1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2002, Linus Torvalds.
7 * Contains all the functions related to writing back and waiting
8 * upon dirty inodes against superblocks, and writing back dirty
9 * pages against inodes. ie: data writeback. Writeout of the
10 * inode itself is not handled here.
12 * 10Apr2002 Andrew Morton
13 * Split out of fs/inode.c
14 * Additions for address_space-based writeback
17 #include <linux/kernel.h>
18 #include <linux/export.h>
19 #include <linux/spinlock.h>
20 #include <linux/slab.h>
21 #include <linux/sched.h>
24 #include <linux/pagemap.h>
25 #include <linux/kthread.h>
26 #include <linux/writeback.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/tracepoint.h>
30 #include <linux/device.h>
31 #include <linux/memcontrol.h>
35 * 4MB minimal write chunk size
37 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
40 * Passed into wb_writeback(), essentially a subset of writeback_control
42 struct wb_writeback_work
{
44 struct super_block
*sb
;
45 enum writeback_sync_modes sync_mode
;
46 unsigned int tagged_writepages
:1;
47 unsigned int for_kupdate
:1;
48 unsigned int range_cyclic
:1;
49 unsigned int for_background
:1;
50 unsigned int for_sync
:1; /* sync(2) WB_SYNC_ALL writeback */
51 unsigned int auto_free
:1; /* free on completion */
52 enum wb_reason reason
; /* why was writeback initiated? */
54 struct list_head list
; /* pending work list */
55 struct wb_completion
*done
; /* set if the caller waits */
59 * If an inode is constantly having its pages dirtied, but then the
60 * updates stop dirtytime_expire_interval seconds in the past, it's
61 * possible for the worst case time between when an inode has its
62 * timestamps updated and when they finally get written out to be two
63 * dirtytime_expire_intervals. We set the default to 12 hours (in
64 * seconds), which means most of the time inodes will have their
65 * timestamps written to disk after 12 hours, but in the worst case a
66 * few inodes might not their timestamps updated for 24 hours.
68 unsigned int dirtytime_expire_interval
= 12 * 60 * 60;
70 static inline struct inode
*wb_inode(struct list_head
*head
)
72 return list_entry(head
, struct inode
, i_io_list
);
76 * Include the creation of the trace points after defining the
77 * wb_writeback_work structure and inline functions so that the definition
78 * remains local to this file.
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/writeback.h>
83 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage
);
85 static bool wb_io_lists_populated(struct bdi_writeback
*wb
)
87 if (wb_has_dirty_io(wb
)) {
90 set_bit(WB_has_dirty_io
, &wb
->state
);
91 WARN_ON_ONCE(!wb
->avg_write_bandwidth
);
92 atomic_long_add(wb
->avg_write_bandwidth
,
93 &wb
->bdi
->tot_write_bandwidth
);
98 static void wb_io_lists_depopulated(struct bdi_writeback
*wb
)
100 if (wb_has_dirty_io(wb
) && list_empty(&wb
->b_dirty
) &&
101 list_empty(&wb
->b_io
) && list_empty(&wb
->b_more_io
)) {
102 clear_bit(WB_has_dirty_io
, &wb
->state
);
103 WARN_ON_ONCE(atomic_long_sub_return(wb
->avg_write_bandwidth
,
104 &wb
->bdi
->tot_write_bandwidth
) < 0);
109 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
110 * @inode: inode to be moved
111 * @wb: target bdi_writeback
112 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
114 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
115 * Returns %true if @inode is the first occupant of the !dirty_time IO
116 * lists; otherwise, %false.
118 static bool inode_io_list_move_locked(struct inode
*inode
,
119 struct bdi_writeback
*wb
,
120 struct list_head
*head
)
122 assert_spin_locked(&wb
->list_lock
);
124 list_move(&inode
->i_io_list
, head
);
126 /* dirty_time doesn't count as dirty_io until expiration */
127 if (head
!= &wb
->b_dirty_time
)
128 return wb_io_lists_populated(wb
);
130 wb_io_lists_depopulated(wb
);
134 static void wb_wakeup(struct bdi_writeback
*wb
)
136 spin_lock_bh(&wb
->work_lock
);
137 if (test_bit(WB_registered
, &wb
->state
))
138 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
139 spin_unlock_bh(&wb
->work_lock
);
142 static void finish_writeback_work(struct bdi_writeback
*wb
,
143 struct wb_writeback_work
*work
)
145 struct wb_completion
*done
= work
->done
;
150 wait_queue_head_t
*waitq
= done
->waitq
;
152 /* @done can't be accessed after the following dec */
153 if (atomic_dec_and_test(&done
->cnt
))
158 static void wb_queue_work(struct bdi_writeback
*wb
,
159 struct wb_writeback_work
*work
)
161 trace_writeback_queue(wb
, work
);
164 atomic_inc(&work
->done
->cnt
);
166 spin_lock_bh(&wb
->work_lock
);
168 if (test_bit(WB_registered
, &wb
->state
)) {
169 list_add_tail(&work
->list
, &wb
->work_list
);
170 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
172 finish_writeback_work(wb
, work
);
174 spin_unlock_bh(&wb
->work_lock
);
178 * wb_wait_for_completion - wait for completion of bdi_writeback_works
179 * @done: target wb_completion
181 * Wait for one or more work items issued to @bdi with their ->done field
182 * set to @done, which should have been initialized with
183 * DEFINE_WB_COMPLETION(). This function returns after all such work items
184 * are completed. Work items which are waited upon aren't freed
185 * automatically on completion.
187 void wb_wait_for_completion(struct wb_completion
*done
)
189 atomic_dec(&done
->cnt
); /* put down the initial count */
190 wait_event(*done
->waitq
, !atomic_read(&done
->cnt
));
193 #ifdef CONFIG_CGROUP_WRITEBACK
196 * Parameters for foreign inode detection, see wbc_detach_inode() to see
199 * These paramters are inherently heuristical as the detection target
200 * itself is fuzzy. All we want to do is detaching an inode from the
201 * current owner if it's being written to by some other cgroups too much.
203 * The current cgroup writeback is built on the assumption that multiple
204 * cgroups writing to the same inode concurrently is very rare and a mode
205 * of operation which isn't well supported. As such, the goal is not
206 * taking too long when a different cgroup takes over an inode while
207 * avoiding too aggressive flip-flops from occasional foreign writes.
209 * We record, very roughly, 2s worth of IO time history and if more than
210 * half of that is foreign, trigger the switch. The recording is quantized
211 * to 16 slots. To avoid tiny writes from swinging the decision too much,
212 * writes smaller than 1/8 of avg size are ignored.
214 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
215 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
216 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
217 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
219 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
220 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
221 /* each slot's duration is 2s / 16 */
222 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
223 /* if foreign slots >= 8, switch */
224 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
225 /* one round can affect upto 5 slots */
226 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
229 * Maximum inodes per isw. A specific value has been chosen to make
230 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
232 #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
233 / sizeof(struct inode *))
235 static atomic_t isw_nr_in_flight
= ATOMIC_INIT(0);
236 static struct workqueue_struct
*isw_wq
;
238 void __inode_attach_wb(struct inode
*inode
, struct page
*page
)
240 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
241 struct bdi_writeback
*wb
= NULL
;
243 if (inode_cgwb_enabled(inode
)) {
244 struct cgroup_subsys_state
*memcg_css
;
247 memcg_css
= mem_cgroup_css_from_page(page
);
248 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
250 /* must pin memcg_css, see wb_get_create() */
251 memcg_css
= task_get_css(current
, memory_cgrp_id
);
252 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
261 * There may be multiple instances of this function racing to
262 * update the same inode. Use cmpxchg() to tell the winner.
264 if (unlikely(cmpxchg(&inode
->i_wb
, NULL
, wb
)))
267 EXPORT_SYMBOL_GPL(__inode_attach_wb
);
270 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
271 * @inode: inode of interest with i_lock held
272 * @wb: target bdi_writeback
274 * Remove the inode from wb's io lists and if necessarily put onto b_attached
275 * list. Only inodes attached to cgwb's are kept on this list.
277 static void inode_cgwb_move_to_attached(struct inode
*inode
,
278 struct bdi_writeback
*wb
)
280 assert_spin_locked(&wb
->list_lock
);
281 assert_spin_locked(&inode
->i_lock
);
283 inode
->i_state
&= ~I_SYNC_QUEUED
;
284 if (wb
!= &wb
->bdi
->wb
)
285 list_move(&inode
->i_io_list
, &wb
->b_attached
);
287 list_del_init(&inode
->i_io_list
);
288 wb_io_lists_depopulated(wb
);
292 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
293 * @inode: inode of interest with i_lock held
295 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
296 * held on entry and is released on return. The returned wb is guaranteed
297 * to stay @inode's associated wb until its list_lock is released.
299 static struct bdi_writeback
*
300 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
301 __releases(&inode
->i_lock
)
302 __acquires(&wb
->list_lock
)
305 struct bdi_writeback
*wb
= inode_to_wb(inode
);
308 * inode_to_wb() association is protected by both
309 * @inode->i_lock and @wb->list_lock but list_lock nests
310 * outside i_lock. Drop i_lock and verify that the
311 * association hasn't changed after acquiring list_lock.
314 spin_unlock(&inode
->i_lock
);
315 spin_lock(&wb
->list_lock
);
317 /* i_wb may have changed inbetween, can't use inode_to_wb() */
318 if (likely(wb
== inode
->i_wb
)) {
319 wb_put(wb
); /* @inode already has ref */
323 spin_unlock(&wb
->list_lock
);
326 spin_lock(&inode
->i_lock
);
331 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
332 * @inode: inode of interest
334 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
337 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
338 __acquires(&wb
->list_lock
)
340 spin_lock(&inode
->i_lock
);
341 return locked_inode_to_wb_and_lock_list(inode
);
344 struct inode_switch_wbs_context
{
345 struct rcu_work work
;
348 * Multiple inodes can be switched at once. The switching procedure
349 * consists of two parts, separated by a RCU grace period. To make
350 * sure that the second part is executed for each inode gone through
351 * the first part, all inode pointers are placed into a NULL-terminated
352 * array embedded into struct inode_switch_wbs_context. Otherwise
353 * an inode could be left in a non-consistent state.
