blob: 518c6294bf6c0ef965e9f56baa4e11d2eccc0165 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/fs-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
10 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16#include <linux/kernel.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050017#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090019#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/sched.h>
21#include <linux/fs.h>
22#include <linux/mm.h>
Wu Fengguangbc31b862012-01-07 20:41:55 -060023#include <linux/pagemap.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020024#include <linux/kthread.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
Dave Chinner455b2862010-07-07 13:24:06 +100028#include <linux/tracepoint.h>
Al Viro719ea2f2013-09-29 11:24:49 -040029#include <linux/device.h>
Tejun Heo21c63212015-05-28 14:50:49 -040030#include <linux/memcontrol.h>
David Howells07f3f052006-09-30 20:52:18 +020031#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
Jens Axboed0bceac2009-05-18 08:20:32 +020033/*
Wu Fengguangbc31b862012-01-07 20:41:55 -060034 * 4MB minimal write chunk size
35 */
36#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
37
Tejun Heocc395d72015-05-22 17:13:58 -040038struct wb_completion {
39 atomic_t cnt;
40};
41
Wu Fengguangbc31b862012-01-07 20:41:55 -060042/*
Jens Axboec4a77a62009-09-16 15:18:25 +020043 * Passed into wb_writeback(), essentially a subset of writeback_control
44 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020045struct wb_writeback_work {
Jens Axboec4a77a62009-09-16 15:18:25 +020046 long nr_pages;
47 struct super_block *sb;
Jan Kara0dc83bd2014-02-21 11:19:04 +010048 unsigned long *older_than_this;
Jens Axboec4a77a62009-09-16 15:18:25 +020049 enum writeback_sync_modes sync_mode;
Wu Fengguang6e6938b2010-06-06 10:38:15 -060050 unsigned int tagged_writepages:1;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050051 unsigned int for_kupdate:1;
52 unsigned int range_cyclic:1;
53 unsigned int for_background:1;
Dave Chinner7747bd42013-07-02 22:38:35 +100054 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
Tejun Heoac7b19a2015-05-22 17:13:57 -040055 unsigned int auto_free:1; /* free on completion */
Tejun Heo98754bf2015-05-22 17:13:59 -040056 unsigned int single_wait:1;
57 unsigned int single_done:1;
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -060058 enum wb_reason reason; /* why was writeback initiated? */
Jens Axboec4a77a62009-09-16 15:18:25 +020059
Jens Axboe8010c3b2009-09-15 20:04:57 +020060 struct list_head list; /* pending work list */
Tejun Heocc395d72015-05-22 17:13:58 -040061 struct wb_completion *done; /* set if the caller waits */
Jens Axboe03ba3782009-09-09 09:08:54 +020062};
63
Theodore Ts'oa2f48702015-03-17 12:23:19 -040064/*
Tejun Heocc395d72015-05-22 17:13:58 -040065 * If one wants to wait for one or more wb_writeback_works, each work's
66 * ->done should be set to a wb_completion defined using the following
67 * macro. Once all work items are issued with wb_queue_work(), the caller
68 * can wait for the completion of all using wb_wait_for_completion(). Work
69 * items which are waited upon aren't freed automatically on completion.
70 */
71#define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
72 struct wb_completion cmpl = { \
73 .cnt = ATOMIC_INIT(1), \
74 }
75
76
77/*
Theodore Ts'oa2f48702015-03-17 12:23:19 -040078 * If an inode is constantly having its pages dirtied, but then the
79 * updates stop dirtytime_expire_interval seconds in the past, it's
80 * possible for the worst case time between when an inode has its
81 * timestamps updated and when they finally get written out to be two
82 * dirtytime_expire_intervals. We set the default to 12 hours (in
83 * seconds), which means most of the time inodes will have their
84 * timestamps written to disk after 12 hours, but in the worst case a
85 * few inodes might not their timestamps updated for 24 hours.
86 */
87unsigned int dirtytime_expire_interval = 12 * 60 * 60;
88
Nick Piggin7ccf19a2010-10-21 11:49:30 +110089static inline struct inode *wb_inode(struct list_head *head)
90{
91 return list_entry(head, struct inode, i_wb_list);
92}
93
Wu Fengguang15eb77a2012-01-17 11:18:56 -060094/*
95 * Include the creation of the trace points after defining the
96 * wb_writeback_work structure and inline functions so that the definition
97 * remains local to this file.
98 */
99#define CREATE_TRACE_POINTS
100#include <trace/events/writeback.h>
101
Steven Whitehouse774016b2014-02-06 15:47:47 +0000102EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
103
Tejun Heod6c10f12015-05-22 17:13:45 -0400104static bool wb_io_lists_populated(struct bdi_writeback *wb)
105{
106 if (wb_has_dirty_io(wb)) {
107 return false;
108 } else {
109 set_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -0400110 WARN_ON_ONCE(!wb->avg_write_bandwidth);
Tejun Heo766a9d62015-05-22 17:13:46 -0400111 atomic_long_add(wb->avg_write_bandwidth,
112 &wb->bdi->tot_write_bandwidth);
Tejun Heod6c10f12015-05-22 17:13:45 -0400113 return true;
114 }
115}
116
117static void wb_io_lists_depopulated(struct bdi_writeback *wb)
118{
119 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
Tejun Heo766a9d62015-05-22 17:13:46 -0400120 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
Tejun Heod6c10f12015-05-22 17:13:45 -0400121 clear_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -0400122 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
123 &wb->bdi->tot_write_bandwidth) < 0);
Tejun Heo766a9d62015-05-22 17:13:46 -0400124 }
Tejun Heod6c10f12015-05-22 17:13:45 -0400125}
126
127/**
128 * inode_wb_list_move_locked - move an inode onto a bdi_writeback IO list
129 * @inode: inode to be moved
130 * @wb: target bdi_writeback
131 * @head: one of @wb->b_{dirty|io|more_io}
132 *
133 * Move @inode->i_wb_list to @list of @wb and set %WB_has_dirty_io.
134 * Returns %true if @inode is the first occupant of the !dirty_time IO
135 * lists; otherwise, %false.
136 */
137static bool inode_wb_list_move_locked(struct inode *inode,
138 struct bdi_writeback *wb,
139 struct list_head *head)
140{
141 assert_spin_locked(&wb->list_lock);
142
143 list_move(&inode->i_wb_list, head);
144
145 /* dirty_time doesn't count as dirty_io until expiration */
146 if (head != &wb->b_dirty_time)
147 return wb_io_lists_populated(wb);
148
149 wb_io_lists_depopulated(wb);
150 return false;
151}
152
153/**
154 * inode_wb_list_del_locked - remove an inode from its bdi_writeback IO list
155 * @inode: inode to be removed
156 * @wb: bdi_writeback @inode is being removed from
157 *
158 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
159 * clear %WB_has_dirty_io if all are empty afterwards.
160 */
161static void inode_wb_list_del_locked(struct inode *inode,
162 struct bdi_writeback *wb)
163{
164 assert_spin_locked(&wb->list_lock);
165
166 list_del_init(&inode->i_wb_list);
167 wb_io_lists_depopulated(wb);
168}
169
Tejun Heof0054bb2015-05-22 17:13:30 -0400170static void wb_wakeup(struct bdi_writeback *wb)
Jan Kara5acda9d2014-04-03 14:46:23 -0700171{
Tejun Heof0054bb2015-05-22 17:13:30 -0400172 spin_lock_bh(&wb->work_lock);
173 if (test_bit(WB_registered, &wb->state))
174 mod_delayed_work(bdi_wq, &wb->dwork, 0);
175 spin_unlock_bh(&wb->work_lock);
Jan Kara5acda9d2014-04-03 14:46:23 -0700176}
177
Tejun Heof0054bb2015-05-22 17:13:30 -0400178static void wb_queue_work(struct bdi_writeback *wb,
179 struct wb_writeback_work *work)
Jan Kara65850272011-01-13 15:45:44 -0800180{
Tejun Heof0054bb2015-05-22 17:13:30 -0400181 trace_writeback_queue(wb->bdi, work);
Jan Kara65850272011-01-13 15:45:44 -0800182
Tejun Heof0054bb2015-05-22 17:13:30 -0400183 spin_lock_bh(&wb->work_lock);
Tejun Heo98754bf2015-05-22 17:13:59 -0400184 if (!test_bit(WB_registered, &wb->state)) {
185 if (work->single_wait)
186 work->single_done = 1;
Jan Kara5acda9d2014-04-03 14:46:23 -0700187 goto out_unlock;
Tejun Heo98754bf2015-05-22 17:13:59 -0400188 }
Tejun Heocc395d72015-05-22 17:13:58 -0400189 if (work->done)
190 atomic_inc(&work->done->cnt);
Tejun Heof0054bb2015-05-22 17:13:30 -0400191 list_add_tail(&work->list, &wb->work_list);
192 mod_delayed_work(bdi_wq, &wb->dwork, 0);
Jan Kara5acda9d2014-04-03 14:46:23 -0700193out_unlock:
Tejun Heof0054bb2015-05-22 17:13:30 -0400194 spin_unlock_bh(&wb->work_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
Tejun Heocc395d72015-05-22 17:13:58 -0400197/**
198 * wb_wait_for_completion - wait for completion of bdi_writeback_works
199 * @bdi: bdi work items were issued to
200 * @done: target wb_completion
201 *
202 * Wait for one or more work items issued to @bdi with their ->done field
203 * set to @done, which should have been defined with
204 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
205 * work items are completed. Work items which are waited upon aren't freed
206 * automatically on completion.
207 */
208static void wb_wait_for_completion(struct backing_dev_info *bdi,
209 struct wb_completion *done)
210{
211 atomic_dec(&done->cnt); /* put down the initial count */
212 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
213}
214
Tejun Heo703c2702015-05-22 17:13:44 -0400215#ifdef CONFIG_CGROUP_WRITEBACK
216
Tejun Heo2a814902015-05-28 14:50:51 -0400217/* parameters for foreign inode detection, see wb_detach_inode() */
218#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
219#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
220#define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
221#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
222
223#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
224#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
225 /* each slot's duration is 2s / 16 */
226#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
227 /* if foreign slots >= 8, switch */
228#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
229 /* one round can affect upto 5 slots */
230
Tejun Heo21c63212015-05-28 14:50:49 -0400231void __inode_attach_wb(struct inode *inode, struct page *page)
232{
233 struct backing_dev_info *bdi = inode_to_bdi(inode);
234 struct bdi_writeback *wb = NULL;
235
236 if (inode_cgwb_enabled(inode)) {
237 struct cgroup_subsys_state *memcg_css;
238
239 if (page) {
240 memcg_css = mem_cgroup_css_from_page(page);
241 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
242 } else {
243 /* must pin memcg_css, see wb_get_create() */
244 memcg_css = task_get_css(current, memory_cgrp_id);
245 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
246 css_put(memcg_css);
247 }
248 }
249
250 if (!wb)
251 wb = &bdi->wb;
252
253 /*
254 * There may be multiple instances of this function racing to
255 * update the same inode. Use cmpxchg() to tell the winner.
256 */
257 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
258 wb_put(wb);
259}
260
Tejun Heo703c2702015-05-22 17:13:44 -0400261/**
Tejun Heo87e1d782015-05-28 14:50:52 -0400262 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
263 * @inode: inode of interest with i_lock held
264 *
265 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
266 * held on entry and is released on return. The returned wb is guaranteed
267 * to stay @inode's associated wb until its list_lock is released.
268 */
269static struct bdi_writeback *
270locked_inode_to_wb_and_lock_list(struct inode *inode)
271 __releases(&inode->i_lock)
272 __acquires(&wb->list_lock)
273{
274 while (true) {
275 struct bdi_writeback *wb = inode_to_wb(inode);
276
277 /*
278 * inode_to_wb() association is protected by both
279 * @inode->i_lock and @wb->list_lock but list_lock nests
280 * outside i_lock. Drop i_lock and verify that the
281 * association hasn't changed after acquiring list_lock.
