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
);
312 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
313 * @wbc: writeback_control of interest
314 * @inode: target inode
316 * @inode is locked and about to be written back under the control of @wbc.
317 * Record @inode's writeback context into @wbc and unlock the i_lock. On
318 * writeback completion, wbc_detach_inode() should be called. This is used
319 * to track the cgroup writeback context.
321 void wbc_attach_and_unlock_inode(struct writeback_control
*wbc
,
324 wbc
->wb
= inode_to_wb(inode
);
327 wbc
->wb_id
= wbc
->wb
->memcg_css
->id
;
328 wbc
->wb_lcand_id
= inode
->i_wb_frn_winner
;
329 wbc
->wb_tcand_id
= 0;
331 wbc
->wb_lcand_bytes
= 0;
332 wbc
->wb_tcand_bytes
= 0;
335 spin_unlock(&inode
->i_lock
);
339 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
340 * @wbc: writeback_control of the just finished writeback
342 * To be called after a writeback attempt of an inode finishes and undoes
343 * wbc_attach_and_unlock_inode(). Can be called under any context.
345 * As concurrent write sharing of an inode is expected to be very rare and
346 * memcg only tracks page ownership on first-use basis severely confining
347 * the usefulness of such sharing, cgroup writeback tracks ownership
348 * per-inode. While the support for concurrent write sharing of an inode
349 * is deemed unnecessary, an inode being written to by different cgroups at
350 * different points in time is a lot more common, and, more importantly,
351 * charging only by first-use can too readily lead to grossly incorrect
352 * behaviors (single foreign page can lead to gigabytes of writeback to be
353 * incorrectly attributed).
355 * To resolve this issue, cgroup writeback detects the majority dirtier of
356 * an inode and transfers the ownership to it. To avoid unnnecessary
357 * oscillation, the detection mechanism keeps track of history and gives
358 * out the switch verdict only if the foreign usage pattern is stable over
359 * a certain amount of time and/or writeback attempts.
361 * On each writeback attempt, @wbc tries to detect the majority writer
362 * using Boyer-Moore majority vote algorithm. In addition to the byte
363 * count from the majority voting, it also counts the bytes written for the
364 * current wb and the last round's winner wb (max of last round's current
365 * wb, the winner from two rounds ago, and the last round's majority
366 * candidate). Keeping track of the historical winner helps the algorithm
367 * to semi-reliably detect the most active writer even when it's not the
370 * Once the winner of the round is determined, whether the winner is
371 * foreign or not and how much IO time the round consumed is recorded in
372 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
373 * over a certain threshold, the switch verdict is given.
375 void wbc_detach_inode(struct writeback_control
*wbc
)
377 struct bdi_writeback
*wb
= wbc
->wb
;
378 struct inode
*inode
= wbc
->inode
;
379 u16 history
= inode
->i_wb_frn_history
;
380 unsigned long avg_time
= inode
->i_wb_frn_avg_time
;
381 unsigned long max_bytes
, max_time
;
384 /* pick the winner of this round */
385 if (wbc
->wb_bytes
>= wbc
->wb_lcand_bytes
&&
386 wbc
->wb_bytes
>= wbc
->wb_tcand_bytes
) {
388 max_bytes
= wbc
->wb_bytes
;
389 } else if (wbc
->wb_lcand_bytes
>= wbc
->wb_tcand_bytes
) {
390 max_id
= wbc
->wb_lcand_id
;
391 max_bytes
= wbc
->wb_lcand_bytes
;
393 max_id
= wbc
->wb_tcand_id
;
394 max_bytes
= wbc
->wb_tcand_bytes
;
398 * Calculate the amount of IO time the winner consumed and fold it
399 * into the running average kept per inode. If the consumed IO
400 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
401 * deciding whether to switch or not. This is to prevent one-off
402 * small dirtiers from skewing the verdict.
404 max_time
= DIV_ROUND_UP((max_bytes
>> PAGE_SHIFT
) << WB_FRN_TIME_SHIFT
,
405 wb
->avg_write_bandwidth
);
407 avg_time
+= (max_time
>> WB_FRN_TIME_AVG_SHIFT
) -
408 (avg_time
>> WB_FRN_TIME_AVG_SHIFT
);
410 avg_time
= max_time
; /* immediate catch up on first run */
412 if (max_time
>= avg_time
/ WB_FRN_TIME_CUT_DIV
) {
416 * The switch verdict is reached if foreign wb's consume
417 * more than a certain proportion of IO time in a
418 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
419 * history mask where each bit represents one sixteenth of
420 * the period. Determine the number of slots to shift into
421 * history from @max_time.
423 slots
= min(DIV_ROUND_UP(max_time
, WB_FRN_HIST_UNIT
),
424 (unsigned long)WB_FRN_HIST_MAX_SLOTS
);
426 if (wbc
->wb_id
!= max_id
)
427 history
|= (1U << slots
) - 1;
430 * Switch if the current wb isn't the consistent winner.
431 * If there are multiple closely competing dirtiers, the
432 * inode may switch across them repeatedly over time, which
433 * is okay. The main goal is avoiding keeping an inode on
434 * the wrong wb for an extended period of time.
436 if (hweight32(history
) > WB_FRN_HIST_THR_SLOTS
) {
445 * Multiple instances of this function may race to update the
446 * following fields but we don't mind occassional inaccuracies.
448 inode
->i_wb_frn_winner
= max_id
;
449 inode
->i_wb_frn_avg_time
= min(avg_time
, (unsigned long)U16_MAX
);
450 inode
->i_wb_frn_history
= history
;
457 * wbc_account_io - account IO issued during writeback
458 * @wbc: writeback_control of the writeback in progress
459 * @page: page being written out
460 * @bytes: number of bytes being written out
462 * @bytes from @page are about to written out during the writeback
463 * controlled by @wbc. Keep the book for foreign inode detection. See
464 * wbc_detach_inode().
466 void wbc_account_io(struct writeback_control
*wbc
, struct page
*page
,
472 * pageout() path doesn't attach @wbc to the inode being written
473 * out. This is intentional as we don't want the function to block
474 * behind a slow cgroup. Ultimately, we want pageout() to kick off
475 * regular writeback instead of writing things out itself.
481 id
= mem_cgroup_css_from_page(page
)->id
;
484 if (id
== wbc
->wb_id
) {
485 wbc
->wb_bytes
+= bytes
;
489 if (id
== wbc
->wb_lcand_id
)
490 wbc
->wb_lcand_bytes
+= bytes
;
492 /* Boyer-Moore majority vote algorithm */
493 if (!wbc
->wb_tcand_bytes
)
494 wbc
->wb_tcand_id
= id
;
495 if (id
== wbc
->wb_tcand_id
)
496 wbc
->wb_tcand_bytes
+= bytes
;
498 wbc
->wb_tcand_bytes
-= min(bytes
, wbc
->wb_tcand_bytes
);
502 * inode_congested - test whether an inode is congested
503 * @inode: inode to test for congestion
504 * @cong_bits: mask of WB_[a]sync_congested bits to test
506 * Tests whether @inode is congested. @cong_bits is the mask of congestion
507 * bits to test and the return value is the mask of set bits.
