1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * fs/fs-writeback.c
4 *
5 * Copyright (C) 2002, Linus Torvalds.
6 *
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.
11 *
12 * 10Apr2002 Andrew Morton
13 * Split out of fs/inode.c
14 * Additions for address_space-based writeback
15 */
16
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>
22 #include <linux/fs.h>
23 #include <linux/mm.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>
32 #include "internal.h"
33
34 /*
35 * 4MB minimal write chunk size
36 */
37 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
38
39 /*
40 * Passed into wb_writeback(), essentially a subset of writeback_control
41 */
42 struct wb_writeback_work {
43 long nr_pages;
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? */
53
54 struct list_head list; /* pending work list */
55 struct wb_completion *done; /* set if the caller waits */
56 };
57
58 /*
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.
67 */
68 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
69
wb_inode(struct list_head * head)70 static inline struct inode *wb_inode(struct list_head *head)
71 {
72 return list_entry(head, struct inode, i_io_list);
73 }
74
75 /*
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.
79 */
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/writeback.h>
82
83 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
84
wb_io_lists_populated(struct bdi_writeback * wb)85 static bool wb_io_lists_populated(struct bdi_writeback *wb)
86 {
87 if (wb_has_dirty_io(wb)) {
88 return false;
89 } else {
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);
94 return true;
95 }
96 }
97
wb_io_lists_depopulated(struct bdi_writeback * wb)98 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
99 {
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);
105 }
106 }
107
108 /**
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}
113 *
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.
117 */
inode_io_list_move_locked(struct inode * inode,struct bdi_writeback * wb,struct list_head * head)118 static bool inode_io_list_move_locked(struct inode *inode,
119 struct bdi_writeback *wb,
120 struct list_head *head)
121 {
122 assert_spin_locked(&wb->list_lock);
123 assert_spin_locked(&inode->i_lock);
124 WARN_ON_ONCE(inode->i_state & I_FREEING);
125
126 list_move(&inode->i_io_list, head);
127
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);
131
132 wb_io_lists_depopulated(wb);
133 return false;
134 }
135
wb_wakeup(struct bdi_writeback * wb)136 static void wb_wakeup(struct bdi_writeback *wb)
137 {
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);
142 }
143
144 /*
145 * This function is used when the first inode for this wb is marked dirty. It
146 * wakes-up the corresponding bdi thread which should then take care of the
147 * periodic background write-out of dirty inodes. Since the write-out would
148 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
149 * set up a timer which wakes the bdi thread up later.
150 *
151 * Note, we wouldn't bother setting up the timer, but this function is on the
152 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
153 * by delaying the wake-up.
154 *
155 * We have to be careful not to postpone flush work if it is scheduled for
156 * earlier. Thus we use queue_delayed_work().
157 */
wb_wakeup_delayed(struct bdi_writeback * wb)158 static void wb_wakeup_delayed(struct bdi_writeback *wb)
159 {
160 unsigned long timeout;
161
162 timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
163 spin_lock_irq(&wb->work_lock);
164 if (test_bit(WB_registered, &wb->state))
165 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
166 spin_unlock_irq(&wb->work_lock);
167 }
168
finish_writeback_work(struct wb_writeback_work * work)169 static void finish_writeback_work(struct wb_writeback_work *work)
170 {
171 struct wb_completion *done = work->done;
172
173 if (work->auto_free)
174 kfree(work);
175 if (done) {
176 wait_queue_head_t *waitq = done->waitq;
177
178 /* @done can't be accessed after the following dec */
179 if (atomic_dec_and_test(&done->cnt))
180 wake_up_all(waitq);
181 }
182 }
183
wb_queue_work(struct bdi_writeback * wb,struct wb_writeback_work * work)184 static void wb_queue_work(struct bdi_writeback *wb,
185 struct wb_writeback_work *work)
186 {
187 trace_writeback_queue(wb, work);
188
189 if (work->done)
190 atomic_inc(&work->done->cnt);
191
192 spin_lock_irq(&wb->work_lock);
193
194 if (test_bit(WB_registered, &wb->state)) {
195 list_add_tail(&work->list, &wb->work_list);
196 mod_delayed_work(bdi_wq, &wb->dwork, 0);
197 } else
198 finish_writeback_work(work);
199
200 spin_unlock_irq(&wb->work_lock);
201 }
202
203 /**
204 * wb_wait_for_completion - wait for completion of bdi_writeback_works
205 * @done: target wb_completion
206 *
207 * Wait for one or more work items issued to @bdi with their ->done field
208 * set to @done, which should have been initialized with
209 * DEFINE_WB_COMPLETION(). This function returns after all such work items
210 * are completed. Work items which are waited upon aren't freed
211 * automatically on completion.
212 */
wb_wait_for_completion(struct wb_completion * done)213 void wb_wait_for_completion(struct wb_completion *done)
214 {
215 atomic_dec(&done->cnt); /* put down the initial count */
216 wait_event(*done->waitq, !atomic_read(&done->cnt));
217 }
218
219 #ifdef CONFIG_CGROUP_WRITEBACK
220
221 /*
222 * Parameters for foreign inode detection, see wbc_detach_inode() to see
223 * how they're used.
224 *
225 * These paramters are inherently heuristical as the detection target
226 * itself is fuzzy. All we want to do is detaching an inode from the
227 * current owner if it's being written to by some other cgroups too much.
228 *
229 * The current cgroup writeback is built on the assumption that multiple
230 * cgroups writing to the same inode concurrently is very rare and a mode
231 * of operation which isn't well supported. As such, the goal is not
232 * taking too long when a different cgroup takes over an inode while
233 * avoiding too aggressive flip-flops from occasional foreign writes.
234 *
235 * We record, very roughly, 2s worth of IO time history and if more than
236 * half of that is foreign, trigger the switch. The recording is quantized
237 * to 16 slots. To avoid tiny writes from swinging the decision too much,
238 * writes smaller than 1/8 of avg size are ignored.
239 */
240 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
241 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
242 #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
243 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
244
245 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
246 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
247 /* each slot's duration is 2s / 16 */
248 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
249 /* if foreign slots >= 8, switch */
250 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
251 /* one round can affect upto 5 slots */
252 #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
253
254 /*
255 * Maximum inodes per isw. A specific value has been chosen to make
256 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
257 */
258 #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
259 / sizeof(struct inode *))
260
261 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
262 static struct workqueue_struct *isw_wq;
263
__inode_attach_wb(struct inode * inode,struct folio * folio)264 void __inode_attach_wb(struct inode *inode, struct folio *folio)
265 {
266 struct backing_dev_info *bdi = inode_to_bdi(inode);
267 struct bdi_writeback *wb = NULL;
268
269 if (inode_cgwb_enabled(inode)) {
270 struct cgroup_subsys_state *memcg_css;
271
272 if (folio) {
273 memcg_css = mem_cgroup_css_from_folio(folio);
274 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
275 } else {
276 /* must pin memcg_css, see wb_get_create() */
277 memcg_css = task_get_css(current, memory_cgrp_id);
278 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
279 css_put(memcg_css);
280 }
281 }
282
283 if (!wb)
284 wb = &bdi->wb;
285
286 /*
287 * There may be multiple instances of this function racing to
288 * update the same inode. Use cmpxchg() to tell the winner.
289 */
290 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
291 wb_put(wb);
292 }
293
294 /**
295 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
296 * @inode: inode of interest with i_lock held
297 * @wb: target bdi_writeback
298 *
299 * Remove the inode from wb's io lists and if necessarily put onto b_attached
300 * list. Only inodes attached to cgwb's are kept on this list.
301 */
inode_cgwb_move_to_attached(struct inode * inode,struct bdi_writeback * wb)302 static void inode_cgwb_move_to_attached(struct inode *inode,
303 struct bdi_writeback *wb)
304 {
305 assert_spin_locked(&wb->list_lock);
306 assert_spin_locked(&inode->i_lock);
307 WARN_ON_ONCE(inode->i_state & I_FREEING);
308
309 inode->i_state &= ~I_SYNC_QUEUED;
310 if (wb != &wb->bdi->wb)
311 list_move(&inode->i_io_list, &wb->b_attached);
312 else
313 list_del_init(&inode->i_io_list);
314 wb_io_lists_depopulated(wb);
315 }
316
317 /**
318 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
319 * @inode: inode of interest with i_lock held
320 *
321 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
322 * held on entry and is released on return. The returned wb is guaranteed
323 * to stay @inode's associated wb until its list_lock is released.
324 */
325 static struct bdi_writeback *
locked_inode_to_wb_and_lock_list(struct inode * inode)326 locked_inode_to_wb_and_lock_list(struct inode *inode)
327 __releases(&inode->i_lock)
328 __acquires(&wb->list_lock)
329 {
330 while (true) {
331 struct bdi_writeback *wb = inode_to_wb(inode);
332
333 /*
334 * inode_to_wb() association is protected by both
335 * @inode->i_lock and @wb->list_lock but list_lock nests
336 * outside i_lock. Drop i_lock and verify that the
337 * association hasn't changed after acquiring list_lock.
338 */
339 wb_get(wb);
340 spin_unlock(&inode->i_lock);
341 spin_lock(&wb->list_lock);
342
343 /* i_wb may have changed inbetween, can't use inode_to_wb() */
344 if (likely(wb == inode->i_wb)) {
345 wb_put(wb); /* @inode already has ref */
346 return wb;
347 }
348
349 spin_unlock(&wb->list_lock);
350 wb_put(wb);
351 cpu_relax();
352 spin_lock(&inode->i_lock);
353 }
354 }
355
356 /**
357 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
358 * @inode: inode of interest
359 *
360 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
361 * on entry.
362 */
inode_to_wb_and_lock_list(struct inode * inode)363 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
364 __acquires(&wb->list_lock)
365 {
366 spin_lock(&inode->i_lock);
367 return locked_inode_to_wb_and_lock_list(inode);
368 }
369
370 struct inode_switch_wbs_context {
371 struct rcu_work work;
372
373 /*
374 * Multiple inodes can be switched at once. The switching procedure
375 * consists of two parts, separated by a RCU grace period. To make
376 * sure that the second part is executed for each inode gone through
377 * the first part, all inode pointers are placed into a NULL-terminated
378 * array embedded into struct inode_switch_wbs_context. Otherwise
379 * an inode could be left in a non-consistent state.
380 */
381 struct bdi_writeback *new_wb;
382 struct inode *inodes[];
383 };
384
bdi_down_write_wb_switch_rwsem(struct backing_dev_info * bdi)385 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
386 {
387 down_write(&bdi->wb_switch_rwsem);
388 }
389
bdi_up_write_wb_switch_rwsem(struct backing_dev_info * bdi)390 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
391 {
392 up_write(&bdi->wb_switch_rwsem);
393 }
394
inode_do_switch_wbs(struct inode * inode,struct bdi_writeback * old_wb,struct bdi_writeback * new_wb)395 static bool inode_do_switch_wbs(struct inode *inode,
396 struct bdi_writeback *old_wb,
397 struct bdi_writeback *new_wb)
398 {
399 struct address_space *mapping = inode->i_mapping;
400 XA_STATE(xas, &mapping->i_pages, 0);
401 struct folio *folio;
402 bool switched = false;
403
404 spin_lock(&inode->i_lock);
405 xa_lock_irq(&mapping->i_pages);
406
407 /*
408 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
409 * path owns the inode and we shouldn't modify ->i_io_list.
410 */
411 if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
412 goto skip_switch;
413
414 trace_inode_switch_wbs(inode, old_wb, new_wb);
415
416 /*
417 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
418 * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
419 * folios actually under writeback.
420 */
421 xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
422 if (folio_test_dirty(folio)) {
423 long nr = folio_nr_pages(folio);
424 wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr);
425 wb_stat_mod(new_wb, WB_RECLAIMABLE, nr);
426 }
427 }
428
429 xas_set(&xas, 0);
430 xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
431 long nr = folio_nr_pages(folio);
432 WARN_ON_ONCE(!folio_test_writeback(folio));
433 wb_stat_mod(old_wb, WB_WRITEBACK, -nr);
434 wb_stat_mod(new_wb, WB_WRITEBACK, nr);
435 }
436
437 if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
438 atomic_dec(&old_wb->writeback_inodes);
439 atomic_inc(&new_wb->writeback_inodes);
440 }
441
442 wb_get(new_wb);
443
444 /*
445 * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
446 * the specific list @inode was on is ignored and the @inode is put on
447 * ->b_dirty which is always correct including from ->b_dirty_time.