282 */
283 wb_get(wb);
284 spin_unlock(&inode->i_lock);
285 spin_lock(&wb->list_lock);
286 wb_put(wb); /* not gonna deref it anymore */
287
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400288 /* i_wb may have changed inbetween, can't use inode_to_wb() */
289 if (likely(wb == inode->i_wb))
Tejun Heo87e1d782015-05-28 14:50:52 -0400290 return wb; /* @inode already has ref */
291
292 spin_unlock(&wb->list_lock);
293 cpu_relax();
294 spin_lock(&inode->i_lock);
295 }
296}
297
298/**
299 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
300 * @inode: inode of interest
301 *
302 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
303 * on entry.
304 */
305static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
306 __acquires(&wb->list_lock)
307{
308 spin_lock(&inode->i_lock);
309 return locked_inode_to_wb_and_lock_list(inode);
310}
311
Tejun Heo682aa8e2015-05-28 14:50:53 -0400312struct inode_switch_wbs_context {
313 struct inode *inode;
314 struct bdi_writeback *new_wb;
315
316 struct rcu_head rcu_head;
317 struct work_struct work;
318};
319
320static void inode_switch_wbs_work_fn(struct work_struct *work)
321{
322 struct inode_switch_wbs_context *isw =
323 container_of(work, struct inode_switch_wbs_context, work);
324 struct inode *inode = isw->inode;
Tejun Heod10c8092015-05-28 14:50:56 -0400325 struct address_space *mapping = inode->i_mapping;
326 struct bdi_writeback *old_wb = inode->i_wb;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400327 struct bdi_writeback *new_wb = isw->new_wb;
Tejun Heod10c8092015-05-28 14:50:56 -0400328 struct radix_tree_iter iter;
329 bool switched = false;
330 void **slot;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400331
332 /*
333 * By the time control reaches here, RCU grace period has passed
334 * since I_WB_SWITCH assertion and all wb stat update transactions
335 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
336 * synchronizing against mapping->tree_lock.
Tejun Heod10c8092015-05-28 14:50:56 -0400337 *
338 * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
339 * gives us exclusion against all wb related operations on @inode
340 * including IO list manipulations and stat updates.
Tejun Heo682aa8e2015-05-28 14:50:53 -0400341 */
Tejun Heod10c8092015-05-28 14:50:56 -0400342 if (old_wb < new_wb) {
343 spin_lock(&old_wb->list_lock);
344 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
345 } else {
346 spin_lock(&new_wb->list_lock);
347 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
348 }
Tejun Heo682aa8e2015-05-28 14:50:53 -0400349 spin_lock(&inode->i_lock);
Tejun Heod10c8092015-05-28 14:50:56 -0400350 spin_lock_irq(&mapping->tree_lock);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400351
Tejun Heod10c8092015-05-28 14:50:56 -0400352 /*
353 * Once I_FREEING is visible under i_lock, the eviction path owns
354 * the inode and we shouldn't modify ->i_wb_list.
355 */
356 if (unlikely(inode->i_state & I_FREEING))
357 goto skip_switch;
358
359 /*
360 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
361 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
362 * pages actually under underwriteback.
363 */
364 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
365 PAGECACHE_TAG_DIRTY) {
366 struct page *page = radix_tree_deref_slot_protected(slot,
367 &mapping->tree_lock);
368 if (likely(page) && PageDirty(page)) {
369 __dec_wb_stat(old_wb, WB_RECLAIMABLE);
370 __inc_wb_stat(new_wb, WB_RECLAIMABLE);
371 }
372 }
373
374 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
375 PAGECACHE_TAG_WRITEBACK) {
376 struct page *page = radix_tree_deref_slot_protected(slot,
377 &mapping->tree_lock);
378 if (likely(page)) {
379 WARN_ON_ONCE(!PageWriteback(page));
380 __dec_wb_stat(old_wb, WB_WRITEBACK);
381 __inc_wb_stat(new_wb, WB_WRITEBACK);
382 }
383 }
384
385 wb_get(new_wb);
386
387 /*
388 * Transfer to @new_wb's IO list if necessary. The specific list
389 * @inode was on is ignored and the inode is put on ->b_dirty which
390 * is always correct including from ->b_dirty_time. The transfer
391 * preserves @inode->dirtied_when ordering.
392 */
393 if (!list_empty(&inode->i_wb_list)) {
394 struct inode *pos;
395
396 inode_wb_list_del_locked(inode, old_wb);
397 inode->i_wb = new_wb;
398 list_for_each_entry(pos, &new_wb->b_dirty, i_wb_list)
399 if (time_after_eq(inode->dirtied_when,
400 pos->dirtied_when))
401 break;
402 inode_wb_list_move_locked(inode, new_wb, pos->i_wb_list.prev);
403 } else {
404 inode->i_wb = new_wb;
405 }
406
407 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400408 inode->i_wb_frn_winner = 0;
409 inode->i_wb_frn_avg_time = 0;
410 inode->i_wb_frn_history = 0;
Tejun Heod10c8092015-05-28 14:50:56 -0400411 switched = true;
412skip_switch:
Tejun Heo682aa8e2015-05-28 14:50:53 -0400413 /*
414 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
415 * ensures that the new wb is visible if they see !I_WB_SWITCH.
416 */
417 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
418
Tejun Heod10c8092015-05-28 14:50:56 -0400419 spin_unlock_irq(&mapping->tree_lock);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400420 spin_unlock(&inode->i_lock);
Tejun Heod10c8092015-05-28 14:50:56 -0400421 spin_unlock(&new_wb->list_lock);
422 spin_unlock(&old_wb->list_lock);
423
424 if (switched) {
425 wb_wakeup(new_wb);
426 wb_put(old_wb);
427 }
428 wb_put(new_wb);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400429
430 iput(inode);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400431 kfree(isw);
432}
433
434static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
435{
436 struct inode_switch_wbs_context *isw = container_of(rcu_head,
437 struct inode_switch_wbs_context, rcu_head);
438
439 /* needs to grab bh-unsafe locks, bounce to work item */
440 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
441 schedule_work(&isw->work);
442}
443
444/**
445 * inode_switch_wbs - change the wb association of an inode
446 * @inode: target inode
447 * @new_wb_id: ID of the new wb
448 *
449 * Switch @inode's wb association to the wb identified by @new_wb_id. The
450 * switching is performed asynchronously and may fail silently.
451 */
452static void inode_switch_wbs(struct inode *inode, int new_wb_id)
453{
454 struct backing_dev_info *bdi = inode_to_bdi(inode);
455 struct cgroup_subsys_state *memcg_css;
456 struct inode_switch_wbs_context *isw;
457
458 /* noop if seems to be already in progress */
459 if (inode->i_state & I_WB_SWITCH)
460 return;
461
462 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
463 if (!isw)
464 return;
465
466 /* find and pin the new wb */
467 rcu_read_lock();
468 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
469 if (memcg_css)
470 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
471 rcu_read_unlock();
472 if (!isw->new_wb)
473 goto out_free;
474
475 /* while holding I_WB_SWITCH, no one else can update the association */
476 spin_lock(&inode->i_lock);
477 if (inode->i_state & (I_WB_SWITCH | I_FREEING) ||
478 inode_to_wb(inode) == isw->new_wb) {
479 spin_unlock(&inode->i_lock);
480 goto out_free;
481 }
482 inode->i_state |= I_WB_SWITCH;
483 spin_unlock(&inode->i_lock);
484
485 ihold(inode);
486 isw->inode = inode;
487
488 /*
489 * In addition to synchronizing among switchers, I_WB_SWITCH tells
490 * the RCU protected stat update paths to grab the mapping's
491 * tree_lock so that stat transfer can synchronize against them.
492 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
493 */
494 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
495 return;
496
497out_free:
498 if (isw->new_wb)
499 wb_put(isw->new_wb);
500 kfree(isw);
501}
502
Tejun Heo87e1d782015-05-28 14:50:52 -0400503/**
Tejun Heob16b1de2015-06-02 08:39:48 -0600504 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
505 * @wbc: writeback_control of interest
506 * @inode: target inode
507 *
508 * @inode is locked and about to be written back under the control of @wbc.
509 * Record @inode's writeback context into @wbc and unlock the i_lock. On
510 * writeback completion, wbc_detach_inode() should be called. This is used
511 * to track the cgroup writeback context.
512 */
513void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
514 struct inode *inode)
515{
Tejun Heodd73e4b2015-06-16 18:48:30 -0400516 if (!inode_cgwb_enabled(inode)) {
517 spin_unlock(&inode->i_lock);
518 return;
519 }
520
Tejun Heob16b1de2015-06-02 08:39:48 -0600521 wbc->wb = inode_to_wb(inode);
Tejun Heo2a814902015-05-28 14:50:51 -0400522 wbc->inode = inode;
523
524 wbc->wb_id = wbc->wb->memcg_css->id;
525 wbc->wb_lcand_id = inode->i_wb_frn_winner;
526 wbc->wb_tcand_id = 0;
527 wbc->wb_bytes = 0;
528 wbc->wb_lcand_bytes = 0;
529 wbc->wb_tcand_bytes = 0;
530
Tejun Heob16b1de2015-06-02 08:39:48 -0600531 wb_get(wbc->wb);
532 spin_unlock(&inode->i_lock);
Tejun Heoe8a7abf2015-05-28 14:50:57 -0400533
534 /*
535 * A dying wb indicates that the memcg-blkcg mapping has changed
536 * and a new wb is already serving the memcg. Switch immediately.
537 */
538 if (unlikely(wb_dying(wbc->wb)))
539 inode_switch_wbs(inode, wbc->wb_id);
Tejun Heob16b1de2015-06-02 08:39:48 -0600540}
541
542/**
Tejun Heo2a814902015-05-28 14:50:51 -0400543 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
544 * @wbc: writeback_control of the just finished writeback
Tejun Heob16b1de2015-06-02 08:39:48 -0600545 *
546 * To be called after a writeback attempt of an inode finishes and undoes
547 * wbc_attach_and_unlock_inode(). Can be called under any context.
Tejun Heo2a814902015-05-28 14:50:51 -0400548 *
549 * As concurrent write sharing of an inode is expected to be very rare and
550 * memcg only tracks page ownership on first-use basis severely confining
551 * the usefulness of such sharing, cgroup writeback tracks ownership
552 * per-inode. While the support for concurrent write sharing of an inode
553 * is deemed unnecessary, an inode being written to by different cgroups at
554 * different points in time is a lot more common, and, more importantly,
555 * charging only by first-use can too readily lead to grossly incorrect
556 * behaviors (single foreign page can lead to gigabytes of writeback to be
557 * incorrectly attributed).
558 *
559 * To resolve this issue, cgroup writeback detects the majority dirtier of
560 * an inode and transfers the ownership to it. To avoid unnnecessary
561 * oscillation, the detection mechanism keeps track of history and gives
562 * out the switch verdict only if the foreign usage pattern is stable over
563 * a certain amount of time and/or writeback attempts.
564 *
565 * On each writeback attempt, @wbc tries to detect the majority writer
566 * using Boyer-Moore majority vote algorithm. In addition to the byte
567 * count from the majority voting, it also counts the bytes written for the
568 * current wb and the last round's winner wb (max of last round's current
569 * wb, the winner from two rounds ago, and the last round's majority
570 * candidate). Keeping track of the historical winner helps the algorithm
571 * to semi-reliably detect the most active writer even when it's not the
572 * absolute majority.
573 *
574 * Once the winner of the round is determined, whether the winner is
575 * foreign or not and how much IO time the round consumed is recorded in
576 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
577 * over a certain threshold, the switch verdict is given.
Tejun Heob16b1de2015-06-02 08:39:48 -0600578 */
579void wbc_detach_inode(struct writeback_control *wbc)
580{
Tejun Heo2a814902015-05-28 14:50:51 -0400581 struct bdi_writeback *wb = wbc->wb;
582 struct inode *inode = wbc->inode;
Tejun Heodd73e4b2015-06-16 18:48:30 -0400583 unsigned long avg_time, max_bytes, max_time;
584 u16 history;
Tejun Heo2a814902015-05-28 14:50:51 -0400585 int max_id;
586
Tejun Heodd73e4b2015-06-16 18:48:30 -0400587 if (!wb)
588 return;
589
590 history = inode->i_wb_frn_history;
591 avg_time = inode->i_wb_frn_avg_time;
592
Tejun Heo2a814902015-05-28 14:50:51 -0400593 /* pick the winner of this round */
594 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
595 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
596 max_id = wbc->wb_id;
597 max_bytes = wbc->wb_bytes;
598 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
599 max_id = wbc->wb_lcand_id;
600 max_bytes = wbc->wb_lcand_bytes;
601 } else {
602 max_id = wbc->wb_tcand_id;
603 max_bytes = wbc->wb_tcand_bytes;
604 }
605
606 /*
607 * Calculate the amount of IO time the winner consumed and fold it
608 * into the running average kept per inode. If the consumed IO
609 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
610 * deciding whether to switch or not. This is to prevent one-off
611 * small dirtiers from skewing the verdict.
