1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2002, Linus Torvalds.
7 * Contains all the functions related to writing back and waiting
8 * upon dirty inodes against superblocks, and writing back dirty
9 * pages against inodes. ie: data writeback. Writeout of the
10 * inode itself is not handled here.
12 * 10Apr2002 Andrew Morton
13 * Split out of fs/inode.c
14 * Additions for address_space-based writeback
17 #include <linux/kernel.h>
18 #include <linux/export.h>
19 #include <linux/spinlock.h>
20 #include <linux/slab.h>
21 #include <linux/sched.h>
24 #include <linux/pagemap.h>
25 #include <linux/kthread.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include <linux/device.h>
31 #include <linux/memcontrol.h>
35 * 4MB minimal write chunk size
37 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
40 * Passed into wb_writeback(), essentially a subset of writeback_control
42 struct wb_writeback_work
{
44 struct super_block
*sb
;
45 enum writeback_sync_modes sync_mode
;
46 unsigned int tagged_writepages
:1;
47 unsigned int for_kupdate
:1;
48 unsigned int range_cyclic
:1;
49 unsigned int for_background
:1;
50 unsigned int for_sync
:1; /* sync(2) WB_SYNC_ALL writeback */
51 unsigned int auto_free
:1; /* free on completion */
52 enum wb_reason reason
; /* why was writeback initiated? */
54 struct list_head list
; /* pending work list */
55 struct wb_completion
*done
; /* set if the caller waits */
59 * If an inode is constantly having its pages dirtied, but then the
60 * updates stop dirtytime_expire_interval seconds in the past, it's
61 * possible for the worst case time between when an inode has its
62 * timestamps updated and when they finally get written out to be two
63 * dirtytime_expire_intervals. We set the default to 12 hours (in
64 * seconds), which means most of the time inodes will have their
65 * timestamps written to disk after 12 hours, but in the worst case a
66 * few inodes might not their timestamps updated for 24 hours.
68 unsigned int dirtytime_expire_interval
= 12 * 60 * 60;
70 static inline struct inode
*wb_inode(struct list_head
*head
)
72 return list_entry(head
, struct inode
, i_io_list
);
76 * Include the creation of the trace points after defining the
77 * wb_writeback_work structure and inline functions so that the definition
78 * remains local to this file.
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/writeback.h>
83 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage
);
85 static bool wb_io_lists_populated(struct bdi_writeback
*wb
)
87 if (wb_has_dirty_io(wb
)) {
90 set_bit(WB_has_dirty_io
, &wb
->state
);
91 WARN_ON_ONCE(!wb
->avg_write_bandwidth
);
92 atomic_long_add(wb
->avg_write_bandwidth
,
93 &wb
->bdi
->tot_write_bandwidth
);
98 static void wb_io_lists_depopulated(struct bdi_writeback
*wb
)
100 if (wb_has_dirty_io(wb
) && list_empty(&wb
->b_dirty
) &&
101 list_empty(&wb
->b_io
) && list_empty(&wb
->b_more_io
)) {
102 clear_bit(WB_has_dirty_io
, &wb
->state
);
103 WARN_ON_ONCE(atomic_long_sub_return(wb
->avg_write_bandwidth
,
104 &wb
->bdi
->tot_write_bandwidth
) < 0);
109 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
110 * @inode: inode to be moved
111 * @wb: target bdi_writeback
112 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
114 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
115 * Returns %true if @inode is the first occupant of the !dirty_time IO
116 * lists; otherwise, %false.
118 static bool inode_io_list_move_locked(struct inode
*inode
,
119 struct bdi_writeback
*wb
,
120 struct list_head
*head
)
122 assert_spin_locked(&wb
->list_lock
);
123 assert_spin_locked(&inode
->i_lock
);
125 list_move(&inode
->i_io_list
, head
);
127 /* dirty_time doesn't count as dirty_io until expiration */
128 if (head
!= &wb
->b_dirty_time
)
129 return wb_io_lists_populated(wb
);
131 wb_io_lists_depopulated(wb
);
135 static void wb_wakeup(struct bdi_writeback
*wb
)
137 spin_lock_irq(&wb
->work_lock
);
138 if (test_bit(WB_registered
, &wb
->state
))
139 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
140 spin_unlock_irq(&wb
->work_lock
);
143 static void finish_writeback_work(struct bdi_writeback
*wb
,
144 struct wb_writeback_work
*work
)
146 struct wb_completion
*done
= work
->done
;
151 wait_queue_head_t
*waitq
= done
->waitq
;
153 /* @done can't be accessed after the following dec */
154 if (atomic_dec_and_test(&done
->cnt
))
159 static void wb_queue_work(struct bdi_writeback
*wb
,
160 struct wb_writeback_work
*work
)
162 trace_writeback_queue(wb
, work
);
165 atomic_inc(&work
->done
->cnt
);
167 spin_lock_irq(&wb
->work_lock
);
169 if (test_bit(WB_registered
, &wb
->state
)) {
170 list_add_tail(&work
->list
, &wb
->work_list
);
171 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
173 finish_writeback_work(wb
, work
);
175 spin_unlock_irq(&wb
->work_lock
);
179 * wb_wait_for_completion - wait for completion of bdi_writeback_works
180 * @done: target wb_completion
182 * Wait for one or more work items issued to @bdi with their ->done field
183 * set to @done, which should have been initialized with
184 * DEFINE_WB_COMPLETION(). This function returns after all such work items
185 * are completed. Work items which are waited upon aren't freed
186 * automatically on completion.
188 void wb_wait_for_completion(struct wb_completion
*done
)
190 atomic_dec(&done
->cnt
); /* put down the initial count */
191 wait_event(*done
->waitq
, !atomic_read(&done
->cnt
));
194 #ifdef CONFIG_CGROUP_WRITEBACK
197 * Parameters for foreign inode detection, see wbc_detach_inode() to see
200 * These paramters are inherently heuristical as the detection target
201 * itself is fuzzy. All we want to do is detaching an inode from the
202 * current owner if it's being written to by some other cgroups too much.
204 * The current cgroup writeback is built on the assumption that multiple
205 * cgroups writing to the same inode concurrently is very rare and a mode
206 * of operation which isn't well supported. As such, the goal is not
207 * taking too long when a different cgroup takes over an inode while
208 * avoiding too aggressive flip-flops from occasional foreign writes.
210 * We record, very roughly, 2s worth of IO time history and if more than
211 * half of that is foreign, trigger the switch. The recording is quantized
212 * to 16 slots. To avoid tiny writes from swinging the decision too much,
213 * writes smaller than 1/8 of avg size are ignored.
215 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
216 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
217 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
218 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
220 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
221 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
222 /* each slot's duration is 2s / 16 */
223 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
224 /* if foreign slots >= 8, switch */
225 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
226 /* one round can affect upto 5 slots */
227 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
230 * Maximum inodes per isw. A specific value has been chosen to make
231 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
233 #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
234 / sizeof(struct inode *))
236 static atomic_t isw_nr_in_flight
= ATOMIC_INIT(0);
237 static struct workqueue_struct
*isw_wq
;
239 void __inode_attach_wb(struct inode
*inode
, struct page
*page
)
241 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
242 struct bdi_writeback
*wb
= NULL
;
244 if (inode_cgwb_enabled(inode
)) {
245 struct cgroup_subsys_state
*memcg_css
;
248 memcg_css
= mem_cgroup_css_from_page(page
);
249 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
251 /* must pin memcg_css, see wb_get_create() */
252 memcg_css
= task_get_css(current
, memory_cgrp_id
);
253 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
262 * There may be multiple instances of this function racing to
263 * update the same inode. Use cmpxchg() to tell the winner.
265 if (unlikely(cmpxchg(&inode
->i_wb
, NULL
, wb
)))
268 EXPORT_SYMBOL_GPL(__inode_attach_wb
);
271 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
272 * @inode: inode of interest with i_lock held
273 * @wb: target bdi_writeback
275 * Remove the inode from wb's io lists and if necessarily put onto b_attached
276 * list. Only inodes attached to cgwb's are kept on this list.
278 static void inode_cgwb_move_to_attached(struct inode
*inode
,
279 struct bdi_writeback
*wb
)
281 assert_spin_locked(&wb
->list_lock
);
282 assert_spin_locked(&inode
->i_lock
);
284 inode
->i_state
&= ~I_SYNC_QUEUED
;
285 if (wb
!= &wb
->bdi
->wb
)
286 list_move(&inode
->i_io_list
, &wb
->b_attached
);
288 list_del_init(&inode
->i_io_list
);
289 wb_io_lists_depopulated(wb
);
293 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
294 * @inode: inode of interest with i_lock held
296 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
297 * held on entry and is released on return. The returned wb is guaranteed
298 * to stay @inode's associated wb until its list_lock is released.
300 static struct bdi_writeback
*
301 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
302 __releases(&inode
->i_lock
)
303 __acquires(&wb
->list_lock
)
306 struct bdi_writeback
*wb
= inode_to_wb(inode
);
309 * inode_to_wb() association is protected by both
310 * @inode->i_lock and @wb->list_lock but list_lock nests
311 * outside i_lock. Drop i_lock and verify that the
312 * association hasn't changed after acquiring list_lock.
315 spin_unlock(&inode
->i_lock
);
316 spin_lock(&wb
->list_lock
);
318 /* i_wb may have changed inbetween, can't use inode_to_wb() */
319 if (likely(wb
== inode
->i_wb
)) {
320 wb_put(wb
); /* @inode already has ref */
324 spin_unlock(&wb
->list_lock
);
327 spin_lock(&inode
->i_lock
);
332 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
333 * @inode: inode of interest
335 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
338 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
339 __acquires(&wb
->list_lock
)
341 spin_lock(&inode
->i_lock
);
342 return locked_inode_to_wb_and_lock_list(inode
);
345 struct inode_switch_wbs_context
{
346 struct rcu_work work
;
349 * Multiple inodes can be switched at once. The switching procedure
350 * consists of two parts, separated by a RCU grace period. To make
351 * sure that the second part is executed for each inode gone through
352 * the first part, all inode pointers are placed into a NULL-terminated
353 * array embedded into struct inode_switch_wbs_context. Otherwise
354 * an inode could be left in a non-consistent state.
