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 2769 : static inline struct inode *wb_inode(struct list_head *head)
71 : {
72 2769 : 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 2163 : static bool wb_io_lists_populated(struct bdi_writeback *wb)
86 : {
87 2163 : if (wb_has_dirty_io(wb)) {
88 : return false;
89 : } else {
90 2 : set_bit(WB_has_dirty_io, &wb->state);
91 2 : WARN_ON_ONCE(!wb->avg_write_bandwidth);
92 4 : atomic_long_add(wb->avg_write_bandwidth,
93 2 : &wb->bdi->tot_write_bandwidth);
94 2 : return true;
95 : }
96 : }
97 :
98 1575 : static void wb_io_lists_depopulated(struct bdi_writeback *wb)
99 : {
100 1575 : if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
101 0 : list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
102 0 : clear_bit(WB_has_dirty_io, &wb->state);
103 0 : WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
104 : &wb->bdi->tot_write_bandwidth) < 0);
105 : }
106 1575 : }
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 2157 : static bool inode_io_list_move_locked(struct inode *inode,
119 : struct bdi_writeback *wb,
120 : struct list_head *head)
121 : {
122 2157 : assert_spin_locked(&wb->list_lock);
123 :
124 2157 : list_move(&inode->i_io_list, head);
125 :
126 : /* dirty_time doesn't count as dirty_io until expiration */
127 2157 : if (head != &wb->b_dirty_time)
128 2157 : return wb_io_lists_populated(wb);
129 :
130 0 : wb_io_lists_depopulated(wb);
131 0 : return false;
132 : }
133 :
134 : /**
135 : * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
136 : * @inode: inode to be removed
137 : * @wb: bdi_writeback @inode is being removed from
138 : *
139 : * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
140 : * clear %WB_has_dirty_io if all are empty afterwards.
141 : */
142 1575 : static void inode_io_list_del_locked(struct inode *inode,
143 : struct bdi_writeback *wb)
144 : {
145 1575 : assert_spin_locked(&wb->list_lock);
146 1575 : assert_spin_locked(&inode->i_lock);
147 :
148 1575 : inode->i_state &= ~I_SYNC_QUEUED;
149 1575 : list_del_init(&inode->i_io_list);
150 1575 : wb_io_lists_depopulated(wb);
151 1575 : }
152 :
153 0 : static void wb_wakeup(struct bdi_writeback *wb)
154 : {
155 0 : spin_lock_bh(&wb->work_lock);
156 0 : if (test_bit(WB_registered, &wb->state))
157 0 : mod_delayed_work(bdi_wq, &wb->dwork, 0);
158 0 : spin_unlock_bh(&wb->work_lock);
159 0 : }
160 :
161 0 : static void finish_writeback_work(struct bdi_writeback *wb,
162 : struct wb_writeback_work *work)
163 : {
164 0 : struct wb_completion *done = work->done;
165 :
166 0 : if (work->auto_free)
167 0 : kfree(work);
168 0 : if (done) {
169 0 : wait_queue_head_t *waitq = done->waitq;
170 :
171 : /* @done can't be accessed after the following dec */
172 0 : if (atomic_dec_and_test(&done->cnt))
173 0 : wake_up_all(waitq);
174 : }
175 0 : }
176 :
177 0 : static void wb_queue_work(struct bdi_writeback *wb,
178 : struct wb_writeback_work *work)
179 : {
180 0 : trace_writeback_queue(wb, work);
181 :
182 0 : if (work->done)
183 0 : atomic_inc(&work->done->cnt);
184 :
185 0 : spin_lock_bh(&wb->work_lock);
186 :
187 0 : if (test_bit(WB_registered, &wb->state)) {
188 0 : list_add_tail(&work->list, &wb->work_list);
189 0 : mod_delayed_work(bdi_wq, &wb->dwork, 0);
190 : } else
191 0 : finish_writeback_work(wb, work);
192 :
193 0 : spin_unlock_bh(&wb->work_lock);
194 0 : }
195 :
196 : /**
197 : * wb_wait_for_completion - wait for completion of bdi_writeback_works
198 : * @done: target wb_completion
199 : *
200 : * Wait for one or more work items issued to @bdi with their ->done field
201 : * set to @done, which should have been initialized with
202 : * DEFINE_WB_COMPLETION(). This function returns after all such work items
203 : * are completed. Work items which are waited upon aren't freed
204 : * automatically on completion.
205 : */
206 0 : void wb_wait_for_completion(struct wb_completion *done)
207 : {
208 0 : atomic_dec(&done->cnt); /* put down the initial count */
209 0 : wait_event(*done->waitq, !atomic_read(&done->cnt));
210 0 : }
211 :
212 : #ifdef CONFIG_CGROUP_WRITEBACK
213 :
214 : /*
215 : * Parameters for foreign inode detection, see wbc_detach_inode() to see
216 : * how they're used.
217 : *
218 : * These paramters are inherently heuristical as the detection target
219 : * itself is fuzzy. All we want to do is detaching an inode from the
220 : * current owner if it's being written to by some other cgroups too much.
221 : *
222 : * The current cgroup writeback is built on the assumption that multiple
223 : * cgroups writing to the same inode concurrently is very rare and a mode
224 : * of operation which isn't well supported. As such, the goal is not
225 : * taking too long when a different cgroup takes over an inode while
226 : * avoiding too aggressive flip-flops from occasional foreign writes.
227 : *
228 : * We record, very roughly, 2s worth of IO time history and if more than
229 : * half of that is foreign, trigger the switch. The recording is quantized
230 : * to 16 slots. To avoid tiny writes from swinging the decision too much,
231 : * writes smaller than 1/8 of avg size are ignored.
232 : */
233 : #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
234 : #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
235 : #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
236 : #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
237 :
238 : #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
239 : #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
240 : /* each slot's duration is 2s / 16 */
241 : #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
242 : /* if foreign slots >= 8, switch */
243 : #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
244 : /* one round can affect upto 5 slots */
245 : #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
246 :
247 : static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
248 : static struct workqueue_struct *isw_wq;
249 :
250 : void __inode_attach_wb(struct inode *inode, struct page *page)
251 : {
252 : struct backing_dev_info *bdi = inode_to_bdi(inode);
253 : struct bdi_writeback *wb = NULL;
254 :
255 : if (inode_cgwb_enabled(inode)) {
256 : struct cgroup_subsys_state *memcg_css;
257 :
258 : if (page) {
259 : memcg_css = mem_cgroup_css_from_page(page);
260 : wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
261 : } else {
262 : /* must pin memcg_css, see wb_get_create() */
263 : memcg_css = task_get_css(current, memory_cgrp_id);
264 : wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
265 : css_put(memcg_css);
266 : }
267 : }
268 :
269 : if (!wb)
270 : wb = &bdi->wb;
271 :
272 : /*
273 : * There may be multiple instances of this function racing to
274 : * update the same inode. Use cmpxchg() to tell the winner.
275 : */
276 : if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
277 : wb_put(wb);
278 : }
279 : EXPORT_SYMBOL_GPL(__inode_attach_wb);
280 :
281 : /**
282 : * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
283 : * @inode: inode of interest with i_lock held
284 : *
285 : * Returns @inode's wb with its list_lock held. @inode->i_lock must be
286 : * held on entry and is released on return. The returned wb is guaranteed
287 : * to stay @inode's associated wb until its list_lock is released.
288 : */
289 : static struct bdi_writeback *
290 : locked_inode_to_wb_and_lock_list(struct inode *inode)
291 : __releases(&inode->i_lock)
292 : __acquires(&wb->list_lock)
293 : {
294 : while (true) {
295 : struct bdi_writeback *wb = inode_to_wb(inode);
296 :
297 : /*
298 : * inode_to_wb() association is protected by both
299 : * @inode->i_lock and @wb->list_lock but list_lock nests
300 : * outside i_lock. Drop i_lock and verify that the
301 : * association hasn't changed after acquiring list_lock.
302 : */
303 : wb_get(wb);
304 : spin_unlock(&inode->i_lock);
305 : spin_lock(&wb->list_lock);
306 :
307 : /* i_wb may have changed inbetween, can't use inode_to_wb() */
308 : if (likely(wb == inode->i_wb)) {
309 : wb_put(wb); /* @inode already has ref */
310 : return wb;
311 : }
312 :
313 : spin_unlock(&wb->list_lock);
314 : wb_put(wb);
315 : cpu_relax();
316 : spin_lock(&inode->i_lock);
317 : }
318 : }
319 :
320 : /**
321 : * inode_to_wb_and_lock_list - determine an inode's wb and lock it
322 : * @inode: inode of interest
323 : *
324 : * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
325 : * on entry.
326 : */
327 : static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
328 : __acquires(&wb->list_lock)
329 : {
330 : spin_lock(&inode->i_lock);
331 : return locked_inode_to_wb_and_lock_list(inode);
332 : }
333 :
334 : struct inode_switch_wbs_context {
335 : struct inode *inode;
336 : struct bdi_writeback *new_wb;
337 :
338 : struct rcu_head rcu_head;
339 : struct work_struct work;
340 : };
341 :
342 : static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
343 : {
344 : down_write(&bdi->wb_switch_rwsem);
345 : }
346 :
347 : static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
348 : {
349 : up_write(&bdi->wb_switch_rwsem);
350 : }
351 :
352 : static void inode_switch_wbs_work_fn(struct work_struct *work)
353 : {
354 : struct inode_switch_wbs_context *isw =
355 : container_of(work, struct inode_switch_wbs_context, work);
356 : struct inode *inode = isw->inode;
357 : struct backing_dev_info *bdi = inode_to_bdi(inode);
358 : struct address_space *mapping = inode->i_mapping;
359 : struct bdi_writeback *old_wb = inode->i_wb;
360 : struct bdi_writeback *new_wb = isw->new_wb;
361 : XA_STATE(xas, &mapping->i_pages, 0);
362 : struct page *page;
363 : bool switched = false;
364 :
365 : /*
366 : * If @inode switches cgwb membership while sync_inodes_sb() is
367 : * being issued, sync_inodes_sb() might miss it. Synchronize.
368 : */
369 : down_read(&bdi->wb_switch_rwsem);
370 :
371 : /*
372 : * By the time control reaches here, RCU grace period has passed
373 : * since I_WB_SWITCH assertion and all wb stat update transactions
374 : * between unlocked_inode_to_wb_begin/end() are guaranteed to be
375 : * synchronizing against the i_pages lock.
376 : *
377 : * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
378 : * gives us exclusion against all wb related operations on @inode
379 : * including IO list manipulations and stat updates.
380 : */
381 : if (old_wb < new_wb) {
382 : spin_lock(&old_wb->list_lock);
383 : spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
384 : } else {
385 : spin_lock(&new_wb->list_lock);
386 : spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
387 : }
388 : spin_lock(&inode->i_lock);
389 : xa_lock_irq(&mapping->i_pages);
390 :
391 : /*
392 : * Once I_FREEING is visible under i_lock, the eviction path owns
393 : * the inode and we shouldn't modify ->i_io_list.
394 : */
395 : if (unlikely(inode->i_state & I_FREEING))
396 : goto skip_switch;
397 :
398 : trace_inode_switch_wbs(inode, old_wb, new_wb);
399 :
400 : /*
401 : * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
402 : * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
403 : * pages actually under writeback.
404 : */
405 : xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
406 : if (PageDirty(page)) {
407 : dec_wb_stat(old_wb, WB_RECLAIMABLE);
408 : inc_wb_stat(new_wb, WB_RECLAIMABLE);
409 : }
410 : }
411 :
412 : xas_set(&xas, 0);
413 : xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
414 : WARN_ON_ONCE(!PageWriteback(page));
415 : dec_wb_stat(old_wb, WB_WRITEBACK);
416 : inc_wb_stat(new_wb, WB_WRITEBACK);
417 : }
418 :
419 : wb_get(new_wb);
420 :
421 : /*
422 : * Transfer to @new_wb's IO list if necessary. The specific list
423 : * @inode was on is ignored and the inode is put on ->b_dirty which
424 : * is always correct including from ->b_dirty_time. The transfer
425 : * preserves @inode->dirtied_when ordering.