612 */
613 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
614 wb->avg_write_bandwidth);
615 if (avg_time)
616 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
617 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
618 else
619 avg_time = max_time; /* immediate catch up on first run */
620
621 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
622 int slots;
623
624 /*
625 * The switch verdict is reached if foreign wb's consume
626 * more than a certain proportion of IO time in a
627 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
628 * history mask where each bit represents one sixteenth of
629 * the period. Determine the number of slots to shift into
630 * history from @max_time.
631 */
632 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
633 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
634 history <<= slots;
635 if (wbc->wb_id != max_id)
636 history |= (1U << slots) - 1;
637
638 /*
639 * Switch if the current wb isn't the consistent winner.
640 * If there are multiple closely competing dirtiers, the
641 * inode may switch across them repeatedly over time, which
642 * is okay. The main goal is avoiding keeping an inode on
643 * the wrong wb for an extended period of time.
644 */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400645 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
646 inode_switch_wbs(inode, max_id);
Tejun Heo2a814902015-05-28 14:50:51 -0400647 }
648
649 /*
650 * Multiple instances of this function may race to update the
651 * following fields but we don't mind occassional inaccuracies.
652 */
653 inode->i_wb_frn_winner = max_id;
654 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
655 inode->i_wb_frn_history = history;
656
Tejun Heob16b1de2015-06-02 08:39:48 -0600657 wb_put(wbc->wb);
658 wbc->wb = NULL;
659}
660
661/**
Tejun Heo2a814902015-05-28 14:50:51 -0400662 * wbc_account_io - account IO issued during writeback
663 * @wbc: writeback_control of the writeback in progress
664 * @page: page being written out
665 * @bytes: number of bytes being written out
666 *
667 * @bytes from @page are about to written out during the writeback
668 * controlled by @wbc. Keep the book for foreign inode detection. See
669 * wbc_detach_inode().
670 */
671void wbc_account_io(struct writeback_control *wbc, struct page *page,
672 size_t bytes)
673{
674 int id;
675
676 /*
677 * pageout() path doesn't attach @wbc to the inode being written
678 * out. This is intentional as we don't want the function to block
679 * behind a slow cgroup. Ultimately, we want pageout() to kick off
680 * regular writeback instead of writing things out itself.
681 */
682 if (!wbc->wb)
683 return;
684
685 rcu_read_lock();
686 id = mem_cgroup_css_from_page(page)->id;
687 rcu_read_unlock();
688
689 if (id == wbc->wb_id) {
690 wbc->wb_bytes += bytes;
691 return;
692 }
693
694 if (id == wbc->wb_lcand_id)
695 wbc->wb_lcand_bytes += bytes;
696
697 /* Boyer-Moore majority vote algorithm */
698 if (!wbc->wb_tcand_bytes)
699 wbc->wb_tcand_id = id;
700 if (id == wbc->wb_tcand_id)
701 wbc->wb_tcand_bytes += bytes;
702 else
703 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
704}
Tejun Heo5aa2a962015-07-23 14:27:09 -0400705EXPORT_SYMBOL_GPL(wbc_account_io);
Tejun Heo2a814902015-05-28 14:50:51 -0400706
707/**
Tejun Heo703c2702015-05-22 17:13:44 -0400708 * inode_congested - test whether an inode is congested
709 * @inode: inode to test for congestion
710 * @cong_bits: mask of WB_[a]sync_congested bits to test
711 *
712 * Tests whether @inode is congested. @cong_bits is the mask of congestion
713 * bits to test and the return value is the mask of set bits.
714 *
715 * If cgroup writeback is enabled for @inode, the congestion state is
716 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
717 * associated with @inode is congested; otherwise, the root wb's congestion
718 * state is used.
719 */
720int inode_congested(struct inode *inode, int cong_bits)
721{
Tejun Heo5cb8b822015-05-28 14:50:54 -0400722 /*
723 * Once set, ->i_wb never becomes NULL while the inode is alive.
724 * Start transaction iff ->i_wb is visible.
725 */
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400726 if (inode && inode_to_wb_is_valid(inode)) {
Tejun Heo5cb8b822015-05-28 14:50:54 -0400727 struct bdi_writeback *wb;
728 bool locked, congested;
729
730 wb = unlocked_inode_to_wb_begin(inode, &locked);
731 congested = wb_congested(wb, cong_bits);
732 unlocked_inode_to_wb_end(inode, locked);
733 return congested;
Tejun Heo703c2702015-05-22 17:13:44 -0400734 }
735
736 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
737}
738EXPORT_SYMBOL_GPL(inode_congested);
739
Tejun Heof2b65122015-05-22 17:13:55 -0400740/**
Tejun Heo98754bf2015-05-22 17:13:59 -0400741 * wb_wait_for_single_work - wait for completion of a single bdi_writeback_work
742 * @bdi: bdi the work item was issued to
743 * @work: work item to wait for
744 *
745 * Wait for the completion of @work which was issued to one of @bdi's
746 * bdi_writeback's. The caller must have set @work->single_wait before
747 * issuing it. This wait operates independently fo
748 * wb_wait_for_completion() and also disables automatic freeing of @work.
749 */
750static void wb_wait_for_single_work(struct backing_dev_info *bdi,
751 struct wb_writeback_work *work)
752{
753 if (WARN_ON_ONCE(!work->single_wait))
754 return;
755
756 wait_event(bdi->wb_waitq, work->single_done);
757
758 /*
759 * Paired with smp_wmb() in wb_do_writeback() and ensures that all
760 * modifications to @work prior to assertion of ->single_done is
761 * visible to the caller once this function returns.
762 */
763 smp_rmb();
764}
765
766/**
Tejun Heof2b65122015-05-22 17:13:55 -0400767 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
768 * @wb: target bdi_writeback to split @nr_pages to
769 * @nr_pages: number of pages to write for the whole bdi
770 *
771 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
772 * relation to the total write bandwidth of all wb's w/ dirty inodes on
773 * @wb->bdi.
774 */
775static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
776{
777 unsigned long this_bw = wb->avg_write_bandwidth;
778 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
779
780 if (nr_pages == LONG_MAX)
781 return LONG_MAX;
782
783 /*
784 * This may be called on clean wb's and proportional distribution
785 * may not make sense, just use the original @nr_pages in those
786 * cases. In general, we wanna err on the side of writing more.
787 */
788 if (!tot_bw || this_bw >= tot_bw)
789 return nr_pages;
790 else
791 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
792}
793
Tejun Heodb125362015-05-22 17:14:01 -0400794/**
795 * wb_clone_and_queue_work - clone a wb_writeback_work and issue it to a wb
796 * @wb: target bdi_writeback
797 * @base_work: source wb_writeback_work
798 *
799 * Try to make a clone of @base_work and issue it to @wb. If cloning
800 * succeeds, %true is returned; otherwise, @base_work is issued directly
801 * and %false is returned. In the latter case, the caller is required to
802 * wait for @base_work's completion using wb_wait_for_single_work().
803 *
804 * A clone is auto-freed on completion. @base_work never is.
805 */
806static bool wb_clone_and_queue_work(struct bdi_writeback *wb,
807 struct wb_writeback_work *base_work)
808{
809 struct wb_writeback_work *work;
810
811 work = kmalloc(sizeof(*work), GFP_ATOMIC);
812 if (work) {
813 *work = *base_work;
814 work->auto_free = 1;
815 work->single_wait = 0;
816 } else {
817 work = base_work;
818 work->auto_free = 0;
819 work->single_wait = 1;
820 }
821 work->single_done = 0;
822 wb_queue_work(wb, work);
823 return work != base_work;
824}
825
826/**
827 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
828 * @bdi: target backing_dev_info
829 * @base_work: wb_writeback_work to issue
830 * @skip_if_busy: skip wb's which already have writeback in progress
831 *
832 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
833 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
834 * distributed to the busy wbs according to each wb's proportion in the
835 * total active write bandwidth of @bdi.
836 */
837static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
838 struct wb_writeback_work *base_work,
839 bool skip_if_busy)
840{
841 long nr_pages = base_work->nr_pages;
842 int next_blkcg_id = 0;
843 struct bdi_writeback *wb;
844 struct wb_iter iter;
845
846 might_sleep();
847
848 if (!bdi_has_dirty_io(bdi))
849 return;
850restart:
851 rcu_read_lock();
852 bdi_for_each_wb(wb, bdi, &iter, next_blkcg_id) {
853 if (!wb_has_dirty_io(wb) ||
854 (skip_if_busy && writeback_in_progress(wb)))
855 continue;
856
857 base_work->nr_pages = wb_split_bdi_pages(wb, nr_pages);
858 if (!wb_clone_and_queue_work(wb, base_work)) {
859 next_blkcg_id = wb->blkcg_css->id + 1;
860 rcu_read_unlock();
861 wb_wait_for_single_work(bdi, base_work);
862 goto restart;
863 }
864 }
865 rcu_read_unlock();
866}
867
Tejun Heof2b65122015-05-22 17:13:55 -0400868#else /* CONFIG_CGROUP_WRITEBACK */
869
Tejun Heo87e1d782015-05-28 14:50:52 -0400870static struct bdi_writeback *
871locked_inode_to_wb_and_lock_list(struct inode *inode)
872 __releases(&inode->i_lock)
873 __acquires(&wb->list_lock)
874{
875 struct bdi_writeback *wb = inode_to_wb(inode);
876
877 spin_unlock(&inode->i_lock);
878 spin_lock(&wb->list_lock);
879 return wb;
880}
881
882static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
883 __acquires(&wb->list_lock)
884{
885 struct bdi_writeback *wb = inode_to_wb(inode);
886
887 spin_lock(&wb->list_lock);
888 return wb;
889}
890
Tejun Heof2b65122015-05-22 17:13:55 -0400891static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
892{
893 return nr_pages;
894}
895
Tejun Heodb125362015-05-22 17:14:01 -0400896static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
897 struct wb_writeback_work *base_work,
898 bool skip_if_busy)
899{
900 might_sleep();
901
902 if (bdi_has_dirty_io(bdi) &&
903 (!skip_if_busy || !writeback_in_progress(&bdi->wb))) {
904 base_work->auto_free = 0;
905 base_work->single_wait = 0;
906 base_work->single_done = 0;
907 wb_queue_work(&bdi->wb, base_work);
908 }
909}
910
Tejun Heo703c2702015-05-22 17:13:44 -0400911#endif /* CONFIG_CGROUP_WRITEBACK */
912
Tejun Heoc00ddad2015-05-22 17:13:51 -0400913void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
914 bool range_cyclic, enum wb_reason reason)
Jens Axboeb6e51312009-09-16 15:13:54 +0200915{
Tejun Heoc00ddad2015-05-22 17:13:51 -0400916 struct wb_writeback_work *work;
917
918 if (!wb_has_dirty_io(wb))
919 return;
920
921 /*
922 * This is WB_SYNC_NONE writeback, so if allocation fails just
923 * wakeup the thread for old dirty data writeback
924 */
925 work = kzalloc(sizeof(*work), GFP_ATOMIC);
926 if (!work) {
927 trace_writeback_nowork(wb->bdi);
928 wb_wakeup(wb);
929 return;
930 }
931
932 work->sync_mode = WB_SYNC_NONE;
933 work->nr_pages = nr_pages;
934 work->range_cyclic = range_cyclic;
935 work->reason = reason;
Tejun Heoac7b19a2015-05-22 17:13:57 -0400936 work->auto_free = 1;
Tejun Heoc00ddad2015-05-22 17:13:51 -0400937
938 wb_queue_work(wb, work);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200939}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800940
Christoph Hellwigc5444192010-06-08 18:15:15 +0200941/**
Tejun Heo9ecf48662015-05-22 17:13:54 -0400942 * wb_start_background_writeback - start background writeback
943 * @wb: bdi_writback to write from
Christoph Hellwigc5444192010-06-08 18:15:15 +0200944 *
945 * Description:
Jan Kara65850272011-01-13 15:45:44 -0800946 * This makes sure WB_SYNC_NONE background writeback happens. When
Tejun Heo9ecf48662015-05-22 17:13:54 -0400947 * this function returns, it is only guaranteed that for given wb
Jan Kara65850272011-01-13 15:45:44 -0800948 * some IO is happening if we are over background dirty threshold.
