Line data Source code
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 :
70 : static inline struct inode *wb_inode(struct list_head *head)
71 : {
72 23143324 : 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 :
85 25015791 : static bool wb_io_lists_populated(struct bdi_writeback *wb)
86 : {
87 50031582 : if (wb_has_dirty_io(wb)) {
88 : return false;
89 : } else {
90 1205950 : set_bit(WB_has_dirty_io, &wb->state);
91 1205949 : WARN_ON_ONCE(!wb->avg_write_bandwidth);
92 1205949 : atomic_long_add(wb->avg_write_bandwidth,
93 1205949 : &wb->bdi->tot_write_bandwidth);
94 1205949 : return true;
95 : }
96 : }
97 :
98 28211947 : static void wb_io_lists_depopulated(struct bdi_writeback *wb)
99 : {
100 56423894 : if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
101 12623213 : list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
102 1205946 : clear_bit(WB_has_dirty_io, &wb->state);
103 2411892 : WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104 : &wb->bdi->tot_write_bandwidth) < 0);
105 : }
106 28211948 : }
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 : */
118 23305900 : static bool inode_io_list_move_locked(struct inode *inode,
119 : struct bdi_writeback *wb,
120 : struct list_head *head)
121 : {
122 23305900 : assert_spin_locked(&wb->list_lock);
123 23305900 : assert_spin_locked(&inode->i_lock);
124 23305900 : WARN_ON_ONCE(inode->i_state & I_FREEING);
125 :
126 23305900 : list_move(&inode->i_io_list, head);
127 :
128 : /* dirty_time doesn't count as dirty_io until expiration */
129 23305898 : if (head != &wb->b_dirty_time)
130 23305886 : return wb_io_lists_populated(wb);
131 :
132 12 : wb_io_lists_depopulated(wb);
133 12 : return false;
134 : }
135 :
136 1023701 : static void wb_wakeup(struct bdi_writeback *wb)
137 : {
138 1023701 : spin_lock_irq(&wb->work_lock);
139 2047402 : if (test_bit(WB_registered, &wb->state))
140 1023701 : mod_delayed_work(bdi_wq, &wb->dwork, 0);
141 1023701 : spin_unlock_irq(&wb->work_lock);
142 1023700 : }
143 :
144 1094558 : static void finish_writeback_work(struct bdi_writeback *wb,
145 : struct wb_writeback_work *work)
146 : {
147 1094558 : struct wb_completion *done = work->done;
148 :
149 1094558 : if (work->auto_free)
150 1094559 : kfree(work);
151 1094564 : if (done) {
152 1094564 : wait_queue_head_t *waitq = done->waitq;
153 :
154 : /* @done can't be accessed after the following dec */
155 2189134 : if (atomic_dec_and_test(&done->cnt))
156 799191 : wake_up_all(waitq);
157 : }
158 1094570 : }
159 :
160 1094576 : static void wb_queue_work(struct bdi_writeback *wb,
161 : struct wb_writeback_work *work)
162 : {
163 1094576 : trace_writeback_queue(wb, work);
164 :
165 1094575 : if (work->done)
166 1094575 : atomic_inc(&work->done->cnt);
167 :
168 1094576 : spin_lock_irq(&wb->work_lock);
169 :
170 2189152 : if (test_bit(WB_registered, &wb->state)) {
171 1094576 : list_add_tail(&work->list, &wb->work_list);
172 1094576 : mod_delayed_work(bdi_wq, &wb->dwork, 0);
173 : } else
174 0 : finish_writeback_work(wb, work);
175 :
176 1094576 : spin_unlock_irq(&wb->work_lock);
177 1094576 : }
178 :
179 : /**
180 : * wb_wait_for_completion - wait for completion of bdi_writeback_works
181 : * @done: target wb_completion
182 : *
183 : * Wait for one or more work items issued to @bdi with their ->done field
184 : * set to @done, which should have been initialized with
185 : * DEFINE_WB_COMPLETION(). This function returns after all such work items
186 : * are completed. Work items which are waited upon aren't freed
187 : * automatically on completion.
188 : */
189 6587682 : void wb_wait_for_completion(struct wb_completion *done)
190 : {
191 6587682 : atomic_dec(&done->cnt); /* put down the initial count */
192 7386705 : wait_event(*done->waitq, !atomic_read(&done->cnt));
193 6587687 : }
194 :
195 : #ifdef CONFIG_CGROUP_WRITEBACK
196 :
197 : /*
198 : * Parameters for foreign inode detection, see wbc_detach_inode() to see
199 : * how they're used.
200 : *
201 : * These paramters are inherently heuristical as the detection target
202 : * itself is fuzzy. All we want to do is detaching an inode from the
203 : * current owner if it's being written to by some other cgroups too much.
204 : *
205 : * The current cgroup writeback is built on the assumption that multiple
206 : * cgroups writing to the same inode concurrently is very rare and a mode
207 : * of operation which isn't well supported. As such, the goal is not
208 : * taking too long when a different cgroup takes over an inode while
209 : * avoiding too aggressive flip-flops from occasional foreign writes.
210 : *
211 : * We record, very roughly, 2s worth of IO time history and if more than
212 : * half of that is foreign, trigger the switch. The recording is quantized
213 : * to 16 slots. To avoid tiny writes from swinging the decision too much,
214 : * writes smaller than 1/8 of avg size are ignored.
215 : */
216 : #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
217 : #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
218 : #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
219 : #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
220 :
221 : #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
222 : #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
223 : /* each slot's duration is 2s / 16 */
224 : #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
225 : /* if foreign slots >= 8, switch */
226 : #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
227 : /* one round can affect upto 5 slots */
228 : #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
229 :
230 : /*
231 : * Maximum inodes per isw. A specific value has been chosen to make
232 : * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
233 : */
234 : #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \
235 : / sizeof(struct inode *))
236 :
237 : static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
238 : static struct workqueue_struct *isw_wq;
239 :
240 8986733 : void __inode_attach_wb(struct inode *inode, struct folio *folio)
241 : {
242 8986733 : struct backing_dev_info *bdi = inode_to_bdi(inode);
243 8986637 : struct bdi_writeback *wb = NULL;
244 :
245 8986637 : if (inode_cgwb_enabled(inode)) {
246 8543505 : struct cgroup_subsys_state *memcg_css;
247 :
248 8543505 : if (folio) {
249 7566245 : memcg_css = mem_cgroup_css_from_folio(folio);
250 7565817 : wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
251 : } else {
252 : /* must pin memcg_css, see wb_get_create() */
253 977260 : memcg_css = task_get_css(current, memory_cgrp_id);
254 977238 : wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
255 977250 : css_put(memcg_css);
256 : }
257 : }
258 :
259 8545493 : if (!wb)
260 440773 : wb = &bdi->wb;
261 :
262 : /*
263 : * There may be multiple instances of this function racing to
264 : * update the same inode. Use cmpxchg() to tell the winner.
265 : */
266 8986266 : if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
267 0 : wb_put(wb);
268 8986329 : }
269 : EXPORT_SYMBOL_GPL(__inode_attach_wb);
270 :
271 : /**
272 : * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
273 : * @inode: inode of interest with i_lock held
274 : * @wb: target bdi_writeback
275 : *
276 : * Remove the inode from wb's io lists and if necessarily put onto b_attached
277 : * list. Only inodes attached to cgwb's are kept on this list.
278 : */
279 19639218 : static void inode_cgwb_move_to_attached(struct inode *inode,
280 : struct bdi_writeback *wb)
281 : {
282 19639218 : assert_spin_locked(&wb->list_lock);
283 19639218 : assert_spin_locked(&inode->i_lock);
284 19639218 : WARN_ON_ONCE(inode->i_state & I_FREEING);
285 :
286 19639218 : inode->i_state &= ~I_SYNC_QUEUED;
287 19639218 : if (wb != &wb->bdi->wb)
288 19549673 : list_move(&inode->i_io_list, &wb->b_attached);
289 : else
290 89545 : list_del_init(&inode->i_io_list);
291 19639215 : wb_io_lists_depopulated(wb);
292 19639199 : }
293 :
294 : /**
295 : * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
296 : * @inode: inode of interest with i_lock held
297 : *
298 : * Returns @inode's wb with its list_lock held. @inode->i_lock must be
299 : * held on entry and is released on return. The returned wb is guaranteed
300 : * to stay @inode's associated wb until its list_lock is released.
301 : */
302 : static struct bdi_writeback *
303 48476982 : locked_inode_to_wb_and_lock_list(struct inode *inode)
304 : __releases(&inode->i_lock)
305 : __acquires(&wb->list_lock)
306 : {
307 48476982 : while (true) {
308 48476982 : struct bdi_writeback *wb = inode_to_wb(inode);
309 :
310 : /*
311 : * inode_to_wb() association is protected by both
312 : * @inode->i_lock and @wb->list_lock but list_lock nests
313 : * outside i_lock. Drop i_lock and verify that the
314 : * association hasn't changed after acquiring list_lock.
315 : */
316 48476982 : wb_get(wb);
317 48475360 : spin_unlock(&inode->i_lock);
318 48475833 : spin_lock(&wb->list_lock);
319 :
320 : /* i_wb may have changed inbetween, can't use inode_to_wb() */
321 48480023 : if (likely(wb == inode->i_wb)) {
322 48480023 : wb_put(wb); /* @inode already has ref */
323 48480020 : return wb;
324 : }
325 :
326 0 : spin_unlock(&wb->list_lock);
327 0 : wb_put(wb);
328 0 : cpu_relax();
329 0 : spin_lock(&inode->i_lock);
330 : }
331 : }
332 :
333 : /**
334 : * inode_to_wb_and_lock_list - determine an inode's wb and lock it
335 : * @inode: inode of interest
336 : *
337 : * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
338 : * on entry.
339 : */
340 : static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
341 : __acquires(&wb->list_lock)
342 : {
343 27906904 : spin_lock(&inode->i_lock);
344 27907031 : return locked_inode_to_wb_and_lock_list(inode);
345 : }
346 :
347 : struct inode_switch_wbs_context {
348 : struct rcu_work work;
349 :
350 : /*
351 : * Multiple inodes can be switched at once. The switching procedure
352 : * consists of two parts, separated by a RCU grace period. To make
353 : * sure that the second part is executed for each inode gone through
354 : * the first part, all inode pointers are placed into a NULL-terminated
355 : * array embedded into struct inode_switch_wbs_context. Otherwise
356 : * an inode could be left in a non-consistent state.
357 : */
358 : struct bdi_writeback *new_wb;
359 : struct inode *inodes[];
360 : };
361 :
362 : static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
363 : {
364 6491874 : down_write(&bdi->wb_switch_rwsem);
365 : }
366 :
367 : static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
368 : {
369 6491857 : up_write(&bdi->wb_switch_rwsem);
370 : }
371 :
372 812321 : static bool inode_do_switch_wbs(struct inode *inode,
373 : struct bdi_writeback *old_wb,
374 : struct bdi_writeback *new_wb)
375 : {
376 812321 : struct address_space *mapping = inode->i_mapping;
377 812321 : XA_STATE(xas, &mapping->i_pages, 0);
378 812321 : struct folio *folio;
379 812321 : bool switched = false;
380 :
381 812321 : spin_lock(&inode->i_lock);
382 812321 : xa_lock_irq(&mapping->i_pages);
383 :
384 : /*
385 : * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
386 : * path owns the inode and we shouldn't modify ->i_io_list.
387 : */
388 812321 : if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE)))
389 0 : goto skip_switch;
390 :
391 812321 : trace_inode_switch_wbs(inode, old_wb, new_wb);
392 :
393 : /*
394 : * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
395 : * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
396 : * folios actually under writeback.
397 : */
398 812563 : xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
399 242 : if (folio_test_dirty(folio)) {
400 242 : long nr = folio_nr_pages(folio);
401 242 : wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr);
402 242 : wb_stat_mod(new_wb, WB_RECLAIMABLE, nr);
403 : }
404 : }
405 :
406 812321 : xas_set(&xas, 0);
407 845627 : xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
408 33306 : long nr = folio_nr_pages(folio);
409 33306 : WARN_ON_ONCE(!folio_test_writeback(folio));
410 33306 : wb_stat_mod(old_wb, WB_WRITEBACK, -nr);
411 33306 : wb_stat_mod(new_wb, WB_WRITEBACK, nr);
412 : }
413 :
414 812321 : if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
415 36 : atomic_dec(&old_wb->writeback_inodes);
416 36 : atomic_inc(&new_wb->writeback_inodes);
417 : }
418 :
419 812321 : wb_get(new_wb);
420 :
421 : /*
422 : * Transfer to @new_wb's IO list if necessary. If the @inode is dirty,
423 : * the specific list @inode was on is ignored and the @inode is put on
424 : * ->b_dirty which is always correct including from ->b_dirty_time.
425 : * The transfer preserves @inode->dirtied_when ordering. If the @inode
426 : * was clean, it means it was on the b_attached list, so move it onto
427 : * the b_attached list of @new_wb.