509 * If cgroup writeback is enabled for @inode, the congestion state is
510 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
511 * associated with @inode is congested; otherwise, the root wb's congestion
514 int inode_congested(struct inode
*inode
, int cong_bits
)
517 struct bdi_writeback
*wb
= inode_to_wb(inode
);
519 return wb_congested(wb
, cong_bits
);
522 return wb_congested(&inode_to_bdi(inode
)->wb
, cong_bits
);
524 EXPORT_SYMBOL_GPL(inode_congested
);
527 * wb_wait_for_single_work - wait for completion of a single bdi_writeback_work
528 * @bdi: bdi the work item was issued to
529 * @work: work item to wait for
531 * Wait for the completion of @work which was issued to one of @bdi's
532 * bdi_writeback's. The caller must have set @work->single_wait before
533 * issuing it. This wait operates independently fo
534 * wb_wait_for_completion() and also disables automatic freeing of @work.
536 static void wb_wait_for_single_work(struct backing_dev_info
*bdi
,
537 struct wb_writeback_work
*work
)
539 if (WARN_ON_ONCE(!work
->single_wait
))
542 wait_event(bdi
->wb_waitq
, work
->single_done
);
545 * Paired with smp_wmb() in wb_do_writeback() and ensures that all
546 * modifications to @work prior to assertion of ->single_done is
547 * visible to the caller once this function returns.
553 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
554 * @wb: target bdi_writeback to split @nr_pages to
555 * @nr_pages: number of pages to write for the whole bdi
557 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
558 * relation to the total write bandwidth of all wb's w/ dirty inodes on
561 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
563 unsigned long this_bw
= wb
->avg_write_bandwidth
;
564 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
566 if (nr_pages
== LONG_MAX
)
570 * This may be called on clean wb's and proportional distribution
571 * may not make sense, just use the original @nr_pages in those
572 * cases. In general, we wanna err on the side of writing more.
574 if (!tot_bw
|| this_bw
>= tot_bw
)
577 return DIV_ROUND_UP_ULL((u64
)nr_pages
* this_bw
, tot_bw
);
581 * wb_clone_and_queue_work - clone a wb_writeback_work and issue it to a wb
582 * @wb: target bdi_writeback
583 * @base_work: source wb_writeback_work
585 * Try to make a clone of @base_work and issue it to @wb. If cloning
586 * succeeds, %true is returned; otherwise, @base_work is issued directly
587 * and %false is returned. In the latter case, the caller is required to
588 * wait for @base_work's completion using wb_wait_for_single_work().
590 * A clone is auto-freed on completion. @base_work never is.
592 static bool wb_clone_and_queue_work(struct bdi_writeback
*wb
,
593 struct wb_writeback_work
*base_work
)
595 struct wb_writeback_work
*work
;
597 work
= kmalloc(sizeof(*work
), GFP_ATOMIC
);
601 work
->single_wait
= 0;
605 work
->single_wait
= 1;
607 work
->single_done
= 0;
608 wb_queue_work(wb
, work
);
609 return work
!= base_work
;
613 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
614 * @bdi: target backing_dev_info
615 * @base_work: wb_writeback_work to issue
616 * @skip_if_busy: skip wb's which already have writeback in progress
618 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
619 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
620 * distributed to the busy wbs according to each wb's proportion in the
621 * total active write bandwidth of @bdi.
623 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
624 struct wb_writeback_work
*base_work
,
627 long nr_pages
= base_work
->nr_pages
;
628 int next_blkcg_id
= 0;
629 struct bdi_writeback
*wb
;
634 if (!bdi_has_dirty_io(bdi
))
638 bdi_for_each_wb(wb
, bdi
, &iter
, next_blkcg_id
) {
639 if (!wb_has_dirty_io(wb
) ||
640 (skip_if_busy
&& writeback_in_progress(wb
)))
643 base_work
->nr_pages
= wb_split_bdi_pages(wb
, nr_pages
);
644 if (!wb_clone_and_queue_work(wb
, base_work
)) {
645 next_blkcg_id
= wb
->blkcg_css
->id
+ 1;
647 wb_wait_for_single_work(bdi
, base_work
);
654 #else /* CONFIG_CGROUP_WRITEBACK */
656 static struct bdi_writeback
*
657 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
658 __releases(&inode
->i_lock
)
659 __acquires(&wb
->list_lock
)
661 struct bdi_writeback
*wb
= inode_to_wb(inode
);
663 spin_unlock(&inode
->i_lock
);
664 spin_lock(&wb
->list_lock
);
668 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
669 __acquires(&wb
->list_lock
)
671 struct bdi_writeback
*wb
= inode_to_wb(inode
);
673 spin_lock(&wb
->list_lock
);
677 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
682 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
683 struct wb_writeback_work
*base_work
,
688 if (bdi_has_dirty_io(bdi
) &&
689 (!skip_if_busy
|| !writeback_in_progress(&bdi
->wb
))) {
690 base_work
->auto_free
= 0;
691 base_work
->single_wait
= 0;
692 base_work
->single_done
= 0;
693 wb_queue_work(&bdi
->wb
, base_work
);
697 #endif /* CONFIG_CGROUP_WRITEBACK */
699 void wb_start_writeback(struct bdi_writeback
*wb
, long nr_pages
,
700 bool range_cyclic
, enum wb_reason reason
)
702 struct wb_writeback_work
*work
;
704 if (!wb_has_dirty_io(wb
))
708 * This is WB_SYNC_NONE writeback, so if allocation fails just
709 * wakeup the thread for old dirty data writeback
711 work
= kzalloc(sizeof(*work
), GFP_ATOMIC
);
713 trace_writeback_nowork(wb
->bdi
);
718 work
->sync_mode
= WB_SYNC_NONE
;
719 work
->nr_pages
= nr_pages
;
720 work
->range_cyclic
= range_cyclic
;
721 work
->reason
= reason
;
724 wb_queue_work(wb
, work
);
728 * wb_start_background_writeback - start background writeback
729 * @wb: bdi_writback to write from
732 * This makes sure WB_SYNC_NONE background writeback happens. When
733 * this function returns, it is only guaranteed that for given wb
734 * some IO is happening if we are over background dirty threshold.
735 * Caller need not hold sb s_umount semaphore.
737 void wb_start_background_writeback(struct bdi_writeback
*wb
)
740 * We just wake up the flusher thread. It will perform background
741 * writeback as soon as there is no other work to do.