949 * Caller need not hold sb s_umount semaphore.
Christoph Hellwigc5444192010-06-08 18:15:15 +0200950 */
Tejun Heo9ecf48662015-05-22 17:13:54 -0400951void wb_start_background_writeback(struct bdi_writeback *wb)
Christoph Hellwigc5444192010-06-08 18:15:15 +0200952{
Jan Kara65850272011-01-13 15:45:44 -0800953 /*
954 * We just wake up the flusher thread. It will perform background
955 * writeback as soon as there is no other work to do.
956 */
Tejun Heo9ecf48662015-05-22 17:13:54 -0400957 trace_writeback_wake_background(wb->bdi);
958 wb_wakeup(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959}
960
961/*
Dave Chinnera66979a2011-03-22 22:23:41 +1100962 * Remove the inode from the writeback list it is on.
963 */
964void inode_wb_list_del(struct inode *inode)
965{
Tejun Heo87e1d782015-05-28 14:50:52 -0400966 struct bdi_writeback *wb;
Dave Chinnera66979a2011-03-22 22:23:41 +1100967
Tejun Heo87e1d782015-05-28 14:50:52 -0400968 wb = inode_to_wb_and_lock_list(inode);
Tejun Heod6c10f12015-05-22 17:13:45 -0400969 inode_wb_list_del_locked(inode, wb);
Tejun Heo52ebea72015-05-22 17:13:37 -0400970 spin_unlock(&wb->list_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600971}
Dave Chinnera66979a2011-03-22 22:23:41 +1100972
973/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700974 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
975 * furthest end of its superblock's dirty-inode list.
976 *
977 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200978 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700979 * the case then the inode must have been redirtied while it was being written
980 * out and we don't reset its dirtied_when.
981 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600982static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
Andrew Morton6610a0b2007-10-16 23:30:32 -0700983{
Jens Axboe03ba3782009-09-09 09:08:54 +0200984 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200985 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700986
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100987 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200988 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700989 inode->dirtied_when = jiffies;
990 }
Tejun Heod6c10f12015-05-22 17:13:45 -0400991 inode_wb_list_move_locked(inode, wb, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700992}
993
994/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200995 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700996 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600997static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700998{
Tejun Heod6c10f12015-05-22 17:13:45 -0400999 inode_wb_list_move_locked(inode, wb, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001000}
1001
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001002static void inode_sync_complete(struct inode *inode)
1003{
Jan Kara365b94ae2012-05-03 14:47:55 +02001004 inode->i_state &= ~I_SYNC;
Jan Kara4eff96d2012-11-26 16:29:51 -08001005 /* If inode is clean an unused, put it into LRU now... */
1006 inode_add_lru(inode);
Jan Kara365b94ae2012-05-03 14:47:55 +02001007 /* Waiters must see I_SYNC cleared before being woken up */
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001008 smp_mb();
1009 wake_up_bit(&inode->i_state, __I_SYNC);
1010}
1011
Jeff Laytond2caa3c52009-04-02 16:56:37 -07001012static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1013{
1014 bool ret = time_after(inode->dirtied_when, t);
1015#ifndef CONFIG_64BIT
1016 /*
1017 * For inodes being constantly redirtied, dirtied_when can get stuck.
1018 * It _appears_ to be in the future, but is actually in distant past.
1019 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +02001020 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -07001021 */
1022 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1023#endif
1024 return ret;
1025}
1026
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001027#define EXPIRE_DIRTY_ATIME 0x0001
1028
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001029/*
Wang Sheng-Hui0e2f2b22012-09-11 08:28:18 +08001030 * Move expired (dirtied before work->older_than_this) dirty inodes from
Jan Kara697e6fe2012-03-09 07:26:22 -08001031 * @delaying_queue to @dispatch_queue.
Fengguang Wu2c136572007-10-16 23:30:39 -07001032 */
Wu Fengguange84d0a42011-04-23 12:27:27 -06001033static int move_expired_inodes(struct list_head *delaying_queue,
Fengguang Wu2c136572007-10-16 23:30:39 -07001034 struct list_head *dispatch_queue,
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001035 int flags,
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001036 struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -07001037{
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001038 unsigned long *older_than_this = NULL;
1039 unsigned long expire_time;
Shaohua Li5c034492009-09-24 14:42:33 +02001040 LIST_HEAD(tmp);
1041 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +02001042 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +02001043 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +02001044 int do_sb_sort = 0;
Wu Fengguange84d0a42011-04-23 12:27:27 -06001045 int moved = 0;
Shaohua Li5c034492009-09-24 14:42:33 +02001046
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001047 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1048 older_than_this = work->older_than_this;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001049 else if (!work->for_sync) {
1050 expire_time = jiffies - (dirtytime_expire_interval * HZ);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001051 older_than_this = &expire_time;
1052 }
Fengguang Wu2c136572007-10-16 23:30:39 -07001053 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001054 inode = wb_inode(delaying_queue->prev);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001055 if (older_than_this &&
1056 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -07001057 break;
Jan Karaa8855992013-07-09 22:36:45 +08001058 list_move(&inode->i_wb_list, &tmp);
1059 moved++;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001060 if (flags & EXPIRE_DIRTY_ATIME)
1061 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
Jan Karaa8855992013-07-09 22:36:45 +08001062 if (sb_is_blkdev_sb(inode->i_sb))
1063 continue;
Jens Axboecf137302009-09-24 15:12:57 +02001064 if (sb && sb != inode->i_sb)
1065 do_sb_sort = 1;
1066 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001067 }
1068
Jens Axboecf137302009-09-24 15:12:57 +02001069 /* just one sb in list, splice to dispatch_queue and we're done */
1070 if (!do_sb_sort) {
1071 list_splice(&tmp, dispatch_queue);
Wu Fengguange84d0a42011-04-23 12:27:27 -06001072 goto out;
Jens Axboecf137302009-09-24 15:12:57 +02001073 }
1074
Shaohua Li5c034492009-09-24 14:42:33 +02001075 /* Move inodes from one superblock together */
1076 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001077 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001078 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001079 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +02001080 if (inode->i_sb == sb)
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001081 list_move(&inode->i_wb_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +02001082 }
Fengguang Wu2c136572007-10-16 23:30:39 -07001083 }
Wu Fengguange84d0a42011-04-23 12:27:27 -06001084out:
1085 return moved;
Fengguang Wu2c136572007-10-16 23:30:39 -07001086}
1087
1088/*
1089 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001090 * Before
1091 * newly dirtied b_dirty b_io b_more_io
1092 * =============> gf edc BA
1093 * After
1094 * newly dirtied b_dirty b_io b_more_io
1095 * =============> g fBAedc
1096 * |
1097 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -07001098 */
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001099static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -07001100{
Wu Fengguange84d0a42011-04-23 12:27:27 -06001101 int moved;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001102
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001103 assert_spin_locked(&wb->list_lock);
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001104 list_splice_init(&wb->b_more_io, &wb->b_io);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001105 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1106 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1107 EXPIRE_DIRTY_ATIME, work);
Tejun Heod6c10f12015-05-22 17:13:45 -04001108 if (moved)
1109 wb_io_lists_populated(wb);
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001110 trace_writeback_queue_io(wb, work, moved);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001111}
1112
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001113static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001114{
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001115 int ret;
1116
1117 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1118 trace_writeback_write_inode_start(inode, wbc);
1119 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1120 trace_writeback_write_inode(inode, wbc);
1121 return ret;
1122 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001123 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -07001124}
1125
1126/*
Jan Kara169ebd92012-05-03 14:48:03 +02001127 * Wait for writeback on an inode to complete. Called with i_lock held.
1128 * Caller must make sure inode cannot go away when we drop i_lock.
Christoph Hellwig01c03192009-06-08 13:35:40 +02001129 */
Jan Kara169ebd92012-05-03 14:48:03 +02001130static void __inode_wait_for_writeback(struct inode *inode)
1131 __releases(inode->i_lock)
1132 __acquires(inode->i_lock)
Christoph Hellwig01c03192009-06-08 13:35:40 +02001133{
1134 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1135 wait_queue_head_t *wqh;
1136
1137 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +11001138 while (inode->i_state & I_SYNC) {
1139 spin_unlock(&inode->i_lock);
NeilBrown74316202014-07-07 15:16:04 +10001140 __wait_on_bit(wqh, &wq, bit_wait,
1141 TASK_UNINTERRUPTIBLE);
Dave Chinner250df6e2011-03-22 22:23:36 +11001142 spin_lock(&inode->i_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -07001143 }
Christoph Hellwig01c03192009-06-08 13:35:40 +02001144}
1145
1146/*
Jan Kara169ebd92012-05-03 14:48:03 +02001147 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1148 */
1149void inode_wait_for_writeback(struct inode *inode)
1150{
1151 spin_lock(&inode->i_lock);
1152 __inode_wait_for_writeback(inode);
1153 spin_unlock(&inode->i_lock);
1154}
1155
1156/*
1157 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1158 * held and drops it. It is aimed for callers not holding any inode reference
1159 * so once i_lock is dropped, inode can go away.
1160 */
1161static void inode_sleep_on_writeback(struct inode *inode)
1162 __releases(inode->i_lock)
1163{
1164 DEFINE_WAIT(wait);
1165 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1166 int sleep;
1167
1168 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1169 sleep = inode->i_state & I_SYNC;
1170 spin_unlock(&inode->i_lock);
1171 if (sleep)
1172 schedule();
1173 finish_wait(wqh, &wait);
1174}
1175
1176/*
Jan Karaccb26b52012-05-03 14:47:58 +02001177 * Find proper writeback list for the inode depending on its current state and
1178 * possibly also change of its state while we were doing writeback. Here we
1179 * handle things such as livelock prevention or fairness of writeback among
1180 * inodes. This function can be called only by flusher thread - noone else
1181 * processes all inodes in writeback lists and requeueing inodes behind flusher
1182 * thread's back can have unexpected consequences.
1183 */
1184static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1185 struct writeback_control *wbc)
1186{
1187 if (inode->i_state & I_FREEING)
1188 return;
1189
1190 /*
1191 * Sync livelock prevention. Each inode is tagged and synced in one
1192 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1193 * the dirty time to prevent enqueue and sync it again.
1194 */
1195 if ((inode->i_state & I_DIRTY) &&
1196 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1197 inode->dirtied_when = jiffies;
1198
Jan Kara4f8ad652012-05-03 14:48:00 +02001199 if (wbc->pages_skipped) {
1200 /*
1201 * writeback is not making progress due to locked
1202 * buffers. Skip this inode for now.
1203 */
1204 redirty_tail(inode, wb);
1205 return;
1206 }
1207
Jan Karaccb26b52012-05-03 14:47:58 +02001208 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1209 /*
1210 * We didn't write back all the pages. nfs_writepages()
1211 * sometimes bales out without doing anything.
1212 */
1213 if (wbc->nr_to_write <= 0) {
1214 /* Slice used up. Queue for next turn. */
1215 requeue_io(inode, wb);
1216 } else {
1217 /*
1218 * Writeback blocked by something other than
1219 * congestion. Delay the inode for some time to
1220 * avoid spinning on the CPU (100% iowait)
1221 * retrying writeback of the dirty page/inode
1222 * that cannot be performed immediately.
1223 */
1224 redirty_tail(inode, wb);
1225 }
1226 } else if (inode->i_state & I_DIRTY) {
1227 /*
1228 * Filesystems can dirty the inode during writeback operations,
1229 * such as delayed allocation during submission or metadata
1230 * updates after data IO completion.
