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
);
124 WARN_ON_ONCE(inode
->i_state
& I_FREEING
);
126 list_move(&inode
->i_io_list
, head
);
128 /* dirty_time doesn't count as dirty_io until expiration */
129 if (head
!= &wb
->b_dirty_time
)
130 return wb_io_lists_populated(wb
);
132 wb_io_lists_depopulated(wb
);
136 static void wb_wakeup(struct bdi_writeback
*wb
)
138 spin_lock_irq(&wb
->work_lock
);
139 if (test_bit(WB_registered
, &wb
->state
))
140 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
141 spin_unlock_irq(&wb
->work_lock
);
144 static void finish_writeback_work(struct bdi_writeback
*wb
,
145 struct wb_writeback_work
*work
)
147 struct wb_completion
*done
= work
->done
;
152 wait_queue_head_t
*waitq
= done
->waitq
;
154 /* @done can't be accessed after the following dec */
155 if (atomic_dec_and_test(&done
->cnt
))
160 static void wb_queue_work(struct bdi_writeback
*wb
,
161 struct wb_writeback_work
*work
)
163 trace_writeback_queue(wb
, work
);
166 atomic_inc(&work
->done
->cnt
);
168 spin_lock_irq(&wb
->work_lock
);
170 if (test_bit(WB_registered
, &wb
->state
)) {
171 list_add_tail(&work
->list
, &wb
->work_list
);
172 mod_delayed_work(bdi_wq
, &wb
->dwork
, 0);
174 finish_writeback_work(wb
, work
);
176 spin_unlock_irq(&wb
->work_lock
);
180 * wb_wait_for_completion - wait for completion of bdi_writeback_works
181 * @done: target wb_completion
183 * Wait for one or more work items issued to @bdi with their ->done field
184 * set to @done, which should have been initialized with
185 * DEFINE_WB_COMPLETION(). This function returns after all such work items
186 * are completed. Work items which are waited upon aren't freed
187 * automatically on completion.
189 void wb_wait_for_completion(struct wb_completion
*done
)
191 atomic_dec(&done
->cnt
); /* put down the initial count */
192 wait_event(*done
->waitq
, !atomic_read(&done
->cnt
));
195 #ifdef CONFIG_CGROUP_WRITEBACK
198 * Parameters for foreign inode detection, see wbc_detach_inode() to see
201 * These paramters are inherently heuristical as the detection target
202 * itself is fuzzy. All we want to do is detaching an inode from the
203 * current owner if it's being written to by some other cgroups too much.
205 * The current cgroup writeback is built on the assumption that multiple
206 * cgroups writing to the same inode concurrently is very rare and a mode
207 * of operation which isn't well supported. As such, the goal is not
208 * taking too long when a different cgroup takes over an inode while
209 * avoiding too aggressive flip-flops from occasional foreign writes.
211 * We record, very roughly, 2s worth of IO time history and if more than
212 * half of that is foreign, trigger the switch. The recording is quantized
213 * to 16 slots. To avoid tiny writes from swinging the decision too much,
214 * writes smaller than 1/8 of avg size are ignored.
216 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
217 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
218 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
219 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
221 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
222 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
223 /* each slot's duration is 2s / 16 */
224 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
225 /* if foreign slots >= 8, switch */
226 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
227 /* one round can affect upto 5 slots */
228 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
231 * Maximum inodes per isw. A specific value has been chosen to make
232 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
234 #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
235 / sizeof(struct inode *))
237 static atomic_t isw_nr_in_flight
= ATOMIC_INIT(0);
238 static struct workqueue_struct
*isw_wq
;
240 void __inode_attach_wb(struct inode
*inode
, struct folio
*folio
)
242 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
243 struct bdi_writeback
*wb
= NULL
;
245 if (inode_cgwb_enabled(inode
)) {
246 struct cgroup_subsys_state
*memcg_css
;
249 memcg_css
= mem_cgroup_css_from_folio(folio
);
250 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
252 /* must pin memcg_css, see wb_get_create() */
253 memcg_css
= task_get_css(current
, memory_cgrp_id
);
254 wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
263 * There may be multiple instances of this function racing to
264 * update the same inode. Use cmpxchg() to tell the winner.
266 if (unlikely(cmpxchg(&inode
->i_wb
, NULL
, wb
)))
269 EXPORT_SYMBOL_GPL(__inode_attach_wb
);
272 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
273 * @inode: inode of interest with i_lock held
274 * @wb: target bdi_writeback
276 * Remove the inode from wb's io lists and if necessarily put onto b_attached
277 * list. Only inodes attached to cgwb's are kept on this list.
279 static void inode_cgwb_move_to_attached(struct inode
*inode
,
280 struct bdi_writeback
*wb
)
282 assert_spin_locked(&wb
->list_lock
);
283 assert_spin_locked(&inode
->i_lock
);
284 WARN_ON_ONCE(inode
->i_state
& I_FREEING
);
286 inode
->i_state
&= ~I_SYNC_QUEUED
;
287 if (wb
!= &wb
->bdi
->wb
)
288 list_move(&inode
->i_io_list
, &wb
->b_attached
);
290 list_del_init(&inode
->i_io_list
);
291 wb_io_lists_depopulated(wb
);
295 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
296 * @inode: inode of interest with i_lock held
298 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
299 * held on entry and is released on return. The returned wb is guaranteed
300 * to stay @inode's associated wb until its list_lock is released.
302 static struct bdi_writeback
*
303 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
304 __releases(&inode
->i_lock
)
305 __acquires(&wb
->list_lock
)
308 struct bdi_writeback
*wb
= inode_to_wb(inode
);
311 * inode_to_wb() association is protected by both
312 * @inode->i_lock and @wb->list_lock but list_lock nests
313 * outside i_lock. Drop i_lock and verify that the
314 * association hasn't changed after acquiring list_lock.
317 spin_unlock(&inode
->i_lock
);
318 spin_lock(&wb
->list_lock
);
320 /* i_wb may have changed inbetween, can't use inode_to_wb() */
321 if (likely(wb
== inode
->i_wb
)) {
322 wb_put(wb
); /* @inode already has ref */
326 spin_unlock(&wb
->list_lock
);
329 spin_lock(&inode
->i_lock
);
334 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
335 * @inode: inode of interest
337 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
340 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
341 __acquires(&wb
->list_lock
)
343 spin_lock(&inode
->i_lock
);
344 return locked_inode_to_wb_and_lock_list(inode
);
347 struct inode_switch_wbs_context
{
348 struct rcu_work work
;
351 * Multiple inodes can be switched at once. The switching procedure
352 * consists of two parts, separated by a RCU grace period. To make
353 * sure that the second part is executed for each inode gone through
354 * the first part, all inode pointers are placed into a NULL-terminated
355 * array embedded into struct inode_switch_wbs_context. Otherwise
356 * an inode could be left in a non-consistent state.
358 struct bdi_writeback
*new_wb
;
359 struct inode
*inodes
[];
362 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
364 down_write(&bdi
->wb_switch_rwsem
);
367 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
)
369 up_write(&bdi
->wb_switch_rwsem
);
372 static bool inode_do_switch_wbs(struct inode
*inode
,
373 struct bdi_writeback
*old_wb
,
374 struct bdi_writeback
*new_wb
)
376 struct address_space
*mapping
= inode
->i_mapping
;
377 XA_STATE(xas
, &mapping
->i_pages
, 0);
379 bool switched
= false;
381 spin_lock(&inode
->i_lock
);
382 xa_lock_irq(&mapping
->i_pages
);
385 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
386 * path owns the inode and we shouldn't modify ->i_io_list.
388 if (unlikely(inode
->i_state
& (I_FREEING
| I_WILL_FREE
)))
391 trace_inode_switch_wbs(inode
, old_wb
, new_wb
);
394 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
395 * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
396 * folios actually under writeback.
398 xas_for_each_marked(&xas
, folio
, ULONG_MAX
, PAGECACHE_TAG_DIRTY
) {
399 if (folio_test_dirty(folio
)) {
400 long nr
= folio_nr_pages(folio
);
401 wb_stat_mod(old_wb
, WB_RECLAIMABLE
, -nr
);
402 wb_stat_mod(new_wb
, WB_RECLAIMABLE
, nr
);
407 xas_for_each_marked(&xas
, folio
, ULONG_MAX
, PAGECACHE_TAG_WRITEBACK
) {
408 long nr
= folio_nr_pages(folio
);
409 WARN_ON_ONCE(!folio_test_writeback(folio
));
410 wb_stat_mod(old_wb
, WB_WRITEBACK
, -nr
);
411 wb_stat_mod(new_wb
, WB_WRITEBACK
, nr
);
414 if (mapping_tagged(mapping
, PAGECACHE_TAG_WRITEBACK
)) {
415 atomic_dec(&old_wb
->writeback_inodes
);
416 atomic_inc(&new_wb
->writeback_inodes
);
422 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
423 * the specific list @inode was on is ignored and the @inode is put on
424 * ->b_dirty which is always correct including from ->b_dirty_time.
425 * The transfer preserves @inode->dirtied_when ordering. If the @inode
426 * was clean, it means it was on the b_attached list, so move it onto
427 * the b_attached list of @new_wb.
429 if (!list_empty(&inode
->i_io_list
)) {
430 inode
->i_wb
= new_wb
;
432 if (inode
->i_state
& I_DIRTY_ALL
) {
435 list_for_each_entry(pos
, &new_wb
->b_dirty
, i_io_list
)
436 if (time_after_eq(inode
->dirtied_when
,
439 inode_io_list_move_locked(inode
, new_wb
,
440 pos
->i_io_list
.prev
);
442 inode_cgwb_move_to_attached(inode
, new_wb
);
445 inode
->i_wb
= new_wb
;
448 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
449 inode
->i_wb_frn_winner
= 0;
450 inode
->i_wb_frn_avg_time
= 0;
451 inode
->i_wb_frn_history
= 0;
455 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
456 * ensures that the new wb is visible if they see !I_WB_SWITCH.