448 * The transfer preserves @inode->dirtied_when ordering. If the @inode
449 * was clean, it means it was on the b_attached list, so move it onto
450 * the b_attached list of @new_wb.
451 */
452 if (!list_empty(&inode->i_io_list)) {
453 inode->i_wb = new_wb;
454
455 if (inode->i_state & I_DIRTY_ALL) {
456 struct inode *pos;
457
458 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
459 if (time_after_eq(inode->dirtied_when,
460 pos->dirtied_when))
461 break;
462 inode_io_list_move_locked(inode, new_wb,
463 pos->i_io_list.prev);
464 } else {
465 inode_cgwb_move_to_attached(inode, new_wb);
466 }
467 } else {
468 inode->i_wb = new_wb;
469 }
470
471 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
472 inode->i_wb_frn_winner = 0;
473 inode->i_wb_frn_avg_time = 0;
474 inode->i_wb_frn_history = 0;
475 switched = true;
476 skip_switch:
477 /*
478 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
479 * ensures that the new wb is visible if they see !I_WB_SWITCH.
480 */
481 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
482
483 xa_unlock_irq(&mapping->i_pages);
484 spin_unlock(&inode->i_lock);
485
486 return switched;
487 }
488
inode_switch_wbs_work_fn(struct work_struct * work)489 static void inode_switch_wbs_work_fn(struct work_struct *work)
490 {
491 struct inode_switch_wbs_context *isw =
492 container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
493 struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
494 struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
495 struct bdi_writeback *new_wb = isw->new_wb;
496 unsigned long nr_switched = 0;
497 struct inode **inodep;
498
499 /*
500 * If @inode switches cgwb membership while sync_inodes_sb() is
501 * being issued, sync_inodes_sb() might miss it. Synchronize.
502 */
503 down_read(&bdi->wb_switch_rwsem);
504
505 /*
506 * By the time control reaches here, RCU grace period has passed
507 * since I_WB_SWITCH assertion and all wb stat update transactions
508 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
509 * synchronizing against the i_pages lock.
510 *
511 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
512 * gives us exclusion against all wb related operations on @inode
513 * including IO list manipulations and stat updates.
514 */
515 if (old_wb < new_wb) {
516 spin_lock(&old_wb->list_lock);
517 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
518 } else {
519 spin_lock(&new_wb->list_lock);
520 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
521 }
522
523 for (inodep = isw->inodes; *inodep; inodep++) {
524 WARN_ON_ONCE((*inodep)->i_wb != old_wb);
525 if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
526 nr_switched++;
527 }
528
529 spin_unlock(&new_wb->list_lock);
530 spin_unlock(&old_wb->list_lock);
531
532 up_read(&bdi->wb_switch_rwsem);
533
534 if (nr_switched) {
535 wb_wakeup(new_wb);
536 wb_put_many(old_wb, nr_switched);
537 }
538
539 for (inodep = isw->inodes; *inodep; inodep++)
540 iput(*inodep);
541 wb_put(new_wb);
542 kfree(isw);
543 atomic_dec(&isw_nr_in_flight);
544 }
545
inode_prepare_wbs_switch(struct inode * inode,struct bdi_writeback * new_wb)546 static bool inode_prepare_wbs_switch(struct inode *inode,
547 struct bdi_writeback *new_wb)
548 {
549 /*
550 * Paired with smp_mb() in cgroup_writeback_umount().
551 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
552 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
553 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
554 */
555 smp_mb();
556
557 if (IS_DAX(inode))
558 return false;
559
560 /* while holding I_WB_SWITCH, no one else can update the association */
561 spin_lock(&inode->i_lock);
562 if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
563 inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
564 inode_to_wb(inode) == new_wb) {
565 spin_unlock(&inode->i_lock);
566 return false;
567 }
568 inode->i_state |= I_WB_SWITCH;
569 __iget(inode);
570 spin_unlock(&inode->i_lock);
571
572 return true;
573 }
574
575 /**
576 * inode_switch_wbs - change the wb association of an inode
577 * @inode: target inode
578 * @new_wb_id: ID of the new wb
579 *
580 * Switch @inode's wb association to the wb identified by @new_wb_id. The
581 * switching is performed asynchronously and may fail silently.
582 */
inode_switch_wbs(struct inode * inode,int new_wb_id)583 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
584 {
585 struct backing_dev_info *bdi = inode_to_bdi(inode);
586 struct cgroup_subsys_state *memcg_css;
587 struct inode_switch_wbs_context *isw;
588
589 /* noop if seems to be already in progress */
590 if (inode->i_state & I_WB_SWITCH)
591 return;
592
593 /* avoid queueing a new switch if too many are already in flight */
594 if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
595 return;
596
597 isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC);
598 if (!isw)
599 return;
600
601 atomic_inc(&isw_nr_in_flight);
602
603 /* find and pin the new wb */
604 rcu_read_lock();
605 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
606 if (memcg_css && !css_tryget(memcg_css))
607 memcg_css = NULL;
608 rcu_read_unlock();
609 if (!memcg_css)
610 goto out_free;
611
612 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
613 css_put(memcg_css);
614 if (!isw->new_wb)
615 goto out_free;
616
617 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
618 goto out_free;
619
620 isw->inodes[0] = inode;
621
622 /*
623 * In addition to synchronizing among switchers, I_WB_SWITCH tells
624 * the RCU protected stat update paths to grab the i_page
625 * lock so that stat transfer can synchronize against them.
626 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
627 */
628 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
629 queue_rcu_work(isw_wq, &isw->work);
630 return;
631
632 out_free:
633 atomic_dec(&isw_nr_in_flight);
634 if (isw->new_wb)
635 wb_put(isw->new_wb);
636 kfree(isw);
637 }
638
isw_prepare_wbs_switch(struct inode_switch_wbs_context * isw,struct list_head * list,int * nr)639 static bool isw_prepare_wbs_switch(struct inode_switch_wbs_context *isw,
640 struct list_head *list, int *nr)
641 {
642 struct inode *inode;
643
644 list_for_each_entry(inode, list, i_io_list) {
645 if (!inode_prepare_wbs_switch(inode, isw->new_wb))
646 continue;
647
648 isw->inodes[*nr] = inode;
649 (*nr)++;
650
651 if (*nr >= WB_MAX_INODES_PER_ISW - 1)
652 return true;
653 }
654 return false;
655 }
656
657 /**
658 * cleanup_offline_cgwb - detach associated inodes
659 * @wb: target wb
660 *
661 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
662 * to eventually release the dying @wb. Returns %true if not all inodes were
663 * switched and the function has to be restarted.
664 */
cleanup_offline_cgwb(struct bdi_writeback * wb)665 bool cleanup_offline_cgwb(struct bdi_writeback *wb)
666 {
667 struct cgroup_subsys_state *memcg_css;
668 struct inode_switch_wbs_context *isw;
669 int nr;
670 bool restart = false;
671
672 isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW),
673 GFP_KERNEL);
674 if (!isw)
675 return restart;
676
677 atomic_inc(&isw_nr_in_flight);
678
679 for (memcg_css = wb->memcg_css->parent; memcg_css;
680 memcg_css = memcg_css->parent) {
681 isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
682 if (isw->new_wb)
683 break;
684 }
685 if (unlikely(!isw->new_wb))
686 isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
687
688 nr = 0;
689 spin_lock(&wb->list_lock);
690 /*
691 * In addition to the inodes that have completed writeback, also switch
692 * cgwbs for those inodes only with dirty timestamps. Otherwise, those
693 * inodes won't be written back for a long time when lazytime is
694 * enabled, and thus pinning the dying cgwbs. It won't break the
695 * bandwidth restrictions, as writeback of inode metadata is not
696 * accounted for.
697 */
698 restart = isw_prepare_wbs_switch(isw, &wb->b_attached, &nr);
699 if (!restart)
700 restart = isw_prepare_wbs_switch(isw, &wb->b_dirty_time, &nr);
701 spin_unlock(&wb->list_lock);
702
703 /* no attached inodes? bail out */
704 if (nr == 0) {
705 atomic_dec(&isw_nr_in_flight);
706 wb_put(isw->new_wb);
707 kfree(isw);
708 return restart;
709 }
710
711 /*
712 * In addition to synchronizing among switchers, I_WB_SWITCH tells
713 * the RCU protected stat update paths to grab the i_page
714 * lock so that stat transfer can synchronize against them.
715 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
716 */
717 INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
718 queue_rcu_work(isw_wq, &isw->work);
719
720 return restart;
721 }
722
723 /**
724 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
725 * @wbc: writeback_control of interest
726 * @inode: target inode
727 *
728 * @inode is locked and about to be written back under the control of @wbc.
729 * Record @inode's writeback context into @wbc and unlock the i_lock. On
730 * writeback completion, wbc_detach_inode() should be called. This is used
731 * to track the cgroup writeback context.
732 */
wbc_attach_and_unlock_inode(struct writeback_control * wbc,struct inode * inode)733 static void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
734 struct inode *inode)
735 __releases(&inode->i_lock)
736 {
737 if (!inode_cgwb_enabled(inode)) {
738 spin_unlock(&inode->i_lock);
739 return;
740 }
741
742 wbc->wb = inode_to_wb(inode);
743 wbc->inode = inode;
744
745 wbc->wb_id = wbc->wb->memcg_css->id;
746 wbc->wb_lcand_id = inode->i_wb_frn_winner;
747 wbc->wb_tcand_id = 0;
748 wbc->wb_bytes = 0;
749 wbc->wb_lcand_bytes = 0;
750 wbc->wb_tcand_bytes = 0;
751
752 wb_get(wbc->wb);
753 spin_unlock(&inode->i_lock);
754
755 /*
756 * A dying wb indicates that either the blkcg associated with the
757 * memcg changed or the associated memcg is dying. In the first
758 * case, a replacement wb should already be available and we should
759 * refresh the wb immediately. In the second case, trying to
760 * refresh will keep failing.
761 */
762 if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
763 inode_switch_wbs(inode, wbc->wb_id);
764 }
765
766 /**
767 * wbc_attach_fdatawrite_inode - associate wbc and inode for fdatawrite
768 * @wbc: writeback_control of interest
769 * @inode: target inode
770 *
771 * This function is to be used by __filemap_fdatawrite_range(), which is an
772 * alternative entry point into writeback code, and first ensures @inode is
773 * associated with a bdi_writeback and attaches it to @wbc.
774 */
wbc_attach_fdatawrite_inode(struct writeback_control * wbc,struct inode * inode)775 void wbc_attach_fdatawrite_inode(struct writeback_control *wbc,
776 struct inode *inode)
777 {
778 spin_lock(&inode->i_lock);
779 inode_attach_wb(inode, NULL);
780 wbc_attach_and_unlock_inode(wbc, inode);
781 }
782 EXPORT_SYMBOL_GPL(wbc_attach_fdatawrite_inode);
783
784 /**
785 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
786 * @wbc: writeback_control of the just finished writeback
787 *
788 * To be called after a writeback attempt of an inode finishes and undoes
789 * wbc_attach_and_unlock_inode(). Can be called under any context.
790 *
791 * As concurrent write sharing of an inode is expected to be very rare and
792 * memcg only tracks page ownership on first-use basis severely confining
793 * the usefulness of such sharing, cgroup writeback tracks ownership
794 * per-inode. While the support for concurrent write sharing of an inode
795 * is deemed unnecessary, an inode being written to by different cgroups at
796 * different points in time is a lot more common, and, more importantly,
797 * charging only by first-use can too readily lead to grossly incorrect
798 * behaviors (single foreign page can lead to gigabytes of writeback to be
799 * incorrectly attributed).