1231 */
1232 redirty_tail(inode, wb);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001233 } else if (inode->i_state & I_DIRTY_TIME) {
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001234 inode->dirtied_when = jiffies;
Tejun Heod6c10f12015-05-22 17:13:45 -04001235 inode_wb_list_move_locked(inode, wb, &wb->b_dirty_time);
Jan Karaccb26b52012-05-03 14:47:58 +02001236 } else {
1237 /* The inode is clean. Remove from writeback lists. */
Tejun Heod6c10f12015-05-22 17:13:45 -04001238 inode_wb_list_del_locked(inode, wb);
Jan Karaccb26b52012-05-03 14:47:58 +02001239 }
1240}
1241
1242/*
Jan Kara4f8ad652012-05-03 14:48:00 +02001243 * Write out an inode and its dirty pages. Do not update the writeback list
1244 * linkage. That is left to the caller. The caller is also responsible for
1245 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 */
1247static int
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001248__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 struct address_space *mapping = inode->i_mapping;
Wu Fengguang251d6a42010-12-01 17:33:37 -06001251 long nr_to_write = wbc->nr_to_write;
Christoph Hellwig01c03192009-06-08 13:35:40 +02001252 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 int ret;
1254
Jan Kara4f8ad652012-05-03 14:48:00 +02001255 WARN_ON(!(inode->i_state & I_SYNC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001257 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1258
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 ret = do_writepages(mapping, wbc);
1260
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001261 /*
1262 * Make sure to wait on the data before writing out the metadata.
1263 * This is important for filesystems that modify metadata on data
Dave Chinner7747bd42013-07-02 22:38:35 +10001264 * I/O completion. We don't do it for sync(2) writeback because it has a
1265 * separate, external IO completion path and ->sync_fs for guaranteeing
1266 * inode metadata is written back correctly.
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001267 */
Dave Chinner7747bd42013-07-02 22:38:35 +10001268 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001269 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 if (ret == 0)
1271 ret = err;
1272 }
1273
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001274 /*
1275 * Some filesystems may redirty the inode during the writeback
1276 * due to delalloc, clear dirty metadata flags right before
1277 * write_inode()
1278 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001279 spin_lock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001280
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001281 dirty = inode->i_state & I_DIRTY;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001282 if (inode->i_state & I_DIRTY_TIME) {
1283 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1284 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1285 unlikely(time_after(jiffies,
1286 (inode->dirtied_time_when +
1287 dirtytime_expire_interval * HZ)))) {
1288 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1289 trace_writeback_lazytime(inode);
1290 }
1291 } else
1292 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001293 inode->i_state &= ~dirty;
Tejun Heo9c6ac782014-10-24 15:38:21 -04001294
1295 /*
1296 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1297 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1298 * either they see the I_DIRTY bits cleared or we see the dirtied
1299 * inode.
1300 *
1301 * I_DIRTY_PAGES is always cleared together above even if @mapping
1302 * still has dirty pages. The flag is reinstated after smp_mb() if
1303 * necessary. This guarantees that either __mark_inode_dirty()
1304 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1305 */
1306 smp_mb();
1307
1308 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1309 inode->i_state |= I_DIRTY_PAGES;
1310
Dave Chinner250df6e2011-03-22 22:23:36 +11001311 spin_unlock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001312
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001313 if (dirty & I_DIRTY_TIME)
1314 mark_inode_dirty_sync(inode);
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001315 /* Don't write the inode if only I_DIRTY_PAGES was set */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001316 if (dirty & ~I_DIRTY_PAGES) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001317 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 if (ret == 0)
1319 ret = err;
1320 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001321 trace_writeback_single_inode(inode, wbc, nr_to_write);
1322 return ret;
1323}
1324
1325/*
1326 * Write out an inode's dirty pages. Either the caller has an active reference
1327 * on the inode or the inode has I_WILL_FREE set.
1328 *
1329 * This function is designed to be called for writing back one inode which
1330 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1331 * and does more profound writeback list handling in writeback_sb_inodes().
1332 */
1333static int
1334writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
1335 struct writeback_control *wbc)
1336{
1337 int ret = 0;
1338
1339 spin_lock(&inode->i_lock);
1340 if (!atomic_read(&inode->i_count))
1341 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1342 else
1343 WARN_ON(inode->i_state & I_WILL_FREE);
1344
1345 if (inode->i_state & I_SYNC) {
1346 if (wbc->sync_mode != WB_SYNC_ALL)
1347 goto out;
1348 /*
Jan Kara169ebd92012-05-03 14:48:03 +02001349 * It's a data-integrity sync. We must wait. Since callers hold
1350 * inode reference or inode has I_WILL_FREE set, it cannot go
1351 * away under us.
Jan Kara4f8ad652012-05-03 14:48:00 +02001352 */
Jan Kara169ebd92012-05-03 14:48:03 +02001353 __inode_wait_for_writeback(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001354 }
1355 WARN_ON(inode->i_state & I_SYNC);
1356 /*
Jan Karaf9b0e052013-12-14 04:21:26 +08001357 * Skip inode if it is clean and we have no outstanding writeback in
1358 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1359 * function since flusher thread may be doing for example sync in
1360 * parallel and if we move the inode, it could get skipped. So here we
1361 * make sure inode is on some writeback list and leave it there unless
1362 * we have completely cleaned the inode.
Jan Kara4f8ad652012-05-03 14:48:00 +02001363 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001364 if (!(inode->i_state & I_DIRTY_ALL) &&
Jan Karaf9b0e052013-12-14 04:21:26 +08001365 (wbc->sync_mode != WB_SYNC_ALL ||
1366 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
Jan Kara4f8ad652012-05-03 14:48:00 +02001367 goto out;
1368 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001369 wbc_attach_and_unlock_inode(wbc, inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001370
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001371 ret = __writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Tejun Heob16b1de2015-06-02 08:39:48 -06001373 wbc_detach_inode(wbc);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001374 spin_lock(&wb->list_lock);
Dave Chinner250df6e2011-03-22 22:23:36 +11001375 spin_lock(&inode->i_lock);
Jan Kara4f8ad652012-05-03 14:48:00 +02001376 /*
1377 * If inode is clean, remove it from writeback lists. Otherwise don't
1378 * touch it. See comment above for explanation.
1379 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001380 if (!(inode->i_state & I_DIRTY_ALL))
Tejun Heod6c10f12015-05-22 17:13:45 -04001381 inode_wb_list_del_locked(inode, wb);
Jan Kara4f8ad652012-05-03 14:48:00 +02001382 spin_unlock(&wb->list_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001383 inode_sync_complete(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001384out:
1385 spin_unlock(&inode->i_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 return ret;
1387}
1388
Tejun Heoa88a3412015-05-22 17:13:28 -04001389static long writeback_chunk_size(struct bdi_writeback *wb,
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001390 struct wb_writeback_work *work)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001391{
1392 long pages;
1393
1394 /*
1395 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1396 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1397 * here avoids calling into writeback_inodes_wb() more than once.
1398 *
1399 * The intended call sequence for WB_SYNC_ALL writeback is:
1400 *
1401 * wb_writeback()
1402 * writeback_sb_inodes() <== called only once
1403 * write_cache_pages() <== called once for each inode
1404 * (quickly) tag currently dirty pages
1405 * (maybe slowly) sync all tagged pages
1406 */
1407 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1408 pages = LONG_MAX;
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001409 else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001410 pages = min(wb->avg_write_bandwidth / 2,
Tejun Heodcc25ae2015-05-22 18:23:22 -04001411 global_wb_domain.dirty_limit / DIRTY_SCOPE);
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001412 pages = min(pages, work->nr_pages);
1413 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1414 MIN_WRITEBACK_PAGES);
1415 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001416
1417 return pages;
1418}
1419
Jens Axboe03ba3782009-09-09 09:08:54 +02001420/*
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001421 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001422 *
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001423 * Return the number of pages and/or inodes written.
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001424 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001425static long writeback_sb_inodes(struct super_block *sb,
1426 struct bdi_writeback *wb,
1427 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001428{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001429 struct writeback_control wbc = {
1430 .sync_mode = work->sync_mode,
1431 .tagged_writepages = work->tagged_writepages,
1432 .for_kupdate = work->for_kupdate,
1433 .for_background = work->for_background,
Dave Chinner7747bd42013-07-02 22:38:35 +10001434 .for_sync = work->for_sync,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001435 .range_cyclic = work->range_cyclic,
1436 .range_start = 0,
1437 .range_end = LLONG_MAX,
1438 };
1439 unsigned long start_time = jiffies;
1440 long write_chunk;
1441 long wrote = 0; /* count both pages and inodes */
1442
Jens Axboe03ba3782009-09-09 09:08:54 +02001443 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001444 struct inode *inode = wb_inode(wb->b_io.prev);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001445
1446 if (inode->i_sb != sb) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001447 if (work->sb) {
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001448 /*
1449 * We only want to write back data for this
1450 * superblock, move all inodes not belonging
1451 * to it back onto the dirty list.
1452 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001453 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001454 continue;
1455 }
1456
1457 /*
1458 * The inode belongs to a different superblock.
1459 * Bounce back to the caller to unpin this and
1460 * pin the next superblock.
1461 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001462 break;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001463 }
1464
Christoph Hellwig9843b762010-10-24 19:40:46 +02001465 /*
Wanpeng Li331cbde2012-06-09 11:10:55 +08001466 * Don't bother with new inodes or inodes being freed, first
1467 * kind does not need periodic writeout yet, and for the latter
Christoph Hellwig9843b762010-10-24 19:40:46 +02001468 * kind writeout is handled by the freer.
1469 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001470 spin_lock(&inode->i_lock);
Christoph Hellwig9843b762010-10-24 19:40:46 +02001471 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Dave Chinner250df6e2011-03-22 22:23:36 +11001472 spin_unlock(&inode->i_lock);
Wu Fengguangfcc5c222011-07-11 23:08:50 -07001473 redirty_tail(inode, wb);
Nick Piggin7ef0d732009-03-12 14:31:38 -07001474 continue;
1475 }
Jan Karacc1676d2012-05-03 14:47:56 +02001476 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1477 /*
1478 * If this inode is locked for writeback and we are not
1479 * doing writeback-for-data-integrity, move it to
1480 * b_more_io so that writeback can proceed with the
1481 * other inodes on s_io.
1482 *
1483 * We'll have another go at writing back this inode
1484 * when we completed a full scan of b_io.
1485 */
1486 spin_unlock(&inode->i_lock);
1487 requeue_io(inode, wb);
1488 trace_writeback_sb_inodes_requeue(inode);
1489 continue;
1490 }
Jan Karaf0d07b72012-05-03 14:47:59 +02001491 spin_unlock(&wb->list_lock);
1492
Jan Kara4f8ad652012-05-03 14:48:00 +02001493 /*
1494 * We already requeued the inode if it had I_SYNC set and we
1495 * are doing WB_SYNC_NONE writeback. So this catches only the
1496 * WB_SYNC_ALL case.
1497 */
Jan Kara169ebd92012-05-03 14:48:03 +02001498 if (inode->i_state & I_SYNC) {
1499 /* Wait for I_SYNC. This function drops i_lock... */
1500 inode_sleep_on_writeback(inode);
1501 /* Inode may be gone, start again */
Jan Karaead188f2012-06-08 17:07:36 +02001502 spin_lock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001503 continue;
1504 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001505 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001506 wbc_attach_and_unlock_inode(&wbc, inode);
Jan Kara169ebd92012-05-03 14:48:03 +02001507
Tejun Heoa88a3412015-05-22 17:13:28 -04001508 write_chunk = writeback_chunk_size(wb, work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001509 wbc.nr_to_write = write_chunk;
1510 wbc.pages_skipped = 0;
Dave Chinner250df6e2011-03-22 22:23:36 +11001511
Jan Kara169ebd92012-05-03 14:48:03 +02001512 /*
1513 * We use I_SYNC to pin the inode in memory. While it is set
1514 * evict_inode() will wait so the inode cannot be freed.
1515 */
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001516 __writeback_single_inode(inode, &wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001517
Tejun Heob16b1de2015-06-02 08:39:48 -06001518 wbc_detach_inode(&wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001519 work->nr_pages -= write_chunk - wbc.nr_to_write;
1520 wrote += write_chunk - wbc.nr_to_write;
Jan Kara4f8ad652012-05-03 14:48:00 +02001521 spin_lock(&wb->list_lock);
1522 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001523 if (!(inode->i_state & I_DIRTY_ALL))
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001524 wrote++;
Jan Kara4f8ad652012-05-03 14:48:00 +02001525 requeue_inode(inode, wb, &wbc);
1526 inode_sync_complete(inode);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001527 spin_unlock(&inode->i_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001528 cond_resched_lock(&wb->list_lock);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001529 /*
1530 * bail out to wb_writeback() often enough to check
1531 * background threshold and other termination conditions.
