blob: fee81e8768c955f0eaf3844d374728f5c42a0140 [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 Howells07f3f05c2006-09-30 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 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -060056 enum wb_reason reason; /* why was writeback initiated? */
Jens Axboec4a77a62009-09-16 15:18:25 +020057
Jens Axboe8010c3b2009-09-15 20:04:57 +020058 struct list_head list; /* pending work list */
Tejun Heocc395d72015-05-22 17:13:58 -040059 struct wb_completion *done; /* set if the caller waits */
Jens Axboe03ba3782009-09-09 09:08:54 +020060};
61
Theodore Ts'oa2f48702015-03-17 12:23:19 -040062/*
Tejun Heocc395d72015-05-22 17:13:58 -040063 * If one wants to wait for one or more wb_writeback_works, each work's
64 * ->done should be set to a wb_completion defined using the following
65 * macro. Once all work items are issued with wb_queue_work(), the caller
66 * can wait for the completion of all using wb_wait_for_completion(). Work
67 * items which are waited upon aren't freed automatically on completion.
68 */
69#define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \
70 struct wb_completion cmpl = { \
71 .cnt = ATOMIC_INIT(1), \
72 }
73
74
75/*
Theodore Ts'oa2f48702015-03-17 12:23:19 -040076 * If an inode is constantly having its pages dirtied, but then the
77 * updates stop dirtytime_expire_interval seconds in the past, it's
78 * possible for the worst case time between when an inode has its
79 * timestamps updated and when they finally get written out to be two
80 * dirtytime_expire_intervals. We set the default to 12 hours (in
81 * seconds), which means most of the time inodes will have their
82 * timestamps written to disk after 12 hours, but in the worst case a
83 * few inodes might not their timestamps updated for 24 hours.
84 */
85unsigned int dirtytime_expire_interval = 12 * 60 * 60;
86
Nick Piggin7ccf19a2010-10-21 11:49:30 +110087static inline struct inode *wb_inode(struct list_head *head)
88{
Dave Chinnerc7f54082015-03-04 14:07:22 -050089 return list_entry(head, struct inode, i_io_list);
Nick Piggin7ccf19a2010-10-21 11:49:30 +110090}
91
Wu Fengguang15eb77a2012-01-17 11:18:56 -060092/*
93 * Include the creation of the trace points after defining the
94 * wb_writeback_work structure and inline functions so that the definition
95 * remains local to this file.
96 */
97#define CREATE_TRACE_POINTS
98#include <trace/events/writeback.h>
99
Steven Whitehouse774016b2014-02-06 15:47:47 +0000100EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
101
Tejun Heod6c10f12015-05-22 17:13:45 -0400102static bool wb_io_lists_populated(struct bdi_writeback *wb)
103{
104 if (wb_has_dirty_io(wb)) {
105 return false;
106 } else {
107 set_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -0400108 WARN_ON_ONCE(!wb->avg_write_bandwidth);
Tejun Heo766a9d62015-05-22 17:13:46 -0400109 atomic_long_add(wb->avg_write_bandwidth,
110 &wb->bdi->tot_write_bandwidth);
Tejun Heod6c10f12015-05-22 17:13:45 -0400111 return true;
112 }
113}
114
115static void wb_io_lists_depopulated(struct bdi_writeback *wb)
116{
117 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
Tejun Heo766a9d62015-05-22 17:13:46 -0400118 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
Tejun Heod6c10f12015-05-22 17:13:45 -0400119 clear_bit(WB_has_dirty_io, &wb->state);
Tejun Heo95a46c62015-05-22 17:13:47 -0400120 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
121 &wb->bdi->tot_write_bandwidth) < 0);
Tejun Heo766a9d62015-05-22 17:13:46 -0400122 }
Tejun Heod6c10f12015-05-22 17:13:45 -0400123}
124
125/**
Dave Chinnerc7f54082015-03-04 14:07:22 -0500126 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
Tejun Heod6c10f12015-05-22 17:13:45 -0400127 * @inode: inode to be moved
128 * @wb: target bdi_writeback
129 * @head: one of @wb->b_{dirty|io|more_io}
130 *
Dave Chinnerc7f54082015-03-04 14:07:22 -0500131 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
Tejun Heod6c10f12015-05-22 17:13:45 -0400132 * Returns %true if @inode is the first occupant of the !dirty_time IO
133 * lists; otherwise, %false.
