attach_wb(struct inode *inode, struct page *page)
233 struct backing_dev_info *bdi = inode_to_bdi(inode);
234 struct bdi_writeback *wb = NULL;
236 if (inode_cgwb_enabled(inode)) {
237 struct cgroup_subsys_state *memcg_css;
240 memcg_css = mem_cgroup_css_from_page(page);
241 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
243 /* must pin memcg_css, see wb_get_create() */
244 memcg_css = task_get_css(current, memory_cgrp_id);
245 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
254 * There may be multiple instances of this function racing to
255 * update the same inode. Use cmpxchg() to tell the winner.
257 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
262 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
263 * @inode: inode of interest with i_lock held
265 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
266 * held on entry and is released on return. The returned wb is guaranteed
267 * to stay @inode's associated wb until its list_lock is released.
269 static struct bdi_writeback *
270 locked_inode_to_wb_and_lock_list(struct inode *inode)
271 __releases(&inode->i_lock)
272 __acquires(&wb->list_lock)
275 struct bdi_writeback *wb = inode_to_wb(inode);
278 * inode_to_wb() association is protected by both
279 * @inode->i_lock and @wb->list_lock but list_lock nests
280 * outside i_lock. Drop i_lock and verify that the
281 * association hasn't changed after acquiring list_lock.
284 spin_unlock(&inode->i_lock);
285 spin_lock(&wb->list_lock);
286 wb_put(wb); /* not gonna deref it anymore */
288 if (likely(wb == inode_to_wb(inode)))
289 return wb; /* @inode already has ref */
291 spin_unlock(&wb->list_lock);
293 spin_lock(&inode->i_lock);
298 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
299 * @inode: inode of interest
301 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
304 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
305 __acquires(&wb->list_lock)
307 spin_lock(&inode->i_lock);
308 return locked_inode_to_wb_and_lock_list(inode);
311 struct inode_switch_wbs_context {
313 struct bdi_writeback *new_wb;
315 struct rcu_head rcu_head;
316 struct work_struct work;
319 static void inode_switch_wbs_work_fn(struct work_struct *work)
321 struct inode_switch_wbs_context *isw =
322 container_of(work, struct inode_switch_wbs_context, work);
323 struct inode *inode = isw->inode;
324 struct bdi_writeback *new_wb = isw->new_wb;
327 * By the time control reaches here, RCU grace period has passed
328 * since I_WB_SWITCH assertion and all wb stat update transactions
329 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
330 * synchronizing against mapping->tree_lock.
332 spin_lock(&inode->i_lock);
334 inode->i_wb_frn_winner = 0;
335 inode->i_wb_frn_avg_time = 0;
336 inode->i_wb_frn_history = 0;
339 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
340 * ensures that the new wb is visible if they see !I_WB_SWITCH.
342 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
344 spin_unlock(&inode->i_lock);
351 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
353 struct inode_switch_wbs_context *isw = container_of(rcu_head,
354 struct inode_switch_wbs_context, rcu_head);
356 /* needs to grab bh-unsafe locks, bounce to work item */
357 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
358 schedule_work(&isw->work);
362 * inode_switch_wbs - change the wb association of an inode
363 * @inode: target inode
364 * @new_wb_id: ID of the new wb
366 * Switch @inode's wb association to the wb identified by @new_wb_id. The
367 * switching is performed asynchronously and may fail silently.
369 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
371 struct backing_dev_info *bdi = inode_to_bdi(inode);
372 struct cgroup_subsys_state *memcg_css;
373 struct inode_switch_wbs_context *isw;
375 /* noop if seems to be already in progress */
376 if (inode->i_state & I_WB_SWITCH)
379 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
383 /* find and pin the new wb */
385 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
387 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
392 /* while holding I_WB_SWITCH, no one else can update the association */
393 spin_lock(&inode->i_lock);
394 if (inode->i_state & (I_WB_SWITCH | I_FREEING) ||
395 inode_to_wb(inode) == isw->new_wb) {
396 spin_unlock(&inode->i_lock);
399 inode->i_state |= I_WB_SWITCH;
400 spin_unlock(&inode->i_lock);
406 * In addition to synchronizing among switchers, I_WB_SWITCH tells
407 * the RCU protected stat update paths to grab the mapping's
408 * tree_lock so that stat transfer can synchronize against them.
409 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
411 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
421 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
422 * @wbc: writeback_control of interest
423 * @inode: target inode
425 * @inode is locked and about to be written back under the control of @wbc.
426 * Record @inode's writeback context into @wbc and unlock the i_lock. On
427 * writeback completion, wbc_detach_inode() should be called. This is used
428 * to track the cgroup writeback context.
430 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
433 wbc->wb = inode_to_wb(inode);
436 wbc->wb_id = wbc->wb->memcg_css->id;
437 wbc->wb_lcand_id = inode->i_wb_frn_winner;
438 wbc->wb_tcand_id = 0;
440 wbc->wb_lcand_bytes = 0;
441 wbc->wb_tcand_bytes = 0;
444 spin_unlock(&inode->i_lock);
448 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
449 * @wbc: writeback_control of the just finished writeback
451 * To be called after a writeback attempt of an inode finishes and undoes
452 * wbc_attach_and_unlock_inode(). Can be called under any context.
454 * As concurrent write sharing of an inode is expected to be very rare and
455 * memcg only tracks page ownership on first-use basis severely confining
456 * the usefulness of such sharing, cgroup writeback tracks ownership
457 * per-inode. While the support for concurrent write sharing of an inode
458 * is deemed unnecessary, an inode being written to by different cgroups at
459 * different points in time is a lot more common, and, more importantly,
460 * charging only by first-use can too readily lead to grossly incorrect
461 * behaviors (single foreign page can lead to gigabytes of writeback to be
462 * incorrectly attributed).
464 * To resolve this issue, cgroup writeback detects the majority dirtier of
465 * an inode and transfers the ownership to it. To avoid unnnecessary
466 * oscillation, the detection mechanism keeps track of history and gives
467 * out the switch verdict only if the foreign usage pattern is stable over
468 * a certain amount of time and/or writeback attempts.
470 * On each writeback attempt, @wbc tries to detect the majority writer
471 * using Boyer-Moore majority vote algorithm. In addition to the byte
472 * count from the majority voting, it also counts the bytes written for the
473 * current wb and the last round's winner wb (max of last round's current
474 * wb, the winner from two rounds ago, and the last round's majority
475 * candidate). Keeping track of the historical winner helps the algorithm
476 * to semi-reliably detect the most active writer even when it's not the
479 * Once the winner of the round is determined, whether the winner is
480 * foreign or not and how much IO time the round consumed is recorded in
481 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
482 * over a certain threshold, the switch verdict is given.