1532 */
1533 if (wrote) {
1534 if (time_is_before_jiffies(start_time + HZ / 10UL))
1535 break;
1536 if (work->nr_pages <= 0)
1537 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -08001538 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001540 return wrote;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001541}
Nick Piggin38f21972009-01-06 14:40:25 -08001542
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001543static long __writeback_inodes_wb(struct bdi_writeback *wb,
1544 struct wb_writeback_work *work)
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001545{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001546 unsigned long start_time = jiffies;
1547 long wrote = 0;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001548
1549 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001550 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001551 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001552
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001553 if (!trylock_super(sb)) {
Wu Fengguang0e9958162011-07-29 22:14:35 -06001554 /*
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001555 * trylock_super() may fail consistently due to
Wu Fengguang0e9958162011-07-29 22:14:35 -06001556 * s_umount being grabbed by someone else. Don't use
1557 * requeue_io() to avoid busy retrying the inode/sb.
1558 */
1559 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001560 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001561 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001562 wrote += writeback_sb_inodes(sb, wb, work);
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001563 up_read(&sb->s_umount);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001564
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001565 /* refer to the same tests at the end of writeback_sb_inodes */
1566 if (wrote) {
1567 if (time_is_before_jiffies(start_time + HZ / 10UL))
1568 break;
1569 if (work->nr_pages <= 0)
1570 break;
1571 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001572 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001573 /* Leave any unwritten inodes on b_io */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001574 return wrote;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001575}
1576
Wanpeng Li7d9f0732013-09-11 14:22:40 -07001577static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001578 enum wb_reason reason)
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001579{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001580 struct wb_writeback_work work = {
1581 .nr_pages = nr_pages,
1582 .sync_mode = WB_SYNC_NONE,
1583 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001584 .reason = reason,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001585 };
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001586
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001587 spin_lock(&wb->list_lock);
Wu Fengguang424b3512010-07-21 20:11:53 -06001588 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001589 queue_io(wb, &work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001590 __writeback_inodes_wb(wb, &work);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001591 spin_unlock(&wb->list_lock);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001592
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001593 return nr_pages - work.nr_pages;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001594}
1595
Jens Axboe03ba3782009-09-09 09:08:54 +02001596/*
1597 * Explicit flushing or periodic writeback of "old" data.
1598 *
1599 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1600 * dirtying-time in the inode's address_space. So this periodic writeback code
1601 * just walks the superblock inode list, writing back any inodes which are
1602 * older than a specific point in time.
1603 *
1604 * Try to run once per dirty_writeback_interval. But if a writeback event
1605 * takes longer than a dirty_writeback_interval interval, then leave a
1606 * one-second gap.
1607 *
1608 * older_than_this takes precedence over nr_to_write. So we'll only write back
1609 * all dirty pages if they are all attached to "old" mappings.
1610 */
Jens Axboec4a77a62009-09-16 15:18:25 +02001611static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001612 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001613{
Wu Fengguange98be2d2010-08-29 11:22:30 -06001614 unsigned long wb_start = jiffies;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001615 long nr_pages = work->nr_pages;
Jan Kara0dc83bd2014-02-21 11:19:04 +01001616 unsigned long oldest_jif;
Jan Karaa5989bd2009-09-16 19:22:48 +02001617 struct inode *inode;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001618 long progress;
Jens Axboe03ba3782009-09-09 09:08:54 +02001619
Jan Kara0dc83bd2014-02-21 11:19:04 +01001620 oldest_jif = jiffies;
1621 work->older_than_this = &oldest_jif;
Jens Axboe03ba3782009-09-09 09:08:54 +02001622
Wu Fengguange8dfc302011-04-21 12:06:32 -06001623 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001624 for (;;) {
1625 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001626 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +02001627 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001628 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +02001629 break;
1630
1631 /*
Jan Karaaa373cf2011-01-13 15:45:47 -08001632 * Background writeout and kupdate-style writeback may
1633 * run forever. Stop them if there is other work to do
1634 * so that e.g. sync can proceed. They'll be restarted
1635 * after the other works are all done.
1636 */
1637 if ((work->for_background || work->for_kupdate) &&
Tejun Heof0054bb2015-05-22 17:13:30 -04001638 !list_empty(&wb->work_list))
Jan Karaaa373cf2011-01-13 15:45:47 -08001639 break;
1640
1641 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001642 * For background writeout, stop when we are below the
1643 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +02001644 */
Tejun Heoaa661bb2015-05-22 18:23:31 -04001645 if (work->for_background && !wb_over_bg_thresh(wb))
Jens Axboe03ba3782009-09-09 09:08:54 +02001646 break;
1647
Jan Kara1bc36b62011-10-19 11:44:41 +02001648 /*
1649 * Kupdate and background works are special and we want to
1650 * include all inodes that need writing. Livelock avoidance is
1651 * handled by these works yielding to any other work so we are
1652 * safe.
1653 */
Wu Fengguangba9aa832010-07-21 20:32:30 -06001654 if (work->for_kupdate) {
Jan Kara0dc83bd2014-02-21 11:19:04 +01001655 oldest_jif = jiffies -
Wu Fengguangba9aa832010-07-21 20:32:30 -06001656 msecs_to_jiffies(dirty_expire_interval * 10);
Jan Kara1bc36b62011-10-19 11:44:41 +02001657 } else if (work->for_background)
Jan Kara0dc83bd2014-02-21 11:19:04 +01001658 oldest_jif = jiffies;
Dave Chinner028c2dd2010-07-07 13:24:07 +10001659
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001660 trace_writeback_start(wb->bdi, work);
Wu Fengguange8dfc302011-04-21 12:06:32 -06001661 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001662 queue_io(wb, work);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001663 if (work->sb)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001664 progress = writeback_sb_inodes(work->sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001665 else
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001666 progress = __writeback_inodes_wb(wb, work);
1667 trace_writeback_written(wb->bdi, work);
Dave Chinner028c2dd2010-07-07 13:24:07 +10001668
Wu Fengguange98be2d2010-08-29 11:22:30 -06001669 wb_update_bandwidth(wb, wb_start);
Jens Axboe03ba3782009-09-09 09:08:54 +02001670
1671 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +02001672 * Did we write something? Try for more
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001673 *
1674 * Dirty inodes are moved to b_io for writeback in batches.
1675 * The completion of the current batch does not necessarily
1676 * mean the overall work is done. So we keep looping as long
1677 * as made some progress on cleaning pages or inodes.
Jens Axboe71fd05a2009-09-23 19:32:26 +02001678 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001679 if (progress)
Jens Axboe71fd05a2009-09-23 19:32:26 +02001680 continue;
1681 /*
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001682 * No more inodes for IO, bail
Jens Axboe71fd05a2009-09-23 19:32:26 +02001683 */
Wu Fengguangb7a24412010-07-21 22:19:51 -06001684 if (list_empty(&wb->b_more_io))
Jens Axboe03ba3782009-09-09 09:08:54 +02001685 break;
1686 /*
Jens Axboe8010c3b2009-09-15 20:04:57 +02001687 * Nothing written. Wait for some inode to
1688 * become available for writeback. Otherwise
1689 * we'll just busyloop.
Jens Axboe03ba3782009-09-09 09:08:54 +02001690 */
Jens Axboe03ba3782009-09-09 09:08:54 +02001691 if (!list_empty(&wb->b_more_io)) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001692 trace_writeback_wait(wb->bdi, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001693 inode = wb_inode(wb->b_more_io.prev);
Dave Chinner250df6e2011-03-22 22:23:36 +11001694 spin_lock(&inode->i_lock);
Jan Karaf0d07b72012-05-03 14:47:59 +02001695 spin_unlock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001696 /* This function drops i_lock... */
1697 inode_sleep_on_writeback(inode);
Jan Karaf0d07b72012-05-03 14:47:59 +02001698 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001699 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001700 }
Wu Fengguange8dfc302011-04-21 12:06:32 -06001701 spin_unlock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001702
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001703 return nr_pages - work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +02001704}
1705
1706/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001707 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +02001708 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001709static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001710{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001711 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +02001712
Tejun Heof0054bb2015-05-22 17:13:30 -04001713 spin_lock_bh(&wb->work_lock);
1714 if (!list_empty(&wb->work_list)) {
1715 work = list_entry(wb->work_list.next,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001716 struct wb_writeback_work, list);
1717 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +02001718 }
Tejun Heof0054bb2015-05-22 17:13:30 -04001719 spin_unlock_bh(&wb->work_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001720 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +02001721}
1722
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001723/*
1724 * Add in the number of potentially dirty inodes, because each inode
1725 * write can dirty pagecache in the underlying blockdev.
1726 */
1727static unsigned long get_nr_dirty_pages(void)
1728{
1729 return global_page_state(NR_FILE_DIRTY) +
1730 global_page_state(NR_UNSTABLE_NFS) +
1731 get_nr_dirty_inodes();
1732}
1733
Jan Kara65850272011-01-13 15:45:44 -08001734static long wb_check_background_flush(struct bdi_writeback *wb)
1735{
Tejun Heoaa661bb2015-05-22 18:23:31 -04001736 if (wb_over_bg_thresh(wb)) {
Jan Kara65850272011-01-13 15:45:44 -08001737
1738 struct wb_writeback_work work = {
1739 .nr_pages = LONG_MAX,
1740 .sync_mode = WB_SYNC_NONE,
1741 .for_background = 1,
1742 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001743 .reason = WB_REASON_BACKGROUND,
Jan Kara65850272011-01-13 15:45:44 -08001744 };
1745
1746 return wb_writeback(wb, &work);
1747 }
1748
1749 return 0;
1750}
1751
Jens Axboe03ba3782009-09-09 09:08:54 +02001752static long wb_check_old_data_flush(struct bdi_writeback *wb)
1753{
1754 unsigned long expired;
1755 long nr_pages;
1756
Jens Axboe69b62d02010-05-17 12:51:03 +02001757 /*
1758 * When set to zero, disable periodic writeback
1759 */
1760 if (!dirty_writeback_interval)
1761 return 0;
1762
Jens Axboe03ba3782009-09-09 09:08:54 +02001763 expired = wb->last_old_flush +
1764 msecs_to_jiffies(dirty_writeback_interval * 10);
1765 if (time_before(jiffies, expired))
1766 return 0;
1767
1768 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001769 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +02001770
Jens Axboec4a77a62009-09-16 15:18:25 +02001771 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001772 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +02001773 .nr_pages = nr_pages,
1774 .sync_mode = WB_SYNC_NONE,
1775 .for_kupdate = 1,
1776 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001777 .reason = WB_REASON_PERIODIC,
Jens Axboec4a77a62009-09-16 15:18:25 +02001778 };
1779
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001780 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +02001781 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001782
1783 return 0;
1784}
1785
1786/*
1787 * Retrieve work items and do the writeback they describe
1788 */
Wanpeng Li25d130b2013-07-08 16:00:14 -07001789static long wb_do_writeback(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001790{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001791 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +02001792 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +02001793
Tejun Heo44522262015-05-22 17:13:26 -04001794 set_bit(WB_writeback_running, &wb->state);
Tejun Heof0054bb2015-05-22 17:13:30 -04001795 while ((work = get_next_work_item(wb)) != NULL) {
Tejun Heocc395d72015-05-22 17:13:58 -04001796 struct wb_completion *done = work->done;
Tejun Heo98754bf2015-05-22 17:13:59 -04001797 bool need_wake_up = false;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001798
Tejun Heof0054bb2015-05-22 17:13:30 -04001799 trace_writeback_exec(wb->bdi, work);
Dave Chinner455b2862010-07-07 13:24:06 +10001800
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001801 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001802
Tejun Heo98754bf2015-05-22 17:13:59 -04001803 if (work->single_wait) {
1804 WARN_ON_ONCE(work->auto_free);
1805 /* paired w/ rmb in wb_wait_for_single_work() */
1806 smp_wmb();
1807 work->single_done = 1;
1808 need_wake_up = true;
1809 } else if (work->auto_free) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001810 kfree(work);
Tejun Heo98754bf2015-05-22 17:13:59 -04001811 }
1812
Tejun Heocc395d72015-05-22 17:13:58 -04001813 if (done && atomic_dec_and_test(&done->cnt))
Tejun Heo98754bf2015-05-22 17:13:59 -04001814 need_wake_up = true;
1815
1816 if (need_wake_up)
Tejun Heocc395d72015-05-22 17:13:58 -04001817 wake_up_all(&wb->bdi->wb_waitq);
Jens Axboe03ba3782009-09-09 09:08:54 +02001818 }
1819
1820 /*
1821 * Check for periodic writeback, kupdated() style
1822 */
1823 wrote += wb_check_old_data_flush(wb);
Jan Kara65850272011-01-13 15:45:44 -08001824 wrote += wb_check_background_flush(wb);
Tejun Heo44522262015-05-22 17:13:26 -04001825 clear_bit(WB_writeback_running, &wb->state);
Jens Axboe03ba3782009-09-09 09:08:54 +02001826
1827 return wrote;
1828}
1829
1830/*
1831 * Handle writeback of dirty data for the device backed by this bdi. Also
Tejun Heo839a8e82013-04-01 19:08:06 -07001832 * reschedules periodically and does kupdated style flushing.