134 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500135static bool inode_io_list_move_locked(struct inode *inode,
Tejun Heod6c10f12015-05-22 17:13:45 -0400136 struct bdi_writeback *wb,
137 struct list_head *head)
138{
139 assert_spin_locked(&wb->list_lock);
140
Dave Chinnerc7f54082015-03-04 14:07:22 -0500141 list_move(&inode->i_io_list, head);
Tejun Heod6c10f12015-05-22 17:13:45 -0400142
143 /* dirty_time doesn't count as dirty_io until expiration */
144 if (head != &wb->b_dirty_time)
145 return wb_io_lists_populated(wb);
146
147 wb_io_lists_depopulated(wb);
148 return false;
149}
150
151/**
Dave Chinnerc7f54082015-03-04 14:07:22 -0500152 * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
Tejun Heod6c10f12015-05-22 17:13:45 -0400153 * @inode: inode to be removed
154 * @wb: bdi_writeback @inode is being removed from
155 *
156 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
157 * clear %WB_has_dirty_io if all are empty afterwards.
158 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500159static void inode_io_list_del_locked(struct inode *inode,
Tejun Heod6c10f12015-05-22 17:13:45 -0400160 struct bdi_writeback *wb)
161{
162 assert_spin_locked(&wb->list_lock);
163
Dave Chinnerc7f54082015-03-04 14:07:22 -0500164 list_del_init(&inode->i_io_list);
Tejun Heod6c10f12015-05-22 17:13:45 -0400165 wb_io_lists_depopulated(wb);
166}
167
Tejun Heof0054bb2015-05-22 17:13:30 -0400168static void wb_wakeup(struct bdi_writeback *wb)
Jan Kara5acda9d2014-04-03 14:46:23 -0700169{
Tejun Heof0054bb2015-05-22 17:13:30 -0400170 spin_lock_bh(&wb->work_lock);
171 if (test_bit(WB_registered, &wb->state))
172 mod_delayed_work(bdi_wq, &wb->dwork, 0);
173 spin_unlock_bh(&wb->work_lock);
Jan Kara5acda9d2014-04-03 14:46:23 -0700174}
175
Tejun Heof0054bb2015-05-22 17:13:30 -0400176static void wb_queue_work(struct bdi_writeback *wb,
177 struct wb_writeback_work *work)
Jan Kara65850272011-01-13 15:45:44 -0800178{
Tejun Heo5634cc22015-08-18 14:54:56 -0700179 trace_writeback_queue(wb, work);
Jan Kara65850272011-01-13 15:45:44 -0800180
Tejun Heof0054bb2015-05-22 17:13:30 -0400181 spin_lock_bh(&wb->work_lock);
Tejun Heo8a1270c2015-08-18 14:54:53 -0700182 if (!test_bit(WB_registered, &wb->state))
Jan Kara5acda9d2014-04-03 14:46:23 -0700183 goto out_unlock;
Tejun Heocc395d72015-05-22 17:13:58 -0400184 if (work->done)
185 atomic_inc(&work->done->cnt);
Tejun Heof0054bb2015-05-22 17:13:30 -0400186 list_add_tail(&work->list, &wb->work_list);
187 mod_delayed_work(bdi_wq, &wb->dwork, 0);
Jan Kara5acda9d2014-04-03 14:46:23 -0700188out_unlock:
Tejun Heof0054bb2015-05-22 17:13:30 -0400189 spin_unlock_bh(&wb->work_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190}
191
Tejun Heocc395d72015-05-22 17:13:58 -0400192/**
193 * wb_wait_for_completion - wait for completion of bdi_writeback_works
194 * @bdi: bdi work items were issued to
195 * @done: target wb_completion
196 *
197 * Wait for one or more work items issued to @bdi with their ->done field
198 * set to @done, which should have been defined with
199 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such
200 * work items are completed. Work items which are waited upon aren't freed
201 * automatically on completion.