484 void wbc_detach_inode(struct writeback_control *wbc)
486 struct bdi_writeback *wb = wbc->wb;
487 struct inode *inode = wbc->inode;
488 u16 history = inode->i_wb_frn_history;
489 unsigned long avg_time = inode->i_wb_frn_avg_time;
490 unsigned long max_bytes, max_time;
493 /* pick the winner of this round */
494 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
495 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
497 max_bytes = wbc->wb_bytes;
498 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
499 max_id = wbc->wb_lcand_id;
500 max_bytes = wbc->wb_lcand_bytes;
502 max_id = wbc->wb_tcand_id;
503 max_bytes = wbc->wb_tcand_bytes;
507 * Calculate the amount of IO time the winner consumed and fold it
508 * into the running average kept per inode. If the consumed IO
509 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
510 * deciding whether to switch or not. This is to prevent one-off
511 * small dirtiers from skewing the verdict.
513 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
514 wb->avg_write_bandwidth);
516 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
517 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
519 avg_time = max_time; /* immediate catch up on first run */
521 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
525 * The switch verdict is reached if foreign wb's consume
526 * more than a certain proportion of IO time in a
527 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
528 * history mask where each bit represents one sixteenth of
529 * the period. Determine the number of slots to shift into
530 * history from @max_time.
532 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
533 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
535 if (wbc->wb_id != max_id)
536 history |= (1U << slots) - 1;
539 * Switch if the current wb isn't the consistent winner.
540 * If there are multiple closely competing dirtiers, the
541 * inode may switch across them repeatedly over time, which
542 * is okay. The main goal is avoiding keeping an inode on
543 * the wrong wb for an extended period of time.
545 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
546 inode_switch_wbs(inode, max_id);
550 * Multiple instances of this function may race to update the
551 * following fields but we don't mind occassional inaccuracies.
553 inode->i_wb_frn_winner = max_id;
554 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
555 inode->i_wb_frn_history = history;
562 * wbc_account_io - account IO issued during writeback
563 * @wbc: writeback_control of the writeback in progress
564 * @page: page being written out
565 * @bytes: number of bytes being written out
567 * @bytes from @page are about to written out during the writeback
568 * controlled by @wbc. Keep the book for foreign inode detection. See
569 * wbc_detach_inode().
571 void wbc_account_io(struct writeback_control *wbc, struct page *page,
577 * pageout() path doesn't attach @wbc to the inode being written
578 * out. This is intentional as we don't want the function to block
579 * behind a slow cgroup. Ultimately, we want pageout() to kick off
580 * regular writeback instead of writing things out itself.
586 id = mem_cgroup_css_from_page(page)->id;
589 if (id == wbc->wb_id) {
590 wbc->wb_bytes += bytes;
594 if (id == wbc->wb_lcand_id)
595 wbc->wb_lcand_bytes += bytes;
597 /* Boyer-Moore majority vote algorithm */
598 if (!wbc->wb_tcand_bytes)
599 wbc->wb_tcand_id = id;
600 if (id == wbc->wb_tcand_id)
601 wbc->wb_tcand_bytes += bytes;
603 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
607 * inode_congested - test whether an inode is congested
608 * @inode: inode to test for congestion
609 * @cong_bits: mask of WB_[a]sync_congested bits to test
611 * Tests whether @inode is congested. @cong_bits is the mask of congestion
612 * bits to test and the return value is the mask of set bits.
614 * If cgroup writeback is enabled for @inode, the congestion state is
615 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
616 * associated with @inode is congested; otherwise, the root wb's congestion
619 int inode_congested(struct inode *inode, int cong_bits)
622 struct bdi_writeback *wb = inode_to_wb(inode);
624 return wb_congested(wb, cong_bits);
627 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
629 EXPORT_SYMBOL_GPL(inode_congested);
632 * wb_wait_for_single_work - wait for completion of a single bdi_writeback_work
633 * @bdi: bdi the work item was issued to
634 * @work: work item to wait for
636 * Wait for the completion of @work which was issued to one of @bdi's
637 * bdi_writeback's. The caller must have set @work->single_wait before
638 * issuing it. This wait operates independently fo
639 * wb_wait_for_completion() and also disables automatic freeing of @work.
641 static void wb_wait_for_single_work(struct backing_dev_info *bdi,
642 struct wb_writeback_work *work)
644 if (WARN_ON_ONCE(!work->single_wait))
647 wait_event(bdi->wb_waitq, work->single_done);
650 * Paired with smp_wmb() in wb_do_writeback() and ensures that all
651 * modifications to @work prior to assertion of ->single_done is
652 * visible to the caller once this function returns.
658 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
659 * @wb: target bdi_writeback to split @nr_pages to
660 * @nr_pages: number of pages to write for the whole bdi
662 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
663 * relation to the total write bandwidth of all wb's w/ dirty inodes on
666 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
668 unsigned long this_bw = wb->avg_write_bandwidth;
669 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
671 if (nr_pages == LONG_MAX)
675 * This may be called on clean wb's and proportional distribution
676 * may not make sense, just use the original @nr_pages in those
677 * cases. In general, we wanna err on the side of writing more.
679 if (!tot_bw || this_bw >= tot_bw)
682 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
686 * wb_clone_and_queue_work - clone a wb_writeback_work and issue it to a wb
687 * @wb: target bdi_writeback
688 * @base_work: source wb_writeback_work
690 * Try to make a clone of @base_work and issue it to @wb. If cloning
691 * succeeds, %true is returned; otherwise, @base_work is issued directly
692 * and %false is returned. In the latter case, the caller is required to
693 * wait for @base_work's completion using wb_wait_for_single_work().
695 * A clone is auto-freed on completion. @base_work never is.
697 static bool wb_clone_and_queue_work(struct bdi_writeback *wb,
698 struct wb_writeback_work *base_work)
700 struct wb_writeback_work *work;
702 work = kmalloc(sizeof(*work), GFP_ATOMIC);
706 work->single_wait = 0;
710 work->single_wait = 1;
712 work->single_done = 0;
713 wb_queue_work(wb, work);
714 return work != base_work;
718 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
719 * @bdi: target backing_dev_info
720 * @base_work: wb_writeback_work to issue
721 * @skip_if_busy: skip wb's which already have writeback in progress
723 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
724 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
725 * distributed to the busy wbs according to each wb's proportion in the
726 * total active write bandwidth of @bdi.