356 struct bdi_writeback
*new_wb
;
357 struct inode
*inodes
[];
360 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
362 down_write(&bdi
->wb_switch_rwsem
);
365 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
367 up_write(&bdi
->wb_switch_rwsem
);
370 static bool inode_do_switch_wbs(struct inode
*inode
,
371 struct bdi_writeback
*old_wb
,
372 struct bdi_writeback
*new_wb
)
374 struct address_space
*mapping
= inode
->i_mapping
;
375 XA_STATE(xas
, &mapping
->i_pages
, 0);
377 bool switched
= false;
379 spin_lock(&inode
->i_lock
);
380 xa_lock_irq(&mapping
->i_pages
);
383 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
384 * path owns the inode and we shouldn't modify ->i_io_list.
386 if (unlikely(inode
->i_state
& (I_FREEING
| I_WILL_FREE
)))
389 trace_inode_switch_wbs(inode
, old_wb
, new_wb
);
392 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
393 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
394 * pages actually under writeback.
396 xas_for_each_marked(&xas
, page
, ULONG_MAX
, PAGECACHE_TAG_DIRTY
) {
397 if (PageDirty(page
)) {
398 dec_wb_stat(old_wb
, WB_RECLAIMABLE
);
399 inc_wb_stat(new_wb
, WB_RECLAIMABLE
);
404 xas_for_each_marked(&xas
, page
, ULONG_MAX
, PAGECACHE_TAG_WRITEBACK
) {
405 WARN_ON_ONCE(!PageWriteback(page
));
406 dec_wb_stat(old_wb
, WB_WRITEBACK
);
407 inc_wb_stat(new_wb
, WB_WRITEBACK
);
410 if (mapping_tagged(mapping
, PAGECACHE_TAG_WRITEBACK
)) {
411 atomic_dec(&old_wb
->writeback_inodes
);
412 atomic_inc(&new_wb
->writeback_inodes
);
418 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
419 * the specific list @inode was on is ignored and the @inode is put on
420 * ->b_dirty which is always correct including from ->b_dirty_time.
421 * The transfer preserves @inode->dirtied_when ordering. If the @inode
422 * was clean, it means it was on the b_attached list, so move it onto
423 * the b_attached list of @new_wb.
425 if (!list_empty(&inode
->i_io_list
)) {
426 inode
->i_wb
= new_wb
;
428 if (inode
->i_state
& I_DIRTY_ALL
) {
431 list_for_each_entry(pos
, &new_wb
->b_dirty
, i_io_list
)
432 if (time_after_eq(inode
->dirtied_when
,
435 inode_io_list_move_locked(inode
, new_wb
,
436 pos
->i_io_list
.prev
);
438 inode_cgwb_move_to_attached(inode
, new_wb
);
441 inode
->i_wb
= new_wb
;
444 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
445 inode
->i_wb_frn_winner
= 0;
446 inode
->i_wb_frn_avg_time
= 0;
447 inode
->i_wb_frn_history
= 0;
451 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
452 * ensures that the new wb is visible if they see !I_WB_SWITCH.
454 smp_store_release(&inode
->i_state
, inode
->i_state
& ~I_WB_SWITCH
);
456 xa_unlock_irq(&mapping
->i_pages
);
457 spin_unlock(&inode
->i_lock
);
462 static void inode_switch_wbs_work_fn(struct work_struct
*work
)
464 struct inode_switch_wbs_context
*isw
=
465 container_of(to_rcu_work(work
), struct inode_switch_wbs_context
, work
);
466 struct backing_dev_info
*bdi
= inode_to_bdi(isw
->inodes
[0]);
467 struct bdi_writeback
*old_wb
= isw
->inodes
[0]->i_wb
;
468 struct bdi_writeback
*new_wb
= isw
->new_wb
;
469 unsigned long nr_switched
= 0;
470 struct inode
**inodep
;
473 * If @inode switches cgwb membership while sync_inodes_sb() is
474 * being issued, sync_inodes_sb() might miss it. Synchronize.
476 down_read(&bdi
->wb_switch_rwsem
);
479 * By the time control reaches here, RCU grace period has passed
480 * since I_WB_SWITCH assertion and all wb stat update transactions
481 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
482 * synchronizing against the i_pages lock.
484 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
485 * gives us exclusion against all wb related operations on @inode
486 * including IO list manipulations and stat updates.
488 if (old_wb
< new_wb
) {
489 spin_lock(&old_wb
->list_lock
);
490 spin_lock_nested(&new_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
492 spin_lock(&new_wb
->list_lock
);
493 spin_lock_nested(&old_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
496 for (inodep
= isw
->inodes
; *inodep
; inodep
++) {
497 WARN_ON_ONCE((*inodep
)->i_wb
!= old_wb
);
498 if (inode_do_switch_wbs(*inodep
, old_wb
, new_wb
))
502 spin_unlock(&new_wb
->list_lock
);
503 spin_unlock(&old_wb
->list_lock
);
505 up_read(&bdi
->wb_switch_rwsem
);
509 wb_put_many(old_wb
, nr_switched
);
512 for (inodep
= isw
->inodes
; *inodep
; inodep
++)
516 atomic_dec(&isw_nr_in_flight
);
519 static bool inode_prepare_wbs_switch(struct inode
*inode
,
520 struct bdi_writeback
*new_wb
)
523 * Paired with smp_mb() in cgroup_writeback_umount().
524 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
525 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
526 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
533 /* while holding I_WB_SWITCH, no one else can update the association */
534 spin_lock(&inode
->i_lock
);
535 if (!(inode
->i_sb
->s_flags
& SB_ACTIVE
) ||
536 inode
->i_state
& (I_WB_SWITCH
| I_FREEING
| I_WILL_FREE
) ||
537 inode_to_wb(inode
) == new_wb
) {
538 spin_unlock(&inode
->i_lock
);
541 inode
->i_state
|= I_WB_SWITCH
;
543 spin_unlock(&inode
->i_lock
);
549 * inode_switch_wbs - change the wb association of an inode
550 * @inode: target inode
551 * @new_wb_id: ID of the new wb
553 * Switch @inode's wb association to the wb identified by @new_wb_id. The
554 * switching is performed asynchronously and may fail silently.
556 static void inode_switch_wbs(struct inode
*inode
, int new_wb_id
)
558 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
559 struct cgroup_subsys_state
*memcg_css
;
560 struct inode_switch_wbs_context
*isw
;
562 /* noop if seems to be already in progress */
563 if (inode
->i_state
& I_WB_SWITCH
)
566 /* avoid queueing a new switch if too many are already in flight */
567 if (atomic_read(&isw_nr_in_flight
) > WB_FRN_MAX_IN_FLIGHT
)
570 isw
= kzalloc(sizeof(*isw
) + 2 * sizeof(struct inode
*), GFP_ATOMIC
);
574 atomic_inc(&isw_nr_in_flight
);
576 /* find and pin the new wb */
578 memcg_css
= css_from_id(new_wb_id
, &memory_cgrp_subsys
);
579 if (memcg_css
&& !css_tryget(memcg_css
))
585 isw
->new_wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
590 if (!inode_prepare_wbs_switch(inode
, isw
->new_wb
))
593 isw
->inodes
[0] = inode
;
596 * In addition to synchronizing among switchers, I_WB_SWITCH tells
597 * the RCU protected stat update paths to grab the i_page
598 * lock so that stat transfer can synchronize against them.
599 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
601 INIT_RCU_WORK(&isw
->work
, inode_switch_wbs_work_fn
);
602 queue_rcu_work(isw_wq
, &isw
->work
);
606 atomic_dec(&isw_nr_in_flight
);
613 * cleanup_offline_cgwb - detach associated inodes
616 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
617 * to eventually release the dying @wb. Returns %true if not all inodes were
618 * switched and the function has to be restarted.
620 bool cleanup_offline_cgwb(struct bdi_writeback
*wb
)
622 struct cgroup_subsys_state
*memcg_css
;
623 struct inode_switch_wbs_context
*isw
;
626 bool restart
= false;
628 isw
= kzalloc(sizeof(*isw
) + WB_MAX_INODES_PER_ISW
*
629 sizeof(struct inode
*), GFP_KERNEL
);
633 atomic_inc(&isw_nr_in_flight
);
635 for (memcg_css
= wb
->memcg_css
->parent
; memcg_css
;
636 memcg_css
= memcg_css
->parent
) {
637 isw
->new_wb
= wb_get_create(wb
->bdi
, memcg_css
, GFP_KERNEL
);
641 if (unlikely(!isw
->new_wb
))
642 isw
->new_wb
= &wb
->bdi
->wb
; /* wb_get() is noop for bdi's wb */
645 spin_lock(&wb
->list_lock
);
646 list_for_each_entry(inode
, &wb
->b_attached
, i_io_list
) {
647 if (!inode_prepare_wbs_switch(inode
, isw
->new_wb
))
650 isw
->inodes
[nr
++] = inode
;
652 if (nr
>= WB_MAX_INODES_PER_ISW
- 1) {
657 spin_unlock(&wb
->list_lock
);
659 /* no attached inodes? bail out */
661 atomic_dec(&isw_nr_in_flight
);
668 * In addition to synchronizing among switchers, I_WB_SWITCH tells
669 * the RCU protected stat update paths to grab the i_page
670 * lock so that stat transfer can synchronize against them.
671 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
673 INIT_RCU_WORK(&isw
->work
, inode_switch_wbs_work_fn
);
674 queue_rcu_work(isw_wq
, &isw
->work
);
680 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
681 * @wbc: writeback_control of interest
682 * @inode: target inode
684 * @inode is locked and about to be written back under the control of @wbc.
685 * Record @inode's writeback context into @wbc and unlock the i_lock. On
686 * writeback completion, wbc_detach_inode() should be called. This is used
687 * to track the cgroup writeback context.