743 trace_writeback_wake_background(wb
->bdi
);
748 * Remove the inode from the writeback list it is on.
750 void inode_wb_list_del(struct inode
*inode
)
752 struct bdi_writeback
*wb
;
754 wb
= inode_to_wb_and_lock_list(inode
);
755 inode_wb_list_del_locked(inode
, wb
);
756 spin_unlock(&wb
->list_lock
);
760 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
761 * furthest end of its superblock's dirty-inode list.
763 * Before stamping the inode's ->dirtied_when, we check to see whether it is
764 * already the most-recently-dirtied inode on the b_dirty list. If that is
765 * the case then the inode must have been redirtied while it was being written
766 * out and we don't reset its dirtied_when.
768 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
770 if (!list_empty(&wb
->b_dirty
)) {
773 tail
= wb_inode(wb
->b_dirty
.next
);
774 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
775 inode
->dirtied_when
= jiffies
;
777 inode_wb_list_move_locked(inode
, wb
, &wb
->b_dirty
);
781 * requeue inode for re-scanning after bdi->b_io list is exhausted.
783 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
785 inode_wb_list_move_locked(inode
, wb
, &wb
->b_more_io
);
788 static void inode_sync_complete(struct inode
*inode
)
790 inode
->i_state
&= ~I_SYNC
;
791 /* If inode is clean an unused, put it into LRU now... */
792 inode_add_lru(inode
);
793 /* Waiters must see I_SYNC cleared before being woken up */
795 wake_up_bit(&inode
->i_state
, __I_SYNC
);
798 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
800 bool ret
= time_after(inode
->dirtied_when
, t
);
803 * For inodes being constantly redirtied, dirtied_when can get stuck.
804 * It _appears_ to be in the future, but is actually in distant past.
805 * This test is necessary to prevent such wrapped-around relative times
806 * from permanently stopping the whole bdi writeback.
808 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
813 #define EXPIRE_DIRTY_ATIME 0x0001
816 * Move expired (dirtied before work->older_than_this) dirty inodes from
817 * @delaying_queue to @dispatch_queue.
819 static int move_expired_inodes(struct list_head
*delaying_queue
,
820 struct list_head
*dispatch_queue
,
822 struct wb_writeback_work
*work
)
824 unsigned long *older_than_this
= NULL
;
825 unsigned long expire_time
;
827 struct list_head
*pos
, *node
;
828 struct super_block
*sb
= NULL
;
833 if ((flags
& EXPIRE_DIRTY_ATIME
) == 0)
834 older_than_this
= work
->older_than_this
;
835 else if (!work
->for_sync
) {
836 expire_time
= jiffies
- (dirtytime_expire_interval
* HZ
);
837 older_than_this
= &expire_time
;
839 while (!list_empty(delaying_queue
)) {
840 inode
= wb_inode(delaying_queue
->prev
);
841 if (older_than_this
&&
842 inode_dirtied_after(inode
, *older_than_this
))
844 list_move(&inode
->i_wb_list
, &tmp
);
846 if (flags
& EXPIRE_DIRTY_ATIME
)
847 set_bit(__I_DIRTY_TIME_EXPIRED
, &inode
->i_state
);
848 if (sb_is_blkdev_sb(inode
->i_sb
))
850 if (sb
&& sb
!= inode
->i_sb
)
855 /* just one sb in list, splice to dispatch_queue and we're done */
857 list_splice(&tmp
, dispatch_queue
);
861 /* Move inodes from one superblock together */
862 while (!list_empty(&tmp
)) {
863 sb
= wb_inode(tmp
.prev
)->i_sb
;
864 list_for_each_prev_safe(pos
, node
, &tmp
) {
865 inode
= wb_inode(pos
);
866 if (inode
->i_sb
== sb
)
867 list_move(&inode
->i_wb_list
, dispatch_queue
);
875 * Queue all expired dirty inodes for io, eldest first.
877 * newly dirtied b_dirty b_io b_more_io
878 * =============> gf edc BA
880 * newly dirtied b_dirty b_io b_more_io
881 * =============> g fBAedc
883 * +--> dequeue for IO
885 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
)
889 assert_spin_locked(&wb
->list_lock
);
890 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
891 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, 0, work
);
892 moved
+= move_expired_inodes(&wb
->b_dirty_time
, &wb
->b_io
,
893 EXPIRE_DIRTY_ATIME
, work
);
895 wb_io_lists_populated(wb
);
896 trace_writeback_queue_io(wb
, work
, moved
);
899 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
903 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
904 trace_writeback_write_inode_start(inode
, wbc
);
905 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
906 trace_writeback_write_inode(inode
, wbc
);
913 * Wait for writeback on an inode to complete. Called with i_lock held.
914 * Caller must make sure inode cannot go away when we drop i_lock.
916 static void __inode_wait_for_writeback(struct inode
*inode
)
917 __releases(inode
->i_lock
)
918 __acquires(inode
->i_lock
)
920 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
921 wait_queue_head_t
*wqh
;
923 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
924 while (inode
->i_state
& I_SYNC
) {
925 spin_unlock(&inode
->i_lock
);
926 __wait_on_bit(wqh
, &wq
, bit_wait
,
927 TASK_UNINTERRUPTIBLE
);
928 spin_lock(&inode
->i_lock
);
933 * Wait for writeback on an inode to complete. Caller must have inode pinned.
935 void inode_wait_for_writeback(struct inode
*inode
)
937 spin_lock(&inode
->i_lock
);
938 __inode_wait_for_writeback(inode
);
939 spin_unlock(&inode
->i_lock
);
943 * Sleep until I_SYNC is cleared. This function must be called with i_lock
944 * held and drops it. It is aimed for callers not holding any inode reference
945 * so once i_lock is dropped, inode can go away.
947 static void inode_sleep_on_writeback(struct inode
*inode
)
948 __releases(inode
->i_lock
)
951 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
954 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
955 sleep
= inode
->i_state
& I_SYNC
;
956 spin_unlock(&inode
->i_lock
);
959 finish_wait(wqh
, &wait
);
963 * Find proper writeback list for the inode depending on its current state and
964 * possibly also change of its state while we were doing writeback. Here we
965 * handle things such as livelock prevention or fairness of writeback among
966 * inodes. This function can be called only by flusher thread - noone else
967 * processes all inodes in writeback lists and requeueing inodes behind flusher
968 * thread's back can have unexpected consequences.
970 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
971 struct writeback_control
*wbc
)
973 if (inode
->i_state
& I_FREEING
)
977 * Sync livelock prevention. Each inode is tagged and synced in one
978 * shot. If still dirty, it will be redirty_tail()'ed below. Update
979 * the dirty time to prevent enqueue and sync it again.
