4 * Copyright (C) 2002, Linus Torvalds.
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
11 * 10Apr2002 Andrew Morton
12 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/spinlock.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
23 #include <linux/pagemap.h>
24 #include <linux/kthread.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/tracepoint.h>
29 #include <linux/device.h>
30 #include <linux/memcontrol.h>
34 * 4MB minimal write chunk size
36 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
38 struct wb_completion
{
43 * Passed into wb_writeback(), essentially a subset of writeback_control
45 struct wb_writeback_work
{
47 struct super_block
*sb
;
48 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 enum wb_reason reason
; /* why was writeback initiated? */
58 struct list_head list
; /* pending work list */
59 struct wb_completion
*done
; /* set if the caller waits */
63 * If one wants to wait for one or more wb_writeback_works, each work's
64 * ->done should be set to a wb_completion defined using the following
65 * macro. Once all work items are issued with wb_queue_work(), the caller
66 * can wait for the completion of all using wb_wait_for_completion(). Work
67 * items which are waited upon aren't freed automatically on completion.
69 #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
70 struct wb_completion cmpl = { \
71 .cnt = ATOMIC_INIT(1), \
76 * If an inode is constantly having its pages dirtied, but then the
77 * updates stop dirtytime_expire_interval seconds in the past, it's
78 * possible for the worst case time between when an inode has its
79 * timestamps updated and when they finally get written out to be two
80 * dirtytime_expire_intervals. We set the default to 12 hours (in
81 * seconds), which means most of the time inodes will have their
82 * timestamps written to disk after 12 hours, but in the worst case a
83 * few inodes might not their timestamps updated for 24 hours.
85 unsigned int dirtytime_expire_interval
= 12 * 60 * 60;
87 static inline struct inode
*wb_inode(struct list_head
*head
)
89 return list_entry(head
, struct inode
, i_io_list
);
93 * Include the creation of the trace points after defining the
94 * wb_writeback_work structure and inline functions so that the definition
95 * remains local to this file.
97 #define CREATE_TRACE_POINTS
98 #include <trace/events/writeback.h>
100 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage
);
102 static bool wb_io_lists_populated(struct bdi_writeback
*wb
)
104 if (wb_has_dirty_io(wb
)) {
107 set_bit(WB_has_dirty_io
, &wb
->state
);
108 WARN_ON_ONCE(!wb
->avg_write_bandwidth
);
109 atomic_long_add(wb
->avg_write_bandwidth
,
110 &wb
->bdi
->tot_write_bandwidth
);
115 static void wb_io_lists_depopulated(struct bdi_writeback
*wb
)
117 if (wb_has_dirty_io(wb
) && list_empty(&wb
->b_dirty
) &&
118 list_empty(&wb
->b_io
) && list_empty(&wb
->b_more_io
)) {
119 clear_bit(WB_has_dirty_io
, &wb
->state
);
120 WARN_ON_ONCE(atomic_long_sub_return(wb
->avg_write_bandwidth
,
121 &wb
->bdi
->tot_write_bandwidth
) < 0);
126 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
127 * @inode: inode to be moved
128 * @wb: target bdi_writeback
129 * @head: one of @wb->b_{dirty|io|more_io}
131 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
132 * Returns %true if @inode is the first occupant of the !dirty_time IO
133 * lists; otherwise, %false.
135 static bool inode_io_list_move_locked(struct inode
*inode
,
136 struct bdi_writeback
*wb
,
137 struct list_head
*head
)
139 assert_spin_locked(&wb
->list_lock
);
141 list_move(&inode
->i_io_list
, head
);
143 /* dirty_time doesn't count as dirty_io until expiration */
144 if (head
!= &wb
->b_dirty_time
)
145 return wb_io_lists_populated(wb
);
147 wb_io_lists_depopulated(wb
);
152 * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
153 * @inode: inode to be removed
154 * @wb: bdi_writeback @inode is being removed from
156 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
157 * clear %WB_has_dirty_io if all are empty afterwards.
159 static void inode_io_list_del_locked(struct inode
*inode
,
160 struct bdi_writeback
*wb
)
162 assert_spin_locked(&wb
->list_lock
);
164 list_del_init(&inode
->i_io_list
);
165 wb_io_lists_depopulated(wb
);
168 static void wb_wakeup(struct bdi_writeback
*wb
)
170 spin_lock_bh(&wb
->work_lock
);
171 if (test_bit(WB_registered
, &wb
->state
))
172 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
173 spin_unlock_bh(&wb
->work_lock
);
176 static void finish_writeback_work(struct bdi_writeback
*wb
,
177 struct wb_writeback_work
*work
)
179 struct wb_completion
*done
= work
->done
;
183 if (done
&& atomic_dec_and_test(&done
->cnt
))
184 wake_up_all(&wb
->bdi
->wb_waitq
);
187 static void wb_queue_work(struct bdi_writeback
*wb
,
188 struct wb_writeback_work
*work
)
190 trace_writeback_queue(wb
, work
);
193 atomic_inc(&work
->done
->cnt
);
195 spin_lock_bh(&wb
->work_lock
);
197 if (test_bit(WB_registered
, &wb
->state
)) {
198 list_add_tail(&work
->list
, &wb
->work_list
);
199 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
201 finish_writeback_work(wb
, work
);
203 spin_unlock_bh(&wb
->work_lock
);
207 * wb_wait_for_completion - wait for completion of bdi_writeback_works
208 * @bdi: bdi work items were issued to
209 * @done: target wb_completion
211 * Wait for one or more work items issued to @bdi with their ->done field
212 * set to @done, which should have been defined with
213 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
214 * work items are completed. Work items which are waited upon aren't freed
215 * automatically on completion.
217 static void wb_wait_for_completion(struct backing_dev_info
*bdi
,
218 struct wb_completion
*done
)
220 atomic_dec(&done
->cnt
); /* put down the initial count */
221 wait_event(bdi
->wb_waitq
, !atomic_read(&done
->cnt
));
224 #ifdef CONFIG_CGROUP_WRITEBACK
226 /* parameters for foreign inode detection, see wb_detach_inode() */
227 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
228 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
229 #define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
230 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
232 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
233 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
234 /* each slot's duration is 2s / 16 */
235 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
236 /* if foreign slots >= 8, switch */
237 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
238 /* one round can affect upto 5 slots */
240 static atomic_t isw_nr_in_flight
= ATOMIC_INIT(0);
241 static struct workqueue_struct
*isw_wq
;
243 void __inode_attach_wb(struct inode
*inode
, struct page
*page
)
245 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
246 struct bdi_writeback
*wb
= NULL
;
248 if (inode_cgwb_enabled(inode
)) {
249 struct cgroup_subsys_state
*memcg_css
;
252 memcg_css
= mem_cgroup_css_from_page(page
);
253 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
255 /* must pin memcg_css, see wb_get_create() */
256 memcg_css
= task_get_css(current
, memory_cgrp_id
);
257 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
266 * There may be multiple instances of this function racing to
267 * update the same inode. Use cmpxchg() to tell the winner.
269 if (unlikely(cmpxchg(&inode
->i_wb
, NULL
, wb
)))
274 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
275 * @inode: inode of interest with i_lock held
277 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
278 * held on entry and is released on return. The returned wb is guaranteed
279 * to stay @inode's associated wb until its list_lock is released.
281 static struct bdi_writeback
*
282 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
283 __releases(&inode
->i_lock
)
284 __acquires(&wb
->list_lock
)
287 struct bdi_writeback
*wb
= inode_to_wb(inode
);
290 * inode_to_wb() association is protected by both
291 * @inode->i_lock and @wb->list_lock but list_lock nests
292 * outside i_lock. Drop i_lock and verify that the
293 * association hasn't changed after acquiring list_lock.
296 spin_unlock(&inode
->i_lock
);
297 spin_lock(&wb
->list_lock
);
299 /* i_wb may have changed inbetween, can't use inode_to_wb() */
300 if (likely(wb
== inode
->i_wb
)) {
301 wb_put(wb
); /* @inode already has ref */
305 spin_unlock(&wb
->list_lock
);
308 spin_lock(&inode
->i_lock
);
313 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
314 * @inode: inode of interest
316 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
319 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
320 __acquires(&wb
->list_lock
)
322 spin_lock(&inode
->i_lock
);
323 return locked_inode_to_wb_and_lock_list(inode
);
326 struct inode_switch_wbs_context
{
328 struct bdi_writeback
*new_wb
;
330 struct rcu_head rcu_head
;
331 struct work_struct work
;
334 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
336 down_write(&bdi
->wb_switch_rwsem
);
339 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
341 up_write(&bdi
->wb_switch_rwsem
);
344 static void inode_switch_wbs_work_fn(struct work_struct
*work
)
346 struct inode_switch_wbs_context
*isw
=
347 container_of(work
, struct inode_switch_wbs_context
, work
);
348 struct inode
*inode
= isw
->inode
;
349 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
350 struct address_space
*mapping
= inode
->i_mapping
;
351 struct bdi_writeback
*old_wb
= inode
->i_wb
;
352 struct bdi_writeback
*new_wb
= isw
->new_wb
;
353 struct radix_tree_iter iter
;
354 bool switched
= false;
358 * If @inode switches cgwb membership while sync_inodes_sb() is
359 * being issued, sync_inodes_sb() might miss it. Synchronize.