689 void wbc_attach_and_unlock_inode(struct writeback_control
*wbc
,
692 if (!inode_cgwb_enabled(inode
)) {
693 spin_unlock(&inode
->i_lock
);
697 wbc
->wb
= inode_to_wb(inode
);
700 wbc
->wb_id
= wbc
->wb
->memcg_css
->id
;
701 wbc
->wb_lcand_id
= inode
->i_wb_frn_winner
;
702 wbc
->wb_tcand_id
= 0;
704 wbc
->wb_lcand_bytes
= 0;
705 wbc
->wb_tcand_bytes
= 0;
708 spin_unlock(&inode
->i_lock
);
711 * A dying wb indicates that either the blkcg associated with the
712 * memcg changed or the associated memcg is dying. In the first
713 * case, a replacement wb should already be available and we should
714 * refresh the wb immediately. In the second case, trying to
715 * refresh will keep failing.
717 if (unlikely(wb_dying(wbc
->wb
) && !css_is_dying(wbc
->wb
->memcg_css
)))
718 inode_switch_wbs(inode
, wbc
->wb_id
);
720 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode
);
723 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
724 * @wbc: writeback_control of the just finished writeback
726 * To be called after a writeback attempt of an inode finishes and undoes
727 * wbc_attach_and_unlock_inode(). Can be called under any context.
729 * As concurrent write sharing of an inode is expected to be very rare and
730 * memcg only tracks page ownership on first-use basis severely confining
731 * the usefulness of such sharing, cgroup writeback tracks ownership
732 * per-inode. While the support for concurrent write sharing of an inode
733 * is deemed unnecessary, an inode being written to by different cgroups at
734 * different points in time is a lot more common, and, more importantly,
735 * charging only by first-use can too readily lead to grossly incorrect
736 * behaviors (single foreign page can lead to gigabytes of writeback to be
737 * incorrectly attributed).
739 * To resolve this issue, cgroup writeback detects the majority dirtier of
740 * an inode and transfers the ownership to it. To avoid unnnecessary
741 * oscillation, the detection mechanism keeps track of history and gives
742 * out the switch verdict only if the foreign usage pattern is stable over
743 * a certain amount of time and/or writeback attempts.
745 * On each writeback attempt, @wbc tries to detect the majority writer
746 * using Boyer-Moore majority vote algorithm. In addition to the byte
747 * count from the majority voting, it also counts the bytes written for the
748 * current wb and the last round's winner wb (max of last round's current
749 * wb, the winner from two rounds ago, and the last round's majority
750 * candidate). Keeping track of the historical winner helps the algorithm
751 * to semi-reliably detect the most active writer even when it's not the
754 * Once the winner of the round is determined, whether the winner is
755 * foreign or not and how much IO time the round consumed is recorded in
756 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
757 * over a certain threshold, the switch verdict is given.
759 void wbc_detach_inode(struct writeback_control
*wbc
)
761 struct bdi_writeback
*wb
= wbc
->wb
;
762 struct inode
*inode
= wbc
->inode
;
763 unsigned long avg_time
, max_bytes
, max_time
;
770 history
= inode
->i_wb_frn_history
;
771 avg_time
= inode
->i_wb_frn_avg_time
;
773 /* pick the winner of this round */
774 if (wbc
->wb_bytes
>= wbc
->wb_lcand_bytes
&&
775 wbc
->wb_bytes
>= wbc
->wb_tcand_bytes
) {
777 max_bytes
= wbc
->wb_bytes
;
778 } else if (wbc
->wb_lcand_bytes
>= wbc
->wb_tcand_bytes
) {
779 max_id
= wbc
->wb_lcand_id
;
780 max_bytes
= wbc
->wb_lcand_bytes
;
782 max_id
= wbc
->wb_tcand_id
;
783 max_bytes
= wbc
->wb_tcand_bytes
;
787 * Calculate the amount of IO time the winner consumed and fold it
788 * into the running average kept per inode. If the consumed IO
789 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
790 * deciding whether to switch or not. This is to prevent one-off
791 * small dirtiers from skewing the verdict.
793 max_time
= DIV_ROUND_UP((max_bytes
>> PAGE_SHIFT
) << WB_FRN_TIME_SHIFT
,
794 wb
->avg_write_bandwidth
);
796 avg_time
+= (max_time
>> WB_FRN_TIME_AVG_SHIFT
) -
797 (avg_time
>> WB_FRN_TIME_AVG_SHIFT
);
799 avg_time
= max_time
; /* immediate catch up on first run */
801 if (max_time
>= avg_time
/ WB_FRN_TIME_CUT_DIV
) {
805 * The switch verdict is reached if foreign wb's consume
806 * more than a certain proportion of IO time in a
807 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
808 * history mask where each bit represents one sixteenth of
809 * the period. Determine the number of slots to shift into
810 * history from @max_time.
812 slots
= min(DIV_ROUND_UP(max_time
, WB_FRN_HIST_UNIT
),
813 (unsigned long)WB_FRN_HIST_MAX_SLOTS
);
815 if (wbc
->wb_id
!= max_id
)
816 history
|= (1U << slots
) - 1;
819 trace_inode_foreign_history(inode
, wbc
, history
);
822 * Switch if the current wb isn't the consistent winner.
823 * If there are multiple closely competing dirtiers, the
824 * inode may switch across them repeatedly over time, which
825 * is okay. The main goal is avoiding keeping an inode on
826 * the wrong wb for an extended period of time.
828 if (hweight32(history
) > WB_FRN_HIST_THR_SLOTS
)
829 inode_switch_wbs(inode
, max_id
);
833 * Multiple instances of this function may race to update the
834 * following fields but we don't mind occassional inaccuracies.
836 inode
->i_wb_frn_winner
= max_id
;
837 inode
->i_wb_frn_avg_time
= min(avg_time
, (unsigned long)U16_MAX
);
838 inode
->i_wb_frn_history
= history
;
843 EXPORT_SYMBOL_GPL(wbc_detach_inode
);
846 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
847 * @wbc: writeback_control of the writeback in progress
848 * @page: page being written out
849 * @bytes: number of bytes being written out
851 * @bytes from @page are about to written out during the writeback
852 * controlled by @wbc. Keep the book for foreign inode detection. See
853 * wbc_detach_inode().
855 void wbc_account_cgroup_owner(struct writeback_control
*wbc
, struct page
*page
,
858 struct cgroup_subsys_state
*css
;
862 * pageout() path doesn't attach @wbc to the inode being written
863 * out. This is intentional as we don't want the function to block
864 * behind a slow cgroup. Ultimately, we want pageout() to kick off
865 * regular writeback instead of writing things out itself.
867 if (!wbc
->wb
|| wbc
->no_cgroup_owner
)
870 css
= mem_cgroup_css_from_page(page
);
871 /* dead cgroups shouldn't contribute to inode ownership arbitration */
872 if (!(css
->flags
& CSS_ONLINE
))
877 if (id
== wbc
->wb_id
) {
878 wbc
->wb_bytes
+= bytes
;
882 if (id
== wbc
->wb_lcand_id
)
883 wbc
->wb_lcand_bytes
+= bytes
;
885 /* Boyer-Moore majority vote algorithm */
886 if (!wbc
->wb_tcand_bytes
)
887 wbc
->wb_tcand_id
= id
;
888 if (id
== wbc
->wb_tcand_id
)
889 wbc
->wb_tcand_bytes
+= bytes
;
891 wbc
->wb_tcand_bytes
-= min(bytes
, wbc
->wb_tcand_bytes
);
893 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner
);
896 * inode_congested - test whether an inode is congested
897 * @inode: inode to test for congestion (may be NULL)
898 * @cong_bits: mask of WB_[a]sync_congested bits to test
900 * Tests whether @inode is congested. @cong_bits is the mask of congestion
901 * bits to test and the return value is the mask of set bits.
903 * If cgroup writeback is enabled for @inode, the congestion state is
904 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
905 * associated with @inode is congested; otherwise, the root wb's congestion
908 * @inode is allowed to be NULL as this function is often called on
909 * mapping->host which is NULL for the swapper space.
911 int inode_congested(struct inode
*inode
, int cong_bits
)
914 * Once set, ->i_wb never becomes NULL while the inode is alive.
915 * Start transaction iff ->i_wb is visible.
917 if (inode
&& inode_to_wb_is_valid(inode
)) {
918 struct bdi_writeback
*wb
;
919 struct wb_lock_cookie lock_cookie
= {};
922 wb
= unlocked_inode_to_wb_begin(inode
, &lock_cookie
);
923 congested
= wb_congested(wb
, cong_bits
);
924 unlocked_inode_to_wb_end(inode
, &lock_cookie
);
928 return wb_congested(&inode_to_bdi(inode
)->wb
, cong_bits
);
930 EXPORT_SYMBOL_GPL(inode_congested
);
933 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
934 * @wb: target bdi_writeback to split @nr_pages to
935 * @nr_pages: number of pages to write for the whole bdi
937 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
938 * relation to the total write bandwidth of all wb's w/ dirty inodes on
941 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
943 unsigned long this_bw
= wb
->avg_write_bandwidth
;
944 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
946 if (nr_pages
== LONG_MAX
)
950 * This may be called on clean wb's and proportional distribution
951 * may not make sense, just use the original @nr_pages in those
952 * cases. In general, we wanna err on the side of writing more.
954 if (!tot_bw
|| this_bw
>= tot_bw
)
957 return DIV_ROUND_UP_ULL((u64
)nr_pages
* this_bw
, tot_bw
);
961 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
962 * @bdi: target backing_dev_info
963 * @base_work: wb_writeback_work to issue
964 * @skip_if_busy: skip wb's which already have writeback in progress
966 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
967 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
968 * distributed to the busy wbs according to each wb's proportion in the
969 * total active write bandwidth of @bdi.