728 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
729 struct wb_writeback_work *base_work,
732 long nr_pages = base_work->nr_pages;
733 int next_blkcg_id = 0;
734 struct bdi_writeback *wb;
739 if (!bdi_has_dirty_io(bdi))
743 bdi_for_each_wb(wb, bdi, &iter, next_blkcg_id) {
744 if (!wb_has_dirty_io(wb) ||
745 (skip_if_busy && writeback_in_progress(wb)))
748 base_work->nr_pages = wb_split_bdi_pages(wb, nr_pages);
749 if (!wb_clone_and_queue_work(wb, base_work)) {
750 next_blkcg_id = wb->blkcg_css->id + 1;
752 wb_wait_for_single_work(bdi, base_work);
759 #else /* CONFIG_CGROUP_WRITEBACK */
761 static struct bdi_writeback *
762 locked_inode_to_wb_and_lock_list(struct inode *inode)
763 __releases(&inode->i_lock)
764 __acquires(&wb->list_lock)
766 struct bdi_writeback *wb = inode_to_wb(inode);
768 spin_unlock(&inode->i_lock);
769 spin_lock(&wb->list_lock);
773 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
774 __acquires(&wb->list_lock)
776 struct bdi_writeback *wb = inode_to_wb(inode);
778 spin_lock(&wb->list_lock);
782 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
787 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
788 struct wb_writeback_work *base_work,
793 if (bdi_has_dirty_io(bdi) &&
794 (!skip_if_busy || !writeback_in_progress(&bdi->wb))) {
795 base_work->auto_free = 0;
796 base_work->single_wait = 0;
797 base_work->single_done = 0;
798 wb_queue_work(&bdi->wb, base_work);
802 #endif /* CONFIG_CGROUP_WRITEBACK */
804 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
805 bool range_cyclic, enum wb_reason reason)
807 struct wb_writeback_work *work;
809 if (!wb_has_dirty_io(wb))
813 * This is WB_SYNC_NONE writeback, so if allocation fails just
814 * wakeup the thread for old dirty data writeback
816 work = kzalloc(sizeof(*work), GFP_ATOMIC);
818 trace_writeback_nowork(wb->bdi);
823 work->sync_mode = WB_SYNC_NONE;
824 work->nr_pages = nr_pages;
825 work->range_cyclic = range_cyclic;
826 work->reason = reason;
829 wb_queue_work(wb, work);
833 * wb_start_background_writeback - start background writeback
834 * @wb: bdi_writback to write from
837 * This makes sure WB_SYNC_NONE background writeback happens. When
838 * this function returns, it is only guaranteed that for given wb
839 * some IO is happening if we are over background dirty threshold.
840 * Caller need not hold sb s_umount semaphore.
842 void wb_start_background_writeback(struct bdi_writeback *wb)
845 * We just wake up the flusher thread. It will perform background
846 * writeback as soon as there is no other work to do.
848 trace_writeback_wake_background(wb->bdi);
853 * Remove the inode from the writeback list it is on.
855 void inode_wb_list_del(struct inode *inode)
857 struct bdi_writeback *wb;
859 wb = inode_to_wb_and_lock_list(inode);
860 inode_wb_list_del_locked(inode, wb);
861 spin_unlock(&wb->list_lock);
865 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
866 * furthest end of its superblock's dirty-inode list.
868 * Before stamping the inode's ->dirtied_when, we check to see whether it is
869 * already the most-recently-dirtied inode on the b_dirty list. If that is
870 * the case then the inode must have been redirtied while it was being written
871 * out and we don't reset its dirtied_when.
873 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
875 if (!list_empty(&wb->b_dirty)) {
878 tail = wb_inode(wb->b_dirty.next);
879 if (time_before(inode->dirtied_when, tail->dirtied_when))
880 inode->dirtied_when = jiffies;
882 inode_wb_list_move_locked(inode, wb, &wb->b_dirty);
886 * requeue inode for re-scanning after bdi->b_io list is exhausted.
888 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
890 inode_wb_list_move_locked(inode, wb, &wb->b_more_io);
893 static void inode_sync_complete(struct inode *inode)
895 inode->i_state &= ~I_SYNC;
896 /* If inode is clean an unused, put it into LRU now... */
897 inode_add_lru(inode);
898 /* Waiters must see I_SYNC cleared before being woken up */
900 wake_up_bit(&inode->i_state, __I_SYNC);
903 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
905 bool ret = time_after(inode->dirtied_when, t);
908 * For inodes being constantly redirtied, dirtied_when can get stuck.
909 * It _appears_ to be in the future, but is actually in distant past.
910 * This test is necessary to prevent such wrapped-around relative times
911 * from permanently stopping the whole bdi writeback.
913 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
918 #define EXPIRE_DIRTY_ATIME 0x0001
921 * Move expired (dirtied before work->older_than_this) dirty inodes from
922 * @delaying_queue to @dispatch_queue.
924 static int move_expired_inodes(struct list_head *delaying_queue,
925 struct list_head *dispatch_queue,
927 struct wb_writeback_work *work)
929 unsigned long *older_than_this = NULL;
930 unsigned long expire_time;
932 struct list_head *pos, *node;
933 struct super_block *sb = NULL;
938 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
939 older_than_this = work->older_than_this;
940 else if (!work->for_sync) {
941 expire_time = jiffies - (dirtytime_expire_interval * HZ);
942 older_than_this = &expire_time;
944 while (!list_empty(delaying_queue)) {
945 inode = wb_inode(delaying_queue->prev);
946 if (older_than_this &&
947 inode_dirtied_after(inode, *older_than_this))
949 list_move(&inode->i_wb_list, &tmp);
951 if (flags & EXPIRE_DIRTY_ATIME)
952 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
953 if (sb_is_blkdev_sb(inode->i_sb))
955 if (sb && sb != inode->i_sb)
960 /* just one sb in list, splice to dispatch_queue and we're done */
962 list_splice(&tmp, dispatch_queue);
966 /* Move inodes from one superblock together */
967 while (!list_empty(&tmp)) {
968 sb = wb_inode(tmp.prev)->i_sb;
969 list_for_each_prev_safe(pos, node, &tmp) {
970 inode = wb_inode(pos);
971 if (inode->i_sb == sb)
972 list_move(&inode->i_wb_list, dispatch_queue);
980 * Queue all expired dirty inodes for io, eldest first.