Jens Axboe03ba3782009-09-09 09:08:54 +02001833 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001834void wb_workfn(struct work_struct *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001835{
Tejun Heo839a8e82013-04-01 19:08:06 -07001836 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1837 struct bdi_writeback, dwork);
Jens Axboe03ba3782009-09-09 09:08:54 +02001838 long pages_written;
1839
Tejun Heof0054bb2015-05-22 17:13:30 -04001840 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
Peter Zijlstra766f9162010-10-26 14:22:45 -07001841 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +02001842
Tejun Heo839a8e82013-04-01 19:08:06 -07001843 if (likely(!current_is_workqueue_rescuer() ||
Tejun Heo44522262015-05-22 17:13:26 -04001844 !test_bit(WB_registered, &wb->state))) {
Artem Bityutskiy64677162010-07-25 14:29:22 +03001845 /*
Tejun Heof0054bb2015-05-22 17:13:30 -04001846 * The normal path. Keep writing back @wb until its
Tejun Heo839a8e82013-04-01 19:08:06 -07001847 * work_list is empty. Note that this path is also taken
Tejun Heof0054bb2015-05-22 17:13:30 -04001848 * if @wb is shutting down even when we're running off the
Tejun Heo839a8e82013-04-01 19:08:06 -07001849 * rescuer as work_list needs to be drained.
Artem Bityutskiy64677162010-07-25 14:29:22 +03001850 */
Tejun Heo839a8e82013-04-01 19:08:06 -07001851 do {
Wanpeng Li25d130b2013-07-08 16:00:14 -07001852 pages_written = wb_do_writeback(wb);
Tejun Heo839a8e82013-04-01 19:08:06 -07001853 trace_writeback_pages_written(pages_written);
Tejun Heof0054bb2015-05-22 17:13:30 -04001854 } while (!list_empty(&wb->work_list));
Tejun Heo839a8e82013-04-01 19:08:06 -07001855 } else {
1856 /*
1857 * bdi_wq can't get enough workers and we're running off
1858 * the emergency worker. Don't hog it. Hopefully, 1024 is
1859 * enough for efficient IO.
1860 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001861 pages_written = writeback_inodes_wb(wb, 1024,
Tejun Heo839a8e82013-04-01 19:08:06 -07001862 WB_REASON_FORKER_THREAD);
Dave Chinner455b2862010-07-07 13:24:06 +10001863 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +02001864 }
1865
Tejun Heof0054bb2015-05-22 17:13:30 -04001866 if (!list_empty(&wb->work_list))
Derek Basehore6ca738d2014-04-03 14:46:22 -07001867 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1868 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Tejun Heof0054bb2015-05-22 17:13:30 -04001869 wb_wakeup_delayed(wb);
Dave Chinner455b2862010-07-07 13:24:06 +10001870
Tejun Heo839a8e82013-04-01 19:08:06 -07001871 current->flags &= ~PF_SWAPWRITE;
Jens Axboe03ba3782009-09-09 09:08:54 +02001872}
1873
1874/*
Jens Axboe03ba3782009-09-09 09:08:54 +02001875 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1876 * the whole world.
1877 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001878void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
Jens Axboe03ba3782009-09-09 09:08:54 +02001879{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001880 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001881
Jan Kara47df3dd2013-09-11 14:22:22 -07001882 if (!nr_pages)
1883 nr_pages = get_nr_dirty_pages();
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001884
1885 rcu_read_lock();
Tejun Heof2b65122015-05-22 17:13:55 -04001886 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1887 struct bdi_writeback *wb;
1888 struct wb_iter iter;
1889
1890 if (!bdi_has_dirty_io(bdi))
1891 continue;
1892
1893 bdi_for_each_wb(wb, bdi, &iter, 0)
1894 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1895 false, reason);
1896 }
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001897 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02001898}
1899
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001900/*
1901 * Wake up bdi's periodically to make sure dirtytime inodes gets
1902 * written back periodically. We deliberately do *not* check the
1903 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1904 * kernel to be constantly waking up once there are any dirtytime
1905 * inodes on the system. So instead we define a separate delayed work
1906 * function which gets called much more rarely. (By default, only
1907 * once every 12 hours.)
1908 *
1909 * If there is any other write activity going on in the file system,
1910 * this function won't be necessary. But if the only thing that has
1911 * happened on the file system is a dirtytime inode caused by an atime
1912 * update, we need this infrastructure below to make sure that inode
1913 * eventually gets pushed out to disk.
1914 */
1915static void wakeup_dirtytime_writeback(struct work_struct *w);
1916static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1917
1918static void wakeup_dirtytime_writeback(struct work_struct *w)
1919{
1920 struct backing_dev_info *bdi;
1921
1922 rcu_read_lock();
1923 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Tejun Heo001fe6f2015-05-22 17:13:56 -04001924 struct bdi_writeback *wb;
1925 struct wb_iter iter;
1926
1927 bdi_for_each_wb(wb, bdi, &iter, 0)
1928 if (!list_empty(&bdi->wb.b_dirty_time))
1929 wb_wakeup(&bdi->wb);
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001930 }
1931 rcu_read_unlock();
1932 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1933}
1934
1935static int __init start_dirtytime_writeback(void)
1936{
1937 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1938 return 0;
1939}
1940__initcall(start_dirtytime_writeback);
1941
Theodore Ts'o1efff912015-03-17 12:23:32 -04001942int dirtytime_interval_handler(struct ctl_table *table, int write,
1943 void __user *buffer, size_t *lenp, loff_t *ppos)
1944{
1945 int ret;
1946
1947 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1948 if (ret == 0 && write)
1949 mod_delayed_work(system_wq, &dirtytime_work, 0);
1950 return ret;
1951}
1952
Jens Axboe03ba3782009-09-09 09:08:54 +02001953static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1954{
1955 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1956 struct dentry *dentry;
1957 const char *name = "?";
1958
1959 dentry = d_find_alias(inode);
1960 if (dentry) {
1961 spin_lock(&dentry->d_lock);
1962 name = (const char *) dentry->d_name.name;
1963 }
1964 printk(KERN_DEBUG
1965 "%s(%d): dirtied inode %lu (%s) on %s\n",
1966 current->comm, task_pid_nr(current), inode->i_ino,
1967 name, inode->i_sb->s_id);
1968 if (dentry) {
1969 spin_unlock(&dentry->d_lock);
1970 dput(dentry);
1971 }
1972 }
1973}
1974
1975/**
1976 * __mark_inode_dirty - internal function
1977 * @inode: inode to mark
1978 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1979 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1980 * mark_inode_dirty_sync.
1981 *
1982 * Put the inode on the super block's dirty list.
1983 *
1984 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1985 * dirty list only if it is hashed or if it refers to a blockdev.
1986 * If it was not hashed, it will never be added to the dirty list
1987 * even if it is later hashed, as it will have been marked dirty already.
1988 *
1989 * In short, make sure you hash any inodes _before_ you start marking
1990 * them dirty.
1991 *
Jens Axboe03ba3782009-09-09 09:08:54 +02001992 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1993 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1994 * the kernel-internal blockdev inode represents the dirtying time of the
1995 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1996 * page->mapping->host, so the page-dirtying time is recorded in the internal
1997 * blockdev inode.
1998 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001999#define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
Jens Axboe03ba3782009-09-09 09:08:54 +02002000void __mark_inode_dirty(struct inode *inode, int flags)
2001{
2002 struct super_block *sb = inode->i_sb;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002003 int dirtytime;
2004
2005 trace_writeback_mark_inode_dirty(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02002006
2007 /*
2008 * Don't do this for I_DIRTY_PAGES - that doesn't actually
2009 * dirty the inode itself
2010 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002011 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002012 trace_writeback_dirty_inode_start(inode, flags);
2013
Jens Axboe03ba3782009-09-09 09:08:54 +02002014 if (sb->s_op->dirty_inode)
Christoph Hellwigaa385722011-05-27 06:53:02 -04002015 sb->s_op->dirty_inode(inode, flags);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002016
2017 trace_writeback_dirty_inode(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02002018 }
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002019 if (flags & I_DIRTY_INODE)
2020 flags &= ~I_DIRTY_TIME;
2021 dirtytime = flags & I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02002022
2023 /*
Tejun Heo9c6ac782014-10-24 15:38:21 -04002024 * Paired with smp_mb() in __writeback_single_inode() for the
2025 * following lockless i_state test. See there for details.
Jens Axboe03ba3782009-09-09 09:08:54 +02002026 */
2027 smp_mb();
2028
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002029 if (((inode->i_state & flags) == flags) ||
2030 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
Jens Axboe03ba3782009-09-09 09:08:54 +02002031 return;
2032
2033 if (unlikely(block_dump))
2034 block_dump___mark_inode_dirty(inode);
2035
Dave Chinner250df6e2011-03-22 22:23:36 +11002036 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002037 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2038 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002039 if ((inode->i_state & flags) != flags) {
2040 const int was_dirty = inode->i_state & I_DIRTY;
2041
Tejun Heo52ebea72015-05-22 17:13:37 -04002042 inode_attach_wb(inode, NULL);
2043
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002044 if (flags & I_DIRTY_INODE)
2045 inode->i_state &= ~I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02002046 inode->i_state |= flags;
2047
2048 /*
2049 * If the inode is being synced, just update its dirty state.
2050 * The unlocker will place the inode on the appropriate
2051 * superblock list, based upon its state.
2052 */
2053 if (inode->i_state & I_SYNC)
Dave Chinner250df6e2011-03-22 22:23:36 +11002054 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002055
2056 /*
2057 * Only add valid (hashed) inodes to the superblock's
2058 * dirty list. Add blockdev inodes as well.
2059 */
2060 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382cb2010-10-23 15:19:20 -04002061 if (inode_unhashed(inode))
Dave Chinner250df6e2011-03-22 22:23:36 +11002062 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002063 }
Al Viroa4ffdde2010-06-02 17:38:30 -04002064 if (inode->i_state & I_FREEING)
Dave Chinner250df6e2011-03-22 22:23:36 +11002065 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002066
2067 /*
2068 * If the inode was already on b_dirty/b_io/b_more_io, don't
2069 * reposition it (that would break b_dirty time-ordering).