428 : */
429 812321 : if (!list_empty(&inode->i_io_list)) {
430 812321 : inode->i_wb = new_wb;
431 :
432 812321 : if (inode->i_state & I_DIRTY_ALL) {
433 53 : struct inode *pos;
434 :
435 53 : list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
436 5 : if (time_after_eq(inode->dirtied_when,
437 : pos->dirtied_when))
438 : break;
439 53 : inode_io_list_move_locked(inode, new_wb,
440 : pos->i_io_list.prev);
441 : } else {
442 812268 : inode_cgwb_move_to_attached(inode, new_wb);
443 : }
444 : } else {
445 0 : inode->i_wb = new_wb;
446 : }
447 :
448 : /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
449 812321 : inode->i_wb_frn_winner = 0;
450 812321 : inode->i_wb_frn_avg_time = 0;
451 812321 : inode->i_wb_frn_history = 0;
452 812321 : switched = true;
453 812321 : skip_switch:
454 : /*
455 : * Paired with load_acquire in unlocked_inode_to_wb_begin() and
456 : * ensures that the new wb is visible if they see !I_WB_SWITCH.
457 : */
458 812321 : smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
459 :
460 812321 : xa_unlock_irq(&mapping->i_pages);
461 812321 : spin_unlock(&inode->i_lock);
462 :
463 812321 : return switched;
464 : }
465 :
466 7338 : static void inode_switch_wbs_work_fn(struct work_struct *work)
467 : {
468 7338 : struct inode_switch_wbs_context *isw =
469 7338 : container_of(to_rcu_work(work), struct inode_switch_wbs_context, work);
470 7338 : struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
471 7338 : struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
472 7338 : struct bdi_writeback *new_wb = isw->new_wb;
473 7338 : unsigned long nr_switched = 0;
474 7338 : struct inode **inodep;
475 :
476 : /*
477 : * If @inode switches cgwb membership while sync_inodes_sb() is
478 : * being issued, sync_inodes_sb() might miss it. Synchronize.
479 : */
480 7338 : down_read(&bdi->wb_switch_rwsem);
481 :
482 : /*
483 : * By the time control reaches here, RCU grace period has passed
484 : * since I_WB_SWITCH assertion and all wb stat update transactions
485 : * between unlocked_inode_to_wb_begin/end() are guaranteed to be
486 : * synchronizing against the i_pages lock.
487 : *
488 : * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
489 : * gives us exclusion against all wb related operations on @inode
490 : * including IO list manipulations and stat updates.
491 : */
492 7338 : if (old_wb < new_wb) {
493 2721 : spin_lock(&old_wb->list_lock);
494 2721 : spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
495 : } else {
496 4617 : spin_lock(&new_wb->list_lock);
497 4617 : spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
498 : }
499 :
500 819659 : for (inodep = isw->inodes; *inodep; inodep++) {
501 812321 : WARN_ON_ONCE((*inodep)->i_wb != old_wb);
502 812321 : if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
503 812321 : nr_switched++;
504 : }
505 :
506 7338 : spin_unlock(&new_wb->list_lock);
507 7338 : spin_unlock(&old_wb->list_lock);
508 :
509 7338 : up_read(&bdi->wb_switch_rwsem);
510 :
511 7338 : if (nr_switched) {
512 7338 : wb_wakeup(new_wb);
513 7338 : wb_put_many(old_wb, nr_switched);
514 : }
515 :
516 819659 : for (inodep = isw->inodes; *inodep; inodep++)
517 812321 : iput(*inodep);
518 7338 : wb_put(new_wb);
519 7338 : kfree(isw);
520 7338 : atomic_dec(&isw_nr_in_flight);
521 7338 : }
522 :
523 29236781 : static bool inode_prepare_wbs_switch(struct inode *inode,
524 : struct bdi_writeback *new_wb)
525 : {
526 : /*
527 : * Paired with smp_mb() in cgroup_writeback_umount().
528 : * isw_nr_in_flight must be increased before checking SB_ACTIVE and
529 : * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
530 : * in cgroup_writeback_umount() and the isw_wq will be not flushed.
531 : */
532 29236781 : smp_mb();
533 :
534 29236781 : if (IS_DAX(inode))
535 : return false;
536 :
537 : /* while holding I_WB_SWITCH, no one else can update the association */
538 29236781 : spin_lock(&inode->i_lock);
539 29236781 : if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
540 29236781 : inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
541 : inode_to_wb(inode) == new_wb) {
542 28424460 : spin_unlock(&inode->i_lock);
543 28424460 : return false;
544 : }
545 812321 : inode->i_state |= I_WB_SWITCH;
546 812321 : __iget(inode);
547 812321 : spin_unlock(&inode->i_lock);
548 :
549 812321 : return true;
550 : }
551 :
552 : /**
553 : * inode_switch_wbs - change the wb association of an inode
554 : * @inode: target inode
555 : * @new_wb_id: ID of the new wb
556 : *
557 : * Switch @inode's wb association to the wb identified by @new_wb_id. The
558 : * switching is performed asynchronously and may fail silently.
559 : */
560 167 : static void inode_switch_wbs(struct inode *inode, int new_wb_id)
561 : {
562 167 : struct backing_dev_info *bdi = inode_to_bdi(inode);
563 167 : struct cgroup_subsys_state *memcg_css;
564 167 : struct inode_switch_wbs_context *isw;
565 :
566 : /* noop if seems to be already in progress */
567 167 : if (inode->i_state & I_WB_SWITCH)
568 : return;
569 :
570 : /* avoid queueing a new switch if too many are already in flight */
571 104 : if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
572 : return;
573 :
574 104 : isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC);
575 104 : if (!isw)
576 : return;
577 :
578 104 : atomic_inc(&isw_nr_in_flight);
579 :
580 : /* find and pin the new wb */
581 104 : rcu_read_lock();
582 104 : memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
583 104 : if (memcg_css && !css_tryget(memcg_css))
584 0 : memcg_css = NULL;
585 104 : rcu_read_unlock();
586 104 : if (!memcg_css)
587 0 : goto out_free;
588 :
589 104 : isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
590 104 : css_put(memcg_css);
591 104 : if (!isw->new_wb)
592 1 : goto out_free;
593 :
594 103 : if (!inode_prepare_wbs_switch(inode, isw->new_wb))
595 0 : goto out_free;
596 :
597 103 : isw->inodes[0] = inode;
598 :
599 : /*
600 : * In addition to synchronizing among switchers, I_WB_SWITCH tells
601 : * the RCU protected stat update paths to grab the i_page
602 : * lock so that stat transfer can synchronize against them.
603 : * Let's continue after I_WB_SWITCH is guaranteed to be visible.
604 : */
605 103 : INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
606 103 : queue_rcu_work(isw_wq, &isw->work);
607 103 : return;
608 :
609 1 : out_free:
610 1 : atomic_dec(&isw_nr_in_flight);
611 1 : if (isw->new_wb)
612 0 : wb_put(isw->new_wb);
613 1 : kfree(isw);
614 : }
615 :
616 : /**
617 : * cleanup_offline_cgwb - detach associated inodes
618 : * @wb: target wb
619 : *
620 : * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
621 : * to eventually release the dying @wb. Returns %true if not all inodes were
622 : * switched and the function has to be restarted.
623 : */
624 8610 : bool cleanup_offline_cgwb(struct bdi_writeback *wb)
625 : {
626 8610 : struct cgroup_subsys_state *memcg_css;
627 8610 : struct inode_switch_wbs_context *isw;
628 8610 : struct inode *inode;
629 8610 : int nr;
630 8610 : bool restart = false;
631 :
632 8610 : isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW),
633 : GFP_KERNEL);
634 8610 : if (!isw)
635 : return restart;
636 :
637 8610 : atomic_inc(&isw_nr_in_flight);
638 :
639 8611 : for (memcg_css = wb->memcg_css->parent; memcg_css;
640 1 : memcg_css = memcg_css->parent) {
641 8611 : isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
642 8611 : if (isw->new_wb)
643 : break;
644 : }
645 8610 : if (unlikely(!isw->new_wb))
646 0 : isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
647 :
648 8610 : nr = 0;
649 8610 : spin_lock(&wb->list_lock);
650 29238509 : list_for_each_entry(inode, &wb->b_attached, i_io_list) {
651 29236678 : if (!inode_prepare_wbs_switch(inode, isw->new_wb))
652 28424460 : continue;
653 :
654 812218 : isw->inodes[nr++] = inode;
655 :
656 812218 : if (nr >= WB_MAX_INODES_PER_ISW - 1) {
657 : restart = true;
658 : break;
659 : }
660 : }
661 8610 : spin_unlock(&wb->list_lock);
662 :
663 : /* no attached inodes? bail out */
664 8610 : if (nr == 0) {
665 1375 : atomic_dec(&isw_nr_in_flight);
666 1375 : wb_put(isw->new_wb);
667 1375 : kfree(isw);
668 1375 : return restart;
669 : }
670 :
671 : /*
672 : * In addition to synchronizing among switchers, I_WB_SWITCH tells
673 : * the RCU protected stat update paths to grab the i_page
674 : * lock so that stat transfer can synchronize against them.
675 : * Let's continue after I_WB_SWITCH is guaranteed to be visible.
676 : */
677 7235 : INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn);
678 7235 : queue_rcu_work(isw_wq, &isw->work);
679 :
680 7235 : return restart;
681 : }
682 :
683 : /**
684 : * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
685 : * @wbc: writeback_control of interest
686 : * @inode: target inode
687 : *
688 : * @inode is locked and about to be written back under the control of @wbc.
689 : * Record @inode's writeback context into @wbc and unlock the i_lock. On
690 : * writeback completion, wbc_detach_inode() should be called. This is used
691 : * to track the cgroup writeback context.
692 : */
693 36759740 : void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
694 : struct inode *inode)
695 : {
696 36759740 : if (!inode_cgwb_enabled(inode)) {
697 105119 : spin_unlock(&inode->i_lock);
698 105119 : return;
699 : }
700 :
701 36654257 : wbc->wb = inode_to_wb(inode);
702 36654257 : wbc->inode = inode;
703 :
704 36654257 : wbc->wb_id = wbc->wb->memcg_css->id;
705 36654257 : wbc->wb_lcand_id = inode->i_wb_frn_winner;
706 36654257 : wbc->wb_tcand_id = 0;
707 36654257 : wbc->wb_bytes = 0;
708 36654257 : wbc->wb_lcand_bytes = 0;
709 36654257 : wbc->wb_tcand_bytes = 0;
710 :
711 36654257 : wb_get(wbc->wb);
712 36654250 : spin_unlock(&inode->i_lock);
713 :
714 : /*
715 : * A dying wb indicates that either the blkcg associated with the
716 : * memcg changed or the associated memcg is dying. In the first
717 : * case, a replacement wb should already be available and we should
718 : * refresh the wb immediately. In the second case, trying to
719 : * refresh will keep failing.
720 : */
721 36685361 : if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
722 0 : inode_switch_wbs(inode, wbc->wb_id);
723 : }
724 : EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
725 :
726 : /**
727 : * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
728 : * @wbc: writeback_control of the just finished writeback
729 : *
730 : * To be called after a writeback attempt of an inode finishes and undoes
731 : * wbc_attach_and_unlock_inode(). Can be called under any context.
732 : *
733 : * As concurrent write sharing of an inode is expected to be very rare and
734 : * memcg only tracks page ownership on first-use basis severely confining
735 : * the usefulness of such sharing, cgroup writeback tracks ownership
736 : * per-inode. While the support for concurrent write sharing of an inode
737 : * is deemed unnecessary, an inode being written to by different cgroups at
738 : * different points in time is a lot more common, and, more importantly,
739 : * charging only by first-use can too readily lead to grossly incorrect
740 : * behaviors (single foreign page can lead to gigabytes of writeback to be
741 : * incorrectly attributed).
742 : *
743 : * To resolve this issue, cgroup writeback detects the majority dirtier of
744 : * an inode and transfers the ownership to it. To avoid unnecessary
745 : * oscillation, the detection mechanism keeps track of history and gives
746 : * out the switch verdict only if the foreign usage pattern is stable over
747 : * a certain amount of time and/or writeback attempts.
748 : *
749 : * On each writeback attempt, @wbc tries to detect the majority writer
750 : * using Boyer-Moore majority vote algorithm. In addition to the byte
751 : * count from the majority voting, it also counts the bytes written for the
752 : * current wb and the last round's winner wb (max of last round's current
753 : * wb, the winner from two rounds ago, and the last round's majority
754 : * candidate). Keeping track of the historical winner helps the algorithm
755 : * to semi-reliably detect the most active writer even when it's not the
756 : * absolute majority.
757 : *
758 : * Once the winner of the round is determined, whether the winner is
759 : * foreign or not and how much IO time the round consumed is recorded in
760 : * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
761 : * over a certain threshold, the switch verdict is given.