361 down_read(&bdi
->wb_switch_rwsem
);
364 * By the time control reaches here, RCU grace period has passed
365 * since I_WB_SWITCH assertion and all wb stat update transactions
366 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
367 * synchronizing against mapping->tree_lock.
369 * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
370 * gives us exclusion against all wb related operations on @inode
371 * including IO list manipulations and stat updates.
373 if (old_wb
< new_wb
) {
374 spin_lock(&old_wb
->list_lock
);
375 spin_lock_nested(&new_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
377 spin_lock(&new_wb
->list_lock
);
378 spin_lock_nested(&old_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
380 spin_lock(&inode
->i_lock
);
381 spin_lock_irq(&mapping
->tree_lock
);
384 * Once I_FREEING is visible under i_lock, the eviction path owns
385 * the inode and we shouldn't modify ->i_io_list.
387 if (unlikely(inode
->i_state
& I_FREEING
))
391 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
392 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
393 * pages actually under underwriteback.
395 radix_tree_for_each_tagged(slot
, &mapping
->page_tree
, &iter
, 0,
396 PAGECACHE_TAG_DIRTY
) {
397 struct page
*page
= radix_tree_deref_slot_protected(slot
,
398 &mapping
->tree_lock
);
399 if (likely(page
) && PageDirty(page
)) {
400 dec_wb_stat(old_wb
, WB_RECLAIMABLE
);
401 inc_wb_stat(new_wb
, WB_RECLAIMABLE
);
405 radix_tree_for_each_tagged(slot
, &mapping
->page_tree
, &iter
, 0,
406 PAGECACHE_TAG_WRITEBACK
) {
407 struct page
*page
= radix_tree_deref_slot_protected(slot
,
408 &mapping
->tree_lock
);
410 WARN_ON_ONCE(!PageWriteback(page
));
411 dec_wb_stat(old_wb
, WB_WRITEBACK
);
412 inc_wb_stat(new_wb
, WB_WRITEBACK
);
419 * Transfer to @new_wb's IO list if necessary. The specific list
420 * @inode was on is ignored and the inode is put on ->b_dirty which
421 * is always correct including from ->b_dirty_time. The transfer
422 * preserves @inode->dirtied_when ordering.
424 if (!list_empty(&inode
->i_io_list
)) {
427 inode_io_list_del_locked(inode
, old_wb
);
428 inode
->i_wb
= new_wb
;
429 list_for_each_entry(pos
, &new_wb
->b_dirty
, i_io_list
)
430 if (time_after_eq(inode
->dirtied_when
,
433 inode_io_list_move_locked(inode
, new_wb
, pos
->i_io_list
.prev
);
435 inode
->i_wb
= new_wb
;
438 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
439 inode
->i_wb_frn_winner
= 0;
440 inode
->i_wb_frn_avg_time
= 0;
441 inode
->i_wb_frn_history
= 0;
445 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
446 * ensures that the new wb is visible if they see !I_WB_SWITCH.
448 smp_store_release(&inode
->i_state
, inode
->i_state
& ~I_WB_SWITCH
);
450 spin_unlock_irq(&mapping
->tree_lock
);
451 spin_unlock(&inode
->i_lock
);
452 spin_unlock(&new_wb
->list_lock
);
453 spin_unlock(&old_wb
->list_lock
);
455 up_read(&bdi
->wb_switch_rwsem
);
466 atomic_dec(&isw_nr_in_flight
);
469 static void inode_switch_wbs_rcu_fn(struct rcu_head
*rcu_head
)
471 struct inode_switch_wbs_context
*isw
= container_of(rcu_head
,
472 struct inode_switch_wbs_context
, rcu_head
);
474 /* needs to grab bh-unsafe locks, bounce to work item */
475 INIT_WORK(&isw
->work
, inode_switch_wbs_work_fn
);
476 queue_work(isw_wq
, &isw
->work
);
480 * inode_switch_wbs - change the wb association of an inode
481 * @inode: target inode
482 * @new_wb_id: ID of the new wb
484 * Switch @inode's wb association to the wb identified by @new_wb_id. The
485 * switching is performed asynchronously and may fail silently.
487 static void inode_switch_wbs(struct inode
*inode
, int new_wb_id
)
489 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
490 struct cgroup_subsys_state
*memcg_css
;
491 struct inode_switch_wbs_context
*isw
;
493 /* noop if seems to be already in progress */
494 if (inode
->i_state
& I_WB_SWITCH
)
498 * Avoid starting new switches while sync_inodes_sb() is in
499 * progress. Otherwise, if the down_write protected issue path
500 * blocks heavily, we might end up starting a large number of
501 * switches which will block on the rwsem.
503 if (!down_read_trylock(&bdi
->wb_switch_rwsem
))
506 isw
= kzalloc(sizeof(*isw
), GFP_ATOMIC
);
510 /* find and pin the new wb */
512 memcg_css
= css_from_id(new_wb_id
, &memory_cgrp_subsys
);
514 isw
->new_wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
519 /* while holding I_WB_SWITCH, no one else can update the association */
520 spin_lock(&inode
->i_lock
);
521 if (!(inode
->i_sb
->s_flags
& SB_ACTIVE
) ||
522 inode
->i_state
& (I_WB_SWITCH
| I_FREEING
) ||
523 inode_to_wb(inode
) == isw
->new_wb
) {
524 spin_unlock(&inode
->i_lock
);
527 inode
->i_state
|= I_WB_SWITCH
;
529 spin_unlock(&inode
->i_lock
);
534 * In addition to synchronizing among switchers, I_WB_SWITCH tells
535 * the RCU protected stat update paths to grab the mapping's
536 * tree_lock so that stat transfer can synchronize against them.
537 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
539 call_rcu(&isw
->rcu_head
, inode_switch_wbs_rcu_fn
);
541 atomic_inc(&isw_nr_in_flight
);
550 up_read(&bdi
->wb_switch_rwsem
);
554 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
555 * @wbc: writeback_control of interest
556 * @inode: target inode
558 * @inode is locked and about to be written back under the control of @wbc.
559 * Record @inode's writeback context into @wbc and unlock the i_lock. On
560 * writeback completion, wbc_detach_inode() should be called. This is used
561 * to track the cgroup writeback context.
563 void wbc_attach_and_unlock_inode(struct writeback_control
*wbc
,
566 if (!inode_cgwb_enabled(inode
)) {
567 spin_unlock(&inode
->i_lock
);
571 wbc
->wb
= inode_to_wb(inode
);
574 wbc
->wb_id
= wbc
->wb
->memcg_css
->id
;
575 wbc
->wb_lcand_id
= inode
->i_wb_frn_winner
;
576 wbc
->wb_tcand_id
= 0;
578 wbc
->wb_lcand_bytes
= 0;
579 wbc
->wb_tcand_bytes
= 0;
582 spin_unlock(&inode
->i_lock
);
585 * A dying wb indicates that either the blkcg associated with the
586 * memcg changed or the associated memcg is dying. In the first
587 * case, a replacement wb should already be available and we should
588 * refresh the wb immediately. In the second case, trying to
589 * refresh will keep failing.
591 if (unlikely(wb_dying(wbc
->wb
) && !css_is_dying(wbc
->wb
->memcg_css
)))
592 inode_switch_wbs(inode
, wbc
->wb_id
);
596 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
597 * @wbc: writeback_control of the just finished writeback
599 * To be called after a writeback attempt of an inode finishes and undoes
600 * wbc_attach_and_unlock_inode(). Can be called under any context.
602 * As concurrent write sharing of an inode is expected to be very rare and
603 * memcg only tracks page ownership on first-use basis severely confining
604 * the usefulness of such sharing, cgroup writeback tracks ownership
605 * per-inode. While the support for concurrent write sharing of an inode
606 * is deemed unnecessary, an inode being written to by different cgroups at
607 * different points in time is a lot more common, and, more importantly,
608 * charging only by first-use can too readily lead to grossly incorrect
609 * behaviors (single foreign page can lead to gigabytes of writeback to be
610 * incorrectly attributed).
612 * To resolve this issue, cgroup writeback detects the majority dirtier of
613 * an inode and transfers the ownership to it. To avoid unnnecessary
614 * oscillation, the detection mechanism keeps track of history and gives
615 * out the switch verdict only if the foreign usage pattern is stable over
616 * a certain amount of time and/or writeback attempts.
618 * On each writeback attempt, @wbc tries to detect the majority writer
619 * using Boyer-Moore majority vote algorithm. In addition to the byte
620 * count from the majority voting, it also counts the bytes written for the
621 * current wb and the last round's winner wb (max of last round's current
622 * wb, the winner from two rounds ago, and the last round's majority
623 * candidate). Keeping track of the historical winner helps the algorithm
624 * to semi-reliably detect the most active writer even when it's not the
627 * Once the winner of the round is determined, whether the winner is
628 * foreign or not and how much IO time the round consumed is recorded in
629 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
630 * over a certain threshold, the switch verdict is given.