981 if ((inode
->i_state
& I_DIRTY
) &&
982 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
983 inode
->dirtied_when
= jiffies
;
985 if (wbc
->pages_skipped
) {
987 * writeback is not making progress due to locked
988 * buffers. Skip this inode for now.
990 redirty_tail(inode
, wb
);
994 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
996 * We didn't write back all the pages. nfs_writepages()
997 * sometimes bales out without doing anything.
999 if (wbc
->nr_to_write
<= 0) {
1000 /* Slice used up. Queue for next turn. */
1001 requeue_io(inode
, wb
);
1004 * Writeback blocked by something other than
1005 * congestion. Delay the inode for some time to
1006 * avoid spinning on the CPU (100% iowait)
1007 * retrying writeback of the dirty page/inode
1008 * that cannot be performed immediately.
1010 redirty_tail(inode
, wb
);
1012 } else if (inode
->i_state
& I_DIRTY
) {
1014 * Filesystems can dirty the inode during writeback operations,
1015 * such as delayed allocation during submission or metadata
1016 * updates after data IO completion.
1018 redirty_tail(inode
, wb
);
1019 } else if (inode
->i_state
& I_DIRTY_TIME
) {
1020 inode
->dirtied_when
= jiffies
;
1021 inode_wb_list_move_locked(inode
, wb
, &wb
->b_dirty_time
);
1023 /* The inode is clean. Remove from writeback lists. */
1024 inode_wb_list_del_locked(inode
, wb
);
1029 * Write out an inode and its dirty pages. Do not update the writeback list
1030 * linkage. That is left to the caller. The caller is also responsible for
1031 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1034 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1036 struct address_space
*mapping
= inode
->i_mapping
;
1037 long nr_to_write
= wbc
->nr_to_write
;
1041 WARN_ON(!(inode
->i_state
& I_SYNC
));
1043 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
1045 ret
= do_writepages(mapping
, wbc
);
1048 * Make sure to wait on the data before writing out the metadata.
1049 * This is important for filesystems that modify metadata on data
1050 * I/O completion. We don't do it for sync(2) writeback because it has a
1051 * separate, external IO completion path and ->sync_fs for guaranteeing
1052 * inode metadata is written back correctly.
1054 if (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
) {
1055 int err
= filemap_fdatawait(mapping
);
1061 * Some filesystems may redirty the inode during the writeback
1062 * due to delalloc, clear dirty metadata flags right before
1065 spin_lock(&inode
->i_lock
);
1067 dirty
= inode
->i_state
& I_DIRTY
;
1068 if (inode
->i_state
& I_DIRTY_TIME
) {
1069 if ((dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) ||
1070 unlikely(inode
->i_state
& I_DIRTY_TIME_EXPIRED
) ||
1071 unlikely(time_after(jiffies
,
1072 (inode
->dirtied_time_when
+
1073 dirtytime_expire_interval
* HZ
)))) {
1074 dirty
|= I_DIRTY_TIME
| I_DIRTY_TIME_EXPIRED
;
1075 trace_writeback_lazytime(inode
);
1078 inode
->i_state
&= ~I_DIRTY_TIME_EXPIRED
;
1079 inode
->i_state
&= ~dirty
;
1082 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1083 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1084 * either they see the I_DIRTY bits cleared or we see the dirtied
1087 * I_DIRTY_PAGES is always cleared together above even if @mapping
1088 * still has dirty pages. The flag is reinstated after smp_mb() if
1089 * necessary. This guarantees that either __mark_inode_dirty()
1090 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1094 if (mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
1095 inode
->i_state
|= I_DIRTY_PAGES
;
1097 spin_unlock(&inode
->i_lock
);
1099 if (dirty
& I_DIRTY_TIME
)
1100 mark_inode_dirty_sync(inode
);
1101 /* Don't write the inode if only I_DIRTY_PAGES was set */
1102 if (dirty
& ~I_DIRTY_PAGES
) {
1103 int err
= write_inode(inode
, wbc
);
1107 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
1112 * Write out an inode's dirty pages. Either the caller has an active reference
1113 * on the inode or the inode has I_WILL_FREE set.
1115 * This function is designed to be called for writing back one inode which
1116 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1117 * and does more profound writeback list handling in writeback_sb_inodes().
1120 writeback_single_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
1121 struct writeback_control
*wbc
)
1125 spin_lock(&inode
->i_lock
);
1126 if (!atomic_read(&inode
->i_count
))
1127 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
1129 WARN_ON(inode
->i_state
& I_WILL_FREE
);
1131 if (inode
->i_state
& I_SYNC
) {
1132 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
1135 * It's a data-integrity sync. We must wait. Since callers hold
1136 * inode reference or inode has I_WILL_FREE set, it cannot go
1139 __inode_wait_for_writeback(inode
);
1141 WARN_ON(inode
->i_state
& I_SYNC
);
1143 * Skip inode if it is clean and we have no outstanding writeback in
1144 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1145 * function since flusher thread may be doing for example sync in
1146 * parallel and if we move the inode, it could get skipped. So here we
1147 * make sure inode is on some writeback list and leave it there unless
1148 * we have completely cleaned the inode.
1150 if (!(inode
->i_state
& I_DIRTY_ALL
) &&
1151 (wbc
->sync_mode
!= WB_SYNC_ALL
||
1152 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
1154 inode
->i_state
|= I_SYNC
;
1155 wbc_attach_and_unlock_inode(wbc
, inode
);
1157 ret
= __writeback_single_inode(inode
, wbc
);
1159 wbc_detach_inode(wbc
);
1160 spin_lock(&wb
->list_lock
);
1161 spin_lock(&inode
->i_lock
);
1163 * If inode is clean, remove it from writeback lists. Otherwise don't
1164 * touch it. See comment above for explanation.
1166 if (!(inode
->i_state
& I_DIRTY_ALL
))
1167 inode_wb_list_del_locked(inode
, wb
);
1168 spin_unlock(&wb
->list_lock
);
1169 inode_sync_complete(inode
);
1171 spin_unlock(&inode
->i_lock
);
1175 static long writeback_chunk_size(struct bdi_writeback
*wb
,
1176 struct wb_writeback_work
*work
)
1181 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1182 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1183 * here avoids calling into writeback_inodes_wb() more than once.
1185 * The intended call sequence for WB_SYNC_ALL writeback is:
1188 * writeback_sb_inodes() <== called only once
1189 * write_cache_pages() <== called once for each inode
1190 * (quickly) tag currently dirty pages
1191 * (maybe slowly) sync all tagged pages
1193 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
1196 pages
= min(wb
->avg_write_bandwidth
/ 2,
1197 global_wb_domain
.dirty_limit
/ DIRTY_SCOPE
);
1198 pages
= min(pages
, work
->nr_pages
);
1199 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
1200 MIN_WRITEBACK_PAGES
);
1207 * Write a portion of b_io inodes which belong to @sb.