202 */
203static void wb_wait_for_completion(struct backing_dev_info *bdi,
204 struct wb_completion *done)
205{
206 atomic_dec(&done->cnt); /* put down the initial count */
207 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
208}
209
Tejun Heo703c2702015-05-22 17:13:44 -0400210#ifdef CONFIG_CGROUP_WRITEBACK
211
Tejun Heo2a814902015-05-28 14:50:51 -0400212/* parameters for foreign inode detection, see wb_detach_inode() */
213#define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
214#define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
215#define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */
216#define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
217
218#define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
219#define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
220 /* each slot's duration is 2s / 16 */
221#define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
222 /* if foreign slots >= 8, switch */
223#define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
224 /* one round can affect upto 5 slots */
225
Tejun Heoa1a0e232016-02-29 18:28:53 -0500226static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
227static struct workqueue_struct *isw_wq;
228
Tejun Heo21c63212015-05-28 14:50:49 -0400229void __inode_attach_wb(struct inode *inode, struct page *page)
230{
231 struct backing_dev_info *bdi = inode_to_bdi(inode);
232 struct bdi_writeback *wb = NULL;
233
234 if (inode_cgwb_enabled(inode)) {
235 struct cgroup_subsys_state *memcg_css;
236
237 if (page) {
238 memcg_css = mem_cgroup_css_from_page(page);
239 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
240 } else {
241 /* must pin memcg_css, see wb_get_create() */
242 memcg_css = task_get_css(current, memory_cgrp_id);
243 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
244 css_put(memcg_css);
245 }
246 }
247
248 if (!wb)
249 wb = &bdi->wb;
250
251 /*
252 * There may be multiple instances of this function racing to
253 * update the same inode. Use cmpxchg() to tell the winner.
254 */
255 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
256 wb_put(wb);
257}
258
Tejun Heo703c2702015-05-22 17:13:44 -0400259/**
Tejun Heo87e1d782015-05-28 14:50:52 -0400260 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
261 * @inode: inode of interest with i_lock held
262 *
263 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
264 * held on entry and is released on return. The returned wb is guaranteed
265 * to stay @inode's associated wb until its list_lock is released.
266 */
267static struct bdi_writeback *
268locked_inode_to_wb_and_lock_list(struct inode *inode)
269 __releases(&inode->i_lock)
270 __acquires(&wb->list_lock)
271{
272 while (true) {
273 struct bdi_writeback *wb = inode_to_wb(inode);
274
275 /*
276 * inode_to_wb() association is protected by both
277 * @inode->i_lock and @wb->list_lock but list_lock nests
278 * outside i_lock. Drop i_lock and verify that the
279 * association hasn't changed after acquiring list_lock.
280 */
281 wb_get(wb);
282 spin_unlock(&inode->i_lock);
283 spin_lock(&wb->list_lock);
Tejun Heo87e1d782015-05-28 14:50:52 -0400284
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400285 /* i_wb may have changed inbetween, can't use inode_to_wb() */
Tejun Heo614a4e32016-03-18 13:50:03 -0400286 if (likely(wb == inode->i_wb)) {
287 wb_put(wb); /* @inode already has ref */
288 return wb;
289 }
Tejun Heo87e1d782015-05-28 14:50:52 -0400290
291 spin_unlock(&wb->list_lock);
Tejun Heo614a4e32016-03-18 13:50:03 -0400292 wb_put(wb);
Tejun Heo87e1d782015-05-28 14:50:52 -0400293 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
Dave Chinnerc7f54082015-03-04 14:07:22 -0500354 * the inode and we shouldn't modify ->i_io_list.
Tejun Heod10c8092015-05-28 14:50:56 -0400355 */
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 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500393 if (!list_empty(&inode->i_io_list)) {
Tejun Heod10c8092015-05-28 14:50:56 -0400394 struct inode *pos;
395
Dave Chinnerc7f54082015-03-04 14:07:22 -0500396 inode_io_list_del_locked(inode, old_wb);
Tejun Heod10c8092015-05-28 14:50:56 -0400397 inode->i_wb = new_wb;
Dave Chinnerc7f54082015-03-04 14:07:22 -0500398 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
Tejun Heod10c8092015-05-28 14:50:56 -0400399 if (time_after_eq(inode->dirtied_when,
400 pos->dirtied_when))
401 break;
Dave Chinnerc7f54082015-03-04 14:07:22 -0500402 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
Tejun Heod10c8092015-05-28 14:50:56 -0400403 } 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);
Tejun Heoa1a0e232016-02-29 18:28:53 -0500432
433 atomic_dec(&isw_nr_in_flight);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400434}
435
436static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
437{
438 struct inode_switch_wbs_context *isw = container_of(rcu_head,
439 struct inode_switch_wbs_context, rcu_head);
440
441 /* needs to grab bh-unsafe locks, bounce to work item */
442 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
Tejun Heoa1a0e232016-02-29 18:28:53 -0500443 queue_work(isw_wq, &isw->work);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400444}
445
446/**
447 * inode_switch_wbs - change the wb association of an inode
448 * @inode: target inode
449 * @new_wb_id: ID of the new wb
450 *
451 * Switch @inode's wb association to the wb identified by @new_wb_id. The
452 * switching is performed asynchronously and may fail silently.