762 : */
763 36759963 : void wbc_detach_inode(struct writeback_control *wbc)
764 : {
765 36759963 : struct bdi_writeback *wb = wbc->wb;
766 36759963 : struct inode *inode = wbc->inode;
767 36759963 : unsigned long avg_time, max_bytes, max_time;
768 36759963 : u16 history;
769 36759963 : int max_id;
770 :
771 36759963 : if (!wb)
772 : return;
773 :
774 36654844 : history = inode->i_wb_frn_history;
775 36654844 : avg_time = inode->i_wb_frn_avg_time;
776 :
777 : /* pick the winner of this round */
778 36654844 : if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
779 36654448 : wbc->wb_bytes >= wbc->wb_tcand_bytes) {
780 36653059 : max_id = wbc->wb_id;
781 36653059 : max_bytes = wbc->wb_bytes;
782 1785 : } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
783 212 : max_id = wbc->wb_lcand_id;
784 212 : max_bytes = wbc->wb_lcand_bytes;
785 : } else {
786 1573 : max_id = wbc->wb_tcand_id;
787 1573 : max_bytes = wbc->wb_tcand_bytes;
788 : }
789 :
790 : /*
791 : * Calculate the amount of IO time the winner consumed and fold it
792 : * into the running average kept per inode. If the consumed IO
793 : * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
794 : * deciding whether to switch or not. This is to prevent one-off
795 : * small dirtiers from skewing the verdict.
796 : */
797 36654844 : max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
798 : wb->avg_write_bandwidth);
799 36654844 : if (avg_time)
800 23797977 : avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
801 23797977 : (avg_time >> WB_FRN_TIME_AVG_SHIFT);
802 : else
803 : avg_time = max_time; /* immediate catch up on first run */
804 :
805 36654844 : if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
806 28529490 : int slots;
807 :
808 : /*
809 : * The switch verdict is reached if foreign wb's consume
810 : * more than a certain proportion of IO time in a
811 : * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
812 : * history mask where each bit represents one sixteenth of
813 : * the period. Determine the number of slots to shift into
814 : * history from @max_time.
815 : */
816 28529490 : slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
817 : (unsigned long)WB_FRN_HIST_MAX_SLOTS);
818 28529490 : history <<= slots;
819 28529490 : if (wbc->wb_id != max_id)
820 1550 : history |= (1U << slots) - 1;
821 :
822 28529490 : if (history)
823 7571 : trace_inode_foreign_history(inode, wbc, history);
824 :
825 : /*
826 : * Switch if the current wb isn't the consistent winner.
827 : * If there are multiple closely competing dirtiers, the
828 : * inode may switch across them repeatedly over time, which
829 : * is okay. The main goal is avoiding keeping an inode on
830 : * the wrong wb for an extended period of time.
831 : */
832 28529490 : if (hweight16(history) > WB_FRN_HIST_THR_SLOTS)
833 167 : inode_switch_wbs(inode, max_id);
834 : }
835 :
836 : /*
837 : * Multiple instances of this function may race to update the
838 : * following fields but we don't mind occassional inaccuracies.
839 : */
840 36654846 : inode->i_wb_frn_winner = max_id;
841 36654846 : inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
842 36654846 : inode->i_wb_frn_history = history;
843 :
844 36654846 : wb_put(wbc->wb);
845 36654770 : wbc->wb = NULL;
846 : }
847 : EXPORT_SYMBOL_GPL(wbc_detach_inode);
848 :
849 : /**
850 : * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
851 : * @wbc: writeback_control of the writeback in progress
852 : * @page: page being written out
853 : * @bytes: number of bytes being written out
854 : *
855 : * @bytes from @page are about to written out during the writeback
856 : * controlled by @wbc. Keep the book for foreign inode detection. See
857 : * wbc_detach_inode().
858 : */
859 402809879 : void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
860 : size_t bytes)
861 : {
862 402809879 : struct folio *folio;
863 402809879 : struct cgroup_subsys_state *css;
864 402809879 : int id;
865 :
866 : /*
867 : * pageout() path doesn't attach @wbc to the inode being written
868 : * out. This is intentional as we don't want the function to block
869 : * behind a slow cgroup. Ultimately, we want pageout() to kick off
870 : * regular writeback instead of writing things out itself.
871 : */
872 402809879 : if (!wbc->wb || wbc->no_cgroup_owner)
873 : return;
874 :
875 402809879 : folio = page_folio(page);
876 402809879 : css = mem_cgroup_css_from_folio(folio);
877 : /* dead cgroups shouldn't contribute to inode ownership arbitration */
878 402811464 : if (!(css->flags & CSS_ONLINE))
879 : return;
880 :
881 401932235 : id = css->id;
882 :
883 401932235 : if (id == wbc->wb_id) {
884 401089441 : wbc->wb_bytes += bytes;
885 401089441 : return;
886 : }
887 :
888 842794 : if (id == wbc->wb_lcand_id)
889 11447 : wbc->wb_lcand_bytes += bytes;
890 :
891 : /* Boyer-Moore majority vote algorithm */
892 842794 : if (!wbc->wb_tcand_bytes)
893 2960 : wbc->wb_tcand_id = id;
894 842794 : if (id == wbc->wb_tcand_id)
895 842713 : wbc->wb_tcand_bytes += bytes;
896 : else
897 81 : wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
898 : }
899 : EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
900 :
901 : /**
902 : * wb_split_bdi_pages - split nr_pages to write according to bandwidth
903 : * @wb: target bdi_writeback to split @nr_pages to
904 : * @nr_pages: number of pages to write for the whole bdi
905 : *
906 : * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
907 : * relation to the total write bandwidth of all wb's w/ dirty inodes on
908 : * @wb->bdi.
909 : */
910 2109610 : static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
911 : {
912 2109610 : unsigned long this_bw = wb->avg_write_bandwidth;
913 2109610 : unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
914 :
915 2109610 : if (nr_pages == LONG_MAX)
916 : return LONG_MAX;
917 :
918 : /*
919 : * This may be called on clean wb's and proportional distribution
920 : * may not make sense, just use the original @nr_pages in those
921 : * cases. In general, we wanna err on the side of writing more.
922 : */
923 1075895 : if (!tot_bw || this_bw >= tot_bw)
924 : return nr_pages;
925 : else
926 136139 : return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
927 : }
928 :
929 : /**
930 : * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
931 : * @bdi: target backing_dev_info
932 : * @base_work: wb_writeback_work to issue
933 : * @skip_if_busy: skip wb's which already have writeback in progress
934 : *
935 : * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
936 : * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
937 : * distributed to the busy wbs according to each wb's proportion in the
938 : * total active write bandwidth of @bdi.
939 : */
940 6550978 : static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
941 : struct wb_writeback_work *base_work,
942 : bool skip_if_busy)
943 : {
944 6550978 : struct bdi_writeback *last_wb = NULL;
945 6550978 : struct bdi_writeback *wb = list_entry(&bdi->wb_list,
946 : struct bdi_writeback, bdi_node);
947 :
948 6550978 : might_sleep();
949 6550979 : restart:
950 6550979 : rcu_read_lock();
951 20767920 : list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
952 14216970 : DEFINE_WB_COMPLETION(fallback_work_done, bdi);
953 14216970 : struct wb_writeback_work fallback_work;
954 14216970 : struct wb_writeback_work *work;
955 14216970 : long nr_pages;
956 :
957 14216970 : if (last_wb) {
958 0 : wb_put(last_wb);
959 0 : last_wb = NULL;
960 : }
961 :
962 : /* SYNC_ALL writes out I_DIRTY_TIME too */
963 28433886 : if (!wb_has_dirty_io(wb) &&
964 13122296 : (base_work->sync_mode == WB_SYNC_NONE ||
965 12926010 : list_empty(&wb->b_dirty_time)))
966 14216944 : continue;
967 1094575 : if (skip_if_busy && writeback_in_progress(wb))
968 0 : continue;
969 :
970 1094575 : nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
971 :
972 1094575 : work = kmalloc(sizeof(*work), GFP_ATOMIC);
973 1094575 : if (work) {
974 1094575 : *work = *base_work;
975 1094575 : work->nr_pages = nr_pages;
976 1094575 : work->auto_free = 1;
977 1094575 : wb_queue_work(wb, work);
978 1094576 : continue;
979 : }
980 :
981 : /*
982 : * If wb_tryget fails, the wb has been shutdown, skip it.
983 : *
984 : * Pin @wb so that it stays on @bdi->wb_list. This allows
985 : * continuing iteration from @wb after dropping and
986 : * regrabbing rcu read lock.
987 : */
988 0 : if (!wb_tryget(wb))
989 0 : continue;
990 :
991 : /* alloc failed, execute synchronously using on-stack fallback */
992 0 : work = &fallback_work;
993 0 : *work = *base_work;
994 0 : work->nr_pages = nr_pages;
995 0 : work->auto_free = 0;
996 0 : work->done = &fallback_work_done;
997 :
998 0 : wb_queue_work(wb, work);
999 0 : last_wb = wb;
1000 :
1001 0 : rcu_read_unlock();
1002 0 : wb_wait_for_completion(&fallback_work_done);
1003 0 : goto restart;
1004 : }
1005 6550950 : rcu_read_unlock();
1006 :
1007 6550953 : if (last_wb)
1008 0 : wb_put(last_wb);
1009 6550953 : }
1010 :
1011 : /**
1012 : * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1013 : * @bdi_id: target bdi id
1014 : * @memcg_id: target memcg css id
1015 : * @reason: reason why some writeback work initiated
1016 : * @done: target wb_completion
1017 : *
1018 : * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1019 : * with the specified parameters.
1020 : */
1021 0 : int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
1022 : enum wb_reason reason, struct wb_completion *done)
1023 : {
1024 0 : struct backing_dev_info *bdi;
1025 0 : struct cgroup_subsys_state *memcg_css;
1026 0 : struct bdi_writeback *wb;
1027 0 : struct wb_writeback_work *work;
1028 0 : unsigned long dirty;
1029 0 : int ret;
1030 :
1031 : /* lookup bdi and memcg */
1032 0 : bdi = bdi_get_by_id(bdi_id);
1033 0 : if (!bdi)
1034 : return -ENOENT;
1035 :
1036 0 : rcu_read_lock();
1037 0 : memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1038 0 : if (memcg_css && !css_tryget(memcg_css))
1039 0 : memcg_css = NULL;
1040 0 : rcu_read_unlock();
1041 0 : if (!memcg_css) {
1042 0 : ret = -ENOENT;
1043 0 : goto out_bdi_put;
1044 : }
1045 :
1046 : /*
1047 : * And find the associated wb. If the wb isn't there already
1048 : * there's nothing to flush, don't create one.
1049 : */
1050 0 : wb = wb_get_lookup(bdi, memcg_css);
1051 0 : if (!wb) {
1052 0 : ret = -ENOENT;
1053 0 : goto out_css_put;
1054 : }
1055 :
1056 : /*
1057 : * The caller is attempting to write out most of
1058 : * the currently dirty pages. Let's take the current dirty page
1059 : * count and inflate it by 25% which should be large enough to
1060 : * flush out most dirty pages while avoiding getting livelocked by
1061 : * concurrent dirtiers.
1062 : *
1063 : * BTW the memcg stats are flushed periodically and this is best-effort
1064 : * estimation, so some potential error is ok.
1065 : */
1066 0 : dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1067 0 : dirty = dirty * 10 / 8;
1068 :
1069 : /* issue the writeback work */
1070 0 : work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
1071 0 : if (work) {
1072 0 : work->nr_pages = dirty;
1073 0 : work->sync_mode = WB_SYNC_NONE;
1074 0 : work->range_cyclic = 1;
1075 0 : work->reason = reason;
1076 0 : work->done = done;
1077 0 : work->auto_free = 1;
1078 0 : wb_queue_work(wb, work);
1079 0 : ret = 0;
1080 : } else {
1081 : ret = -ENOMEM;
1082 : }
1083 :
1084 0 : wb_put(wb);
1085 0 : out_css_put:
1086 0 : css_put(memcg_css);
1087 0 : out_bdi_put:
1088 0 : bdi_put(bdi);
1089 0 : return ret;
1090 : }
1091 :
1092 : /**
1093 : * cgroup_writeback_umount - flush inode wb switches for umount
1094 : *
1095 : * This function is called when a super_block is about to be destroyed and
1096 : * flushes in-flight inode wb switches. An inode wb switch goes through
1097 : * RCU and then workqueue, so the two need to be flushed in order to ensure
1098 : * that all previously scheduled switches are finished. As wb switches are
1099 : * rare occurrences and synchronize_rcu() can take a while, perform
1100 : * flushing iff wb switches are in flight.
1101 : */
1102 32784 : void cgroup_writeback_umount(void)
1103 : {
1104 : /*
1105 : * SB_ACTIVE should be reliably cleared before checking
1106 : * isw_nr_in_flight, see generic_shutdown_super().
1107 : */
1108 32784 : smp_mb();
1109 :
1110 32777 : if (atomic_read(&isw_nr_in_flight)) {
1111 : /*
1112 : * Use rcu_barrier() to wait for all pending callbacks to
1113 : * ensure that all in-flight wb switches are in the workqueue.