971 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
972 struct wb_writeback_work
*base_work
,
975 struct bdi_writeback
*last_wb
= NULL
;
976 struct bdi_writeback
*wb
= list_entry(&bdi
->wb_list
,
977 struct bdi_writeback
, bdi_node
);
982 list_for_each_entry_continue_rcu(wb
, &bdi
->wb_list
, bdi_node
) {
983 DEFINE_WB_COMPLETION(fallback_work_done
, bdi
);
984 struct wb_writeback_work fallback_work
;
985 struct wb_writeback_work
*work
;
993 /* SYNC_ALL writes out I_DIRTY_TIME too */
994 if (!wb_has_dirty_io(wb
) &&
995 (base_work
->sync_mode
== WB_SYNC_NONE
||
996 list_empty(&wb
->b_dirty_time
)))
998 if (skip_if_busy
&& writeback_in_progress(wb
))
1001 nr_pages
= wb_split_bdi_pages(wb
, base_work
->nr_pages
);
1003 work
= kmalloc(sizeof(*work
), GFP_ATOMIC
);
1006 work
->nr_pages
= nr_pages
;
1007 work
->auto_free
= 1;
1008 wb_queue_work(wb
, work
);
1012 /* alloc failed, execute synchronously using on-stack fallback */
1013 work
= &fallback_work
;
1015 work
->nr_pages
= nr_pages
;
1016 work
->auto_free
= 0;
1017 work
->done
= &fallback_work_done
;
1019 wb_queue_work(wb
, work
);
1022 * Pin @wb so that it stays on @bdi->wb_list. This allows
1023 * continuing iteration from @wb after dropping and
1024 * regrabbing rcu read lock.
1030 wb_wait_for_completion(&fallback_work_done
);
1040 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1041 * @bdi_id: target bdi id
1042 * @memcg_id: target memcg css id
1043 * @reason: reason why some writeback work initiated
1044 * @done: target wb_completion
1046 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1047 * with the specified parameters.
1049 int cgroup_writeback_by_id(u64 bdi_id
, int memcg_id
,
1050 enum wb_reason reason
, struct wb_completion
*done
)
1052 struct backing_dev_info
*bdi
;
1053 struct cgroup_subsys_state
*memcg_css
;
1054 struct bdi_writeback
*wb
;
1055 struct wb_writeback_work
*work
;
1056 unsigned long dirty
;
1059 /* lookup bdi and memcg */
1060 bdi
= bdi_get_by_id(bdi_id
);
1065 memcg_css
= css_from_id(memcg_id
, &memory_cgrp_subsys
);
1066 if (memcg_css
&& !css_tryget(memcg_css
))
1075 * And find the associated wb. If the wb isn't there already
1076 * there's nothing to flush, don't create one.
1078 wb
= wb_get_lookup(bdi
, memcg_css
);
1085 * The caller is attempting to write out most of
1086 * the currently dirty pages. Let's take the current dirty page
1087 * count and inflate it by 25% which should be large enough to
1088 * flush out most dirty pages while avoiding getting livelocked by
1089 * concurrent dirtiers.
1091 * BTW the memcg stats are flushed periodically and this is best-effort
1092 * estimation, so some potential error is ok.
1094 dirty
= memcg_page_state(mem_cgroup_from_css(memcg_css
), NR_FILE_DIRTY
);
1095 dirty
= dirty
* 10 / 8;
1097 /* issue the writeback work */
1098 work
= kzalloc(sizeof(*work
), GFP_NOWAIT
| __GFP_NOWARN
);
1100 work
->nr_pages
= dirty
;
1101 work
->sync_mode
= WB_SYNC_NONE
;
1102 work
->range_cyclic
= 1;
1103 work
->reason
= reason
;
1105 work
->auto_free
= 1;
1106 wb_queue_work(wb
, work
);
1121 * cgroup_writeback_umount - flush inode wb switches for umount
1123 * This function is called when a super_block is about to be destroyed and
1124 * flushes in-flight inode wb switches. An inode wb switch goes through
1125 * RCU and then workqueue, so the two need to be flushed in order to ensure
1126 * that all previously scheduled switches are finished. As wb switches are
1127 * rare occurrences and synchronize_rcu() can take a while, perform
1128 * flushing iff wb switches are in flight.
1130 void cgroup_writeback_umount(void)
1133 * SB_ACTIVE should be reliably cleared before checking
1134 * isw_nr_in_flight, see generic_shutdown_super().
1138 if (atomic_read(&isw_nr_in_flight
)) {
1140 * Use rcu_barrier() to wait for all pending callbacks to
1141 * ensure that all in-flight wb switches are in the workqueue.
1144 flush_workqueue(isw_wq
);
1148 static int __init
cgroup_writeback_init(void)
1150 isw_wq
= alloc_workqueue("inode_switch_wbs", 0, 0);
1155 fs_initcall(cgroup_writeback_init
);
1157 #else /* CONFIG_CGROUP_WRITEBACK */
1159 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
1160 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
1162 static void inode_cgwb_move_to_attached(struct inode
*inode
,
1163 struct bdi_writeback
*wb
)
1165 assert_spin_locked(&wb
->list_lock
);
1166 assert_spin_locked(&inode
->i_lock
);
1168 inode
->i_state
&= ~I_SYNC_QUEUED
;
1169 list_del_init(&inode
->i_io_list
);
1170 wb_io_lists_depopulated(wb
);
1173 static struct bdi_writeback
*
1174 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
1175 __releases(&inode
->i_lock
)
1176 __acquires(&wb
->list_lock
)
1178 struct bdi_writeback
*wb
= inode_to_wb(inode
);
1180 spin_unlock(&inode
->i_lock
);
1181 spin_lock(&wb
->list_lock
);
1185 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
1186 __acquires(&wb
->list_lock
)
1188 struct bdi_writeback
*wb
= inode_to_wb(inode
);
1190 spin_lock(&wb
->list_lock
);
1194 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
1199 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
1200 struct wb_writeback_work
*base_work
,
1205 if (!skip_if_busy
|| !writeback_in_progress(&bdi
->wb
)) {
1206 base_work
->auto_free
= 0;
1207 wb_queue_work(&bdi
->wb
, base_work
);
1211 #endif /* CONFIG_CGROUP_WRITEBACK */
1214 * Add in the number of potentially dirty inodes, because each inode
1215 * write can dirty pagecache in the underlying blockdev.
1217 static unsigned long get_nr_dirty_pages(void)
1219 return global_node_page_state(NR_FILE_DIRTY
) +
1220 get_nr_dirty_inodes();
1223 static void wb_start_writeback(struct bdi_writeback
*wb
, enum wb_reason reason
)
1225 if (!wb_has_dirty_io(wb
))
1229 * All callers of this function want to start writeback of all
1230 * dirty pages. Places like vmscan can call this at a very
1231 * high frequency, causing pointless allocations of tons of
1232 * work items and keeping the flusher threads busy retrieving
1233 * that work. Ensure that we only allow one of them pending and
1234 * inflight at the time.
1236 if (test_bit(WB_start_all
, &wb
->state
) ||
1237 test_and_set_bit(WB_start_all
, &wb
->state
))
1240 wb
->start_all_reason
= reason
;
1245 * wb_start_background_writeback - start background writeback
1246 * @wb: bdi_writback to write from
1249 * This makes sure WB_SYNC_NONE background writeback happens. When
1250 * this function returns, it is only guaranteed that for given wb
1251 * some IO is happening if we are over background dirty threshold.
1252 * Caller need not hold sb s_umount semaphore.
1254 void wb_start_background_writeback(struct bdi_writeback
*wb
)
1257 * We just wake up the flusher thread. It will perform background
1258 * writeback as soon as there is no other work to do.
1260 trace_writeback_wake_background(wb
);
1265 * Remove the inode from the writeback list it is on.
1267 void inode_io_list_del(struct inode
*inode
)
1269 struct bdi_writeback
*wb
;
1271 wb
= inode_to_wb_and_lock_list(inode
);
1272 spin_lock(&inode
->i_lock
);
1274 inode
->i_state
&= ~I_SYNC_QUEUED
;
1275 list_del_init(&inode
->i_io_list
);
1276 wb_io_lists_depopulated(wb
);
1278 spin_unlock(&inode
->i_lock
);
1279 spin_unlock(&wb
->list_lock
);
1281 EXPORT_SYMBOL(inode_io_list_del
);
1284 * mark an inode as under writeback on the sb
1286 void sb_mark_inode_writeback(struct inode
*inode
)
1288 struct super_block
*sb
= inode
->i_sb
;
1289 unsigned long flags
;
1291 if (list_empty(&inode
->i_wb_list
)) {
1292 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1293 if (list_empty(&inode
->i_wb_list
)) {
1294 list_add_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
1295 trace_sb_mark_inode_writeback(inode
);
1297 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1302 * clear an inode as under writeback on the sb
1304 void sb_clear_inode_writeback(struct inode
*inode
)
1306 struct super_block
*sb
= inode
->i_sb
;
1307 unsigned long flags
;
1309 if (!list_empty(&inode
->i_wb_list
)) {
1310 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1311 if (!list_empty(&inode
->i_wb_list
)) {
1312 list_del_init(&inode
->i_wb_list
);
1313 trace_sb_clear_inode_writeback(inode
);
1315 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1320 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1321 * furthest end of its superblock's dirty-inode list.
1323 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1324 * already the most-recently-dirtied inode on the b_dirty list. If that is
1325 * the case then the inode must have been redirtied while it was being written
1326 * out and we don't reset its dirtied_when.
1328 static void redirty_tail_locked(struct inode
*inode
, struct bdi_writeback
*wb
)
1330 assert_spin_locked(&inode
->i_lock
);
1332 if (!list_empty(&wb
->b_dirty
)) {
1335 tail
= wb_inode(wb
->b_dirty
.next
);
1336 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
1337 inode
->dirtied_when
= jiffies
;
1339 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty
);
1340 inode
->i_state
&= ~I_SYNC_QUEUED
;
1343 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
1345 spin_lock(&inode
->i_lock
);
1346 redirty_tail_locked(inode
, wb
);
1347 spin_unlock(&inode
->i_lock
);
1351 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1353 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
1355 inode_io_list_move_locked(inode
, wb
, &wb
->b_more_io
);
1358 static void inode_sync_complete(struct inode
*inode
)
1360 inode
->i_state
&= ~I_SYNC
;
1361 /* If inode is clean an unused, put it into LRU now... */
1362 inode_add_lru(inode
);
1363 /* Waiters must see I_SYNC cleared before being woken up */
1365 wake_up_bit(&inode
->i_state
, __I_SYNC
);
1368 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
1370 bool ret
= time_after(inode
->dirtied_when
, t
);
1371 #ifndef CONFIG_64BIT
1373 * For inodes being constantly redirtied, dirtied_when can get stuck.