800 *
801 * To resolve this issue, cgroup writeback detects the majority dirtier of
802 * an inode and transfers the ownership to it. To avoid unnecessary
803 * oscillation, the detection mechanism keeps track of history and gives
804 * out the switch verdict only if the foreign usage pattern is stable over
805 * a certain amount of time and/or writeback attempts.
806 *
807 * On each writeback attempt, @wbc tries to detect the majority writer
808 * using Boyer-Moore majority vote algorithm. In addition to the byte
809 * count from the majority voting, it also counts the bytes written for the
810 * current wb and the last round's winner wb (max of last round's current
811 * wb, the winner from two rounds ago, and the last round's majority
812 * candidate). Keeping track of the historical winner helps the algorithm
813 * to semi-reliably detect the most active writer even when it's not the
814 * absolute majority.
815 *
816 * Once the winner of the round is determined, whether the winner is
817 * foreign or not and how much IO time the round consumed is recorded in
818 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
819 * over a certain threshold, the switch verdict is given.
820 */
wbc_detach_inode(struct writeback_control * wbc)821 void wbc_detach_inode(struct writeback_control *wbc)
822 {
823 struct bdi_writeback *wb = wbc->wb;
824 struct inode *inode = wbc->inode;
825 unsigned long avg_time, max_bytes, max_time;
826 u16 history;
827 int max_id;
828
829 if (!wb)
830 return;
831
832 history = inode->i_wb_frn_history;
833 avg_time = inode->i_wb_frn_avg_time;
834
835 /* pick the winner of this round */
836 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
837 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
838 max_id = wbc->wb_id;
839 max_bytes = wbc->wb_bytes;
840 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
841 max_id = wbc->wb_lcand_id;
842 max_bytes = wbc->wb_lcand_bytes;
843 } else {
844 max_id = wbc->wb_tcand_id;
845 max_bytes = wbc->wb_tcand_bytes;
846 }
847
848 /*
849 * Calculate the amount of IO time the winner consumed and fold it
850 * into the running average kept per inode. If the consumed IO
851 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
852 * deciding whether to switch or not. This is to prevent one-off
853 * small dirtiers from skewing the verdict.
854 */
855 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
856 wb->avg_write_bandwidth);
857 if (avg_time)
858 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
859 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
860 else
861 avg_time = max_time; /* immediate catch up on first run */
862
863 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
864 int slots;
865
866 /*
867 * The switch verdict is reached if foreign wb's consume
868 * more than a certain proportion of IO time in a
869 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
870 * history mask where each bit represents one sixteenth of
871 * the period. Determine the number of slots to shift into
872 * history from @max_time.
873 */
874 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
875 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
876 history <<= slots;
877 if (wbc->wb_id != max_id)
878 history |= (1U << slots) - 1;
879
880 if (history)
881 trace_inode_foreign_history(inode, wbc, history);
882
883 /*
884 * Switch if the current wb isn't the consistent winner.
885 * If there are multiple closely competing dirtiers, the
886 * inode may switch across them repeatedly over time, which
887 * is okay. The main goal is avoiding keeping an inode on
888 * the wrong wb for an extended period of time.
889 */
890 if (hweight16(history) > WB_FRN_HIST_THR_SLOTS)
891 inode_switch_wbs(inode, max_id);
892 }
893
894 /*
895 * Multiple instances of this function may race to update the
896 * following fields but we don't mind occassional inaccuracies.
897 */
898 inode->i_wb_frn_winner = max_id;
899 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
900 inode->i_wb_frn_history = history;
901
902 wb_put(wbc->wb);
903 wbc->wb = NULL;
904 }
905 EXPORT_SYMBOL_GPL(wbc_detach_inode);
906
907 /**
908 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
909 * @wbc: writeback_control of the writeback in progress
910 * @folio: folio being written out
911 * @bytes: number of bytes being written out
912 *
913 * @bytes from @folio are about to written out during the writeback
914 * controlled by @wbc. Keep the book for foreign inode detection. See
915 * wbc_detach_inode().
916 */
wbc_account_cgroup_owner(struct writeback_control * wbc,struct folio * folio,size_t bytes)917 void wbc_account_cgroup_owner(struct writeback_control *wbc, struct folio *folio,
918 size_t bytes)
919 {
920 struct cgroup_subsys_state *css;
921 int id;
922
923 /*
924 * pageout() path doesn't attach @wbc to the inode being written
925 * out. This is intentional as we don't want the function to block
926 * behind a slow cgroup. Ultimately, we want pageout() to kick off
927 * regular writeback instead of writing things out itself.
928 */
929 if (!wbc->wb || wbc->no_cgroup_owner)
930 return;
931
932 css = mem_cgroup_css_from_folio(folio);
933 /* dead cgroups shouldn't contribute to inode ownership arbitration */
934 if (!(css->flags & CSS_ONLINE))
935 return;
936
937 id = css->id;
938
939 if (id == wbc->wb_id) {
940 wbc->wb_bytes += bytes;
941 return;
942 }
943
944 if (id == wbc->wb_lcand_id)
945 wbc->wb_lcand_bytes += bytes;
946
947 /* Boyer-Moore majority vote algorithm */
948 if (!wbc->wb_tcand_bytes)
949 wbc->wb_tcand_id = id;
950 if (id == wbc->wb_tcand_id)
951 wbc->wb_tcand_bytes += bytes;
952 else
953 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
954 }
955 EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
956
957 /**
958 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
959 * @wb: target bdi_writeback to split @nr_pages to
960 * @nr_pages: number of pages to write for the whole bdi
961 *
962 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
963 * relation to the total write bandwidth of all wb's w/ dirty inodes on
964 * @wb->bdi.
965 */
wb_split_bdi_pages(struct bdi_writeback * wb,long nr_pages)966 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
967 {
968 unsigned long this_bw = wb->avg_write_bandwidth;
969 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
970
971 if (nr_pages == LONG_MAX)
972 return LONG_MAX;
973
974 /*
975 * This may be called on clean wb's and proportional distribution
976 * may not make sense, just use the original @nr_pages in those
977 * cases. In general, we wanna err on the side of writing more.
978 */
979 if (!tot_bw || this_bw >= tot_bw)
980 return nr_pages;
981 else
982 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
983 }
984
985 /**
986 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
987 * @bdi: target backing_dev_info
988 * @base_work: wb_writeback_work to issue
989 * @skip_if_busy: skip wb's which already have writeback in progress
990 *
991 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
992 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
993 * distributed to the busy wbs according to each wb's proportion in the
994 * total active write bandwidth of @bdi.
995 */
bdi_split_work_to_wbs(struct backing_dev_info * bdi,struct wb_writeback_work * base_work,bool skip_if_busy)996 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
997 struct wb_writeback_work *base_work,
998 bool skip_if_busy)
999 {
1000 struct bdi_writeback *last_wb = NULL;
1001 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
1002 struct bdi_writeback, bdi_node);
1003
1004 might_sleep();
1005 restart:
1006 rcu_read_lock();
1007 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
1008 DEFINE_WB_COMPLETION(fallback_work_done, bdi);
1009 struct wb_writeback_work fallback_work;
1010 struct wb_writeback_work *work;
1011 long nr_pages;
1012
1013 if (last_wb) {
1014 wb_put(last_wb);
1015 last_wb = NULL;
1016 }
1017
1018 /* SYNC_ALL writes out I_DIRTY_TIME too */
1019 if (!wb_has_dirty_io(wb) &&
1020 (base_work->sync_mode == WB_SYNC_NONE ||
1021 list_empty(&wb->b_dirty_time)))
1022 continue;
1023 if (skip_if_busy && writeback_in_progress(wb))
1024 continue;
1025
1026 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
1027
1028 work = kmalloc(sizeof(*work), GFP_ATOMIC);
1029 if (work) {
1030 *work = *base_work;
1031 work->nr_pages = nr_pages;
1032 work->auto_free = 1;
1033 wb_queue_work(wb, work);
1034 continue;
1035 }
1036
1037 /*
1038 * If wb_tryget fails, the wb has been shutdown, skip it.
1039 *
1040 * Pin @wb so that it stays on @bdi->wb_list. This allows
1041 * continuing iteration from @wb after dropping and
1042 * regrabbing rcu read lock.
1043 */
1044 if (!wb_tryget(wb))
1045 continue;
1046
1047 /* alloc failed, execute synchronously using on-stack fallback */
1048 work = &fallback_work;
1049 *work = *base_work;
1050 work->nr_pages = nr_pages;
1051 work->auto_free = 0;
1052 work->done = &fallback_work_done;
1053
1054 wb_queue_work(wb, work);
1055 last_wb = wb;
1056
1057 rcu_read_unlock();
1058 wb_wait_for_completion(&fallback_work_done);
1059 goto restart;
1060 }
1061 rcu_read_unlock();
1062
1063 if (last_wb)
1064 wb_put(last_wb);
1065 }
1066
1067 /**
1068 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1069 * @bdi_id: target bdi id
1070 * @memcg_id: target memcg css id
1071 * @reason: reason why some writeback work initiated
1072 * @done: target wb_completion
1073 *
1074 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1075 * with the specified parameters.
1076 */
cgroup_writeback_by_id(u64 bdi_id,int memcg_id,enum wb_reason reason,struct wb_completion * done)1077 int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
1078 enum wb_reason reason, struct wb_completion *done)
1079 {
1080 struct backing_dev_info *bdi;
1081 struct cgroup_subsys_state *memcg_css;
1082 struct bdi_writeback *wb;
1083 struct wb_writeback_work *work;
1084 unsigned long dirty;
1085 int ret;
1086
1087 /* lookup bdi and memcg */
1088 bdi = bdi_get_by_id(bdi_id);
1089 if (!bdi)
1090 return -ENOENT;
1091
1092 rcu_read_lock();
1093 memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1094 if (memcg_css && !css_tryget(memcg_css))
1095 memcg_css = NULL;
1096 rcu_read_unlock();
1097 if (!memcg_css) {
1098 ret = -ENOENT;
1099 goto out_bdi_put;
1100 }
1101
1102 /*
1103 * And find the associated wb. If the wb isn't there already
1104 * there's nothing to flush, don't create one.
1105 */
1106 wb = wb_get_lookup(bdi, memcg_css);
1107 if (!wb) {
1108 ret = -ENOENT;
1109 goto out_css_put;
1110 }
1111
1112 /*
1113 * The caller is attempting to write out most of
1114 * the currently dirty pages. Let's take the current dirty page
1115 * count and inflate it by 25% which should be large enough to
1116 * flush out most dirty pages while avoiding getting livelocked by
1117 * concurrent dirtiers.
1118 *
1119 * BTW the memcg stats are flushed periodically and this is best-effort
1120 * estimation, so some potential error is ok.
1121 */
1122 dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1123 dirty = dirty * 10 / 8;
1124
1125 /* issue the writeback work */
1126 work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1127 if (work) {
1128 work->nr_pages = dirty;
1129 work->sync_mode = WB_SYNC_NONE;
1130 work->range_cyclic = 1;
1131 work->reason = reason;
1132 work->done = done;
1133 work->auto_free = 1;
1134 wb_queue_work(wb, work);
1135 ret = 0;
1136 } else {
1137 ret = -ENOMEM;
1138 }
1139
1140 wb_put(wb);
1141 out_css_put:
1142 css_put(memcg_css);
1143 out_bdi_put:
1144 bdi_put(bdi);
1145 return ret;
1146 }
1147
1148 /**
1149 * cgroup_writeback_umount - flush inode wb switches for umount
1150 * @sb: target super_block
1151 *
1152 * This function is called when a super_block is about to be destroyed and
1153 * flushes in-flight inode wb switches. An inode wb switch goes through
1154 * RCU and then workqueue, so the two need to be flushed in order to ensure
1155 * that all previously scheduled switches are finished. As wb switches are
1156 * rare occurrences and synchronize_rcu() can take a while, perform
1157 * flushing iff wb switches are in flight.
1158 */
cgroup_writeback_umount(struct super_block * sb)1159 void cgroup_writeback_umount(struct super_block *sb)
1160 {
1161
1162 if (!(sb->s_bdi->capabilities & BDI_CAP_WRITEBACK))
1163 return;
1164
1165 /*
1166 * SB_ACTIVE should be reliably cleared before checking
1167 * isw_nr_in_flight, see generic_shutdown_super().