1114 : */
1115 44 : rcu_barrier();
1116 44 : flush_workqueue(isw_wq);
1117 : }
1118 32777 : }
1119 :
1120 0 : static int __init cgroup_writeback_init(void)
1121 : {
1122 0 : isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1123 0 : if (!isw_wq)
1124 0 : return -ENOMEM;
1125 : return 0;
1126 : }
1127 : fs_initcall(cgroup_writeback_init);
1128 :
1129 : #else /* CONFIG_CGROUP_WRITEBACK */
1130 :
1131 : static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1132 : static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1133 :
1134 : static void inode_cgwb_move_to_attached(struct inode *inode,
1135 : struct bdi_writeback *wb)
1136 : {
1137 : assert_spin_locked(&wb->list_lock);
1138 : assert_spin_locked(&inode->i_lock);
1139 : WARN_ON_ONCE(inode->i_state & I_FREEING);
1140 :
1141 : inode->i_state &= ~I_SYNC_QUEUED;
1142 : list_del_init(&inode->i_io_list);
1143 : wb_io_lists_depopulated(wb);
1144 : }
1145 :
1146 : static struct bdi_writeback *
1147 : locked_inode_to_wb_and_lock_list(struct inode *inode)
1148 : __releases(&inode->i_lock)
1149 : __acquires(&wb->list_lock)
1150 : {
1151 : struct bdi_writeback *wb = inode_to_wb(inode);
1152 :
1153 : spin_unlock(&inode->i_lock);
1154 : spin_lock(&wb->list_lock);
1155 : return wb;
1156 : }
1157 :
1158 : static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1159 : __acquires(&wb->list_lock)
1160 : {
1161 : struct bdi_writeback *wb = inode_to_wb(inode);
1162 :
1163 : spin_lock(&wb->list_lock);
1164 : return wb;
1165 : }
1166 :
1167 : static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1168 : {
1169 : return nr_pages;
1170 : }
1171 :
1172 : static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1173 : struct wb_writeback_work *base_work,
1174 : bool skip_if_busy)
1175 : {
1176 : might_sleep();
1177 :
1178 : if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1179 : base_work->auto_free = 0;
1180 : wb_queue_work(&bdi->wb, base_work);
1181 : }
1182 : }
1183 :
1184 : #endif /* CONFIG_CGROUP_WRITEBACK */
1185 :
1186 : /*
1187 : * Add in the number of potentially dirty inodes, because each inode
1188 : * write can dirty pagecache in the underlying blockdev.
1189 : */
1190 : static unsigned long get_nr_dirty_pages(void)
1191 : {
1192 2284340 : return global_node_page_state(NR_FILE_DIRTY) +
1193 1142170 : get_nr_dirty_inodes();
1194 : }
1195 :
1196 3823336 : static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1197 : {
1198 7646672 : if (!wb_has_dirty_io(wb))
1199 : return;
1200 :
1201 : /*
1202 : * All callers of this function want to start writeback of all
1203 : * dirty pages. Places like vmscan can call this at a very
1204 : * high frequency, causing pointless allocations of tons of
1205 : * work items and keeping the flusher threads busy retrieving
1206 : * that work. Ensure that we only allow one of them pending and
1207 : * inflight at the time.
1208 : */
1209 2061581 : if (test_bit(WB_start_all, &wb->state) ||
1210 : test_and_set_bit(WB_start_all, &wb->state))
1211 31506 : return;
1212 :
1213 1015038 : wb->start_all_reason = reason;
1214 1015038 : wb_wakeup(wb);
1215 : }
1216 :
1217 : /**
1218 : * wb_start_background_writeback - start background writeback
1219 : * @wb: bdi_writback to write from
1220 : *
1221 : * Description:
1222 : * This makes sure WB_SYNC_NONE background writeback happens. When
1223 : * this function returns, it is only guaranteed that for given wb
1224 : * some IO is happening if we are over background dirty threshold.
1225 : * Caller need not hold sb s_umount semaphore.
1226 : */
1227 1319 : void wb_start_background_writeback(struct bdi_writeback *wb)
1228 : {
1229 : /*
1230 : * We just wake up the flusher thread. It will perform background
1231 : * writeback as soon as there is no other work to do.
1232 : */
1233 1319 : trace_writeback_wake_background(wb);
1234 1319 : wb_wakeup(wb);
1235 1319 : }
1236 :
1237 : /*
1238 : * Remove the inode from the writeback list it is on.
1239 : */
1240 8572423 : void inode_io_list_del(struct inode *inode)
1241 : {
1242 8572423 : struct bdi_writeback *wb;
1243 :
1244 8572423 : wb = inode_to_wb_and_lock_list(inode);
1245 8572725 : spin_lock(&inode->i_lock);
1246 :
1247 8572725 : inode->i_state &= ~I_SYNC_QUEUED;
1248 8572725 : list_del_init(&inode->i_io_list);
1249 8572725 : wb_io_lists_depopulated(wb);
1250 :
1251 8572725 : spin_unlock(&inode->i_lock);
1252 8572725 : spin_unlock(&wb->list_lock);
1253 8572725 : }
1254 : EXPORT_SYMBOL(inode_io_list_del);
1255 :
1256 : /*
1257 : * mark an inode as under writeback on the sb
1258 : */
1259 21989288 : void sb_mark_inode_writeback(struct inode *inode)
1260 : {
1261 21989288 : struct super_block *sb = inode->i_sb;
1262 21989288 : unsigned long flags;
1263 :
1264 21989288 : if (list_empty(&inode->i_wb_list)) {
1265 21989886 : spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1266 21989916 : if (list_empty(&inode->i_wb_list)) {
1267 21989916 : list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1268 21989915 : trace_sb_mark_inode_writeback(inode);
1269 : }
1270 21989912 : spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1271 : }
1272 21989307 : }
1273 :
1274 : /*
1275 : * clear an inode as under writeback on the sb
1276 : */
1277 21989921 : void sb_clear_inode_writeback(struct inode *inode)
1278 : {
1279 21989921 : struct super_block *sb = inode->i_sb;
1280 21989921 : unsigned long flags;
1281 :
1282 21989921 : if (!list_empty(&inode->i_wb_list)) {
1283 21989921 : spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1284 21989925 : if (!list_empty(&inode->i_wb_list)) {
1285 21989925 : list_del_init(&inode->i_wb_list);
1286 21989923 : trace_sb_clear_inode_writeback(inode);
1287 : }
1288 21989923 : spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1289 : }
1290 21989921 : }
1291 :
1292 : /*
1293 : * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1294 : * furthest end of its superblock's dirty-inode list.
1295 : *
1296 : * Before stamping the inode's ->dirtied_when, we check to see whether it is
1297 : * already the most-recently-dirtied inode on the b_dirty list. If that is
1298 : * the case then the inode must have been redirtied while it was being written
1299 : * out and we don't reset its dirtied_when.
1300 : */
1301 3153154 : static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1302 : {
1303 3153154 : assert_spin_locked(&inode->i_lock);
1304 :
1305 3153154 : inode->i_state &= ~I_SYNC_QUEUED;
1306 : /*
1307 : * When the inode is being freed just don't bother with dirty list
1308 : * tracking. Flush worker will ignore this inode anyway and it will
1309 : * trigger assertions in inode_io_list_move_locked().
1310 : */
1311 3153154 : if (inode->i_state & I_FREEING) {
1312 3 : list_del_init(&inode->i_io_list);
1313 3 : wb_io_lists_depopulated(wb);
1314 3 : return;
1315 : }
1316 3153151 : if (!list_empty(&wb->b_dirty)) {
1317 2808308 : struct inode *tail;
1318 :
1319 2808308 : tail = wb_inode(wb->b_dirty.next);
1320 2808308 : if (time_before(inode->dirtied_when, tail->dirtied_when))
1321 46994 : inode->dirtied_when = jiffies;
1322 : }
1323 3153151 : inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1324 : }
1325 :
1326 2648756 : static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1327 : {
1328 2648756 : spin_lock(&inode->i_lock);
1329 2648756 : redirty_tail_locked(inode, wb);
1330 2648756 : spin_unlock(&inode->i_lock);
1331 2648756 : }
1332 :
1333 : /*
1334 : * requeue inode for re-scanning after bdi->b_io list is exhausted.
1335 : */
1336 : static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1337 : {
1338 3041 : inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1339 3041 : }
1340 :
1341 19334462 : static void inode_sync_complete(struct inode *inode)
1342 : {
1343 19334462 : inode->i_state &= ~I_SYNC;
1344 : /* If inode is clean an unused, put it into LRU now... */
1345 19334462 : inode_add_lru(inode);
1346 : /* Waiters must see I_SYNC cleared before being woken up */
1347 19334464 : smp_mb();
1348 19334465 : wake_up_bit(&inode->i_state, __I_SYNC);
1349 19334470 : }
1350 :
1351 : static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1352 : {
1353 44243230 : bool ret = time_after(inode->dirtied_when, t);
1354 : #ifndef CONFIG_64BIT
1355 : /*
1356 : * For inodes being constantly redirtied, dirtied_when can get stuck.
1357 : * It _appears_ to be in the future, but is actually in distant past.
1358 : * This test is necessary to prevent such wrapped-around relative times
1359 : * from permanently stopping the whole bdi writeback.
1360 : */
1361 : ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1362 : #endif
1363 22121615 : return ret;
1364 : }
1365 :
1366 : /*
1367 : * Move expired (dirtied before dirtied_before) dirty inodes from
1368 : * @delaying_queue to @dispatch_queue.
1369 : */
1370 7302905 : static int move_expired_inodes(struct list_head *delaying_queue,
1371 : struct list_head *dispatch_queue,
1372 : unsigned long dirtied_before)
1373 : {
1374 7302905 : LIST_HEAD(tmp);
1375 7302905 : struct list_head *pos, *node;
1376 7302905 : struct super_block *sb = NULL;
1377 7302905 : struct inode *inode;
1378 7302905 : int do_sb_sort = 0;
1379 7302905 : int moved = 0;
1380 :
1381 29300069 : while (!list_empty(delaying_queue)) {
1382 22121615 : inode = wb_inode(delaying_queue->prev);
1383 44118224 : if (inode_dirtied_after(inode, dirtied_before))
1384 : break;
1385 21997034 : spin_lock(&inode->i_lock);
1386 21996669 : list_move(&inode->i_io_list, &tmp);
1387 21995863 : moved++;
1388 21995863 : inode->i_state |= I_SYNC_QUEUED;
1389 21995863 : spin_unlock(&inode->i_lock);
1390 21997164 : if (sb_is_blkdev_sb(inode->i_sb))
1391 289 : continue;
1392 21996875 : if (sb && sb != inode->i_sb)
1393 12 : do_sb_sort = 1;
1394 : sb = inode->i_sb;
1395 : }
1396 :
1397 : /* just one sb in list, splice to dispatch_queue and we're done */
1398 7303035 : if (!do_sb_sort) {
1399 7303023 : list_splice(&tmp, dispatch_queue);
1400 7303023 : goto out;
1401 : }
1402 :
1403 : /*
1404 : * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1405 : * we don't take inode->i_lock here because it is just a pointless overhead.
1406 : * Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1407 : * fully under our control.
1408 : */
1409 36 : while (!list_empty(&tmp)) {
1410 24 : sb = wb_inode(tmp.prev)->i_sb;
1411 74 : list_for_each_prev_safe(pos, node, &tmp) {
1412 50 : inode = wb_inode(pos);
1413 50 : if (inode->i_sb == sb)
1414 34 : list_move(&inode->i_io_list, dispatch_queue);
1415 : }
1416 : }
1417 12 : out:
1418 7303035 : return moved;
1419 : }
1420 :
1421 : /*
1422 : * Queue all expired dirty inodes for io, eldest first.
1423 : * Before
1424 : * newly dirtied b_dirty b_io b_more_io
1425 : * =============> gf edc BA
1426 : * After
1427 : * newly dirtied b_dirty b_io b_more_io
1428 : * =============> g fBAedc
1429 : * |
1430 : * +--> dequeue for IO
1431 : */
1432 3651520 : static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1433 : unsigned long dirtied_before)
1434 : {
1435 3651520 : int moved;
1436 3651520 : unsigned long time_expire_jif = dirtied_before;
1437 :
1438 3651520 : assert_spin_locked(&wb->list_lock);
1439 3651520 : list_splice_init(&wb->b_more_io, &wb->b_io);
1440 3651520 : moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1441 3651514 : if (!work->for_sync)
1442 2298290 : time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1443 3651514 : moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1444 : time_expire_jif);
1445 3651515 : if (moved)
1446 1709906 : wb_io_lists_populated(wb);
1447 3651526 : trace_writeback_queue_io(wb, work, dirtied_before, moved);
1448 3651525 : }
1449 :
1450 50501 : static int write_inode(struct inode *inode, struct writeback_control *wbc)
1451 : {
1452 50501 : int ret;
1453 :
1454 50501 : if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1455 45309 : trace_writeback_write_inode_start(inode, wbc);
1456 45309 : ret = inode->i_sb->s_op->write_inode(inode, wbc);
1457 45309 : trace_writeback_write_inode(inode, wbc);
1458 45309 : return ret;
1459 : }
1460 : return 0;
1461 : }
1462 :
1463 : /*
1464 : * Wait for writeback on an inode to complete. Called with i_lock held.
1465 : * Caller must make sure inode cannot go away when we drop i_lock.