426 : */
427 : if (!list_empty(&inode->i_io_list)) {
428 : struct inode *pos;
429 :
430 : inode_io_list_del_locked(inode, old_wb);
431 : inode->i_wb = new_wb;
432 : list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
433 : if (time_after_eq(inode->dirtied_when,
434 : pos->dirtied_when))
435 : break;
436 : inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
437 : } else {
438 : inode->i_wb = new_wb;
439 : }
440 :
441 : /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
442 : inode->i_wb_frn_winner = 0;
443 : inode->i_wb_frn_avg_time = 0;
444 : inode->i_wb_frn_history = 0;
445 : switched = true;
446 : skip_switch:
447 : /*
448 : * Paired with load_acquire in unlocked_inode_to_wb_begin() and
449 : * ensures that the new wb is visible if they see !I_WB_SWITCH.
450 : */
451 : smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
452 :
453 : xa_unlock_irq(&mapping->i_pages);
454 : spin_unlock(&inode->i_lock);
455 : spin_unlock(&new_wb->list_lock);
456 : spin_unlock(&old_wb->list_lock);
457 :
458 : up_read(&bdi->wb_switch_rwsem);
459 :
460 : if (switched) {
461 : wb_wakeup(new_wb);
462 : wb_put(old_wb);
463 : }
464 : wb_put(new_wb);
465 :
466 : iput(inode);
467 : kfree(isw);
468 :
469 : atomic_dec(&isw_nr_in_flight);
470 : }
471 :
472 : static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
473 : {
474 : struct inode_switch_wbs_context *isw = container_of(rcu_head,
475 : struct inode_switch_wbs_context, rcu_head);
476 :
477 : /* needs to grab bh-unsafe locks, bounce to work item */
478 : INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
479 : queue_work(isw_wq, &isw->work);
480 : }
481 :
482 : /**
483 : * inode_switch_wbs - change the wb association of an inode
484 : * @inode: target inode
485 : * @new_wb_id: ID of the new wb
486 : *
487 : * Switch @inode's wb association to the wb identified by @new_wb_id. The
488 : * switching is performed asynchronously and may fail silently.
489 : */
490 : static void inode_switch_wbs(struct inode *inode, int new_wb_id)
491 : {
492 : struct backing_dev_info *bdi = inode_to_bdi(inode);
493 : struct cgroup_subsys_state *memcg_css;
494 : struct inode_switch_wbs_context *isw;
495 :
496 : /* noop if seems to be already in progress */
497 : if (inode->i_state & I_WB_SWITCH)
498 : return;
499 :
500 : /* avoid queueing a new switch if too many are already in flight */
501 : if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
502 : return;
503 :
504 : isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
505 : if (!isw)
506 : return;
507 :
508 : /* find and pin the new wb */
509 : rcu_read_lock();
510 : memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
511 : if (memcg_css)
512 : isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
513 : rcu_read_unlock();
514 : if (!isw->new_wb)
515 : goto out_free;
516 :
517 : /* while holding I_WB_SWITCH, no one else can update the association */
518 : spin_lock(&inode->i_lock);
519 : if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
520 : inode->i_state & (I_WB_SWITCH | I_FREEING) ||
521 : inode_to_wb(inode) == isw->new_wb) {
522 : spin_unlock(&inode->i_lock);
523 : goto out_free;
524 : }
525 : inode->i_state |= I_WB_SWITCH;
526 : __iget(inode);
527 : spin_unlock(&inode->i_lock);
528 :
529 : isw->inode = inode;
530 :
531 : /*
532 : * In addition to synchronizing among switchers, I_WB_SWITCH tells
533 : * the RCU protected stat update paths to grab the i_page
534 : * lock so that stat transfer can synchronize against them.
535 : * Let's continue after I_WB_SWITCH is guaranteed to be visible.
536 : */
537 : call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
538 :
539 : atomic_inc(&isw_nr_in_flight);
540 : return;
541 :
542 : out_free:
543 : if (isw->new_wb)
544 : wb_put(isw->new_wb);
545 : kfree(isw);
546 : }
547 :
548 : /**
549 : * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
550 : * @wbc: writeback_control of interest
551 : * @inode: target inode
552 : *
553 : * @inode is locked and about to be written back under the control of @wbc.
554 : * Record @inode's writeback context into @wbc and unlock the i_lock. On
555 : * writeback completion, wbc_detach_inode() should be called. This is used
556 : * to track the cgroup writeback context.
557 : */
558 : void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
559 : struct inode *inode)
560 : {
561 : if (!inode_cgwb_enabled(inode)) {
562 : spin_unlock(&inode->i_lock);
563 : return;
564 : }
565 :
566 : wbc->wb = inode_to_wb(inode);
567 : wbc->inode = inode;
568 :
569 : wbc->wb_id = wbc->wb->memcg_css->id;
570 : wbc->wb_lcand_id = inode->i_wb_frn_winner;
571 : wbc->wb_tcand_id = 0;
572 : wbc->wb_bytes = 0;
573 : wbc->wb_lcand_bytes = 0;
574 : wbc->wb_tcand_bytes = 0;
575 :
576 : wb_get(wbc->wb);
577 : spin_unlock(&inode->i_lock);
578 :
579 : /*
580 : * A dying wb indicates that either the blkcg associated with the
581 : * memcg changed or the associated memcg is dying. In the first
582 : * case, a replacement wb should already be available and we should
583 : * refresh the wb immediately. In the second case, trying to
584 : * refresh will keep failing.
585 : */
586 : if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
587 : inode_switch_wbs(inode, wbc->wb_id);
588 : }
589 : EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
590 :
591 : /**
592 : * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
593 : * @wbc: writeback_control of the just finished writeback
594 : *
595 : * To be called after a writeback attempt of an inode finishes and undoes
596 : * wbc_attach_and_unlock_inode(). Can be called under any context.
597 : *
598 : * As concurrent write sharing of an inode is expected to be very rare and
599 : * memcg only tracks page ownership on first-use basis severely confining
600 : * the usefulness of such sharing, cgroup writeback tracks ownership
601 : * per-inode. While the support for concurrent write sharing of an inode
602 : * is deemed unnecessary, an inode being written to by different cgroups at
603 : * different points in time is a lot more common, and, more importantly,
604 : * charging only by first-use can too readily lead to grossly incorrect
605 : * behaviors (single foreign page can lead to gigabytes of writeback to be
606 : * incorrectly attributed).
607 : *
608 : * To resolve this issue, cgroup writeback detects the majority dirtier of
609 : * an inode and transfers the ownership to it. To avoid unnnecessary
610 : * oscillation, the detection mechanism keeps track of history and gives
611 : * out the switch verdict only if the foreign usage pattern is stable over
612 : * a certain amount of time and/or writeback attempts.
613 : *
614 : * On each writeback attempt, @wbc tries to detect the majority writer
615 : * using Boyer-Moore majority vote algorithm. In addition to the byte
616 : * count from the majority voting, it also counts the bytes written for the
617 : * current wb and the last round's winner wb (max of last round's current
618 : * wb, the winner from two rounds ago, and the last round's majority
619 : * candidate). Keeping track of the historical winner helps the algorithm
620 : * to semi-reliably detect the most active writer even when it's not the
621 : * absolute majority.
622 : *
623 : * Once the winner of the round is determined, whether the winner is
624 : * foreign or not and how much IO time the round consumed is recorded in
625 : * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
626 : * over a certain threshold, the switch verdict is given.
627 : */
628 : void wbc_detach_inode(struct writeback_control *wbc)
629 : {
630 : struct bdi_writeback *wb = wbc->wb;
631 : struct inode *inode = wbc->inode;
632 : unsigned long avg_time, max_bytes, max_time;
633 : u16 history;
634 : int max_id;
635 :
636 : if (!wb)
637 : return;
638 :
639 : history = inode->i_wb_frn_history;
640 : avg_time = inode->i_wb_frn_avg_time;
641 :
642 : /* pick the winner of this round */
643 : if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
644 : wbc->wb_bytes >= wbc->wb_tcand_bytes) {
645 : max_id = wbc->wb_id;
646 : max_bytes = wbc->wb_bytes;
647 : } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
648 : max_id = wbc->wb_lcand_id;
649 : max_bytes = wbc->wb_lcand_bytes;
650 : } else {
651 : max_id = wbc->wb_tcand_id;
652 : max_bytes = wbc->wb_tcand_bytes;
653 : }
654 :
655 : /*
656 : * Calculate the amount of IO time the winner consumed and fold it
657 : * into the running average kept per inode. If the consumed IO
658 : * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
659 : * deciding whether to switch or not. This is to prevent one-off
660 : * small dirtiers from skewing the verdict.
661 : */
662 : max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
663 : wb->avg_write_bandwidth);
664 : if (avg_time)
665 : avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
666 : (avg_time >> WB_FRN_TIME_AVG_SHIFT);
667 : else
668 : avg_time = max_time; /* immediate catch up on first run */
669 :
670 : if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
671 : int slots;
672 :
673 : /*
674 : * The switch verdict is reached if foreign wb's consume
675 : * more than a certain proportion of IO time in a
676 : * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
677 : * history mask where each bit represents one sixteenth of
678 : * the period. Determine the number of slots to shift into
679 : * history from @max_time.
680 : */
681 : slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
682 : (unsigned long)WB_FRN_HIST_MAX_SLOTS);
683 : history <<= slots;
684 : if (wbc->wb_id != max_id)
685 : history |= (1U << slots) - 1;
686 :
687 : if (history)
688 : trace_inode_foreign_history(inode, wbc, history);
689 :
690 : /*
691 : * Switch if the current wb isn't the consistent winner.
692 : * If there are multiple closely competing dirtiers, the
693 : * inode may switch across them repeatedly over time, which
694 : * is okay. The main goal is avoiding keeping an inode on
695 : * the wrong wb for an extended period of time.
696 : */
697 : if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
698 : inode_switch_wbs(inode, max_id);
699 : }
700 :
701 : /*
702 : * Multiple instances of this function may race to update the
703 : * following fields but we don't mind occassional inaccuracies.
704 : */
705 : inode->i_wb_frn_winner = max_id;
706 : inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
707 : inode->i_wb_frn_history = history;
708 :
709 : wb_put(wbc->wb);
710 : wbc->wb = NULL;
711 : }
712 : EXPORT_SYMBOL_GPL(wbc_detach_inode);
713 :
714 : /**
715 : * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
716 : * @wbc: writeback_control of the writeback in progress
717 : * @page: page being written out
718 : * @bytes: number of bytes being written out
719 : *
720 : * @bytes from @page are about to written out during the writeback
721 : * controlled by @wbc. Keep the book for foreign inode detection. See
722 : * wbc_detach_inode().
723 : */
724 : void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
725 : size_t bytes)
726 : {
727 : struct cgroup_subsys_state *css;
728 : int id;
729 :
730 : /*
731 : * pageout() path doesn't attach @wbc to the inode being written
732 : * out. This is intentional as we don't want the function to block
733 : * behind a slow cgroup. Ultimately, we want pageout() to kick off
734 : * regular writeback instead of writing things out itself.
735 : */
736 : if (!wbc->wb || wbc->no_cgroup_owner)
737 : return;
738 :
739 : css = mem_cgroup_css_from_page(page);
740 : /* dead cgroups shouldn't contribute to inode ownership arbitration */
741 : if (!(css->flags & CSS_ONLINE))
742 : return;
743 :
744 : id = css->id;
745 :
746 : if (id == wbc->wb_id) {
747 : wbc->wb_bytes += bytes;
748 : return;
749 : }
750 :
751 : if (id == wbc->wb_lcand_id)
752 : wbc->wb_lcand_bytes += bytes;
753 :
754 : /* Boyer-Moore majority vote algorithm */
755 : if (!wbc->wb_tcand_bytes)
756 : wbc->wb_tcand_id = id;
757 : if (id == wbc->wb_tcand_id)
758 : wbc->wb_tcand_bytes += bytes;
759 : else
760 : wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
761 : }
762 : EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
763 :
764 : /**
765 : * inode_congested - test whether an inode is congested
766 : * @inode: inode to test for congestion (may be NULL)
767 : * @cong_bits: mask of WB_[a]sync_congested bits to test
768 : *
769 : * Tests whether @inode is congested. @cong_bits is the mask of congestion
770 : * bits to test and the return value is the mask of set bits.