1168 */
1169 smp_mb();
1170
1171 if (atomic_read(&isw_nr_in_flight)) {
1172 /*
1173 * Use rcu_barrier() to wait for all pending callbacks to
1174 * ensure that all in-flight wb switches are in the workqueue.
1175 */
1176 rcu_barrier();
1177 flush_workqueue(isw_wq);
1178 }
1179 }
1180
cgroup_writeback_init(void)1181 static int __init cgroup_writeback_init(void)
1182 {
1183 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1184 if (!isw_wq)
1185 return -ENOMEM;
1186 return 0;
1187 }
1188 fs_initcall(cgroup_writeback_init);
1189
1190 #else /* CONFIG_CGROUP_WRITEBACK */
1191
bdi_down_write_wb_switch_rwsem(struct backing_dev_info * bdi)1192 static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
bdi_up_write_wb_switch_rwsem(struct backing_dev_info * bdi)1193 static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1194
inode_cgwb_move_to_attached(struct inode * inode,struct bdi_writeback * wb)1195 static void inode_cgwb_move_to_attached(struct inode *inode,
1196 struct bdi_writeback *wb)
1197 {
1198 assert_spin_locked(&wb->list_lock);
1199 assert_spin_locked(&inode->i_lock);
1200 WARN_ON_ONCE(inode->i_state & I_FREEING);
1201
1202 inode->i_state &= ~I_SYNC_QUEUED;
1203 list_del_init(&inode->i_io_list);
1204 wb_io_lists_depopulated(wb);
1205 }
1206
1207 static struct bdi_writeback *
locked_inode_to_wb_and_lock_list(struct inode * inode)1208 locked_inode_to_wb_and_lock_list(struct inode *inode)
1209 __releases(&inode->i_lock)
1210 __acquires(&wb->list_lock)
1211 {
1212 struct bdi_writeback *wb = inode_to_wb(inode);
1213
1214 spin_unlock(&inode->i_lock);
1215 spin_lock(&wb->list_lock);
1216 return wb;
1217 }
1218
inode_to_wb_and_lock_list(struct inode * inode)1219 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1220 __acquires(&wb->list_lock)
1221 {
1222 struct bdi_writeback *wb = inode_to_wb(inode);
1223
1224 spin_lock(&wb->list_lock);
1225 return wb;
1226 }
1227
wb_split_bdi_pages(struct bdi_writeback * wb,long nr_pages)1228 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1229 {
1230 return nr_pages;
1231 }
1232
bdi_split_work_to_wbs(struct backing_dev_info * bdi,struct wb_writeback_work * base_work,bool skip_if_busy)1233 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1234 struct wb_writeback_work *base_work,
1235 bool skip_if_busy)
1236 {
1237 might_sleep();
1238
1239 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1240 base_work->auto_free = 0;
1241 wb_queue_work(&bdi->wb, base_work);
1242 }
1243 }
1244
wbc_attach_and_unlock_inode(struct writeback_control * wbc,struct inode * inode)1245 static inline void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
1246 struct inode *inode)
1247 __releases(&inode->i_lock)
1248 {
1249 spin_unlock(&inode->i_lock);
1250 }
1251
1252 #endif /* CONFIG_CGROUP_WRITEBACK */
1253
1254 /*
1255 * Add in the number of potentially dirty inodes, because each inode
1256 * write can dirty pagecache in the underlying blockdev.
1257 */
get_nr_dirty_pages(void)1258 static unsigned long get_nr_dirty_pages(void)
1259 {
1260 return global_node_page_state(NR_FILE_DIRTY) +
1261 get_nr_dirty_inodes();
1262 }
1263
wb_start_writeback(struct bdi_writeback * wb,enum wb_reason reason)1264 static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1265 {
1266 if (!wb_has_dirty_io(wb))
1267 return;
1268
1269 /*
1270 * All callers of this function want to start writeback of all
1271 * dirty pages. Places like vmscan can call this at a very
1272 * high frequency, causing pointless allocations of tons of
1273 * work items and keeping the flusher threads busy retrieving
1274 * that work. Ensure that we only allow one of them pending and
1275 * inflight at the time.
1276 */
1277 if (test_bit(WB_start_all, &wb->state) ||
1278 test_and_set_bit(WB_start_all, &wb->state))
1279 return;
1280
1281 wb->start_all_reason = reason;
1282 wb_wakeup(wb);
1283 }
1284
1285 /**
1286 * wb_start_background_writeback - start background writeback
1287 * @wb: bdi_writback to write from
1288 *
1289 * Description:
1290 * This makes sure WB_SYNC_NONE background writeback happens. When
1291 * this function returns, it is only guaranteed that for given wb
1292 * some IO is happening if we are over background dirty threshold.
1293 * Caller need not hold sb s_umount semaphore.
1294 */
wb_start_background_writeback(struct bdi_writeback * wb)1295 void wb_start_background_writeback(struct bdi_writeback *wb)
1296 {
1297 /*
1298 * We just wake up the flusher thread. It will perform background
1299 * writeback as soon as there is no other work to do.
1300 */
1301 trace_writeback_wake_background(wb);
1302 wb_wakeup(wb);
1303 }
1304
1305 /*
1306 * Remove the inode from the writeback list it is on.
1307 */
inode_io_list_del(struct inode * inode)1308 void inode_io_list_del(struct inode *inode)
1309 {
1310 struct bdi_writeback *wb;
1311
1312 wb = inode_to_wb_and_lock_list(inode);
1313 spin_lock(&inode->i_lock);
1314
1315 inode->i_state &= ~I_SYNC_QUEUED;
1316 list_del_init(&inode->i_io_list);
1317 wb_io_lists_depopulated(wb);
1318
1319 spin_unlock(&inode->i_lock);
1320 spin_unlock(&wb->list_lock);
1321 }
1322 EXPORT_SYMBOL(inode_io_list_del);
1323
1324 /*
1325 * mark an inode as under writeback on the sb
1326 */
sb_mark_inode_writeback(struct inode * inode)1327 void sb_mark_inode_writeback(struct inode *inode)
1328 {
1329 struct super_block *sb = inode->i_sb;
1330 unsigned long flags;
1331
1332 if (list_empty(&inode->i_wb_list)) {
1333 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1334 if (list_empty(&inode->i_wb_list)) {
1335 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1336 trace_sb_mark_inode_writeback(inode);
1337 }
1338 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1339 }
1340 }
1341
1342 /*
1343 * clear an inode as under writeback on the sb
1344 */
sb_clear_inode_writeback(struct inode * inode)1345 void sb_clear_inode_writeback(struct inode *inode)
1346 {
1347 struct super_block *sb = inode->i_sb;
1348 unsigned long flags;
1349
1350 if (!list_empty(&inode->i_wb_list)) {
1351 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1352 if (!list_empty(&inode->i_wb_list)) {
1353 list_del_init(&inode->i_wb_list);
1354 trace_sb_clear_inode_writeback(inode);
1355 }
1356 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1357 }
1358 }
1359
1360 /*
1361 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1362 * furthest end of its superblock's dirty-inode list.
1363 *
1364 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1365 * already the most-recently-dirtied inode on the b_dirty list. If that is
1366 * the case then the inode must have been redirtied while it was being written
1367 * out and we don't reset its dirtied_when.
1368 */
redirty_tail_locked(struct inode * inode,struct bdi_writeback * wb)1369 static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1370 {
1371 assert_spin_locked(&inode->i_lock);
1372
1373 inode->i_state &= ~I_SYNC_QUEUED;
1374 /*
1375 * When the inode is being freed just don't bother with dirty list
1376 * tracking. Flush worker will ignore this inode anyway and it will
1377 * trigger assertions in inode_io_list_move_locked().
1378 */
1379 if (inode->i_state & I_FREEING) {
1380 list_del_init(&inode->i_io_list);
1381 wb_io_lists_depopulated(wb);
1382 return;
1383 }
1384 if (!list_empty(&wb->b_dirty)) {
1385 struct inode *tail;
1386
1387 tail = wb_inode(wb->b_dirty.next);
1388 if (time_before(inode->dirtied_when, tail->dirtied_when))
1389 inode->dirtied_when = jiffies;
1390 }
1391 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1392 }
1393
redirty_tail(struct inode * inode,struct bdi_writeback * wb)1394 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1395 {
1396 spin_lock(&inode->i_lock);
1397 redirty_tail_locked(inode, wb);
1398 spin_unlock(&inode->i_lock);
1399 }
1400
1401 /*
1402 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1403 */
requeue_io(struct inode * inode,struct bdi_writeback * wb)1404 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1405 {
1406 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1407 }
1408
inode_sync_complete(struct inode * inode)1409 static void inode_sync_complete(struct inode *inode)
1410 {
1411 assert_spin_locked(&inode->i_lock);
1412
1413 inode->i_state &= ~I_SYNC;
1414 /* If inode is clean an unused, put it into LRU now... */
1415 inode_add_lru(inode);
1416 /* Called with inode->i_lock which ensures memory ordering. */
1417 inode_wake_up_bit(inode, __I_SYNC);
1418 }
1419
inode_dirtied_after(struct inode * inode,unsigned long t)1420 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1421 {
1422 bool ret = time_after(inode->dirtied_when, t);
1423 #ifndef CONFIG_64BIT
1424 /*
1425 * For inodes being constantly redirtied, dirtied_when can get stuck.
1426 * It _appears_ to be in the future, but is actually in distant past.
1427 * This test is necessary to prevent such wrapped-around relative times
1428 * from permanently stopping the whole bdi writeback.
1429 */
1430 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1431 #endif
1432 return ret;
1433 }
1434
1435 /*
1436 * Move expired (dirtied before dirtied_before) dirty inodes from
1437 * @delaying_queue to @dispatch_queue.
1438 */
move_expired_inodes(struct list_head * delaying_queue,struct list_head * dispatch_queue,unsigned long dirtied_before)1439 static int move_expired_inodes(struct list_head *delaying_queue,
1440 struct list_head *dispatch_queue,
1441 unsigned long dirtied_before)
1442 {
1443 LIST_HEAD(tmp);
1444 struct list_head *pos, *node;
1445 struct super_block *sb = NULL;
1446 struct inode *inode;
1447 int do_sb_sort = 0;
1448 int moved = 0;
1449
1450 while (!list_empty(delaying_queue)) {
1451 inode = wb_inode(delaying_queue->prev);
1452 if (inode_dirtied_after(inode, dirtied_before))
1453 break;
1454 spin_lock(&inode->i_lock);
1455 list_move(&inode->i_io_list, &tmp);
1456 moved++;
1457 inode->i_state |= I_SYNC_QUEUED;
1458 spin_unlock(&inode->i_lock);
1459 if (sb_is_blkdev_sb(inode->i_sb))
1460 continue;
1461 if (sb && sb != inode->i_sb)
1462 do_sb_sort = 1;
1463 sb = inode->i_sb;
1464 }
1465
1466 /* just one sb in list, splice to dispatch_queue and we're done */
1467 if (!do_sb_sort) {
1468 list_splice(&tmp, dispatch_queue);
1469 goto out;
1470 }
1471
1472 /*
1473 * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1474 * we don't take inode->i_lock here because it is just a pointless overhead.
1475 * Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1476 * fully under our control.
1477 */
1478 while (!list_empty(&tmp)) {
1479 sb = wb_inode(tmp.prev)->i_sb;
1480 list_for_each_prev_safe(pos, node, &tmp) {
1481 inode = wb_inode(pos);
1482 if (inode->i_sb == sb)
1483 list_move(&inode->i_io_list, dispatch_queue);
1484 }
1485 }
1486 out:
1487 return moved;
1488 }
1489
1490 /*
1491 * Queue all expired dirty inodes for io, eldest first.