458 smp_store_release(&inode
->i_state
, inode
->i_state
& ~I_WB_SWITCH
);
460 xa_unlock_irq(&mapping
->i_pages
);
461 spin_unlock(&inode
->i_lock
);
466 static void inode_switch_wbs_work_fn(struct work_struct
*work
)
468 struct inode_switch_wbs_context
*isw
=
469 container_of(to_rcu_work(work
), struct inode_switch_wbs_context
, work
);
470 struct backing_dev_info
*bdi
= inode_to_bdi(isw
->inodes
[0]);
471 struct bdi_writeback
*old_wb
= isw
->inodes
[0]->i_wb
;
472 struct bdi_writeback
*new_wb
= isw
->new_wb
;
473 unsigned long nr_switched
= 0;
474 struct inode
**inodep
;
477 * If @inode switches cgwb membership while sync_inodes_sb() is
478 * being issued, sync_inodes_sb() might miss it. Synchronize.
480 down_read(&bdi
->wb_switch_rwsem
);
483 * By the time control reaches here, RCU grace period has passed
484 * since I_WB_SWITCH assertion and all wb stat update transactions
485 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
486 * synchronizing against the i_pages lock.
488 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
489 * gives us exclusion against all wb related operations on @inode
490 * including IO list manipulations and stat updates.
492 if (old_wb
< new_wb
) {
493 spin_lock(&old_wb
->list_lock
);
494 spin_lock_nested(&new_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
496 spin_lock(&new_wb
->list_lock
);
497 spin_lock_nested(&old_wb
->list_lock
, SINGLE_DEPTH_NESTING
);
500 for (inodep
= isw
->inodes
; *inodep
; inodep
++) {
501 WARN_ON_ONCE((*inodep
)->i_wb
!= old_wb
);
502 if (inode_do_switch_wbs(*inodep
, old_wb
, new_wb
))
506 spin_unlock(&new_wb
->list_lock
);
507 spin_unlock(&old_wb
->list_lock
);
509 up_read(&bdi
->wb_switch_rwsem
);
513 wb_put_many(old_wb
, nr_switched
);
516 for (inodep
= isw
->inodes
; *inodep
; inodep
++)
520 atomic_dec(&isw_nr_in_flight
);
523 static bool inode_prepare_wbs_switch(struct inode
*inode
,
524 struct bdi_writeback
*new_wb
)
527 * Paired with smp_mb() in cgroup_writeback_umount().
528 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
529 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
530 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
537 /* while holding I_WB_SWITCH, no one else can update the association */
538 spin_lock(&inode
->i_lock
);
539 if (!(inode
->i_sb
->s_flags
& SB_ACTIVE
) ||
540 inode
->i_state
& (I_WB_SWITCH
| I_FREEING
| I_WILL_FREE
) ||
541 inode_to_wb(inode
) == new_wb
) {
542 spin_unlock(&inode
->i_lock
);
545 inode
->i_state
|= I_WB_SWITCH
;
547 spin_unlock(&inode
->i_lock
);
553 * inode_switch_wbs - change the wb association of an inode
554 * @inode: target inode
555 * @new_wb_id: ID of the new wb
557 * Switch @inode's wb association to the wb identified by @new_wb_id. The
558 * switching is performed asynchronously and may fail silently.
560 static void inode_switch_wbs(struct inode
*inode
, int new_wb_id
)
562 struct backing_dev_info
*bdi
= inode_to_bdi(inode
);
563 struct cgroup_subsys_state
*memcg_css
;
564 struct inode_switch_wbs_context
*isw
;
566 /* noop if seems to be already in progress */
567 if (inode
->i_state
& I_WB_SWITCH
)
570 /* avoid queueing a new switch if too many are already in flight */
571 if (atomic_read(&isw_nr_in_flight
) > WB_FRN_MAX_IN_FLIGHT
)
574 isw
= kzalloc(struct_size(isw
, inodes
, 2), GFP_ATOMIC
);
578 atomic_inc(&isw_nr_in_flight
);
580 /* find and pin the new wb */
582 memcg_css
= css_from_id(new_wb_id
, &memory_cgrp_subsys
);
583 if (memcg_css
&& !css_tryget(memcg_css
))
589 isw
->new_wb
= wb_get_create(bdi
, memcg_css
, GFP_ATOMIC
);
594 if (!inode_prepare_wbs_switch(inode
, isw
->new_wb
))
597 isw
->inodes
[0] = inode
;
600 * In addition to synchronizing among switchers, I_WB_SWITCH tells
601 * the RCU protected stat update paths to grab the i_page
602 * lock so that stat transfer can synchronize against them.
603 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
605 INIT_RCU_WORK(&isw
->work
, inode_switch_wbs_work_fn
);
606 queue_rcu_work(isw_wq
, &isw
->work
);
610 atomic_dec(&isw_nr_in_flight
);
617 * cleanup_offline_cgwb - detach associated inodes
620 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
621 * to eventually release the dying @wb. Returns %true if not all inodes were
622 * switched and the function has to be restarted.
624 bool cleanup_offline_cgwb(struct bdi_writeback
*wb
)
626 struct cgroup_subsys_state
*memcg_css
;
627 struct inode_switch_wbs_context
*isw
;
630 bool restart
= false;
632 isw
= kzalloc(struct_size(isw
, inodes
, WB_MAX_INODES_PER_ISW
),
637 atomic_inc(&isw_nr_in_flight
);
639 for (memcg_css
= wb
->memcg_css
->parent
; memcg_css
;
640 memcg_css
= memcg_css
->parent
) {
641 isw
->new_wb
= wb_get_create(wb
->bdi
, memcg_css
, GFP_KERNEL
);
645 if (unlikely(!isw
->new_wb
))
646 isw
->new_wb
= &wb
->bdi
->wb
; /* wb_get() is noop for bdi's wb */
649 spin_lock(&wb
->list_lock
);
650 list_for_each_entry(inode
, &wb
->b_attached
, i_io_list
) {
651 if (!inode_prepare_wbs_switch(inode
, isw
->new_wb
))
654 isw
->inodes
[nr
++] = inode
;
656 if (nr
>= WB_MAX_INODES_PER_ISW
- 1) {
661 spin_unlock(&wb
->list_lock
);
663 /* no attached inodes? bail out */
665 atomic_dec(&isw_nr_in_flight
);
672 * In addition to synchronizing among switchers, I_WB_SWITCH tells
673 * the RCU protected stat update paths to grab the i_page
674 * lock so that stat transfer can synchronize against them.
675 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
677 INIT_RCU_WORK(&isw
->work
, inode_switch_wbs_work_fn
);
678 queue_rcu_work(isw_wq
, &isw
->work
);
684 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
685 * @wbc: writeback_control of interest
686 * @inode: target inode
688 * @inode is locked and about to be written back under the control of @wbc.
689 * Record @inode's writeback context into @wbc and unlock the i_lock. On
690 * writeback completion, wbc_detach_inode() should be called. This is used
691 * to track the cgroup writeback context.
693 void wbc_attach_and_unlock_inode(struct writeback_control
*wbc
,
696 if (!inode_cgwb_enabled(inode
)) {
697 spin_unlock(&inode
->i_lock
);
701 wbc
->wb
= inode_to_wb(inode
);
704 wbc
->wb_id
= wbc
->wb
->memcg_css
->id
;
705 wbc
->wb_lcand_id
= inode
->i_wb_frn_winner
;
706 wbc
->wb_tcand_id
= 0;
708 wbc
->wb_lcand_bytes
= 0;
709 wbc
->wb_tcand_bytes
= 0;
712 spin_unlock(&inode
->i_lock
);
715 * A dying wb indicates that either the blkcg associated with the
716 * memcg changed or the associated memcg is dying. In the first
717 * case, a replacement wb should already be available and we should
718 * refresh the wb immediately. In the second case, trying to
719 * refresh will keep failing.
721 if (unlikely(wb_dying(wbc
->wb
) && !css_is_dying(wbc
->wb
->memcg_css
)))
722 inode_switch_wbs(inode
, wbc
->wb_id
);
724 EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode
);
727 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
728 * @wbc: writeback_control of the just finished writeback
730 * To be called after a writeback attempt of an inode finishes and undoes
731 * wbc_attach_and_unlock_inode(). Can be called under any context.
733 * As concurrent write sharing of an inode is expected to be very rare and
734 * memcg only tracks page ownership on first-use basis severely confining
735 * the usefulness of such sharing, cgroup writeback tracks ownership
736 * per-inode. While the support for concurrent write sharing of an inode
737 * is deemed unnecessary, an inode being written to by different cgroups at
738 * different points in time is a lot more common, and, more importantly,
739 * charging only by first-use can too readily lead to grossly incorrect
740 * behaviors (single foreign page can lead to gigabytes of writeback to be
741 * incorrectly attributed).
743 * To resolve this issue, cgroup writeback detects the majority dirtier of
744 * an inode and transfers the ownership to it. To avoid unnecessary
745 * oscillation, the detection mechanism keeps track of history and gives
746 * out the switch verdict only if the foreign usage pattern is stable over
747 * a certain amount of time and/or writeback attempts.
749 * On each writeback attempt, @wbc tries to detect the majority writer
750 * using Boyer-Moore majority vote algorithm. In addition to the byte
751 * count from the majority voting, it also counts the bytes written for the
752 * current wb and the last round's winner wb (max of last round's current
753 * wb, the winner from two rounds ago, and the last round's majority
754 * candidate). Keeping track of the historical winner helps the algorithm
755 * to semi-reliably detect the most active writer even when it's not the
758 * Once the winner of the round is determined, whether the winner is
759 * foreign or not and how much IO time the round consumed is recorded in
760 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
761 * over a certain threshold, the switch verdict is given.
763 void wbc_detach_inode(struct writeback_control
*wbc
)
765 struct bdi_writeback
*wb
= wbc
->wb
;
766 struct inode
*inode
= wbc
->inode
;
767 unsigned long avg_time
, max_bytes
, max_time
;
774 history
= inode
->i_wb_frn_history
;
775 avg_time
= inode
->i_wb_frn_avg_time
;
777 /* pick the winner of this round */
778 if (wbc
->wb_bytes
>= wbc
->wb_lcand_bytes
&&
779 wbc
->wb_bytes
>= wbc
->wb_tcand_bytes
) {
781 max_bytes
= wbc
->wb_bytes
;
782 } else if (wbc
->wb_lcand_bytes
>= wbc
->wb_tcand_bytes
) {
783 max_id
= wbc
->wb_lcand_id
;
784 max_bytes
= wbc
->wb_lcand_bytes
;
786 max_id
= wbc
->wb_tcand_id
;
787 max_bytes
= wbc
->wb_tcand_bytes
;
791 * Calculate the amount of IO time the winner consumed and fold it
792 * into the running average kept per inode. If the consumed IO
793 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
794 * deciding whether to switch or not. This is to prevent one-off
795 * small dirtiers from skewing the verdict.