771 : *
772 : * If cgroup writeback is enabled for @inode, the congestion state is
773 : * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
774 : * associated with @inode is congested; otherwise, the root wb's congestion
775 : * state is used.
776 : *
777 : * @inode is allowed to be NULL as this function is often called on
778 : * mapping->host which is NULL for the swapper space.
779 : */
780 : int inode_congested(struct inode *inode, int cong_bits)
781 : {
782 : /*
783 : * Once set, ->i_wb never becomes NULL while the inode is alive.
784 : * Start transaction iff ->i_wb is visible.
785 : */
786 : if (inode && inode_to_wb_is_valid(inode)) {
787 : struct bdi_writeback *wb;
788 : struct wb_lock_cookie lock_cookie = {};
789 : bool congested;
790 :
791 : wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
792 : congested = wb_congested(wb, cong_bits);
793 : unlocked_inode_to_wb_end(inode, &lock_cookie);
794 : return congested;
795 : }
796 :
797 : return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
798 : }
799 : EXPORT_SYMBOL_GPL(inode_congested);
800 :
801 : /**
802 : * wb_split_bdi_pages - split nr_pages to write according to bandwidth
803 : * @wb: target bdi_writeback to split @nr_pages to
804 : * @nr_pages: number of pages to write for the whole bdi
805 : *
806 : * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
807 : * relation to the total write bandwidth of all wb's w/ dirty inodes on
808 : * @wb->bdi.
809 : */
810 : static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
811 : {
812 : unsigned long this_bw = wb->avg_write_bandwidth;
813 : unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
814 :
815 : if (nr_pages == LONG_MAX)
816 : return LONG_MAX;
817 :
818 : /*
819 : * This may be called on clean wb's and proportional distribution
820 : * may not make sense, just use the original @nr_pages in those
821 : * cases. In general, we wanna err on the side of writing more.
822 : */
823 : if (!tot_bw || this_bw >= tot_bw)
824 : return nr_pages;
825 : else
826 : return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
827 : }
828 :
829 : /**
830 : * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
831 : * @bdi: target backing_dev_info
832 : * @base_work: wb_writeback_work to issue
833 : * @skip_if_busy: skip wb's which already have writeback in progress
834 : *
835 : * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
836 : * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
837 : * distributed to the busy wbs according to each wb's proportion in the
838 : * total active write bandwidth of @bdi.
839 : */
840 : static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
841 : struct wb_writeback_work *base_work,
842 : bool skip_if_busy)
843 : {
844 : struct bdi_writeback *last_wb = NULL;
845 : struct bdi_writeback *wb = list_entry(&bdi->wb_list,
846 : struct bdi_writeback, bdi_node);
847 :
848 : might_sleep();
849 : restart:
850 : rcu_read_lock();
851 : list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
852 : DEFINE_WB_COMPLETION(fallback_work_done, bdi);
853 : struct wb_writeback_work fallback_work;
854 : struct wb_writeback_work *work;
855 : long nr_pages;
856 :
857 : if (last_wb) {
858 : wb_put(last_wb);
859 : last_wb = NULL;
860 : }
861 :
862 : /* SYNC_ALL writes out I_DIRTY_TIME too */
863 : if (!wb_has_dirty_io(wb) &&
864 : (base_work->sync_mode == WB_SYNC_NONE ||
865 : list_empty(&wb->b_dirty_time)))
866 : continue;
867 : if (skip_if_busy && writeback_in_progress(wb))
868 : continue;
869 :
870 : nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
871 :
872 : work = kmalloc(sizeof(*work), GFP_ATOMIC);
873 : if (work) {
874 : *work = *base_work;
875 : work->nr_pages = nr_pages;
876 : work->auto_free = 1;
877 : wb_queue_work(wb, work);
878 : continue;
879 : }
880 :
881 : /* alloc failed, execute synchronously using on-stack fallback */
882 : work = &fallback_work;
883 : *work = *base_work;
884 : work->nr_pages = nr_pages;
885 : work->auto_free = 0;
886 : work->done = &fallback_work_done;
887 :
888 : wb_queue_work(wb, work);
889 :
890 : /*
891 : * Pin @wb so that it stays on @bdi->wb_list. This allows
892 : * continuing iteration from @wb after dropping and
893 : * regrabbing rcu read lock.
894 : */
895 : wb_get(wb);
896 : last_wb = wb;
897 :
898 : rcu_read_unlock();
899 : wb_wait_for_completion(&fallback_work_done);
900 : goto restart;
901 : }
902 : rcu_read_unlock();
903 :
904 : if (last_wb)
905 : wb_put(last_wb);
906 : }
907 :
908 : /**
909 : * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
910 : * @bdi_id: target bdi id
911 : * @memcg_id: target memcg css id
912 : * @nr: number of pages to write, 0 for best-effort dirty flushing
913 : * @reason: reason why some writeback work initiated
914 : * @done: target wb_completion
915 : *
916 : * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
917 : * with the specified parameters.
918 : */
919 : int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, unsigned long nr,
920 : enum wb_reason reason, struct wb_completion *done)
921 : {
922 : struct backing_dev_info *bdi;
923 : struct cgroup_subsys_state *memcg_css;
924 : struct bdi_writeback *wb;
925 : struct wb_writeback_work *work;
926 : int ret;
927 :
928 : /* lookup bdi and memcg */
929 : bdi = bdi_get_by_id(bdi_id);
930 : if (!bdi)
931 : return -ENOENT;
932 :
933 : rcu_read_lock();
934 : memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
935 : if (memcg_css && !css_tryget(memcg_css))
936 : memcg_css = NULL;
937 : rcu_read_unlock();
938 : if (!memcg_css) {
939 : ret = -ENOENT;
940 : goto out_bdi_put;
941 : }
942 :
943 : /*
944 : * And find the associated wb. If the wb isn't there already
945 : * there's nothing to flush, don't create one.
946 : */
947 : wb = wb_get_lookup(bdi, memcg_css);
948 : if (!wb) {
949 : ret = -ENOENT;
950 : goto out_css_put;
951 : }
952 :
953 : /*
954 : * If @nr is zero, the caller is attempting to write out most of
955 : * the currently dirty pages. Let's take the current dirty page
956 : * count and inflate it by 25% which should be large enough to
957 : * flush out most dirty pages while avoiding getting livelocked by
958 : * concurrent dirtiers.
959 : */
960 : if (!nr) {
961 : unsigned long filepages, headroom, dirty, writeback;
962 :
963 : mem_cgroup_wb_stats(wb, &filepages, &headroom, &dirty,
964 : &writeback);
965 : nr = dirty * 10 / 8;
966 : }
967 :
968 : /* issue the writeback work */
969 : work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
970 : if (work) {
971 : work->nr_pages = nr;
972 : work->sync_mode = WB_SYNC_NONE;
973 : work->range_cyclic = 1;
974 : work->reason = reason;
975 : work->done = done;
976 : work->auto_free = 1;
977 : wb_queue_work(wb, work);
978 : ret = 0;
979 : } else {
980 : ret = -ENOMEM;
981 : }
982 :
983 : wb_put(wb);
984 : out_css_put:
985 : css_put(memcg_css);
986 : out_bdi_put:
987 : bdi_put(bdi);
988 : return ret;
989 : }
990 :
991 : /**
992 : * cgroup_writeback_umount - flush inode wb switches for umount
993 : *
994 : * This function is called when a super_block is about to be destroyed and
995 : * flushes in-flight inode wb switches. An inode wb switch goes through
996 : * RCU and then workqueue, so the two need to be flushed in order to ensure
997 : * that all previously scheduled switches are finished. As wb switches are
998 : * rare occurrences and synchronize_rcu() can take a while, perform
999 : * flushing iff wb switches are in flight.
1000 : */
1001 : void cgroup_writeback_umount(void)
1002 : {
1003 : if (atomic_read(&isw_nr_in_flight)) {
1004 : /*
1005 : * Use rcu_barrier() to wait for all pending callbacks to
1006 : * ensure that all in-flight wb switches are in the workqueue.
1007 : */
1008 : rcu_barrier();
1009 : flush_workqueue(isw_wq);
1010 : }
1011 : }
1012 :
1013 : static int __init cgroup_writeback_init(void)
1014 : {
1015 : isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
1016 : if (!isw_wq)
1017 : return -ENOMEM;
1018 : return 0;
1019 : }
1020 : fs_initcall(cgroup_writeback_init);
1021 :
1022 : #else /* CONFIG_CGROUP_WRITEBACK */
1023 :
1024 0 : static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1025 0 : static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1026 :
1027 : static struct bdi_writeback *
1028 2157 : locked_inode_to_wb_and_lock_list(struct inode *inode)
1029 : __releases(&inode->i_lock)
1030 : __acquires(&wb->list_lock)
1031 : {
1032 2157 : struct bdi_writeback *wb = inode_to_wb(inode);
1033 :
1034 2157 : spin_unlock(&inode->i_lock);
1035 2157 : spin_lock(&wb->list_lock);
1036 2157 : return wb;
1037 : }
1038 :
1039 1575 : static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1040 : __acquires(&wb->list_lock)
1041 : {
1042 3150 : struct bdi_writeback *wb = inode_to_wb(inode);
1043 :
1044 1575 : spin_lock(&wb->list_lock);
1045 1575 : return wb;
1046 : }
1047 :
1048 0 : static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1049 : {
1050 0 : return nr_pages;
1051 : }
1052 :
1053 0 : static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1054 : struct wb_writeback_work *base_work,
1055 : bool skip_if_busy)
1056 : {
1057 0 : might_sleep();
1058 :
1059 0 : if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1060 0 : base_work->auto_free = 0;
1061 0 : wb_queue_work(&bdi->wb, base_work);
1062 : }
1063 0 : }
1064 :
1065 : #endif /* CONFIG_CGROUP_WRITEBACK */
1066 :
1067 : /*
1068 : * Add in the number of potentially dirty inodes, because each inode
1069 : * write can dirty pagecache in the underlying blockdev.
1070 : */
1071 109 : static unsigned long get_nr_dirty_pages(void)
1072 : {
1073 109 : return global_node_page_state(NR_FILE_DIRTY) +
1074 109 : get_nr_dirty_inodes();
1075 : }
1076 :
1077 0 : static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1078 : {
1079 0 : if (!wb_has_dirty_io(wb))
1080 : return;
1081 :
1082 : /*
1083 : * All callers of this function want to start writeback of all
1084 : * dirty pages. Places like vmscan can call this at a very
1085 : * high frequency, causing pointless allocations of tons of
1086 : * work items and keeping the flusher threads busy retrieving
1087 : * that work. Ensure that we only allow one of them pending and
1088 : * inflight at the time.
1089 : */
1090 0 : if (test_bit(WB_start_all, &wb->state) ||
1091 0 : test_and_set_bit(WB_start_all, &wb->state))
1092 0 : return;
1093 :
1094 0 : wb->start_all_reason = reason;
1095 0 : wb_wakeup(wb);
1096 : }
1097 :
1098 : /**
1099 : * wb_start_background_writeback - start background writeback
1100 : * @wb: bdi_writback to write from
1101 : *
1102 : * Description:
1103 : * This makes sure WB_SYNC_NONE background writeback happens. When
1104 : * this function returns, it is only guaranteed that for given wb
1105 : * some IO is happening if we are over background dirty threshold.
1106 : * Caller need not hold sb s_umount semaphore.