1209 * Return the number of pages and/or inodes written.
1211 static long writeback_sb_inodes(struct super_block
*sb
,
1212 struct bdi_writeback
*wb
,
1213 struct wb_writeback_work
*work
)
1215 struct writeback_control wbc
= {
1216 .sync_mode
= work
->sync_mode
,
1217 .tagged_writepages
= work
->tagged_writepages
,
1218 .for_kupdate
= work
->for_kupdate
,
1219 .for_background
= work
->for_background
,
1220 .for_sync
= work
->for_sync
,
1221 .range_cyclic
= work
->range_cyclic
,
1223 .range_end
= LLONG_MAX
,
1225 unsigned long start_time
= jiffies
;
1227 long wrote
= 0; /* count both pages and inodes */
1229 while (!list_empty(&wb
->b_io
)) {
1230 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1232 if (inode
->i_sb
!= sb
) {
1235 * We only want to write back data for this
1236 * superblock, move all inodes not belonging
1237 * to it back onto the dirty list.
1239 redirty_tail(inode
, wb
);
1244 * The inode belongs to a different superblock.
1245 * Bounce back to the caller to unpin this and
1246 * pin the next superblock.
1252 * Don't bother with new inodes or inodes being freed, first
1253 * kind does not need periodic writeout yet, and for the latter
1254 * kind writeout is handled by the freer.
1256 spin_lock(&inode
->i_lock
);
1257 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
1258 spin_unlock(&inode
->i_lock
);
1259 redirty_tail(inode
, wb
);
1262 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
1264 * If this inode is locked for writeback and we are not
1265 * doing writeback-for-data-integrity, move it to
1266 * b_more_io so that writeback can proceed with the
1267 * other inodes on s_io.
1269 * We'll have another go at writing back this inode
1270 * when we completed a full scan of b_io.
1272 spin_unlock(&inode
->i_lock
);
1273 requeue_io(inode
, wb
);
1274 trace_writeback_sb_inodes_requeue(inode
);
1277 spin_unlock(&wb
->list_lock
);
1280 * We already requeued the inode if it had I_SYNC set and we
1281 * are doing WB_SYNC_NONE writeback. So this catches only the
1284 if (inode
->i_state
& I_SYNC
) {
1285 /* Wait for I_SYNC. This function drops i_lock... */
1286 inode_sleep_on_writeback(inode
);
1287 /* Inode may be gone, start again */
1288 spin_lock(&wb
->list_lock
);
1291 inode
->i_state
|= I_SYNC
;
1292 wbc_attach_and_unlock_inode(&wbc
, inode
);
1294 write_chunk
= writeback_chunk_size(wb
, work
);
1295 wbc
.nr_to_write
= write_chunk
;
1296 wbc
.pages_skipped
= 0;
1299 * We use I_SYNC to pin the inode in memory. While it is set
1300 * evict_inode() will wait so the inode cannot be freed.
1302 __writeback_single_inode(inode
, &wbc
);
1304 wbc_detach_inode(&wbc
);
1305 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
1306 wrote
+= write_chunk
- wbc
.nr_to_write
;
1307 spin_lock(&wb
->list_lock
);
1308 spin_lock(&inode
->i_lock
);
1309 if (!(inode
->i_state
& I_DIRTY_ALL
))
1311 requeue_inode(inode
, wb
, &wbc
);
1312 inode_sync_complete(inode
);
1313 spin_unlock(&inode
->i_lock
);
1314 cond_resched_lock(&wb
->list_lock
);
1316 * bail out to wb_writeback() often enough to check
1317 * background threshold and other termination conditions.
1320 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1322 if (work
->nr_pages
<= 0)
1329 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
1330 struct wb_writeback_work
*work
)
1332 unsigned long start_time
= jiffies
;
1335 while (!list_empty(&wb
->b_io
)) {
1336 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1337 struct super_block
*sb
= inode
->i_sb
;
1339 if (!trylock_super(sb
)) {
1341 * trylock_super() may fail consistently due to
1342 * s_umount being grabbed by someone else. Don't use
1343 * requeue_io() to avoid busy retrying the inode/sb.
1345 redirty_tail(inode
, wb
);
1348 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
1349 up_read(&sb
->s_umount
);
1351 /* refer to the same tests at the end of writeback_sb_inodes */
1353 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1355 if (work
->nr_pages
<= 0)
1359 /* Leave any unwritten inodes on b_io */
1363 static long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
1364 enum wb_reason reason
)
1366 struct wb_writeback_work work
= {
1367 .nr_pages
= nr_pages
,
1368 .sync_mode
= WB_SYNC_NONE
,
1373 spin_lock(&wb
->list_lock
);
1374 if (list_empty(&wb
->b_io
))
1375 queue_io(wb
, &work
);
1376 __writeback_inodes_wb(wb
, &work
);
1377 spin_unlock(&wb
->list_lock
);
1379 return nr_pages
- work
.nr_pages
;
1383 * Explicit flushing or periodic writeback of "old" data.
1385 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1386 * dirtying-time in the inode's address_space. So this periodic writeback code
1387 * just walks the superblock inode list, writing back any inodes which are
1388 * older than a specific point in time.
1390 * Try to run once per dirty_writeback_interval. But if a writeback event
1391 * takes longer than a dirty_writeback_interval interval, then leave a
1394 * older_than_this takes precedence over nr_to_write. So we'll only write back
1395 * all dirty pages if they are all attached to "old" mappings.
1397 static long wb_writeback(struct bdi_writeback
*wb
,
1398 struct wb_writeback_work
*work
)
1400 unsigned long wb_start
= jiffies
;
1401 long nr_pages
= work
->nr_pages
;
1402 unsigned long oldest_jif
;
1403 struct inode
*inode
;
1406 oldest_jif
= jiffies
;
1407 work
->older_than_this
= &oldest_jif
;
1409 spin_lock(&wb
->list_lock
);
1412 * Stop writeback when nr_pages has been consumed
1414 if (work
->nr_pages
<= 0)
1418 * Background writeout and kupdate-style writeback may
1419 * run forever. Stop them if there is other work to do
1420 * so that e.g. sync can proceed. They'll be restarted
1421 * after the other works are all done.