797 max_time
= DIV_ROUND_UP((max_bytes
>> PAGE_SHIFT
) << WB_FRN_TIME_SHIFT
,
798 wb
->avg_write_bandwidth
);
800 avg_time
+= (max_time
>> WB_FRN_TIME_AVG_SHIFT
) -
801 (avg_time
>> WB_FRN_TIME_AVG_SHIFT
);
803 avg_time
= max_time
; /* immediate catch up on first run */
805 if (max_time
>= avg_time
/ WB_FRN_TIME_CUT_DIV
) {
809 * The switch verdict is reached if foreign wb's consume
810 * more than a certain proportion of IO time in a
811 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
812 * history mask where each bit represents one sixteenth of
813 * the period. Determine the number of slots to shift into
814 * history from @max_time.
816 slots
= min(DIV_ROUND_UP(max_time
, WB_FRN_HIST_UNIT
),
817 (unsigned long)WB_FRN_HIST_MAX_SLOTS
);
819 if (wbc
->wb_id
!= max_id
)
820 history
|= (1U << slots
) - 1;
823 trace_inode_foreign_history(inode
, wbc
, history
);
826 * Switch if the current wb isn't the consistent winner.
827 * If there are multiple closely competing dirtiers, the
828 * inode may switch across them repeatedly over time, which
829 * is okay. The main goal is avoiding keeping an inode on
830 * the wrong wb for an extended period of time.
832 if (hweight32(history
) > WB_FRN_HIST_THR_SLOTS
)
833 inode_switch_wbs(inode
, max_id
);
837 * Multiple instances of this function may race to update the
838 * following fields but we don't mind occassional inaccuracies.
840 inode
->i_wb_frn_winner
= max_id
;
841 inode
->i_wb_frn_avg_time
= min(avg_time
, (unsigned long)U16_MAX
);
842 inode
->i_wb_frn_history
= history
;
847 EXPORT_SYMBOL_GPL(wbc_detach_inode
);
850 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
851 * @wbc: writeback_control of the writeback in progress
852 * @page: page being written out
853 * @bytes: number of bytes being written out
855 * @bytes from @page are about to written out during the writeback
856 * controlled by @wbc. Keep the book for foreign inode detection. See
857 * wbc_detach_inode().
859 void wbc_account_cgroup_owner(struct writeback_control
*wbc
, struct page
*page
,
863 struct cgroup_subsys_state
*css
;
867 * pageout() path doesn't attach @wbc to the inode being written
868 * out. This is intentional as we don't want the function to block
869 * behind a slow cgroup. Ultimately, we want pageout() to kick off
870 * regular writeback instead of writing things out itself.
872 if (!wbc
->wb
|| wbc
->no_cgroup_owner
)
875 folio
= page_folio(page
);
876 css
= mem_cgroup_css_from_folio(folio
);
877 /* dead cgroups shouldn't contribute to inode ownership arbitration */
878 if (!(css
->flags
& CSS_ONLINE
))
883 if (id
== wbc
->wb_id
) {
884 wbc
->wb_bytes
+= bytes
;
888 if (id
== wbc
->wb_lcand_id
)
889 wbc
->wb_lcand_bytes
+= bytes
;
891 /* Boyer-Moore majority vote algorithm */
892 if (!wbc
->wb_tcand_bytes
)
893 wbc
->wb_tcand_id
= id
;
894 if (id
== wbc
->wb_tcand_id
)
895 wbc
->wb_tcand_bytes
+= bytes
;
897 wbc
->wb_tcand_bytes
-= min(bytes
, wbc
->wb_tcand_bytes
);
899 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner
);
902 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
903 * @wb: target bdi_writeback to split @nr_pages to
904 * @nr_pages: number of pages to write for the whole bdi
906 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
907 * relation to the total write bandwidth of all wb's w/ dirty inodes on
910 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
912 unsigned long this_bw
= wb
->avg_write_bandwidth
;
913 unsigned long tot_bw
= atomic_long_read(&wb
->bdi
->tot_write_bandwidth
);
915 if (nr_pages
== LONG_MAX
)
919 * This may be called on clean wb's and proportional distribution
920 * may not make sense, just use the original @nr_pages in those
921 * cases. In general, we wanna err on the side of writing more.
923 if (!tot_bw
|| this_bw
>= tot_bw
)
926 return DIV_ROUND_UP_ULL((u64
)nr_pages
* this_bw
, tot_bw
);
930 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
931 * @bdi: target backing_dev_info
932 * @base_work: wb_writeback_work to issue
933 * @skip_if_busy: skip wb's which already have writeback in progress
935 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
936 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
937 * distributed to the busy wbs according to each wb's proportion in the
938 * total active write bandwidth of @bdi.
940 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
941 struct wb_writeback_work
*base_work
,
944 struct bdi_writeback
*last_wb
= NULL
;
945 struct bdi_writeback
*wb
= list_entry(&bdi
->wb_list
,
946 struct bdi_writeback
, bdi_node
);
951 list_for_each_entry_continue_rcu(wb
, &bdi
->wb_list
, bdi_node
) {
952 DEFINE_WB_COMPLETION(fallback_work_done
, bdi
);
953 struct wb_writeback_work fallback_work
;
954 struct wb_writeback_work
*work
;
962 /* SYNC_ALL writes out I_DIRTY_TIME too */
963 if (!wb_has_dirty_io(wb
) &&
964 (base_work
->sync_mode
== WB_SYNC_NONE
||
965 list_empty(&wb
->b_dirty_time
)))
967 if (skip_if_busy
&& writeback_in_progress(wb
))
970 nr_pages
= wb_split_bdi_pages(wb
, base_work
->nr_pages
);
972 work
= kmalloc(sizeof(*work
), GFP_ATOMIC
);
975 work
->nr_pages
= nr_pages
;
977 wb_queue_work(wb
, work
);
981 /* alloc failed, execute synchronously using on-stack fallback */
982 work
= &fallback_work
;
984 work
->nr_pages
= nr_pages
;
986 work
->done
= &fallback_work_done
;
988 wb_queue_work(wb
, work
);
991 * Pin @wb so that it stays on @bdi->wb_list. This allows
992 * continuing iteration from @wb after dropping and
993 * regrabbing rcu read lock.
999 wb_wait_for_completion(&fallback_work_done
);
1009 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1010 * @bdi_id: target bdi id
1011 * @memcg_id: target memcg css id
1012 * @reason: reason why some writeback work initiated
1013 * @done: target wb_completion
1015 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1016 * with the specified parameters.
1018 int cgroup_writeback_by_id(u64 bdi_id
, int memcg_id
,
1019 enum wb_reason reason
, struct wb_completion
*done
)
1021 struct backing_dev_info
*bdi
;
1022 struct cgroup_subsys_state
*memcg_css
;
1023 struct bdi_writeback
*wb
;
1024 struct wb_writeback_work
*work
;
1025 unsigned long dirty
;
1028 /* lookup bdi and memcg */
1029 bdi
= bdi_get_by_id(bdi_id
);
1034 memcg_css
= css_from_id(memcg_id
, &memory_cgrp_subsys
);
1035 if (memcg_css
&& !css_tryget(memcg_css
))
1044 * And find the associated wb. If the wb isn't there already
1045 * there's nothing to flush, don't create one.
1047 wb
= wb_get_lookup(bdi
, memcg_css
);
1054 * The caller is attempting to write out most of
1055 * the currently dirty pages. Let's take the current dirty page
1056 * count and inflate it by 25% which should be large enough to
1057 * flush out most dirty pages while avoiding getting livelocked by
1058 * concurrent dirtiers.
1060 * BTW the memcg stats are flushed periodically and this is best-effort
1061 * estimation, so some potential error is ok.
1063 dirty
= memcg_page_state(mem_cgroup_from_css(memcg_css
), NR_FILE_DIRTY
);
1064 dirty
= dirty
* 10 / 8;
1066 /* issue the writeback work */
1067 work
= kzalloc(sizeof(*work
), GFP_NOWAIT
| __GFP_NOWARN
);
1069 work
->nr_pages
= dirty
;
1070 work
->sync_mode
= WB_SYNC_NONE
;
1071 work
->range_cyclic
= 1;
1072 work
->reason
= reason
;
1074 work
->auto_free
= 1;
1075 wb_queue_work(wb
, work
);
1090 * cgroup_writeback_umount - flush inode wb switches for umount
1092 * This function is called when a super_block is about to be destroyed and
1093 * flushes in-flight inode wb switches. An inode wb switch goes through
1094 * RCU and then workqueue, so the two need to be flushed in order to ensure
1095 * that all previously scheduled switches are finished. As wb switches are
1096 * rare occurrences and synchronize_rcu() can take a while, perform
1097 * flushing iff wb switches are in flight.
1099 void cgroup_writeback_umount(void)
1102 * SB_ACTIVE should be reliably cleared before checking
1103 * isw_nr_in_flight, see generic_shutdown_super().
1107 if (atomic_read(&isw_nr_in_flight
)) {
1109 * Use rcu_barrier() to wait for all pending callbacks to
1110 * ensure that all in-flight wb switches are in the workqueue.