982 * newly dirtied b_dirty b_io b_more_io
983 * =============> gf edc BA
985 * newly dirtied b_dirty b_io b_more_io
986 * =============> g fBAedc
988 * +--> dequeue for IO
990 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
994 assert_spin_locked(&wb->list_lock);
995 list_splice_init(&wb->b_more_io, &wb->b_io);
996 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
997 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
998 EXPIRE_DIRTY_ATIME, work);
1000 wb_io_lists_populated(wb);
1001 trace_writeback_queue_io(wb, work, moved);
1004 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1008 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1009 trace_writeback_write_inode_start(inode, wbc);
1010 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1011 trace_writeback_write_inode(inode, wbc);
1018 * Wait for writeback on an inode to complete. Called with i_lock held.
1019 * Caller must make sure inode cannot go away when we drop i_lock.
1021 static void __inode_wait_for_writeback(struct inode *inode)
1022 __releases(inode->i_lock)
1023 __acquires(inode->i_lock)
1025 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1026 wait_queue_head_t *wqh;
1028 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1029 while (inode->i_state & I_SYNC) {
1030 spin_unlock(&inode->i_lock);
1031 __wait_on_bit(wqh, &wq, bit_wait,
1032 TASK_UNINTERRUPTIBLE);
1033 spin_lock(&inode->i_lock);
1038 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1040 void inode_wait_for_writeback(struct inode *inode)
1042 spin_lock(&inode->i_lock);
1043 __inode_wait_for_writeback(inode);
1044 spin_unlock(&inode->i_lock);
1048 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1049 * held and drops it. It is aimed for callers not holding any inode reference
1050 * so once i_lock is dropped, inode can go away.
1052 static void inode_sleep_on_writeback(struct inode *inode)
1053 __releases(inode->i_lock)
1056 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1059 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1060 sleep = inode->i_state & I_SYNC;
1061 spin_unlock(&inode->i_lock);
1064 finish_wait(wqh, &wait);
1068 * Find proper writeback list for the inode depending on its current state and
1069 * possibly also change of its state while we were doing writeback. Here we
1070 * handle things such as livelock prevention or fairness of writeback among
1071 * inodes. This function can be called only by flusher thread - noone else
1072 * processes all inodes in writeback lists and requeueing inodes behind flusher
1073 * thread's back can have unexpected consequences.
1075 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1076 struct writeback_control *wbc)
1078 if (inode->i_state & I_FREEING)
1082 * Sync livelock prevention. Each inode is tagged and synced in one
1083 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1084 * the dirty time to prevent enqueue and sync it again.
1086 if ((inode->i_state & I_DIRTY) &&
1087 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1088 inode->dirtied_when = jiffies;
1090 if (wbc->pages_skipped) {
1092 * writeback is not making progress due to locked
1093 * buffers. Skip this inode for now.
1095 redirty_tail(inode, wb);
1099 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1101 * We didn't write back all the pages. nfs_writepages()
1102 * sometimes bales out without doing anything.
1104 if (wbc->nr_to_write <= 0) {
1105 /* Slice used up. Queue for next turn. */
1106 requeue_io(inode, wb);
1109 * Writeback blocked by something other than
1110 * congestion. Delay the inode for some time to
1111 * avoid spinning on the CPU (100% iowait)
1112 * retrying writeback of the dirty page/inode
1113 * that cannot be performed immediately.
1115 redirty_tail(inode, wb);
1117 } else if (inode->i_state & I_DIRTY) {
1119 * Filesystems can dirty the inode during writeback operations,
1120 * such as delayed allocation during submission or metadata
1121 * updates after data IO completion.
1123 redirty_tail(inode, wb);
1124 } else if (inode->i_state & I_DIRTY_TIME) {
1125 inode->dirtied_when = jiffies;
1126 inode_wb_list_move_locked(inode, wb, &wb->b_dirty_time);
1128 /* The inode is clean. Remove from writeback lists. */
1129 inode_wb_list_del_locked(inode, wb);
1134 * Write out an inode and its dirty pages. Do not update the writeback list
1135 * linkage. That is left to the caller. The caller is also responsible for
1136 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1139 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1141 struct address_space *mapping = inode->i_mapping;
1142 long nr_to_write = wbc->nr_to_write;
1146 WARN_ON(!(inode->i_state & I_SYNC));
1148 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1150 ret = do_writepages(mapping, wbc);
1153 * Make sure to wait on the data before writing out the metadata.
1154 * This is important for filesystems that modify metadata on data
1155 * I/O completion. We don't do it for sync(2) writeback because it has a
1156 * separate, external IO completion path and ->sync_fs for guaranteeing
1157 * inode metadata is written back correctly.
1159 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1160 int err = filemap_fdatawait(mapping);
1166 * Some filesystems may redirty the inode during the writeback
1167 * due to delalloc, clear dirty metadata flags right before
1170 spin_lock(&inode->i_lock);
1172 dirty = inode->i_state & I_DIRTY;
1173 if (inode->i_state & I_DIRTY_TIME) {
1174 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1175 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1176 unlikely(time_after(jiffies,
1177 (inode->dirtied_time_when +
1178 dirtytime_expire_interval * HZ)))) {
1179 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1180 trace_writeback_lazytime(inode);
1183 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1184 inode->i_state &= ~dirty;
1187 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1188 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1189 * either they see the I_DIRTY bits cleared or we see the dirtied
1192 * I_DIRTY_PAGES is always cleared together above even if @mapping
1193 * still has dirty pages. The flag is reinstated after smp_mb() if
1194 * necessary. This guarantees that either __mark_inode_dirty()
1195 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1199 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1200 inode->i_state |= I_DIRTY_PAGES;
1202 spin_unlock(&inode->i_lock);
1204 if (dirty & I_DIRTY_TIME)
1205 mark_inode_dirty_sync(inode);
1206 /* Don't write the inode if only I_DIRTY_PAGES was set */
1207 if (dirty & ~I_DIRTY_PAGES) {
1208 int err = write_inode(inode, wbc);
1212 trace_writeback_single_inode(inode, wbc, nr_to_write);
1217 * Write out an inode's dirty pages. Either the caller has an active reference
1218 * on the inode or the inode has I_WILL_FREE set.
1220 * This function is designed to be called for writing back one inode which
1221 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1222 * and does more profound writeback list handling in writeback_sb_inodes().
1225 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
1226 struct writeback_control *wbc)
1230 spin_lock(&inode->i_lock);
1231 if (!atomic_read(&inode->i_count))
1232 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1234 WARN_ON(inode->i_state & I_WILL_FREE);
1236 if (inode->i_state & I_SYNC) {
1237 if (wbc->sync_mode != WB_SYNC_ALL)
1240 * It's a data-integrity sync. We must wait. Since callers hold
1241 * inode reference or inode has I_WILL_FREE set, it cannot go
1244 __inode_wait_for_writeback(inode);
1246 WARN_ON(inode->i_state & I_SYNC);
1248 * Skip inode if it is clean and we have no outstanding writeback in
1249 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1250 * function since flusher thread may be doing for example sync in
1251 * parallel and if we move the inode, it could get skipped. So here we
1252 * make sure inode is on some writeback list and leave it there unless
1253 * we have completely cleaned the inode.