632 void wbc_detach_inode(struct writeback_control
*wbc
)
634 struct bdi_writeback
*wb
= wbc
->wb
;
635 struct inode
*inode
= wbc
->inode
;
636 unsigned long avg_time
, max_bytes
, max_time
;
643 history
= inode
->i_wb_frn_history
;
644 avg_time
= inode
->i_wb_frn_avg_time
;
646 /* pick the winner of this round */
647 if (wbc
->wb_bytes
>= wbc
->wb_lcand_bytes
&&
648 wbc
->wb_bytes
>= wbc
->wb_tcand_bytes
) {
650 max_bytes
= wbc
->wb_bytes
;
651 } else if (wbc
->wb_lcand_bytes
>= wbc
->wb_tcand_bytes
) {
652 max_id
= wbc
->wb_lcand_id
;
653 max_bytes
= wbc
->wb_lcand_bytes
;
655 max_id
= wbc
->wb_tcand_id
;
656 max_bytes
= wbc
->wb_tcand_bytes
;
660 * Calculate the amount of IO time the winner consumed and fold it
661 * into the running average kept per inode. If the consumed IO
662 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
663 * deciding whether to switch or not. This is to prevent one-off
664 * small dirtiers from skewing the verdict.
666 max_time
= DIV_ROUND_UP((max_bytes
>> PAGE_SHIFT
) << WB_FRN_TIME_SHIFT
,
667 wb
->avg_write_bandwidth
);
669 avg_time
+= (max_time
>> WB_FRN_TIME_AVG_SHIFT
) -
670 (avg_time
>> WB_FRN_TIME_AVG_SHIFT
);
672 avg_time
= max_time
; /* immediate catch up on first run */
674 if (max_time
>= avg_time
/ WB_FRN_TIME_CUT_DIV
) {
678 * The switch verdict is reached if foreign wb's consume
679 * more than a certain proportion of IO time in a
680 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
681 * history mask where each bit represents one sixteenth of
682 * the period. Determine the number of slots to shift into
683 * history from @max_time.
685 slots
= min(DIV_ROUND_UP(max_time
, WB_FRN_HIST_UNIT
),
686 (unsigned long)WB_FRN_HIST_MAX_SLOTS
);
688 if (wbc
->wb_id
!= max_id
)
689 history
|= (1U << slots
) - 1;
692 * Switch if the current wb isn't the consistent winner.
693 * If there are multiple closely competing dirtiers, the
694 * inode may switch across them repeatedly over time, which
695 * is okay. The main goal is avoiding keeping an inode on
696 * the wrong wb for an extended period of time.
698 if (hweight32(history
) > WB_FRN_HIST_THR_SLOTS
)
699 inode_switch_wbs(inode
, max_id
);
703 * Multiple instances of this function may race to update the
704 * following fields but we don't mind occassional inaccuracies.
706 inode
->i_wb_frn_winner
= max_id
;
707 inode
->i_wb_frn_avg_time
= min(avg_time
, (unsigned long)U16_MAX
);
708 inode
->i_wb_frn_history
= history
;
715 * wbc_account_io - account IO issued during writeback
716 * @wbc: writeback_control of the writeback in progress
717 * @page: page being written out
718 * @bytes: number of bytes being written out
720 * @bytes from @page are about to written out during the writeback
721 * controlled by @wbc. Keep the book for foreign inode detection. See
722 * wbc_detach_inode().
724 void wbc_account_io(struct writeback_control
*wbc
, struct page
*page
,
727 struct cgroup_subsys_state
*css
;
731 * pageout() path doesn't attach @wbc to the inode being written
732 * out. This is intentional as we don't want the function to block
733 * behind a slow cgroup. Ultimately, we want pageout() to kick off
734 * regular writeback instead of writing things out itself.
739 css
= mem_cgroup_css_from_page(page
);
740 /* dead cgroups shouldn't contribute to inode ownership arbitration */
741 if (!(css
->flags
& CSS_ONLINE
))
746 if (id
== wbc
->wb_id
) {
747 wbc
->wb_bytes
+= bytes
;
751 if (id
== wbc
->wb_lcand_id
)
752 wbc
->wb_lcand_bytes
+= bytes
;
754 /* Boyer-Moore majority vote algorithm */
755 if (!wbc
->wb_tcand_bytes
)
756 wbc
->wb_tcand_id
= id
;
757 if (id
== wbc
->wb_tcand_id
)
758 wbc
->wb_tcand_bytes
+= bytes
;
760 wbc
->wb_tcand_bytes
-= min(bytes
, wbc
->wb_tcand_bytes
);
762 EXPORT_SYMBOL_GPL(wbc_account_io
);
765 * inode_congested - test whether an inode is congested
766 * @inode: inode to test for congestion (may be NULL)
767 * @cong_bits: mask of WB_[a]sync_congested bits to test
769 * Tests whether @inode is congested. @cong_bits is the mask of congestion
770 * bits to test and the return value is the mask of set bits.
772 * If cgroup writeback is enabled for @inode, the congestion state is
773 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
774 * associated with @inode is congested; otherwise, the root wb's congestion
777 * @inode is allowed to be NULL as this function is often called on
778 * mapping->host which is NULL for the swapper space.
780 int inode_congested(struct inode
*inode
, int cong_bits
)
783 * Once set, ->i_wb never becomes NULL while the inode is alive.
784 * Start transaction iff ->i_wb is visible.
786 if (inode
&& inode_to_wb_is_valid(inode
)) {
787 struct bdi_writeback
*wb
;
788 struct wb_lock_cookie lock_cookie
= {};
791 wb
= unlocked_inode_to_wb_begin(inode
, &lock_cookie
);
792 congested
= wb_congested(wb
, cong_bits
);
793 unlocked_inode_to_wb_end(inode
, &lock_cookie
);
797 return wb_congested(&inode_to_bdi(inode
)->wb
, cong_bits
);
799 EXPORT_SYMBOL_GPL(inode_congested
);
802 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
803 * @wb: target bdi_writeback to split @nr_pages to
804 * @nr_pages: number of pages to write for the whole bdi
806 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
807 * relation to the total write bandwidth of all wb's w/ dirty inodes on
810 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
812 unsigned long this_bw
= wb
->avg_write_bandwidth
;
813 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
815 if (nr_pages
== LONG_MAX
)
819 * This may be called on clean wb's and proportional distribution
820 * may not make sense, just use the original @nr_pages in those
821 * cases. In general, we wanna err on the side of writing more.
823 if (!tot_bw
|| this_bw
>= tot_bw
)
826 return DIV_ROUND_UP_ULL((u64
)nr_pages
* this_bw
, tot_bw
);
830 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
831 * @bdi: target backing_dev_info
832 * @base_work: wb_writeback_work to issue
833 * @skip_if_busy: skip wb's which already have writeback in progress
835 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
836 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
837 * distributed to the busy wbs according to each wb's proportion in the
838 * total active write bandwidth of @bdi.
840 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
841 struct wb_writeback_work
*base_work
,
844 struct bdi_writeback
*last_wb
= NULL
;
845 struct bdi_writeback
*wb
= list_entry(&bdi
->wb_list
,
846 struct bdi_writeback
, bdi_node
);
851 list_for_each_entry_continue_rcu(wb
, &bdi
->wb_list
, bdi_node
) {
852 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done
);
853 struct wb_writeback_work fallback_work
;
854 struct wb_writeback_work
*work
;
862 /* SYNC_ALL writes out I_DIRTY_TIME too */
863 if (!wb_has_dirty_io(wb
) &&
864 (base_work
->sync_mode
== WB_SYNC_NONE
||
865 list_empty(&wb
->b_dirty_time
)))
867 if (skip_if_busy
&& writeback_in_progress(wb
))
870 nr_pages
= wb_split_bdi_pages(wb
, base_work
->nr_pages
);
872 work
= kmalloc(sizeof(*work
), GFP_ATOMIC
);
875 work
->nr_pages
= nr_pages
;
877 wb_queue_work(wb
, work
);
881 /* alloc failed, execute synchronously using on-stack fallback */
882 work
= &fallback_work
;
884 work
->nr_pages
= nr_pages
;
886 work
->done
= &fallback_work_done
;
888 wb_queue_work(wb
, work
);
891 * Pin @wb so that it stays on @bdi->wb_list. This allows
892 * continuing iteration from @wb after dropping and
893 * regrabbing rcu read lock.
899 wb_wait_for_completion(bdi
, &fallback_work_done
);
909 * cgroup_writeback_umount - flush inode wb switches for umount
911 * This function is called when a super_block is about to be destroyed and
912 * flushes in-flight inode wb switches. An inode wb switch goes through
913 * RCU and then workqueue, so the two need to be flushed in order to ensure
914 * that all previously scheduled switches are finished. As wb switches are
915 * rare occurrences and synchronize_rcu() can take a while, perform
916 * flushing iff wb switches are in flight.
918 void cgroup_writeback_umount(void)
920 if (atomic_read(&isw_nr_in_flight
)) {
922 * Use rcu_barrier() to wait for all pending callbacks to
923 * ensure that all in-flight wb switches are in the workqueue.