453 */
454static void inode_switch_wbs(struct inode *inode, int new_wb_id)
455{
456 struct backing_dev_info *bdi = inode_to_bdi(inode);
457 struct cgroup_subsys_state *memcg_css;
458 struct inode_switch_wbs_context *isw;
459
460 /* noop if seems to be already in progress */
461 if (inode->i_state & I_WB_SWITCH)
462 return;
463
464 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
465 if (!isw)
466 return;
467
468 /* find and pin the new wb */
469 rcu_read_lock();
470 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
471 if (memcg_css)
472 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
473 rcu_read_unlock();
474 if (!isw->new_wb)
475 goto out_free;
476
477 /* while holding I_WB_SWITCH, no one else can update the association */
478 spin_lock(&inode->i_lock);
Tejun Heoa1a0e232016-02-29 18:28:53 -0500479 if (!(inode->i_sb->s_flags & MS_ACTIVE) ||
480 inode->i_state & (I_WB_SWITCH | I_FREEING) ||
481 inode_to_wb(inode) == isw->new_wb) {
482 spin_unlock(&inode->i_lock);
483 goto out_free;
484 }
Tejun Heo682aa8e2015-05-28 14:50:53 -0400485 inode->i_state |= I_WB_SWITCH;
486 spin_unlock(&inode->i_lock);
487
488 ihold(inode);
489 isw->inode = inode;
490
Tejun Heoa1a0e232016-02-29 18:28:53 -0500491 atomic_inc(&isw_nr_in_flight);
492
Tejun Heo682aa8e2015-05-28 14:50:53 -0400493 /*
494 * In addition to synchronizing among switchers, I_WB_SWITCH tells
495 * the RCU protected stat update paths to grab the mapping's
496 * tree_lock so that stat transfer can synchronize against them.
497 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
498 */
499 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
500 return;
501
502out_free:
503 if (isw->new_wb)
504 wb_put(isw->new_wb);
505 kfree(isw);
506}
507
Tejun Heo87e1d782015-05-28 14:50:52 -0400508/**
Tejun Heob16b1de2015-06-02 08:39:48 -0600509 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
510 * @wbc: writeback_control of interest
511 * @inode: target inode
512 *
513 * @inode is locked and about to be written back under the control of @wbc.
514 * Record @inode's writeback context into @wbc and unlock the i_lock. On
515 * writeback completion, wbc_detach_inode() should be called. This is used
516 * to track the cgroup writeback context.
517 */
518void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
519 struct inode *inode)
520{
Tejun Heodd73e4b2015-06-16 18:48:30 -0400521 if (!inode_cgwb_enabled(inode)) {
522 spin_unlock(&inode->i_lock);
523 return;
524 }
525
Tejun Heob16b1de2015-06-02 08:39:48 -0600526 wbc->wb = inode_to_wb(inode);
Tejun Heo2a814902015-05-28 14:50:51 -0400527 wbc->inode = inode;
528
529 wbc->wb_id = wbc->wb->memcg_css->id;
530 wbc->wb_lcand_id = inode->i_wb_frn_winner;
531 wbc->wb_tcand_id = 0;
532 wbc->wb_bytes = 0;
533 wbc->wb_lcand_bytes = 0;
534 wbc->wb_tcand_bytes = 0;
535
Tejun Heob16b1de2015-06-02 08:39:48 -0600536 wb_get(wbc->wb);
537 spin_unlock(&inode->i_lock);
Tejun Heoe8a7abf2015-05-28 14:50:57 -0400538
539 /*
540 * A dying wb indicates that the memcg-blkcg mapping has changed
541 * and a new wb is already serving the memcg. Switch immediately.
542 */
543 if (unlikely(wb_dying(wbc->wb)))
544 inode_switch_wbs(inode, wbc->wb_id);
Tejun Heob16b1de2015-06-02 08:39:48 -0600545}
546
547/**
Tejun Heo2a814902015-05-28 14:50:51 -0400548 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
549 * @wbc: writeback_control of the just finished writeback
Tejun Heob16b1de2015-06-02 08:39:48 -0600550 *
551 * To be called after a writeback attempt of an inode finishes and undoes
552 * wbc_attach_and_unlock_inode(). Can be called under any context.