1374 * It _appears_ to be in the future, but is actually in distant past.
1375 * This test is necessary to prevent such wrapped-around relative times
1376 * from permanently stopping the whole bdi writeback.
1378 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
1383 #define EXPIRE_DIRTY_ATIME 0x0001
1386 * Move expired (dirtied before dirtied_before) dirty inodes from
1387 * @delaying_queue to @dispatch_queue.
1389 static int move_expired_inodes(struct list_head
*delaying_queue
,
1390 struct list_head
*dispatch_queue
,
1391 unsigned long dirtied_before
)
1394 struct list_head
*pos
, *node
;
1395 struct super_block
*sb
= NULL
;
1396 struct inode
*inode
;
1400 while (!list_empty(delaying_queue
)) {
1401 inode
= wb_inode(delaying_queue
->prev
);
1402 if (inode_dirtied_after(inode
, dirtied_before
))
1404 spin_lock(&inode
->i_lock
);
1405 list_move(&inode
->i_io_list
, &tmp
);
1407 inode
->i_state
|= I_SYNC_QUEUED
;
1408 spin_unlock(&inode
->i_lock
);
1409 if (sb_is_blkdev_sb(inode
->i_sb
))
1411 if (sb
&& sb
!= inode
->i_sb
)
1416 /* just one sb in list, splice to dispatch_queue and we're done */
1418 list_splice(&tmp
, dispatch_queue
);
1423 * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1424 * we don't take inode->i_lock here because it is just a pointless overhead.
1425 * Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1426 * fully under our control.
1428 while (!list_empty(&tmp
)) {
1429 sb
= wb_inode(tmp
.prev
)->i_sb
;
1430 list_for_each_prev_safe(pos
, node
, &tmp
) {
1431 inode
= wb_inode(pos
);
1432 if (inode
->i_sb
== sb
)
1433 list_move(&inode
->i_io_list
, dispatch_queue
);
1441 * Queue all expired dirty inodes for io, eldest first.
1443 * newly dirtied b_dirty b_io b_more_io
1444 * =============> gf edc BA
1446 * newly dirtied b_dirty b_io b_more_io
1447 * =============> g fBAedc
1449 * +--> dequeue for IO
1451 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
,
1452 unsigned long dirtied_before
)
1455 unsigned long time_expire_jif
= dirtied_before
;
1457 assert_spin_locked(&wb
->list_lock
);
1458 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
1459 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, dirtied_before
);
1460 if (!work
->for_sync
)
1461 time_expire_jif
= jiffies
- dirtytime_expire_interval
* HZ
;
1462 moved
+= move_expired_inodes(&wb
->b_dirty_time
, &wb
->b_io
,
1465 wb_io_lists_populated(wb
);
1466 trace_writeback_queue_io(wb
, work
, dirtied_before
, moved
);
1469 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1473 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
1474 trace_writeback_write_inode_start(inode
, wbc
);
1475 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
1476 trace_writeback_write_inode(inode
, wbc
);
1483 * Wait for writeback on an inode to complete. Called with i_lock held.
1484 * Caller must make sure inode cannot go away when we drop i_lock.
1486 static void __inode_wait_for_writeback(struct inode
*inode
)
1487 __releases(inode
->i_lock
)
1488 __acquires(inode
->i_lock
)
1490 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
1491 wait_queue_head_t
*wqh
;
1493 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1494 while (inode
->i_state
& I_SYNC
) {
1495 spin_unlock(&inode
->i_lock
);
1496 __wait_on_bit(wqh
, &wq
, bit_wait
,
1497 TASK_UNINTERRUPTIBLE
);
1498 spin_lock(&inode
->i_lock
);
1503 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1505 void inode_wait_for_writeback(struct inode
*inode
)
1507 spin_lock(&inode
->i_lock
);
1508 __inode_wait_for_writeback(inode
);
1509 spin_unlock(&inode
->i_lock
);
1513 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1514 * held and drops it. It is aimed for callers not holding any inode reference
1515 * so once i_lock is dropped, inode can go away.
1517 static void inode_sleep_on_writeback(struct inode
*inode
)
1518 __releases(inode
->i_lock
)
1521 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1524 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
1525 sleep
= inode
->i_state
& I_SYNC
;
1526 spin_unlock(&inode
->i_lock
);
1529 finish_wait(wqh
, &wait
);
1533 * Find proper writeback list for the inode depending on its current state and
1534 * possibly also change of its state while we were doing writeback. Here we
1535 * handle things such as livelock prevention or fairness of writeback among
1536 * inodes. This function can be called only by flusher thread - noone else
1537 * processes all inodes in writeback lists and requeueing inodes behind flusher
1538 * thread's back can have unexpected consequences.
1540 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
1541 struct writeback_control
*wbc
)
1543 if (inode
->i_state
& I_FREEING
)
1547 * Sync livelock prevention. Each inode is tagged and synced in one
1548 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1549 * the dirty time to prevent enqueue and sync it again.
1551 if ((inode
->i_state
& I_DIRTY
) &&
1552 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
1553 inode
->dirtied_when
= jiffies
;
1555 if (wbc
->pages_skipped
) {
1557 * writeback is not making progress due to locked
1558 * buffers. Skip this inode for now.
1560 redirty_tail_locked(inode
, wb
);
1564 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
1566 * We didn't write back all the pages. nfs_writepages()
1567 * sometimes bales out without doing anything.
1569 if (wbc
->nr_to_write
<= 0) {
1570 /* Slice used up. Queue for next turn. */
1571 requeue_io(inode
, wb
);
1574 * Writeback blocked by something other than
1575 * congestion. Delay the inode for some time to
1576 * avoid spinning on the CPU (100% iowait)
1577 * retrying writeback of the dirty page/inode
1578 * that cannot be performed immediately.
1580 redirty_tail_locked(inode
, wb
);
1582 } else if (inode
->i_state
& I_DIRTY
) {
1584 * Filesystems can dirty the inode during writeback operations,
1585 * such as delayed allocation during submission or metadata
1586 * updates after data IO completion.
1588 redirty_tail_locked(inode
, wb
);
1589 } else if (inode
->i_state
& I_DIRTY_TIME
) {
1590 inode
->dirtied_when
= jiffies
;
1591 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty_time
);
1592 inode
->i_state
&= ~I_SYNC_QUEUED
;
1594 /* The inode is clean. Remove from writeback lists. */
1595 inode_cgwb_move_to_attached(inode
, wb
);
1600 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1601 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1603 * This doesn't remove the inode from the writeback list it is on, except
1604 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1605 * expiration. The caller is otherwise responsible for writeback list handling.
1607 * The caller is also responsible for setting the I_SYNC flag beforehand and
1608 * calling inode_sync_complete() to clear it afterwards.
1611 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1613 struct address_space
*mapping
= inode
->i_mapping
;
1614 long nr_to_write
= wbc
->nr_to_write
;
1618 WARN_ON(!(inode
->i_state
& I_SYNC
));
1620 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
1622 ret
= do_writepages(mapping
, wbc
);
1625 * Make sure to wait on the data before writing out the metadata.
1626 * This is important for filesystems that modify metadata on data
1627 * I/O completion. We don't do it for sync(2) writeback because it has a
1628 * separate, external IO completion path and ->sync_fs for guaranteeing
1629 * inode metadata is written back correctly.
1631 if (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
) {
1632 int err
= filemap_fdatawait(mapping
);
1638 * If the inode has dirty timestamps and we need to write them, call
1639 * mark_inode_dirty_sync() to notify the filesystem about it and to
1640 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1642 if ((inode
->i_state
& I_DIRTY_TIME
) &&
1643 (wbc
->sync_mode
== WB_SYNC_ALL
||
1644 time_after(jiffies
, inode
->dirtied_time_when
+
1645 dirtytime_expire_interval
* HZ
))) {
1646 trace_writeback_lazytime(inode
);
1647 mark_inode_dirty_sync(inode
);
1651 * Get and clear the dirty flags from i_state. This needs to be done
1652 * after calling writepages because some filesystems may redirty the
1653 * inode during writepages due to delalloc. It also needs to be done
1654 * after handling timestamp expiration, as that may dirty the inode too.
1656 spin_lock(&inode
->i_lock
);
1657 dirty
= inode
->i_state
& I_DIRTY
;
1658 inode
->i_state
&= ~dirty
;
1661 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1662 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1663 * either they see the I_DIRTY bits cleared or we see the dirtied
1666 * I_DIRTY_PAGES is always cleared together above even if @mapping
1667 * still has dirty pages. The flag is reinstated after smp_mb() if
1668 * necessary. This guarantees that either __mark_inode_dirty()
1669 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1673 if (mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
1674 inode
->i_state
|= I_DIRTY_PAGES
;
1676 spin_unlock(&inode
->i_lock
);
1678 /* Don't write the inode if only I_DIRTY_PAGES was set */
1679 if (dirty
& ~I_DIRTY_PAGES
) {
1680 int err
= write_inode(inode
, wbc
);
1684 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
1689 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1690 * the regular batched writeback done by the flusher threads in
1691 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1692 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1694 * To prevent the inode from going away, either the caller must have a reference
1695 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1697 static int writeback_single_inode(struct inode
*inode
,
1698 struct writeback_control
*wbc
)
1700 struct bdi_writeback
*wb
;
1703 spin_lock(&inode
->i_lock
);
1704 if (!atomic_read(&inode
->i_count
))
1705 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
1707 WARN_ON(inode
->i_state
& I_WILL_FREE
);
1709 if (inode
->i_state
& I_SYNC
) {
1711 * Writeback is already running on the inode. For WB_SYNC_NONE,
1712 * that's enough and we can just return. For WB_SYNC_ALL, we
1713 * must wait for the existing writeback to complete, then do
1714 * writeback again if there's anything left.