926 flush_workqueue(isw_wq
);
930 static int __init
cgroup_writeback_init(void)
932 isw_wq
= alloc_workqueue("inode_switch_wbs", 0, 0);
937 fs_initcall(cgroup_writeback_init
);
939 #else /* CONFIG_CGROUP_WRITEBACK */
941 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
942 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
944 static struct bdi_writeback
*
945 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
946 __releases(&inode
->i_lock
)
947 __acquires(&wb
->list_lock
)
949 struct bdi_writeback
*wb
= inode_to_wb(inode
);
951 spin_unlock(&inode
->i_lock
);
952 spin_lock(&wb
->list_lock
);
956 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
957 __acquires(&wb
->list_lock
)
959 struct bdi_writeback
*wb
= inode_to_wb(inode
);
961 spin_lock(&wb
->list_lock
);
965 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
970 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
971 struct wb_writeback_work
*base_work
,
976 if (!skip_if_busy
|| !writeback_in_progress(&bdi
->wb
)) {
977 base_work
->auto_free
= 0;
978 wb_queue_work(&bdi
->wb
, base_work
);
982 #endif /* CONFIG_CGROUP_WRITEBACK */
985 * Add in the number of potentially dirty inodes, because each inode
986 * write can dirty pagecache in the underlying blockdev.
988 static unsigned long get_nr_dirty_pages(void)
990 return global_node_page_state(NR_FILE_DIRTY
) +
991 global_node_page_state(NR_UNSTABLE_NFS
) +
992 get_nr_dirty_inodes();
995 static void wb_start_writeback(struct bdi_writeback
*wb
, enum wb_reason reason
)
997 if (!wb_has_dirty_io(wb
))
1001 * All callers of this function want to start writeback of all
1002 * dirty pages. Places like vmscan can call this at a very
1003 * high frequency, causing pointless allocations of tons of
1004 * work items and keeping the flusher threads busy retrieving
1005 * that work. Ensure that we only allow one of them pending and
1006 * inflight at the time.
1008 if (test_bit(WB_start_all
, &wb
->state
) ||
1009 test_and_set_bit(WB_start_all
, &wb
->state
))
1012 wb
->start_all_reason
= reason
;
1017 * wb_start_background_writeback - start background writeback
1018 * @wb: bdi_writback to write from
1021 * This makes sure WB_SYNC_NONE background writeback happens. When
1022 * this function returns, it is only guaranteed that for given wb
1023 * some IO is happening if we are over background dirty threshold.
1024 * Caller need not hold sb s_umount semaphore.
1026 void wb_start_background_writeback(struct bdi_writeback
*wb
)
1029 * We just wake up the flusher thread. It will perform background
1030 * writeback as soon as there is no other work to do.
1032 trace_writeback_wake_background(wb
);
1037 * Remove the inode from the writeback list it is on.
1039 void inode_io_list_del(struct inode
*inode
)
1041 struct bdi_writeback
*wb
;
1043 wb
= inode_to_wb_and_lock_list(inode
);
1044 inode_io_list_del_locked(inode
, wb
);
1045 spin_unlock(&wb
->list_lock
);
1049 * mark an inode as under writeback on the sb
1051 void sb_mark_inode_writeback(struct inode
*inode
)
1053 struct super_block
*sb
= inode
->i_sb
;
1054 unsigned long flags
;
1056 if (list_empty(&inode
->i_wb_list
)) {
1057 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1058 if (list_empty(&inode
->i_wb_list
)) {
1059 list_add_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
1060 trace_sb_mark_inode_writeback(inode
);
1062 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1067 * clear an inode as under writeback on the sb
1069 void sb_clear_inode_writeback(struct inode
*inode
)
1071 struct super_block
*sb
= inode
->i_sb
;
1072 unsigned long flags
;
1074 if (!list_empty(&inode
->i_wb_list
)) {
1075 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1076 if (!list_empty(&inode
->i_wb_list
)) {
1077 list_del_init(&inode
->i_wb_list
);
1078 trace_sb_clear_inode_writeback(inode
);
1080 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1085 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1086 * furthest end of its superblock's dirty-inode list.
1088 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1089 * already the most-recently-dirtied inode on the b_dirty list. If that is
1090 * the case then the inode must have been redirtied while it was being written
1091 * out and we don't reset its dirtied_when.
1093 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
1095 if (!list_empty(&wb
->b_dirty
)) {
1098 tail
= wb_inode(wb
->b_dirty
.next
);
1099 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
1100 inode
->dirtied_when
= jiffies
;
1102 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty
);
1106 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1108 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
1110 inode_io_list_move_locked(inode
, wb
, &wb
->b_more_io
);
1113 static void inode_sync_complete(struct inode
*inode
)
1115 inode
->i_state
&= ~I_SYNC
;
1116 /* If inode is clean an unused, put it into LRU now... */
1117 inode_add_lru(inode
);
1118 /* Waiters must see I_SYNC cleared before being woken up */
1120 wake_up_bit(&inode
->i_state
, __I_SYNC
);
1123 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
1125 bool ret
= time_after(inode
->dirtied_when
, t
);
1126 #ifndef CONFIG_64BIT
1128 * For inodes being constantly redirtied, dirtied_when can get stuck.
1129 * It _appears_ to be in the future, but is actually in distant past.
1130 * This test is necessary to prevent such wrapped-around relative times
1131 * from permanently stopping the whole bdi writeback.
1133 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
1138 #define EXPIRE_DIRTY_ATIME 0x0001
1141 * Move expired (dirtied before work->older_than_this) dirty inodes from
1142 * @delaying_queue to @dispatch_queue.
1144 static int move_expired_inodes(struct list_head
*delaying_queue
,
1145 struct list_head
*dispatch_queue
,
1147 struct wb_writeback_work
*work
)
1149 unsigned long *older_than_this
= NULL
;
1150 unsigned long expire_time
;
1152 struct list_head
*pos
, *node
;
1153 struct super_block
*sb
= NULL
;
1154 struct inode
*inode
;
1158 if ((flags
& EXPIRE_DIRTY_ATIME
) == 0)
1159 older_than_this
= work
->older_than_this
;
1160 else if (!work
->for_sync
) {
1161 expire_time
= jiffies
- (dirtytime_expire_interval
* HZ
);
1162 older_than_this
= &expire_time
;
1164 while (!list_empty(delaying_queue
)) {
1165 inode
= wb_inode(delaying_queue
->prev
);
1166 if (older_than_this
&&
1167 inode_dirtied_after(inode
, *older_than_this
))
1169 list_move(&inode
->i_io_list
, &tmp
);
1171 if (flags
& EXPIRE_DIRTY_ATIME
)
1172 set_bit(__I_DIRTY_TIME_EXPIRED
, &inode
->i_state
);
1173 if (sb_is_blkdev_sb(inode
->i_sb
))
1175 if (sb
&& sb
!= inode
->i_sb
)
1180 /* just one sb in list, splice to dispatch_queue and we're done */
1182 list_splice(&tmp
, dispatch_queue
);
1186 /* Move inodes from one superblock together */
1187 while (!list_empty(&tmp
)) {
1188 sb
= wb_inode(tmp
.prev
)->i_sb
;
1189 list_for_each_prev_safe(pos
, node
, &tmp
) {
1190 inode
= wb_inode(pos
);
1191 if (inode
->i_sb
== sb
)
1192 list_move(&inode
->i_io_list
, dispatch_queue
);
1200 * Queue all expired dirty inodes for io, eldest first.
1202 * newly dirtied b_dirty b_io b_more_io
1203 * =============> gf edc BA
1205 * newly dirtied b_dirty b_io b_more_io
1206 * =============> g fBAedc
1208 * +--> dequeue for IO
1210 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
)
1214 assert_spin_locked(&wb
->list_lock
);
1215 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
1216 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, 0, work
);
1217 moved
+= move_expired_inodes(&wb
->b_dirty_time
, &wb
->b_io
,
1218 EXPIRE_DIRTY_ATIME
, work
);
1220 wb_io_lists_populated(wb
);
1221 trace_writeback_queue_io(wb
, work
, moved
);
1224 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1228 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
1229 trace_writeback_write_inode_start(inode
, wbc
);
1230 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
1231 trace_writeback_write_inode(inode
, wbc
);
1238 * Wait for writeback on an inode to complete. Called with i_lock held.
1239 * Caller must make sure inode cannot go away when we drop i_lock.
1241 static void __inode_wait_for_writeback(struct inode
*inode
)
1242 __releases(inode
->i_lock
)
1243 __acquires(inode
->i_lock
)
1245 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
1246 wait_queue_head_t
*wqh
;
1248 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1249 while (inode
->i_state
& I_SYNC
) {
1250 spin_unlock(&inode
->i_lock
);
1251 __wait_on_bit(wqh
, &wq
, bit_wait
,
1252 TASK_UNINTERRUPTIBLE
);
1253 spin_lock(&inode
->i_lock
);
1258 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1260 void inode_wait_for_writeback(struct inode
*inode
)
1262 spin_lock(&inode
->i_lock
);
1263 __inode_wait_for_writeback(inode
);
1264 spin_unlock(&inode
->i_lock
);
1268 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1269 * held and drops it. It is aimed for callers not holding any inode reference
1270 * so once i_lock is dropped, inode can go away.