1492 * Before
1493 * newly dirtied b_dirty b_io b_more_io
1494 * =============> gf edc BA
1495 * After
1496 * newly dirtied b_dirty b_io b_more_io
1497 * =============> g fBAedc
1498 * |
1499 * +--> dequeue for IO
1500 */
queue_io(struct bdi_writeback * wb,struct wb_writeback_work * work,unsigned long dirtied_before)1501 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1502 unsigned long dirtied_before)
1503 {
1504 int moved;
1505 unsigned long time_expire_jif = dirtied_before;
1506
1507 assert_spin_locked(&wb->list_lock);
1508 list_splice_init(&wb->b_more_io, &wb->b_io);
1509 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1510 if (!work->for_sync)
1511 time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1512 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1513 time_expire_jif);
1514 if (moved)
1515 wb_io_lists_populated(wb);
1516 trace_writeback_queue_io(wb, work, dirtied_before, moved);
1517 }
1518
write_inode(struct inode * inode,struct writeback_control * wbc)1519 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1520 {
1521 int ret;
1522
1523 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1524 trace_writeback_write_inode_start(inode, wbc);
1525 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1526 trace_writeback_write_inode(inode, wbc);
1527 return ret;
1528 }
1529 return 0;
1530 }
1531
1532 /*
1533 * Wait for writeback on an inode to complete. Called with i_lock held.
1534 * Caller must make sure inode cannot go away when we drop i_lock.
1535 */
inode_wait_for_writeback(struct inode * inode)1536 void inode_wait_for_writeback(struct inode *inode)
1537 {
1538 struct wait_bit_queue_entry wqe;
1539 struct wait_queue_head *wq_head;
1540
1541 assert_spin_locked(&inode->i_lock);
1542
1543 if (!(inode->i_state & I_SYNC))
1544 return;
1545
1546 wq_head = inode_bit_waitqueue(&wqe, inode, __I_SYNC);
1547 for (;;) {
1548 prepare_to_wait_event(wq_head, &wqe.wq_entry, TASK_UNINTERRUPTIBLE);
1549 /* Checking I_SYNC with inode->i_lock guarantees memory ordering. */
1550 if (!(inode->i_state & I_SYNC))
1551 break;
1552 spin_unlock(&inode->i_lock);
1553 schedule();
1554 spin_lock(&inode->i_lock);
1555 }
1556 finish_wait(wq_head, &wqe.wq_entry);
1557 }
1558
1559 /*
1560 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1561 * held and drops it. It is aimed for callers not holding any inode reference
1562 * so once i_lock is dropped, inode can go away.
1563 */
inode_sleep_on_writeback(struct inode * inode)1564 static void inode_sleep_on_writeback(struct inode *inode)
1565 __releases(inode->i_lock)
1566 {
1567 struct wait_bit_queue_entry wqe;
1568 struct wait_queue_head *wq_head;
1569 bool sleep;
1570
1571 assert_spin_locked(&inode->i_lock);
1572
1573 wq_head = inode_bit_waitqueue(&wqe, inode, __I_SYNC);
1574 prepare_to_wait_event(wq_head, &wqe.wq_entry, TASK_UNINTERRUPTIBLE);
1575 /* Checking I_SYNC with inode->i_lock guarantees memory ordering. */
1576 sleep = !!(inode->i_state & I_SYNC);
1577 spin_unlock(&inode->i_lock);
1578 if (sleep)
1579 schedule();
1580 finish_wait(wq_head, &wqe.wq_entry);
1581 }
1582
1583 /*
1584 * Find proper writeback list for the inode depending on its current state and
1585 * possibly also change of its state while we were doing writeback. Here we
1586 * handle things such as livelock prevention or fairness of writeback among
1587 * inodes. This function can be called only by flusher thread - noone else
1588 * processes all inodes in writeback lists and requeueing inodes behind flusher
1589 * thread's back can have unexpected consequences.
1590 */
requeue_inode(struct inode * inode,struct bdi_writeback * wb,struct writeback_control * wbc,unsigned long dirtied_before)1591 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1592 struct writeback_control *wbc,
1593 unsigned long dirtied_before)
1594 {
1595 if (inode->i_state & I_FREEING)
1596 return;
1597
1598 /*
1599 * Sync livelock prevention. Each inode is tagged and synced in one
1600 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1601 * the dirty time to prevent enqueue and sync it again.
1602 */
1603 if ((inode->i_state & I_DIRTY) &&
1604 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1605 inode->dirtied_when = jiffies;
1606
1607 if (wbc->pages_skipped) {
1608 /*
1609 * Writeback is not making progress due to locked buffers.
1610 * Skip this inode for now. Although having skipped pages
1611 * is odd for clean inodes, it can happen for some
1612 * filesystems so handle that gracefully.
1613 */
1614 if (inode->i_state & I_DIRTY_ALL)
1615 redirty_tail_locked(inode, wb);
1616 else
1617 inode_cgwb_move_to_attached(inode, wb);
1618 return;
1619 }
1620
1621 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1622 /*
1623 * We didn't write back all the pages. nfs_writepages()
1624 * sometimes bales out without doing anything.
1625 */
1626 if (wbc->nr_to_write <= 0 &&
1627 !inode_dirtied_after(inode, dirtied_before)) {
1628 /* Slice used up. Queue for next turn. */
1629 requeue_io(inode, wb);
1630 } else {
1631 /*
1632 * Writeback blocked by something other than
1633 * congestion. Delay the inode for some time to
1634 * avoid spinning on the CPU (100% iowait)
1635 * retrying writeback of the dirty page/inode
1636 * that cannot be performed immediately.
1637 */
1638 redirty_tail_locked(inode, wb);
1639 }
1640 } else if (inode->i_state & I_DIRTY) {
1641 /*
1642 * Filesystems can dirty the inode during writeback operations,
1643 * such as delayed allocation during submission or metadata
1644 * updates after data IO completion.
1645 */
1646 redirty_tail_locked(inode, wb);
1647 } else if (inode->i_state & I_DIRTY_TIME) {
1648 inode->dirtied_when = jiffies;
1649 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1650 inode->i_state &= ~I_SYNC_QUEUED;
1651 } else {
1652 /* The inode is clean. Remove from writeback lists. */
1653 inode_cgwb_move_to_attached(inode, wb);
1654 }
1655 }
1656
1657 /*
1658 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1659 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1660 *
1661 * This doesn't remove the inode from the writeback list it is on, except
1662 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1663 * expiration. The caller is otherwise responsible for writeback list handling.
1664 *
1665 * The caller is also responsible for setting the I_SYNC flag beforehand and
1666 * calling inode_sync_complete() to clear it afterwards.
1667 */
1668 static int
__writeback_single_inode(struct inode * inode,struct writeback_control * wbc)1669 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1670 {
1671 struct address_space *mapping = inode->i_mapping;
1672 long nr_to_write = wbc->nr_to_write;
1673 unsigned dirty;
1674 int ret;
1675
1676 WARN_ON(!(inode->i_state & I_SYNC));
1677
1678 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1679
1680 ret = do_writepages(mapping, wbc);
1681
1682 /*
1683 * Make sure to wait on the data before writing out the metadata.
1684 * This is important for filesystems that modify metadata on data
1685 * I/O completion. We don't do it for sync(2) writeback because it has a
1686 * separate, external IO completion path and ->sync_fs for guaranteeing
1687 * inode metadata is written back correctly.
1688 */
1689 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1690 int err = filemap_fdatawait(mapping);
1691 if (ret == 0)
1692 ret = err;
1693 }
1694
1695 /*
1696 * If the inode has dirty timestamps and we need to write them, call
1697 * mark_inode_dirty_sync() to notify the filesystem about it and to
1698 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1699 */
1700 if ((inode->i_state & I_DIRTY_TIME) &&
1701 (wbc->sync_mode == WB_SYNC_ALL ||
1702 time_after(jiffies, inode->dirtied_time_when +
1703 dirtytime_expire_interval * HZ))) {
1704 trace_writeback_lazytime(inode);
1705 mark_inode_dirty_sync(inode);
1706 }
1707
1708 /*
1709 * Get and clear the dirty flags from i_state. This needs to be done
1710 * after calling writepages because some filesystems may redirty the
1711 * inode during writepages due to delalloc. It also needs to be done
1712 * after handling timestamp expiration, as that may dirty the inode too.
1713 */
1714 spin_lock(&inode->i_lock);
1715 dirty = inode->i_state & I_DIRTY;
1716 inode->i_state &= ~dirty;
1717
1718 /*
1719 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1720 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1721 * either they see the I_DIRTY bits cleared or we see the dirtied
1722 * inode.
1723 *
1724 * I_DIRTY_PAGES is always cleared together above even if @mapping
1725 * still has dirty pages. The flag is reinstated after smp_mb() if
1726 * necessary. This guarantees that either __mark_inode_dirty()
1727 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1728 */
1729 smp_mb();
1730
1731 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1732 inode->i_state |= I_DIRTY_PAGES;
1733 else if (unlikely(inode->i_state & I_PINNING_NETFS_WB)) {
1734 if (!(inode->i_state & I_DIRTY_PAGES)) {
1735 inode->i_state &= ~I_PINNING_NETFS_WB;
1736 wbc->unpinned_netfs_wb = true;
1737 dirty |= I_PINNING_NETFS_WB; /* Cause write_inode */
1738 }
1739 }
1740
1741 spin_unlock(&inode->i_lock);
1742
1743 /* Don't write the inode if only I_DIRTY_PAGES was set */
1744 if (dirty & ~I_DIRTY_PAGES) {
1745 int err = write_inode(inode, wbc);
1746 if (ret == 0)
1747 ret = err;
1748 }
1749 wbc->unpinned_netfs_wb = false;
1750 trace_writeback_single_inode(inode, wbc, nr_to_write);
1751 return ret;
1752 }
1753
1754 /*
1755 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1756 * the regular batched writeback done by the flusher threads in
1757 * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1758 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1759 *
1760 * To prevent the inode from going away, either the caller must have a reference
1761 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1762 */
writeback_single_inode(struct inode * inode,struct writeback_control * wbc)1763 static int writeback_single_inode(struct inode *inode,
1764 struct writeback_control *wbc)
1765 {
1766 struct bdi_writeback *wb;
1767 int ret = 0;
1768
1769 spin_lock(&inode->i_lock);
1770 if (!atomic_read(&inode->i_count))
1771 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1772 else
1773 WARN_ON(inode->i_state & I_WILL_FREE);
1774
1775 if (inode->i_state & I_SYNC) {
1776 /*
1777 * Writeback is already running on the inode. For WB_SYNC_NONE,
1778 * that's enough and we can just return. For WB_SYNC_ALL, we
1779 * must wait for the existing writeback to complete, then do
1780 * writeback again if there's anything left.
1781 */
1782 if (wbc->sync_mode != WB_SYNC_ALL)
1783 goto out;
1784 inode_wait_for_writeback(inode);
1785 }
1786 WARN_ON(inode->i_state & I_SYNC);
1787 /*
1788 * If the inode is already fully clean, then there's nothing to do.
1789 *
1790 * For data-integrity syncs we also need to check whether any pages are
1791 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1792 * there are any such pages, we'll need to wait for them.
1793 */
1794 if (!(inode->i_state & I_DIRTY_ALL) &&
1795 (wbc->sync_mode != WB_SYNC_ALL ||
1796 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1797 goto out;
1798 inode->i_state |= I_SYNC;
1799 wbc_attach_and_unlock_inode(wbc, inode);
1800
1801 ret = __writeback_single_inode(inode, wbc);
1802
1803 wbc_detach_inode(wbc);
1804
1805 wb = inode_to_wb_and_lock_list(inode);
1806 spin_lock(&inode->i_lock);
1807 /*
1808 * If the inode is freeing, its i_io_list shoudn't be updated
1809 * as it can be finally deleted at this moment.
1810 */
1811 if (!(inode->i_state & I_FREEING)) {
1812 /*
1813 * If the inode is now fully clean, then it can be safely
1814 * removed from its writeback list (if any). Otherwise the
1815 * flusher threads are responsible for the writeback lists.