1423 if ((work
->for_background
|| work
->for_kupdate
) &&
1424 !list_empty(&wb
->work_list
))
1428 * For background writeout, stop when we are below the
1429 * background dirty threshold
1431 if (work
->for_background
&& !wb_over_bg_thresh(wb
))
1435 * Kupdate and background works are special and we want to
1436 * include all inodes that need writing. Livelock avoidance is
1437 * handled by these works yielding to any other work so we are
1440 if (work
->for_kupdate
) {
1441 oldest_jif
= jiffies
-
1442 msecs_to_jiffies(dirty_expire_interval
* 10);
1443 } else if (work
->for_background
)
1444 oldest_jif
= jiffies
;
1446 trace_writeback_start(wb
->bdi
, work
);
1447 if (list_empty(&wb
->b_io
))
1450 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
1452 progress
= __writeback_inodes_wb(wb
, work
);
1453 trace_writeback_written(wb
->bdi
, work
);
1455 wb_update_bandwidth(wb
, wb_start
);
1458 * Did we write something? Try for more
1460 * Dirty inodes are moved to b_io for writeback in batches.
1461 * The completion of the current batch does not necessarily
1462 * mean the overall work is done. So we keep looping as long
1463 * as made some progress on cleaning pages or inodes.
1468 * No more inodes for IO, bail
1470 if (list_empty(&wb
->b_more_io
))
1473 * Nothing written. Wait for some inode to
1474 * become available for writeback. Otherwise
1475 * we'll just busyloop.
1477 if (!list_empty(&wb
->b_more_io
)) {
1478 trace_writeback_wait(wb
->bdi
, work
);
1479 inode
= wb_inode(wb
->b_more_io
.prev
);
1480 spin_lock(&inode
->i_lock
);
1481 spin_unlock(&wb
->list_lock
);
1482 /* This function drops i_lock... */
1483 inode_sleep_on_writeback(inode
);
1484 spin_lock(&wb
->list_lock
);
1487 spin_unlock(&wb
->list_lock
);
1489 return nr_pages
- work
->nr_pages
;
1493 * Return the next wb_writeback_work struct that hasn't been processed yet.
1495 static struct wb_writeback_work
*get_next_work_item(struct bdi_writeback
*wb
)
1497 struct wb_writeback_work
*work
= NULL
;
1499 spin_lock_bh(&wb
->work_lock
);
1500 if (!list_empty(&wb
->work_list
)) {
1501 work
= list_entry(wb
->work_list
.next
,
1502 struct wb_writeback_work
, list
);
1503 list_del_init(&work
->list
);
1505 spin_unlock_bh(&wb
->work_lock
);
1510 * Add in the number of potentially dirty inodes, because each inode
1511 * write can dirty pagecache in the underlying blockdev.
1513 static unsigned long get_nr_dirty_pages(void)
1515 return global_page_state(NR_FILE_DIRTY
) +
1516 global_page_state(NR_UNSTABLE_NFS
) +
1517 get_nr_dirty_inodes();
1520 static long wb_check_background_flush(struct bdi_writeback
*wb
)
1522 if (wb_over_bg_thresh(wb
)) {
1524 struct wb_writeback_work work
= {
1525 .nr_pages
= LONG_MAX
,
1526 .sync_mode
= WB_SYNC_NONE
,
1527 .for_background
= 1,
1529 .reason
= WB_REASON_BACKGROUND
,
1532 return wb_writeback(wb
, &work
);
1538 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
1540 unsigned long expired
;
1544 * When set to zero, disable periodic writeback
1546 if (!dirty_writeback_interval
)
1549 expired
= wb
->last_old_flush
+
1550 msecs_to_jiffies(dirty_writeback_interval
* 10);
1551 if (time_before(jiffies
, expired
))
1554 wb
->last_old_flush
= jiffies
;
1555 nr_pages
= get_nr_dirty_pages();
1558 struct wb_writeback_work work
= {
1559 .nr_pages
= nr_pages
,
1560 .sync_mode
= WB_SYNC_NONE
,
1563 .reason
= WB_REASON_PERIODIC
,
1566 return wb_writeback(wb
, &work
);
1573 * Retrieve work items and do the writeback they describe
1575 static long wb_do_writeback(struct bdi_writeback
*wb
)
1577 struct wb_writeback_work
*work
;
1580 set_bit(WB_writeback_running
, &wb
->state
);
1581 while ((work
= get_next_work_item(wb
)) != NULL
) {
1582 struct wb_completion
*done
= work
->done
;
1583 bool need_wake_up
= false;
1585 trace_writeback_exec(wb
->bdi
, work
);
1587 wrote
+= wb_writeback(wb
, work
);
1589 if (work
->single_wait
) {
1590 WARN_ON_ONCE(work
->auto_free
);
1591 /* paired w/ rmb in wb_wait_for_single_work() */
1593 work
->single_done
= 1;
1594 need_wake_up
= true;
1595 } else if (work
->auto_free
) {
1599 if (done
&& atomic_dec_and_test(&done
->cnt
))
1600 need_wake_up
= true;
1603 wake_up_all(&wb
->bdi
->wb_waitq
);
1607 * Check for periodic writeback, kupdated() style
1609 wrote
+= wb_check_old_data_flush(wb
);
1610 wrote
+= wb_check_background_flush(wb
);
1611 clear_bit(WB_writeback_running
, &wb
->state
);
1617 * Handle writeback of dirty data for the device backed by this bdi. Also
1618 * reschedules periodically and does kupdated style flushing.
1620 void wb_workfn(struct work_struct
*work
)
1622 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
1623 struct bdi_writeback
, dwork
);
1626 set_worker_desc("flush-%s", dev_name(wb
->bdi
->dev
));
1627 current
->flags
|= PF_SWAPWRITE
;
1629 if (likely(!current_is_workqueue_rescuer() ||
1630 !test_bit(WB_registered
, &wb
->state
))) {
1632 * The normal path. Keep writing back @wb until its
1633 * work_list is empty. Note that this path is also taken
1634 * if @wb is shutting down even when we're running off the
1635 * rescuer as work_list needs to be drained.
1638 pages_written
= wb_do_writeback(wb
);
1639 trace_writeback_pages_written(pages_written
);
1640 } while (!list_empty(&wb
->work_list
));
1643 * bdi_wq can't get enough workers and we're running off
1644 * the emergency worker. Don't hog it. Hopefully, 1024 is
1645 * enough for efficient IO.
1647 pages_written
= writeback_inodes_wb(wb
, 1024,
1648 WB_REASON_FORKER_THREAD
);
1649 trace_writeback_pages_written(pages_written
);
1652 if (!list_empty(&wb
->work_list
))
1653 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
1654 else if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
1655 wb_wakeup_delayed(wb
);
1657 current
->flags
&= ~PF_SWAPWRITE
;
1661 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1664 void wakeup_flusher_threads(long nr_pages
, enum wb_reason reason
)
1666 struct backing_dev_info
*bdi
;
1669 nr_pages
= get_nr_dirty_pages();
1672 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
1673 struct bdi_writeback
*wb
;
1674 struct wb_iter iter
;
1676 if (!bdi_has_dirty_io(bdi
))
1679 bdi_for_each_wb(wb
, bdi
, &iter
, 0)
1680 wb_start_writeback(wb
, wb_split_bdi_pages(wb
, nr_pages
),
1687 * Wake up bdi's periodically to make sure dirtytime inodes gets
1688 * written back periodically. We deliberately do *not* check the
1689 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1690 * kernel to be constantly waking up once there are any dirtytime
1691 * inodes on the system. So instead we define a separate delayed work
1692 * function which gets called much more rarely. (By default, only
1693 * once every 12 hours.)