2070 */
2071 if (!was_dirty) {
Tejun Heo87e1d782015-05-28 14:50:52 -04002072 struct bdi_writeback *wb;
Tejun Heod6c10f12015-05-22 17:13:45 -04002073 struct list_head *dirty_list;
Dave Chinnera66979a2011-03-22 22:23:41 +11002074 bool wakeup_bdi = false;
Jens Axboe500b0672009-09-09 09:10:25 +02002075
Tejun Heo87e1d782015-05-28 14:50:52 -04002076 wb = locked_inode_to_wb_and_lock_list(inode);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002077
Tejun Heo07472592015-05-22 17:14:02 -04002078 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2079 !test_bit(WB_registered, &wb->state),
2080 "bdi-%s not registered\n", wb->bdi->name);
Jens Axboe03ba3782009-09-09 09:08:54 +02002081
2082 inode->dirtied_when = jiffies;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002083 if (dirtytime)
2084 inode->dirtied_time_when = jiffies;
Tejun Heod6c10f12015-05-22 17:13:45 -04002085
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002086 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
Tejun Heo07472592015-05-22 17:14:02 -04002087 dirty_list = &wb->b_dirty;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002088 else
Tejun Heo07472592015-05-22 17:14:02 -04002089 dirty_list = &wb->b_dirty_time;
Tejun Heod6c10f12015-05-22 17:13:45 -04002090
Tejun Heo07472592015-05-22 17:14:02 -04002091 wakeup_bdi = inode_wb_list_move_locked(inode, wb,
Tejun Heod6c10f12015-05-22 17:13:45 -04002092 dirty_list);
2093
Tejun Heo07472592015-05-22 17:14:02 -04002094 spin_unlock(&wb->list_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002095 trace_writeback_dirty_inode_enqueue(inode);
Dave Chinnera66979a2011-03-22 22:23:41 +11002096
Tejun Heod6c10f12015-05-22 17:13:45 -04002097 /*
2098 * If this is the first dirty inode for this bdi,
2099 * we have to wake-up the corresponding bdi thread
2100 * to make sure background write-back happens
2101 * later.
2102 */
Tejun Heo07472592015-05-22 17:14:02 -04002103 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2104 wb_wakeup_delayed(wb);
Dave Chinnera66979a2011-03-22 22:23:41 +11002105 return;
Jens Axboe03ba3782009-09-09 09:08:54 +02002106 }
2107 }
Dave Chinner250df6e2011-03-22 22:23:36 +11002108out_unlock_inode:
2109 spin_unlock(&inode->i_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002110
Jens Axboe03ba3782009-09-09 09:08:54 +02002111}
2112EXPORT_SYMBOL(__mark_inode_dirty);
2113
Jens Axboeb6e51312009-09-16 15:13:54 +02002114static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002115{
Jens Axboe03ba3782009-09-09 09:08:54 +02002116 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002117
Jens Axboe03ba3782009-09-09 09:08:54 +02002118 /*
2119 * We need to be protected against the filesystem going from
2120 * r/o to r/w or vice versa.
2121 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002122 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002123
Dave Chinner55fa6092011-03-22 22:23:40 +11002124 spin_lock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002125
2126 /*
2127 * Data integrity sync. Must wait for all pages under writeback,
2128 * because there may have been pages dirtied before our sync
2129 * call, but which had writeout started before we write it out.
2130 * In which case, the inode may not be on the dirty list, but
2131 * we still have to wait for that writeout.
2132 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002133 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Dave Chinner250df6e2011-03-22 22:23:36 +11002134 struct address_space *mapping = inode->i_mapping;
Jens Axboe03ba3782009-09-09 09:08:54 +02002135
Dave Chinner250df6e2011-03-22 22:23:36 +11002136 spin_lock(&inode->i_lock);
2137 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2138 (mapping->nrpages == 0)) {
2139 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002140 continue;
Dave Chinner250df6e2011-03-22 22:23:36 +11002141 }
Jens Axboe03ba3782009-09-09 09:08:54 +02002142 __iget(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11002143 spin_unlock(&inode->i_lock);
Dave Chinner55fa6092011-03-22 22:23:40 +11002144 spin_unlock(&inode_sb_list_lock);
2145
Jens Axboe03ba3782009-09-09 09:08:54 +02002146 /*
Dave Chinner55fa6092011-03-22 22:23:40 +11002147 * We hold a reference to 'inode' so it couldn't have been
2148 * removed from s_inodes list while we dropped the
2149 * inode_sb_list_lock. We cannot iput the inode now as we can
2150 * be holding the last reference and we cannot iput it under
2151 * inode_sb_list_lock. So we keep the reference and iput it
2152 * later.
Jens Axboe03ba3782009-09-09 09:08:54 +02002153 */
2154 iput(old_inode);
2155 old_inode = inode;
2156
2157 filemap_fdatawait(mapping);
2158
2159 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08002160
Dave Chinner55fa6092011-03-22 22:23:40 +11002161 spin_lock(&inode_sb_list_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002162 }
Dave Chinner55fa6092011-03-22 22:23:40 +11002163 spin_unlock(&inode_sb_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002164 iput(old_inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165}
2166
Tejun Heof30a7d02015-05-22 17:14:00 -04002167static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2168 enum wb_reason reason, bool skip_if_busy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169{
Tejun Heocc395d72015-05-22 17:13:58 -04002170 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002171 struct wb_writeback_work work = {
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002172 .sb = sb,
2173 .sync_mode = WB_SYNC_NONE,
2174 .tagged_writepages = 1,
2175 .done = &done,
2176 .nr_pages = nr,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002177 .reason = reason,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002178 };
Tejun Heoe7972912015-05-22 17:13:48 -04002179 struct backing_dev_info *bdi = sb->s_bdi;
Jens Axboe0e3c9a22010-06-01 11:08:43 +02002180
Tejun Heoe7972912015-05-22 17:13:48 -04002181 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002182 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002183 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Tejun Heof30a7d02015-05-22 17:14:00 -04002184
Tejun Heodb125362015-05-22 17:14:01 -04002185 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
Tejun Heocc395d72015-05-22 17:13:58 -04002186 wb_wait_for_completion(bdi, &done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187}
Tejun Heof30a7d02015-05-22 17:14:00 -04002188
2189/**
2190 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2191 * @sb: the superblock
2192 * @nr: the number of pages to write
2193 * @reason: reason why some writeback work initiated
2194 *
2195 * Start writeback on some inodes on this super_block. No guarantees are made
2196 * on how many (if any) will be written, and this function does not wait
2197 * for IO completion of submitted IO.
2198 */
2199void writeback_inodes_sb_nr(struct super_block *sb,
2200 unsigned long nr,
2201 enum wb_reason reason)
2202{
2203 __writeback_inodes_sb_nr(sb, nr, reason, false);
2204}
Chris Mason3259f8b2010-10-29 11:16:17 -04002205EXPORT_SYMBOL(writeback_inodes_sb_nr);
2206
2207/**
2208 * writeback_inodes_sb - writeback dirty inodes from given super_block
2209 * @sb: the superblock
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +08002210 * @reason: reason why some writeback work was initiated
Chris Mason3259f8b2010-10-29 11:16:17 -04002211 *
2212 * Start writeback on some inodes on this super_block. No guarantees are made
2213 * on how many (if any) will be written, and this function does not wait
2214 * for IO completion of submitted IO.
2215 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002216void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002217{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002218 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04002219}
Jens Axboed8a85592009-09-02 12:34:32 +02002220EXPORT_SYMBOL(writeback_inodes_sb);
2221
2222/**
Miao Xie10ee27a2013-01-10 13:47:57 +08002223 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
Chris Mason3259f8b2010-10-29 11:16:17 -04002224 * @sb: the superblock
2225 * @nr: the number of pages to write
Miao Xie10ee27a2013-01-10 13:47:57 +08002226 * @reason: the reason of writeback
Chris Mason3259f8b2010-10-29 11:16:17 -04002227 *
Miao Xie10ee27a2013-01-10 13:47:57 +08002228 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
Chris Mason3259f8b2010-10-29 11:16:17 -04002229 * Returns 1 if writeback was started, 0 if not.
2230 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002231bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2232 enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002233{
Miao Xie10ee27a2013-01-10 13:47:57 +08002234 if (!down_read_trylock(&sb->s_umount))
Tejun Heof30a7d02015-05-22 17:14:00 -04002235 return false;
Miao Xie10ee27a2013-01-10 13:47:57 +08002236
Tejun Heof30a7d02015-05-22 17:14:00 -04002237 __writeback_inodes_sb_nr(sb, nr, reason, true);
Miao Xie10ee27a2013-01-10 13:47:57 +08002238 up_read(&sb->s_umount);
Tejun Heof30a7d02015-05-22 17:14:00 -04002239 return true;
Chris Mason3259f8b2010-10-29 11:16:17 -04002240}
Miao Xie10ee27a2013-01-10 13:47:57 +08002241EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2242
2243/**
2244 * try_to_writeback_inodes_sb - try to start writeback if none underway
2245 * @sb: the superblock
2246 * @reason: reason why some writeback work was initiated
2247 *
2248 * Implement by try_to_writeback_inodes_sb_nr()
2249 * Returns 1 if writeback was started, 0 if not.
2250 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002251bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Miao Xie10ee27a2013-01-10 13:47:57 +08002252{
2253 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2254}
2255EXPORT_SYMBOL(try_to_writeback_inodes_sb);
Chris Mason3259f8b2010-10-29 11:16:17 -04002256
2257/**
Jens Axboed8a85592009-09-02 12:34:32 +02002258 * sync_inodes_sb - sync sb inode pages
Jan Kara0dc83bd2014-02-21 11:19:04 +01002259 * @sb: the superblock
Jens Axboed8a85592009-09-02 12:34:32 +02002260 *
2261 * This function writes and waits on any dirty inode belonging to this
Jan Kara0dc83bd2014-02-21 11:19:04 +01002262 * super_block.
Jens Axboed8a85592009-09-02 12:34:32 +02002263 */
Jan Kara0dc83bd2014-02-21 11:19:04 +01002264void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02002265{
Tejun Heocc395d72015-05-22 17:13:58 -04002266 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002267 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002268 .sb = sb,
2269 .sync_mode = WB_SYNC_ALL,
2270 .nr_pages = LONG_MAX,
2271 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002272 .done = &done,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002273 .reason = WB_REASON_SYNC,
Dave Chinner7747bd42013-07-02 22:38:35 +10002274 .for_sync = 1,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002275 };
Tejun Heoe7972912015-05-22 17:13:48 -04002276 struct backing_dev_info *bdi = sb->s_bdi;
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002277
Jan Kara6eedc702012-07-03 16:45:27 +02002278 /* Nothing to do? */
Tejun Heoe7972912015-05-22 17:13:48 -04002279 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002280 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002281 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2282
Tejun Heodb125362015-05-22 17:14:01 -04002283 bdi_split_work_to_wbs(bdi, &work, false);
Tejun Heocc395d72015-05-22 17:13:58 -04002284 wb_wait_for_completion(bdi, &done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002285
Jens Axboeb6e51312009-09-16 15:13:54 +02002286 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02002287}
2288EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002291 * write_inode_now - write an inode to disk
2292 * @inode: inode to write to disk
2293 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002295 * This function commits an inode to disk immediately if it is dirty. This is
2296 * primarily needed by knfsd.
2297 *
2298 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300int write_inode_now(struct inode *inode, int sync)
2301{
Christoph Hellwigf758eea2011-04-21 18:19:44 -06002302 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303 struct writeback_control wbc = {
2304 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08002305 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07002306 .range_start = 0,
2307 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 };
2309
2310 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08002311 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312
2313 might_sleep();
Jan Kara4f8ad652012-05-03 14:48:00 +02002314 return writeback_single_inode(inode, wb, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315}
2316EXPORT_SYMBOL(write_inode_now);
2317
2318/**
2319 * sync_inode - write an inode and its pages to disk.
2320 * @inode: the inode to sync
2321 * @wbc: controls the writeback mode
2322 *
2323 * sync_inode() will write an inode and its pages to disk. It will also
2324 * correctly update the inode on its superblock's dirty inode lists and will
2325 * update inode->i_state.
2326 *
2327 * The caller must have a ref on the inode.
2328 */
2329int sync_inode(struct inode *inode, struct writeback_control *wbc)
2330{
Jan Kara4f8ad652012-05-03 14:48:00 +02002331 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332}
2333EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc37650162010-10-06 10:48:20 +02002334
2335/**
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002336 * sync_inode_metadata - write an inode to disk
Christoph Hellwigc37650162010-10-06 10:48:20 +02002337 * @inode: the inode to sync
2338 * @wait: wait for I/O to complete.
2339 *
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002340 * Write an inode to disk and adjust its dirty state after completion.
Christoph Hellwigc37650162010-10-06 10:48:20 +02002341 *
2342 * Note: only writes the actual inode, no associated data or other metadata.
2343 */
2344int sync_inode_metadata(struct inode *inode, int wait)
2345{
2346 struct writeback_control wbc = {
2347 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2348 .nr_to_write = 0, /* metadata-only */
2349 };
2350
2351 return sync_inode(inode, &wbc);
2352}
2353EXPORT_SYMBOL(sync_inode_metadata);