1816 */
1817 if (!(inode->i_state & I_DIRTY_ALL))
1818 inode_cgwb_move_to_attached(inode, wb);
1819 else if (!(inode->i_state & I_SYNC_QUEUED)) {
1820 if ((inode->i_state & I_DIRTY))
1821 redirty_tail_locked(inode, wb);
1822 else if (inode->i_state & I_DIRTY_TIME) {
1823 inode->dirtied_when = jiffies;
1824 inode_io_list_move_locked(inode,
1825 wb,
1826 &wb->b_dirty_time);
1827 }
1828 }
1829 }
1830
1831 spin_unlock(&wb->list_lock);
1832 inode_sync_complete(inode);
1833 out:
1834 spin_unlock(&inode->i_lock);
1835 return ret;
1836 }
1837
writeback_chunk_size(struct bdi_writeback * wb,struct wb_writeback_work * work)1838 static long writeback_chunk_size(struct bdi_writeback *wb,
1839 struct wb_writeback_work *work)
1840 {
1841 long pages;
1842
1843 /*
1844 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1845 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1846 * here avoids calling into writeback_inodes_wb() more than once.
1847 *
1848 * The intended call sequence for WB_SYNC_ALL writeback is:
1849 *
1850 * wb_writeback()
1851 * writeback_sb_inodes() <== called only once
1852 * write_cache_pages() <== called once for each inode
1853 * (quickly) tag currently dirty pages
1854 * (maybe slowly) sync all tagged pages
1855 */
1856 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1857 pages = LONG_MAX;
1858 else {
1859 pages = min(wb->avg_write_bandwidth / 2,
1860 global_wb_domain.dirty_limit / DIRTY_SCOPE);
1861 pages = min(pages, work->nr_pages);
1862 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1863 MIN_WRITEBACK_PAGES);
1864 }
1865
1866 return pages;
1867 }
1868
1869 /*
1870 * Write a portion of b_io inodes which belong to @sb.
1871 *
1872 * Return the number of pages and/or inodes written.
1873 *
1874 * NOTE! This is called with wb->list_lock held, and will
1875 * unlock and relock that for each inode it ends up doing
1876 * IO for.
1877 */
writeback_sb_inodes(struct super_block * sb,struct bdi_writeback * wb,struct wb_writeback_work * work)1878 static long writeback_sb_inodes(struct super_block *sb,
1879 struct bdi_writeback *wb,
1880 struct wb_writeback_work *work)
1881 {
1882 struct writeback_control wbc = {
1883 .sync_mode = work->sync_mode,
1884 .tagged_writepages = work->tagged_writepages,
1885 .for_kupdate = work->for_kupdate,
1886 .for_background = work->for_background,
1887 .for_sync = work->for_sync,
1888 .range_cyclic = work->range_cyclic,
1889 .range_start = 0,
1890 .range_end = LLONG_MAX,
1891 };
1892 unsigned long start_time = jiffies;
1893 long write_chunk;
1894 long total_wrote = 0; /* count both pages and inodes */
1895 unsigned long dirtied_before = jiffies;
1896
1897 if (work->for_kupdate)
1898 dirtied_before = jiffies -
1899 msecs_to_jiffies(dirty_expire_interval * 10);
1900
1901 while (!list_empty(&wb->b_io)) {
1902 struct inode *inode = wb_inode(wb->b_io.prev);
1903 struct bdi_writeback *tmp_wb;
1904 long wrote;
1905
1906 if (inode->i_sb != sb) {
1907 if (work->sb) {
1908 /*
1909 * We only want to write back data for this
1910 * superblock, move all inodes not belonging
1911 * to it back onto the dirty list.
1912 */
1913 redirty_tail(inode, wb);
1914 continue;
1915 }
1916
1917 /*
1918 * The inode belongs to a different superblock.
1919 * Bounce back to the caller to unpin this and
1920 * pin the next superblock.
1921 */
1922 break;
1923 }
1924
1925 /*
1926 * Don't bother with new inodes or inodes being freed, first
1927 * kind does not need periodic writeout yet, and for the latter
1928 * kind writeout is handled by the freer.
1929 */
1930 spin_lock(&inode->i_lock);
1931 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1932 redirty_tail_locked(inode, wb);
1933 spin_unlock(&inode->i_lock);
1934 continue;
1935 }
1936 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1937 /*
1938 * If this inode is locked for writeback and we are not
1939 * doing writeback-for-data-integrity, move it to
1940 * b_more_io so that writeback can proceed with the
1941 * other inodes on s_io.
1942 *
1943 * We'll have another go at writing back this inode
1944 * when we completed a full scan of b_io.
1945 */
1946 requeue_io(inode, wb);
1947 spin_unlock(&inode->i_lock);
1948 trace_writeback_sb_inodes_requeue(inode);
1949 continue;
1950 }
1951 spin_unlock(&wb->list_lock);
1952
1953 /*
1954 * We already requeued the inode if it had I_SYNC set and we
1955 * are doing WB_SYNC_NONE writeback. So this catches only the
1956 * WB_SYNC_ALL case.
1957 */
1958 if (inode->i_state & I_SYNC) {
1959 /* Wait for I_SYNC. This function drops i_lock... */
1960 inode_sleep_on_writeback(inode);
1961 /* Inode may be gone, start again */
1962 spin_lock(&wb->list_lock);
1963 continue;
1964 }
1965 inode->i_state |= I_SYNC;
1966 wbc_attach_and_unlock_inode(&wbc, inode);
1967
1968 write_chunk = writeback_chunk_size(wb, work);
1969 wbc.nr_to_write = write_chunk;
1970 wbc.pages_skipped = 0;
1971
1972 /*
1973 * We use I_SYNC to pin the inode in memory. While it is set
1974 * evict_inode() will wait so the inode cannot be freed.
1975 */
1976 __writeback_single_inode(inode, &wbc);
1977
1978 wbc_detach_inode(&wbc);
1979 work->nr_pages -= write_chunk - wbc.nr_to_write;
1980 wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
1981 wrote = wrote < 0 ? 0 : wrote;
1982 total_wrote += wrote;
1983
1984 if (need_resched()) {
1985 /*
1986 * We're trying to balance between building up a nice
1987 * long list of IOs to improve our merge rate, and
1988 * getting those IOs out quickly for anyone throttling
1989 * in balance_dirty_pages(). cond_resched() doesn't
1990 * unplug, so get our IOs out the door before we
1991 * give up the CPU.
1992 */
1993 blk_flush_plug(current->plug, false);
1994 cond_resched();
1995 }
1996
1997 /*
1998 * Requeue @inode if still dirty. Be careful as @inode may
1999 * have been switched to another wb in the meantime.
2000 */
2001 tmp_wb = inode_to_wb_and_lock_list(inode);
2002 spin_lock(&inode->i_lock);
2003 if (!(inode->i_state & I_DIRTY_ALL))
2004 total_wrote++;
2005 requeue_inode(inode, tmp_wb, &wbc, dirtied_before);
2006 inode_sync_complete(inode);
2007 spin_unlock(&inode->i_lock);
2008
2009 if (unlikely(tmp_wb != wb)) {
2010 spin_unlock(&tmp_wb->list_lock);
2011 spin_lock(&wb->list_lock);
2012 }
2013
2014 /*
2015 * bail out to wb_writeback() often enough to check
2016 * background threshold and other termination conditions.
2017 */
2018 if (total_wrote) {
2019 if (time_is_before_jiffies(start_time + HZ / 10UL))
2020 break;
2021 if (work->nr_pages <= 0)
2022 break;
2023 }
2024 }
2025 return total_wrote;
2026 }
2027
__writeback_inodes_wb(struct bdi_writeback * wb,struct wb_writeback_work * work)2028 static long __writeback_inodes_wb(struct bdi_writeback *wb,
2029 struct wb_writeback_work *work)
2030 {
2031 unsigned long start_time = jiffies;
2032 long wrote = 0;
2033
2034 while (!list_empty(&wb->b_io)) {
2035 struct inode *inode = wb_inode(wb->b_io.prev);
2036 struct super_block *sb = inode->i_sb;
2037
2038 if (!super_trylock_shared(sb)) {
2039 /*
2040 * super_trylock_shared() may fail consistently due to
2041 * s_umount being grabbed by someone else. Don't use
2042 * requeue_io() to avoid busy retrying the inode/sb.
2043 */
2044 redirty_tail(inode, wb);
2045 continue;
2046 }
2047 wrote += writeback_sb_inodes(sb, wb, work);
2048 up_read(&sb->s_umount);
2049
2050 /* refer to the same tests at the end of writeback_sb_inodes */
2051 if (wrote) {
2052 if (time_is_before_jiffies(start_time + HZ / 10UL))
2053 break;
2054 if (work->nr_pages <= 0)
2055 break;
2056 }
2057 }
2058 /* Leave any unwritten inodes on b_io */
2059 return wrote;
2060 }
2061
writeback_inodes_wb(struct bdi_writeback * wb,long nr_pages,enum wb_reason reason)2062 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
2063 enum wb_reason reason)
2064 {
2065 struct wb_writeback_work work = {
2066 .nr_pages = nr_pages,
2067 .sync_mode = WB_SYNC_NONE,
2068 .range_cyclic = 1,
2069 .reason = reason,
2070 };
2071 struct blk_plug plug;
2072
2073 blk_start_plug(&plug);
2074 spin_lock(&wb->list_lock);
2075 if (list_empty(&wb->b_io))
2076 queue_io(wb, &work, jiffies);
2077 __writeback_inodes_wb(wb, &work);
2078 spin_unlock(&wb->list_lock);
2079 blk_finish_plug(&plug);
2080
2081 return nr_pages - work.nr_pages;
2082 }
2083
2084 /*
2085 * Explicit flushing or periodic writeback of "old" data.
2086 *
2087 * Define "old": the first time one of an inode's pages is dirtied, we mark the
2088 * dirtying-time in the inode's address_space. So this periodic writeback code
2089 * just walks the superblock inode list, writing back any inodes which are
2090 * older than a specific point in time.
2091 *
2092 * Try to run once per dirty_writeback_interval. But if a writeback event
2093 * takes longer than a dirty_writeback_interval interval, then leave a
2094 * one-second gap.
2095 *
2096 * dirtied_before takes precedence over nr_to_write. So we'll only write back
2097 * all dirty pages if they are all attached to "old" mappings.
2098 */
wb_writeback(struct bdi_writeback * wb,struct wb_writeback_work * work)2099 static long wb_writeback(struct bdi_writeback *wb,
2100 struct wb_writeback_work *work)
2101 {
2102 long nr_pages = work->nr_pages;
2103 unsigned long dirtied_before = jiffies;
2104 struct inode *inode;
2105 long progress;
2106 struct blk_plug plug;
2107 bool queued = false;
2108
2109 blk_start_plug(&plug);
2110 for (;;) {
2111 /*
2112 * Stop writeback when nr_pages has been consumed
2113 */
2114 if (work->nr_pages <= 0)
2115 break;
2116
2117 /*
2118 * Background writeout and kupdate-style writeback may
2119 * run forever. Stop them if there is other work to do
2120 * so that e.g. sync can proceed. They'll be restarted
2121 * after the other works are all done.
2122 */
2123 if ((work->for_background || work->for_kupdate) &&
2124 !list_empty(&wb->work_list))
2125 break;
2126
2127 /*
2128 * For background writeout, stop when we are below the
2129 * background dirty threshold
2130 */
2131 if (work->for_background && !wb_over_bg_thresh(wb))
2132 break;
2133
2134
2135 spin_lock(&wb->list_lock);
2136
2137 trace_writeback_start(wb, work);
2138 if (list_empty(&wb->b_io)) {
2139 /*
2140 * Kupdate and background works are special and we want
2141 * to include all inodes that need writing. Livelock
2142 * avoidance is handled by these works yielding to any
2143 * other work so we are safe.
2144 */
2145 if (work->for_kupdate) {
2146 dirtied_before = jiffies -
2147 msecs_to_jiffies(dirty_expire_interval *
2148 10);
2149 } else if (work->for_background)
2150 dirtied_before = jiffies;
2151
2152 queue_io(wb, work, dirtied_before);
2153 queued = true;
2154 }
2155 if (work->sb)
2156 progress = writeback_sb_inodes(work->sb, wb, work);
2157 else
2158 progress = __writeback_inodes_wb(wb, work);
2159 trace_writeback_written(wb, work);
2160
2161 /*
2162 * Did we write something? Try for more
2163 *
2164 * Dirty inodes are moved to b_io for writeback in batches.