1113 flush_workqueue(isw_wq
);
1117 static int __init
cgroup_writeback_init(void)
1119 isw_wq
= alloc_workqueue("inode_switch_wbs", 0, 0);
1124 fs_initcall(cgroup_writeback_init
);
1126 #else /* CONFIG_CGROUP_WRITEBACK */
1128 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
1129 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info
*bdi
) { }
1131 static void inode_cgwb_move_to_attached(struct inode
*inode
,
1132 struct bdi_writeback
*wb
)
1134 assert_spin_locked(&wb
->list_lock
);
1135 assert_spin_locked(&inode
->i_lock
);
1136 WARN_ON_ONCE(inode
->i_state
& I_FREEING
);
1138 inode
->i_state
&= ~I_SYNC_QUEUED
;
1139 list_del_init(&inode
->i_io_list
);
1140 wb_io_lists_depopulated(wb
);
1143 static struct bdi_writeback
*
1144 locked_inode_to_wb_and_lock_list(struct inode
*inode
)
1145 __releases(&inode
->i_lock
)
1146 __acquires(&wb
->list_lock
)
1148 struct bdi_writeback
*wb
= inode_to_wb(inode
);
1150 spin_unlock(&inode
->i_lock
);
1151 spin_lock(&wb
->list_lock
);
1155 static struct bdi_writeback
*inode_to_wb_and_lock_list(struct inode
*inode
)
1156 __acquires(&wb
->list_lock
)
1158 struct bdi_writeback
*wb
= inode_to_wb(inode
);
1160 spin_lock(&wb
->list_lock
);
1164 static long wb_split_bdi_pages(struct bdi_writeback
*wb
, long nr_pages
)
1169 static void bdi_split_work_to_wbs(struct backing_dev_info
*bdi
,
1170 struct wb_writeback_work
*base_work
,
1175 if (!skip_if_busy
|| !writeback_in_progress(&bdi
->wb
)) {
1176 base_work
->auto_free
= 0;
1177 wb_queue_work(&bdi
->wb
, base_work
);
1181 #endif /* CONFIG_CGROUP_WRITEBACK */
1184 * Add in the number of potentially dirty inodes, because each inode
1185 * write can dirty pagecache in the underlying blockdev.
1187 static unsigned long get_nr_dirty_pages(void)
1189 return global_node_page_state(NR_FILE_DIRTY
) +
1190 get_nr_dirty_inodes();
1193 static void wb_start_writeback(struct bdi_writeback
*wb
, enum wb_reason reason
)
1195 if (!wb_has_dirty_io(wb
))
1199 * All callers of this function want to start writeback of all
1200 * dirty pages. Places like vmscan can call this at a very
1201 * high frequency, causing pointless allocations of tons of
1202 * work items and keeping the flusher threads busy retrieving
1203 * that work. Ensure that we only allow one of them pending and
1204 * inflight at the time.
1206 if (test_bit(WB_start_all
, &wb
->state
) ||
1207 test_and_set_bit(WB_start_all
, &wb
->state
))
1210 wb
->start_all_reason
= reason
;
1215 * wb_start_background_writeback - start background writeback
1216 * @wb: bdi_writback to write from
1219 * This makes sure WB_SYNC_NONE background writeback happens. When
1220 * this function returns, it is only guaranteed that for given wb
1221 * some IO is happening if we are over background dirty threshold.
1222 * Caller need not hold sb s_umount semaphore.
1224 void wb_start_background_writeback(struct bdi_writeback
*wb
)
1227 * We just wake up the flusher thread. It will perform background
1228 * writeback as soon as there is no other work to do.
1230 trace_writeback_wake_background(wb
);
1235 * Remove the inode from the writeback list it is on.
1237 void inode_io_list_del(struct inode
*inode
)
1239 struct bdi_writeback
*wb
;
1241 wb
= inode_to_wb_and_lock_list(inode
);
1242 spin_lock(&inode
->i_lock
);
1244 inode
->i_state
&= ~I_SYNC_QUEUED
;
1245 list_del_init(&inode
->i_io_list
);
1246 wb_io_lists_depopulated(wb
);
1248 spin_unlock(&inode
->i_lock
);
1249 spin_unlock(&wb
->list_lock
);
1251 EXPORT_SYMBOL(inode_io_list_del
);
1254 * mark an inode as under writeback on the sb
1256 void sb_mark_inode_writeback(struct inode
*inode
)
1258 struct super_block
*sb
= inode
->i_sb
;
1259 unsigned long flags
;
1261 if (list_empty(&inode
->i_wb_list
)) {
1262 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1263 if (list_empty(&inode
->i_wb_list
)) {
1264 list_add_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
1265 trace_sb_mark_inode_writeback(inode
);
1267 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1272 * clear an inode as under writeback on the sb
1274 void sb_clear_inode_writeback(struct inode
*inode
)
1276 struct super_block
*sb
= inode
->i_sb
;
1277 unsigned long flags
;
1279 if (!list_empty(&inode
->i_wb_list
)) {
1280 spin_lock_irqsave(&sb
->s_inode_wblist_lock
, flags
);
1281 if (!list_empty(&inode
->i_wb_list
)) {
1282 list_del_init(&inode
->i_wb_list
);
1283 trace_sb_clear_inode_writeback(inode
);
1285 spin_unlock_irqrestore(&sb
->s_inode_wblist_lock
, flags
);
1290 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1291 * furthest end of its superblock's dirty-inode list.
1293 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1294 * already the most-recently-dirtied inode on the b_dirty list. If that is
1295 * the case then the inode must have been redirtied while it was being written
1296 * out and we don't reset its dirtied_when.
1298 static void redirty_tail_locked(struct inode
*inode
, struct bdi_writeback
*wb
)
1300 assert_spin_locked(&inode
->i_lock
);
1302 inode
->i_state
&= ~I_SYNC_QUEUED
;
1304 * When the inode is being freed just don't bother with dirty list
1305 * tracking. Flush worker will ignore this inode anyway and it will
1306 * trigger assertions in inode_io_list_move_locked().
1308 if (inode
->i_state
& I_FREEING
) {
1309 list_del_init(&inode
->i_io_list
);
1310 wb_io_lists_depopulated(wb
);
1313 if (!list_empty(&wb
->b_dirty
)) {
1316 tail
= wb_inode(wb
->b_dirty
.next
);
1317 if (time_before(inode
->dirtied_when
, tail
->dirtied_when
))
1318 inode
->dirtied_when
= jiffies
;
1320 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty
);
1323 static void redirty_tail(struct inode
*inode
, struct bdi_writeback
*wb
)
1325 spin_lock(&inode
->i_lock
);
1326 redirty_tail_locked(inode
, wb
);
1327 spin_unlock(&inode
->i_lock
);
1331 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1333 static void requeue_io(struct inode
*inode
, struct bdi_writeback
*wb
)
1335 inode_io_list_move_locked(inode
, wb
, &wb
->b_more_io
);
1338 static void inode_sync_complete(struct inode
*inode
)
1340 inode
->i_state
&= ~I_SYNC
;
1341 /* If inode is clean an unused, put it into LRU now... */
1342 inode_add_lru(inode
);
1343 /* Waiters must see I_SYNC cleared before being woken up */
1345 wake_up_bit(&inode
->i_state
, __I_SYNC
);
1348 static bool inode_dirtied_after(struct inode
*inode
, unsigned long t
)
1350 bool ret
= time_after(inode
->dirtied_when
, t
);
1351 #ifndef CONFIG_64BIT
1353 * For inodes being constantly redirtied, dirtied_when can get stuck.
1354 * It _appears_ to be in the future, but is actually in distant past.
1355 * This test is necessary to prevent such wrapped-around relative times
1356 * from permanently stopping the whole bdi writeback.
1358 ret
= ret
&& time_before_eq(inode
->dirtied_when
, jiffies
);
1364 * Move expired (dirtied before dirtied_before) dirty inodes from
1365 * @delaying_queue to @dispatch_queue.
1367 static int move_expired_inodes(struct list_head
*delaying_queue
,
1368 struct list_head
*dispatch_queue
,
1369 unsigned long dirtied_before
)
1372 struct list_head
*pos
, *node
;
1373 struct super_block
*sb
= NULL
;
1374 struct inode
*inode
;
1378 while (!list_empty(delaying_queue
)) {
1379 inode
= wb_inode(delaying_queue
->prev
);
1380 if (inode_dirtied_after(inode
, dirtied_before
))
1382 spin_lock(&inode
->i_lock
);
1383 list_move(&inode
->i_io_list
, &tmp
);
1385 inode
->i_state
|= I_SYNC_QUEUED
;
1386 spin_unlock(&inode
->i_lock
);
1387 if (sb_is_blkdev_sb(inode
->i_sb
))
1389 if (sb
&& sb
!= inode
->i_sb
)
1394 /* just one sb in list, splice to dispatch_queue and we're done */
1396 list_splice(&tmp
, dispatch_queue
);
1401 * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1402 * we don't take inode->i_lock here because it is just a pointless overhead.
1403 * Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1404 * fully under our control.
1406 while (!list_empty(&tmp
)) {
1407 sb
= wb_inode(tmp
.prev
)->i_sb
;
1408 list_for_each_prev_safe(pos
, node
, &tmp
) {
1409 inode
= wb_inode(pos
);
1410 if (inode
->i_sb
== sb
)
1411 list_move(&inode
->i_io_list
, dispatch_queue
);
1419 * Queue all expired dirty inodes for io, eldest first.
1421 * newly dirtied b_dirty b_io b_more_io
1422 * =============> gf edc BA
1424 * newly dirtied b_dirty b_io b_more_io
1425 * =============> g fBAedc
1427 * +--> dequeue for IO
1429 static void queue_io(struct bdi_writeback
*wb
, struct wb_writeback_work
*work
,
1430 unsigned long dirtied_before
)
1433 unsigned long time_expire_jif
= dirtied_before
;
1435 assert_spin_locked(&wb
->list_lock
);
1436 list_splice_init(&wb
->b_more_io
, &wb
->b_io
);
1437 moved
= move_expired_inodes(&wb
->b_dirty
, &wb
->b_io
, dirtied_before
);
1438 if (!work
->for_sync
)
1439 time_expire_jif
= jiffies
- dirtytime_expire_interval
* HZ
;
1440 moved
+= move_expired_inodes(&wb
->b_dirty_time
, &wb
->b_io
,
1443 wb_io_lists_populated(wb
);
1444 trace_writeback_queue_io(wb
, work
, dirtied_before
, moved
);
1447 static int write_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1451 if (inode
->i_sb
->s_op
->write_inode
&& !is_bad_inode(inode
)) {
1452 trace_writeback_write_inode_start(inode
, wbc
);
1453 ret
= inode
->i_sb
->s_op
->write_inode(inode
, wbc
);
1454 trace_writeback_write_inode(inode
, wbc
);
1461 * Wait for writeback on an inode to complete. Called with i_lock held.