1107 : */
1108 0 : void wb_start_background_writeback(struct bdi_writeback *wb)
1109 : {
1110 : /*
1111 : * We just wake up the flusher thread. It will perform background
1112 : * writeback as soon as there is no other work to do.
1113 : */
1114 0 : trace_writeback_wake_background(wb);
1115 0 : wb_wakeup(wb);
1116 0 : }
1117 :
1118 : /*
1119 : * Remove the inode from the writeback list it is on.
1120 : */
1121 201 : void inode_io_list_del(struct inode *inode)
1122 : {
1123 201 : struct bdi_writeback *wb;
1124 :
1125 201 : wb = inode_to_wb_and_lock_list(inode);
1126 201 : spin_lock(&inode->i_lock);
1127 201 : inode_io_list_del_locked(inode, wb);
1128 201 : spin_unlock(&inode->i_lock);
1129 201 : spin_unlock(&wb->list_lock);
1130 201 : }
1131 : EXPORT_SYMBOL(inode_io_list_del);
1132 :
1133 : /*
1134 : * mark an inode as under writeback on the sb
1135 : */
1136 84 : void sb_mark_inode_writeback(struct inode *inode)
1137 : {
1138 84 : struct super_block *sb = inode->i_sb;
1139 84 : unsigned long flags;
1140 :
1141 84 : if (list_empty(&inode->i_wb_list)) {
1142 84 : spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1143 84 : if (list_empty(&inode->i_wb_list)) {
1144 84 : list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1145 84 : trace_sb_mark_inode_writeback(inode);
1146 : }
1147 84 : spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1148 : }
1149 84 : }
1150 :
1151 : /*
1152 : * clear an inode as under writeback on the sb
1153 : */
1154 84 : void sb_clear_inode_writeback(struct inode *inode)
1155 : {
1156 84 : struct super_block *sb = inode->i_sb;
1157 84 : unsigned long flags;
1158 :
1159 84 : if (!list_empty(&inode->i_wb_list)) {
1160 84 : spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1161 84 : if (!list_empty(&inode->i_wb_list)) {
1162 84 : list_del_init(&inode->i_wb_list);
1163 84 : trace_sb_clear_inode_writeback(inode);
1164 : }
1165 84 : spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1166 : }
1167 84 : }
1168 :
1169 : /*
1170 : * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1171 : * furthest end of its superblock's dirty-inode list.
1172 : *
1173 : * Before stamping the inode's ->dirtied_when, we check to see whether it is
1174 : * already the most-recently-dirtied inode on the b_dirty list. If that is
1175 : * the case then the inode must have been redirtied while it was being written
1176 : * out and we don't reset its dirtied_when.
1177 : */
1178 0 : static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1179 : {
1180 0 : assert_spin_locked(&inode->i_lock);
1181 :
1182 0 : if (!list_empty(&wb->b_dirty)) {
1183 0 : struct inode *tail;
1184 :
1185 0 : tail = wb_inode(wb->b_dirty.next);
1186 0 : if (time_before(inode->dirtied_when, tail->dirtied_when))
1187 0 : inode->dirtied_when = jiffies;
1188 : }
1189 0 : inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1190 0 : inode->i_state &= ~I_SYNC_QUEUED;
1191 0 : }
1192 :
1193 0 : static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1194 : {
1195 0 : spin_lock(&inode->i_lock);
1196 0 : redirty_tail_locked(inode, wb);
1197 0 : spin_unlock(&inode->i_lock);
1198 0 : }
1199 :
1200 : /*
1201 : * requeue inode for re-scanning after bdi->b_io list is exhausted.
1202 : */
1203 0 : static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1204 : {
1205 0 : inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1206 0 : }
1207 :
1208 1374 : static void inode_sync_complete(struct inode *inode)
1209 : {
1210 1374 : inode->i_state &= ~I_SYNC;
1211 : /* If inode is clean an unused, put it into LRU now... */
1212 1374 : inode_add_lru(inode);
1213 : /* Waiters must see I_SYNC cleared before being woken up */
1214 1374 : smp_mb();
1215 1374 : wake_up_bit(&inode->i_state, __I_SYNC);
1216 1374 : }
1217 :
1218 1388 : static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1219 : {
1220 2776 : bool ret = time_after(inode->dirtied_when, t);
1221 : #ifndef CONFIG_64BIT
1222 : /*
1223 : * For inodes being constantly redirtied, dirtied_when can get stuck.
1224 : * It _appears_ to be in the future, but is actually in distant past.
1225 : * This test is necessary to prevent such wrapped-around relative times
1226 : * from permanently stopping the whole bdi writeback.
1227 : */
1228 : ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1229 : #endif
1230 1388 : return ret;
1231 : }
1232 :
1233 : #define EXPIRE_DIRTY_ATIME 0x0001
1234 :
1235 : /*
1236 : * Move expired (dirtied before dirtied_before) dirty inodes from
1237 : * @delaying_queue to @dispatch_queue.
1238 : */
1239 28 : static int move_expired_inodes(struct list_head *delaying_queue,
1240 : struct list_head *dispatch_queue,
1241 : unsigned long dirtied_before)
1242 : {
1243 28 : LIST_HEAD(tmp);
1244 28 : struct list_head *pos, *node;
1245 28 : struct super_block *sb = NULL;
1246 28 : struct inode *inode;
1247 28 : int do_sb_sort = 0;
1248 28 : int moved = 0;
1249 :
1250 1402 : while (!list_empty(delaying_queue)) {
1251 1388 : inode = wb_inode(delaying_queue->prev);
1252 2762 : if (inode_dirtied_after(inode, dirtied_before))
1253 : break;
1254 1374 : list_move(&inode->i_io_list, &tmp);
1255 1374 : moved++;
1256 1374 : spin_lock(&inode->i_lock);
1257 1374 : inode->i_state |= I_SYNC_QUEUED;
1258 1374 : spin_unlock(&inode->i_lock);
1259 1374 : if (sb_is_blkdev_sb(inode->i_sb))
1260 1 : continue;
1261 1373 : if (sb && sb != inode->i_sb)
1262 0 : do_sb_sort = 1;
1263 : sb = inode->i_sb;
1264 : }
1265 :
1266 : /* just one sb in list, splice to dispatch_queue and we're done */
1267 28 : if (!do_sb_sort) {
1268 28 : list_splice(&tmp, dispatch_queue);
1269 28 : goto out;
1270 : }
1271 :
1272 : /* Move inodes from one superblock together */
1273 0 : while (!list_empty(&tmp)) {
1274 0 : sb = wb_inode(tmp.prev)->i_sb;
1275 0 : list_for_each_prev_safe(pos, node, &tmp) {
1276 0 : inode = wb_inode(pos);
1277 0 : if (inode->i_sb == sb)
1278 0 : list_move(&inode->i_io_list, dispatch_queue);
1279 : }
1280 : }
1281 0 : out:
1282 28 : return moved;
1283 : }
1284 :
1285 : /*
1286 : * Queue all expired dirty inodes for io, eldest first.
1287 : * Before
1288 : * newly dirtied b_dirty b_io b_more_io
1289 : * =============> gf edc BA
1290 : * After
1291 : * newly dirtied b_dirty b_io b_more_io
1292 : * =============> g fBAedc
1293 : * |
1294 : * +--> dequeue for IO
1295 : */
1296 14 : static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1297 : unsigned long dirtied_before)
1298 : {
1299 14 : int moved;
1300 14 : unsigned long time_expire_jif = dirtied_before;
1301 :
1302 14 : assert_spin_locked(&wb->list_lock);
1303 14 : list_splice_init(&wb->b_more_io, &wb->b_io);
1304 14 : moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1305 14 : if (!work->for_sync)
1306 14 : time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1307 14 : moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1308 : time_expire_jif);
1309 14 : if (moved)
1310 6 : wb_io_lists_populated(wb);
1311 14 : trace_writeback_queue_io(wb, work, dirtied_before, moved);
1312 14 : }
1313 :
1314 1373 : static int write_inode(struct inode *inode, struct writeback_control *wbc)
1315 : {
1316 1373 : int ret;
1317 :
1318 1373 : if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1319 1373 : trace_writeback_write_inode_start(inode, wbc);
1320 1373 : ret = inode->i_sb->s_op->write_inode(inode, wbc);
1321 1373 : trace_writeback_write_inode(inode, wbc);
1322 1373 : return ret;
1323 : }
1324 : return 0;
1325 : }
1326 :
1327 : /*
1328 : * Wait for writeback on an inode to complete. Called with i_lock held.
1329 : * Caller must make sure inode cannot go away when we drop i_lock.
1330 : */
1331 5287 : static void __inode_wait_for_writeback(struct inode *inode)
1332 : __releases(inode->i_lock)
1333 : __acquires(inode->i_lock)
1334 : {
1335 5287 : DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1336 5287 : wait_queue_head_t *wqh;
1337 :
1338 5287 : wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1339 5287 : while (inode->i_state & I_SYNC) {
1340 0 : spin_unlock(&inode->i_lock);
1341 0 : __wait_on_bit(wqh, &wq, bit_wait,
1342 : TASK_UNINTERRUPTIBLE);
1343 5287 : spin_lock(&inode->i_lock);
1344 : }
1345 5287 : }
1346 :
1347 : /*
1348 : * Wait for writeback on an inode to complete. Caller must have inode pinned.
1349 : */
1350 5287 : void inode_wait_for_writeback(struct inode *inode)
1351 : {
1352 5287 : spin_lock(&inode->i_lock);
1353 5288 : __inode_wait_for_writeback(inode);
1354 5288 : spin_unlock(&inode->i_lock);
1355 5288 : }
1356 :
1357 : /*
1358 : * Sleep until I_SYNC is cleared. This function must be called with i_lock
1359 : * held and drops it. It is aimed for callers not holding any inode reference
1360 : * so once i_lock is dropped, inode can go away.
1361 : */
1362 0 : static void inode_sleep_on_writeback(struct inode *inode)
1363 : __releases(inode->i_lock)
1364 : {
1365 0 : DEFINE_WAIT(wait);
1366 0 : wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1367 0 : int sleep;
1368 :
1369 0 : prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1370 0 : sleep = inode->i_state & I_SYNC;
1371 0 : spin_unlock(&inode->i_lock);
1372 0 : if (sleep)
1373 0 : schedule();
1374 0 : finish_wait(wqh, &wait);
1375 0 : }
1376 :
1377 : /*
1378 : * Find proper writeback list for the inode depending on its current state and
1379 : * possibly also change of its state while we were doing writeback. Here we
1380 : * handle things such as livelock prevention or fairness of writeback among
1381 : * inodes. This function can be called only by flusher thread - noone else
1382 : * processes all inodes in writeback lists and requeueing inodes behind flusher
1383 : * thread's back can have unexpected consequences.
1384 : */
1385 1374 : static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1386 : struct writeback_control *wbc)
1387 : {
1388 1374 : if (inode->i_state & I_FREEING)
1389 : return;
1390 :
1391 : /*
1392 : * Sync livelock prevention. Each inode is tagged and synced in one
1393 : * shot. If still dirty, it will be redirty_tail()'ed below. Update
1394 : * the dirty time to prevent enqueue and sync it again.
1395 : */
1396 1374 : if ((inode->i_state & I_DIRTY) &&
1397 0 : (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1398 0 : inode->dirtied_when = jiffies;
1399 :
1400 1374 : if (wbc->pages_skipped) {
1401 : /*
1402 : * writeback is not making progress due to locked
1403 : * buffers. Skip this inode for now.
1404 : */
1405 0 : redirty_tail_locked(inode, wb);
1406 0 : return;
1407 : }
1408 :
1409 1374 : if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1410 : /*
1411 : * We didn't write back all the pages. nfs_writepages()
1412 : * sometimes bales out without doing anything.