355 struct bdi_writeback
*new_wb
;
356 struct inode
*inodes
[];
359 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
361 down_write(&bdi
->wb_switch_rwsem
);
364 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
366 up_write(&bdi
->wb_switch_rwsem
);
369 static bool inode_do_switch_wbs(struct inode
*inode
,
370 struct bdi_writeback
*old_wb
,
371 struct bdi_writeback
*new_wb
)
373 struct address_space
*mapping
= inode
->i_mapping
;
374 XA_STATE(xas
, &mapping
->i_pages
, 0);
376 bool switched
= false;
378 spin_lock(&inode
->i_lock
);
379 xa_lock_irq(&mapping
->i_pages
);
382 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
383 * path owns the inode and we shouldn't modify ->i_io_list.
385 if (unlikely(inode
->i_state
& (I_FREEING
| I_WILL_FREE
)))
388 trace_inode_switch_wbs(inode
, old_wb
, new_wb
);
391 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
392 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
393 * pages actually under writeback.
395 xas_for_each_marked(&xas
, page
, ULONG_MAX
, PAGECACHE_TAG_DIRTY
) {
396 if (PageDirty(page
)) {
397 dec_wb_stat(old_wb
, WB_RECLAIMABLE
);
398 inc_wb_stat(new_wb
, WB_RECLAIMABLE
);
403 xas_for_each_marked(&xas
, page
, ULONG_MAX
, PAGECACHE_TAG_WRITEBACK
) {
404 WARN_ON_ONCE(!PageWriteback(page
));
405 dec_wb_stat(old_wb
, WB_WRITEBACK
);
406 inc_wb_stat(new_wb
, WB_WRITEBACK
);
412 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
413 * the specific list @inode was on is ignored and the @inode is put on
414 * ->b_dirty which is always correct including from ->b_dirty_time.
415 * The transfer preserves @inode->dirtied_when ordering. If the @inode
416 * was clean, it means it was on the b_attached list, so move it onto
417 * the b_attached list of @new_wb.
419 if (!list_empty(&inode
->i_io_list
)) {
420 inode
->i_wb
= new_wb
;
422 if (inode
->i_state
& I_DIRTY_ALL
) {
425 list_for_each_entry(pos
, &new_wb
->b_dirty
, i_io_list
)
426 if (time_after_eq(inode
->dirtied_when
,
429 inode_io_list_move_locked(inode
, new_wb
,
430 pos
->i_io_list
.prev
);
432 inode_cgwb_move_to_attached(inode
, new_wb
);
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 xa_unlock_irq(&mapping
->i_pages
);
451 spin_unlock(&inode
->i_lock
);
456 static void inode_switch_wbs_work_fn(struct work_struct
*work
)
458 struct inode_switch_wbs_context
*isw
=
459 container_of(to_rcu_work(work
), struct inode_switch_wbs_context
, work
);
460 struct backing_dev_info
*bdi
= inode_to_bdi(isw
->inodes
[0]);
461 struct bdi_writeback
*old_wb
= isw
->inodes
[0]->i_wb
;
462 struct bdi_writeback
*new_wb
= isw
->new_wb
;
463 unsigned long nr_switched
= 0;
464 struct inode
**inodep
;
467 * If @inode switches cgwb membership while sync_inodes_sb() is
468 * being issued, sync_inodes_sb() might miss it. Synchronize.
470 down_read(&bdi
->wb_switch_rwsem
);
473 * By the time control reaches here, RCU grace period has passed
474 * since I_WB_SWITCH assertion and all wb stat update transactions
475 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
476 * synchronizing against the i_pages lock.
478 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
479 * gives us exclusion against all wb related operations on @inode
480 * including IO list manipulations and stat updates.
482 if (old_wb
< new_wb
) {
483 spin_lock(&old_wb
->list_lock
);
484 spin_lock_nested(&new_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
486 spin_lock(&new_wb
->list_lock
);
487 spin_lock_nested(&old_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
490 for (inodep
= isw
->inodes
; *inodep
; inodep
++) {
491 WARN_ON_ONCE((*inodep
)->i_wb
!= old_wb
);
492 if (inode_do_switch_wbs(*inodep
, old_wb
, new_wb
))
496 spin_unlock(&new_wb
->list_lock
);
497 spin_unlock(&old_wb
->list_lock
);
499 up_read(&bdi
->wb_switch_rwsem
);
503 wb_put_many(old_wb
, nr_switched
);
506 for (inodep
= isw
->inodes
; *inodep
; inodep
++)
510 atomic_dec(&isw_nr_in_flight
);
513 static bool inode_prepare_wbs_switch(struct inode
*inode
,
514 struct bdi_writeback
*new_wb
)
517 * Paired with smp_mb() in cgroup_writeback_umount().
518 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
519 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
520 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
524 /* while holding I_WB_SWITCH, no one else can update the association */
525 spin_lock(&inode
->i_lock
);
526 if (!(inode
->i_sb
->s_flags
& SB_ACTIVE
) ||
527 inode
->i_state
& (I_WB_SWITCH
| I_FREEING
| I_WILL_FREE
) ||
528 inode_to_wb(inode
) == new_wb
) {
529 spin_unlock(&inode
->i_lock
);
532 inode
->i_state
|= I_WB_SWITCH
;
534 spin_unlock(&inode
->i_lock
);
540 * inode_switch_wbs - change the wb association of an inode
541 * @inode: target inode
542 * @new_wb_id: ID of the new wb
544 * Switch @inode's wb association to the wb identified by @new_wb_id. The
545 * switching is performed asynchronously and may fail silently.
547 static void inode_switch_wbs(struct inode
*inode
, int new_wb_id
)
549 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
550 struct cgroup_subsys_state
*memcg_css
;
551 struct inode_switch_wbs_context
*isw
;
553 /* noop if seems to be already in progress */
554 if (inode
->i_state
& I_WB_SWITCH
)
557 /* avoid queueing a new switch if too many are already in flight */
558 if (atomic_read(&isw_nr_in_flight
) > WB_FRN_MAX_IN_FLIGHT
)
561 isw
= kzalloc(sizeof(*isw
) + 2 * sizeof(struct inode
*), GFP_ATOMIC
);
565 atomic_inc(&isw_nr_in_flight
);
567 /* find and pin the new wb */
569 memcg_css
= css_from_id(new_wb_id
, &memory_cgrp_subsys
);
570 if (memcg_css
&& !css_tryget(memcg_css
))
576 isw
->new_wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
581 if (!inode_prepare_wbs_switch(inode
, isw
->new_wb
))
584 isw
->inodes
[0] = inode
;
587 * In addition to synchronizing among switchers, I_WB_SWITCH tells
588 * the RCU protected stat update paths to grab the i_page
589 * lock so that stat transfer can synchronize against them.
590 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
592 INIT_RCU_WORK(&isw
->work
, inode_switch_wbs_work_fn
);
593 queue_rcu_work(isw_wq
, &isw
->work
);
597 atomic_dec(&isw_nr_in_flight
);
604 * cleanup_offline_cgwb - detach associated inodes
607 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
608 * to eventually release the dying @wb. Returns %true if not all inodes were
609 * switched and the function has to be restarted.
611 bool cleanup_offline_cgwb(struct bdi_writeback
*wb
)
613 struct cgroup_subsys_state
*memcg_css
;
614 struct inode_switch_wbs_context
*isw
;
617 bool restart
= false;
619 isw
= kzalloc(sizeof(*isw
) + WB_MAX_INODES_PER_ISW
*
620 sizeof(struct inode
*), GFP_KERNEL
);
624 atomic_inc(&isw_nr_in_flight
);
626 for (memcg_css
= wb
->memcg_css
->parent
; memcg_css
;
627 memcg_css
= memcg_css
->parent
) {
628 isw
->new_wb
= wb_get_create(wb
->bdi
, memcg_css
, GFP_KERNEL
);
632 if (unlikely(!isw
->new_wb
))
633 isw
->new_wb
= &wb
->bdi
->wb
; /* wb_get() is noop for bdi's wb */
636 spin_lock(&wb
->list_lock
);
637 list_for_each_entry(inode
, &wb
->b_attached
, i_io_list
) {
638 if (!inode_prepare_wbs_switch(inode
, isw
->new_wb
))
641 isw
->inodes
[nr
++] = inode
;
643 if (nr
>= WB_MAX_INODES_PER_ISW
- 1) {
648 spin_unlock(&wb
->list_lock
);
650 /* no attached inodes? bail out */
652 atomic_dec(&isw_nr_in_flight
);
659 * In addition to synchronizing among switchers, I_WB_SWITCH tells
660 * the RCU protected stat update paths to grab the i_page
661 * lock so that stat transfer can synchronize against them.
662 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
664 INIT_RCU_WORK(&isw
->work
, inode_switch_wbs_work_fn
);
665 queue_rcu_work(isw_wq
, &isw
->work
);
671 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
672 * @wbc: writeback_control of interest
673 * @inode: target inode
675 * @inode is locked and about to be written back under the control of @wbc.
676 * Record @inode's writeback context into @wbc and unlock the i_lock. On
677 * writeback completion, wbc_detach_inode() should be called. This is used
678 * to track the cgroup writeback context.