1462 * Caller must make sure inode cannot go away when we drop i_lock.
1464 static void __inode_wait_for_writeback(struct inode
*inode
)
1465 __releases(inode
->i_lock
)
1466 __acquires(inode
->i_lock
)
1468 DEFINE_WAIT_BIT(wq
, &inode
->i_state
, __I_SYNC
);
1469 wait_queue_head_t
*wqh
;
1471 wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1472 while (inode
->i_state
& I_SYNC
) {
1473 spin_unlock(&inode
->i_lock
);
1474 __wait_on_bit(wqh
, &wq
, bit_wait
,
1475 TASK_UNINTERRUPTIBLE
);
1476 spin_lock(&inode
->i_lock
);
1481 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1483 void inode_wait_for_writeback(struct inode
*inode
)
1485 spin_lock(&inode
->i_lock
);
1486 __inode_wait_for_writeback(inode
);
1487 spin_unlock(&inode
->i_lock
);
1491 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1492 * held and drops it. It is aimed for callers not holding any inode reference
1493 * so once i_lock is dropped, inode can go away.
1495 static void inode_sleep_on_writeback(struct inode
*inode
)
1496 __releases(inode
->i_lock
)
1499 wait_queue_head_t
*wqh
= bit_waitqueue(&inode
->i_state
, __I_SYNC
);
1502 prepare_to_wait(wqh
, &wait
, TASK_UNINTERRUPTIBLE
);
1503 sleep
= inode
->i_state
& I_SYNC
;
1504 spin_unlock(&inode
->i_lock
);
1507 finish_wait(wqh
, &wait
);
1511 * Find proper writeback list for the inode depending on its current state and
1512 * possibly also change of its state while we were doing writeback. Here we
1513 * handle things such as livelock prevention or fairness of writeback among
1514 * inodes. This function can be called only by flusher thread - noone else
1515 * processes all inodes in writeback lists and requeueing inodes behind flusher
1516 * thread's back can have unexpected consequences.
1518 static void requeue_inode(struct inode
*inode
, struct bdi_writeback
*wb
,
1519 struct writeback_control
*wbc
)
1521 if (inode
->i_state
& I_FREEING
)
1525 * Sync livelock prevention. Each inode is tagged and synced in one
1526 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1527 * the dirty time to prevent enqueue and sync it again.
1529 if ((inode
->i_state
& I_DIRTY
) &&
1530 (wbc
->sync_mode
== WB_SYNC_ALL
|| wbc
->tagged_writepages
))
1531 inode
->dirtied_when
= jiffies
;
1533 if (wbc
->pages_skipped
) {
1535 * writeback is not making progress due to locked
1536 * buffers. Skip this inode for now.
1538 redirty_tail_locked(inode
, wb
);
1542 if (mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_DIRTY
)) {
1544 * We didn't write back all the pages. nfs_writepages()
1545 * sometimes bales out without doing anything.
1547 if (wbc
->nr_to_write
<= 0) {
1548 /* Slice used up. Queue for next turn. */
1549 requeue_io(inode
, wb
);
1552 * Writeback blocked by something other than
1553 * congestion. Delay the inode for some time to
1554 * avoid spinning on the CPU (100% iowait)
1555 * retrying writeback of the dirty page/inode
1556 * that cannot be performed immediately.
1558 redirty_tail_locked(inode
, wb
);
1560 } else if (inode
->i_state
& I_DIRTY
) {
1562 * Filesystems can dirty the inode during writeback operations,
1563 * such as delayed allocation during submission or metadata
1564 * updates after data IO completion.
1566 redirty_tail_locked(inode
, wb
);
1567 } else if (inode
->i_state
& I_DIRTY_TIME
) {
1568 inode
->dirtied_when
= jiffies
;
1569 inode_io_list_move_locked(inode
, wb
, &wb
->b_dirty_time
);
1570 inode
->i_state
&= ~I_SYNC_QUEUED
;
1572 /* The inode is clean. Remove from writeback lists. */
1573 inode_cgwb_move_to_attached(inode
, wb
);
1578 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1579 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1581 * This doesn't remove the inode from the writeback list it is on, except
1582 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1583 * expiration. The caller is otherwise responsible for writeback list handling.
1585 * The caller is also responsible for setting the I_SYNC flag beforehand and
1586 * calling inode_sync_complete() to clear it afterwards.
1589 __writeback_single_inode(struct inode
*inode
, struct writeback_control
*wbc
)
1591 struct address_space
*mapping
= inode
->i_mapping
;
1592 long nr_to_write
= wbc
->nr_to_write
;
1596 WARN_ON(!(inode
->i_state
& I_SYNC
));
1598 trace_writeback_single_inode_start(inode
, wbc
, nr_to_write
);
1600 ret
= do_writepages(mapping
, wbc
);
1603 * Make sure to wait on the data before writing out the metadata.
1604 * This is important for filesystems that modify metadata on data
1605 * I/O completion. We don't do it for sync(2) writeback because it has a
1606 * separate, external IO completion path and ->sync_fs for guaranteeing
1607 * inode metadata is written back correctly.
1609 if (wbc
->sync_mode
== WB_SYNC_ALL
&& !wbc
->for_sync
) {
1610 int err
= filemap_fdatawait(mapping
);
1616 * If the inode has dirty timestamps and we need to write them, call
1617 * mark_inode_dirty_sync() to notify the filesystem about it and to
1618 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1620 if ((inode
->i_state
& I_DIRTY_TIME
) &&
1621 (wbc
->sync_mode
== WB_SYNC_ALL
||
1622 time_after(jiffies
, inode
->dirtied_time_when
+
1623 dirtytime_expire_interval
* HZ
))) {
1624 trace_writeback_lazytime(inode
);
1625 mark_inode_dirty_sync(inode
);
1629 * Get and clear the dirty flags from i_state. This needs to be done
1630 * after calling writepages because some filesystems may redirty the
1631 * inode during writepages due to delalloc. It also needs to be done
1632 * after handling timestamp expiration, as that may dirty the inode too.
1634 spin_lock(&inode
->i_lock
);
1635 dirty
= inode
->i_state
& I_DIRTY
;
1636 inode
->i_state
&= ~dirty
;
1639 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1640 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1641 * either they see the I_DIRTY bits cleared or we see the dirtied
1644 * I_DIRTY_PAGES is always cleared together above even if @mapping
1645 * still has dirty pages. The flag is reinstated after smp_mb() if
1646 * necessary. This guarantees that either __mark_inode_dirty()
1647 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1651 if (mapping_tagged(mapping
, PAGECACHE_TAG_DIRTY
))
1652 inode
->i_state
|= I_DIRTY_PAGES
;
1653 else if (unlikely(inode
->i_state
& I_PINNING_FSCACHE_WB
)) {
1654 if (!(inode
->i_state
& I_DIRTY_PAGES
)) {
1655 inode
->i_state
&= ~I_PINNING_FSCACHE_WB
;
1656 wbc
->unpinned_fscache_wb
= true;
1657 dirty
|= I_PINNING_FSCACHE_WB
; /* Cause write_inode */
1661 spin_unlock(&inode
->i_lock
);
1663 /* Don't write the inode if only I_DIRTY_PAGES was set */
1664 if (dirty
& ~I_DIRTY_PAGES
) {
1665 int err
= write_inode(inode
, wbc
);
1669 wbc
->unpinned_fscache_wb
= false;
1670 trace_writeback_single_inode(inode
, wbc
, nr_to_write
);
1675 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1676 * the regular batched writeback done by the flusher threads in
1677 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1678 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1680 * To prevent the inode from going away, either the caller must have a reference
1681 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1683 static int writeback_single_inode(struct inode
*inode
,
1684 struct writeback_control
*wbc
)
1686 struct bdi_writeback
*wb
;
1689 spin_lock(&inode
->i_lock
);
1690 if (!atomic_read(&inode
->i_count
))
1691 WARN_ON(!(inode
->i_state
& (I_WILL_FREE
|I_FREEING
)));
1693 WARN_ON(inode
->i_state
& I_WILL_FREE
);
1695 if (inode
->i_state
& I_SYNC
) {
1697 * Writeback is already running on the inode. For WB_SYNC_NONE,
1698 * that's enough and we can just return. For WB_SYNC_ALL, we
1699 * must wait for the existing writeback to complete, then do
1700 * writeback again if there's anything left.
1702 if (wbc
->sync_mode
!= WB_SYNC_ALL
)
1704 __inode_wait_for_writeback(inode
);
1706 WARN_ON(inode
->i_state
& I_SYNC
);
1708 * If the inode is already fully clean, then there's nothing to do.
1710 * For data-integrity syncs we also need to check whether any pages are
1711 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1712 * there are any such pages, we'll need to wait for them.
1714 if (!(inode
->i_state
& I_DIRTY_ALL
) &&
1715 (wbc
->sync_mode
!= WB_SYNC_ALL
||
1716 !mapping_tagged(inode
->i_mapping
, PAGECACHE_TAG_WRITEBACK
)))
1718 inode
->i_state
|= I_SYNC
;
1719 wbc_attach_and_unlock_inode(wbc
, inode
);
1721 ret
= __writeback_single_inode(inode
, wbc
);
1723 wbc_detach_inode(wbc
);
1725 wb
= inode_to_wb_and_lock_list(inode
);
1726 spin_lock(&inode
->i_lock
);
1728 * If the inode is freeing, its i_io_list shoudn't be updated
1729 * as it can be finally deleted at this moment.
1731 if (!(inode
->i_state
& I_FREEING
)) {
1733 * If the inode is now fully clean, then it can be safely
1734 * removed from its writeback list (if any). Otherwise the
1735 * flusher threads are responsible for the writeback lists.