1466 : */
1467 1341415890 : static void __inode_wait_for_writeback(struct inode *inode)
1468 : __releases(inode->i_lock)
1469 : __acquires(inode->i_lock)
1470 : {
1471 1341415890 : DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1472 1341415890 : wait_queue_head_t *wqh;
1473 :
1474 1341415890 : wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1475 1341429766 : while (inode->i_state & I_SYNC) {
1476 108 : spin_unlock(&inode->i_lock);
1477 108 : __wait_on_bit(wqh, &wq, bit_wait,
1478 : TASK_UNINTERRUPTIBLE);
1479 108 : spin_lock(&inode->i_lock);
1480 : }
1481 1341434850 : }
1482 :
1483 : /*
1484 : * Wait for writeback on an inode to complete. Caller must have inode pinned.
1485 : */
1486 1341293487 : void inode_wait_for_writeback(struct inode *inode)
1487 : {
1488 1341293487 : spin_lock(&inode->i_lock);
1489 1341421103 : __inode_wait_for_writeback(inode);
1490 1341389757 : spin_unlock(&inode->i_lock);
1491 1341379869 : }
1492 :
1493 : /*
1494 : * Sleep until I_SYNC is cleared. This function must be called with i_lock
1495 : * held and drops it. It is aimed for callers not holding any inode reference
1496 : * so once i_lock is dropped, inode can go away.
1497 : */
1498 5 : static void inode_sleep_on_writeback(struct inode *inode)
1499 : __releases(inode->i_lock)
1500 : {
1501 5 : DEFINE_WAIT(wait);
1502 5 : wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1503 5 : int sleep;
1504 :
1505 5 : prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1506 5 : sleep = inode->i_state & I_SYNC;
1507 5 : spin_unlock(&inode->i_lock);
1508 5 : if (sleep)
1509 0 : schedule();
1510 5 : finish_wait(wqh, &wait);
1511 5 : }
1512 :
1513 : /*
1514 : * Find proper writeback list for the inode depending on its current state and
1515 : * possibly also change of its state while we were doing writeback. Here we
1516 : * handle things such as livelock prevention or fairness of writeback among
1517 : * inodes. This function can be called only by flusher thread - noone else
1518 : * processes all inodes in writeback lists and requeueing inodes behind flusher
1519 : * thread's back can have unexpected consequences.
1520 : */
1521 19320984 : static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1522 : struct writeback_control *wbc)
1523 : {
1524 19320984 : if (inode->i_state & I_FREEING)
1525 : return;
1526 :
1527 : /*
1528 : * Sync livelock prevention. Each inode is tagged and synced in one
1529 : * shot. If still dirty, it will be redirty_tail()'ed below. Update
1530 : * the dirty time to prevent enqueue and sync it again.
1531 : */
1532 19320876 : if ((inode->i_state & I_DIRTY) &&
1533 507402 : (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1534 231699 : inode->dirtied_when = jiffies;
1535 :
1536 19320876 : if (wbc->pages_skipped) {
1537 : /*
1538 : * writeback is not making progress due to locked
1539 : * buffers. Skip this inode for now.
1540 : */
1541 0 : redirty_tail_locked(inode, wb);
1542 0 : return;
1543 : }
1544 :
1545 19320876 : if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1546 : /*
1547 : * We didn't write back all the pages. nfs_writepages()
1548 : * sometimes bales out without doing anything.
1549 : */
1550 505069 : if (wbc->nr_to_write <= 0) {
1551 : /* Slice used up. Queue for next turn. */
1552 3041 : requeue_io(inode, wb);
1553 : } else {
1554 : /*
1555 : * Writeback blocked by something other than
1556 : * congestion. Delay the inode for some time to
1557 : * avoid spinning on the CPU (100% iowait)
1558 : * retrying writeback of the dirty page/inode
1559 : * that cannot be performed immediately.
1560 : */
1561 502028 : redirty_tail_locked(inode, wb);
1562 : }
1563 18815807 : } else if (inode->i_state & I_DIRTY) {
1564 : /*
1565 : * Filesystems can dirty the inode during writeback operations,
1566 : * such as delayed allocation during submission or metadata
1567 : * updates after data IO completion.
1568 : */
1569 2362 : redirty_tail_locked(inode, wb);
1570 18813445 : } else if (inode->i_state & I_DIRTY_TIME) {
1571 0 : inode->dirtied_when = jiffies;
1572 0 : inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1573 0 : inode->i_state &= ~I_SYNC_QUEUED;
1574 : } else {
1575 : /* The inode is clean. Remove from writeback lists. */
1576 18813445 : inode_cgwb_move_to_attached(inode, wb);
1577 : }
1578 : }
1579 :
1580 : /*
1581 : * Write out an inode and its dirty pages (or some of its dirty pages, depending
1582 : * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1583 : *
1584 : * This doesn't remove the inode from the writeback list it is on, except
1585 : * potentially to move it from b_dirty_time to b_dirty due to timestamp
1586 : * expiration. The caller is otherwise responsible for writeback list handling.
1587 : *
1588 : * The caller is also responsible for setting the I_SYNC flag beforehand and
1589 : * calling inode_sync_complete() to clear it afterwards.
1590 : */
1591 : static int
1592 19334350 : __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1593 : {
1594 19334350 : struct address_space *mapping = inode->i_mapping;
1595 19334350 : long nr_to_write = wbc->nr_to_write;
1596 19334350 : unsigned dirty;
1597 19334350 : int ret;
1598 :
1599 19334350 : WARN_ON(!(inode->i_state & I_SYNC));
1600 :
1601 19334350 : trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1602 :
1603 19334320 : ret = do_writepages(mapping, wbc);
1604 :
1605 : /*
1606 : * Make sure to wait on the data before writing out the metadata.
1607 : * This is important for filesystems that modify metadata on data
1608 : * I/O completion. We don't do it for sync(2) writeback because it has a
1609 : * separate, external IO completion path and ->sync_fs for guaranteeing
1610 : * inode metadata is written back correctly.
1611 : */
1612 19334481 : if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1613 13502 : int err = filemap_fdatawait(mapping);
1614 13502 : if (ret == 0)
1615 13502 : ret = err;
1616 : }
1617 :
1618 : /*
1619 : * If the inode has dirty timestamps and we need to write them, call
1620 : * mark_inode_dirty_sync() to notify the filesystem about it and to
1621 : * change I_DIRTY_TIME into I_DIRTY_SYNC.
1622 : */
1623 19334481 : if ((inode->i_state & I_DIRTY_TIME) &&
1624 12 : (wbc->sync_mode == WB_SYNC_ALL ||
1625 8 : time_after(jiffies, inode->dirtied_time_when +
1626 : dirtytime_expire_interval * HZ))) {
1627 12 : trace_writeback_lazytime(inode);
1628 12 : mark_inode_dirty_sync(inode);
1629 : }
1630 :
1631 : /*
1632 : * Get and clear the dirty flags from i_state. This needs to be done
1633 : * after calling writepages because some filesystems may redirty the
1634 : * inode during writepages due to delalloc. It also needs to be done
1635 : * after handling timestamp expiration, as that may dirty the inode too.
1636 : */
1637 19334481 : spin_lock(&inode->i_lock);
1638 19334481 : dirty = inode->i_state & I_DIRTY;
1639 19334481 : inode->i_state &= ~dirty;
1640 :
1641 : /*
1642 : * Paired with smp_mb() in __mark_inode_dirty(). This allows
1643 : * __mark_inode_dirty() to test i_state without grabbing i_lock -
1644 : * either they see the I_DIRTY bits cleared or we see the dirtied
1645 : * inode.
1646 : *
1647 : * I_DIRTY_PAGES is always cleared together above even if @mapping
1648 : * still has dirty pages. The flag is reinstated after smp_mb() if
1649 : * necessary. This guarantees that either __mark_inode_dirty()
1650 : * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1651 : */
1652 19334481 : smp_mb();
1653 :
1654 19334483 : if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1655 504765 : inode->i_state |= I_DIRTY_PAGES;
1656 18829718 : else if (unlikely(inode->i_state & I_PINNING_FSCACHE_WB)) {
1657 0 : if (!(inode->i_state & I_DIRTY_PAGES)) {
1658 0 : inode->i_state &= ~I_PINNING_FSCACHE_WB;
1659 0 : wbc->unpinned_fscache_wb = true;
1660 0 : dirty |= I_PINNING_FSCACHE_WB; /* Cause write_inode */
1661 : }
1662 : }
1663 :
1664 19334483 : spin_unlock(&inode->i_lock);
1665 :
1666 : /* Don't write the inode if only I_DIRTY_PAGES was set */
1667 19334478 : if (dirty & ~I_DIRTY_PAGES) {
1668 50501 : int err = write_inode(inode, wbc);
1669 50501 : if (ret == 0)
1670 50501 : ret = err;
1671 : }
1672 19334478 : wbc->unpinned_fscache_wb = false;
1673 19334478 : trace_writeback_single_inode(inode, wbc, nr_to_write);
1674 19334477 : return ret;
1675 : }
1676 :
1677 : /*
1678 : * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1679 : * the regular batched writeback done by the flusher threads in
1680 : * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1681 : * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1682 : *
1683 : * To prevent the inode from going away, either the caller must have a reference
1684 : * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1685 : */
1686 516908762 : static int writeback_single_inode(struct inode *inode,
1687 : struct writeback_control *wbc)
1688 : {
1689 516908762 : struct bdi_writeback *wb;
1690 516908762 : int ret = 0;
1691 :
1692 516908762 : spin_lock(&inode->i_lock);
1693 516928973 : if (!atomic_read(&inode->i_count))
1694 516912082 : WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1695 : else
1696 17842 : WARN_ON(inode->i_state & I_WILL_FREE);
1697 :
1698 516928973 : if (inode->i_state & I_SYNC) {
1699 : /*
1700 : * Writeback is already running on the inode. For WB_SYNC_NONE,
1701 : * that's enough and we can just return. For WB_SYNC_ALL, we
1702 : * must wait for the existing writeback to complete, then do
1703 : * writeback again if there's anything left.
1704 : */
1705 0 : if (wbc->sync_mode != WB_SYNC_ALL)
1706 0 : goto out;
1707 0 : __inode_wait_for_writeback(inode);
1708 : }
1709 516928973 : WARN_ON(inode->i_state & I_SYNC);
1710 : /*
1711 : * If the inode is already fully clean, then there's nothing to do.
1712 : *
1713 : * For data-integrity syncs we also need to check whether any pages are
1714 : * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1715 : * there are any such pages, we'll need to wait for them.
1716 : */
1717 516928973 : if (!(inode->i_state & I_DIRTY_ALL) &&
1718 516915471 : (wbc->sync_mode != WB_SYNC_ALL ||
1719 516915477 : !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1720 516916890 : goto out;
1721 12979 : inode->i_state |= I_SYNC;
1722 12979 : wbc_attach_and_unlock_inode(wbc, inode);
1723 :
1724 13502 : ret = __writeback_single_inode(inode, wbc);
1725 :
1726 13502 : wbc_detach_inode(wbc);
1727 :
1728 13502 : wb = inode_to_wb_and_lock_list(inode);
1729 13502 : spin_lock(&inode->i_lock);
1730 : /*
1731 : * If the inode is freeing, its i_io_list shoudn't be updated
1732 : * as it can be finally deleted at this moment.
1733 : */
1734 13502 : if (!(inode->i_state & I_FREEING)) {
1735 : /*
1736 : * If the inode is now fully clean, then it can be safely
1737 : * removed from its writeback list (if any). Otherwise the
1738 : * flusher threads are responsible for the writeback lists.
1739 : */
1740 13502 : if (!(inode->i_state & I_DIRTY_ALL))
1741 13502 : inode_cgwb_move_to_attached(inode, wb);
1742 0 : else if (!(inode->i_state & I_SYNC_QUEUED)) {
1743 0 : if ((inode->i_state & I_DIRTY))
1744 0 : redirty_tail_locked(inode, wb);
1745 0 : else if (inode->i_state & I_DIRTY_TIME) {
1746 0 : inode->dirtied_when = jiffies;
1747 0 : inode_io_list_move_locked(inode,
1748 : wb,
1749 : &wb->b_dirty_time);
1750 : }
1751 : }
1752 : }
1753 :
1754 13502 : spin_unlock(&wb->list_lock);
1755 13502 : inode_sync_complete(inode);
1756 516930392 : out:
1757 516930392 : spin_unlock(&inode->i_lock);
1758 516926496 : return ret;
1759 : }
1760 :
1761 19320929 : static long writeback_chunk_size(struct bdi_writeback *wb,
1762 : struct wb_writeback_work *work)
1763 : {
1764 19320929 : long pages;
1765 :
1766 : /*
1767 : * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1768 : * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1769 : * here avoids calling into writeback_inodes_wb() more than once.