1255 if (!(inode->i_state & I_DIRTY_ALL) &&
1256 (wbc->sync_mode != WB_SYNC_ALL ||
1257 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1259 inode->i_state |= I_SYNC;
1260 wbc_attach_and_unlock_inode(wbc, inode);
1262 ret = __writeback_single_inode(inode, wbc);
1264 wbc_detach_inode(wbc);
1265 spin_lock(&wb->list_lock);
1266 spin_lock(&inode->i_lock);
1268 * If inode is clean, remove it from writeback lists. Otherwise don't
1269 * touch it. See comment above for explanation.
1271 if (!(inode->i_state & I_DIRTY_ALL))
1272 inode_wb_list_del_locked(inode, wb);
1273 spin_unlock(&wb->list_lock);
1274 inode_sync_complete(inode);
1276 spin_unlock(&inode->i_lock);
1280 static long writeback_chunk_size(struct bdi_writeback *wb,
1281 struct wb_writeback_work *work)
1286 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1287 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1288 * here avoids calling into writeback_inodes_wb() more than once.
1290 * The intended call sequence for WB_SYNC_ALL writeback is:
1293 * writeback_sb_inodes() <== called only once
1294 * write_cache_pages() <== called once for each inode
1295 * (quickly) tag currently dirty pages
1296 * (maybe slowly) sync all tagged pages
1298 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1301 pages = min(wb->avg_write_bandwidth / 2,
1302 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1303 pages = min(pages, work->nr_pages);
1304 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1305 MIN_WRITEBACK_PAGES);
1312 * Write a portion of b_io inodes which belong to @sb.
1314 * Return the number of pages and/or inodes written.
1316 static long writeback_sb_inodes(struct super_block *sb,
1317 struct bdi_writeback *wb,
1318 struct wb_writeback_work *work)
1320 struct writeback_control wbc = {
1321 .sync_mode = work->sync_mode,
1322 .tagged_writepages = work->tagged_writepages,
1323 .for_kupdate = work->for_kupdate,
1324 .for_background = work->for_background,
1325 .for_sync = work->for_sync,
1326 .range_cyclic = work->range_cyclic,
1328 .range_end = LLONG_MAX,
1330 unsigned long start_time = jiffies;
1332 long wrote = 0; /* count both pages and inodes */
1334 while (!list_empty(&wb->b_io)) {
1335 struct inode *inode = wb_inode(wb->b_io.prev);
1337 if (inode->i_sb != sb) {
1340 * We only want to write back data for this
1341 * superblock, move all inodes not belonging
1342 * to it back onto the dirty list.
1344 redirty_tail(inode, wb);
1349 * The inode belongs to a different superblock.
1350 * Bounce back to the caller to unpin this and
1351 * pin the next superblock.
1357 * Don't bother with new inodes or inodes being freed, first
1358 * kind does not need periodic writeout yet, and for the latter
1359 * kind writeout is handled by the freer.
1361 spin_lock(&inode->i_lock);
1362 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1363 spin_unlock(&inode->i_lock);
1364 redirty_tail(inode, wb);
1367 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1369 * If this inode is locked for writeback and we are not
1370 * doing writeback-for-data-integrity, move it to
1371 * b_more_io so that writeback can proceed with the
1372 * other inodes on s_io.
1374 * We'll have another go at writing back this inode
1375 * when we completed a full scan of b_io.
1377 spin_unlock(&inode->i_lock);
1378 requeue_io(inode, wb);
1379 trace_writeback_sb_inodes_requeue(inode);
1382 spin_unlock(&wb->list_lock);
1385 * We already requeued the inode if it had I_SYNC set and we
1386 * are doing WB_SYNC_NONE writeback. So this catches only the
1389 if (inode->i_state & I_SYNC) {
1390 /* Wait for I_SYNC. This function drops i_lock... */
1391 inode_sleep_on_writeback(inode);
1392 /* Inode may be gone, start again */
1393 spin_lock(&wb->list_lock);
1396 inode->i_state |= I_SYNC;
1397 wbc_attach_and_unlock_inode(&wbc, inode);
1399 write_chunk = writeback_chunk_size(wb, work);
1400 wbc.nr_to_write = write_chunk;
1401 wbc.pages_skipped = 0;
1404 * We use I_SYNC to pin the inode in memory. While it is set
1405 * evict_inode() will wait so the inode cannot be freed.
1407 __writeback_single_inode(inode, &wbc);
1409 wbc_detach_inode(&wbc);
1410 work->nr_pages -= write_chunk - wbc.nr_to_write;
1411 wrote += write_chunk - wbc.nr_to_write;
1412 spin_lock(&wb->list_lock);
1413 spin_lock(&inode->i_lock);
1414 if (!(inode->i_state & I_DIRTY_ALL))
1416 requeue_inode(inode, wb, &wbc);
1417 inode_sync_complete(inode);
1418 spin_unlock(&inode->i_lock);
1419 cond_resched_lock(&wb->list_lock);
1421 * bail out to wb_writeback() often enough to check
1422 * background threshold and other termination conditions.
1425 if (time_is_before_jiffies(start_time + HZ / 10UL))
1427 if (work->nr_pages <= 0)
1434 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1435 struct wb_writeback_work *work)
1437 unsigned long start_time = jiffies;
1440 while (!list_empty(&wb->b_io)) {
1441 struct inode *inode = wb_inode(wb->b_io.prev);
1442 struct super_block *sb = inode->i_sb;
1444 if (!trylock_super(sb)) {
1446 * trylock_super() may fail consistently due to
1447 * s_umount being grabbed by someone else. Don't use
1448 * requeue_io() to avoid busy retrying the inode/sb.
1450 redirty_tail(inode, wb);
1453 wrote += writeback_sb_inodes(sb, wb, work);
1454 up_read(&sb->s_umount);
1456 /* refer to the same tests at the end of writeback_sb_inodes */
1458 if (time_is_before_jiffies(start_time + HZ / 10UL))
1460 if (work->nr_pages <= 0)
1464 /* Leave any unwritten inodes on b_io */
1468 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1469 enum wb_reason reason)
1471 struct wb_writeback_work work = {
1472 .nr_pages = nr_pages,
1473 .sync_mode = WB_SYNC_NONE,
1478 spin_lock(&wb->list_lock);
1479 if (list_empty(&wb->b_io))
1480 queue_io(wb, &work);
1481 __writeback_inodes_wb(wb, &work);
1482 spin_unlock(&wb->list_lock);
1484 return nr_pages - work.nr_pages;
1488 * Explicit flushing or periodic writeback of "old" data.