1737 if (!(inode
->i_state
& I_DIRTY_ALL
))
1738 inode_cgwb_move_to_attached(inode
, wb
);
1739 else if (!(inode
->i_state
& I_SYNC_QUEUED
)) {
1740 if ((inode
->i_state
& I_DIRTY
))
1741 redirty_tail_locked(inode
, wb
);
1742 else if (inode
->i_state
& I_DIRTY_TIME
) {
1743 inode
->dirtied_when
= jiffies
;
1744 inode_io_list_move_locked(inode
,
1751 spin_unlock(&wb
->list_lock
);
1752 inode_sync_complete(inode
);
1754 spin_unlock(&inode
->i_lock
);
1758 static long writeback_chunk_size(struct bdi_writeback
*wb
,
1759 struct wb_writeback_work
*work
)
1764 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1765 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1766 * here avoids calling into writeback_inodes_wb() more than once.
1768 * The intended call sequence for WB_SYNC_ALL writeback is:
1771 * writeback_sb_inodes() <== called only once
1772 * write_cache_pages() <== called once for each inode
1773 * (quickly) tag currently dirty pages
1774 * (maybe slowly) sync all tagged pages
1776 if (work
->sync_mode
== WB_SYNC_ALL
|| work
->tagged_writepages
)
1779 pages
= min(wb
->avg_write_bandwidth
/ 2,
1780 global_wb_domain
.dirty_limit
/ DIRTY_SCOPE
);
1781 pages
= min(pages
, work
->nr_pages
);
1782 pages
= round_down(pages
+ MIN_WRITEBACK_PAGES
,
1783 MIN_WRITEBACK_PAGES
);
1790 * Write a portion of b_io inodes which belong to @sb.
1792 * Return the number of pages and/or inodes written.
1794 * NOTE! This is called with wb->list_lock held, and will
1795 * unlock and relock that for each inode it ends up doing
1798 static long writeback_sb_inodes(struct super_block
*sb
,
1799 struct bdi_writeback
*wb
,
1800 struct wb_writeback_work
*work
)
1802 struct writeback_control wbc
= {
1803 .sync_mode
= work
->sync_mode
,
1804 .tagged_writepages
= work
->tagged_writepages
,
1805 .for_kupdate
= work
->for_kupdate
,
1806 .for_background
= work
->for_background
,
1807 .for_sync
= work
->for_sync
,
1808 .range_cyclic
= work
->range_cyclic
,
1810 .range_end
= LLONG_MAX
,
1812 unsigned long start_time
= jiffies
;
1814 long total_wrote
= 0; /* count both pages and inodes */
1816 while (!list_empty(&wb
->b_io
)) {
1817 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1818 struct bdi_writeback
*tmp_wb
;
1821 if (inode
->i_sb
!= sb
) {
1824 * We only want to write back data for this
1825 * superblock, move all inodes not belonging
1826 * to it back onto the dirty list.
1828 redirty_tail(inode
, wb
);
1833 * The inode belongs to a different superblock.
1834 * Bounce back to the caller to unpin this and
1835 * pin the next superblock.
1841 * Don't bother with new inodes or inodes being freed, first
1842 * kind does not need periodic writeout yet, and for the latter
1843 * kind writeout is handled by the freer.
1845 spin_lock(&inode
->i_lock
);
1846 if (inode
->i_state
& (I_NEW
| I_FREEING
| I_WILL_FREE
)) {
1847 redirty_tail_locked(inode
, wb
);
1848 spin_unlock(&inode
->i_lock
);
1851 if ((inode
->i_state
& I_SYNC
) && wbc
.sync_mode
!= WB_SYNC_ALL
) {
1853 * If this inode is locked for writeback and we are not
1854 * doing writeback-for-data-integrity, move it to
1855 * b_more_io so that writeback can proceed with the
1856 * other inodes on s_io.
1858 * We'll have another go at writing back this inode
1859 * when we completed a full scan of b_io.
1861 requeue_io(inode
, wb
);
1862 spin_unlock(&inode
->i_lock
);
1863 trace_writeback_sb_inodes_requeue(inode
);
1866 spin_unlock(&wb
->list_lock
);
1869 * We already requeued the inode if it had I_SYNC set and we
1870 * are doing WB_SYNC_NONE writeback. So this catches only the
1873 if (inode
->i_state
& I_SYNC
) {
1874 /* Wait for I_SYNC. This function drops i_lock... */
1875 inode_sleep_on_writeback(inode
);
1876 /* Inode may be gone, start again */
1877 spin_lock(&wb
->list_lock
);
1880 inode
->i_state
|= I_SYNC
;
1881 wbc_attach_and_unlock_inode(&wbc
, inode
);
1883 write_chunk
= writeback_chunk_size(wb
, work
);
1884 wbc
.nr_to_write
= write_chunk
;
1885 wbc
.pages_skipped
= 0;
1888 * We use I_SYNC to pin the inode in memory. While it is set
1889 * evict_inode() will wait so the inode cannot be freed.
1891 __writeback_single_inode(inode
, &wbc
);
1893 wbc_detach_inode(&wbc
);
1894 work
->nr_pages
-= write_chunk
- wbc
.nr_to_write
;
1895 wrote
= write_chunk
- wbc
.nr_to_write
- wbc
.pages_skipped
;
1896 wrote
= wrote
< 0 ? 0 : wrote
;
1897 total_wrote
+= wrote
;
1899 if (need_resched()) {
1901 * We're trying to balance between building up a nice
1902 * long list of IOs to improve our merge rate, and
1903 * getting those IOs out quickly for anyone throttling
1904 * in balance_dirty_pages(). cond_resched() doesn't
1905 * unplug, so get our IOs out the door before we
1908 blk_flush_plug(current
->plug
, false);
1913 * Requeue @inode if still dirty. Be careful as @inode may
1914 * have been switched to another wb in the meantime.
1916 tmp_wb
= inode_to_wb_and_lock_list(inode
);
1917 spin_lock(&inode
->i_lock
);
1918 if (!(inode
->i_state
& I_DIRTY_ALL
))
1920 requeue_inode(inode
, tmp_wb
, &wbc
);
1921 inode_sync_complete(inode
);
1922 spin_unlock(&inode
->i_lock
);
1924 if (unlikely(tmp_wb
!= wb
)) {
1925 spin_unlock(&tmp_wb
->list_lock
);
1926 spin_lock(&wb
->list_lock
);
1930 * bail out to wb_writeback() often enough to check
1931 * background threshold and other termination conditions.
1934 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1936 if (work
->nr_pages
<= 0)
1943 static long __writeback_inodes_wb(struct bdi_writeback
*wb
,
1944 struct wb_writeback_work
*work
)
1946 unsigned long start_time
= jiffies
;
1949 while (!list_empty(&wb
->b_io
)) {
1950 struct inode
*inode
= wb_inode(wb
->b_io
.prev
);
1951 struct super_block
*sb
= inode
->i_sb
;
1953 if (!trylock_super(sb
)) {
1955 * trylock_super() may fail consistently due to
1956 * s_umount being grabbed by someone else. Don't use
1957 * requeue_io() to avoid busy retrying the inode/sb.
1959 redirty_tail(inode
, wb
);
1962 wrote
+= writeback_sb_inodes(sb
, wb
, work
);
1963 up_read(&sb
->s_umount
);
1965 /* refer to the same tests at the end of writeback_sb_inodes */
1967 if (time_is_before_jiffies(start_time
+ HZ
/ 10UL))
1969 if (work
->nr_pages
<= 0)
1973 /* Leave any unwritten inodes on b_io */
1977 static long writeback_inodes_wb(struct bdi_writeback
*wb
, long nr_pages
,
1978 enum wb_reason reason
)
1980 struct wb_writeback_work work
= {
1981 .nr_pages
= nr_pages
,
1982 .sync_mode
= WB_SYNC_NONE
,
1986 struct blk_plug plug
;
1988 blk_start_plug(&plug
);
1989 spin_lock(&wb
->list_lock
);
1990 if (list_empty(&wb
->b_io
))
1991 queue_io(wb
, &work
, jiffies
);
1992 __writeback_inodes_wb(wb
, &work
);
1993 spin_unlock(&wb
->list_lock
);
1994 blk_finish_plug(&plug
);
1996 return nr_pages
- work
.nr_pages
;
2000 * Explicit flushing or periodic writeback of "old" data.
2002 * Define "old": the first time one of an inode's pages is dirtied, we mark the
2003 * dirtying-time in the inode's address_space. So this periodic writeback code
2004 * just walks the superblock inode list, writing back any inodes which are
2005 * older than a specific point in time.
2007 * Try to run once per dirty_writeback_interval. But if a writeback event
2008 * takes longer than a dirty_writeback_interval interval, then leave a
2011 * dirtied_before takes precedence over nr_to_write. So we'll only write back
2012 * all dirty pages if they are all attached to "old" mappings.
2014 static long wb_writeback(struct bdi_writeback
*wb
,
2015 struct wb_writeback_work
*work
)
2017 long nr_pages
= work
->nr_pages
;
2018 unsigned long dirtied_before
= jiffies
;
2019 struct inode
*inode
;
2021 struct blk_plug plug
;
2023 blk_start_plug(&plug
);
2024 spin_lock(&wb
->list_lock
);
2027 * Stop writeback when nr_pages has been consumed
2029 if (work
->nr_pages
<= 0)
2033 * Background writeout and kupdate-style writeback may
2034 * run forever. Stop them if there is other work to do
2035 * so that e.g. sync can proceed. They'll be restarted
2036 * after the other works are all done.
2038 if ((work
->for_background
|| work
->for_kupdate
) &&
2039 !list_empty(&wb
->work_list
))
2043 * For background writeout, stop when we are below the
2044 * background dirty threshold
2046 if (work
->for_background
&& !wb_over_bg_thresh(wb
))
2050 * Kupdate and background works are special and we want to
2051 * include all inodes that need writing. Livelock avoidance is
2052 * handled by these works yielding to any other work so we are
2055 if (work
->for_kupdate
) {
2056 dirtied_before
= jiffies
-
2057 msecs_to_jiffies(dirty_expire_interval
* 10);
2058 } else if (work
->for_background
)
2059 dirtied_before
= jiffies
;
2061 trace_writeback_start(wb
, work
);
2062 if (list_empty(&wb
->b_io
))
2063 queue_io(wb
, work
, dirtied_before
);
2065 progress
= writeback_sb_inodes(work
->sb
, wb
, work
);
2067 progress
= __writeback_inodes_wb(wb
, work
);
2068 trace_writeback_written(wb
, work
);
2071 * Did we write something? Try for more
2073 * Dirty inodes are moved to b_io for writeback in batches.