1716 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
1718 __inode_wait_for_writeback(inode
);
1720 WARN_ON(inode
->i_state
& I_SYNC
);
1722 * If the inode is already fully clean, then there's nothing to do.
1724 * For data-integrity syncs we also need to check whether any pages are
1725 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1726 * there are any such pages, we'll need to wait for them.
1728 if (!(inode
->i_state
& I_DIRTY_ALL
) &&
1729 (wbc
->sync_mode
!= WB_SYNC_ALL
||
1730 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
1732 inode
->i_state
|= I_SYNC
;
1733 wbc_attach_and_unlock_inode(wbc
, inode
);
1735 ret
= __writeback_single_inode(inode
, wbc
);
1737 wbc_detach_inode(wbc
);
1739 wb
= inode_to_wb_and_lock_list(inode
);
1740 spin_lock(&inode
->i_lock
);
1742 * If the inode is freeing, its i_io_list shoudn't be updated
1743 * as it can be finally deleted at this moment.
1745 if (!(inode
->i_state
& I_FREEING
)) {
1747 * If the inode is now fully clean, then it can be safely
1748 * removed from its writeback list (if any). Otherwise the
1749 * flusher threads are responsible for the writeback lists.
1751 if (!(inode
->i_state
& I_DIRTY_ALL
))
1752 inode_cgwb_move_to_attached(inode
, wb
);
1753 else if (!(inode
->i_state
& I_SYNC_QUEUED
)) {
1754 if ((inode
->i_state
& I_DIRTY
))
1755 redirty_tail_locked(inode
, wb
);
1756 else if (inode
->i_state
& I_DIRTY_TIME
) {
1757 inode
->dirtied_when
= jiffies
;
1758 inode_io_list_move_locked(inode
,
1765 spin_unlock(&wb
->list_lock
);
1766 inode_sync_complete(inode
);
1768 spin_unlock(&inode
->i_lock
);
1772 static long writeback_chunk_size(struct bdi_writeback
*wb
,
1773 struct wb_writeback_work
*work
)
1778 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1779 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1780 * here avoids calling into writeback_inodes_wb() more than once.
1782 * The intended call sequence for WB_SYNC_ALL writeback is:
1785 * writeback_sb_inodes() <== called only once
1786 * write_cache_pages() <== called once for each inode
1787 * (quickly) tag currently dirty pages
1788 * (maybe slowly) sync all tagged pages
1790 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
1793 pages
= min(wb
->avg_write_bandwidth
/ 2,
1794 global_wb_domain
.dirty_limit
/ DIRTY_SCOPE
);
1795 pages
= min(pages
, work
->nr_pages
);
1796 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
1797 MIN_WRITEBACK_PAGES
);
1804 * Write a portion of b_io inodes which belong to @sb.
1806 * Return the number of pages and/or inodes written.
1808 * NOTE! This is called with wb->list_lock held, and will
1809 * unlock and relock that for each inode it ends up doing
1812 static long writeback_sb_inodes(struct super_block
*sb
,
1813 struct bdi_writeback
*wb
,
1814 struct wb_writeback_work
*work
)
1816 struct writeback_control wbc
= {
1817 .sync_mode
= work
->sync_mode
,
1818 .tagged_writepages
= work
->tagged_writepages
,
1819 .for_kupdate
= work
->for_kupdate
,
1820 .for_background
= work
->for_background
,
1821 .for_sync
= work
->for_sync
,
1822 .range_cyclic
= work
->range_cyclic
,
1824 .range_end
= LLONG_MAX
,
1826 unsigned long start_time
= jiffies
;
1828 long total_wrote
= 0; /* count both pages and inodes */
1830 while (!list_empty(&wb
->b_io
)) {
1831 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1832 struct bdi_writeback
*tmp_wb
;
1835 if (inode
->i_sb
!= sb
) {
1838 * We only want to write back data for this
1839 * superblock, move all inodes not belonging
1840 * to it back onto the dirty list.
1842 redirty_tail(inode
, wb
);
1847 * The inode belongs to a different superblock.
1848 * Bounce back to the caller to unpin this and
1849 * pin the next superblock.
1855 * Don't bother with new inodes or inodes being freed, first
1856 * kind does not need periodic writeout yet, and for the latter
1857 * kind writeout is handled by the freer.
1859 spin_lock(&inode
->i_lock
);
1860 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
1861 redirty_tail_locked(inode
, wb
);
1862 spin_unlock(&inode
->i_lock
);
1865 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
1867 * If this inode is locked for writeback and we are not
1868 * doing writeback-for-data-integrity, move it to
1869 * b_more_io so that writeback can proceed with the
1870 * other inodes on s_io.
1872 * We'll have another go at writing back this inode
1873 * when we completed a full scan of b_io.
1875 requeue_io(inode
, wb
);
1876 spin_unlock(&inode
->i_lock
);
1877 trace_writeback_sb_inodes_requeue(inode
);
1880 spin_unlock(&wb
->list_lock
);
1883 * We already requeued the inode if it had I_SYNC set and we
1884 * are doing WB_SYNC_NONE writeback. So this catches only the
1887 if (inode
->i_state
& I_SYNC
) {
1888 /* Wait for I_SYNC. This function drops i_lock... */
1889 inode_sleep_on_writeback(inode
);
1890 /* Inode may be gone, start again */
1891 spin_lock(&wb
->list_lock
);
1894 inode
->i_state
|= I_SYNC
;
1895 wbc_attach_and_unlock_inode(&wbc
, inode
);
1897 write_chunk
= writeback_chunk_size(wb
, work
);
1898 wbc
.nr_to_write
= write_chunk
;
1899 wbc
.pages_skipped
= 0;
1902 * We use I_SYNC to pin the inode in memory. While it is set
1903 * evict_inode() will wait so the inode cannot be freed.
1905 __writeback_single_inode(inode
, &wbc
);
1907 wbc_detach_inode(&wbc
);
1908 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
1909 wrote
= write_chunk
- wbc
.nr_to_write
- wbc
.pages_skipped
;
1910 wrote
= wrote
< 0 ? 0 : wrote
;
1911 total_wrote
+= wrote
;
1913 if (need_resched()) {
1915 * We're trying to balance between building up a nice
1916 * long list of IOs to improve our merge rate, and
1917 * getting those IOs out quickly for anyone throttling
1918 * in balance_dirty_pages(). cond_resched() doesn't
1919 * unplug, so get our IOs out the door before we
1922 blk_flush_plug(current
);
1927 * Requeue @inode if still dirty. Be careful as @inode may
1928 * have been switched to another wb in the meantime.
1930 tmp_wb
= inode_to_wb_and_lock_list(inode
);
1931 spin_lock(&inode
->i_lock
);
1932 if (!(inode
->i_state
& I_DIRTY_ALL
))
1934 requeue_inode(inode
, tmp_wb
, &wbc
);
1935 inode_sync_complete(inode
);
1936 spin_unlock(&inode
->i_lock
);
1938 if (unlikely(tmp_wb
!= wb
)) {
1939 spin_unlock(&tmp_wb
->list_lock
);
1940 spin_lock(&wb
->list_lock
);
1944 * bail out to wb_writeback() often enough to check
1945 * background threshold and other termination conditions.
1948 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1950 if (work
->nr_pages
<= 0)
1957 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
1958 struct wb_writeback_work
*work
)
1960 unsigned long start_time
= jiffies
;
1963 while (!list_empty(&wb
->b_io
)) {
1964 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1965 struct super_block
*sb
= inode
->i_sb
;
1967 if (!trylock_super(sb
)) {
1969 * trylock_super() may fail consistently due to
1970 * s_umount being grabbed by someone else. Don't use
1971 * requeue_io() to avoid busy retrying the inode/sb.
1973 redirty_tail(inode
, wb
);
1976 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
1977 up_read(&sb
->s_umount
);
1979 /* refer to the same tests at the end of writeback_sb_inodes */
1981 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1983 if (work
->nr_pages
<= 0)
1987 /* Leave any unwritten inodes on b_io */
1991 static long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
1992 enum wb_reason reason
)
1994 struct wb_writeback_work work
= {
1995 .nr_pages
= nr_pages
,
1996 .sync_mode
= WB_SYNC_NONE
,
2000 struct blk_plug plug
;
2002 blk_start_plug(&plug
);
2003 spin_lock(&wb
->list_lock
);
2004 if (list_empty(&wb
->b_io
))
2005 queue_io(wb
, &work
, jiffies
);
2006 __writeback_inodes_wb(wb
, &work
);
2007 spin_unlock(&wb
->list_lock
);
2008 blk_finish_plug(&plug
);
2010 return nr_pages
- work
.nr_pages
;
2014 * Explicit flushing or periodic writeback of "old" data.
2016 * Define "old": the first time one of an inode's pages is dirtied, we mark the
2017 * dirtying-time in the inode's address_space. So this periodic writeback code
2018 * just walks the superblock inode list, writing back any inodes which are
2019 * older than a specific point in time.
2021 * Try to run once per dirty_writeback_interval. But if a writeback event
2022 * takes longer than a dirty_writeback_interval interval, then leave a
2025 * dirtied_before takes precedence over nr_to_write. So we'll only write back
2026 * all dirty pages if they are all attached to "old" mappings.
2028 static long wb_writeback(struct bdi_writeback
*wb
,
2029 struct wb_writeback_work
*work
)
2031 long nr_pages
= work
->nr_pages
;
2032 unsigned long dirtied_before
= jiffies
;
2033 struct inode
*inode
;
2035 struct blk_plug plug
;
2037 blk_start_plug(&plug
);
2038 spin_lock(&wb
->list_lock
);
2041 * Stop writeback when nr_pages has been consumed
2043 if (work
->nr_pages
<= 0)
2047 * Background writeout and kupdate-style writeback may
2048 * run forever. Stop them if there is other work to do
2049 * so that e.g. sync can proceed. They'll be restarted
2050 * after the other works are all done.