1413 : */
1414 0 : if (wbc->nr_to_write <= 0) {
1415 : /* Slice used up. Queue for next turn. */
1416 0 : requeue_io(inode, wb);
1417 : } else {
1418 : /*
1419 : * Writeback blocked by something other than
1420 : * congestion. Delay the inode for some time to
1421 : * avoid spinning on the CPU (100% iowait)
1422 : * retrying writeback of the dirty page/inode
1423 : * that cannot be performed immediately.
1424 : */
1425 0 : redirty_tail_locked(inode, wb);
1426 : }
1427 1374 : } else if (inode->i_state & I_DIRTY) {
1428 : /*
1429 : * Filesystems can dirty the inode during writeback operations,
1430 : * such as delayed allocation during submission or metadata
1431 : * updates after data IO completion.
1432 : */
1433 0 : redirty_tail_locked(inode, wb);
1434 1374 : } else if (inode->i_state & I_DIRTY_TIME) {
1435 0 : inode->dirtied_when = jiffies;
1436 0 : inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1437 0 : inode->i_state &= ~I_SYNC_QUEUED;
1438 : } else {
1439 : /* The inode is clean. Remove from writeback lists. */
1440 1374 : inode_io_list_del_locked(inode, wb);
1441 : }
1442 : }
1443 :
1444 : /*
1445 : * Write out an inode and its dirty pages (or some of its dirty pages, depending
1446 : * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1447 : *
1448 : * This doesn't remove the inode from the writeback list it is on, except
1449 : * potentially to move it from b_dirty_time to b_dirty due to timestamp
1450 : * expiration. The caller is otherwise responsible for writeback list handling.
1451 : *
1452 : * The caller is also responsible for setting the I_SYNC flag beforehand and
1453 : * calling inode_sync_complete() to clear it afterwards.
1454 : */
1455 : static int
1456 1374 : __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1457 : {
1458 1374 : struct address_space *mapping = inode->i_mapping;
1459 1374 : long nr_to_write = wbc->nr_to_write;
1460 1374 : unsigned dirty;
1461 1374 : int ret;
1462 :
1463 1374 : WARN_ON(!(inode->i_state & I_SYNC));
1464 :
1465 1374 : trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1466 :
1467 1374 : ret = do_writepages(mapping, wbc);
1468 :
1469 : /*
1470 : * Make sure to wait on the data before writing out the metadata.
1471 : * This is important for filesystems that modify metadata on data
1472 : * I/O completion. We don't do it for sync(2) writeback because it has a
1473 : * separate, external IO completion path and ->sync_fs for guaranteeing
1474 : * inode metadata is written back correctly.
1475 : */
1476 1374 : if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1477 0 : int err = filemap_fdatawait(mapping);
1478 0 : if (ret == 0)
1479 0 : ret = err;
1480 : }
1481 :
1482 : /*
1483 : * If the inode has dirty timestamps and we need to write them, call
1484 : * mark_inode_dirty_sync() to notify the filesystem about it and to
1485 : * change I_DIRTY_TIME into I_DIRTY_SYNC.
1486 : */
1487 1374 : if ((inode->i_state & I_DIRTY_TIME) &&
1488 0 : (wbc->sync_mode == WB_SYNC_ALL ||
1489 0 : time_after(jiffies, inode->dirtied_time_when +
1490 : dirtytime_expire_interval * HZ))) {
1491 0 : trace_writeback_lazytime(inode);
1492 0 : mark_inode_dirty_sync(inode);
1493 : }
1494 :
1495 : /*
1496 : * Get and clear the dirty flags from i_state. This needs to be done
1497 : * after calling writepages because some filesystems may redirty the
1498 : * inode during writepages due to delalloc. It also needs to be done
1499 : * after handling timestamp expiration, as that may dirty the inode too.
1500 : */
1501 1374 : spin_lock(&inode->i_lock);
1502 1374 : dirty = inode->i_state & I_DIRTY;
1503 1374 : inode->i_state &= ~dirty;
1504 :
1505 : /*
1506 : * Paired with smp_mb() in __mark_inode_dirty(). This allows
1507 : * __mark_inode_dirty() to test i_state without grabbing i_lock -
1508 : * either they see the I_DIRTY bits cleared or we see the dirtied
1509 : * inode.
1510 : *
1511 : * I_DIRTY_PAGES is always cleared together above even if @mapping
1512 : * still has dirty pages. The flag is reinstated after smp_mb() if
1513 : * necessary. This guarantees that either __mark_inode_dirty()
1514 : * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1515 : */
1516 1374 : smp_mb();
1517 :
1518 1374 : if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1519 0 : inode->i_state |= I_DIRTY_PAGES;
1520 :
1521 1374 : spin_unlock(&inode->i_lock);
1522 :
1523 : /* Don't write the inode if only I_DIRTY_PAGES was set */
1524 1374 : if (dirty & ~I_DIRTY_PAGES) {
1525 1373 : int err = write_inode(inode, wbc);
1526 1373 : if (ret == 0)
1527 1373 : ret = err;
1528 : }
1529 1374 : trace_writeback_single_inode(inode, wbc, nr_to_write);
1530 1374 : return ret;
1531 : }
1532 :
1533 : /*
1534 : * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1535 : * the regular batched writeback done by the flusher threads in
1536 : * writeback_sb_inodes(). @wbc controls various aspects of the write, such as
1537 : * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1538 : *
1539 : * To prevent the inode from going away, either the caller must have a reference
1540 : * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1541 : */
1542 0 : static int writeback_single_inode(struct inode *inode,
1543 : struct writeback_control *wbc)
1544 : {
1545 0 : struct bdi_writeback *wb;
1546 0 : int ret = 0;
1547 :
1548 0 : spin_lock(&inode->i_lock);
1549 0 : if (!atomic_read(&inode->i_count))
1550 0 : WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1551 : else
1552 0 : WARN_ON(inode->i_state & I_WILL_FREE);
1553 :
1554 0 : if (inode->i_state & I_SYNC) {
1555 : /*
1556 : * Writeback is already running on the inode. For WB_SYNC_NONE,
1557 : * that's enough and we can just return. For WB_SYNC_ALL, we
1558 : * must wait for the existing writeback to complete, then do
1559 : * writeback again if there's anything left.
1560 : */
1561 0 : if (wbc->sync_mode != WB_SYNC_ALL)
1562 0 : goto out;
1563 0 : __inode_wait_for_writeback(inode);
1564 : }
1565 0 : WARN_ON(inode->i_state & I_SYNC);
1566 : /*
1567 : * If the inode is already fully clean, then there's nothing to do.
1568 : *
1569 : * For data-integrity syncs we also need to check whether any pages are
1570 : * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If
1571 : * there are any such pages, we'll need to wait for them.
1572 : */
1573 0 : if (!(inode->i_state & I_DIRTY_ALL) &&
1574 0 : (wbc->sync_mode != WB_SYNC_ALL ||
1575 0 : !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1576 0 : goto out;
1577 0 : inode->i_state |= I_SYNC;
1578 0 : wbc_attach_and_unlock_inode(wbc, inode);
1579 :
1580 0 : ret = __writeback_single_inode(inode, wbc);
1581 :
1582 0 : wbc_detach_inode(wbc);
1583 :
1584 0 : wb = inode_to_wb_and_lock_list(inode);
1585 0 : spin_lock(&inode->i_lock);
1586 : /*
1587 : * If the inode is now fully clean, then it can be safely removed from
1588 : * its writeback list (if any). Otherwise the flusher threads are
1589 : * responsible for the writeback lists.
1590 : */
1591 0 : if (!(inode->i_state & I_DIRTY_ALL))
1592 0 : inode_io_list_del_locked(inode, wb);
1593 0 : spin_unlock(&wb->list_lock);
1594 0 : inode_sync_complete(inode);
1595 0 : out:
1596 0 : spin_unlock(&inode->i_lock);
1597 0 : return ret;
1598 : }
1599 :
1600 1374 : static long writeback_chunk_size(struct bdi_writeback *wb,
1601 : struct wb_writeback_work *work)
1602 : {
1603 1374 : long pages;
1604 :
1605 : /*
1606 : * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1607 : * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1608 : * here avoids calling into writeback_inodes_wb() more than once.
1609 : *
1610 : * The intended call sequence for WB_SYNC_ALL writeback is:
1611 : *
1612 : * wb_writeback()
1613 : * writeback_sb_inodes() <== called only once
1614 : * write_cache_pages() <== called once for each inode
1615 : * (quickly) tag currently dirty pages
1616 : * (maybe slowly) sync all tagged pages
1617 : */
1618 1374 : if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1619 : pages = LONG_MAX;
1620 : else {
1621 1374 : pages = min(wb->avg_write_bandwidth / 2,
1622 : global_wb_domain.dirty_limit / DIRTY_SCOPE);
1623 1374 : pages = min(pages, work->nr_pages);
1624 1374 : pages = round_down(pages + MIN_WRITEBACK_PAGES,
1625 : MIN_WRITEBACK_PAGES);
1626 : }
1627 :
1628 1374 : return pages;
1629 : }
1630 :
1631 : /*
1632 : * Write a portion of b_io inodes which belong to @sb.
1633 : *
1634 : * Return the number of pages and/or inodes written.
1635 : *
1636 : * NOTE! This is called with wb->list_lock held, and will
1637 : * unlock and relock that for each inode it ends up doing
1638 : * IO for.
1639 : */
1640 7 : static long writeback_sb_inodes(struct super_block *sb,
1641 : struct bdi_writeback *wb,
1642 : struct wb_writeback_work *work)
1643 : {
1644 7 : struct writeback_control wbc = {
1645 7 : .sync_mode = work->sync_mode,
1646 7 : .tagged_writepages = work->tagged_writepages,
1647 7 : .for_kupdate = work->for_kupdate,
1648 7 : .for_background = work->for_background,
1649 7 : .for_sync = work->for_sync,
1650 7 : .range_cyclic = work->range_cyclic,
1651 : .range_start = 0,
1652 : .range_end = LLONG_MAX,
1653 : };
1654 7 : unsigned long start_time = jiffies;
1655 7 : long write_chunk;
1656 7 : long wrote = 0; /* count both pages and inodes */
1657 :
1658 1380 : while (!list_empty(&wb->b_io)) {
1659 1374 : struct inode *inode = wb_inode(wb->b_io.prev);
1660 1374 : struct bdi_writeback *tmp_wb;
1661 :
1662 1374 : if (inode->i_sb != sb) {
1663 0 : if (work->sb) {
1664 : /*
1665 : * We only want to write back data for this
1666 : * superblock, move all inodes not belonging
1667 : * to it back onto the dirty list.
1668 : */
1669 0 : redirty_tail(inode, wb);
1670 0 : continue;
1671 : }
1672 :
1673 : /*
1674 : * The inode belongs to a different superblock.
1675 : * Bounce back to the caller to unpin this and
1676 : * pin the next superblock.
1677 : */
1678 : break;
1679 : }
1680 :
1681 : /*
1682 : * Don't bother with new inodes or inodes being freed, first
1683 : * kind does not need periodic writeout yet, and for the latter
1684 : * kind writeout is handled by the freer.
1685 : */
1686 1374 : spin_lock(&inode->i_lock);
1687 1374 : if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1688 0 : redirty_tail_locked(inode, wb);
1689 0 : spin_unlock(&inode->i_lock);
1690 0 : continue;
1691 : }
1692 1374 : if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1693 : /*
1694 : * If this inode is locked for writeback and we are not
1695 : * doing writeback-for-data-integrity, move it to
1696 : * b_more_io so that writeback can proceed with the
1697 : * other inodes on s_io.
1698 : *
1699 : * We'll have another go at writing back this inode
1700 : * when we completed a full scan of b_io.
1701 : */
1702 0 : spin_unlock(&inode->i_lock);
1703 0 : requeue_io(inode, wb);
1704 0 : trace_writeback_sb_inodes_requeue(inode);
1705 0 : continue;
1706 : }
1707 1374 : spin_unlock(&wb->list_lock);
1708 :
1709 : /*
1710 : * We already requeued the inode if it had I_SYNC set and we
1711 : * are doing WB_SYNC_NONE writeback. So this catches only the
1712 : * WB_SYNC_ALL case.