1490 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1491 * dirtying-time in the inode's address_space. So this periodic writeback code
1492 * just walks the superblock inode list, writing back any inodes which are
1493 * older than a specific point in time.
1495 * Try to run once per dirty_writeback_interval. But if a writeback event
1496 * takes longer than a dirty_writeback_interval interval, then leave a
1499 * older_than_this takes precedence over nr_to_write. So we'll only write back
1500 * all dirty pages if they are all attached to "old" mappings.
1502 static long wb_writeback(struct bdi_writeback *wb,
1503 struct wb_writeback_work *work)
1505 unsigned long wb_start = jiffies;
1506 long nr_pages = work->nr_pages;
1507 unsigned long oldest_jif;
1508 struct inode *inode;
1511 oldest_jif = jiffies;
1512 work->older_than_this = &oldest_jif;
1514 spin_lock(&wb->list_lock);
1517 * Stop writeback when nr_pages has been consumed
1519 if (work->nr_pages <= 0)
1523 * Background writeout and kupdate-style writeback may
1524 * run forever. Stop them if there is other work to do
1525 * so that e.g. sync can proceed. They'll be restarted
1526 * after the other works are all done.
1528 if ((work->for_background || work->for_kupdate) &&
1529 !list_empty(&wb->work_list))
1533 * For background writeout, stop when we are below the
1534 * background dirty threshold
1536 if (work->for_background && !wb_over_bg_thresh(wb))
1540 * Kupdate and background works are special and we want to
1541 * include all inodes that need writing. Livelock avoidance is
1542 * handled by these works yielding to any other work so we are
1545 if (work->for_kupdate) {
1546 oldest_jif = jiffies -
1547 msecs_to_jiffies(dirty_expire_interval * 10);
1548 } else if (work->for_background)
1549 oldest_jif = jiffies;
1551 trace_writeback_start(wb->bdi, work);
1552 if (list_empty(&wb->b_io))
1555 progress = writeback_sb_inodes(work->sb, wb, work);
1557 progress = __writeback_inodes_wb(wb, work);
1558 trace_writeback_written(wb->bdi, work);
1560 wb_update_bandwidth(wb, wb_start);
1563 * Did we write something? Try for more
1565 * Dirty inodes are moved to b_io for writeback in batches.
1566 * The completion of the current batch does not necessarily
1567 * mean the overall work is done. So we keep looping as long
1568 * as made some progress on cleaning pages or inodes.
1573 * No more inodes for IO, bail
1575 if (list_empty(&wb->b_more_io))
1578 * Nothing written. Wait for some inode to
1579 * become available for writeback. Otherwise
1580 * we'll just busyloop.
1582 if (!list_empty(&wb->b_more_io)) {
1583 trace_writeback_wait(wb->bdi, work);
1584 inode = wb_inode(wb->b_more_io.prev);
1585 spin_lock(&inode->i_lock);
1586 spin_unlock(&wb->list_lock);
1587 /* This function drops i_lock... */
1588 inode_sleep_on_writeback(inode);
1589 spin_lock(&wb->list_lock);
1592 spin_unlock(&wb->list_lock);
1594 return nr_pages - work->nr_pages;
1598 * Return the next wb_writeback_work struct that hasn't been processed yet.
1600 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1602 struct wb_writeback_work *work = NULL;
1604 spin_lock_bh(&wb->work_lock);
1605 if (!list_empty(&wb->work_list)) {
1606 work = list_entry(wb->work_list.next,
1607 struct wb_writeback_work, list);
1608 list_del_init(&work->list);
1610 spin_unlock_bh(&wb->work_lock);
1615 * Add in the number of potentially dirty inodes, because each inode
1616 * write can dirty pagecache in the underlying blockdev.
1618 static unsigned long get_nr_dirty_pages(void)
1620 return global_page_state(NR_FILE_DIRTY) +
1621 global_page_state(NR_UNSTABLE_NFS) +
1622 get_nr_dirty_inodes();
1625 static long wb_check_background_flush(struct bdi_writeback *wb)
1627 if (wb_over_bg_thresh(wb)) {
1629 struct wb_writeback_work work = {
1630 .nr_pages = LONG_MAX,
1631 .sync_mode = WB_SYNC_NONE,
1632 .for_background = 1,
1634 .reason = WB_REASON_BACKGROUND,
1637 return wb_writeback(wb, &work);
1643 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1645 unsigned long expired;
1649 * When set to zero, disable periodic writeback
1651 if (!dirty_writeback_interval)
1654 expired = wb->last_old_flush +
1655 msecs_to_jiffies(dirty_writeback_interval * 10);
1656 if (time_before(jiffies, expired))
1659 wb->last_old_flush = jiffies;
1660 nr_pages = get_nr_dirty_pages();
1663 struct wb_writeback_work work = {
1664 .nr_pages = nr_pages,
1665 .sync_mode = WB_SYNC_NONE,
1668 .reason = WB_REASON_PERIODIC,
1671 return wb_writeback(wb, &work);
1678 * Retrieve work items and do the writeback they describe
1680 static long wb_do_writeback(struct bdi_writeback *wb)
1682 struct wb_writeback_work *work;
1685 set_bit(WB_writeback_running, &wb->state);
1686 while ((work = get_next_work_item(wb)) != NULL) {
1687 struct wb_completion *done = work->done;
1688 bool need_wake_up = false;
1690 trace_writeback_exec(wb->bdi, work);
1692 wrote += wb_writeback(wb, work);
1694 if (work->single_wait) {
1695 WARN_ON_ONCE(work->auto_free);
1696 /* paired w/ rmb in wb_wait_for_single_work() */
1698 work->single_done = 1;
1699 need_wake_up = true;
1700 } else if (work->auto_free) {
1704 if (done && atomic_dec_and_test(&done->cnt))
1705 need_wake_up = true;
1708 wake_up_all(&wb->bdi->wb_waitq);
1712 * Check for periodic writeback, kupdated() style
1714 wrote += wb_check_old_data_flush(wb);
1715 wrote += wb_check_background_flush(wb);
1716 clear_bit(WB_writeback_running, &wb->state);
1722 * Handle writeback of dirty data for the device backed by this bdi. Also
1723 * reschedules periodically and does kupdated style flushing.