680 void wbc_attach_and_unlock_inode(struct writeback_control
*wbc
,
683 if (!inode_cgwb_enabled(inode
)) {
684 spin_unlock(&inode
->i_lock
);
688 wbc
->wb
= inode_to_wb(inode
);
691 wbc
->wb_id
= wbc
->wb
->memcg_css
->id
;
692 wbc
->wb_lcand_id
= inode
->i_wb_frn_winner
;
693 wbc
->wb_tcand_id
= 0;
695 wbc
->wb_lcand_bytes
= 0;
696 wbc
->wb_tcand_bytes
= 0;
699 spin_unlock(&inode
->i_lock
);
702 * A dying wb indicates that either the blkcg associated with the
703 * memcg changed or the associated memcg is dying. In the first
704 * case, a replacement wb should already be available and we should
705 * refresh the wb immediately. In the second case, trying to
706 * refresh will keep failing.
708 if (unlikely(wb_dying(wbc
->wb
) && !css_is_dying(wbc
->wb
->memcg_css
)))
709 inode_switch_wbs(inode
, wbc
->wb_id
);
711 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode
);
714 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
715 * @wbc: writeback_control of the just finished writeback
717 * To be called after a writeback attempt of an inode finishes and undoes
718 * wbc_attach_and_unlock_inode(). Can be called under any context.
720 * As concurrent write sharing of an inode is expected to be very rare and
721 * memcg only tracks page ownership on first-use basis severely confining
722 * the usefulness of such sharing, cgroup writeback tracks ownership
723 * per-inode. While the support for concurrent write sharing of an inode
724 * is deemed unnecessary, an inode being written to by different cgroups at
725 * different points in time is a lot more common, and, more importantly,
726 * charging only by first-use can too readily lead to grossly incorrect
727 * behaviors (single foreign page can lead to gigabytes of writeback to be
728 * incorrectly attributed).
730 * To resolve this issue, cgroup writeback detects the majority dirtier of
731 * an inode and transfers the ownership to it. To avoid unnnecessary
732 * oscillation, the detection mechanism keeps track of history and gives
733 * out the switch verdict only if the foreign usage pattern is stable over
734 * a certain amount of time and/or writeback attempts.
736 * On each writeback attempt, @wbc tries to detect the majority writer
737 * using Boyer-Moore majority vote algorithm. In addition to the byte
738 * count from the majority voting, it also counts the bytes written for the
739 * current wb and the last round's winner wb (max of last round's current
740 * wb, the winner from two rounds ago, and the last round's majority
741 * candidate). Keeping track of the historical winner helps the algorithm
742 * to semi-reliably detect the most active writer even when it's not the
745 * Once the winner of the round is determined, whether the winner is
746 * foreign or not and how much IO time the round consumed is recorded in
747 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
748 * over a certain threshold, the switch verdict is given.
750 void wbc_detach_inode(struct writeback_control
*wbc
)
752 struct bdi_writeback
*wb
= wbc
->wb
;
753 struct inode
*inode
= wbc
->inode
;
754 unsigned long avg_time
, max_bytes
, max_time
;
761 history
= inode
->i_wb_frn_history
;
762 avg_time
= inode
->i_wb_frn_avg_time
;
764 /* pick the winner of this round */
765 if (wbc
->wb_bytes
>= wbc
->wb_lcand_bytes
&&
766 wbc
->wb_bytes
>= wbc
->wb_tcand_bytes
) {
768 max_bytes
= wbc
->wb_bytes
;
769 } else if (wbc
->wb_lcand_bytes
>= wbc
->wb_tcand_bytes
) {
770 max_id
= wbc
->wb_lcand_id
;
771 max_bytes
= wbc
->wb_lcand_bytes
;
773 max_id
= wbc
->wb_tcand_id
;
774 max_bytes
= wbc
->wb_tcand_bytes
;
778 * Calculate the amount of IO time the winner consumed and fold it
779 * into the running average kept per inode. If the consumed IO
780 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
781 * deciding whether to switch or not. This is to prevent one-off
782 * small dirtiers from skewing the verdict.
784 max_time
= DIV_ROUND_UP((max_bytes
>> PAGE_SHIFT
) << WB_FRN_TIME_SHIFT
,
785 wb
->avg_write_bandwidth
);
787 avg_time
+= (max_time
>> WB_FRN_TIME_AVG_SHIFT
) -
788 (avg_time
>> WB_FRN_TIME_AVG_SHIFT
);
790 avg_time
= max_time
; /* immediate catch up on first run */
792 if (max_time
>= avg_time
/ WB_FRN_TIME_CUT_DIV
) {
796 * The switch verdict is reached if foreign wb's consume
797 * more than a certain proportion of IO time in a
798 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
799 * history mask where each bit represents one sixteenth of
800 * the period. Determine the number of slots to shift into
801 * history from @max_time.
803 slots
= min(DIV_ROUND_UP(max_time
, WB_FRN_HIST_UNIT
),
804 (unsigned long)WB_FRN_HIST_MAX_SLOTS
);
806 if (wbc
->wb_id
!= max_id
)
807 history
|= (1U << slots
) - 1;
810 trace_inode_foreign_history(inode
, wbc
, history
);
813 * Switch if the current wb isn't the consistent winner.
814 * If there are multiple closely competing dirtiers, the
815 * inode may switch across them repeatedly over time, which
816 * is okay. The main goal is avoiding keeping an inode on
817 * the wrong wb for an extended period of time.
819 if (hweight32(history
) > WB_FRN_HIST_THR_SLOTS
)
820 inode_switch_wbs(inode
, max_id
);
824 * Multiple instances of this function may race to update the
825 * following fields but we don't mind occassional inaccuracies.
827 inode
->i_wb_frn_winner
= max_id
;
828 inode
->i_wb_frn_avg_time
= min(avg_time
, (unsigned long)U16_MAX
);
829 inode
->i_wb_frn_history
= history
;
834 EXPORT_SYMBOL_GPL(wbc_detach_inode
);
837 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
838 * @wbc: writeback_control of the writeback in progress
839 * @page: page being written out
840 * @bytes: number of bytes being written out
842 * @bytes from @page are about to written out during the writeback
843 * controlled by @wbc. Keep the book for foreign inode detection. See
844 * wbc_detach_inode().
846 void wbc_account_cgroup_owner(struct writeback_control
*wbc
, struct page
*page
,
849 struct cgroup_subsys_state
*css
;
853 * pageout() path doesn't attach @wbc to the inode being written
854 * out. This is intentional as we don't want the function to block
855 * behind a slow cgroup. Ultimately, we want pageout() to kick off
856 * regular writeback instead of writing things out itself.
858 if (!wbc
->wb
|| wbc
->no_cgroup_owner
)
861 css
= mem_cgroup_css_from_page(page
);
862 /* dead cgroups shouldn't contribute to inode ownership arbitration */
863 if (!(css
->flags
& CSS_ONLINE
))
868 if (id
== wbc
->wb_id
) {
869 wbc
->wb_bytes
+= bytes
;
873 if (id
== wbc
->wb_lcand_id
)
874 wbc
->wb_lcand_bytes
+= bytes
;
876 /* Boyer-Moore majority vote algorithm */
877 if (!wbc
->wb_tcand_bytes
)
878 wbc
->wb_tcand_id
= id
;
879 if (id
== wbc
->wb_tcand_id
)
880 wbc
->wb_tcand_bytes
+= bytes
;
882 wbc
->wb_tcand_bytes
-= min(bytes
, wbc
->wb_tcand_bytes
);
884 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner
);
887 * inode_congested - test whether an inode is congested
888 * @inode: inode to test for congestion (may be NULL)
889 * @cong_bits: mask of WB_[a]sync_congested bits to test
891 * Tests whether @inode is congested. @cong_bits is the mask of congestion
892 * bits to test and the return value is the mask of set bits.
894 * If cgroup writeback is enabled for @inode, the congestion state is
895 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
896 * associated with @inode is congested; otherwise, the root wb's congestion
899 * @inode is allowed to be NULL as this function is often called on
900 * mapping->host which is NULL for the swapper space.
902 int inode_congested(struct inode
*inode
, int cong_bits
)
905 * Once set, ->i_wb never becomes NULL while the inode is alive.
906 * Start transaction iff ->i_wb is visible.
908 if (inode
&& inode_to_wb_is_valid(inode
)) {
909 struct bdi_writeback
*wb
;
910 struct wb_lock_cookie lock_cookie
= {};
913 wb
= unlocked_inode_to_wb_begin(inode
, &lock_cookie
);
914 congested
= wb_congested(wb
, cong_bits
);
915 unlocked_inode_to_wb_end(inode
, &lock_cookie
);
919 return wb_congested(&inode_to_bdi(inode
)->wb
, cong_bits
);
921 EXPORT_SYMBOL_GPL(inode_congested
);
924 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
925 * @wb: target bdi_writeback to split @nr_pages to
926 * @nr_pages: number of pages to write for the whole bdi
928 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
929 * relation to the total write bandwidth of all wb's w/ dirty inodes on
932 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
934 unsigned long this_bw
= wb
->avg_write_bandwidth
;
935 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
937 if (nr_pages
== LONG_MAX
)
941 * This may be called on clean wb's and proportional distribution
942 * may not make sense, just use the original @nr_pages in those
943 * cases. In general, we wanna err on the side of writing more.
945 if (!tot_bw
|| this_bw
>= tot_bw
)
948 return DIV_ROUND_UP_ULL((u64
)nr_pages
* this_bw
, tot_bw
);
952 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
953 * @bdi: target backing_dev_info
954 * @base_work: wb_writeback_work to issue
955 * @skip_if_busy: skip wb's which already have writeback in progress
957 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
958 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
959 * distributed to the busy wbs according to each wb's proportion in the
960 * total active write bandwidth of @bdi.