1272 static void inode_sleep_on_writeback(struct inode
*inode
)
1273 __releases(inode
->i_lock
)
1276 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1279 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
1280 sleep
= inode
->i_state
& I_SYNC
;
1281 spin_unlock(&inode
->i_lock
);
1284 finish_wait(wqh
, &wait
);
1288 * Find proper writeback list for the inode depending on its current state and
1289 * possibly also change of its state while we were doing writeback. Here we
1290 * handle things such as livelock prevention or fairness of writeback among
1291 * inodes. This function can be called only by flusher thread - noone else
1292 * processes all inodes in writeback lists and requeueing inodes behind flusher
1293 * thread's back can have unexpected consequences.
1295 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
1296 struct writeback_control
*wbc
)
1298 if (inode
->i_state
& I_FREEING
)
1302 * Sync livelock prevention. Each inode is tagged and synced in one
1303 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1304 * the dirty time to prevent enqueue and sync it again.
1306 if ((inode
->i_state
& I_DIRTY
) &&
1307 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
1308 inode
->dirtied_when
= jiffies
;
1310 if (wbc
->pages_skipped
) {
1312 * writeback is not making progress due to locked
1313 * buffers. Skip this inode for now.
1315 redirty_tail(inode
, wb
);
1319 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
1321 * We didn't write back all the pages. nfs_writepages()
1322 * sometimes bales out without doing anything.
1324 if (wbc
->nr_to_write
<= 0) {
1325 /* Slice used up. Queue for next turn. */
1326 requeue_io(inode
, wb
);
1329 * Writeback blocked by something other than
1330 * congestion. Delay the inode for some time to
1331 * avoid spinning on the CPU (100% iowait)
1332 * retrying writeback of the dirty page/inode
1333 * that cannot be performed immediately.
1335 redirty_tail(inode
, wb
);
1337 } else if (inode
->i_state
& I_DIRTY
) {
1339 * Filesystems can dirty the inode during writeback operations,
1340 * such as delayed allocation during submission or metadata
1341 * updates after data IO completion.
1343 redirty_tail(inode
, wb
);
1344 } else if (inode
->i_state
& I_DIRTY_TIME
) {
1345 inode
->dirtied_when
= jiffies
;
1346 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty_time
);
1348 /* The inode is clean. Remove from writeback lists. */
1349 inode_io_list_del_locked(inode
, wb
);
1354 * Write out an inode and its dirty pages. Do not update the writeback list
1355 * linkage. That is left to the caller. The caller is also responsible for
1356 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1359 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1361 struct address_space
*mapping
= inode
->i_mapping
;
1362 long nr_to_write
= wbc
->nr_to_write
;
1366 WARN_ON(!(inode
->i_state
& I_SYNC
));
1368 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
1370 ret
= do_writepages(mapping
, wbc
);
1373 * Make sure to wait on the data before writing out the metadata.
1374 * This is important for filesystems that modify metadata on data
1375 * I/O completion. We don't do it for sync(2) writeback because it has a
1376 * separate, external IO completion path and ->sync_fs for guaranteeing
1377 * inode metadata is written back correctly.
1379 if (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
) {
1380 int err
= filemap_fdatawait(mapping
);
1386 * Some filesystems may redirty the inode during the writeback
1387 * due to delalloc, clear dirty metadata flags right before
1390 spin_lock(&inode
->i_lock
);
1392 dirty
= inode
->i_state
& I_DIRTY
;
1393 if (inode
->i_state
& I_DIRTY_TIME
) {
1394 if ((dirty
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
)) ||
1395 wbc
->sync_mode
== WB_SYNC_ALL
||
1396 unlikely(inode
->i_state
& I_DIRTY_TIME_EXPIRED
) ||
1397 unlikely(time_after(jiffies
,
1398 (inode
->dirtied_time_when
+
1399 dirtytime_expire_interval
* HZ
)))) {
1400 dirty
|= I_DIRTY_TIME
| I_DIRTY_TIME_EXPIRED
;
1401 trace_writeback_lazytime(inode
);
1404 inode
->i_state
&= ~I_DIRTY_TIME_EXPIRED
;
1405 inode
->i_state
&= ~dirty
;
1408 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1409 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1410 * either they see the I_DIRTY bits cleared or we see the dirtied
1413 * I_DIRTY_PAGES is always cleared together above even if @mapping
1414 * still has dirty pages. The flag is reinstated after smp_mb() if
1415 * necessary. This guarantees that either __mark_inode_dirty()
1416 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1420 if (mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
1421 inode
->i_state
|= I_DIRTY_PAGES
;
1423 spin_unlock(&inode
->i_lock
);
1425 if (dirty
& I_DIRTY_TIME
)
1426 mark_inode_dirty_sync(inode
);
1427 /* Don't write the inode if only I_DIRTY_PAGES was set */
1428 if (dirty
& ~I_DIRTY_PAGES
) {
1429 int err
= write_inode(inode
, wbc
);
1433 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
1438 * Write out an inode's dirty pages. Either the caller has an active reference
1439 * on the inode or the inode has I_WILL_FREE set.
1441 * This function is designed to be called for writing back one inode which
1442 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1443 * and does more profound writeback list handling in writeback_sb_inodes().
1445 static int writeback_single_inode(struct inode
*inode
,
1446 struct writeback_control
*wbc
)
1448 struct bdi_writeback
*wb
;
1451 spin_lock(&inode
->i_lock
);
1452 if (!atomic_read(&inode
->i_count
))
1453 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
1455 WARN_ON(inode
->i_state
& I_WILL_FREE
);
1457 if (inode
->i_state
& I_SYNC
) {
1458 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
1461 * It's a data-integrity sync. We must wait. Since callers hold
1462 * inode reference or inode has I_WILL_FREE set, it cannot go
1465 __inode_wait_for_writeback(inode
);
1467 WARN_ON(inode
->i_state
& I_SYNC
);
1469 * Skip inode if it is clean and we have no outstanding writeback in
1470 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1471 * function since flusher thread may be doing for example sync in
1472 * parallel and if we move the inode, it could get skipped. So here we
1473 * make sure inode is on some writeback list and leave it there unless
1474 * we have completely cleaned the inode.
1476 if (!(inode
->i_state
& I_DIRTY_ALL
) &&
1477 (wbc
->sync_mode
!= WB_SYNC_ALL
||
1478 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
1480 inode
->i_state
|= I_SYNC
;
1481 wbc_attach_and_unlock_inode(wbc
, inode
);
1483 ret
= __writeback_single_inode(inode
, wbc
);
1485 wbc_detach_inode(wbc
);
1487 wb
= inode_to_wb_and_lock_list(inode
);
1488 spin_lock(&inode
->i_lock
);
1490 * If inode is clean, remove it from writeback lists. Otherwise don't
1491 * touch it. See comment above for explanation.
1493 if (!(inode
->i_state
& I_DIRTY_ALL
))
1494 inode_io_list_del_locked(inode
, wb
);
1495 spin_unlock(&wb
->list_lock
);
1496 inode_sync_complete(inode
);
1498 spin_unlock(&inode
->i_lock
);
1502 static long writeback_chunk_size(struct bdi_writeback
*wb
,
1503 struct wb_writeback_work
*work
)
1508 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1509 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1510 * here avoids calling into writeback_inodes_wb() more than once.
1512 * The intended call sequence for WB_SYNC_ALL writeback is:
1515 * writeback_sb_inodes() <== called only once
1516 * write_cache_pages() <== called once for each inode
1517 * (quickly) tag currently dirty pages
1518 * (maybe slowly) sync all tagged pages
1520 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
1523 pages
= min(wb
->avg_write_bandwidth
/ 2,
1524 global_wb_domain
.dirty_limit
/ DIRTY_SCOPE
);
1525 pages
= min(pages
, work
->nr_pages
);
1526 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
1527 MIN_WRITEBACK_PAGES
);
1534 * Write a portion of b_io inodes which belong to @sb.
1536 * Return the number of pages and/or inodes written.
1538 * NOTE! This is called with wb->list_lock held, and will
1539 * unlock and relock that for each inode it ends up doing
1542 static long writeback_sb_inodes(struct super_block
*sb
,
1543 struct bdi_writeback
*wb
,
1544 struct wb_writeback_work
*work
)
1546 struct writeback_control wbc
= {
1547 .sync_mode
= work
->sync_mode
,
1548 .tagged_writepages
= work
->tagged_writepages
,
1549 .for_kupdate
= work
->for_kupdate
,
1550 .for_background
= work
->for_background
,
1551 .for_sync
= work
->for_sync
,
1552 .range_cyclic
= work
->range_cyclic
,
1554 .range_end
= LLONG_MAX
,
1556 unsigned long start_time
= jiffies
;
1558 long wrote
= 0; /* count both pages and inodes */
1560 while (!list_empty(&wb
->b_io
)) {
1561 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1562 struct bdi_writeback
*tmp_wb
;
1564 if (inode
->i_sb
!= sb
) {
1567 * We only want to write back data for this
1568 * superblock, move all inodes not belonging
1569 * to it back onto the dirty list.