1725 void wb_workfn(struct work_struct *work)
1727 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1728 struct bdi_writeback, dwork);
1731 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1732 current->flags |= PF_SWAPWRITE;
1734 if (likely(!current_is_workqueue_rescuer() ||
1735 !test_bit(WB_registered, &wb->state))) {
1737 * The normal path. Keep writing back @wb until its
1738 * work_list is empty. Note that this path is also taken
1739 * if @wb is shutting down even when we're running off the
1740 * rescuer as work_list needs to be drained.
1743 pages_written = wb_do_writeback(wb);
1744 trace_writeback_pages_written(pages_written);
1745 } while (!list_empty(&wb->work_list));
1748 * bdi_wq can't get enough workers and we're running off
1749 * the emergency worker. Don't hog it. Hopefully, 1024 is
1750 * enough for efficient IO.
1752 pages_written = writeback_inodes_wb(wb, 1024,
1753 WB_REASON_FORKER_THREAD);
1754 trace_writeback_pages_written(pages_written);
1757 if (!list_empty(&wb->work_list))
1758 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1759 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1760 wb_wakeup_delayed(wb);
1762 current->flags &= ~PF_SWAPWRITE;
1766 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1769 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1771 struct backing_dev_info *bdi;
1774 nr_pages = get_nr_dirty_pages();
1777 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1778 struct bdi_writeback *wb;
1779 struct wb_iter iter;
1781 if (!bdi_has_dirty_io(bdi))
1784 bdi_for_each_wb(wb, bdi, &iter, 0)
1785 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1792 * Wake up bdi's periodically to make sure dirtytime inodes gets
1793 * written back periodically. We deliberately do *not* check the
1794 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1795 * kernel to be constantly waking up once there are any dirtytime
1796 * inodes on the system. So instead we define a separate delayed work
1797 * function which gets called much more rarely. (By default, only
1798 * once every 12 hours.)
1800 * If there is any other write activity going on in the file system,
1801 * this function won't be necessary. But if the only thing that has
1802 * happened on the file system is a dirtytime inode caused by an atime
1803 * update, we need this infrastructure below to make sure that inode
1804 * eventually gets pushed out to disk.
1806 static void wakeup_dirtytime_writeback(struct work_struct *w);
1807 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1809 static void wakeup_dirtytime_writeback(struct work_struct *w)
1811 struct backing_dev_info *bdi;
1814 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1815 struct bdi_writeback *wb;
1816 struct wb_iter iter;
1818 bdi_for_each_wb(wb, bdi, &iter, 0)
1819 if (!list_empty(&bdi->wb.b_dirty_time))
1820 wb_wakeup(&bdi->wb);
1823 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1826 static int __init start_dirtytime_writeback(void)
1828 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1831 __initcall(start_dirtytime_writeback);
1833 int dirtytime_interval_handler(struct ctl_table *table, int write,
1834 void __user *buffer, size_t *lenp, loff_t *ppos)
1838 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1839 if (ret == 0 && write)
1840 mod_delayed_work(system_wq, &dirtytime_work, 0);
1844 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1846 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1847 struct dentry *dentry;
1848 const char *name = "?";
1850 dentry = d_find_alias(inode);
1852 spin_lock(&dentry->d_lock);
1853 name = (const char *) dentry->d_name.name;
1856 "%s(%d): dirtied inode %lu (%s) on %s\n",
1857 current->comm, task_pid_nr(current), inode->i_ino,
1858 name, inode->i_sb->s_id);
1860 spin_unlock(&dentry->d_lock);
1867 * __mark_inode_dirty - internal function
1868 * @inode: inode to mark
1869 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1870 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1871 * mark_inode_dirty_sync.
1873 * Put the inode on the super block's dirty list.
1875 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1876 * dirty list only if it is hashed or if it refers to a blockdev.
1877 * If it was not hashed, it will never be added to the dirty list
1878 * even if it is later hashed, as it will have been marked dirty already.
1880 * In short, make sure you hash any inodes _before_ you start marking
1883 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1884 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1885 * the kernel-internal blockdev inode represents the dirtying time of the
1886 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1887 * page->mapping->host, so the page-dirtying time is recorded in the internal
1890 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
1891 void __mark_inode_dirty(struct inode *inode, int flags)
1893 struct super_block *sb = inode->i_sb;
1896 trace_writeback_mark_inode_dirty(inode, flags);
1899 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1900 * dirty the inode itself
1902 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
1903 trace_writeback_dirty_inode_start(inode, flags);
1905 if (sb->s_op->dirty_inode)
1906 sb->s_op->dirty_inode(inode, flags);
1908 trace_writeback_dirty_inode(inode, flags);
1910 if (flags & I_DIRTY_INODE)
1911 flags &= ~I_DIRTY_TIME;
1912 dirtytime = flags & I_DIRTY_TIME;
1915 * Paired with smp_mb() in __writeback_single_inode() for the
1916 * following lockless i_state test. See there for details.
1920 if (((inode->i_state & flags) == flags) ||
1921 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
1924 if (unlikely(block_dump))
1925 block_dump___mark_inode_dirty(inode);
1927 spin_lock(&inode->i_lock);
1928 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
1929 goto out_unlock_inode;
1930 if ((inode->i_state & flags) != flags) {
1931 const int was_dirty = inode->i_state & I_DIRTY;
1933 inode_attach_wb(inode, NULL);
1935 if (flags & I_DIRTY_INODE)
1936 inode->i_state &= ~I_DIRTY_TIME;
1937 inode->i_state |= flags;
1940 * If the inode is being synced, just update its dirty state.
1941 * The unlocker will place the inode on the appropriate
1942 * superblock list, based upon its state.
1944 if (inode->i_state & I_SYNC)
1945 goto out_unlock_inode;
1948 * Only add valid (hashed) inodes to the superblock's
1949 * dirty list. Add blockdev inodes as well.
1951 if (!S_ISBLK(inode->i_mode)) {
1952 if (inode_unhashed(inode))
1953 goto out_unlock_inode;
1955 if (inode->i_state & I_FREEING)
1956 goto out_unlock_inode;
1959 * If the inode was already on b_dirty/b_io/b_more_io, don't
1960 * reposition it (that would break b_dirty time-ordering).