962 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
963 struct wb_writeback_work
*base_work
,
966 struct bdi_writeback
*last_wb
= NULL
;
967 struct bdi_writeback
*wb
= list_entry(&bdi
->wb_list
,
968 struct bdi_writeback
, bdi_node
);
973 list_for_each_entry_continue_rcu(wb
, &bdi
->wb_list
, bdi_node
) {
974 DEFINE_WB_COMPLETION(fallback_work_done
, bdi
);
975 struct wb_writeback_work fallback_work
;
976 struct wb_writeback_work
*work
;
984 /* SYNC_ALL writes out I_DIRTY_TIME too */
985 if (!wb_has_dirty_io(wb
) &&
986 (base_work
->sync_mode
== WB_SYNC_NONE
||
987 list_empty(&wb
->b_dirty_time
)))
989 if (skip_if_busy
&& writeback_in_progress(wb
))
992 nr_pages
= wb_split_bdi_pages(wb
, base_work
->nr_pages
);
994 work
= kmalloc(sizeof(*work
), GFP_ATOMIC
);
997 work
->nr_pages
= nr_pages
;
999 wb_queue_work(wb
, work
);
1003 /* alloc failed, execute synchronously using on-stack fallback */
1004 work
= &fallback_work
;
1006 work
->nr_pages
= nr_pages
;
1007 work
->auto_free
= 0;
1008 work
->done
= &fallback_work_done
;
1010 wb_queue_work(wb
, work
);
1013 * Pin @wb so that it stays on @bdi->wb_list. This allows
1014 * continuing iteration from @wb after dropping and
1015 * regrabbing rcu read lock.
1021 wb_wait_for_completion(&fallback_work_done
);
1031 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1032 * @bdi_id: target bdi id
1033 * @memcg_id: target memcg css id
1034 * @nr: number of pages to write, 0 for best-effort dirty flushing
1035 * @reason: reason why some writeback work initiated
1036 * @done: target wb_completion
1038 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1039 * with the specified parameters.
1041 int cgroup_writeback_by_id(u64 bdi_id
, int memcg_id
, unsigned long nr
,
1042 enum wb_reason reason
, struct wb_completion
*done
)
1044 struct backing_dev_info
*bdi
;
1045 struct cgroup_subsys_state
*memcg_css
;
1046 struct bdi_writeback
*wb
;
1047 struct wb_writeback_work
*work
;
1050 /* lookup bdi and memcg */
1051 bdi
= bdi_get_by_id(bdi_id
);
1056 memcg_css
= css_from_id(memcg_id
, &memory_cgrp_subsys
);
1057 if (memcg_css
&& !css_tryget(memcg_css
))
1066 * And find the associated wb. If the wb isn't there already
1067 * there's nothing to flush, don't create one.
1069 wb
= wb_get_lookup(bdi
, memcg_css
);
1076 * If @nr is zero, the caller is attempting to write out most of
1077 * the currently dirty pages. Let's take the current dirty page
1078 * count and inflate it by 25% which should be large enough to
1079 * flush out most dirty pages while avoiding getting livelocked by
1080 * concurrent dirtiers.
1083 unsigned long filepages
, headroom
, dirty
, writeback
;
1085 mem_cgroup_wb_stats(wb
, &filepages
, &headroom
, &dirty
,
1087 nr
= dirty
* 10 / 8;
1090 /* issue the writeback work */
1091 work
= kzalloc(sizeof(*work
), GFP_NOWAIT
| __GFP_NOWARN
);
1093 work
->nr_pages
= nr
;
1094 work
->sync_mode
= WB_SYNC_NONE
;
1095 work
->range_cyclic
= 1;
1096 work
->reason
= reason
;
1098 work
->auto_free
= 1;
1099 wb_queue_work(wb
, work
);
1114 * cgroup_writeback_umount - flush inode wb switches for umount
1116 * This function is called when a super_block is about to be destroyed and
1117 * flushes in-flight inode wb switches. An inode wb switch goes through
1118 * RCU and then workqueue, so the two need to be flushed in order to ensure
1119 * that all previously scheduled switches are finished. As wb switches are
1120 * rare occurrences and synchronize_rcu() can take a while, perform
1121 * flushing iff wb switches are in flight.
1123 void cgroup_writeback_umount(void)
1126 * SB_ACTIVE should be reliably cleared before checking
1127 * isw_nr_in_flight, see generic_shutdown_super().
1131 if (atomic_read(&isw_nr_in_flight
)) {
1133 * Use rcu_barrier() to wait for all pending callbacks to
1134 * ensure that all in-flight wb switches are in the workqueue.
1137 flush_workqueue(isw_wq
);
1141 static int __init
cgroup_writeback_init(void)
1143 isw_wq
= alloc_workqueue("inode_switch_wbs", 0, 0);
1148 fs_initcall(cgroup_writeback_init
);
1150 #else /* CONFIG_CGROUP_WRITEBACK */
1152 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
1153 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
1155 static void inode_cgwb_move_to_attached(struct inode
*inode
,
1156 struct bdi_writeback
*wb
)
1158 assert_spin_locked(&wb
->list_lock
);
1159 assert_spin_locked(&inode
->i_lock
);
1161 inode
->i_state
&= ~I_SYNC_QUEUED
;
1162 list_del_init(&inode
->i_io_list
);
1163 wb_io_lists_depopulated(wb
);
1166 static struct bdi_writeback
*
1167 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
1168 __releases(&inode
->i_lock
)
1169 __acquires(&wb
->list_lock
)
1171 struct bdi_writeback
*wb
= inode_to_wb(inode
);
1173 spin_unlock(&inode
->i_lock
);
1174 spin_lock(&wb
->list_lock
);
1178 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
1179 __acquires(&wb
->list_lock
)
1181 struct bdi_writeback
*wb
= inode_to_wb(inode
);
1183 spin_lock(&wb
->list_lock
);
1187 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
1192 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
1193 struct wb_writeback_work
*base_work
,
1198 if (!skip_if_busy
|| !writeback_in_progress(&bdi
->wb
)) {
1199 base_work
->auto_free
= 0;
1200 wb_queue_work(&bdi
->wb
, base_work
);
1204 #endif /* CONFIG_CGROUP_WRITEBACK */
1207 * Add in the number of potentially dirty inodes, because each inode
1208 * write can dirty pagecache in the underlying blockdev.
1210 static unsigned long get_nr_dirty_pages(void)
1212 return global_node_page_state(NR_FILE_DIRTY
) +
1213 get_nr_dirty_inodes();
1216 static void wb_start_writeback(struct bdi_writeback
*wb
, enum wb_reason reason
)
1218 if (!wb_has_dirty_io(wb
))
1222 * All callers of this function want to start writeback of all
1223 * dirty pages. Places like vmscan can call this at a very
1224 * high frequency, causing pointless allocations of tons of
1225 * work items and keeping the flusher threads busy retrieving
1226 * that work. Ensure that we only allow one of them pending and
1227 * inflight at the time.
1229 if (test_bit(WB_start_all
, &wb
->state
) ||
1230 test_and_set_bit(WB_start_all
, &wb
->state
))
1233 wb
->start_all_reason
= reason
;
1238 * wb_start_background_writeback - start background writeback
1239 * @wb: bdi_writback to write from
1242 * This makes sure WB_SYNC_NONE background writeback happens. When
1243 * this function returns, it is only guaranteed that for given wb
1244 * some IO is happening if we are over background dirty threshold.
1245 * Caller need not hold sb s_umount semaphore.
1247 void wb_start_background_writeback(struct bdi_writeback
*wb
)
1250 * We just wake up the flusher thread. It will perform background
1251 * writeback as soon as there is no other work to do.
1253 trace_writeback_wake_background(wb
);
1258 * Remove the inode from the writeback list it is on.
1260 void inode_io_list_del(struct inode
*inode
)
1262 struct bdi_writeback
*wb
;
1264 wb
= inode_to_wb_and_lock_list(inode
);
1265 spin_lock(&inode
->i_lock
);
1267 inode
->i_state
&= ~I_SYNC_QUEUED
;
1268 list_del_init(&inode
->i_io_list
);
1269 wb_io_lists_depopulated(wb
);
1271 spin_unlock(&inode
->i_lock
);
1272 spin_unlock(&wb
->list_lock
);
1274 EXPORT_SYMBOL(inode_io_list_del
);
1277 * mark an inode as under writeback on the sb
1279 void sb_mark_inode_writeback(struct inode
*inode
)
1281 struct super_block
*sb
= inode
->i_sb
;
1282 unsigned long flags
;
1284 if (list_empty(&inode
->i_wb_list
)) {
1285 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1286 if (list_empty(&inode
->i_wb_list
)) {
1287 list_add_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
1288 trace_sb_mark_inode_writeback(inode
);
1290 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1295 * clear an inode as under writeback on the sb
1297 void sb_clear_inode_writeback(struct inode
*inode
)
1299 struct super_block
*sb
= inode
->i_sb
;
1300 unsigned long flags
;
1302 if (!list_empty(&inode
->i_wb_list
)) {
1303 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1304 if (!list_empty(&inode
->i_wb_list
)) {
1305 list_del_init(&inode
->i_wb_list
);
1306 trace_sb_clear_inode_writeback(inode
);
1308 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1313 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1314 * furthest end of its superblock's dirty-inode list.
1316 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1317 * already the most-recently-dirtied inode on the b_dirty list. If that is
1318 * the case then the inode must have been redirtied while it was being written
1319 * out and we don't reset its dirtied_when.
1321 static void redirty_tail_locked(struct inode
*inode
, struct bdi_writeback
*wb
)
1323 assert_spin_locked(&inode
->i_lock
);
1325 if (!list_empty(&wb
->b_dirty
)) {
1328 tail
= wb_inode(wb
->b_dirty
.next
);
1329 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
1330 inode
->dirtied_when
= jiffies
;
1332 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty
);
1333 inode
->i_state
&= ~I_SYNC_QUEUED
;
1336 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
1338 spin_lock(&inode
->i_lock
);
1339 redirty_tail_locked(inode
, wb
);
1340 spin_unlock(&inode
->i_lock
);
1344 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1346 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
1348 inode_io_list_move_locked(inode
, wb
, &wb
->b_more_io
);
1351 static void inode_sync_complete(struct inode
*inode
)
1353 inode
->i_state
&= ~I_SYNC
;
1354 /* If inode is clean an unused, put it into LRU now... */
1355 inode_add_lru(inode
);
1356 /* Waiters must see I_SYNC cleared before being woken up */
1358 wake_up_bit(&inode
->i_state
, __I_SYNC
);
1361 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
1363 bool ret
= time_after(inode
->dirtied_when
, t
);
1364 #ifndef CONFIG_64BIT
1366 * For inodes being constantly redirtied, dirtied_when can get stuck.