2165 * The completion of the current batch does not necessarily
2166 * mean the overall work is done. So we keep looping as long
2167 * as made some progress on cleaning pages or inodes.
2168 */
2169 if (progress || !queued) {
2170 spin_unlock(&wb->list_lock);
2171 continue;
2172 }
2173
2174 /*
2175 * No more inodes for IO, bail
2176 */
2177 if (list_empty(&wb->b_more_io)) {
2178 spin_unlock(&wb->list_lock);
2179 break;
2180 }
2181
2182 /*
2183 * Nothing written. Wait for some inode to
2184 * become available for writeback. Otherwise
2185 * we'll just busyloop.
2186 */
2187 trace_writeback_wait(wb, work);
2188 inode = wb_inode(wb->b_more_io.prev);
2189 spin_lock(&inode->i_lock);
2190 spin_unlock(&wb->list_lock);
2191 /* This function drops i_lock... */
2192 inode_sleep_on_writeback(inode);
2193 }
2194 blk_finish_plug(&plug);
2195
2196 return nr_pages - work->nr_pages;
2197 }
2198
2199 /*
2200 * Return the next wb_writeback_work struct that hasn't been processed yet.
2201 */
get_next_work_item(struct bdi_writeback * wb)2202 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2203 {
2204 struct wb_writeback_work *work = NULL;
2205
2206 spin_lock_irq(&wb->work_lock);
2207 if (!list_empty(&wb->work_list)) {
2208 work = list_entry(wb->work_list.next,
2209 struct wb_writeback_work, list);
2210 list_del_init(&work->list);
2211 }
2212 spin_unlock_irq(&wb->work_lock);
2213 return work;
2214 }
2215
wb_check_background_flush(struct bdi_writeback * wb)2216 static long wb_check_background_flush(struct bdi_writeback *wb)
2217 {
2218 if (wb_over_bg_thresh(wb)) {
2219
2220 struct wb_writeback_work work = {
2221 .nr_pages = LONG_MAX,
2222 .sync_mode = WB_SYNC_NONE,
2223 .for_background = 1,
2224 .range_cyclic = 1,
2225 .reason = WB_REASON_BACKGROUND,
2226 };
2227
2228 return wb_writeback(wb, &work);
2229 }
2230
2231 return 0;
2232 }
2233
wb_check_old_data_flush(struct bdi_writeback * wb)2234 static long wb_check_old_data_flush(struct bdi_writeback *wb)
2235 {
2236 unsigned long expired;
2237 long nr_pages;
2238
2239 /*
2240 * When set to zero, disable periodic writeback
2241 */
2242 if (!dirty_writeback_interval)
2243 return 0;
2244
2245 expired = wb->last_old_flush +
2246 msecs_to_jiffies(dirty_writeback_interval * 10);
2247 if (time_before(jiffies, expired))
2248 return 0;
2249
2250 wb->last_old_flush = jiffies;
2251 nr_pages = get_nr_dirty_pages();
2252
2253 if (nr_pages) {
2254 struct wb_writeback_work work = {
2255 .nr_pages = nr_pages,
2256 .sync_mode = WB_SYNC_NONE,
2257 .for_kupdate = 1,
2258 .range_cyclic = 1,
2259 .reason = WB_REASON_PERIODIC,
2260 };
2261
2262 return wb_writeback(wb, &work);
2263 }
2264
2265 return 0;
2266 }
2267
wb_check_start_all(struct bdi_writeback * wb)2268 static long wb_check_start_all(struct bdi_writeback *wb)
2269 {
2270 long nr_pages;
2271
2272 if (!test_bit(WB_start_all, &wb->state))
2273 return 0;
2274
2275 nr_pages = get_nr_dirty_pages();
2276 if (nr_pages) {
2277 struct wb_writeback_work work = {
2278 .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2279 .sync_mode = WB_SYNC_NONE,
2280 .range_cyclic = 1,
2281 .reason = wb->start_all_reason,
2282 };
2283
2284 nr_pages = wb_writeback(wb, &work);
2285 }
2286
2287 clear_bit(WB_start_all, &wb->state);
2288 return nr_pages;
2289 }
2290
2291
2292 /*
2293 * Retrieve work items and do the writeback they describe
2294 */
wb_do_writeback(struct bdi_writeback * wb)2295 static long wb_do_writeback(struct bdi_writeback *wb)
2296 {
2297 struct wb_writeback_work *work;
2298 long wrote = 0;
2299
2300 set_bit(WB_writeback_running, &wb->state);
2301 while ((work = get_next_work_item(wb)) != NULL) {
2302 trace_writeback_exec(wb, work);
2303 wrote += wb_writeback(wb, work);
2304 finish_writeback_work(work);
2305 }
2306
2307 /*
2308 * Check for a flush-everything request
2309 */
2310 wrote += wb_check_start_all(wb);
2311
2312 /*
2313 * Check for periodic writeback, kupdated() style
2314 */
2315 wrote += wb_check_old_data_flush(wb);
2316 wrote += wb_check_background_flush(wb);
2317 clear_bit(WB_writeback_running, &wb->state);
2318
2319 return wrote;
2320 }
2321
2322 /*
2323 * Handle writeback of dirty data for the device backed by this bdi. Also
2324 * reschedules periodically and does kupdated style flushing.
2325 */
wb_workfn(struct work_struct * work)2326 void wb_workfn(struct work_struct *work)
2327 {
2328 struct bdi_writeback *wb = container_of(to_delayed_work(work),
2329 struct bdi_writeback, dwork);
2330 long pages_written;
2331
2332 set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2333
2334 if (likely(!current_is_workqueue_rescuer() ||
2335 !test_bit(WB_registered, &wb->state))) {
2336 /*
2337 * The normal path. Keep writing back @wb until its
2338 * work_list is empty. Note that this path is also taken
2339 * if @wb is shutting down even when we're running off the
2340 * rescuer as work_list needs to be drained.
2341 */
2342 do {
2343 pages_written = wb_do_writeback(wb);
2344 trace_writeback_pages_written(pages_written);
2345 } while (!list_empty(&wb->work_list));
2346 } else {
2347 /*
2348 * bdi_wq can't get enough workers and we're running off
2349 * the emergency worker. Don't hog it. Hopefully, 1024 is
2350 * enough for efficient IO.
2351 */
2352 pages_written = writeback_inodes_wb(wb, 1024,
2353 WB_REASON_FORKER_THREAD);
2354 trace_writeback_pages_written(pages_written);
2355 }
2356
2357 if (!list_empty(&wb->work_list))
2358 wb_wakeup(wb);
2359 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2360 wb_wakeup_delayed(wb);
2361 }
2362
2363 /*
2364 * Start writeback of all dirty pages on this bdi.
2365 */
__wakeup_flusher_threads_bdi(struct backing_dev_info * bdi,enum wb_reason reason)2366 static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2367 enum wb_reason reason)
2368 {
2369 struct bdi_writeback *wb;
2370
2371 if (!bdi_has_dirty_io(bdi))
2372 return;
2373
2374 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2375 wb_start_writeback(wb, reason);
2376 }
2377
wakeup_flusher_threads_bdi(struct backing_dev_info * bdi,enum wb_reason reason)2378 void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2379 enum wb_reason reason)
2380 {
2381 rcu_read_lock();
2382 __wakeup_flusher_threads_bdi(bdi, reason);
2383 rcu_read_unlock();
2384 }
2385
2386 /*
2387 * Wakeup the flusher threads to start writeback of all currently dirty pages
2388 */
wakeup_flusher_threads(enum wb_reason reason)2389 void wakeup_flusher_threads(enum wb_reason reason)
2390 {
2391 struct backing_dev_info *bdi;
2392
2393 /*
2394 * If we are expecting writeback progress we must submit plugged IO.
2395 */
2396 blk_flush_plug(current->plug, true);
2397
2398 rcu_read_lock();
2399 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2400 __wakeup_flusher_threads_bdi(bdi, reason);
2401 rcu_read_unlock();
2402 }
2403
2404 /*
2405 * Wake up bdi's periodically to make sure dirtytime inodes gets
2406 * written back periodically. We deliberately do *not* check the
2407 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2408 * kernel to be constantly waking up once there are any dirtytime
2409 * inodes on the system. So instead we define a separate delayed work
2410 * function which gets called much more rarely. (By default, only
2411 * once every 12 hours.)
2412 *
2413 * If there is any other write activity going on in the file system,
2414 * this function won't be necessary. But if the only thing that has
2415 * happened on the file system is a dirtytime inode caused by an atime
2416 * update, we need this infrastructure below to make sure that inode
2417 * eventually gets pushed out to disk.
2418 */
2419 static void wakeup_dirtytime_writeback(struct work_struct *w);
2420 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2421
wakeup_dirtytime_writeback(struct work_struct * w)2422 static void wakeup_dirtytime_writeback(struct work_struct *w)
2423 {
2424 struct backing_dev_info *bdi;
2425
2426 rcu_read_lock();
2427 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2428 struct bdi_writeback *wb;
2429
2430 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2431 if (!list_empty(&wb->b_dirty_time))
2432 wb_wakeup(wb);
2433 }
2434 rcu_read_unlock();
2435 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2436 }
2437
start_dirtytime_writeback(void)2438 static int __init start_dirtytime_writeback(void)
2439 {
2440 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2441 return 0;
2442 }
2443 __initcall(start_dirtytime_writeback);
2444
dirtytime_interval_handler(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)2445 int dirtytime_interval_handler(const struct ctl_table *table, int write,
2446 void *buffer, size_t *lenp, loff_t *ppos)
2447 {
2448 int ret;
2449
2450 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2451 if (ret == 0 && write)
2452 mod_delayed_work(system_wq, &dirtytime_work, 0);
2453 return ret;
2454 }
2455
2456 /**
2457 * __mark_inode_dirty - internal function to mark an inode dirty
2458 *
2459 * @inode: inode to mark
2460 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2461 * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2462 * with I_DIRTY_PAGES.
2463 *
2464 * Mark an inode as dirty. We notify the filesystem, then update the inode's
2465 * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2466 *
2467 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2468 * instead of calling this directly.
2469 *
2470 * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2471 * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2472 * even if they are later hashed, as they will have been marked dirty already.
2473 *
2474 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2475 *
2476 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2477 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2478 * the kernel-internal blockdev inode represents the dirtying time of the
2479 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2480 * page->mapping->host, so the page-dirtying time is recorded in the internal
2481 * blockdev inode.
2482 */
__mark_inode_dirty(struct inode * inode,int flags)2483 void __mark_inode_dirty(struct inode *inode, int flags)
2484 {
2485 struct super_block *sb = inode->i_sb;
2486 int dirtytime = 0;
2487 struct bdi_writeback *wb = NULL;
2488
2489 trace_writeback_mark_inode_dirty(inode, flags);
2490
2491 if (flags & I_DIRTY_INODE) {
2492 /*
2493 * Inode timestamp update will piggback on this dirtying.
2494 * We tell ->dirty_inode callback that timestamps need to
2495 * be updated by setting I_DIRTY_TIME in flags.
2496 */
2497 if (inode->i_state & I_DIRTY_TIME) {
2498 spin_lock(&inode->i_lock);
2499 if (inode->i_state & I_DIRTY_TIME) {
2500 inode->i_state &= ~I_DIRTY_TIME;
2501 flags |= I_DIRTY_TIME;
2502 }
2503 spin_unlock(&inode->i_lock);
2504 }
2505
2506 /*
2507 * Notify the filesystem about the inode being dirtied, so that
2508 * (if needed) it can update on-disk fields and journal the
2509 * inode. This is only needed when the inode itself is being
2510 * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2511 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2512 */
2513 trace_writeback_dirty_inode_start(inode, flags);
2514 if (sb->s_op->dirty_inode)
2515 sb->s_op->dirty_inode(inode,
2516 flags & (I_DIRTY_INODE | I_DIRTY_TIME));
2517 trace_writeback_dirty_inode(inode, flags);
2518
2519 /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2520 flags &= ~I_DIRTY_TIME;
2521 } else {
2522 /*
2523 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2524 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2525 * in one call to __mark_inode_dirty().)