1695 * If there is any other write activity going on in the file system,
1696 * this function won't be necessary. But if the only thing that has
1697 * happened on the file system is a dirtytime inode caused by an atime
1698 * update, we need this infrastructure below to make sure that inode
1699 * eventually gets pushed out to disk.
1701 static void wakeup_dirtytime_writeback(struct work_struct
*w
);
1702 static DECLARE_DELAYED_WORK(dirtytime_work
, wakeup_dirtytime_writeback
);
1704 static void wakeup_dirtytime_writeback(struct work_struct
*w
)
1706 struct backing_dev_info
*bdi
;
1709 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
1710 struct bdi_writeback
*wb
;
1711 struct wb_iter iter
;
1713 bdi_for_each_wb(wb
, bdi
, &iter
, 0)
1714 if (!list_empty(&bdi
->wb
.b_dirty_time
))
1715 wb_wakeup(&bdi
->wb
);
1718 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
1721 static int __init
start_dirtytime_writeback(void)
1723 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
1726 __initcall(start_dirtytime_writeback
);
1728 int dirtytime_interval_handler(struct ctl_table
*table
, int write
,
1729 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1733 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
1734 if (ret
== 0 && write
)
1735 mod_delayed_work(system_wq
, &dirtytime_work
, 0);
1739 static noinline
void block_dump___mark_inode_dirty(struct inode
*inode
)
1741 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev")) {
1742 struct dentry
*dentry
;
1743 const char *name
= "?";
1745 dentry
= d_find_alias(inode
);
1747 spin_lock(&dentry
->d_lock
);
1748 name
= (const char *) dentry
->d_name
.name
;
1751 "%s(%d): dirtied inode %lu (%s) on %s\n",
1752 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
1753 name
, inode
->i_sb
->s_id
);
1755 spin_unlock(&dentry
->d_lock
);
1762 * __mark_inode_dirty - internal function
1763 * @inode: inode to mark
1764 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1765 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1766 * mark_inode_dirty_sync.
1768 * Put the inode on the super block's dirty list.
1770 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1771 * dirty list only if it is hashed or if it refers to a blockdev.
1772 * If it was not hashed, it will never be added to the dirty list
1773 * even if it is later hashed, as it will have been marked dirty already.
1775 * In short, make sure you hash any inodes _before_ you start marking
1778 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1779 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1780 * the kernel-internal blockdev inode represents the dirtying time of the
1781 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1782 * page->mapping->host, so the page-dirtying time is recorded in the internal
1785 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
1786 void __mark_inode_dirty(struct inode
*inode
, int flags
)
1788 struct super_block
*sb
= inode
->i_sb
;
1791 trace_writeback_mark_inode_dirty(inode
, flags
);
1794 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1795 * dirty the inode itself
1797 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
| I_DIRTY_TIME
)) {
1798 trace_writeback_dirty_inode_start(inode
, flags
);
1800 if (sb
->s_op
->dirty_inode
)
1801 sb
->s_op
->dirty_inode(inode
, flags
);
1803 trace_writeback_dirty_inode(inode
, flags
);
1805 if (flags
& I_DIRTY_INODE
)
1806 flags
&= ~I_DIRTY_TIME
;
1807 dirtytime
= flags
& I_DIRTY_TIME
;
1810 * Paired with smp_mb() in __writeback_single_inode() for the
1811 * following lockless i_state test. See there for details.
1815 if (((inode
->i_state
& flags
) == flags
) ||
1816 (dirtytime
&& (inode
->i_state
& I_DIRTY_INODE
)))
1819 if (unlikely(block_dump
))
1820 block_dump___mark_inode_dirty(inode
);
1822 spin_lock(&inode
->i_lock
);
1823 if (dirtytime
&& (inode
->i_state
& I_DIRTY_INODE
))
1824 goto out_unlock_inode
;
1825 if ((inode
->i_state
& flags
) != flags
) {
1826 const int was_dirty
= inode
->i_state
& I_DIRTY
;
1828 inode_attach_wb(inode
, NULL
);
1830 if (flags
& I_DIRTY_INODE
)
1831 inode
->i_state
&= ~I_DIRTY_TIME
;
1832 inode
->i_state
|= flags
;
1835 * If the inode is being synced, just update its dirty state.
1836 * The unlocker will place the inode on the appropriate
1837 * superblock list, based upon its state.
1839 if (inode
->i_state
& I_SYNC
)
1840 goto out_unlock_inode
;
1843 * Only add valid (hashed) inodes to the superblock's
1844 * dirty list. Add blockdev inodes as well.
1846 if (!S_ISBLK(inode
->i_mode
)) {
1847 if (inode_unhashed(inode
))
1848 goto out_unlock_inode
;
1850 if (inode
->i_state
& I_FREEING
)
1851 goto out_unlock_inode
;
1854 * If the inode was already on b_dirty/b_io/b_more_io, don't
1855 * reposition it (that would break b_dirty time-ordering).
1858 struct bdi_writeback
*wb
;
1859 struct list_head
*dirty_list
;
1860 bool wakeup_bdi
= false;
1862 wb
= locked_inode_to_wb_and_lock_list(inode
);
1864 WARN(bdi_cap_writeback_dirty(wb
->bdi
) &&
1865 !test_bit(WB_registered
, &wb
->state
),
1866 "bdi-%s not registered\n", wb
->bdi
->name
);
1868 inode
->dirtied_when
= jiffies
;
1870 inode
->dirtied_time_when
= jiffies
;
1872 if (inode
->i_state
& (I_DIRTY_INODE
| I_DIRTY_PAGES
))
1873 dirty_list
= &wb
->b_dirty
;
1875 dirty_list
= &wb
->b_dirty_time
;
1877 wakeup_bdi
= inode_wb_list_move_locked(inode
, wb
,
1880 spin_unlock(&wb
->list_lock
);
1881 trace_writeback_dirty_inode_enqueue(inode
);
1884 * If this is the first dirty inode for this bdi,
1885 * we have to wake-up the corresponding bdi thread
1886 * to make sure background write-back happens
1889 if (bdi_cap_writeback_dirty(wb
->bdi
) && wakeup_bdi
)
1890 wb_wakeup_delayed(wb
);
1895 spin_unlock(&inode
->i_lock
);
1898 EXPORT_SYMBOL(__mark_inode_dirty
);
1900 static void wait_sb_inodes(struct super_block
*sb
)
1902 struct inode
*inode
, *old_inode
= NULL
;
1905 * We need to be protected against the filesystem going from
1906 * r/o to r/w or vice versa.