1367 * It _appears_ to be in the future, but is actually in distant past.
1368 * This test is necessary to prevent such wrapped-around relative times
1369 * from permanently stopping the whole bdi writeback.
1371 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
1376 #define EXPIRE_DIRTY_ATIME 0x0001
1379 * Move expired (dirtied before dirtied_before) dirty inodes from
1380 * @delaying_queue to @dispatch_queue.
1382 static int move_expired_inodes(struct list_head
*delaying_queue
,
1383 struct list_head
*dispatch_queue
,
1384 unsigned long dirtied_before
)
1387 struct list_head
*pos
, *node
;
1388 struct super_block
*sb
= NULL
;
1389 struct inode
*inode
;
1393 while (!list_empty(delaying_queue
)) {
1394 inode
= wb_inode(delaying_queue
->prev
);
1395 if (inode_dirtied_after(inode
, dirtied_before
))
1397 list_move(&inode
->i_io_list
, &tmp
);
1399 spin_lock(&inode
->i_lock
);
1400 inode
->i_state
|= I_SYNC_QUEUED
;
1401 spin_unlock(&inode
->i_lock
);
1402 if (sb_is_blkdev_sb(inode
->i_sb
))
1404 if (sb
&& sb
!= inode
->i_sb
)
1409 /* just one sb in list, splice to dispatch_queue and we're done */
1411 list_splice(&tmp
, dispatch_queue
);
1415 /* Move inodes from one superblock together */
1416 while (!list_empty(&tmp
)) {
1417 sb
= wb_inode(tmp
.prev
)->i_sb
;
1418 list_for_each_prev_safe(pos
, node
, &tmp
) {
1419 inode
= wb_inode(pos
);
1420 if (inode
->i_sb
== sb
)
1421 list_move(&inode
->i_io_list
, dispatch_queue
);
1429 * Queue all expired dirty inodes for io, eldest first.
1431 * newly dirtied b_dirty b_io b_more_io
1432 * =============> gf edc BA
1434 * newly dirtied b_dirty b_io b_more_io
1435 * =============> g fBAedc
1437 * +--> dequeue for IO
1439 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
,
1440 unsigned long dirtied_before
)
1443 unsigned long time_expire_jif
= dirtied_before
;
1445 assert_spin_locked(&wb
->list_lock
);
1446 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
1447 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, dirtied_before
);
1448 if (!work
->for_sync
)
1449 time_expire_jif
= jiffies
- dirtytime_expire_interval
* HZ
;
1450 moved
+= move_expired_inodes(&wb
->b_dirty_time
, &wb
->b_io
,
1453 wb_io_lists_populated(wb
);
1454 trace_writeback_queue_io(wb
, work
, dirtied_before
, moved
);
1457 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1461 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
1462 trace_writeback_write_inode_start(inode
, wbc
);
1463 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
1464 trace_writeback_write_inode(inode
, wbc
);
1471 * Wait for writeback on an inode to complete. Called with i_lock held.
1472 * Caller must make sure inode cannot go away when we drop i_lock.
1474 static void __inode_wait_for_writeback(struct inode
*inode
)
1475 __releases(inode
->i_lock
)
1476 __acquires(inode
->i_lock
)
1478 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
1479 wait_queue_head_t
*wqh
;
1481 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1482 while (inode
->i_state
& I_SYNC
) {
1483 spin_unlock(&inode
->i_lock
);
1484 __wait_on_bit(wqh
, &wq
, bit_wait
,
1485 TASK_UNINTERRUPTIBLE
);
1486 spin_lock(&inode
->i_lock
);
1491 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1493 void inode_wait_for_writeback(struct inode
*inode
)
1495 spin_lock(&inode
->i_lock
);
1496 __inode_wait_for_writeback(inode
);
1497 spin_unlock(&inode
->i_lock
);
1501 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1502 * held and drops it. It is aimed for callers not holding any inode reference
1503 * so once i_lock is dropped, inode can go away.
1505 static void inode_sleep_on_writeback(struct inode
*inode
)
1506 __releases(inode
->i_lock
)
1509 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1512 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
1513 sleep
= inode
->i_state
& I_SYNC
;
1514 spin_unlock(&inode
->i_lock
);
1517 finish_wait(wqh
, &wait
);
1521 * Find proper writeback list for the inode depending on its current state and
1522 * possibly also change of its state while we were doing writeback. Here we
1523 * handle things such as livelock prevention or fairness of writeback among
1524 * inodes. This function can be called only by flusher thread - noone else
1525 * processes all inodes in writeback lists and requeueing inodes behind flusher
1526 * thread's back can have unexpected consequences.
1528 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
1529 struct writeback_control
*wbc
)
1531 if (inode
->i_state
& I_FREEING
)
1535 * Sync livelock prevention. Each inode is tagged and synced in one
1536 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1537 * the dirty time to prevent enqueue and sync it again.
1539 if ((inode
->i_state
& I_DIRTY
) &&
1540 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
1541 inode
->dirtied_when
= jiffies
;
1543 if (wbc
->pages_skipped
) {
1545 * writeback is not making progress due to locked
1546 * buffers. Skip this inode for now.
1548 redirty_tail_locked(inode
, wb
);
1552 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
1554 * We didn't write back all the pages. nfs_writepages()
1555 * sometimes bales out without doing anything.
1557 if (wbc
->nr_to_write
<= 0) {
1558 /* Slice used up. Queue for next turn. */
1559 requeue_io(inode
, wb
);
1562 * Writeback blocked by something other than
1563 * congestion. Delay the inode for some time to
1564 * avoid spinning on the CPU (100% iowait)
1565 * retrying writeback of the dirty page/inode
1566 * that cannot be performed immediately.
1568 redirty_tail_locked(inode
, wb
);
1570 } else if (inode
->i_state
& I_DIRTY
) {
1572 * Filesystems can dirty the inode during writeback operations,
1573 * such as delayed allocation during submission or metadata
1574 * updates after data IO completion.
1576 redirty_tail_locked(inode
, wb
);
1577 } else if (inode
->i_state
& I_DIRTY_TIME
) {
1578 inode
->dirtied_when
= jiffies
;
1579 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty_time
);
1580 inode
->i_state
&= ~I_SYNC_QUEUED
;
1582 /* The inode is clean. Remove from writeback lists. */
1583 inode_cgwb_move_to_attached(inode
, wb
);
1588 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1589 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1591 * This doesn't remove the inode from the writeback list it is on, except
1592 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1593 * expiration. The caller is otherwise responsible for writeback list handling.
1595 * The caller is also responsible for setting the I_SYNC flag beforehand and
1596 * calling inode_sync_complete() to clear it afterwards.
1599 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1601 struct address_space
*mapping
= inode
->i_mapping
;
1602 long nr_to_write
= wbc
->nr_to_write
;
1606 WARN_ON(!(inode
->i_state
& I_SYNC
));
1608 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
1610 ret
= do_writepages(mapping
, wbc
);
1613 * Make sure to wait on the data before writing out the metadata.
1614 * This is important for filesystems that modify metadata on data
1615 * I/O completion. We don't do it for sync(2) writeback because it has a
1616 * separate, external IO completion path and ->sync_fs for guaranteeing
1617 * inode metadata is written back correctly.
1619 if (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
) {
1620 int err
= filemap_fdatawait(mapping
);
1626 * If the inode has dirty timestamps and we need to write them, call
1627 * mark_inode_dirty_sync() to notify the filesystem about it and to
1628 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1630 if ((inode
->i_state
& I_DIRTY_TIME
) &&
1631 (wbc
->sync_mode
== WB_SYNC_ALL
||
1632 time_after(jiffies
, inode
->dirtied_time_when
+
1633 dirtytime_expire_interval
* HZ
))) {
1634 trace_writeback_lazytime(inode
);
1635 mark_inode_dirty_sync(inode
);
1639 * Get and clear the dirty flags from i_state. This needs to be done
1640 * after calling writepages because some filesystems may redirty the
1641 * inode during writepages due to delalloc. It also needs to be done
1642 * after handling timestamp expiration, as that may dirty the inode too.
1644 spin_lock(&inode
->i_lock
);
1645 dirty
= inode
->i_state
& I_DIRTY
;
1646 inode
->i_state
&= ~dirty
;
1649 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1650 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1651 * either they see the I_DIRTY bits cleared or we see the dirtied
1654 * I_DIRTY_PAGES is always cleared together above even if @mapping
1655 * still has dirty pages. The flag is reinstated after smp_mb() if
1656 * necessary. This guarantees that either __mark_inode_dirty()
1657 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1661 if (mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
1662 inode
->i_state
|= I_DIRTY_PAGES
;
1664 spin_unlock(&inode
->i_lock
);
1666 /* Don't write the inode if only I_DIRTY_PAGES was set */
1667 if (dirty
& ~I_DIRTY_PAGES
) {
1668 int err
= write_inode(inode
, wbc
);
1672 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
1677 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1678 * the regular batched writeback done by the flusher threads in
1679 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1680 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1682 * To prevent the inode from going away, either the caller must have a reference
1683 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1685 static int writeback_single_inode(struct inode
*inode
,
1686 struct writeback_control
*wbc
)
1688 struct bdi_writeback
*wb
;
1691 spin_lock(&inode
->i_lock
);
1692 if (!atomic_read(&inode
->i_count
))
1693 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
1695 WARN_ON(inode
->i_state
& I_WILL_FREE
);
1697 if (inode
->i_state
& I_SYNC
) {
1699 * Writeback is already running on the inode. For WB_SYNC_NONE,
1700 * that's enough and we can just return. For WB_SYNC_ALL, we
1701 * must wait for the existing writeback to complete, then do
1702 * writeback again if there's anything left.