2074 * The completion of the current batch does not necessarily
2075 * mean the overall work is done. So we keep looping as long
2076 * as made some progress on cleaning pages or inodes.
2081 * No more inodes for IO, bail
2083 if (list_empty(&wb
->b_more_io
))
2086 * Nothing written. Wait for some inode to
2087 * become available for writeback. Otherwise
2088 * we'll just busyloop.
2090 trace_writeback_wait(wb
, work
);
2091 inode
= wb_inode(wb
->b_more_io
.prev
);
2092 spin_lock(&inode
->i_lock
);
2093 spin_unlock(&wb
->list_lock
);
2094 /* This function drops i_lock... */
2095 inode_sleep_on_writeback(inode
);
2096 spin_lock(&wb
->list_lock
);
2098 spin_unlock(&wb
->list_lock
);
2099 blk_finish_plug(&plug
);
2101 return nr_pages
- work
->nr_pages
;
2105 * Return the next wb_writeback_work struct that hasn't been processed yet.
2107 static struct wb_writeback_work
*get_next_work_item(struct bdi_writeback
*wb
)
2109 struct wb_writeback_work
*work
= NULL
;
2111 spin_lock_irq(&wb
->work_lock
);
2112 if (!list_empty(&wb
->work_list
)) {
2113 work
= list_entry(wb
->work_list
.next
,
2114 struct wb_writeback_work
, list
);
2115 list_del_init(&work
->list
);
2117 spin_unlock_irq(&wb
->work_lock
);
2121 static long wb_check_background_flush(struct bdi_writeback
*wb
)
2123 if (wb_over_bg_thresh(wb
)) {
2125 struct wb_writeback_work work
= {
2126 .nr_pages
= LONG_MAX
,
2127 .sync_mode
= WB_SYNC_NONE
,
2128 .for_background
= 1,
2130 .reason
= WB_REASON_BACKGROUND
,
2133 return wb_writeback(wb
, &work
);
2139 static long wb_check_old_data_flush(struct bdi_writeback
*wb
)
2141 unsigned long expired
;
2145 * When set to zero, disable periodic writeback
2147 if (!dirty_writeback_interval
)
2150 expired
= wb
->last_old_flush
+
2151 msecs_to_jiffies(dirty_writeback_interval
* 10);
2152 if (time_before(jiffies
, expired
))
2155 wb
->last_old_flush
= jiffies
;
2156 nr_pages
= get_nr_dirty_pages();
2159 struct wb_writeback_work work
= {
2160 .nr_pages
= nr_pages
,
2161 .sync_mode
= WB_SYNC_NONE
,
2164 .reason
= WB_REASON_PERIODIC
,
2167 return wb_writeback(wb
, &work
);
2173 static long wb_check_start_all(struct bdi_writeback
*wb
)
2177 if (!test_bit(WB_start_all
, &wb
->state
))
2180 nr_pages
= get_nr_dirty_pages();
2182 struct wb_writeback_work work
= {
2183 .nr_pages
= wb_split_bdi_pages(wb
, nr_pages
),
2184 .sync_mode
= WB_SYNC_NONE
,
2186 .reason
= wb
->start_all_reason
,
2189 nr_pages
= wb_writeback(wb
, &work
);
2192 clear_bit(WB_start_all
, &wb
->state
);
2198 * Retrieve work items and do the writeback they describe
2200 static long wb_do_writeback(struct bdi_writeback
*wb
)
2202 struct wb_writeback_work
*work
;
2205 set_bit(WB_writeback_running
, &wb
->state
);
2206 while ((work
= get_next_work_item(wb
)) != NULL
) {
2207 trace_writeback_exec(wb
, work
);
2208 wrote
+= wb_writeback(wb
, work
);
2209 finish_writeback_work(wb
, work
);
2213 * Check for a flush-everything request
2215 wrote
+= wb_check_start_all(wb
);
2218 * Check for periodic writeback, kupdated() style
2220 wrote
+= wb_check_old_data_flush(wb
);
2221 wrote
+= wb_check_background_flush(wb
);
2222 clear_bit(WB_writeback_running
, &wb
->state
);
2228 * Handle writeback of dirty data for the device backed by this bdi. Also
2229 * reschedules periodically and does kupdated style flushing.
2231 void wb_workfn(struct work_struct
*work
)
2233 struct bdi_writeback
*wb
= container_of(to_delayed_work(work
),
2234 struct bdi_writeback
, dwork
);
2237 set_worker_desc("flush-%s", bdi_dev_name(wb
->bdi
));
2239 if (likely(!current_is_workqueue_rescuer() ||
2240 !test_bit(WB_registered
, &wb
->state
))) {
2242 * The normal path. Keep writing back @wb until its
2243 * work_list is empty. Note that this path is also taken
2244 * if @wb is shutting down even when we're running off the
2245 * rescuer as work_list needs to be drained.
2248 pages_written
= wb_do_writeback(wb
);
2249 trace_writeback_pages_written(pages_written
);
2250 } while (!list_empty(&wb
->work_list
));
2253 * bdi_wq can't get enough workers and we're running off
2254 * the emergency worker. Don't hog it. Hopefully, 1024 is
2255 * enough for efficient IO.
2257 pages_written
= writeback_inodes_wb(wb
, 1024,
2258 WB_REASON_FORKER_THREAD
);
2259 trace_writeback_pages_written(pages_written
);
2262 if (!list_empty(&wb
->work_list
))
2264 else if (wb_has_dirty_io(wb
) && dirty_writeback_interval
)
2265 wb_wakeup_delayed(wb
);
2269 * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2270 * write back the whole world.
2272 static void __wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2273 enum wb_reason reason
)
2275 struct bdi_writeback
*wb
;
2277 if (!bdi_has_dirty_io(bdi
))
2280 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2281 wb_start_writeback(wb
, reason
);
2284 void wakeup_flusher_threads_bdi(struct backing_dev_info
*bdi
,
2285 enum wb_reason reason
)
2288 __wakeup_flusher_threads_bdi(bdi
, reason
);
2293 * Wakeup the flusher threads to start writeback of all currently dirty pages
2295 void wakeup_flusher_threads(enum wb_reason reason
)
2297 struct backing_dev_info
*bdi
;
2300 * If we are expecting writeback progress we must submit plugged IO.
2302 blk_flush_plug(current
->plug
, true);
2305 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
)
2306 __wakeup_flusher_threads_bdi(bdi
, reason
);
2311 * Wake up bdi's periodically to make sure dirtytime inodes gets
2312 * written back periodically. We deliberately do *not* check the
2313 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2314 * kernel to be constantly waking up once there are any dirtytime
2315 * inodes on the system. So instead we define a separate delayed work
2316 * function which gets called much more rarely. (By default, only
2317 * once every 12 hours.)
2319 * If there is any other write activity going on in the file system,
2320 * this function won't be necessary. But if the only thing that has
2321 * happened on the file system is a dirtytime inode caused by an atime
2322 * update, we need this infrastructure below to make sure that inode
2323 * eventually gets pushed out to disk.
2325 static void wakeup_dirtytime_writeback(struct work_struct
*w
);
2326 static DECLARE_DELAYED_WORK(dirtytime_work
, wakeup_dirtytime_writeback
);
2328 static void wakeup_dirtytime_writeback(struct work_struct
*w
)
2330 struct backing_dev_info
*bdi
;
2333 list_for_each_entry_rcu(bdi
, &bdi_list
, bdi_list
) {
2334 struct bdi_writeback
*wb
;
2336 list_for_each_entry_rcu(wb
, &bdi
->wb_list
, bdi_node
)
2337 if (!list_empty(&wb
->b_dirty_time
))
2341 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2344 static int __init
start_dirtytime_writeback(void)
2346 schedule_delayed_work(&dirtytime_work
, dirtytime_expire_interval
* HZ
);
2349 __initcall(start_dirtytime_writeback
);
2351 int dirtytime_interval_handler(struct ctl_table
*table
, int write
,
2352 void *buffer
, size_t *lenp
, loff_t
*ppos
)
2356 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
2357 if (ret
== 0 && write
)
2358 mod_delayed_work(system_wq
, &dirtytime_work
, 0);
2363 * __mark_inode_dirty - internal function to mark an inode dirty
2365 * @inode: inode to mark
2366 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2367 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2368 * with I_DIRTY_PAGES.
2370 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2371 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2373 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2374 * instead of calling this directly.
2376 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2377 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2378 * even if they are later hashed, as they will have been marked dirty already.
2380 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2382 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2383 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2384 * the kernel-internal blockdev inode represents the dirtying time of the
2385 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2386 * page->mapping->host, so the page-dirtying time is recorded in the internal
2389 void __mark_inode_dirty(struct inode
*inode
, int flags
)
2391 struct super_block
*sb
= inode
->i_sb
;
2393 struct bdi_writeback
*wb
= NULL
;
2395 trace_writeback_mark_inode_dirty(inode
, flags
);
2397 if (flags
& I_DIRTY_INODE
) {
2399 * Inode timestamp update will piggback on this dirtying.
2400 * We tell ->dirty_inode callback that timestamps need to
2401 * be updated by setting I_DIRTY_TIME in flags.
2403 if (inode
->i_state
& I_DIRTY_TIME
) {
2404 spin_lock(&inode
->i_lock
);
2405 if (inode
->i_state
& I_DIRTY_TIME
) {
2406 inode
->i_state
&= ~I_DIRTY_TIME
;
2407 flags
|= I_DIRTY_TIME
;
2409 spin_unlock(&inode
->i_lock
);
2413 * Notify the filesystem about the inode being dirtied, so that
2414 * (if needed) it can update on-disk fields and journal the
2415 * inode. This is only needed when the inode itself is being
2416 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2417 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2419 trace_writeback_dirty_inode_start(inode
, flags
);
2420 if (sb
->s_op
->dirty_inode
)
2421 sb
->s_op
->dirty_inode(inode
,
2422 flags
& (I_DIRTY_INODE
| I_DIRTY_TIME
));
2423 trace_writeback_dirty_inode(inode
, flags
);
2425 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2426 flags
&= ~I_DIRTY_TIME
;
2429 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2430 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2431 * in one call to __mark_inode_dirty().)