1770 : *
1771 : * The intended call sequence for WB_SYNC_ALL writeback is:
1772 : *
1773 : * wb_writeback()
1774 : * writeback_sb_inodes() <== called only once
1775 : * write_cache_pages() <== called once for each inode
1776 : * (quickly) tag currently dirty pages
1777 : * (maybe slowly) sync all tagged pages
1778 : */
1779 19320929 : if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1780 : pages = LONG_MAX;
1781 : else {
1782 17409993 : pages = min(wb->avg_write_bandwidth / 2,
1783 : global_wb_domain.dirty_limit / DIRTY_SCOPE);
1784 17409993 : pages = min(pages, work->nr_pages);
1785 17409993 : pages = round_down(pages + MIN_WRITEBACK_PAGES,
1786 : MIN_WRITEBACK_PAGES);
1787 : }
1788 :
1789 19320929 : return pages;
1790 : }
1791 :
1792 : /*
1793 : * Write a portion of b_io inodes which belong to @sb.
1794 : *
1795 : * Return the number of pages and/or inodes written.
1796 : *
1797 : * NOTE! This is called with wb->list_lock held, and will
1798 : * unlock and relock that for each inode it ends up doing
1799 : * IO for.
1800 : */
1801 2632405 : static long writeback_sb_inodes(struct super_block *sb,
1802 : struct bdi_writeback *wb,
1803 : struct wb_writeback_work *work)
1804 : {
1805 2632405 : struct writeback_control wbc = {
1806 2632405 : .sync_mode = work->sync_mode,
1807 2632405 : .tagged_writepages = work->tagged_writepages,
1808 2632405 : .for_kupdate = work->for_kupdate,
1809 2632405 : .for_background = work->for_background,
1810 2632405 : .for_sync = work->for_sync,
1811 2632405 : .range_cyclic = work->range_cyclic,
1812 : .range_start = 0,
1813 : .range_end = LLONG_MAX,
1814 : };
1815 2632405 : unsigned long start_time = jiffies;
1816 2632405 : long write_chunk;
1817 2632405 : long total_wrote = 0; /* count both pages and inodes */
1818 :
1819 24497020 : while (!list_empty(&wb->b_io)) {
1820 21876250 : struct inode *inode = wb_inode(wb->b_io.prev);
1821 21876250 : struct bdi_writeback *tmp_wb;
1822 21876250 : long wrote;
1823 :
1824 21876250 : if (inode->i_sb != sb) {
1825 2555304 : if (work->sb) {
1826 : /*
1827 : * We only want to write back data for this
1828 : * superblock, move all inodes not belonging
1829 : * to it back onto the dirty list.
1830 : */
1831 2555270 : redirty_tail(inode, wb);
1832 2555270 : continue;
1833 : }
1834 :
1835 : /*
1836 : * The inode belongs to a different superblock.
1837 : * Bounce back to the caller to unpin this and
1838 : * pin the next superblock.
1839 : */
1840 : break;
1841 : }
1842 :
1843 : /*
1844 : * Don't bother with new inodes or inodes being freed, first
1845 : * kind does not need periodic writeout yet, and for the latter
1846 : * kind writeout is handled by the freer.
1847 : */
1848 19320946 : spin_lock(&inode->i_lock);
1849 19320970 : if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1850 8 : redirty_tail_locked(inode, wb);
1851 8 : spin_unlock(&inode->i_lock);
1852 8 : continue;
1853 : }
1854 19320962 : if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1855 : /*
1856 : * If this inode is locked for writeback and we are not
1857 : * doing writeback-for-data-integrity, move it to
1858 : * b_more_io so that writeback can proceed with the
1859 : * other inodes on s_io.
1860 : *
1861 : * We'll have another go at writing back this inode
1862 : * when we completed a full scan of b_io.
1863 : */
1864 0 : requeue_io(inode, wb);
1865 0 : spin_unlock(&inode->i_lock);
1866 0 : trace_writeback_sb_inodes_requeue(inode);
1867 0 : continue;
1868 : }
1869 19320962 : spin_unlock(&wb->list_lock);
1870 :
1871 : /*
1872 : * We already requeued the inode if it had I_SYNC set and we
1873 : * are doing WB_SYNC_NONE writeback. So this catches only the
1874 : * WB_SYNC_ALL case.
1875 : */
1876 19320904 : if (inode->i_state & I_SYNC) {
1877 : /* Wait for I_SYNC. This function drops i_lock... */
1878 0 : inode_sleep_on_writeback(inode);
1879 : /* Inode may be gone, start again */
1880 0 : spin_lock(&wb->list_lock);
1881 0 : continue;
1882 : }
1883 19320904 : inode->i_state |= I_SYNC;
1884 19320904 : wbc_attach_and_unlock_inode(&wbc, inode);
1885 :
1886 19320878 : write_chunk = writeback_chunk_size(wb, work);
1887 19320878 : wbc.nr_to_write = write_chunk;
1888 19320878 : wbc.pages_skipped = 0;
1889 :
1890 : /*
1891 : * We use I_SYNC to pin the inode in memory. While it is set
1892 : * evict_inode() will wait so the inode cannot be freed.
1893 : */
1894 19320878 : __writeback_single_inode(inode, &wbc);
1895 :
1896 19320978 : wbc_detach_inode(&wbc);
1897 19320979 : work->nr_pages -= write_chunk - wbc.nr_to_write;
1898 19320979 : wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
1899 19320979 : wrote = wrote < 0 ? 0 : wrote;
1900 19320979 : total_wrote += wrote;
1901 :
1902 19320979 : if (need_resched()) {
1903 : /*
1904 : * We're trying to balance between building up a nice
1905 : * long list of IOs to improve our merge rate, and
1906 : * getting those IOs out quickly for anyone throttling
1907 : * in balance_dirty_pages(). cond_resched() doesn't
1908 : * unplug, so get our IOs out the door before we
1909 : * give up the CPU.
1910 : */
1911 2401 : blk_flush_plug(current->plug, false);
1912 2401 : cond_resched();
1913 : }
1914 :
1915 : /*
1916 : * Requeue @inode if still dirty. Be careful as @inode may
1917 : * have been switched to another wb in the meantime.
1918 : */
1919 19320979 : tmp_wb = inode_to_wb_and_lock_list(inode);
1920 19320977 : spin_lock(&inode->i_lock);
1921 19320981 : if (!(inode->i_state & I_DIRTY_ALL))
1922 18813565 : total_wrote++;
1923 19320981 : requeue_inode(inode, tmp_wb, &wbc);
1924 19320973 : inode_sync_complete(inode);
1925 19320960 : spin_unlock(&inode->i_lock);
1926 :
1927 19320966 : if (unlikely(tmp_wb != wb)) {
1928 0 : spin_unlock(&tmp_wb->list_lock);
1929 0 : spin_lock(&wb->list_lock);
1930 : }
1931 :
1932 : /*
1933 : * bail out to wb_writeback() often enough to check
1934 : * background threshold and other termination conditions.
1935 : */
1936 19320966 : if (total_wrote) {
1937 19150005 : if (time_is_before_jiffies(start_time + HZ / 10UL))
1938 : break;
1939 19138376 : if (work->nr_pages <= 0)
1940 : break;
1941 : }
1942 : }
1943 2632433 : return total_wrote;
1944 : }
1945 :
1946 2201191 : static long __writeback_inodes_wb(struct bdi_writeback *wb,
1947 : struct wb_writeback_work *work)
1948 : {
1949 2201191 : unsigned long start_time = jiffies;
1950 2201191 : long wrote = 0;
1951 :
1952 3458817 : while (!list_empty(&wb->b_io)) {
1953 1267069 : struct inode *inode = wb_inode(wb->b_io.prev);
1954 1267069 : struct super_block *sb = inode->i_sb;
1955 :
1956 1267069 : if (!trylock_super(sb)) {
1957 : /*
1958 : * trylock_super() may fail consistently due to
1959 : * s_umount being grabbed by someone else. Don't use
1960 : * requeue_io() to avoid busy retrying the inode/sb.
1961 : */
1962 93486 : redirty_tail(inode, wb);
1963 93486 : continue;
1964 : }
1965 1173607 : wrote += writeback_sb_inodes(sb, wb, work);
1966 1173612 : up_read(&sb->s_umount);
1967 :
1968 : /* refer to the same tests at the end of writeback_sb_inodes */
1969 1173612 : if (wrote) {
1970 1167133 : if (time_is_before_jiffies(start_time + HZ / 10UL))
1971 : break;
1972 1157661 : if (work->nr_pages <= 0)
1973 : break;
1974 : }
1975 : }
1976 : /* Leave any unwritten inodes on b_io */
1977 2201220 : return wrote;
1978 : }
1979 :
1980 0 : static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1981 : enum wb_reason reason)
1982 : {
1983 0 : struct wb_writeback_work work = {
1984 : .nr_pages = nr_pages,
1985 : .sync_mode = WB_SYNC_NONE,
1986 : .range_cyclic = 1,
1987 : .reason = reason,
1988 : };
1989 0 : struct blk_plug plug;
1990 :
1991 0 : blk_start_plug(&plug);
1992 0 : spin_lock(&wb->list_lock);
1993 0 : if (list_empty(&wb->b_io))
1994 0 : queue_io(wb, &work, jiffies);
1995 0 : __writeback_inodes_wb(wb, &work);
1996 0 : spin_unlock(&wb->list_lock);
1997 0 : blk_finish_plug(&plug);
1998 :
1999 0 : return nr_pages - work.nr_pages;
2000 : }
2001 :
2002 : /*
2003 : * Explicit flushing or periodic writeback of "old" data.
2004 : *
2005 : * Define "old": the first time one of an inode's pages is dirtied, we mark the
2006 : * dirtying-time in the inode's address_space. So this periodic writeback code
2007 : * just walks the superblock inode list, writing back any inodes which are
2008 : * older than a specific point in time.
2009 : *
2010 : * Try to run once per dirty_writeback_interval. But if a writeback event
2011 : * takes longer than a dirty_writeback_interval interval, then leave a
2012 : * one-second gap.
2013 : *
2014 : * dirtied_before takes precedence over nr_to_write. So we'll only write back
2015 : * all dirty pages if they are all attached to "old" mappings.
2016 : */
2017 2136107 : static long wb_writeback(struct bdi_writeback *wb,
2018 : struct wb_writeback_work *work)
2019 : {
2020 2136107 : long nr_pages = work->nr_pages;
2021 2136107 : unsigned long dirtied_before = jiffies;
2022 2136107 : struct inode *inode;
2023 2136107 : long progress;
2024 2136107 : struct blk_plug plug;
2025 :
2026 2136107 : blk_start_plug(&plug);
2027 3667473 : for (;;) {
2028 : /*
2029 : * Stop writeback when nr_pages has been consumed
2030 : */
2031 3667473 : if (work->nr_pages <= 0)
2032 : break;
2033 :
2034 : /*
2035 : * Background writeout and kupdate-style writeback may
2036 : * run forever. Stop them if there is other work to do
2037 : * so that e.g. sync can proceed. They'll be restarted
2038 : * after the other works are all done.
2039 : */
2040 3667467 : if ((work->for_background || work->for_kupdate) &&
2041 39602 : !list_empty(&wb->work_list))
2042 : break;
2043 :
2044 : /*
2045 : * For background writeout, stop when we are below the
2046 : * background dirty threshold
2047 : */
2048 3663167 : if (work->for_background && !wb_over_bg_thresh(wb))
2049 : break;
2050 :
2051 :
2052 3660004 : spin_lock(&wb->list_lock);
2053 :
2054 : /*
2055 : * Kupdate and background works are special and we want to
2056 : * include all inodes that need writing. Livelock avoidance is
2057 : * handled by these works yielding to any other work so we are
2058 : * safe.
2059 : */
2060 3660007 : if (work->for_kupdate) {
2061 24099 : dirtied_before = jiffies -
2062 24099 : msecs_to_jiffies(dirty_expire_interval * 10);
2063 3635908 : } else if (work->for_background)
2064 8040 : dirtied_before = jiffies;
2065 :
2066 3660007 : trace_writeback_start(wb, work);
2067 3659997 : if (list_empty(&wb->b_io))
2068 3651511 : queue_io(wb, work, dirtied_before);
2069 3660007 : if (work->sb)
2070 1458815 : progress = writeback_sb_inodes(work->sb, wb, work);
2071 : else
2072 2201192 : progress = __writeback_inodes_wb(wb, work);
2073 3660042 : trace_writeback_written(wb, work);
2074 :
2075 : /*
2076 : * Did we write something? Try for more
2077 : *
2078 : * Dirty inodes are moved to b_io for writeback in batches.
2079 : * The completion of the current batch does not necessarily
2080 : * mean the overall work is done. So we keep looping as long
2081 : * as made some progress on cleaning pages or inodes.
2082 : */
2083 3660042 : if (progress) {
2084 1531356 : spin_unlock(&wb->list_lock);
2085 1531357 : continue;
2086 : }
2087 :
2088 : /*
2089 : * No more inodes for IO, bail
2090 : */
2091 2128686 : if (list_empty(&wb->b_more_io)) {
2092 2128681 : spin_unlock(&wb->list_lock);
2093 : break;
2094 : }
2095 :
2096 : /*
2097 : * Nothing written. Wait for some inode to
2098 : * become available for writeback. Otherwise
2099 : * we'll just busyloop.