1704 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
1706 __inode_wait_for_writeback(inode
);
1708 WARN_ON(inode
->i_state
& I_SYNC
);
1710 * If the inode is already fully clean, then there's nothing to do.
1712 * For data-integrity syncs we also need to check whether any pages are
1713 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1714 * there are any such pages, we'll need to wait for them.
1716 if (!(inode
->i_state
& I_DIRTY_ALL
) &&
1717 (wbc
->sync_mode
!= WB_SYNC_ALL
||
1718 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
1720 inode
->i_state
|= I_SYNC
;
1721 wbc_attach_and_unlock_inode(wbc
, inode
);
1723 ret
= __writeback_single_inode(inode
, wbc
);
1725 wbc_detach_inode(wbc
);
1727 wb
= inode_to_wb_and_lock_list(inode
);
1728 spin_lock(&inode
->i_lock
);
1730 * If the inode is now fully clean, then it can be safely removed from
1731 * its writeback list (if any). Otherwise the flusher threads are
1732 * responsible for the writeback lists.
1734 if (!(inode
->i_state
& I_DIRTY_ALL
))
1735 inode_cgwb_move_to_attached(inode
, wb
);
1736 spin_unlock(&wb
->list_lock
);
1737 inode_sync_complete(inode
);
1739 spin_unlock(&inode
->i_lock
);
1743 static long writeback_chunk_size(struct bdi_writeback
*wb
,
1744 struct wb_writeback_work
*work
)
1749 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1750 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1751 * here avoids calling into writeback_inodes_wb() more than once.
1753 * The intended call sequence for WB_SYNC_ALL writeback is:
1756 * writeback_sb_inodes() <== called only once
1757 * write_cache_pages() <== called once for each inode
1758 * (quickly) tag currently dirty pages
1759 * (maybe slowly) sync all tagged pages
1761 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
1764 pages
= min(wb
->avg_write_bandwidth
/ 2,
1765 global_wb_domain
.dirty_limit
/ DIRTY_SCOPE
);
1766 pages
= min(pages
, work
->nr_pages
);
1767 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
1768 MIN_WRITEBACK_PAGES
);
1775 * Write a portion of b_io inodes which belong to @sb.
1777 * Return the number of pages and/or inodes written.
1779 * NOTE! This is called with wb->list_lock held, and will
1780 * unlock and relock that for each inode it ends up doing
1783 static long writeback_sb_inodes(struct super_block
*sb
,
1784 struct bdi_writeback
*wb
,
1785 struct wb_writeback_work
*work
)
1787 struct writeback_control wbc
= {
1788 .sync_mode
= work
->sync_mode
,
1789 .tagged_writepages
= work
->tagged_writepages
,
1790 .for_kupdate
= work
->for_kupdate
,
1791 .for_background
= work
->for_background
,
1792 .for_sync
= work
->for_sync
,
1793 .range_cyclic
= work
->range_cyclic
,
1795 .range_end
= LLONG_MAX
,
1797 unsigned long start_time
= jiffies
;
1799 long wrote
= 0; /* count both pages and inodes */
1801 while (!list_empty(&wb
->b_io
)) {
1802 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1803 struct bdi_writeback
*tmp_wb
;
1805 if (inode
->i_sb
!= sb
) {
1808 * We only want to write back data for this
1809 * superblock, move all inodes not belonging
1810 * to it back onto the dirty list.
1812 redirty_tail(inode
, wb
);
1817 * The inode belongs to a different superblock.
1818 * Bounce back to the caller to unpin this and
1819 * pin the next superblock.
1825 * Don't bother with new inodes or inodes being freed, first
1826 * kind does not need periodic writeout yet, and for the latter
1827 * kind writeout is handled by the freer.
1829 spin_lock(&inode
->i_lock
);
1830 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
1831 redirty_tail_locked(inode
, wb
);
1832 spin_unlock(&inode
->i_lock
);
1835 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
1837 * If this inode is locked for writeback and we are not
1838 * doing writeback-for-data-integrity, move it to
1839 * b_more_io so that writeback can proceed with the
1840 * other inodes on s_io.
1842 * We'll have another go at writing back this inode
1843 * when we completed a full scan of b_io.
1845 spin_unlock(&inode
->i_lock
);
1846 requeue_io(inode
, wb
);
1847 trace_writeback_sb_inodes_requeue(inode
);
1850 spin_unlock(&wb
->list_lock
);
1853 * We already requeued the inode if it had I_SYNC set and we
1854 * are doing WB_SYNC_NONE writeback. So this catches only the
1857 if (inode
->i_state
& I_SYNC
) {
1858 /* Wait for I_SYNC. This function drops i_lock... */
1859 inode_sleep_on_writeback(inode
);
1860 /* Inode may be gone, start again */
1861 spin_lock(&wb
->list_lock
);
1864 inode
->i_state
|= I_SYNC
;
1865 wbc_attach_and_unlock_inode(&wbc
, inode
);
1867 write_chunk
= writeback_chunk_size(wb
, work
);
1868 wbc
.nr_to_write
= write_chunk
;
1869 wbc
.pages_skipped
= 0;
1872 * We use I_SYNC to pin the inode in memory. While it is set
1873 * evict_inode() will wait so the inode cannot be freed.
1875 __writeback_single_inode(inode
, &wbc
);
1877 wbc_detach_inode(&wbc
);
1878 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
1879 wrote
+= write_chunk
- wbc
.nr_to_write
;
1881 if (need_resched()) {
1883 * We're trying to balance between building up a nice
1884 * long list of IOs to improve our merge rate, and
1885 * getting those IOs out quickly for anyone throttling
1886 * in balance_dirty_pages(). cond_resched() doesn't
1887 * unplug, so get our IOs out the door before we
1890 blk_flush_plug(current
);
1895 * Requeue @inode if still dirty. Be careful as @inode may
1896 * have been switched to another wb in the meantime.
1898 tmp_wb
= inode_to_wb_and_lock_list(inode
);
1899 spin_lock(&inode
->i_lock
);
1900 if (!(inode
->i_state
& I_DIRTY_ALL
))
1902 requeue_inode(inode
, tmp_wb
, &wbc
);
1903 inode_sync_complete(inode
);
1904 spin_unlock(&inode
->i_lock
);
1906 if (unlikely(tmp_wb
!= wb
)) {
1907 spin_unlock(&tmp_wb
->list_lock
);
1908 spin_lock(&wb
->list_lock
);
1912 * bail out to wb_writeback() often enough to check
1913 * background threshold and other termination conditions.
1916 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1918 if (work
->nr_pages
<= 0)
1925 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
1926 struct wb_writeback_work
*work
)
1928 unsigned long start_time
= jiffies
;
1931 while (!list_empty(&wb
->b_io
)) {
1932 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1933 struct super_block
*sb
= inode
->i_sb
;
1935 if (!trylock_super(sb
)) {
1937 * trylock_super() may fail consistently due to
1938 * s_umount being grabbed by someone else. Don't use
1939 * requeue_io() to avoid busy retrying the inode/sb.
1941 redirty_tail(inode
, wb
);
1944 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
1945 up_read(&sb
->s_umount
);
1947 /* refer to the same tests at the end of writeback_sb_inodes */
1949 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1951 if (work
->nr_pages
<= 0)
1955 /* Leave any unwritten inodes on b_io */
1959 static long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
1960 enum wb_reason reason
)
1962 struct wb_writeback_work work
= {
1963 .nr_pages
= nr_pages
,
1964 .sync_mode
= WB_SYNC_NONE
,
1968 struct blk_plug plug
;
1970 blk_start_plug(&plug
);
1971 spin_lock(&wb
->list_lock
);
1972 if (list_empty(&wb
->b_io
))
1973 queue_io(wb
, &work
, jiffies
);
1974 __writeback_inodes_wb(wb
, &work
);
1975 spin_unlock(&wb
->list_lock
);
1976 blk_finish_plug(&plug
);
1978 return nr_pages
- work
.nr_pages
;
1982 * Explicit flushing or periodic writeback of "old" data.
1984 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1985 * dirtying-time in the inode's address_space. So this periodic writeback code
1986 * just walks the superblock inode list, writing back any inodes which are
1987 * older than a specific point in time.
1989 * Try to run once per dirty_writeback_interval. But if a writeback event
1990 * takes longer than a dirty_writeback_interval interval, then leave a
1993 * dirtied_before takes precedence over nr_to_write. So we'll only write back
1994 * all dirty pages if they are all attached to "old" mappings.
1996 static long wb_writeback(struct bdi_writeback
*wb
,
1997 struct wb_writeback_work
*work
)
1999 unsigned long wb_start
= jiffies
;
2000 long nr_pages
= work
->nr_pages
;
2001 unsigned long dirtied_before
= jiffies
;
2002 struct inode
*inode
;
2004 struct blk_plug plug
;
2006 blk_start_plug(&plug
);
2007 spin_lock(&wb
->list_lock
);
2010 * Stop writeback when nr_pages has been consumed
2012 if (work
->nr_pages
<= 0)
2016 * Background writeout and kupdate-style writeback may
2017 * run forever. Stop them if there is other work to do
2018 * so that e.g. sync can proceed. They'll be restarted
2019 * after the other works are all done.