2052 if ((work
->for_background
|| work
->for_kupdate
) &&
2053 !list_empty(&wb
->work_list
))
2057 * For background writeout, stop when we are below the
2058 * background dirty threshold
2060 if (work
->for_background
&& !wb_over_bg_thresh(wb
))
2064 * Kupdate and background works are special and we want to
2065 * include all inodes that need writing. Livelock avoidance is
2066 * handled by these works yielding to any other work so we are
2069 if (work
->for_kupdate
) {
2070 dirtied_before
= jiffies
-
2071 msecs_to_jiffies(dirty_expire_interval
* 10);
2072 } else if (work
->for_background
)
2073 dirtied_before
= jiffies
;
2075 trace_writeback_start(wb
, work
);
2076 if (list_empty(&wb
->b_io
))
2077 queue_io(wb
, work
, dirtied_before
);
2079 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
2081 progress
= __writeback_inodes_wb(wb
, work
);
2082 trace_writeback_written(wb
, work
);
2085 * Did we write something? Try for more
2087 * Dirty inodes are moved to b_io for writeback in batches.
2088 * The completion of the current batch does not necessarily
2089 * mean the overall work is done. So we keep looping as long
2090 * as made some progress on cleaning pages or inodes.
2095 * No more inodes for IO, bail
2097 if (list_empty(&wb
->b_more_io
))
2100 * Nothing written. Wait for some inode to
2101 * become available for writeback. Otherwise
2102 * we'll just busyloop.
2104 trace_writeback_wait(wb
, work
);
2105 inode
= wb_inode(wb
->b_more_io
.prev
);
2106 spin_lock(&inode
->i_lock
);
2107 spin_unlock(&wb
->list_lock
);
2108 /* This function drops i_lock... */
2109 inode_sleep_on_writeback(inode
);
2110 spin_lock(&wb
->list_lock
);
2112 spin_unlock(&wb
->list_lock
);
2113 blk_finish_plug(&plug
);
2115 return nr_pages
- work
->nr_pages
;
2119 * Return the next wb_writeback_work struct that hasn't been processed yet.
2121 static struct wb_writeback_work
*get_next_work_item(struct bdi_writeback
*wb
)
2123 struct wb_writeback_work
*work
= NULL
;
2125 spin_lock_irq(&wb
->work_lock
);
2126 if (!list_empty(&wb
->work_list
)) {
2127 work
= list_entry(wb
->work_list
.next
,
2128 struct wb_writeback_work
, list
);
2129 list_del_init(&work
->list
);
2131 spin_unlock_irq(&wb
->work_lock
);
2135 static long wb_check_background_flush(struct bdi_writeback
*wb
)
2137 if (wb_over_bg_thresh(wb
)) {
2139 struct wb_writeback_work work
= {
2140 .nr_pages
= LONG_MAX
,
2141 .sync_mode
= WB_SYNC_NONE
,
2142 .for_background
= 1,
2144 .reason
= WB_REASON_BACKGROUND
,
2147 return wb_writeback(wb
, &work
);
2153 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
2155 unsigned long expired
;
2159 * When set to zero, disable periodic writeback
2161 if (!dirty_writeback_interval
)
2164 expired
= wb
->last_old_flush
+
2165 msecs_to_jiffies(dirty_writeback_interval
* 10);
2166 if (time_before(jiffies
, expired
))
2169 wb
->last_old_flush
= jiffies
;
2170 nr_pages
= get_nr_dirty_pages();
2173 struct wb_writeback_work work
= {
2174 .nr_pages
= nr_pages
,
2175 .sync_mode
= WB_SYNC_NONE
,
2178 .reason
= WB_REASON_PERIODIC
,
2181 return wb_writeback(wb
, &work
);
2187 static long wb_check_start_all(struct bdi_writeback
*wb
)
2191 if (!test_bit(WB_start_all
, &wb
->state
))
2194 nr_pages
= get_nr_dirty_pages();
2196 struct wb_writeback_work work
= {
2197 .nr_pages
= wb_split_bdi_pages(wb
, nr_pages
),
2198 .sync_mode
= WB_SYNC_NONE
,
2200 .reason
= wb
->start_all_reason
,
2203 nr_pages
= wb_writeback(wb
, &work
);
2206 clear_bit(WB_start_all
, &wb
->state
);
2212 * Retrieve work items and do the writeback they describe
2214 static long wb_do_writeback(struct bdi_writeback
*wb
)
2216 struct wb_writeback_work
*work
;
2219 set_bit(WB_writeback_running
, &wb
->state
);
2220 while ((work
= get_next_work_item(wb
)) != NULL
) {
2221 trace_writeback_exec(wb
, work
);
2222 wrote
+= wb_writeback(wb
, work
);
2223 finish_writeback_work(wb
, work
);
2227 * Check for a flush-everything request
2229 wrote
+= wb_check_start_all(wb
);
2232 * Check for periodic writeback, kupdated() style
2234 wrote
+= wb_check_old_data_flush(wb
);
2235 wrote
+= wb_check_background_flush(wb
);
2236 clear_bit(WB_writeback_running
, &wb
->state
);
2242 * Handle writeback of dirty data for the device backed by this bdi. Also
2243 * reschedules periodically and does kupdated style flushing.
2245 void wb_workfn(struct work_struct
*work
)
2247 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
2248 struct bdi_writeback
, dwork
);
2251 set_worker_desc("flush-%s", bdi_dev_name(wb
->bdi
));
2252 current
->flags
|= PF_SWAPWRITE
;
2254 if (likely(!current_is_workqueue_rescuer() ||
2255 !test_bit(WB_registered
, &wb
->state
))) {
2257 * The normal path. Keep writing back @wb until its
2258 * work_list is empty. Note that this path is also taken
2259 * if @wb is shutting down even when we're running off the
2260 * rescuer as work_list needs to be drained.
2263 pages_written
= wb_do_writeback(wb
);
2264 trace_writeback_pages_written(pages_written
);
2265 } while (!list_empty(&wb
->work_list
));
2268 * bdi_wq can't get enough workers and we're running off
2269 * the emergency worker. Don't hog it. Hopefully, 1024 is
2270 * enough for efficient IO.
2272 pages_written
= writeback_inodes_wb(wb
, 1024,
2273 WB_REASON_FORKER_THREAD
);
2274 trace_writeback_pages_written(pages_written
);
2277 if (!list_empty(&wb
->work_list
))
2279 else if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
2280 wb_wakeup_delayed(wb
);
2282 current
->flags
&= ~PF_SWAPWRITE
;
2286 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2287 * write back the whole world.
2289 static void __wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2290 enum wb_reason reason
)
2292 struct bdi_writeback
*wb
;
2294 if (!bdi_has_dirty_io(bdi
))
2297 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2298 wb_start_writeback(wb
, reason
);
2301 void wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2302 enum wb_reason reason
)
2305 __wakeup_flusher_threads_bdi(bdi
, reason
);
2310 * Wakeup the flusher threads to start writeback of all currently dirty pages
2312 void wakeup_flusher_threads(enum wb_reason reason
)
2314 struct backing_dev_info
*bdi
;
2317 * If we are expecting writeback progress we must submit plugged IO.
2319 if (blk_needs_flush_plug(current
))
2320 blk_schedule_flush_plug(current
);
2323 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
2324 __wakeup_flusher_threads_bdi(bdi
, reason
);
2329 * Wake up bdi's periodically to make sure dirtytime inodes gets
2330 * written back periodically. We deliberately do *not* check the
2331 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2332 * kernel to be constantly waking up once there are any dirtytime
2333 * inodes on the system. So instead we define a separate delayed work
2334 * function which gets called much more rarely. (By default, only
2335 * once every 12 hours.)
2337 * If there is any other write activity going on in the file system,
2338 * this function won't be necessary. But if the only thing that has
2339 * happened on the file system is a dirtytime inode caused by an atime
2340 * update, we need this infrastructure below to make sure that inode
2341 * eventually gets pushed out to disk.
2343 static void wakeup_dirtytime_writeback(struct work_struct
*w
);
2344 static DECLARE_DELAYED_WORK(dirtytime_work
, wakeup_dirtytime_writeback
);
2346 static void wakeup_dirtytime_writeback(struct work_struct
*w
)
2348 struct backing_dev_info
*bdi
;
2351 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
2352 struct bdi_writeback
*wb
;
2354 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2355 if (!list_empty(&wb
->b_dirty_time
))
2359 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2362 static int __init
start_dirtytime_writeback(void)
2364 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2367 __initcall(start_dirtytime_writeback
);
2369 int dirtytime_interval_handler(struct ctl_table
*table
, int write
,
2370 void *buffer
, size_t *lenp
, loff_t
*ppos
)
2374 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
2375 if (ret
== 0 && write
)
2376 mod_delayed_work(system_wq
, &dirtytime_work
, 0);
2381 * __mark_inode_dirty - internal function to mark an inode dirty
2383 * @inode: inode to mark
2384 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2385 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2386 * with I_DIRTY_PAGES.
2388 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2389 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2391 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2392 * instead of calling this directly.
2394 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2395 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2396 * even if they are later hashed, as they will have been marked dirty already.
2398 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2400 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2401 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2402 * the kernel-internal blockdev inode represents the dirtying time of the
2403 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2404 * page->mapping->host, so the page-dirtying time is recorded in the internal
2407 void __mark_inode_dirty(struct inode
*inode
, int flags
)
2409 struct super_block
*sb
= inode
->i_sb
;
2411 struct bdi_writeback
*wb
= NULL
;
2413 trace_writeback_mark_inode_dirty(inode
, flags
);
2415 if (flags
& I_DIRTY_INODE
) {
2417 * Inode timestamp update will piggback on this dirtying.
2418 * We tell ->dirty_inode callback that timestamps need to
2419 * be updated by setting I_DIRTY_TIME in flags.