1713 : */
1714 1374 : if (inode->i_state & I_SYNC) {
1715 : /* Wait for I_SYNC. This function drops i_lock... */
1716 0 : inode_sleep_on_writeback(inode);
1717 : /* Inode may be gone, start again */
1718 0 : spin_lock(&wb->list_lock);
1719 0 : continue;
1720 : }
1721 1374 : inode->i_state |= I_SYNC;
1722 1374 : wbc_attach_and_unlock_inode(&wbc, inode);
1723 :
1724 1374 : write_chunk = writeback_chunk_size(wb, work);
1725 1374 : wbc.nr_to_write = write_chunk;
1726 1374 : wbc.pages_skipped = 0;
1727 :
1728 : /*
1729 : * We use I_SYNC to pin the inode in memory. While it is set
1730 : * evict_inode() will wait so the inode cannot be freed.
1731 : */
1732 1374 : __writeback_single_inode(inode, &wbc);
1733 :
1734 1374 : wbc_detach_inode(&wbc);
1735 1374 : work->nr_pages -= write_chunk - wbc.nr_to_write;
1736 1374 : wrote += write_chunk - wbc.nr_to_write;
1737 :
1738 1374 : if (need_resched()) {
1739 : /*
1740 : * We're trying to balance between building up a nice
1741 : * long list of IOs to improve our merge rate, and
1742 : * getting those IOs out quickly for anyone throttling
1743 : * in balance_dirty_pages(). cond_resched() doesn't
1744 : * unplug, so get our IOs out the door before we
1745 : * give up the CPU.
1746 : */
1747 2 : blk_flush_plug(current);
1748 2 : cond_resched();
1749 : }
1750 :
1751 : /*
1752 : * Requeue @inode if still dirty. Be careful as @inode may
1753 : * have been switched to another wb in the meantime.
1754 : */
1755 1374 : tmp_wb = inode_to_wb_and_lock_list(inode);
1756 1374 : spin_lock(&inode->i_lock);
1757 1374 : if (!(inode->i_state & I_DIRTY_ALL))
1758 1374 : wrote++;
1759 1374 : requeue_inode(inode, tmp_wb, &wbc);
1760 1374 : inode_sync_complete(inode);
1761 1374 : spin_unlock(&inode->i_lock);
1762 :
1763 1374 : if (unlikely(tmp_wb != wb)) {
1764 0 : spin_unlock(&tmp_wb->list_lock);
1765 0 : spin_lock(&wb->list_lock);
1766 : }
1767 :
1768 : /*
1769 : * bail out to wb_writeback() often enough to check
1770 : * background threshold and other termination conditions.
1771 : */
1772 1374 : if (wrote) {
1773 1374 : if (time_is_before_jiffies(start_time + HZ / 10UL))
1774 : break;
1775 1373 : if (work->nr_pages <= 0)
1776 : break;
1777 : }
1778 : }
1779 7 : return wrote;
1780 : }
1781 :
1782 15 : static long __writeback_inodes_wb(struct bdi_writeback *wb,
1783 : struct wb_writeback_work *work)
1784 : {
1785 15 : unsigned long start_time = jiffies;
1786 15 : long wrote = 0;
1787 :
1788 21 : while (!list_empty(&wb->b_io)) {
1789 7 : struct inode *inode = wb_inode(wb->b_io.prev);
1790 7 : struct super_block *sb = inode->i_sb;
1791 :
1792 7 : if (!trylock_super(sb)) {
1793 : /*
1794 : * trylock_super() may fail consistently due to
1795 : * s_umount being grabbed by someone else. Don't use
1796 : * requeue_io() to avoid busy retrying the inode/sb.
1797 : */
1798 0 : redirty_tail(inode, wb);
1799 0 : continue;
1800 : }
1801 7 : wrote += writeback_sb_inodes(sb, wb, work);
1802 7 : up_read(&sb->s_umount);
1803 :
1804 : /* refer to the same tests at the end of writeback_sb_inodes */
1805 7 : if (wrote) {
1806 7 : if (time_is_before_jiffies(start_time + HZ / 10UL))
1807 : break;
1808 6 : if (work->nr_pages <= 0)
1809 : break;
1810 : }
1811 : }
1812 : /* Leave any unwritten inodes on b_io */
1813 15 : return wrote;
1814 : }
1815 :
1816 0 : static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1817 : enum wb_reason reason)
1818 : {
1819 0 : struct wb_writeback_work work = {
1820 : .nr_pages = nr_pages,
1821 : .sync_mode = WB_SYNC_NONE,
1822 : .range_cyclic = 1,
1823 : .reason = reason,
1824 : };
1825 0 : struct blk_plug plug;
1826 :
1827 0 : blk_start_plug(&plug);
1828 0 : spin_lock(&wb->list_lock);
1829 0 : if (list_empty(&wb->b_io))
1830 0 : queue_io(wb, &work, jiffies);
1831 0 : __writeback_inodes_wb(wb, &work);
1832 0 : spin_unlock(&wb->list_lock);
1833 0 : blk_finish_plug(&plug);
1834 :
1835 0 : return nr_pages - work.nr_pages;
1836 : }
1837 :
1838 : /*
1839 : * Explicit flushing or periodic writeback of "old" data.
1840 : *
1841 : * Define "old": the first time one of an inode's pages is dirtied, we mark the
1842 : * dirtying-time in the inode's address_space. So this periodic writeback code
1843 : * just walks the superblock inode list, writing back any inodes which are
1844 : * older than a specific point in time.
1845 : *
1846 : * Try to run once per dirty_writeback_interval. But if a writeback event
1847 : * takes longer than a dirty_writeback_interval interval, then leave a
1848 : * one-second gap.
1849 : *
1850 : * dirtied_before takes precedence over nr_to_write. So we'll only write back
1851 : * all dirty pages if they are all attached to "old" mappings.
1852 : */
1853 8 : static long wb_writeback(struct bdi_writeback *wb,
1854 : struct wb_writeback_work *work)
1855 : {
1856 8 : unsigned long wb_start = jiffies;
1857 8 : long nr_pages = work->nr_pages;
1858 8 : unsigned long dirtied_before = jiffies;
1859 8 : struct inode *inode;
1860 8 : long progress;
1861 8 : struct blk_plug plug;
1862 :
1863 8 : blk_start_plug(&plug);
1864 8 : spin_lock(&wb->list_lock);
1865 15 : for (;;) {
1866 : /*
1867 : * Stop writeback when nr_pages has been consumed
1868 : */
1869 15 : if (work->nr_pages <= 0)
1870 : break;
1871 :
1872 : /*
1873 : * Background writeout and kupdate-style writeback may
1874 : * run forever. Stop them if there is other work to do
1875 : * so that e.g. sync can proceed. They'll be restarted
1876 : * after the other works are all done.
1877 : */
1878 15 : if ((work->for_background || work->for_kupdate) &&
1879 15 : !list_empty(&wb->work_list))
1880 : break;
1881 :
1882 : /*
1883 : * For background writeout, stop when we are below the
1884 : * background dirty threshold
1885 : */
1886 15 : if (work->for_background && !wb_over_bg_thresh(wb))
1887 : break;
1888 :
1889 : /*
1890 : * Kupdate and background works are special and we want to
1891 : * include all inodes that need writing. Livelock avoidance is
1892 : * handled by these works yielding to any other work so we are
1893 : * safe.
1894 : */
1895 15 : if (work->for_kupdate) {
1896 30 : dirtied_before = jiffies -
1897 30 : msecs_to_jiffies(dirty_expire_interval * 10);
1898 0 : } else if (work->for_background)
1899 0 : dirtied_before = jiffies;
1900 :
1901 15 : trace_writeback_start(wb, work);
1902 15 : if (list_empty(&wb->b_io))
1903 14 : queue_io(wb, work, dirtied_before);
1904 15 : if (work->sb)
1905 0 : progress = writeback_sb_inodes(work->sb, wb, work);
1906 : else
1907 15 : progress = __writeback_inodes_wb(wb, work);
1908 15 : trace_writeback_written(wb, work);
1909 :
1910 15 : wb_update_bandwidth(wb, wb_start);
1911 :
1912 : /*
1913 : * Did we write something? Try for more
1914 : *
1915 : * Dirty inodes are moved to b_io for writeback in batches.
1916 : * The completion of the current batch does not necessarily
1917 : * mean the overall work is done. So we keep looping as long
1918 : * as made some progress on cleaning pages or inodes.
1919 : */
1920 15 : if (progress)
1921 7 : continue;
1922 : /*
1923 : * No more inodes for IO, bail
1924 : */
1925 8 : if (list_empty(&wb->b_more_io))
1926 : break;
1927 : /*
1928 : * Nothing written. Wait for some inode to
1929 : * become available for writeback. Otherwise
1930 : * we'll just busyloop.
1931 : */
1932 0 : trace_writeback_wait(wb, work);
1933 0 : inode = wb_inode(wb->b_more_io.prev);
1934 0 : spin_lock(&inode->i_lock);
1935 0 : spin_unlock(&wb->list_lock);
1936 : /* This function drops i_lock... */
1937 0 : inode_sleep_on_writeback(inode);
1938 15 : spin_lock(&wb->list_lock);
1939 : }
1940 8 : spin_unlock(&wb->list_lock);
1941 8 : blk_finish_plug(&plug);
1942 :
1943 8 : return nr_pages - work->nr_pages;
1944 : }
1945 :
1946 : /*
1947 : * Return the next wb_writeback_work struct that hasn't been processed yet.
1948 : */
1949 8 : static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1950 : {
1951 8 : struct wb_writeback_work *work = NULL;
1952 :
1953 8 : spin_lock_bh(&wb->work_lock);
1954 8 : if (!list_empty(&wb->work_list)) {
1955 0 : work = list_entry(wb->work_list.next,
1956 : struct wb_writeback_work, list);
1957 0 : list_del_init(&work->list);
1958 : }
1959 8 : spin_unlock_bh(&wb->work_lock);
1960 8 : return work;
1961 : }
1962 :
1963 8 : static long wb_check_background_flush(struct bdi_writeback *wb)
1964 : {
1965 8 : if (wb_over_bg_thresh(wb)) {
1966 :
1967 0 : struct wb_writeback_work work = {
1968 : .nr_pages = LONG_MAX,
1969 : .sync_mode = WB_SYNC_NONE,
1970 : .for_background = 1,
1971 : .range_cyclic = 1,
1972 : .reason = WB_REASON_BACKGROUND,
1973 : };
1974 :
1975 0 : return wb_writeback(wb, &work);
1976 : }
1977 :
1978 : return 0;
1979 : }
1980 :
1981 8 : static long wb_check_old_data_flush(struct bdi_writeback *wb)
1982 : {
1983 8 : unsigned long expired;
1984 8 : long nr_pages;
1985 :
1986 : /*
1987 : * When set to zero, disable periodic writeback
1988 : */
1989 8 : if (!dirty_writeback_interval)
1990 : return 0;
1991 :
1992 16 : expired = wb->last_old_flush +
1993 8 : msecs_to_jiffies(dirty_writeback_interval * 10);
1994 8 : if (time_before(jiffies, expired))
1995 : return 0;
1996 :
1997 8 : wb->last_old_flush = jiffies;
1998 8 : nr_pages = get_nr_dirty_pages();
1999 :
2000 8 : if (nr_pages) {
2001 8 : struct wb_writeback_work work = {
2002 : .nr_pages = nr_pages,
2003 : .sync_mode = WB_SYNC_NONE,
2004 : .for_kupdate = 1,
2005 : .range_cyclic = 1,
2006 : .reason = WB_REASON_PERIODIC,
2007 : };
2008 :
2009 8 : return wb_writeback(wb, &work);
2010 : }
2011 :
2012 : return 0;
2013 : }
2014 :
2015 8 : static long wb_check_start_all(struct bdi_writeback *wb)
2016 : {
2017 8 : long nr_pages;
2018 :
2019 8 : if (!test_bit(WB_start_all, &wb->state))
2020 : return 0;
2021 :
2022 0 : nr_pages = get_nr_dirty_pages();
2023 0 : if (nr_pages) {
2024 0 : struct wb_writeback_work work = {
2025 0 : .nr_pages = wb_split_bdi_pages(wb, nr_pages),
2026 : .sync_mode = WB_SYNC_NONE,
2027 : .range_cyclic = 1,
2028 0 : .reason = wb->start_all_reason,
2029 : };
2030 :
2031 0 : nr_pages = wb_writeback(wb, &work);
2032 : }
2033 :
2034 0 : clear_bit(WB_start_all, &wb->state);
2035 0 : return nr_pages;
2036 : }
2037 :
2038 :
2039 : /*
2040 : * Retrieve work items and do the writeback they describe
2041 : */
2042 8 : static long wb_do_writeback(struct bdi_writeback *wb)
2043 : {
2044 8 : struct wb_writeback_work *work;
2045 8 : long wrote = 0;
2046 :
2047 8 : set_bit(WB_writeback_running, &wb->state);
2048 8 : while ((work = get_next_work_item(wb)) != NULL) {
2049 0 : trace_writeback_exec(wb, work);
2050 0 : wrote += wb_writeback(wb, work);
2051 0 : finish_writeback_work(wb, work);
2052 : }
2053 :
2054 : /*
2055 : * Check for a flush-everything request
2056 : */
2057 8 : wrote += wb_check_start_all(wb);
2058 :
2059 : /*
2060 : * Check for periodic writeback, kupdated() style
2061 : */
2062 8 : wrote += wb_check_old_data_flush(wb);
2063 8 : wrote += wb_check_background_flush(wb);
2064 8 : clear_bit(WB_writeback_running, &wb->state);
2065 :
2066 8 : return wrote;
2067 : }
2068 :
2069 : /*
2070 : * Handle writeback of dirty data for the device backed by this bdi. Also
2071 : * reschedules periodically and does kupdated style flushing.