2100 : */
2101 5 : trace_writeback_wait(wb, work);
2102 5 : inode = wb_inode(wb->b_more_io.prev);
2103 5 : spin_lock(&inode->i_lock);
2104 5 : spin_unlock(&wb->list_lock);
2105 : /* This function drops i_lock... */
2106 5 : inode_sleep_on_writeback(inode);
2107 : }
2108 2136142 : blk_finish_plug(&plug);
2109 :
2110 2136136 : return nr_pages - work->nr_pages;
2111 : }
2112 :
2113 : /*
2114 : * Return the next wb_writeback_work struct that hasn't been processed yet.
2115 : */
2116 3588136 : static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2117 : {
2118 3588136 : struct wb_writeback_work *work = NULL;
2119 :
2120 3588136 : spin_lock_irq(&wb->work_lock);
2121 3588245 : if (!list_empty(&wb->work_list)) {
2122 1094552 : work = list_entry(wb->work_list.next,
2123 : struct wb_writeback_work, list);
2124 1094552 : list_del_init(&work->list);
2125 : }
2126 3588232 : spin_unlock_irq(&wb->work_lock);
2127 3588092 : return work;
2128 : }
2129 :
2130 2493696 : static long wb_check_background_flush(struct bdi_writeback *wb)
2131 : {
2132 2493696 : if (wb_over_bg_thresh(wb)) {
2133 :
2134 3460 : struct wb_writeback_work work = {
2135 : .nr_pages = LONG_MAX,
2136 : .sync_mode = WB_SYNC_NONE,
2137 : .for_background = 1,
2138 : .range_cyclic = 1,
2139 : .reason = WB_REASON_BACKGROUND,
2140 : };
2141 :
2142 3460 : return wb_writeback(wb, &work);
2143 : }
2144 :
2145 : return 0;
2146 : }
2147 :
2148 2493698 : static long wb_check_old_data_flush(struct bdi_writeback *wb)
2149 : {
2150 2493698 : unsigned long expired;
2151 2493698 : long nr_pages;
2152 :
2153 : /*
2154 : * When set to zero, disable periodic writeback
2155 : */
2156 2493698 : if (!dirty_writeback_interval)
2157 : return 0;
2158 :
2159 4987392 : expired = wb->last_old_flush +
2160 2493698 : msecs_to_jiffies(dirty_writeback_interval * 10);
2161 2493694 : if (time_before(jiffies, expired))
2162 : return 0;
2163 :
2164 23087 : wb->last_old_flush = jiffies;
2165 23087 : nr_pages = get_nr_dirty_pages();
2166 :
2167 23087 : if (nr_pages) {
2168 23087 : struct wb_writeback_work work = {
2169 : .nr_pages = nr_pages,
2170 : .sync_mode = WB_SYNC_NONE,
2171 : .for_kupdate = 1,
2172 : .range_cyclic = 1,
2173 : .reason = WB_REASON_PERIODIC,
2174 : };
2175 :
2176 23087 : return wb_writeback(wb, &work);
2177 : }
2178 :
2179 : return 0;
2180 : }
2181 :
2182 2493687 : static long wb_check_start_all(struct bdi_writeback *wb)
2183 : {
2184 2493687 : long nr_pages;
2185 :
2186 2493687 : if (!test_bit(WB_start_all, &wb->state))
2187 : return 0;
2188 :
2189 1015034 : nr_pages = get_nr_dirty_pages();
2190 1015034 : if (nr_pages) {
2191 2030060 : struct wb_writeback_work work = {
2192 1015032 : .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2193 : .sync_mode = WB_SYNC_NONE,
2194 : .range_cyclic = 1,
2195 1015028 : .reason = wb->start_all_reason,
2196 : };
2197 :
2198 1015028 : nr_pages = wb_writeback(wb, &work);
2199 : }
2200 :
2201 1015037 : clear_bit(WB_start_all, &wb->state);
2202 : return nr_pages;
2203 : }
2204 :
2205 :
2206 : /*
2207 : * Retrieve work items and do the writeback they describe
2208 : */
2209 2493654 : static long wb_do_writeback(struct bdi_writeback *wb)
2210 : {
2211 2493654 : struct wb_writeback_work *work;
2212 2493654 : long wrote = 0;
2213 :
2214 2493654 : set_bit(WB_writeback_running, &wb->state);
2215 3588232 : while ((work = get_next_work_item(wb)) != NULL) {
2216 1094557 : trace_writeback_exec(wb, work);
2217 1094552 : wrote += wb_writeback(wb, work);
2218 1094556 : finish_writeback_work(wb, work);
2219 : }
2220 :
2221 : /*
2222 : * Check for a flush-everything request
2223 : */
2224 2493686 : wrote += wb_check_start_all(wb);
2225 :
2226 : /*
2227 : * Check for periodic writeback, kupdated() style
2228 : */
2229 2493693 : wrote += wb_check_old_data_flush(wb);
2230 2493694 : wrote += wb_check_background_flush(wb);
2231 2493692 : clear_bit(WB_writeback_running, &wb->state);
2232 :
2233 2493695 : return wrote;
2234 : }
2235 :
2236 : /*
2237 : * Handle writeback of dirty data for the device backed by this bdi. Also
2238 : * reschedules periodically and does kupdated style flushing.
2239 : */
2240 2047639 : void wb_workfn(struct work_struct *work)
2241 : {
2242 2047639 : struct bdi_writeback *wb = container_of(to_delayed_work(work),
2243 : struct bdi_writeback, dwork);
2244 2047639 : long pages_written;
2245 :
2246 2047639 : set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2247 :
2248 2047578 : if (likely(!current_is_workqueue_rescuer() ||
2249 : !test_bit(WB_registered, &wb->state))) {
2250 : /*
2251 : * The normal path. Keep writing back @wb until its
2252 : * work_list is empty. Note that this path is also taken
2253 : * if @wb is shutting down even when we're running off the
2254 : * rescuer as work_list needs to be drained.
2255 : */
2256 2493654 : do {
2257 2493654 : pages_written = wb_do_writeback(wb);
2258 2493692 : trace_writeback_pages_written(pages_written);
2259 2493690 : } while (!list_empty(&wb->work_list));
2260 : } else {
2261 : /*
2262 : * bdi_wq can't get enough workers and we're running off
2263 : * the emergency worker. Don't hog it. Hopefully, 1024 is
2264 : * enough for efficient IO.
2265 : */
2266 0 : pages_written = writeback_inodes_wb(wb, 1024,
2267 : WB_REASON_FORKER_THREAD);
2268 0 : trace_writeback_pages_written(pages_written);
2269 : }
2270 :
2271 2047634 : if (!list_empty(&wb->work_list))
2272 0 : wb_wakeup(wb);
2273 4095268 : else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2274 509766 : wb_wakeup_delayed(wb);
2275 2047618 : }
2276 :
2277 : /*
2278 : * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2279 : * write back the whole world.
2280 : */
2281 14221442 : static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2282 : enum wb_reason reason)
2283 : {
2284 14221442 : struct bdi_writeback *wb;
2285 :
2286 14221442 : if (!bdi_has_dirty_io(bdi))
2287 : return;
2288 :
2289 4819212 : list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2290 3823318 : wb_start_writeback(wb, reason);
2291 : }
2292 :
2293 0 : void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2294 : enum wb_reason reason)
2295 : {
2296 0 : rcu_read_lock();
2297 0 : __wakeup_flusher_threads_bdi(bdi, reason);
2298 0 : rcu_read_unlock();
2299 0 : }
2300 :
2301 : /*
2302 : * Wakeup the flusher threads to start writeback of all currently dirty pages
2303 : */
2304 1010066 : void wakeup_flusher_threads(enum wb_reason reason)
2305 : {
2306 1010066 : struct backing_dev_info *bdi;
2307 :
2308 : /*
2309 : * If we are expecting writeback progress we must submit plugged IO.
2310 : */
2311 1010066 : blk_flush_plug(current->plug, true);
2312 :
2313 1010066 : rcu_read_lock();
2314 15231468 : list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2315 14221397 : __wakeup_flusher_threads_bdi(bdi, reason);
2316 1010071 : rcu_read_unlock();
2317 1010071 : }
2318 :
2319 : /*
2320 : * Wake up bdi's periodically to make sure dirtytime inodes gets
2321 : * written back periodically. We deliberately do *not* check the
2322 : * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2323 : * kernel to be constantly waking up once there are any dirtytime
2324 : * inodes on the system. So instead we define a separate delayed work
2325 : * function which gets called much more rarely. (By default, only
2326 : * once every 12 hours.)
2327 : *
2328 : * If there is any other write activity going on in the file system,
2329 : * this function won't be necessary. But if the only thing that has
2330 : * happened on the file system is a dirtytime inode caused by an atime
2331 : * update, we need this infrastructure below to make sure that inode
2332 : * eventually gets pushed out to disk.
2333 : */
2334 : static void wakeup_dirtytime_writeback(struct work_struct *w);
2335 : static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2336 :
2337 33 : static void wakeup_dirtytime_writeback(struct work_struct *w)
2338 : {
2339 33 : struct backing_dev_info *bdi;
2340 :
2341 33 : rcu_read_lock();
2342 495 : list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2343 462 : struct bdi_writeback *wb;
2344 :
2345 1150 : list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2346 688 : if (!list_empty(&wb->b_dirty_time))
2347 6 : wb_wakeup(wb);
2348 : }
2349 33 : rcu_read_unlock();
2350 33 : schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2351 33 : }
2352 :
2353 0 : static int __init start_dirtytime_writeback(void)
2354 : {
2355 0 : schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2356 0 : return 0;
2357 : }
2358 : __initcall(start_dirtytime_writeback);
2359 :
2360 30 : int dirtytime_interval_handler(struct ctl_table *table, int write,
2361 : void *buffer, size_t *lenp, loff_t *ppos)
2362 : {
2363 30 : int ret;
2364 :
2365 30 : ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2366 30 : if (ret == 0 && write)
2367 26 : mod_delayed_work(system_wq, &dirtytime_work, 0);
2368 30 : return ret;
2369 : }
2370 :
2371 : /**
2372 : * __mark_inode_dirty - internal function to mark an inode dirty
2373 : *
2374 : * @inode: inode to mark
2375 : * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2376 : * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2377 : * with I_DIRTY_PAGES.
2378 : *
2379 : * Mark an inode as dirty. We notify the filesystem, then update the inode's
2380 : * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2381 : *
2382 : * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2383 : * instead of calling this directly.
2384 : *
2385 : * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2386 : * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2387 : * even if they are later hashed, as they will have been marked dirty already.
2388 : *
2389 : * In short, ensure you hash any inodes _before_ you start marking them dirty.
2390 : *
2391 : * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2392 : * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2393 : * the kernel-internal blockdev inode represents the dirtying time of the
2394 : * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2395 : * page->mapping->host, so the page-dirtying time is recorded in the internal
2396 : * blockdev inode.
2397 : */
2398 59443251 : void __mark_inode_dirty(struct inode *inode, int flags)
2399 : {
2400 59443251 : struct super_block *sb = inode->i_sb;
2401 59443251 : int dirtytime = 0;
2402 59443251 : struct bdi_writeback *wb = NULL;
2403 :
2404 59443251 : trace_writeback_mark_inode_dirty(inode, flags);
2405 :
2406 59441470 : if (flags & I_DIRTY_INODE) {
2407 : /*
2408 : * Inode timestamp update will piggback on this dirtying.
2409 : * We tell ->dirty_inode callback that timestamps need to
2410 : * be updated by setting I_DIRTY_TIME in flags.
2411 : */
2412 11817297 : if (inode->i_state & I_DIRTY_TIME) {
2413 20 : spin_lock(&inode->i_lock);
2414 20 : if (inode->i_state & I_DIRTY_TIME) {
2415 20 : inode->i_state &= ~I_DIRTY_TIME;
2416 20 : flags |= I_DIRTY_TIME;
2417 : }
2418 20 : spin_unlock(&inode->i_lock);
2419 : }
2420 :
2421 : /*
2422 : * Notify the filesystem about the inode being dirtied, so that
2423 : * (if needed) it can update on-disk fields and journal the
2424 : * inode. This is only needed when the inode itself is being
2425 : * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2426 : * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2427 : */
2428 11817297 : trace_writeback_dirty_inode_start(inode, flags);
2429 11819704 : if (sb->s_op->dirty_inode)
2430 977061 : sb->s_op->dirty_inode(inode,
2431 : flags & (I_DIRTY_INODE | I_DIRTY_TIME));
2432 11819644 : trace_writeback_dirty_inode(inode, flags);
2433 :
2434 : /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2435 11819861 : flags &= ~I_DIRTY_TIME;
2436 : } else {
2437 : /*
2438 : * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2439 : * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2440 : * in one call to __mark_inode_dirty().)
2441 : */
2442 47624173 : dirtytime = flags & I_DIRTY_TIME;
2443 47624173 : WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2444 : }
2445 :
2446 : /*
2447 : * Paired with smp_mb() in __writeback_single_inode() for the
2448 : * following lockless i_state test. See there for details.