1571 redirty_tail(inode
, wb
);
1576 * The inode belongs to a different superblock.
1577 * Bounce back to the caller to unpin this and
1578 * pin the next superblock.
1584 * Don't bother with new inodes or inodes being freed, first
1585 * kind does not need periodic writeout yet, and for the latter
1586 * kind writeout is handled by the freer.
1588 spin_lock(&inode
->i_lock
);
1589 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
1590 spin_unlock(&inode
->i_lock
);
1591 redirty_tail(inode
, wb
);
1594 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
1596 * If this inode is locked for writeback and we are not
1597 * doing writeback-for-data-integrity, move it to
1598 * b_more_io so that writeback can proceed with the
1599 * other inodes on s_io.
1601 * We'll have another go at writing back this inode
1602 * when we completed a full scan of b_io.
1604 spin_unlock(&inode
->i_lock
);
1605 requeue_io(inode
, wb
);
1606 trace_writeback_sb_inodes_requeue(inode
);
1609 spin_unlock(&wb
->list_lock
);
1612 * We already requeued the inode if it had I_SYNC set and we
1613 * are doing WB_SYNC_NONE writeback. So this catches only the
1616 if (inode
->i_state
& I_SYNC
) {
1617 /* Wait for I_SYNC. This function drops i_lock... */
1618 inode_sleep_on_writeback(inode
);
1619 /* Inode may be gone, start again */
1620 spin_lock(&wb
->list_lock
);
1623 inode
->i_state
|= I_SYNC
;
1624 wbc_attach_and_unlock_inode(&wbc
, inode
);
1626 write_chunk
= writeback_chunk_size(wb
, work
);
1627 wbc
.nr_to_write
= write_chunk
;
1628 wbc
.pages_skipped
= 0;
1631 * We use I_SYNC to pin the inode in memory. While it is set
1632 * evict_inode() will wait so the inode cannot be freed.
1634 __writeback_single_inode(inode
, &wbc
);
1636 wbc_detach_inode(&wbc
);
1637 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
1638 wrote
+= write_chunk
- wbc
.nr_to_write
;
1640 if (need_resched()) {
1642 * We're trying to balance between building up a nice
1643 * long list of IOs to improve our merge rate, and
1644 * getting those IOs out quickly for anyone throttling
1645 * in balance_dirty_pages(). cond_resched() doesn't
1646 * unplug, so get our IOs out the door before we
1649 blk_flush_plug(current
);
1654 * Requeue @inode if still dirty. Be careful as @inode may
1655 * have been switched to another wb in the meantime.
1657 tmp_wb
= inode_to_wb_and_lock_list(inode
);
1658 spin_lock(&inode
->i_lock
);
1659 if (!(inode
->i_state
& I_DIRTY_ALL
))
1661 requeue_inode(inode
, tmp_wb
, &wbc
);
1662 inode_sync_complete(inode
);
1663 spin_unlock(&inode
->i_lock
);
1665 if (unlikely(tmp_wb
!= wb
)) {
1666 spin_unlock(&tmp_wb
->list_lock
);
1667 spin_lock(&wb
->list_lock
);
1671 * bail out to wb_writeback() often enough to check
1672 * background threshold and other termination conditions.
1675 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1677 if (work
->nr_pages
<= 0)
1684 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
1685 struct wb_writeback_work
*work
)
1687 unsigned long start_time
= jiffies
;
1690 while (!list_empty(&wb
->b_io
)) {
1691 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1692 struct super_block
*sb
= inode
->i_sb
;
1694 if (!trylock_super(sb
)) {
1696 * trylock_super() may fail consistently due to
1697 * s_umount being grabbed by someone else. Don't use
1698 * requeue_io() to avoid busy retrying the inode/sb.
1700 redirty_tail(inode
, wb
);
1703 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
1704 up_read(&sb
->s_umount
);
1706 /* refer to the same tests at the end of writeback_sb_inodes */
1708 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1710 if (work
->nr_pages
<= 0)
1714 /* Leave any unwritten inodes on b_io */
1718 static long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
1719 enum wb_reason reason
)
1721 struct wb_writeback_work work
= {
1722 .nr_pages
= nr_pages
,
1723 .sync_mode
= WB_SYNC_NONE
,
1727 struct blk_plug plug
;
1729 blk_start_plug(&plug
);
1730 spin_lock(&wb
->list_lock
);
1731 if (list_empty(&wb
->b_io
))
1732 queue_io(wb
, &work
);
1733 __writeback_inodes_wb(wb
, &work
);
1734 spin_unlock(&wb
->list_lock
);
1735 blk_finish_plug(&plug
);
1737 return nr_pages
- work
.nr_pages
;
1741 * Explicit flushing or periodic writeback of "old" data.
1743 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1744 * dirtying-time in the inode's address_space. So this periodic writeback code
1745 * just walks the superblock inode list, writing back any inodes which are
1746 * older than a specific point in time.
1748 * Try to run once per dirty_writeback_interval. But if a writeback event
1749 * takes longer than a dirty_writeback_interval interval, then leave a
1752 * older_than_this takes precedence over nr_to_write. So we'll only write back
1753 * all dirty pages if they are all attached to "old" mappings.
1755 static long wb_writeback(struct bdi_writeback
*wb
,
1756 struct wb_writeback_work
*work
)
1758 unsigned long wb_start
= jiffies
;
1759 long nr_pages
= work
->nr_pages
;
1760 unsigned long oldest_jif
;
1761 struct inode
*inode
;
1763 struct blk_plug plug
;
1765 oldest_jif
= jiffies
;
1766 work
->older_than_this
= &oldest_jif
;
1768 blk_start_plug(&plug
);
1769 spin_lock(&wb
->list_lock
);
1772 * Stop writeback when nr_pages has been consumed
1774 if (work
->nr_pages
<= 0)
1778 * Background writeout and kupdate-style writeback may
1779 * run forever. Stop them if there is other work to do
1780 * so that e.g. sync can proceed. They'll be restarted
1781 * after the other works are all done.
1783 if ((work
->for_background
|| work
->for_kupdate
) &&
1784 !list_empty(&wb
->work_list
))
1788 * For background writeout, stop when we are below the
1789 * background dirty threshold
1791 if (work
->for_background
&& !wb_over_bg_thresh(wb
))
1795 * Kupdate and background works are special and we want to
1796 * include all inodes that need writing. Livelock avoidance is
1797 * handled by these works yielding to any other work so we are
1800 if (work
->for_kupdate
) {
1801 oldest_jif
= jiffies
-
1802 msecs_to_jiffies(dirty_expire_interval
* 10);
1803 } else if (work
->for_background
)
1804 oldest_jif
= jiffies
;
1806 trace_writeback_start(wb
, work
);
1807 if (list_empty(&wb
->b_io
))
1810 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
1812 progress
= __writeback_inodes_wb(wb
, work
);
1813 trace_writeback_written(wb
, work
);
1815 wb_update_bandwidth(wb
, wb_start
);
1818 * Did we write something? Try for more
1820 * Dirty inodes are moved to b_io for writeback in batches.
1821 * The completion of the current batch does not necessarily
1822 * mean the overall work is done. So we keep looping as long
1823 * as made some progress on cleaning pages or inodes.
1828 * No more inodes for IO, bail
1830 if (list_empty(&wb
->b_more_io
))
1833 * Nothing written. Wait for some inode to
1834 * become available for writeback. Otherwise
1835 * we'll just busyloop.
1837 trace_writeback_wait(wb
, work
);
1838 inode
= wb_inode(wb
->b_more_io
.prev
);
1839 spin_lock(&inode
->i_lock
);
1840 spin_unlock(&wb
->list_lock
);
1841 /* This function drops i_lock... */
1842 inode_sleep_on_writeback(inode
);
1843 spin_lock(&wb
->list_lock
);
1845 spin_unlock(&wb
->list_lock
);
1846 blk_finish_plug(&plug
);
1848 return nr_pages
- work
->nr_pages
;
1852 * Return the next wb_writeback_work struct that hasn't been processed yet.