2072 : */
2073 8 : void wb_workfn(struct work_struct *work)
2074 : {
2075 8 : struct bdi_writeback *wb = container_of(to_delayed_work(work),
2076 : struct bdi_writeback, dwork);
2077 8 : long pages_written;
2078 :
2079 8 : set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2080 8 : current->flags |= PF_SWAPWRITE;
2081 :
2082 8 : if (likely(!current_is_workqueue_rescuer() ||
2083 : !test_bit(WB_registered, &wb->state))) {
2084 : /*
2085 : * The normal path. Keep writing back @wb until its
2086 : * work_list is empty. Note that this path is also taken
2087 : * if @wb is shutting down even when we're running off the
2088 : * rescuer as work_list needs to be drained.
2089 : */
2090 8 : do {
2091 8 : pages_written = wb_do_writeback(wb);
2092 8 : trace_writeback_pages_written(pages_written);
2093 8 : } while (!list_empty(&wb->work_list));
2094 : } else {
2095 : /*
2096 : * bdi_wq can't get enough workers and we're running off
2097 : * the emergency worker. Don't hog it. Hopefully, 1024 is
2098 : * enough for efficient IO.
2099 : */
2100 0 : pages_written = writeback_inodes_wb(wb, 1024,
2101 : WB_REASON_FORKER_THREAD);
2102 0 : trace_writeback_pages_written(pages_written);
2103 : }
2104 :
2105 8 : if (!list_empty(&wb->work_list))
2106 0 : wb_wakeup(wb);
2107 8 : else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2108 8 : wb_wakeup_delayed(wb);
2109 :
2110 8 : current->flags &= ~PF_SWAPWRITE;
2111 8 : }
2112 :
2113 : /*
2114 : * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
2115 : * write back the whole world.
2116 : */
2117 0 : static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2118 : enum wb_reason reason)
2119 : {
2120 0 : struct bdi_writeback *wb;
2121 :
2122 0 : if (!bdi_has_dirty_io(bdi))
2123 : return;
2124 :
2125 0 : list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2126 0 : wb_start_writeback(wb, reason);
2127 : }
2128 :
2129 0 : void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2130 : enum wb_reason reason)
2131 : {
2132 0 : rcu_read_lock();
2133 0 : __wakeup_flusher_threads_bdi(bdi, reason);
2134 0 : rcu_read_unlock();
2135 0 : }
2136 :
2137 : /*
2138 : * Wakeup the flusher threads to start writeback of all currently dirty pages
2139 : */
2140 0 : void wakeup_flusher_threads(enum wb_reason reason)
2141 : {
2142 0 : struct backing_dev_info *bdi;
2143 :
2144 : /*
2145 : * If we are expecting writeback progress we must submit plugged IO.
2146 : */
2147 0 : if (blk_needs_flush_plug(current))
2148 0 : blk_schedule_flush_plug(current);
2149 :
2150 0 : rcu_read_lock();
2151 0 : list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2152 0 : __wakeup_flusher_threads_bdi(bdi, reason);
2153 0 : rcu_read_unlock();
2154 0 : }
2155 :
2156 : /*
2157 : * Wake up bdi's periodically to make sure dirtytime inodes gets
2158 : * written back periodically. We deliberately do *not* check the
2159 : * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2160 : * kernel to be constantly waking up once there are any dirtytime
2161 : * inodes on the system. So instead we define a separate delayed work
2162 : * function which gets called much more rarely. (By default, only
2163 : * once every 12 hours.)
2164 : *
2165 : * If there is any other write activity going on in the file system,
2166 : * this function won't be necessary. But if the only thing that has
2167 : * happened on the file system is a dirtytime inode caused by an atime
2168 : * update, we need this infrastructure below to make sure that inode
2169 : * eventually gets pushed out to disk.
2170 : */
2171 : static void wakeup_dirtytime_writeback(struct work_struct *w);
2172 : static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2173 :
2174 0 : static void wakeup_dirtytime_writeback(struct work_struct *w)
2175 : {
2176 0 : struct backing_dev_info *bdi;
2177 :
2178 0 : rcu_read_lock();
2179 0 : list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2180 0 : struct bdi_writeback *wb;
2181 :
2182 0 : list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2183 0 : if (!list_empty(&wb->b_dirty_time))
2184 0 : wb_wakeup(wb);
2185 : }
2186 0 : rcu_read_unlock();
2187 0 : schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2188 0 : }
2189 :
2190 1 : static int __init start_dirtytime_writeback(void)
2191 : {
2192 1 : schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2193 1 : return 0;
2194 : }
2195 : __initcall(start_dirtytime_writeback);
2196 :
2197 0 : int dirtytime_interval_handler(struct ctl_table *table, int write,
2198 : void *buffer, size_t *lenp, loff_t *ppos)
2199 : {
2200 0 : int ret;
2201 :
2202 0 : ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2203 0 : if (ret == 0 && write)
2204 0 : mod_delayed_work(system_wq, &dirtytime_work, 0);
2205 0 : return ret;
2206 : }
2207 :
2208 0 : static noinline void block_dump___mark_inode_dirty(struct inode *inode)
2209 : {
2210 0 : if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
2211 0 : struct dentry *dentry;
2212 0 : const char *name = "?";
2213 :
2214 0 : dentry = d_find_alias(inode);
2215 0 : if (dentry) {
2216 0 : spin_lock(&dentry->d_lock);
2217 0 : name = (const char *) dentry->d_name.name;
2218 : }
2219 0 : printk(KERN_DEBUG
2220 : "%s(%d): dirtied inode %lu (%s) on %s\n",
2221 0 : current->comm, task_pid_nr(current), inode->i_ino,
2222 0 : name, inode->i_sb->s_id);
2223 0 : if (dentry) {
2224 0 : spin_unlock(&dentry->d_lock);
2225 0 : dput(dentry);
2226 : }
2227 : }
2228 0 : }
2229 :
2230 : /**
2231 : * __mark_inode_dirty - internal function to mark an inode dirty
2232 : *
2233 : * @inode: inode to mark
2234 : * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of
2235 : * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2236 : * with I_DIRTY_PAGES.
2237 : *
2238 : * Mark an inode as dirty. We notify the filesystem, then update the inode's
2239 : * dirty flags. Then, if needed we add the inode to the appropriate dirty list.
2240 : *
2241 : * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2242 : * instead of calling this directly.
2243 : *
2244 : * CAREFUL! We only add the inode to the dirty list if it is hashed or if it
2245 : * refers to a blockdev. Unhashed inodes will never be added to the dirty list
2246 : * even if they are later hashed, as they will have been marked dirty already.
2247 : *
2248 : * In short, ensure you hash any inodes _before_ you start marking them dirty.
2249 : *
2250 : * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2251 : * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2252 : * the kernel-internal blockdev inode represents the dirtying time of the
2253 : * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2254 : * page->mapping->host, so the page-dirtying time is recorded in the internal
2255 : * blockdev inode.
2256 : */
2257 10239 : void __mark_inode_dirty(struct inode *inode, int flags)
2258 : {
2259 10239 : struct super_block *sb = inode->i_sb;
2260 10239 : int dirtytime = 0;
2261 :
2262 10239 : trace_writeback_mark_inode_dirty(inode, flags);
2263 :
2264 10239 : if (flags & I_DIRTY_INODE) {
2265 : /*
2266 : * Notify the filesystem about the inode being dirtied, so that
2267 : * (if needed) it can update on-disk fields and journal the
2268 : * inode. This is only needed when the inode itself is being
2269 : * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not
2270 : * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2271 : */
2272 7541 : trace_writeback_dirty_inode_start(inode, flags);
2273 7541 : if (sb->s_op->dirty_inode)
2274 5007 : sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE);
2275 7541 : trace_writeback_dirty_inode(inode, flags);
2276 :
2277 : /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2278 7541 : flags &= ~I_DIRTY_TIME;
2279 : } else {
2280 : /*
2281 : * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2282 : * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2283 : * in one call to __mark_inode_dirty().)
2284 : */
2285 2698 : dirtytime = flags & I_DIRTY_TIME;
2286 2698 : WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2287 : }
2288 :
2289 : /*
2290 : * Paired with smp_mb() in __writeback_single_inode() for the
2291 : * following lockless i_state test. See there for details.
2292 : */
2293 10239 : smp_mb();
2294 :
2295 10239 : if (((inode->i_state & flags) == flags) ||
2296 0 : (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2297 : return;
2298 :
2299 3231 : if (unlikely(block_dump))
2300 0 : block_dump___mark_inode_dirty(inode);
2301 :
2302 3231 : spin_lock(&inode->i_lock);
2303 3231 : if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2304 0 : goto out_unlock_inode;
2305 3231 : if ((inode->i_state & flags) != flags) {
2306 3231 : const int was_dirty = inode->i_state & I_DIRTY;
2307 :
2308 3231 : inode_attach_wb(inode, NULL);
2309 :
2310 : /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2311 3231 : if (flags & I_DIRTY_INODE)
2312 2881 : inode->i_state &= ~I_DIRTY_TIME;
2313 3231 : inode->i_state |= flags;
2314 :
2315 : /*
2316 : * If the inode is queued for writeback by flush worker, just
2317 : * update its dirty state. Once the flush worker is done with
2318 : * the inode it will place it on the appropriate superblock
2319 : * list, based upon its state.
2320 : */
2321 3231 : if (inode->i_state & I_SYNC_QUEUED)
2322 0 : goto out_unlock_inode;
2323 :
2324 : /*
2325 : * Only add valid (hashed) inodes to the superblock's
2326 : * dirty list. Add blockdev inodes as well.
2327 : */
2328 3231 : if (!S_ISBLK(inode->i_mode)) {
2329 3219 : if (inode_unhashed(inode))
2330 416 : goto out_unlock_inode;
2331 : }
2332 2815 : if (inode->i_state & I_FREEING)
2333 27 : goto out_unlock_inode;
2334 :
2335 : /*
2336 : * If the inode was already on b_dirty/b_io/b_more_io, don't
2337 : * reposition it (that would break b_dirty time-ordering).