2433 dirtytime
= flags
& I_DIRTY_TIME
;
2434 WARN_ON_ONCE(dirtytime
&& flags
!= I_DIRTY_TIME
);
2438 * Paired with smp_mb() in __writeback_single_inode() for the
2439 * following lockless i_state test. See there for details.
2443 if ((inode
->i_state
& flags
) == flags
)
2446 spin_lock(&inode
->i_lock
);
2447 if ((inode
->i_state
& flags
) != flags
) {
2448 const int was_dirty
= inode
->i_state
& I_DIRTY
;
2450 inode_attach_wb(inode
, NULL
);
2452 inode
->i_state
|= flags
;
2455 * Grab inode's wb early because it requires dropping i_lock and we
2456 * need to make sure following checks happen atomically with dirty
2457 * list handling so that we don't move inodes under flush worker's
2461 wb
= locked_inode_to_wb_and_lock_list(inode
);
2462 spin_lock(&inode
->i_lock
);
2466 * If the inode is queued for writeback by flush worker, just
2467 * update its dirty state. Once the flush worker is done with
2468 * the inode it will place it on the appropriate superblock
2469 * list, based upon its state.
2471 if (inode
->i_state
& I_SYNC_QUEUED
)
2475 * Only add valid (hashed) inodes to the superblock's
2476 * dirty list. Add blockdev inodes as well.
2478 if (!S_ISBLK(inode
->i_mode
)) {
2479 if (inode_unhashed(inode
))
2482 if (inode
->i_state
& I_FREEING
)
2486 * If the inode was already on b_dirty/b_io/b_more_io, don't
2487 * reposition it (that would break b_dirty time-ordering).
2490 struct list_head
*dirty_list
;
2491 bool wakeup_bdi
= false;
2493 inode
->dirtied_when
= jiffies
;
2495 inode
->dirtied_time_when
= jiffies
;
2497 if (inode
->i_state
& I_DIRTY
)
2498 dirty_list
= &wb
->b_dirty
;
2500 dirty_list
= &wb
->b_dirty_time
;
2502 wakeup_bdi
= inode_io_list_move_locked(inode
, wb
,
2505 spin_unlock(&wb
->list_lock
);
2506 spin_unlock(&inode
->i_lock
);
2507 trace_writeback_dirty_inode_enqueue(inode
);
2510 * If this is the first dirty inode for this bdi,
2511 * we have to wake-up the corresponding bdi thread
2512 * to make sure background write-back happens
2516 (wb
->bdi
->capabilities
& BDI_CAP_WRITEBACK
))
2517 wb_wakeup_delayed(wb
);
2523 spin_unlock(&wb
->list_lock
);
2524 spin_unlock(&inode
->i_lock
);
2526 EXPORT_SYMBOL(__mark_inode_dirty
);
2529 * The @s_sync_lock is used to serialise concurrent sync operations
2530 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2531 * Concurrent callers will block on the s_sync_lock rather than doing contending
2532 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2533 * has been issued up to the time this function is enter is guaranteed to be
2534 * completed by the time we have gained the lock and waited for all IO that is
2535 * in progress regardless of the order callers are granted the lock.
2537 static void wait_sb_inodes(struct super_block
*sb
)
2539 LIST_HEAD(sync_list
);
2542 * We need to be protected against the filesystem going from
2543 * r/o to r/w or vice versa.
2545 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2547 mutex_lock(&sb
->s_sync_lock
);
2550 * Splice the writeback list onto a temporary list to avoid waiting on
2551 * inodes that have started writeback after this point.
2553 * Use rcu_read_lock() to keep the inodes around until we have a
2554 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2555 * the local list because inodes can be dropped from either by writeback
2559 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2560 list_splice_init(&sb
->s_inodes_wb
, &sync_list
);
2563 * Data integrity sync. Must wait for all pages under writeback, because
2564 * there may have been pages dirtied before our sync call, but which had
2565 * writeout started before we write it out. In which case, the inode
2566 * may not be on the dirty list, but we still have to wait for that
2569 while (!list_empty(&sync_list
)) {
2570 struct inode
*inode
= list_first_entry(&sync_list
, struct inode
,
2572 struct address_space
*mapping
= inode
->i_mapping
;
2575 * Move each inode back to the wb list before we drop the lock
2576 * to preserve consistency between i_wb_list and the mapping
2577 * writeback tag. Writeback completion is responsible to remove
2578 * the inode from either list once the writeback tag is cleared.
2580 list_move_tail(&inode
->i_wb_list
, &sb
->s_inodes_wb
);
2583 * The mapping can appear untagged while still on-list since we
2584 * do not have the mapping lock. Skip it here, wb completion
2587 if (!mapping_tagged(mapping
, PAGECACHE_TAG_WRITEBACK
))
2590 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2592 spin_lock(&inode
->i_lock
);
2593 if (inode
->i_state
& (I_FREEING
|I_WILL_FREE
|I_NEW
)) {
2594 spin_unlock(&inode
->i_lock
);
2596 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2600 spin_unlock(&inode
->i_lock
);
2604 * We keep the error status of individual mapping so that
2605 * applications can catch the writeback error using fsync(2).
2606 * See filemap_fdatawait_keep_errors() for details.
2608 filemap_fdatawait_keep_errors(mapping
);
2615 spin_lock_irq(&sb
->s_inode_wblist_lock
);
2617 spin_unlock_irq(&sb
->s_inode_wblist_lock
);
2619 mutex_unlock(&sb
->s_sync_lock
);
2622 static void __writeback_inodes_sb_nr(struct super_block
*sb
, unsigned long nr
,
2623 enum wb_reason reason
, bool skip_if_busy
)
2625 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2626 DEFINE_WB_COMPLETION(done
, bdi
);
2627 struct wb_writeback_work work
= {
2629 .sync_mode
= WB_SYNC_NONE
,
2630 .tagged_writepages
= 1,
2636 if (!bdi_has_dirty_io(bdi
) || bdi
== &noop_backing_dev_info
)
2638 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2640 bdi_split_work_to_wbs(sb
->s_bdi
, &work
, skip_if_busy
);
2641 wb_wait_for_completion(&done
);
2645 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2646 * @sb: the superblock
2647 * @nr: the number of pages to write
2648 * @reason: reason why some writeback work initiated
2650 * Start writeback on some inodes on this super_block. No guarantees are made
2651 * on how many (if any) will be written, and this function does not wait
2652 * for IO completion of submitted IO.
2654 void writeback_inodes_sb_nr(struct super_block
*sb
,
2656 enum wb_reason reason
)
2658 __writeback_inodes_sb_nr(sb
, nr
, reason
, false);
2660 EXPORT_SYMBOL(writeback_inodes_sb_nr
);
2663 * writeback_inodes_sb - writeback dirty inodes from given super_block
2664 * @sb: the superblock
2665 * @reason: reason why some writeback work was initiated
2667 * Start writeback on some inodes on this super_block. No guarantees are made
2668 * on how many (if any) will be written, and this function does not wait
2669 * for IO completion of submitted IO.
2671 void writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2673 return writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
);
2675 EXPORT_SYMBOL(writeback_inodes_sb
);
2678 * try_to_writeback_inodes_sb - try to start writeback if none underway
2679 * @sb: the superblock
2680 * @reason: reason why some writeback work was initiated
2682 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2684 void try_to_writeback_inodes_sb(struct super_block
*sb
, enum wb_reason reason
)
2686 if (!down_read_trylock(&sb
->s_umount
))
2689 __writeback_inodes_sb_nr(sb
, get_nr_dirty_pages(), reason
, true);
2690 up_read(&sb
->s_umount
);
2692 EXPORT_SYMBOL(try_to_writeback_inodes_sb
);
2695 * sync_inodes_sb - sync sb inode pages
2696 * @sb: the superblock
2698 * This function writes and waits on any dirty inode belonging to this
2701 void sync_inodes_sb(struct super_block
*sb
)
2703 struct backing_dev_info
*bdi
= sb
->s_bdi
;
2704 DEFINE_WB_COMPLETION(done
, bdi
);
2705 struct wb_writeback_work work
= {
2707 .sync_mode
= WB_SYNC_ALL
,
2708 .nr_pages
= LONG_MAX
,
2711 .reason
= WB_REASON_SYNC
,
2716 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2717 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2718 * bdi_has_dirty() need to be written out too.
2720 if (bdi
== &noop_backing_dev_info
)
2722 WARN_ON(!rwsem_is_locked(&sb
->s_umount
));
2724 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2725 bdi_down_write_wb_switch_rwsem(bdi
);
2726 bdi_split_work_to_wbs(bdi
, &work
, false);
2727 wb_wait_for_completion(&done
);
2728 bdi_up_write_wb_switch_rwsem(bdi
);
2732 EXPORT_SYMBOL(sync_inodes_sb
);
2735 * write_inode_now - write an inode to disk
2736 * @inode: inode to write to disk
2737 * @sync: whether the write should be synchronous or not
2739 * This function commits an inode to disk immediately if it is dirty. This is
2740 * primarily needed by knfsd.
2742 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2744 int write_inode_now(struct inode
*inode
, int sync
)
2746 struct writeback_control wbc
= {
2747 .nr_to_write
= LONG_MAX
,
2748 .sync_mode
= sync
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2750 .range_end
= LLONG_MAX
,
2753 if (!mapping_can_writeback(inode
->i_mapping
))
2754 wbc
.nr_to_write
= 0;
2757 return writeback_single_inode(inode
, &wbc
);
2759 EXPORT_SYMBOL(write_inode_now
);
2762 * sync_inode_metadata - write an inode to disk
2763 * @inode: the inode to sync
2764 * @wait: wait for I/O to complete.
2766 * Write an inode to disk and adjust its dirty state after completion.
2768 * Note: only writes the actual inode, no associated data or other metadata.
2770 int sync_inode_metadata(struct inode
*inode
, int wait
)
2772 struct writeback_control wbc
= {
2773 .sync_mode
= wait
? WB_SYNC_ALL
: WB_SYNC_NONE
,
2774 .nr_to_write
= 0, /* metadata-only */
2777 return writeback_single_inode(inode
, &wbc
);
2779 EXPORT_SYMBOL(sync_inode_metadata
);