1854 static struct wb_writeback_work
*get_next_work_item(struct bdi_writeback
*wb
)
1856 struct wb_writeback_work
*work
= NULL
;
1858 spin_lock_bh(&wb
->work_lock
);
1859 if (!list_empty(&wb
->work_list
)) {
1860 work
= list_entry(wb
->work_list
.next
,
1861 struct wb_writeback_work
, list
);
1862 list_del_init(&work
->list
);
1864 spin_unlock_bh(&wb
->work_lock
);
1868 static long wb_check_background_flush(struct bdi_writeback
*wb
)
1870 if (wb_over_bg_thresh(wb
)) {
1872 struct wb_writeback_work work
= {
1873 .nr_pages
= LONG_MAX
,
1874 .sync_mode
= WB_SYNC_NONE
,
1875 .for_background
= 1,
1877 .reason
= WB_REASON_BACKGROUND
,
1880 return wb_writeback(wb
, &work
);
1886 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
1888 unsigned long expired
;
1892 * When set to zero, disable periodic writeback
1894 if (!dirty_writeback_interval
)
1897 expired
= wb
->last_old_flush
+
1898 msecs_to_jiffies(dirty_writeback_interval
* 10);
1899 if (time_before(jiffies
, expired
))
1902 wb
->last_old_flush
= jiffies
;
1903 nr_pages
= get_nr_dirty_pages();
1906 struct wb_writeback_work work
= {
1907 .nr_pages
= nr_pages
,
1908 .sync_mode
= WB_SYNC_NONE
,
1911 .reason
= WB_REASON_PERIODIC
,
1914 return wb_writeback(wb
, &work
);
1920 static long wb_check_start_all(struct bdi_writeback
*wb
)
1924 if (!test_bit(WB_start_all
, &wb
->state
))
1927 nr_pages
= get_nr_dirty_pages();
1929 struct wb_writeback_work work
= {
1930 .nr_pages
= wb_split_bdi_pages(wb
, nr_pages
),
1931 .sync_mode
= WB_SYNC_NONE
,
1933 .reason
= wb
->start_all_reason
,
1936 nr_pages
= wb_writeback(wb
, &work
);
1939 clear_bit(WB_start_all
, &wb
->state
);
1945 * Retrieve work items and do the writeback they describe
1947 static long wb_do_writeback(struct bdi_writeback
*wb
)
1949 struct wb_writeback_work
*work
;
1952 set_bit(WB_writeback_running
, &wb
->state
);
1953 while ((work
= get_next_work_item(wb
)) != NULL
) {
1954 trace_writeback_exec(wb
, work
);
1955 wrote
+= wb_writeback(wb
, work
);
1956 finish_writeback_work(wb
, work
);
1960 * Check for a flush-everything request
1962 wrote
+= wb_check_start_all(wb
);
1965 * Check for periodic writeback, kupdated() style
1967 wrote
+= wb_check_old_data_flush(wb
);
1968 wrote
+= wb_check_background_flush(wb
);
1969 clear_bit(WB_writeback_running
, &wb
->state
);
1975 * Handle writeback of dirty data for the device backed by this bdi. Also
1976 * reschedules periodically and does kupdated style flushing.
1978 void wb_workfn(struct work_struct
*work
)
1980 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
1981 struct bdi_writeback
, dwork
);
1984 set_worker_desc("flush-%s", dev_name(wb
->bdi
->dev
));
1985 current
->flags
|= PF_SWAPWRITE
;
1987 if (likely(!current_is_workqueue_rescuer() ||
1988 !test_bit(WB_registered
, &wb
->state
))) {
1990 * The normal path. Keep writing back @wb until its
1991 * work_list is empty. Note that this path is also taken
1992 * if @wb is shutting down even when we're running off the
1993 * rescuer as work_list needs to be drained.
1996 pages_written
= wb_do_writeback(wb
);
1997 trace_writeback_pages_written(pages_written
);
1998 } while (!list_empty(&wb
->work_list
));
2001 * bdi_wq can't get enough workers and we're running off
2002 * the emergency worker. Don't hog it. Hopefully, 1024 is
2003 * enough for efficient IO.
2005 pages_written
= writeback_inodes_wb(wb
, 1024,
2006 WB_REASON_FORKER_THREAD
);
2007 trace_writeback_pages_written(pages_written
);
2010 if (!list_empty(&wb
->work_list
))
2012 else if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
2013 wb_wakeup_delayed(wb
);
2015 current
->flags
&= ~PF_SWAPWRITE
;
2019 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2020 * write back the whole world.
2022 static void __wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2023 enum wb_reason reason
)
2025 struct bdi_writeback
*wb
;
2027 if (!bdi_has_dirty_io(bdi
))
2030 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2031 wb_start_writeback(wb
, reason
);
2034 void wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2035 enum wb_reason reason
)
2038 __wakeup_flusher_threads_bdi(bdi
, reason
);
2043 * Wakeup the flusher threads to start writeback of all currently dirty pages
2045 void wakeup_flusher_threads(enum wb_reason reason
)
2047 struct backing_dev_info
*bdi
;
2050 * If we are expecting writeback progress we must submit plugged IO.
2052 if (blk_needs_flush_plug(current
))
2053 blk_schedule_flush_plug(current
);
2056 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
2057 __wakeup_flusher_threads_bdi(bdi
, reason
);
2062 * Wake up bdi's periodically to make sure dirtytime inodes gets
2063 * written back periodically. We deliberately do *not* check the
2064 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2065 * kernel to be constantly waking up once there are any dirtytime
2066 * inodes on the system. So instead we define a separate delayed work
2067 * function which gets called much more rarely. (By default, only
2068 * once every 12 hours.)
2070 * If there is any other write activity going on in the file system,
2071 * this function won't be necessary. But if the only thing that has
2072 * happened on the file system is a dirtytime inode caused by an atime
2073 * update, we need this infrastructure below to make sure that inode
2074 * eventually gets pushed out to disk.
2076 static void wakeup_dirtytime_writeback(struct work_struct
*w
);
2077 static DECLARE_DELAYED_WORK(dirtytime_work
, wakeup_dirtytime_writeback
);
2079 static void wakeup_dirtytime_writeback(struct work_struct
*w
)
2081 struct backing_dev_info
*bdi
;
2084 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
2085 struct bdi_writeback
*wb
;
2087 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2088 if (!list_empty(&wb
->b_dirty_time
))
2092 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2095 static int __init
start_dirtytime_writeback(void)
2097 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2100 __initcall(start_dirtytime_writeback
);
2102 int dirtytime_interval_handler(struct ctl_table
*table
, int write
,
2103 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
2107 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
2108 if (ret
== 0 && write
)
2109 mod_delayed_work(system_wq
, &dirtytime_work
, 0);
2113 static noinline
void block_dump___mark_inode_dirty(struct inode
*inode
)
2115 if (inode
->i_ino
|| strcmp(inode
->i_sb
->s_id
, "bdev")) {
2116 struct dentry
*dentry
;
2117 const char *name
= "?";
2119 dentry
= d_find_alias(inode
);
2121 spin_lock(&dentry
->d_lock
);
2122 name
= (const char *) dentry
->d_name
.name
;
2125 "%s(%d): dirtied inode %lu (%s) on %s\n",
2126 current
->comm
, task_pid_nr(current
), inode
->i_ino
,
2127 name
, inode
->i_sb
->s_id
);
2129 spin_unlock(&dentry
->d_lock
);
2136 * __mark_inode_dirty - internal function
2138 * @inode: inode to mark
2139 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2141 * Mark an inode as dirty. Callers should use mark_inode_dirty or
2142 * mark_inode_dirty_sync.
2144 * Put the inode on the super block's dirty list.
2146 * CAREFUL! We mark it dirty unconditionally, but move it onto the
2147 * dirty list only if it is hashed or if it refers to a blockdev.
2148 * If it was not hashed, it will never be added to the dirty list
2149 * even if it is later hashed, as it will have been marked dirty already.
2151 * In short, make sure you hash any inodes _before_ you start marking
2154 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2155 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2156 * the kernel-internal blockdev inode represents the dirtying time of the
2157 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2158 * page->mapping->host, so the page-dirtying time is recorded in the internal
2161 void __mark_inode_dirty(struct inode
*inode
, int flags
)
2163 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
2164 struct super_block
*sb
= inode
->i_sb
;
2167 trace_writeback_mark_inode_dirty(inode
, flags
);
2170 * Don't do this for I_DIRTY_PAGES - that doesn't actually
2171 * dirty the inode itself
2173 if (flags
& (I_DIRTY_SYNC
| I_DIRTY_DATASYNC
| I_DIRTY_TIME
)) {
2174 trace_writeback_dirty_inode_start(inode
, flags
);
2176 if (sb
->s_op
->dirty_inode
)
2177 sb
->s_op
->dirty_inode(inode
, flags
);
2179 trace_writeback_dirty_inode(inode
, flags
);
2181 if (flags
& I_DIRTY_INODE
)
2182 flags
&= ~I_DIRTY_TIME
;
2183 dirtytime
= flags
& I_DIRTY_TIME
;
2186 * Paired with smp_mb() in __writeback_single_inode() for the
2187 * following lockless i_state test. See there for details.
2191 if (((inode
->i_state
& flags
) == flags
) ||
2192 (dirtytime
&& (inode
->i_state
& I_DIRTY_INODE
)))
2195 if (unlikely(block_dump
))
2196 block_dump___mark_inode_dirty(inode
);
2198 spin_lock(&inode
->i_lock
);
2199 if (dirtytime
&& (inode
->i_state
& I_DIRTY_INODE
))
2200 goto out_unlock_inode
;
2201 if ((inode
->i_state
& flags
) != flags
) {
2202 const int was_dirty
= inode
->i_state
& I_DIRTY
;
2204 inode_attach_wb(inode
, NULL
);
2206 if (flags
& I_DIRTY_INODE
)
2207 inode
->i_state
&= ~I_DIRTY_TIME
;
2208 inode
->i_state
|= flags
;
2211 * If the inode is being synced, just update its dirty state.
2212 * The unlocker will place the inode on the appropriate
2213 * superblock list, based upon its state.