2449 : */
2450 59444034 : smp_mb();
2451 :
2452 59448421 : if ((inode->i_state & flags) == flags)
2453 : return;
2454 :
2455 20590267 : spin_lock(&inode->i_lock);
2456 20588757 : if ((inode->i_state & flags) != flags) {
2457 20586037 : const int was_dirty = inode->i_state & I_DIRTY;
2458 :
2459 20586037 : inode_attach_wb(inode, NULL);
2460 :
2461 20585962 : inode->i_state |= flags;
2462 :
2463 : /*
2464 : * Grab inode's wb early because it requires dropping i_lock and we
2465 : * need to make sure following checks happen atomically with dirty
2466 : * list handling so that we don't move inodes under flush worker's
2467 : * hands.
2468 : */
2469 20585962 : if (!was_dirty) {
2470 20569681 : wb = locked_inode_to_wb_and_lock_list(inode);
2471 20572815 : spin_lock(&inode->i_lock);
2472 : }
2473 :
2474 : /*
2475 : * If the inode is queued for writeback by flush worker, just
2476 : * update its dirty state. Once the flush worker is done with
2477 : * the inode it will place it on the appropriate superblock
2478 : * list, based upon its state.
2479 : */
2480 20589099 : if (inode->i_state & I_SYNC_QUEUED)
2481 2797 : goto out_unlock;
2482 :
2483 : /*
2484 : * Only add valid (hashed) inodes to the superblock's
2485 : * dirty list. Add blockdev inodes as well.
2486 : */
2487 20586302 : if (!S_ISBLK(inode->i_mode)) {
2488 20573085 : if (inode_unhashed(inode))
2489 420530 : goto out_unlock;
2490 : }
2491 20165772 : if (inode->i_state & I_FREEING)
2492 0 : goto out_unlock;
2493 :
2494 : /*
2495 : * If the inode was already on b_dirty/b_io/b_more_io, don't
2496 : * reposition it (that would break b_dirty time-ordering).
2497 : */
2498 20165772 : if (!was_dirty) {
2499 20149657 : struct list_head *dirty_list;
2500 20149657 : bool wakeup_bdi = false;
2501 :
2502 20149657 : inode->dirtied_when = jiffies;
2503 20149657 : if (dirtytime)
2504 12 : inode->dirtied_time_when = jiffies;
2505 :
2506 20149657 : if (inode->i_state & I_DIRTY)
2507 20149645 : dirty_list = &wb->b_dirty;
2508 : else
2509 12 : dirty_list = &wb->b_dirty_time;
2510 :
2511 20149657 : wakeup_bdi = inode_io_list_move_locked(inode, wb,
2512 : dirty_list);
2513 :
2514 20149652 : spin_unlock(&wb->list_lock);
2515 20149656 : spin_unlock(&inode->i_lock);
2516 20148623 : trace_writeback_dirty_inode_enqueue(inode);
2517 :
2518 : /*
2519 : * If this is the first dirty inode for this bdi,
2520 : * we have to wake-up the corresponding bdi thread
2521 : * to make sure background write-back happens
2522 : * later.
2523 : */
2524 20149557 : if (wakeup_bdi &&
2525 1205907 : (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2526 1205907 : wb_wakeup_delayed(wb);
2527 20149558 : return;
2528 : }
2529 : }
2530 18835 : out_unlock:
2531 439442 : if (wb)
2532 423161 : spin_unlock(&wb->list_lock);
2533 442162 : spin_unlock(&inode->i_lock);
2534 : }
2535 : EXPORT_SYMBOL(__mark_inode_dirty);
2536 :
2537 : /*
2538 : * The @s_sync_lock is used to serialise concurrent sync operations
2539 : * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2540 : * Concurrent callers will block on the s_sync_lock rather than doing contending
2541 : * walks. The queueing maintains sync(2) required behaviour as all the IO that
2542 : * has been issued up to the time this function is enter is guaranteed to be
2543 : * completed by the time we have gained the lock and waited for all IO that is
2544 : * in progress regardless of the order callers are granted the lock.
2545 : */
2546 6491838 : static void wait_sb_inodes(struct super_block *sb)
2547 : {
2548 6491838 : LIST_HEAD(sync_list);
2549 :
2550 : /*
2551 : * We need to be protected against the filesystem going from
2552 : * r/o to r/w or vice versa.
2553 : */
2554 6491838 : WARN_ON(!rwsem_is_locked(&sb->s_umount));
2555 :
2556 6491838 : mutex_lock(&sb->s_sync_lock);
2557 :
2558 : /*
2559 : * Splice the writeback list onto a temporary list to avoid waiting on
2560 : * inodes that have started writeback after this point.
2561 : *
2562 : * Use rcu_read_lock() to keep the inodes around until we have a
2563 : * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2564 : * the local list because inodes can be dropped from either by writeback
2565 : * completion.
2566 : */
2567 6491880 : rcu_read_lock();
2568 6491880 : spin_lock_irq(&sb->s_inode_wblist_lock);
2569 6491876 : list_splice_init(&sb->s_inodes_wb, &sync_list);
2570 :
2571 : /*
2572 : * Data integrity sync. Must wait for all pages under writeback, because
2573 : * there may have been pages dirtied before our sync call, but which had
2574 : * writeout started before we write it out. In which case, the inode
2575 : * may not be on the dirty list, but we still have to wait for that
2576 : * writeout.
2577 : */
2578 9503584 : while (!list_empty(&sync_list)) {
2579 3011704 : struct inode *inode = list_first_entry(&sync_list, struct inode,
2580 : i_wb_list);
2581 3011704 : struct address_space *mapping = inode->i_mapping;
2582 :
2583 : /*
2584 : * Move each inode back to the wb list before we drop the lock
2585 : * to preserve consistency between i_wb_list and the mapping
2586 : * writeback tag. Writeback completion is responsible to remove
2587 : * the inode from either list once the writeback tag is cleared.
2588 : */
2589 3011704 : list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2590 :
2591 : /*
2592 : * The mapping can appear untagged while still on-list since we
2593 : * do not have the mapping lock. Skip it here, wb completion
2594 : * will remove it.
2595 : */
2596 3011708 : if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2597 2558 : continue;
2598 :
2599 3009150 : spin_unlock_irq(&sb->s_inode_wblist_lock);
2600 :
2601 3009150 : spin_lock(&inode->i_lock);
2602 3009150 : if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2603 856 : spin_unlock(&inode->i_lock);
2604 :
2605 856 : spin_lock_irq(&sb->s_inode_wblist_lock);
2606 856 : continue;
2607 : }
2608 3008294 : __iget(inode);
2609 3008294 : spin_unlock(&inode->i_lock);
2610 3008294 : rcu_read_unlock();
2611 :
2612 : /*
2613 : * We keep the error status of individual mapping so that
2614 : * applications can catch the writeback error using fsync(2).
2615 : * See filemap_fdatawait_keep_errors() for details.
2616 : */
2617 3008294 : filemap_fdatawait_keep_errors(mapping);
2618 :
2619 3008294 : cond_resched();
2620 :
2621 3008294 : iput(inode);
2622 :
2623 3008294 : rcu_read_lock();
2624 3008294 : spin_lock_irq(&sb->s_inode_wblist_lock);
2625 : }
2626 6491880 : spin_unlock_irq(&sb->s_inode_wblist_lock);
2627 6491882 : rcu_read_unlock();
2628 6491880 : mutex_unlock(&sb->s_sync_lock);
2629 6491848 : }
2630 :
2631 104049 : static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2632 : enum wb_reason reason, bool skip_if_busy)
2633 : {
2634 104049 : struct backing_dev_info *bdi = sb->s_bdi;
2635 104049 : DEFINE_WB_COMPLETION(done, bdi);
2636 104049 : struct wb_writeback_work work = {
2637 : .sb = sb,
2638 : .sync_mode = WB_SYNC_NONE,
2639 : .tagged_writepages = 1,
2640 : .done = &done,
2641 : .nr_pages = nr,
2642 : .reason = reason,
2643 : };
2644 :
2645 104049 : if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2646 44951 : return;
2647 59098 : WARN_ON(!rwsem_is_locked(&sb->s_umount));
2648 :
2649 59098 : bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2650 59098 : wb_wait_for_completion(&done);
2651 : }
2652 :
2653 : /**
2654 : * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2655 : * @sb: the superblock
2656 : * @nr: the number of pages to write
2657 : * @reason: reason why some writeback work initiated
2658 : *
2659 : * Start writeback on some inodes on this super_block. No guarantees are made
2660 : * on how many (if any) will be written, and this function does not wait
2661 : * for IO completion of submitted IO.
2662 : */
2663 0 : void writeback_inodes_sb_nr(struct super_block *sb,
2664 : unsigned long nr,
2665 : enum wb_reason reason)
2666 : {
2667 104049 : __writeback_inodes_sb_nr(sb, nr, reason, false);
2668 0 : }
2669 : EXPORT_SYMBOL(writeback_inodes_sb_nr);
2670 :
2671 : /**
2672 : * writeback_inodes_sb - writeback dirty inodes from given super_block
2673 : * @sb: the superblock
2674 : * @reason: reason why some writeback work was initiated
2675 : *
2676 : * Start writeback on some inodes on this super_block. No guarantees are made
2677 : * on how many (if any) will be written, and this function does not wait
2678 : * for IO completion of submitted IO.
2679 : */
2680 104049 : void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2681 : {
2682 104049 : return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2683 : }
2684 : EXPORT_SYMBOL(writeback_inodes_sb);
2685 :
2686 : /**
2687 : * try_to_writeback_inodes_sb - try to start writeback if none underway
2688 : * @sb: the superblock
2689 : * @reason: reason why some writeback work was initiated
2690 : *
2691 : * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2692 : */
2693 0 : void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2694 : {
2695 0 : if (!down_read_trylock(&sb->s_umount))
2696 : return;
2697 :
2698 0 : __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2699 0 : up_read(&sb->s_umount);
2700 : }
2701 : EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2702 :
2703 : /**
2704 : * sync_inodes_sb - sync sb inode pages
2705 : * @sb: the superblock
2706 : *
2707 : * This function writes and waits on any dirty inode belonging to this
2708 : * super_block.
2709 : */
2710 61044240 : void sync_inodes_sb(struct super_block *sb)
2711 : {
2712 61044240 : struct backing_dev_info *bdi = sb->s_bdi;
2713 61044240 : DEFINE_WB_COMPLETION(done, bdi);
2714 61044240 : struct wb_writeback_work work = {
2715 : .sb = sb,
2716 : .sync_mode = WB_SYNC_ALL,
2717 : .nr_pages = LONG_MAX,
2718 : .range_cyclic = 0,
2719 : .done = &done,
2720 : .reason = WB_REASON_SYNC,
2721 : .for_sync = 1,
2722 : };
2723 :
2724 : /*
2725 : * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2726 : * inodes under writeback and I_DIRTY_TIME inodes ignored by
2727 : * bdi_has_dirty() need to be written out too.
2728 : */
2729 61044240 : if (bdi == &noop_backing_dev_info)
2730 54552366 : return;
2731 6491874 : WARN_ON(!rwsem_is_locked(&sb->s_umount));
2732 :
2733 : /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2734 6491874 : bdi_down_write_wb_switch_rwsem(bdi);
2735 6491886 : bdi_split_work_to_wbs(bdi, &work, false);
2736 6491857 : wb_wait_for_completion(&done);
2737 6491857 : bdi_up_write_wb_switch_rwsem(bdi);
2738 :
2739 6491832 : wait_sb_inodes(sb);
2740 : }
2741 : EXPORT_SYMBOL(sync_inodes_sb);
2742 :
2743 : /**
2744 : * write_inode_now - write an inode to disk
2745 : * @inode: inode to write to disk
2746 : * @sync: whether the write should be synchronous or not
2747 : *
2748 : * This function commits an inode to disk immediately if it is dirty. This is
2749 : * primarily needed by knfsd.
2750 : *
2751 : * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2752 : */
2753 516899237 : int write_inode_now(struct inode *inode, int sync)
2754 : {
2755 516899237 : struct writeback_control wbc = {
2756 : .nr_to_write = LONG_MAX,
2757 516899237 : .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2758 : .range_start = 0,
2759 : .range_end = LLONG_MAX,
2760 : };
2761 :
2762 516899237 : if (!mapping_can_writeback(inode->i_mapping))
2763 0 : wbc.nr_to_write = 0;
2764 :
2765 516921038 : might_sleep();
2766 516928941 : return writeback_single_inode(inode, &wbc);
2767 : }
2768 : EXPORT_SYMBOL(write_inode_now);
2769 :
2770 : /**
2771 : * sync_inode_metadata - write an inode to disk
2772 : * @inode: the inode to sync
2773 : * @wait: wait for I/O to complete.
2774 : *
2775 : * Write an inode to disk and adjust its dirty state after completion.
2776 : *
2777 : * Note: only writes the actual inode, no associated data or other metadata.
2778 : */
2779 597 : int sync_inode_metadata(struct inode *inode, int wait)
2780 : {
2781 597 : struct writeback_control wbc = {
2782 597 : .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2783 : .nr_to_write = 0, /* metadata-only */
2784 : };
2785 :
2786 597 : return writeback_single_inode(inode, &wbc);
2787 : }
2788 : EXPORT_SYMBOL(sync_inode_metadata);
|