2526 */
2527 dirtytime = flags & I_DIRTY_TIME;
2528 WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2529 }
2530
2531 /*
2532 * Paired with smp_mb() in __writeback_single_inode() for the
2533 * following lockless i_state test. See there for details.
2534 */
2535 smp_mb();
2536
2537 if ((inode->i_state & flags) == flags)
2538 return;
2539
2540 spin_lock(&inode->i_lock);
2541 if ((inode->i_state & flags) != flags) {
2542 const int was_dirty = inode->i_state & I_DIRTY;
2543
2544 inode_attach_wb(inode, NULL);
2545
2546 inode->i_state |= flags;
2547
2548 /*
2549 * Grab inode's wb early because it requires dropping i_lock and we
2550 * need to make sure following checks happen atomically with dirty
2551 * list handling so that we don't move inodes under flush worker's
2552 * hands.
2553 */
2554 if (!was_dirty) {
2555 wb = locked_inode_to_wb_and_lock_list(inode);
2556 spin_lock(&inode->i_lock);
2557 }
2558
2559 /*
2560 * If the inode is queued for writeback by flush worker, just
2561 * update its dirty state. Once the flush worker is done with
2562 * the inode it will place it on the appropriate superblock
2563 * list, based upon its state.
2564 */
2565 if (inode->i_state & I_SYNC_QUEUED)
2566 goto out_unlock;
2567
2568 /*
2569 * Only add valid (hashed) inodes to the superblock's
2570 * dirty list. Add blockdev inodes as well.
2571 */
2572 if (!S_ISBLK(inode->i_mode)) {
2573 if (inode_unhashed(inode))
2574 goto out_unlock;
2575 }
2576 if (inode->i_state & I_FREEING)
2577 goto out_unlock;
2578
2579 /*
2580 * If the inode was already on b_dirty/b_io/b_more_io, don't
2581 * reposition it (that would break b_dirty time-ordering).
2582 */
2583 if (!was_dirty) {
2584 struct list_head *dirty_list;
2585 bool wakeup_bdi = false;
2586
2587 inode->dirtied_when = jiffies;
2588 if (dirtytime)
2589 inode->dirtied_time_when = jiffies;
2590
2591 if (inode->i_state & I_DIRTY)
2592 dirty_list = &wb->b_dirty;
2593 else
2594 dirty_list = &wb->b_dirty_time;
2595
2596 wakeup_bdi = inode_io_list_move_locked(inode, wb,
2597 dirty_list);
2598
2599 spin_unlock(&wb->list_lock);
2600 spin_unlock(&inode->i_lock);
2601 trace_writeback_dirty_inode_enqueue(inode);
2602
2603 /*
2604 * If this is the first dirty inode for this bdi,
2605 * we have to wake-up the corresponding bdi thread
2606 * to make sure background write-back happens
2607 * later.
2608 */
2609 if (wakeup_bdi &&
2610 (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2611 wb_wakeup_delayed(wb);
2612 return;
2613 }
2614 }
2615 out_unlock:
2616 if (wb)
2617 spin_unlock(&wb->list_lock);
2618 spin_unlock(&inode->i_lock);
2619 }
2620 EXPORT_SYMBOL(__mark_inode_dirty);
2621
2622 /*
2623 * The @s_sync_lock is used to serialise concurrent sync operations
2624 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2625 * Concurrent callers will block on the s_sync_lock rather than doing contending
2626 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2627 * has been issued up to the time this function is enter is guaranteed to be
2628 * completed by the time we have gained the lock and waited for all IO that is
2629 * in progress regardless of the order callers are granted the lock.
2630 */
wait_sb_inodes(struct super_block * sb)2631 static void wait_sb_inodes(struct super_block *sb)
2632 {
2633 LIST_HEAD(sync_list);
2634
2635 /*
2636 * We need to be protected against the filesystem going from
2637 * r/o to r/w or vice versa.
2638 */
2639 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2640
2641 mutex_lock(&sb->s_sync_lock);
2642
2643 /*
2644 * Splice the writeback list onto a temporary list to avoid waiting on
2645 * inodes that have started writeback after this point.
2646 *
2647 * Use rcu_read_lock() to keep the inodes around until we have a
2648 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2649 * the local list because inodes can be dropped from either by writeback
2650 * completion.
2651 */
2652 rcu_read_lock();
2653 spin_lock_irq(&sb->s_inode_wblist_lock);
2654 list_splice_init(&sb->s_inodes_wb, &sync_list);
2655
2656 /*
2657 * Data integrity sync. Must wait for all pages under writeback, because
2658 * there may have been pages dirtied before our sync call, but which had
2659 * writeout started before we write it out. In which case, the inode
2660 * may not be on the dirty list, but we still have to wait for that
2661 * writeout.
2662 */
2663 while (!list_empty(&sync_list)) {
2664 struct inode *inode = list_first_entry(&sync_list, struct inode,
2665 i_wb_list);
2666 struct address_space *mapping = inode->i_mapping;
2667
2668 /*
2669 * Move each inode back to the wb list before we drop the lock
2670 * to preserve consistency between i_wb_list and the mapping
2671 * writeback tag. Writeback completion is responsible to remove
2672 * the inode from either list once the writeback tag is cleared.
2673 */
2674 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2675
2676 /*
2677 * The mapping can appear untagged while still on-list since we
2678 * do not have the mapping lock. Skip it here, wb completion
2679 * will remove it.
2680 */
2681 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2682 continue;
2683
2684 spin_unlock_irq(&sb->s_inode_wblist_lock);
2685
2686 spin_lock(&inode->i_lock);
2687 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2688 spin_unlock(&inode->i_lock);
2689
2690 spin_lock_irq(&sb->s_inode_wblist_lock);
2691 continue;
2692 }
2693 __iget(inode);
2694 spin_unlock(&inode->i_lock);
2695 rcu_read_unlock();
2696
2697 /*
2698 * We keep the error status of individual mapping so that
2699 * applications can catch the writeback error using fsync(2).
2700 * See filemap_fdatawait_keep_errors() for details.
2701 */
2702 filemap_fdatawait_keep_errors(mapping);
2703
2704 cond_resched();
2705
2706 iput(inode);
2707
2708 rcu_read_lock();
2709 spin_lock_irq(&sb->s_inode_wblist_lock);
2710 }
2711 spin_unlock_irq(&sb->s_inode_wblist_lock);
2712 rcu_read_unlock();
2713 mutex_unlock(&sb->s_sync_lock);
2714 }
2715
__writeback_inodes_sb_nr(struct super_block * sb,unsigned long nr,enum wb_reason reason,bool skip_if_busy)2716 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2717 enum wb_reason reason, bool skip_if_busy)
2718 {
2719 struct backing_dev_info *bdi = sb->s_bdi;
2720 DEFINE_WB_COMPLETION(done, bdi);
2721 struct wb_writeback_work work = {
2722 .sb = sb,
2723 .sync_mode = WB_SYNC_NONE,
2724 .tagged_writepages = 1,
2725 .done = &done,
2726 .nr_pages = nr,
2727 .reason = reason,
2728 };
2729
2730 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2731 return;
2732 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2733
2734 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2735 wb_wait_for_completion(&done);
2736 }
2737
2738 /**
2739 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2740 * @sb: the superblock
2741 * @nr: the number of pages to write
2742 * @reason: reason why some writeback work initiated
2743 *
2744 * Start writeback on some inodes on this super_block. No guarantees are made
2745 * on how many (if any) will be written, and this function does not wait
2746 * for IO completion of submitted IO.
2747 */
writeback_inodes_sb_nr(struct super_block * sb,unsigned long nr,enum wb_reason reason)2748 void writeback_inodes_sb_nr(struct super_block *sb,
2749 unsigned long nr,
2750 enum wb_reason reason)
2751 {
2752 __writeback_inodes_sb_nr(sb, nr, reason, false);
2753 }
2754 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2755
2756 /**
2757 * writeback_inodes_sb - writeback dirty inodes from given super_block
2758 * @sb: the superblock
2759 * @reason: reason why some writeback work was initiated
2760 *
2761 * Start writeback on some inodes on this super_block. No guarantees are made
2762 * on how many (if any) will be written, and this function does not wait
2763 * for IO completion of submitted IO.
2764 */
writeback_inodes_sb(struct super_block * sb,enum wb_reason reason)2765 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2766 {
2767 writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2768 }
2769 EXPORT_SYMBOL(writeback_inodes_sb);
2770
2771 /**
2772 * try_to_writeback_inodes_sb - try to start writeback if none underway
2773 * @sb: the superblock
2774 * @reason: reason why some writeback work was initiated
2775 *
2776 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2777 */
try_to_writeback_inodes_sb(struct super_block * sb,enum wb_reason reason)2778 void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2779 {
2780 if (!down_read_trylock(&sb->s_umount))
2781 return;
2782
2783 __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2784 up_read(&sb->s_umount);
2785 }
2786 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2787
2788 /**
2789 * sync_inodes_sb - sync sb inode pages
2790 * @sb: the superblock
2791 *
2792 * This function writes and waits on any dirty inode belonging to this
2793 * super_block.
2794 */
sync_inodes_sb(struct super_block * sb)2795 void sync_inodes_sb(struct super_block *sb)
2796 {
2797 struct backing_dev_info *bdi = sb->s_bdi;
2798 DEFINE_WB_COMPLETION(done, bdi);
2799 struct wb_writeback_work work = {
2800 .sb = sb,
2801 .sync_mode = WB_SYNC_ALL,
2802 .nr_pages = LONG_MAX,
2803 .range_cyclic = 0,
2804 .done = &done,
2805 .reason = WB_REASON_SYNC,
2806 .for_sync = 1,
2807 };
2808
2809 /*
2810 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2811 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2812 * bdi_has_dirty() need to be written out too.
2813 */
2814 if (bdi == &noop_backing_dev_info)
2815 return;
2816 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2817
2818 /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2819 bdi_down_write_wb_switch_rwsem(bdi);
2820 bdi_split_work_to_wbs(bdi, &work, false);
2821 wb_wait_for_completion(&done);
2822 bdi_up_write_wb_switch_rwsem(bdi);
2823
2824 wait_sb_inodes(sb);
2825 }
2826 EXPORT_SYMBOL(sync_inodes_sb);
2827
2828 /**
2829 * write_inode_now - write an inode to disk
2830 * @inode: inode to write to disk
2831 * @sync: whether the write should be synchronous or not
2832 *
2833 * This function commits an inode to disk immediately if it is dirty. This is
2834 * primarily needed by knfsd.
2835 *
2836 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2837 */
write_inode_now(struct inode * inode,int sync)2838 int write_inode_now(struct inode *inode, int sync)
2839 {
2840 struct writeback_control wbc = {
2841 .nr_to_write = LONG_MAX,
2842 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2843 .range_start = 0,
2844 .range_end = LLONG_MAX,
2845 };
2846
2847 if (!mapping_can_writeback(inode->i_mapping))
2848 wbc.nr_to_write = 0;
2849
2850 might_sleep();
2851 return writeback_single_inode(inode, &wbc);
2852 }
2853 EXPORT_SYMBOL(write_inode_now);
2854
2855 /**
2856 * sync_inode_metadata - write an inode to disk
2857 * @inode: the inode to sync
2858 * @wait: wait for I/O to complete.
2859 *
2860 * Write an inode to disk and adjust its dirty state after completion.
2861 *
2862 * Note: only writes the actual inode, no associated data or other metadata.
2863 */
sync_inode_metadata(struct inode * inode,int wait)2864 int sync_inode_metadata(struct inode *inode, int wait)
2865 {
2866 struct writeback_control wbc = {
2867 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2868 .nr_to_write = 0, /* metadata-only */
2869 };
2870
2871 return writeback_single_inode(inode, &wbc);
2872 }
2873 EXPORT_SYMBOL(sync_inode_metadata);
2874