2338 : */
2339 2788 : if (!was_dirty) {
2340 2157 : struct bdi_writeback *wb;
2341 2157 : struct list_head *dirty_list;
2342 2157 : bool wakeup_bdi = false;
2343 :
2344 2157 : wb = locked_inode_to_wb_and_lock_list(inode);
2345 :
2346 2157 : inode->dirtied_when = jiffies;
2347 2157 : if (dirtytime)
2348 0 : inode->dirtied_time_when = jiffies;
2349 :
2350 2157 : if (inode->i_state & I_DIRTY)
2351 2157 : dirty_list = &wb->b_dirty;
2352 : else
2353 0 : dirty_list = &wb->b_dirty_time;
2354 :
2355 2157 : wakeup_bdi = inode_io_list_move_locked(inode, wb,
2356 : dirty_list);
2357 :
2358 2157 : spin_unlock(&wb->list_lock);
2359 2157 : trace_writeback_dirty_inode_enqueue(inode);
2360 :
2361 : /*
2362 : * If this is the first dirty inode for this bdi,
2363 : * we have to wake-up the corresponding bdi thread
2364 : * to make sure background write-back happens
2365 : * later.
2366 : */
2367 2157 : if (wakeup_bdi &&
2368 2 : (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2369 1 : wb_wakeup_delayed(wb);
2370 2157 : return;
2371 : }
2372 : }
2373 631 : out_unlock_inode:
2374 1074 : spin_unlock(&inode->i_lock);
2375 : }
2376 : EXPORT_SYMBOL(__mark_inode_dirty);
2377 :
2378 : /*
2379 : * The @s_sync_lock is used to serialise concurrent sync operations
2380 : * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2381 : * Concurrent callers will block on the s_sync_lock rather than doing contending
2382 : * walks. The queueing maintains sync(2) required behaviour as all the IO that
2383 : * has been issued up to the time this function is enter is guaranteed to be
2384 : * completed by the time we have gained the lock and waited for all IO that is
2385 : * in progress regardless of the order callers are granted the lock.
2386 : */
2387 0 : static void wait_sb_inodes(struct super_block *sb)
2388 : {
2389 0 : LIST_HEAD(sync_list);
2390 :
2391 : /*
2392 : * We need to be protected against the filesystem going from
2393 : * r/o to r/w or vice versa.
2394 : */
2395 0 : WARN_ON(!rwsem_is_locked(&sb->s_umount));
2396 :
2397 0 : mutex_lock(&sb->s_sync_lock);
2398 :
2399 : /*
2400 : * Splice the writeback list onto a temporary list to avoid waiting on
2401 : * inodes that have started writeback after this point.
2402 : *
2403 : * Use rcu_read_lock() to keep the inodes around until we have a
2404 : * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2405 : * the local list because inodes can be dropped from either by writeback
2406 : * completion.
2407 : */
2408 0 : rcu_read_lock();
2409 0 : spin_lock_irq(&sb->s_inode_wblist_lock);
2410 0 : list_splice_init(&sb->s_inodes_wb, &sync_list);
2411 :
2412 : /*
2413 : * Data integrity sync. Must wait for all pages under writeback, because
2414 : * there may have been pages dirtied before our sync call, but which had
2415 : * writeout started before we write it out. In which case, the inode
2416 : * may not be on the dirty list, but we still have to wait for that
2417 : * writeout.
2418 : */
2419 0 : while (!list_empty(&sync_list)) {
2420 0 : struct inode *inode = list_first_entry(&sync_list, struct inode,
2421 : i_wb_list);
2422 0 : struct address_space *mapping = inode->i_mapping;
2423 :
2424 : /*
2425 : * Move each inode back to the wb list before we drop the lock
2426 : * to preserve consistency between i_wb_list and the mapping
2427 : * writeback tag. Writeback completion is responsible to remove
2428 : * the inode from either list once the writeback tag is cleared.
2429 : */
2430 0 : list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2431 :
2432 : /*
2433 : * The mapping can appear untagged while still on-list since we
2434 : * do not have the mapping lock. Skip it here, wb completion
2435 : * will remove it.
2436 : */
2437 0 : if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2438 0 : continue;
2439 :
2440 0 : spin_unlock_irq(&sb->s_inode_wblist_lock);
2441 :
2442 0 : spin_lock(&inode->i_lock);
2443 0 : if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2444 0 : spin_unlock(&inode->i_lock);
2445 :
2446 0 : spin_lock_irq(&sb->s_inode_wblist_lock);
2447 0 : continue;
2448 : }
2449 0 : __iget(inode);
2450 0 : spin_unlock(&inode->i_lock);
2451 0 : rcu_read_unlock();
2452 :
2453 : /*
2454 : * We keep the error status of individual mapping so that
2455 : * applications can catch the writeback error using fsync(2).
2456 : * See filemap_fdatawait_keep_errors() for details.
2457 : */
2458 0 : filemap_fdatawait_keep_errors(mapping);
2459 :
2460 0 : cond_resched();
2461 :
2462 0 : iput(inode);
2463 :
2464 0 : rcu_read_lock();
2465 0 : spin_lock_irq(&sb->s_inode_wblist_lock);
2466 : }
2467 0 : spin_unlock_irq(&sb->s_inode_wblist_lock);
2468 0 : rcu_read_unlock();
2469 0 : mutex_unlock(&sb->s_sync_lock);
2470 0 : }
2471 :
2472 101 : static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2473 : enum wb_reason reason, bool skip_if_busy)
2474 : {
2475 101 : struct backing_dev_info *bdi = sb->s_bdi;
2476 101 : DEFINE_WB_COMPLETION(done, bdi);
2477 101 : struct wb_writeback_work work = {
2478 : .sb = sb,
2479 : .sync_mode = WB_SYNC_NONE,
2480 : .tagged_writepages = 1,
2481 : .done = &done,
2482 : .nr_pages = nr,
2483 : .reason = reason,
2484 : };
2485 :
2486 101 : if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2487 101 : return;
2488 0 : WARN_ON(!rwsem_is_locked(&sb->s_umount));
2489 :
2490 0 : bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2491 0 : wb_wait_for_completion(&done);
2492 : }
2493 :
2494 : /**
2495 : * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2496 : * @sb: the superblock
2497 : * @nr: the number of pages to write
2498 : * @reason: reason why some writeback work initiated
2499 : *
2500 : * Start writeback on some inodes on this super_block. No guarantees are made
2501 : * on how many (if any) will be written, and this function does not wait
2502 : * for IO completion of submitted IO.
2503 : */
2504 101 : void writeback_inodes_sb_nr(struct super_block *sb,
2505 : unsigned long nr,
2506 : enum wb_reason reason)
2507 : {
2508 0 : __writeback_inodes_sb_nr(sb, nr, reason, false);
2509 0 : }
2510 : EXPORT_SYMBOL(writeback_inodes_sb_nr);
2511 :
2512 : /**
2513 : * writeback_inodes_sb - writeback dirty inodes from given super_block
2514 : * @sb: the superblock
2515 : * @reason: reason why some writeback work was initiated
2516 : *
2517 : * Start writeback on some inodes on this super_block. No guarantees are made
2518 : * on how many (if any) will be written, and this function does not wait
2519 : * for IO completion of submitted IO.
2520 : */
2521 101 : void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2522 : {
2523 101 : return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2524 : }
2525 : EXPORT_SYMBOL(writeback_inodes_sb);
2526 :
2527 : /**
2528 : * try_to_writeback_inodes_sb - try to start writeback if none underway
2529 : * @sb: the superblock
2530 : * @reason: reason why some writeback work was initiated
2531 : *
2532 : * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2533 : */
2534 0 : void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2535 : {
2536 0 : if (!down_read_trylock(&sb->s_umount))
2537 : return;
2538 :
2539 0 : __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2540 0 : up_read(&sb->s_umount);
2541 : }
2542 : EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2543 :
2544 : /**
2545 : * sync_inodes_sb - sync sb inode pages
2546 : * @sb: the superblock
2547 : *
2548 : * This function writes and waits on any dirty inode belonging to this
2549 : * super_block.
2550 : */
2551 101 : void sync_inodes_sb(struct super_block *sb)
2552 : {
2553 101 : struct backing_dev_info *bdi = sb->s_bdi;
2554 101 : DEFINE_WB_COMPLETION(done, bdi);
2555 101 : struct wb_writeback_work work = {
2556 : .sb = sb,
2557 : .sync_mode = WB_SYNC_ALL,
2558 : .nr_pages = LONG_MAX,
2559 : .range_cyclic = 0,
2560 : .done = &done,
2561 : .reason = WB_REASON_SYNC,
2562 : .for_sync = 1,
2563 : };
2564 :
2565 : /*
2566 : * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2567 : * inodes under writeback and I_DIRTY_TIME inodes ignored by
2568 : * bdi_has_dirty() need to be written out too.
2569 : */
2570 101 : if (bdi == &noop_backing_dev_info)
2571 101 : return;
2572 0 : WARN_ON(!rwsem_is_locked(&sb->s_umount));
2573 :
2574 : /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2575 0 : bdi_down_write_wb_switch_rwsem(bdi);
2576 0 : bdi_split_work_to_wbs(bdi, &work, false);
2577 0 : wb_wait_for_completion(&done);
2578 0 : bdi_up_write_wb_switch_rwsem(bdi);
2579 :
2580 0 : wait_sb_inodes(sb);
2581 : }
2582 : EXPORT_SYMBOL(sync_inodes_sb);
2583 :
2584 : /**
2585 : * write_inode_now - write an inode to disk
2586 : * @inode: inode to write to disk
2587 : * @sync: whether the write should be synchronous or not
2588 : *
2589 : * This function commits an inode to disk immediately if it is dirty. This is
2590 : * primarily needed by knfsd.
2591 : *
2592 : * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2593 : */
2594 0 : int write_inode_now(struct inode *inode, int sync)
2595 : {
2596 0 : struct writeback_control wbc = {
2597 : .nr_to_write = LONG_MAX,
2598 0 : .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2599 : .range_start = 0,
2600 : .range_end = LLONG_MAX,
2601 : };
2602 :
2603 0 : if (!mapping_can_writeback(inode->i_mapping))
2604 0 : wbc.nr_to_write = 0;
2605 :
2606 0 : might_sleep();
2607 0 : return writeback_single_inode(inode, &wbc);
2608 : }
2609 : EXPORT_SYMBOL(write_inode_now);
2610 :
2611 : /**
2612 : * sync_inode - write an inode and its pages to disk.
2613 : * @inode: the inode to sync
2614 : * @wbc: controls the writeback mode
2615 : *
2616 : * sync_inode() will write an inode and its pages to disk. It will also
2617 : * correctly update the inode on its superblock's dirty inode lists and will
2618 : * update inode->i_state.
2619 : *
2620 : * The caller must have a ref on the inode.
2621 : */
2622 0 : int sync_inode(struct inode *inode, struct writeback_control *wbc)
2623 : {
2624 0 : return writeback_single_inode(inode, wbc);
2625 : }
2626 : EXPORT_SYMBOL(sync_inode);
2627 :
2628 : /**
2629 : * sync_inode_metadata - write an inode to disk
2630 : * @inode: the inode to sync
2631 : * @wait: wait for I/O to complete.
2632 : *
2633 : * Write an inode to disk and adjust its dirty state after completion.
2634 : *
2635 : * Note: only writes the actual inode, no associated data or other metadata.
2636 : */
2637 0 : int sync_inode_metadata(struct inode *inode, int wait)
2638 : {
2639 0 : struct writeback_control wbc = {
2640 0 : .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2641 : .nr_to_write = 0, /* metadata-only */
2642 : };
2643 :
2644 0 : return sync_inode(inode, &wbc);
2645 : }
2646 : EXPORT_SYMBOL(sync_inode_metadata);
|