1963 struct bdi_writeback *wb;
1964 struct list_head *dirty_list;
1965 bool wakeup_bdi = false;
1967 wb = locked_inode_to_wb_and_lock_list(inode);
1969 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
1970 !test_bit(WB_registered, &wb->state),
1971 "bdi-%s not registered\n", wb->bdi->name);
1973 inode->dirtied_when = jiffies;
1975 inode->dirtied_time_when = jiffies;
1977 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
1978 dirty_list = &wb->b_dirty;
1980 dirty_list = &wb->b_dirty_time;
1982 wakeup_bdi = inode_wb_list_move_locked(inode, wb,
1985 spin_unlock(&wb->list_lock);
1986 trace_writeback_dirty_inode_enqueue(inode);
1989 * If this is the first dirty inode for this bdi,
1990 * we have to wake-up the corresponding bdi thread
1991 * to make sure background write-back happens
1994 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
1995 wb_wakeup_delayed(wb);
2000 spin_unlock(&inode->i_lock);
2003 EXPORT_SYMBOL(__mark_inode_dirty);
2005 static void wait_sb_inodes(struct super_block *sb)
2007 struct inode *inode, *old_inode = NULL;
2010 * We need to be protected against the filesystem going from
2011 * r/o to r/w or vice versa.
2013 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2015 spin_lock(&inode_sb_list_lock);
2018 * Data integrity sync. Must wait for all pages under writeback,
2019 * because there may have been pages dirtied before our sync
2020 * call, but which had writeout started before we write it out.
2021 * In which case, the inode may not be on the dirty list, but
2022 * we still have to wait for that writeout.
2024 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
2025 struct address_space *mapping = inode->i_mapping;
2027 spin_lock(&inode->i_lock);
2028 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2029 (mapping->nrpages == 0)) {
2030 spin_unlock(&inode->i_lock);
2034 spin_unlock(&inode->i_lock);
2035 spin_unlock(&inode_sb_list_lock);
2038 * We hold a reference to 'inode' so it couldn't have been
2039 * removed from s_inodes list while we dropped the
2040 * inode_sb_list_lock. We cannot iput the inode now as we can
2041 * be holding the last reference and we cannot iput it under
2042 * inode_sb_list_lock. So we keep the reference and iput it
2048 filemap_fdatawait(mapping);
2052 spin_lock(&inode_sb_list_lock);
2054 spin_unlock(&inode_sb_list_lock);
2058 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2059 enum wb_reason reason, bool skip_if_busy)
2061 DEFINE_WB_COMPLETION_ONSTACK(done);
2062 struct wb_writeback_work work = {
2064 .sync_mode = WB_SYNC_NONE,
2065 .tagged_writepages = 1,
2070 struct backing_dev_info *bdi = sb->s_bdi;
2072 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2074 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2076 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2077 wb_wait_for_completion(bdi, &done);
2081 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2082 * @sb: the superblock
2083 * @nr: the number of pages to write
2084 * @reason: reason why some writeback work initiated
2086 * Start writeback on some inodes on this super_block. No guarantees are made
2087 * on how many (if any) will be written, and this function does not wait
2088 * for IO completion of submitted IO.
2090 void writeback_inodes_sb_nr(struct super_block *sb,
2092 enum wb_reason reason)
2094 __writeback_inodes_sb_nr(sb, nr, reason, false);
2096 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2099 * writeback_inodes_sb - writeback dirty inodes from given super_block
2100 * @sb: the superblock
2101 * @reason: reason why some writeback work was initiated
2103 * Start writeback on some inodes on this super_block. No guarantees are made
2104 * on how many (if any) will be written, and this function does not wait
2105 * for IO completion of submitted IO.
2107 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2109 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2111 EXPORT_SYMBOL(writeback_inodes_sb);
2114 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2115 * @sb: the superblock
2116 * @nr: the number of pages to write
2117 * @reason: the reason of writeback
2119 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2120 * Returns 1 if writeback was started, 0 if not.
2122 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2123 enum wb_reason reason)
2125 if (!down_read_trylock(&sb->s_umount))
2128 __writeback_inodes_sb_nr(sb, nr, reason, true);
2129 up_read(&sb->s_umount);
2132 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2135 * try_to_writeback_inodes_sb - try to start writeback if none underway
2136 * @sb: the superblock
2137 * @reason: reason why some writeback work was initiated
2139 * Implement by try_to_writeback_inodes_sb_nr()
2140 * Returns 1 if writeback was started, 0 if not.
2142 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2144 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2146 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2149 * sync_inodes_sb - sync sb inode pages
2150 * @sb: the superblock
2152 * This function writes and waits on any dirty inode belonging to this
2155 void sync_inodes_sb(struct super_block *sb)
2157 DEFINE_WB_COMPLETION_ONSTACK(done);
2158 struct wb_writeback_work work = {
2160 .sync_mode = WB_SYNC_ALL,
2161 .nr_pages = LONG_MAX,
2164 .reason = WB_REASON_SYNC,
2167 struct backing_dev_info *bdi = sb->s_bdi;
2169 /* Nothing to do? */
2170 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2172 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2174 bdi_split_work_to_wbs(bdi, &work, false);
2175 wb_wait_for_completion(bdi, &done);
2179 EXPORT_SYMBOL(sync_inodes_sb);
2182 * write_inode_now - write an inode to disk
2183 * @inode: inode to write to disk
2184 * @sync: whether the write should be synchronous or not
2186 * This function commits an inode to disk immediately if it is dirty. This is
2187 * primarily needed by knfsd.
2189 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2191 int write_inode_now(struct inode *inode, int sync)
2193 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
2194 struct writeback_control wbc = {
2195 .nr_to_write = LONG_MAX,
2196 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2198 .range_end = LLONG_MAX,
2201 if (!mapping_cap_writeback_dirty(inode->i_mapping))
2202 wbc.nr_to_write = 0;
2205 return writeback_single_inode(inode, wb, &wbc);
2207 EXPORT_SYMBOL(write_inode_now);
2210 * sync_inode - write an inode and its pages to disk.
2211 * @inode: the inode to sync
2212 * @wbc: controls the writeback mode
2214 * sync_inode() will write an inode and its pages to disk. It will also
2215 * correctly update the inode on its superblock's dirty inode lists and will
2216 * update inode->i_state.
2218 * The caller must have a ref on the inode.
2220 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2222 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
2224 EXPORT_SYMBOL(sync_inode);
2227 * sync_inode_metadata - write an inode to disk
2228 * @inode: the inode to sync
2229 * @wait: wait for I/O to complete.
2231 * Write an inode to disk and adjust its dirty state after completion.
2233 * Note: only writes the actual inode, no associated data or other metadata.
2235 int sync_inode_metadata(struct inode *inode, int wait)
2237 struct writeback_control wbc = {
2238 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2239 .nr_to_write = 0, /* metadata-only */
2242 return sync_inode(inode, &wbc);
2244 EXPORT_SYMBOL(sync_inode_metadata);