2021 if ((work
->for_background
|| work
->for_kupdate
) &&
2022 !list_empty(&wb
->work_list
))
2026 * For background writeout, stop when we are below the
2027 * background dirty threshold
2029 if (work
->for_background
&& !wb_over_bg_thresh(wb
))
2033 * Kupdate and background works are special and we want to
2034 * include all inodes that need writing. Livelock avoidance is
2035 * handled by these works yielding to any other work so we are
2038 if (work
->for_kupdate
) {
2039 dirtied_before
= jiffies
-
2040 msecs_to_jiffies(dirty_expire_interval
* 10);
2041 } else if (work
->for_background
)
2042 dirtied_before
= jiffies
;
2044 trace_writeback_start(wb
, work
);
2045 if (list_empty(&wb
->b_io
))
2046 queue_io(wb
, work
, dirtied_before
);
2048 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
2050 progress
= __writeback_inodes_wb(wb
, work
);
2051 trace_writeback_written(wb
, work
);
2053 wb_update_bandwidth(wb
, wb_start
);
2056 * Did we write something? Try for more
2058 * Dirty inodes are moved to b_io for writeback in batches.
2059 * The completion of the current batch does not necessarily
2060 * mean the overall work is done. So we keep looping as long
2061 * as made some progress on cleaning pages or inodes.
2066 * No more inodes for IO, bail
2068 if (list_empty(&wb
->b_more_io
))
2071 * Nothing written. Wait for some inode to
2072 * become available for writeback. Otherwise
2073 * we'll just busyloop.
2075 trace_writeback_wait(wb
, work
);
2076 inode
= wb_inode(wb
->b_more_io
.prev
);
2077 spin_lock(&inode
->i_lock
);
2078 spin_unlock(&wb
->list_lock
);
2079 /* This function drops i_lock... */
2080 inode_sleep_on_writeback(inode
);
2081 spin_lock(&wb
->list_lock
);
2083 spin_unlock(&wb
->list_lock
);
2084 blk_finish_plug(&plug
);
2086 return nr_pages
- work
->nr_pages
;
2090 * Return the next wb_writeback_work struct that hasn't been processed yet.
2092 static struct wb_writeback_work
*get_next_work_item(struct bdi_writeback
*wb
)
2094 struct wb_writeback_work
*work
= NULL
;
2096 spin_lock_bh(&wb
->work_lock
);
2097 if (!list_empty(&wb
->work_list
)) {
2098 work
= list_entry(wb
->work_list
.next
,
2099 struct wb_writeback_work
, list
);
2100 list_del_init(&work
->list
);
2102 spin_unlock_bh(&wb
->work_lock
);
2106 static long wb_check_background_flush(struct bdi_writeback
*wb
)
2108 if (wb_over_bg_thresh(wb
)) {
2110 struct wb_writeback_work work
= {
2111 .nr_pages
= LONG_MAX
,
2112 .sync_mode
= WB_SYNC_NONE
,
2113 .for_background
= 1,
2115 .reason
= WB_REASON_BACKGROUND
,
2118 return wb_writeback(wb
, &work
);
2124 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
2126 unsigned long expired
;
2130 * When set to zero, disable periodic writeback
2132 if (!dirty_writeback_interval
)
2135 expired
= wb
->last_old_flush
+
2136 msecs_to_jiffies(dirty_writeback_interval
* 10);
2137 if (time_before(jiffies
, expired
))
2140 wb
->last_old_flush
= jiffies
;
2141 nr_pages
= get_nr_dirty_pages();
2144 struct wb_writeback_work work
= {
2145 .nr_pages
= nr_pages
,
2146 .sync_mode
= WB_SYNC_NONE
,
2149 .reason
= WB_REASON_PERIODIC
,
2152 return wb_writeback(wb
, &work
);
2158 static long wb_check_start_all(struct bdi_writeback
*wb
)
2162 if (!test_bit(WB_start_all
, &wb
->state
))
2165 nr_pages
= get_nr_dirty_pages();
2167 struct wb_writeback_work work
= {
2168 .nr_pages
= wb_split_bdi_pages(wb
, nr_pages
),
2169 .sync_mode
= WB_SYNC_NONE
,
2171 .reason
= wb
->start_all_reason
,
2174 nr_pages
= wb_writeback(wb
, &work
);
2177 clear_bit(WB_start_all
, &wb
->state
);
2183 * Retrieve work items and do the writeback they describe
2185 static long wb_do_writeback(struct bdi_writeback
*wb
)
2187 struct wb_writeback_work
*work
;
2190 set_bit(WB_writeback_running
, &wb
->state
);
2191 while ((work
= get_next_work_item(wb
)) != NULL
) {
2192 trace_writeback_exec(wb
, work
);
2193 wrote
+= wb_writeback(wb
, work
);
2194 finish_writeback_work(wb
, work
);
2198 * Check for a flush-everything request
2200 wrote
+= wb_check_start_all(wb
);
2203 * Check for periodic writeback, kupdated() style
2205 wrote
+= wb_check_old_data_flush(wb
);
2206 wrote
+= wb_check_background_flush(wb
);
2207 clear_bit(WB_writeback_running
, &wb
->state
);
2213 * Handle writeback of dirty data for the device backed by this bdi. Also
2214 * reschedules periodically and does kupdated style flushing.
2216 void wb_workfn(struct work_struct
*work
)
2218 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
2219 struct bdi_writeback
, dwork
);
2222 set_worker_desc("flush-%s", bdi_dev_name(wb
->bdi
));
2223 current
->flags
|= PF_SWAPWRITE
;
2225 if (likely(!current_is_workqueue_rescuer() ||
2226 !test_bit(WB_registered
, &wb
->state
))) {
2228 * The normal path. Keep writing back @wb until its
2229 * work_list is empty. Note that this path is also taken
2230 * if @wb is shutting down even when we're running off the
2231 * rescuer as work_list needs to be drained.
2234 pages_written
= wb_do_writeback(wb
);
2235 trace_writeback_pages_written(pages_written
);
2236 } while (!list_empty(&wb
->work_list
));
2239 * bdi_wq can't get enough workers and we're running off
2240 * the emergency worker. Don't hog it. Hopefully, 1024 is
2241 * enough for efficient IO.
2243 pages_written
= writeback_inodes_wb(wb
, 1024,
2244 WB_REASON_FORKER_THREAD
);
2245 trace_writeback_pages_written(pages_written
);
2248 if (!list_empty(&wb
->work_list
))
2250 else if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
2251 wb_wakeup_delayed(wb
);
2253 current
->flags
&= ~PF_SWAPWRITE
;
2257 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2258 * write back the whole world.
2260 static void __wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2261 enum wb_reason reason
)
2263 struct bdi_writeback
*wb
;
2265 if (!bdi_has_dirty_io(bdi
))
2268 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2269 wb_start_writeback(wb
, reason
);
2272 void wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2273 enum wb_reason reason
)
2276 __wakeup_flusher_threads_bdi(bdi
, reason
);
2281 * Wakeup the flusher threads to start writeback of all currently dirty pages
2283 void wakeup_flusher_threads(enum wb_reason reason
)
2285 struct backing_dev_info
*bdi
;
2288 * If we are expecting writeback progress we must submit plugged IO.
2290 if (blk_needs_flush_plug(current
))
2291 blk_schedule_flush_plug(current
);
2294 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
2295 __wakeup_flusher_threads_bdi(bdi
, reason
);
2300 * Wake up bdi's periodically to make sure dirtytime inodes gets
2301 * written back periodically. We deliberately do *not* check the
2302 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2303 * kernel to be constantly waking up once there are any dirtytime
2304 * inodes on the system. So instead we define a separate delayed work
2305 * function which gets called much more rarely. (By default, only
2306 * once every 12 hours.)
2308 * If there is any other write activity going on in the file system,
2309 * this function won't be necessary. But if the only thing that has
2310 * happened on the file system is a dirtytime inode caused by an atime
2311 * update, we need this infrastructure below to make sure that inode
2312 * eventually gets pushed out to disk.
2314 static void wakeup_dirtytime_writeback(struct work_struct
*w
);
2315 static DECLARE_DELAYED_WORK(dirtytime_work
, wakeup_dirtytime_writeback
);
2317 static void wakeup_dirtytime_writeback(struct work_struct
*w
)
2319 struct backing_dev_info
*bdi
;
2322 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
2323 struct bdi_writeback
*wb
;
2325 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2326 if (!list_empty(&wb
->b_dirty_time
))
2330 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2333 static int __init
start_dirtytime_writeback(void)
2335 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2338 __initcall(start_dirtytime_writeback
);
2340 int dirtytime_interval_handler(struct ctl_table
*table
, int write
,
2341 void *buffer
, size_t *lenp
, loff_t
*ppos
)
2345 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
2346 if (ret
== 0 && write
)
2347 mod_delayed_work(system_wq
, &dirtytime_work
, 0);
2352 * __mark_inode_dirty - internal function to mark an inode dirty
2354 * @inode: inode to mark
2355 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2356 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2357 * with I_DIRTY_PAGES.
2359 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2360 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2362 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2363 * instead of calling this directly.
2365 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2366 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2367 * even if they are later hashed, as they will have been marked dirty already.
2369 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2371 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2372 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2373 * the kernel-internal blockdev inode represents the dirtying time of the
2374 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2375 * page->mapping->host, so the page-dirtying time is recorded in the internal
2378 void __mark_inode_dirty(struct inode
*inode
, int flags
)
2380 struct super_block
*sb
= inode
->i_sb
;
2383 trace_writeback_mark_inode_dirty(inode
, flags
);
2385 if (flags
& I_DIRTY_INODE
) {
2387 * Notify the filesystem about the inode being dirtied, so that
2388 * (if needed) it can update on-disk fields and journal the
2389 * inode. This is only needed when the inode itself is being
2390 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2391 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2393 trace_writeback_dirty_inode_start(inode
, flags
);
2394 if (sb
->s_op
->dirty_inode
)
2395 sb
->s_op
->dirty_inode(inode
, flags
& I_DIRTY_INODE
);
2396 trace_writeback_dirty_inode(inode
, flags
);
2398 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2399 flags
&= ~I_DIRTY_TIME
;
2402 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2403 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2404 * in one call to __mark_inode_dirty().)
2406 dirtytime
= flags
& I_DIRTY_TIME
;
2407 WARN_ON_ONCE(dirtytime
&& flags
!= I_DIRTY_TIME
);
2411 * Paired with smp_mb() in __writeback_single_inode() for the
2412 * following lockless i_state test. See there for details.