1908 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1910 spin_lock(&inode_sb_list_lock
);
1913 * Data integrity sync. Must wait for all pages under writeback,
1914 * because there may have been pages dirtied before our sync
1915 * call, but which had writeout started before we write it out.
1916 * In which case, the inode may not be on the dirty list, but
1917 * we still have to wait for that writeout.
1919 list_for_each_entry(inode
, &sb
->s_inodes
, i_sb_list
) {
1920 struct address_space
*mapping
= inode
->i_mapping
;
1922 spin_lock(&inode
->i_lock
);
1923 if ((inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) ||
1924 (mapping
->nrpages
== 0)) {
1925 spin_unlock(&inode
->i_lock
);
1929 spin_unlock(&inode
->i_lock
);
1930 spin_unlock(&inode_sb_list_lock
);
1933 * We hold a reference to 'inode' so it couldn't have been
1934 * removed from s_inodes list while we dropped the
1935 * inode_sb_list_lock. We cannot iput the inode now as we can
1936 * be holding the last reference and we cannot iput it under
1937 * inode_sb_list_lock. So we keep the reference and iput it
1943 filemap_fdatawait(mapping
);
1947 spin_lock(&inode_sb_list_lock
);
1949 spin_unlock(&inode_sb_list_lock
);
1953 static void __writeback_inodes_sb_nr(struct super_block
*sb
, unsigned long nr
,
1954 enum wb_reason reason
, bool skip_if_busy
)
1956 DEFINE_WB_COMPLETION_ONSTACK(done
);
1957 struct wb_writeback_work work
= {
1959 .sync_mode
= WB_SYNC_NONE
,
1960 .tagged_writepages
= 1,
1965 struct backing_dev_info
*bdi
= sb
->s_bdi
;
1967 if (!bdi_has_dirty_io(bdi
) || bdi
== &noop_backing_dev_info
)
1969 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
1971 bdi_split_work_to_wbs(sb
->s_bdi
, &work
, skip_if_busy
);
1972 wb_wait_for_completion(bdi
, &done
);
1976 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
1977 * @sb: the superblock
1978 * @nr: the number of pages to write
1979 * @reason: reason why some writeback work initiated
1981 * Start writeback on some inodes on this super_block. No guarantees are made
1982 * on how many (if any) will be written, and this function does not wait
1983 * for IO completion of submitted IO.
1985 void writeback_inodes_sb_nr(struct super_block
*sb
,
1987 enum wb_reason reason
)
1989 __writeback_inodes_sb_nr(sb
, nr
, reason
, false);
1991 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
1994 * writeback_inodes_sb - writeback dirty inodes from given super_block
1995 * @sb: the superblock
1996 * @reason: reason why some writeback work was initiated
1998 * Start writeback on some inodes on this super_block. No guarantees are made
1999 * on how many (if any) will be written, and this function does not wait
2000 * for IO completion of submitted IO.
2002 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2004 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
2006 EXPORT_SYMBOL(writeback_inodes_sb
);
2009 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2010 * @sb: the superblock
2011 * @nr: the number of pages to write
2012 * @reason: the reason of writeback
2014 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2015 * Returns 1 if writeback was started, 0 if not.
2017 bool try_to_writeback_inodes_sb_nr(struct super_block
*sb
, unsigned long nr
,
2018 enum wb_reason reason
)
2020 if (!down_read_trylock(&sb
->s_umount
))
2023 __writeback_inodes_sb_nr(sb
, nr
, reason
, true);
2024 up_read(&sb
->s_umount
);
2027 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr
);
2030 * try_to_writeback_inodes_sb - try to start writeback if none underway
2031 * @sb: the superblock
2032 * @reason: reason why some writeback work was initiated
2034 * Implement by try_to_writeback_inodes_sb_nr()
2035 * Returns 1 if writeback was started, 0 if not.
2037 bool try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2039 return try_to_writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
2041 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
2044 * sync_inodes_sb - sync sb inode pages
2045 * @sb: the superblock
2047 * This function writes and waits on any dirty inode belonging to this
2050 void sync_inodes_sb(struct super_block
*sb
)
2052 DEFINE_WB_COMPLETION_ONSTACK(done
);
2053 struct wb_writeback_work work
= {
2055 .sync_mode
= WB_SYNC_ALL
,
2056 .nr_pages
= LONG_MAX
,
2059 .reason
= WB_REASON_SYNC
,
2062 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2064 /* Nothing to do? */
2065 if (!bdi_has_dirty_io(bdi
) || bdi
== &noop_backing_dev_info
)
2067 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2069 bdi_split_work_to_wbs(bdi
, &work
, false);
2070 wb_wait_for_completion(bdi
, &done
);
2074 EXPORT_SYMBOL(sync_inodes_sb
);
2077 * write_inode_now - write an inode to disk
2078 * @inode: inode to write to disk
2079 * @sync: whether the write should be synchronous or not
2081 * This function commits an inode to disk immediately if it is dirty. This is
2082 * primarily needed by knfsd.
2084 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2086 int write_inode_now(struct inode
*inode
, int sync
)
2088 struct bdi_writeback
*wb
= &inode_to_bdi(inode
)->wb
;
2089 struct writeback_control wbc
= {
2090 .nr_to_write
= LONG_MAX
,
2091 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2093 .range_end
= LLONG_MAX
,
2096 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
2097 wbc
.nr_to_write
= 0;
2100 return writeback_single_inode(inode
, wb
, &wbc
);
2102 EXPORT_SYMBOL(write_inode_now
);
2105 * sync_inode - write an inode and its pages to disk.
2106 * @inode: the inode to sync
2107 * @wbc: controls the writeback mode
2109 * sync_inode() will write an inode and its pages to disk. It will also
2110 * correctly update the inode on its superblock's dirty inode lists and will
2111 * update inode->i_state.
2113 * The caller must have a ref on the inode.
2115 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2117 return writeback_single_inode(inode
, &inode_to_bdi(inode
)->wb
, wbc
);
2119 EXPORT_SYMBOL(sync_inode
);
2122 * sync_inode_metadata - write an inode to disk
2123 * @inode: the inode to sync
2124 * @wait: wait for I/O to complete.
2126 * Write an inode to disk and adjust its dirty state after completion.
2128 * Note: only writes the actual inode, no associated data or other metadata.
2130 int sync_inode_metadata(struct inode
*inode
, int wait
)
2132 struct writeback_control wbc
= {
2133 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2134 .nr_to_write
= 0, /* metadata-only */
2137 return sync_inode(inode
, &wbc
);
2139 EXPORT_SYMBOL(sync_inode_metadata
);