2215 if (inode
->i_state
& I_SYNC
)
2216 goto out_unlock_inode
;
2219 * Only add valid (hashed) inodes to the superblock's
2220 * dirty list. Add blockdev inodes as well.
2222 if (!S_ISBLK(inode
->i_mode
)) {
2223 if (inode_unhashed(inode
))
2224 goto out_unlock_inode
;
2226 if (inode
->i_state
& I_FREEING
)
2227 goto out_unlock_inode
;
2230 * If the inode was already on b_dirty/b_io/b_more_io, don't
2231 * reposition it (that would break b_dirty time-ordering).
2234 struct bdi_writeback
*wb
;
2235 struct list_head
*dirty_list
;
2236 bool wakeup_bdi
= false;
2238 wb
= locked_inode_to_wb_and_lock_list(inode
);
2240 WARN(bdi_cap_writeback_dirty(wb
->bdi
) &&
2241 !test_bit(WB_registered
, &wb
->state
),
2242 "bdi-%s not registered\n", wb
->bdi
->name
);
2244 inode
->dirtied_when
= jiffies
;
2246 inode
->dirtied_time_when
= jiffies
;
2248 if (inode
->i_state
& (I_DIRTY_INODE
| I_DIRTY_PAGES
))
2249 dirty_list
= &wb
->b_dirty
;
2251 dirty_list
= &wb
->b_dirty_time
;
2253 wakeup_bdi
= inode_io_list_move_locked(inode
, wb
,
2256 spin_unlock(&wb
->list_lock
);
2257 trace_writeback_dirty_inode_enqueue(inode
);
2260 * If this is the first dirty inode for this bdi,
2261 * we have to wake-up the corresponding bdi thread
2262 * to make sure background write-back happens
2265 if (bdi_cap_writeback_dirty(wb
->bdi
) && wakeup_bdi
)
2266 wb_wakeup_delayed(wb
);
2271 spin_unlock(&inode
->i_lock
);
2273 #undef I_DIRTY_INODE
2275 EXPORT_SYMBOL(__mark_inode_dirty
);
2278 * The @s_sync_lock is used to serialise concurrent sync operations
2279 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2280 * Concurrent callers will block on the s_sync_lock rather than doing contending
2281 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2282 * has been issued up to the time this function is enter is guaranteed to be
2283 * completed by the time we have gained the lock and waited for all IO that is
2284 * in progress regardless of the order callers are granted the lock.
2286 static void wait_sb_inodes(struct super_block
*sb
)
2288 LIST_HEAD(sync_list
);
2291 * We need to be protected against the filesystem going from
2292 * r/o to r/w or vice versa.
2294 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2296 mutex_lock(&sb
->s_sync_lock
);
2299 * Splice the writeback list onto a temporary list to avoid waiting on
2300 * inodes that have started writeback after this point.
2302 * Use rcu_read_lock() to keep the inodes around until we have a
2303 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2304 * the local list because inodes can be dropped from either by writeback
2308 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2309 list_splice_init(&sb
->s_inodes_wb
, &sync_list
);
2312 * Data integrity sync. Must wait for all pages under writeback, because
2313 * there may have been pages dirtied before our sync call, but which had
2314 * writeout started before we write it out. In which case, the inode
2315 * may not be on the dirty list, but we still have to wait for that
2318 while (!list_empty(&sync_list
)) {
2319 struct inode
*inode
= list_first_entry(&sync_list
, struct inode
,
2321 struct address_space
*mapping
= inode
->i_mapping
;
2324 * Move each inode back to the wb list before we drop the lock
2325 * to preserve consistency between i_wb_list and the mapping
2326 * writeback tag. Writeback completion is responsible to remove
2327 * the inode from either list once the writeback tag is cleared.
2329 list_move_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
2332 * The mapping can appear untagged while still on-list since we
2333 * do not have the mapping lock. Skip it here, wb completion
2336 if (!mapping_tagged(mapping
, PAGECACHE_TAG_WRITEBACK
))
2339 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2341 spin_lock(&inode
->i_lock
);
2342 if (inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) {
2343 spin_unlock(&inode
->i_lock
);
2345 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2349 spin_unlock(&inode
->i_lock
);
2353 * We keep the error status of individual mapping so that
2354 * applications can catch the writeback error using fsync(2).
2355 * See filemap_fdatawait_keep_errors() for details.
2357 filemap_fdatawait_keep_errors(mapping
);
2364 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2366 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2368 mutex_unlock(&sb
->s_sync_lock
);
2371 static void __writeback_inodes_sb_nr(struct super_block
*sb
, unsigned long nr
,
2372 enum wb_reason reason
, bool skip_if_busy
)
2374 DEFINE_WB_COMPLETION_ONSTACK(done
);
2375 struct wb_writeback_work work
= {
2377 .sync_mode
= WB_SYNC_NONE
,
2378 .tagged_writepages
= 1,
2383 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2385 if (!bdi_has_dirty_io(bdi
) || bdi
== &noop_backing_dev_info
)
2387 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2389 bdi_split_work_to_wbs(sb
->s_bdi
, &work
, skip_if_busy
);
2390 wb_wait_for_completion(bdi
, &done
);
2394 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2395 * @sb: the superblock
2396 * @nr: the number of pages to write
2397 * @reason: reason why some writeback work initiated
2399 * Start writeback on some inodes on this super_block. No guarantees are made
2400 * on how many (if any) will be written, and this function does not wait
2401 * for IO completion of submitted IO.
2403 void writeback_inodes_sb_nr(struct super_block
*sb
,
2405 enum wb_reason reason
)
2407 __writeback_inodes_sb_nr(sb
, nr
, reason
, false);
2409 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
2412 * writeback_inodes_sb - writeback dirty inodes from given super_block
2413 * @sb: the superblock
2414 * @reason: reason why some writeback work was initiated
2416 * Start writeback on some inodes on this super_block. No guarantees are made
2417 * on how many (if any) will be written, and this function does not wait
2418 * for IO completion of submitted IO.
2420 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2422 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
2424 EXPORT_SYMBOL(writeback_inodes_sb
);
2427 * try_to_writeback_inodes_sb - try to start writeback if none underway
2428 * @sb: the superblock
2429 * @reason: reason why some writeback work was initiated
2431 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2433 void try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2435 if (!down_read_trylock(&sb
->s_umount
))
2438 __writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
, true);
2439 up_read(&sb
->s_umount
);
2441 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
2444 * sync_inodes_sb - sync sb inode pages
2445 * @sb: the superblock
2447 * This function writes and waits on any dirty inode belonging to this
2450 void sync_inodes_sb(struct super_block
*sb
)
2452 DEFINE_WB_COMPLETION_ONSTACK(done
);
2453 struct wb_writeback_work work
= {
2455 .sync_mode
= WB_SYNC_ALL
,
2456 .nr_pages
= LONG_MAX
,
2459 .reason
= WB_REASON_SYNC
,
2462 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2465 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2466 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2467 * bdi_has_dirty() need to be written out too.
2469 if (bdi
== &noop_backing_dev_info
)
2471 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2473 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2474 bdi_down_write_wb_switch_rwsem(bdi
);
2475 bdi_split_work_to_wbs(bdi
, &work
, false);
2476 wb_wait_for_completion(bdi
, &done
);
2477 bdi_up_write_wb_switch_rwsem(bdi
);
2481 EXPORT_SYMBOL(sync_inodes_sb
);
2484 * write_inode_now - write an inode to disk
2485 * @inode: inode to write to disk
2486 * @sync: whether the write should be synchronous or not
2488 * This function commits an inode to disk immediately if it is dirty. This is
2489 * primarily needed by knfsd.
2491 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2493 int write_inode_now(struct inode
*inode
, int sync
)
2495 struct writeback_control wbc
= {
2496 .nr_to_write
= LONG_MAX
,
2497 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2499 .range_end
= LLONG_MAX
,
2502 if (!mapping_cap_writeback_dirty(inode
->i_mapping
))
2503 wbc
.nr_to_write
= 0;
2506 return writeback_single_inode(inode
, &wbc
);
2508 EXPORT_SYMBOL(write_inode_now
);
2511 * sync_inode - write an inode and its pages to disk.
2512 * @inode: the inode to sync
2513 * @wbc: controls the writeback mode
2515 * sync_inode() will write an inode and its pages to disk. It will also
2516 * correctly update the inode on its superblock's dirty inode lists and will
2517 * update inode->i_state.
2519 * The caller must have a ref on the inode.
2521 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2523 return writeback_single_inode(inode
, wbc
);
2525 EXPORT_SYMBOL(sync_inode
);
2528 * sync_inode_metadata - write an inode to disk
2529 * @inode: the inode to sync
2530 * @wait: wait for I/O to complete.
2532 * Write an inode to disk and adjust its dirty state after completion.
2534 * Note: only writes the actual inode, no associated data or other metadata.
2536 int sync_inode_metadata(struct inode
*inode
, int wait
)
2538 struct writeback_control wbc
= {
2539 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2540 .nr_to_write
= 0, /* metadata-only */
2543 return sync_inode(inode
, &wbc
);
2545 EXPORT_SYMBOL(sync_inode_metadata
);