2416 if (((inode
->i_state
& flags
) == flags
) ||
2417 (dirtytime
&& (inode
->i_state
& I_DIRTY_INODE
)))
2420 spin_lock(&inode
->i_lock
);
2421 if (dirtytime
&& (inode
->i_state
& I_DIRTY_INODE
))
2422 goto out_unlock_inode
;
2423 if ((inode
->i_state
& flags
) != flags
) {
2424 const int was_dirty
= inode
->i_state
& I_DIRTY
;
2426 inode_attach_wb(inode
, NULL
);
2428 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2429 if (flags
& I_DIRTY_INODE
)
2430 inode
->i_state
&= ~I_DIRTY_TIME
;
2431 inode
->i_state
|= flags
;
2434 * If the inode is queued for writeback by flush worker, just
2435 * update its dirty state. Once the flush worker is done with
2436 * the inode it will place it on the appropriate superblock
2437 * list, based upon its state.
2439 if (inode
->i_state
& I_SYNC_QUEUED
)
2440 goto out_unlock_inode
;
2443 * Only add valid (hashed) inodes to the superblock's
2444 * dirty list. Add blockdev inodes as well.
2446 if (!S_ISBLK(inode
->i_mode
)) {
2447 if (inode_unhashed(inode
))
2448 goto out_unlock_inode
;
2450 if (inode
->i_state
& I_FREEING
)
2451 goto out_unlock_inode
;
2454 * If the inode was already on b_dirty/b_io/b_more_io, don't
2455 * reposition it (that would break b_dirty time-ordering).
2458 struct bdi_writeback
*wb
;
2459 struct list_head
*dirty_list
;
2460 bool wakeup_bdi
= false;
2462 wb
= locked_inode_to_wb_and_lock_list(inode
);
2464 inode
->dirtied_when
= jiffies
;
2466 inode
->dirtied_time_when
= jiffies
;
2468 if (inode
->i_state
& I_DIRTY
)
2469 dirty_list
= &wb
->b_dirty
;
2471 dirty_list
= &wb
->b_dirty_time
;
2473 wakeup_bdi
= inode_io_list_move_locked(inode
, wb
,
2476 spin_unlock(&wb
->list_lock
);
2477 trace_writeback_dirty_inode_enqueue(inode
);
2480 * If this is the first dirty inode for this bdi,
2481 * we have to wake-up the corresponding bdi thread
2482 * to make sure background write-back happens
2486 (wb
->bdi
->capabilities
& BDI_CAP_WRITEBACK
))
2487 wb_wakeup_delayed(wb
);
2492 spin_unlock(&inode
->i_lock
);
2494 EXPORT_SYMBOL(__mark_inode_dirty
);
2497 * The @s_sync_lock is used to serialise concurrent sync operations
2498 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2499 * Concurrent callers will block on the s_sync_lock rather than doing contending
2500 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2501 * has been issued up to the time this function is enter is guaranteed to be
2502 * completed by the time we have gained the lock and waited for all IO that is
2503 * in progress regardless of the order callers are granted the lock.
2505 static void wait_sb_inodes(struct super_block
*sb
)
2507 LIST_HEAD(sync_list
);
2510 * We need to be protected against the filesystem going from
2511 * r/o to r/w or vice versa.
2513 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2515 mutex_lock(&sb
->s_sync_lock
);
2518 * Splice the writeback list onto a temporary list to avoid waiting on
2519 * inodes that have started writeback after this point.
2521 * Use rcu_read_lock() to keep the inodes around until we have a
2522 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2523 * the local list because inodes can be dropped from either by writeback
2527 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2528 list_splice_init(&sb
->s_inodes_wb
, &sync_list
);
2531 * Data integrity sync. Must wait for all pages under writeback, because
2532 * there may have been pages dirtied before our sync call, but which had
2533 * writeout started before we write it out. In which case, the inode
2534 * may not be on the dirty list, but we still have to wait for that
2537 while (!list_empty(&sync_list
)) {
2538 struct inode
*inode
= list_first_entry(&sync_list
, struct inode
,
2540 struct address_space
*mapping
= inode
->i_mapping
;
2543 * Move each inode back to the wb list before we drop the lock
2544 * to preserve consistency between i_wb_list and the mapping
2545 * writeback tag. Writeback completion is responsible to remove
2546 * the inode from either list once the writeback tag is cleared.
2548 list_move_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
2551 * The mapping can appear untagged while still on-list since we
2552 * do not have the mapping lock. Skip it here, wb completion
2555 if (!mapping_tagged(mapping
, PAGECACHE_TAG_WRITEBACK
))
2558 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2560 spin_lock(&inode
->i_lock
);
2561 if (inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) {
2562 spin_unlock(&inode
->i_lock
);
2564 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2568 spin_unlock(&inode
->i_lock
);
2572 * We keep the error status of individual mapping so that
2573 * applications can catch the writeback error using fsync(2).
2574 * See filemap_fdatawait_keep_errors() for details.
2576 filemap_fdatawait_keep_errors(mapping
);
2583 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2585 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2587 mutex_unlock(&sb
->s_sync_lock
);
2590 static void __writeback_inodes_sb_nr(struct super_block
*sb
, unsigned long nr
,
2591 enum wb_reason reason
, bool skip_if_busy
)
2593 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2594 DEFINE_WB_COMPLETION(done
, bdi
);
2595 struct wb_writeback_work work
= {
2597 .sync_mode
= WB_SYNC_NONE
,
2598 .tagged_writepages
= 1,
2604 if (!bdi_has_dirty_io(bdi
) || bdi
== &noop_backing_dev_info
)
2606 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2608 bdi_split_work_to_wbs(sb
->s_bdi
, &work
, skip_if_busy
);
2609 wb_wait_for_completion(&done
);
2613 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2614 * @sb: the superblock
2615 * @nr: the number of pages to write
2616 * @reason: reason why some writeback work initiated
2618 * Start writeback on some inodes on this super_block. No guarantees are made
2619 * on how many (if any) will be written, and this function does not wait
2620 * for IO completion of submitted IO.
2622 void writeback_inodes_sb_nr(struct super_block
*sb
,
2624 enum wb_reason reason
)
2626 __writeback_inodes_sb_nr(sb
, nr
, reason
, false);
2628 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
2631 * writeback_inodes_sb - writeback dirty inodes from given super_block
2632 * @sb: the superblock
2633 * @reason: reason why some writeback work was initiated
2635 * Start writeback on some inodes on this super_block. No guarantees are made
2636 * on how many (if any) will be written, and this function does not wait
2637 * for IO completion of submitted IO.
2639 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2641 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
2643 EXPORT_SYMBOL(writeback_inodes_sb
);
2646 * try_to_writeback_inodes_sb - try to start writeback if none underway
2647 * @sb: the superblock
2648 * @reason: reason why some writeback work was initiated
2650 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2652 void try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2654 if (!down_read_trylock(&sb
->s_umount
))
2657 __writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
, true);
2658 up_read(&sb
->s_umount
);
2660 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
2663 * sync_inodes_sb - sync sb inode pages
2664 * @sb: the superblock
2666 * This function writes and waits on any dirty inode belonging to this
2669 void sync_inodes_sb(struct super_block
*sb
)
2671 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2672 DEFINE_WB_COMPLETION(done
, bdi
);
2673 struct wb_writeback_work work
= {
2675 .sync_mode
= WB_SYNC_ALL
,
2676 .nr_pages
= LONG_MAX
,
2679 .reason
= WB_REASON_SYNC
,
2684 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2685 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2686 * bdi_has_dirty() need to be written out too.
2688 if (bdi
== &noop_backing_dev_info
)
2690 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2692 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2693 bdi_down_write_wb_switch_rwsem(bdi
);
2694 bdi_split_work_to_wbs(bdi
, &work
, false);
2695 wb_wait_for_completion(&done
);
2696 bdi_up_write_wb_switch_rwsem(bdi
);
2700 EXPORT_SYMBOL(sync_inodes_sb
);
2703 * write_inode_now - write an inode to disk
2704 * @inode: inode to write to disk
2705 * @sync: whether the write should be synchronous or not
2707 * This function commits an inode to disk immediately if it is dirty. This is
2708 * primarily needed by knfsd.
2710 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2712 int write_inode_now(struct inode
*inode
, int sync
)
2714 struct writeback_control wbc
= {
2715 .nr_to_write
= LONG_MAX
,
2716 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2718 .range_end
= LLONG_MAX
,
2721 if (!mapping_can_writeback(inode
->i_mapping
))
2722 wbc
.nr_to_write
= 0;
2725 return writeback_single_inode(inode
, &wbc
);
2727 EXPORT_SYMBOL(write_inode_now
);
2730 * sync_inode - write an inode and its pages to disk.
2731 * @inode: the inode to sync
2732 * @wbc: controls the writeback mode
2734 * sync_inode() will write an inode and its pages to disk. It will also
2735 * correctly update the inode on its superblock's dirty inode lists and will
2736 * update inode->i_state.
2738 * The caller must have a ref on the inode.
2740 int sync_inode(struct inode
*inode
, struct writeback_control
*wbc
)
2742 return writeback_single_inode(inode
, wbc
);
2744 EXPORT_SYMBOL(sync_inode
);
2747 * sync_inode_metadata - write an inode to disk
2748 * @inode: the inode to sync
2749 * @wait: wait for I/O to complete.
2751 * Write an inode to disk and adjust its dirty state after completion.
2753 * Note: only writes the actual inode, no associated data or other metadata.
2755 int sync_inode_metadata(struct inode
*inode
, int wait
)
2757 struct writeback_control wbc
= {
2758 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2759 .nr_to_write
= 0, /* metadata-only */
2762 return sync_inode(inode
, &wbc
);
2764 EXPORT_SYMBOL(sync_inode_metadata
);