2421 if (inode
->i_state
& I_DIRTY_TIME
) {
2422 spin_lock(&inode
->i_lock
);
2423 if (inode
->i_state
& I_DIRTY_TIME
) {
2424 inode
->i_state
&= ~I_DIRTY_TIME
;
2425 flags
|= I_DIRTY_TIME
;
2427 spin_unlock(&inode
->i_lock
);
2431 * Notify the filesystem about the inode being dirtied, so that
2432 * (if needed) it can update on-disk fields and journal the
2433 * inode. This is only needed when the inode itself is being
2434 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2435 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2437 trace_writeback_dirty_inode_start(inode
, flags
);
2438 if (sb
->s_op
->dirty_inode
)
2439 sb
->s_op
->dirty_inode(inode
,
2440 flags
& (I_DIRTY_INODE
| I_DIRTY_TIME
));
2441 trace_writeback_dirty_inode(inode
, flags
);
2443 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2444 flags
&= ~I_DIRTY_TIME
;
2447 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2448 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2449 * in one call to __mark_inode_dirty().)
2451 dirtytime
= flags
& I_DIRTY_TIME
;
2452 WARN_ON_ONCE(dirtytime
&& flags
!= I_DIRTY_TIME
);
2456 * Paired with smp_mb() in __writeback_single_inode() for the
2457 * following lockless i_state test. See there for details.
2461 if ((inode
->i_state
& flags
) == flags
)
2464 spin_lock(&inode
->i_lock
);
2465 if ((inode
->i_state
& flags
) != flags
) {
2466 const int was_dirty
= inode
->i_state
& I_DIRTY
;
2468 inode_attach_wb(inode
, NULL
);
2470 inode
->i_state
|= flags
;
2473 * Grab inode's wb early because it requires dropping i_lock and we
2474 * need to make sure following checks happen atomically with dirty
2475 * list handling so that we don't move inodes under flush worker's
2479 wb
= locked_inode_to_wb_and_lock_list(inode
);
2480 spin_lock(&inode
->i_lock
);
2484 * If the inode is queued for writeback by flush worker, just
2485 * update its dirty state. Once the flush worker is done with
2486 * the inode it will place it on the appropriate superblock
2487 * list, based upon its state.
2489 if (inode
->i_state
& I_SYNC_QUEUED
)
2493 * Only add valid (hashed) inodes to the superblock's
2494 * dirty list. Add blockdev inodes as well.
2496 if (!S_ISBLK(inode
->i_mode
)) {
2497 if (inode_unhashed(inode
))
2500 if (inode
->i_state
& I_FREEING
)
2504 * If the inode was already on b_dirty/b_io/b_more_io, don't
2505 * reposition it (that would break b_dirty time-ordering).
2508 struct list_head
*dirty_list
;
2509 bool wakeup_bdi
= false;
2511 inode
->dirtied_when
= jiffies
;
2513 inode
->dirtied_time_when
= jiffies
;
2515 if (inode
->i_state
& I_DIRTY
)
2516 dirty_list
= &wb
->b_dirty
;
2518 dirty_list
= &wb
->b_dirty_time
;
2520 wakeup_bdi
= inode_io_list_move_locked(inode
, wb
,
2523 spin_unlock(&wb
->list_lock
);
2524 spin_unlock(&inode
->i_lock
);
2525 trace_writeback_dirty_inode_enqueue(inode
);
2528 * If this is the first dirty inode for this bdi,
2529 * we have to wake-up the corresponding bdi thread
2530 * to make sure background write-back happens
2534 (wb
->bdi
->capabilities
& BDI_CAP_WRITEBACK
))
2535 wb_wakeup_delayed(wb
);
2541 spin_unlock(&wb
->list_lock
);
2542 spin_unlock(&inode
->i_lock
);
2544 EXPORT_SYMBOL(__mark_inode_dirty
);
2547 * The @s_sync_lock is used to serialise concurrent sync operations
2548 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2549 * Concurrent callers will block on the s_sync_lock rather than doing contending
2550 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2551 * has been issued up to the time this function is enter is guaranteed to be
2552 * completed by the time we have gained the lock and waited for all IO that is
2553 * in progress regardless of the order callers are granted the lock.
2555 static void wait_sb_inodes(struct super_block
*sb
)
2557 LIST_HEAD(sync_list
);
2560 * We need to be protected against the filesystem going from
2561 * r/o to r/w or vice versa.
2563 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2565 mutex_lock(&sb
->s_sync_lock
);
2568 * Splice the writeback list onto a temporary list to avoid waiting on
2569 * inodes that have started writeback after this point.
2571 * Use rcu_read_lock() to keep the inodes around until we have a
2572 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2573 * the local list because inodes can be dropped from either by writeback
2577 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2578 list_splice_init(&sb
->s_inodes_wb
, &sync_list
);
2581 * Data integrity sync. Must wait for all pages under writeback, because
2582 * there may have been pages dirtied before our sync call, but which had
2583 * writeout started before we write it out. In which case, the inode
2584 * may not be on the dirty list, but we still have to wait for that
2587 while (!list_empty(&sync_list
)) {
2588 struct inode
*inode
= list_first_entry(&sync_list
, struct inode
,
2590 struct address_space
*mapping
= inode
->i_mapping
;
2593 * Move each inode back to the wb list before we drop the lock
2594 * to preserve consistency between i_wb_list and the mapping
2595 * writeback tag. Writeback completion is responsible to remove
2596 * the inode from either list once the writeback tag is cleared.
2598 list_move_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
2601 * The mapping can appear untagged while still on-list since we
2602 * do not have the mapping lock. Skip it here, wb completion
2605 if (!mapping_tagged(mapping
, PAGECACHE_TAG_WRITEBACK
))
2608 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2610 spin_lock(&inode
->i_lock
);
2611 if (inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) {
2612 spin_unlock(&inode
->i_lock
);
2614 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2618 spin_unlock(&inode
->i_lock
);
2622 * We keep the error status of individual mapping so that
2623 * applications can catch the writeback error using fsync(2).
2624 * See filemap_fdatawait_keep_errors() for details.
2626 filemap_fdatawait_keep_errors(mapping
);
2633 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2635 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2637 mutex_unlock(&sb
->s_sync_lock
);
2640 static void __writeback_inodes_sb_nr(struct super_block
*sb
, unsigned long nr
,
2641 enum wb_reason reason
, bool skip_if_busy
)
2643 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2644 DEFINE_WB_COMPLETION(done
, bdi
);
2645 struct wb_writeback_work work
= {
2647 .sync_mode
= WB_SYNC_NONE
,
2648 .tagged_writepages
= 1,
2654 if (!bdi_has_dirty_io(bdi
) || bdi
== &noop_backing_dev_info
)
2656 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2658 bdi_split_work_to_wbs(sb
->s_bdi
, &work
, skip_if_busy
);
2659 wb_wait_for_completion(&done
);
2663 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2664 * @sb: the superblock
2665 * @nr: the number of pages to write
2666 * @reason: reason why some writeback work initiated
2668 * Start writeback on some inodes on this super_block. No guarantees are made
2669 * on how many (if any) will be written, and this function does not wait
2670 * for IO completion of submitted IO.
2672 void writeback_inodes_sb_nr(struct super_block
*sb
,
2674 enum wb_reason reason
)
2676 __writeback_inodes_sb_nr(sb
, nr
, reason
, false);
2678 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
2681 * writeback_inodes_sb - writeback dirty inodes from given super_block
2682 * @sb: the superblock
2683 * @reason: reason why some writeback work was initiated
2685 * Start writeback on some inodes on this super_block. No guarantees are made
2686 * on how many (if any) will be written, and this function does not wait
2687 * for IO completion of submitted IO.
2689 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2691 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
2693 EXPORT_SYMBOL(writeback_inodes_sb
);
2696 * try_to_writeback_inodes_sb - try to start writeback if none underway
2697 * @sb: the superblock
2698 * @reason: reason why some writeback work was initiated
2700 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2702 void try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2704 if (!down_read_trylock(&sb
->s_umount
))
2707 __writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
, true);
2708 up_read(&sb
->s_umount
);
2710 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
2713 * sync_inodes_sb - sync sb inode pages
2714 * @sb: the superblock
2716 * This function writes and waits on any dirty inode belonging to this
2719 void sync_inodes_sb(struct super_block
*sb
)
2721 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2722 DEFINE_WB_COMPLETION(done
, bdi
);
2723 struct wb_writeback_work work
= {
2725 .sync_mode
= WB_SYNC_ALL
,
2726 .nr_pages
= LONG_MAX
,
2729 .reason
= WB_REASON_SYNC
,
2734 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2735 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2736 * bdi_has_dirty() need to be written out too.
2738 if (bdi
== &noop_backing_dev_info
)
2740 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2742 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2743 bdi_down_write_wb_switch_rwsem(bdi
);
2744 bdi_split_work_to_wbs(bdi
, &work
, false);
2745 wb_wait_for_completion(&done
);
2746 bdi_up_write_wb_switch_rwsem(bdi
);
2750 EXPORT_SYMBOL(sync_inodes_sb
);
2753 * write_inode_now - write an inode to disk
2754 * @inode: inode to write to disk
2755 * @sync: whether the write should be synchronous or not
2757 * This function commits an inode to disk immediately if it is dirty. This is
2758 * primarily needed by knfsd.
2760 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2762 int write_inode_now(struct inode
*inode
, int sync
)
2764 struct writeback_control wbc
= {
2765 .nr_to_write
= LONG_MAX
,
2766 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2768 .range_end
= LLONG_MAX
,
2771 if (!mapping_can_writeback(inode
->i_mapping
))
2772 wbc
.nr_to_write
= 0;
2775 return writeback_single_inode(inode
, &wbc
);
2777 EXPORT_SYMBOL(write_inode_now
);
2780 * sync_inode_metadata - write an inode to disk
2781 * @inode: the inode to sync
2782 * @wait: wait for I/O to complete.
2784 * Write an inode to disk and adjust its dirty state after completion.
2786 * Note: only writes the actual inode, no associated data or other metadata.
2788 int sync_inode_metadata(struct inode
*inode
, int wait
)
2790 struct writeback_control wbc
= {
2791 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2792 .nr_to_write
= 0, /* metadata-only */
2795 return writeback_single_inode(inode
, &wbc
);
2797 EXPORT_SYMBOL(sync_inode_metadata
);