Tejun Heo2a814902015-05-28 14:50:51 -0400553 *
554 * As concurrent write sharing of an inode is expected to be very rare and
555 * memcg only tracks page ownership on first-use basis severely confining
556 * the usefulness of such sharing, cgroup writeback tracks ownership
557 * per-inode. While the support for concurrent write sharing of an inode
558 * is deemed unnecessary, an inode being written to by different cgroups at
559 * different points in time is a lot more common, and, more importantly,
560 * charging only by first-use can too readily lead to grossly incorrect
561 * behaviors (single foreign page can lead to gigabytes of writeback to be
562 * incorrectly attributed).
563 *
564 * To resolve this issue, cgroup writeback detects the majority dirtier of
565 * an inode and transfers the ownership to it. To avoid unnnecessary
566 * oscillation, the detection mechanism keeps track of history and gives
567 * out the switch verdict only if the foreign usage pattern is stable over
568 * a certain amount of time and/or writeback attempts.
569 *
570 * On each writeback attempt, @wbc tries to detect the majority writer
571 * using Boyer-Moore majority vote algorithm. In addition to the byte
572 * count from the majority voting, it also counts the bytes written for the
573 * current wb and the last round's winner wb (max of last round's current
574 * wb, the winner from two rounds ago, and the last round's majority
575 * candidate). Keeping track of the historical winner helps the algorithm
576 * to semi-reliably detect the most active writer even when it's not the
577 * absolute majority.
578 *
579 * Once the winner of the round is determined, whether the winner is
580 * foreign or not and how much IO time the round consumed is recorded in
581 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
582 * over a certain threshold, the switch verdict is given.
Tejun Heob16b1de2015-06-02 08:39:48 -0600583 */
584void wbc_detach_inode(struct writeback_control *wbc)
585{
Tejun Heo2a814902015-05-28 14:50:51 -0400586 struct bdi_writeback *wb = wbc->wb;
587 struct inode *inode = wbc->inode;
Tejun Heodd73e4b2015-06-16 18:48:30 -0400588 unsigned long avg_time, max_bytes, max_time;
589 u16 history;
Tejun Heo2a814902015-05-28 14:50:51 -0400590 int max_id;
591
Tejun Heodd73e4b2015-06-16 18:48:30 -0400592 if (!wb)
593 return;
594
595 history = inode->i_wb_frn_history;
596 avg_time = inode->i_wb_frn_avg_time;
597
Tejun Heo2a814902015-05-28 14:50:51 -0400598 /* pick the winner of this round */
599 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
600 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
601 max_id = wbc->wb_id;
602 max_bytes = wbc->wb_bytes;
603 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
604 max_id = wbc->wb_lcand_id;
605 max_bytes = wbc->wb_lcand_bytes;
606 } else {
607 max_id = wbc->wb_tcand_id;
608 max_bytes = wbc->wb_tcand_bytes;
609 }
610
611 /*
612 * Calculate the amount of IO time the winner consumed and fold it
613 * into the running average kept per inode. If the consumed IO
614 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
615 * deciding whether to switch or not. This is to prevent one-off
616 * small dirtiers from skewing the verdict.
617 */
618 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
619 wb->avg_write_bandwidth);
620 if (avg_time)
621 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
622 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
623 else
624 avg_time = max_time; /* immediate catch up on first run */
625
626 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
627 int slots;
628
629 /*
630 * The switch verdict is reached if foreign wb's consume
631 * more than a certain proportion of IO time in a
632 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
633 * history mask where each bit represents one sixteenth of
634 * the period. Determine the number of slots to shift into
635 * history from @max_time.
636 */
637 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
638 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
639 history <<= slots;
640 if (wbc->wb_id != max_id)
641 history |= (1U << slots) - 1;
642
643 /*
644 * Switch if the current wb isn't the consistent winner.
645 * If there are multiple closely competing dirtiers, the
646 * inode may switch across them repeatedly over time, which
647 * is okay. The main goal is avoiding keeping an inode on
648 * the wrong wb for an extended period of time.
649 */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400650 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
651 inode_switch_wbs(inode, max_id);
Tejun Heo2a814902015-05-28 14:50:51 -0400652 }
653
654 /*
655 * Multiple instances of this function may race to update the
656 * following fields but we don't mind occassional inaccuracies.
657 */
658 inode->i_wb_frn_winner = max_id;
659 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
660 inode->i_wb_frn_history = history;
661
Tejun Heob16b1de2015-06-02 08:39:48 -0600662 wb_put(wbc->wb);
663 wbc->wb = NULL;
664}
665
666/**
Tejun Heo2a814902015-05-28 14:50:51 -0400667 * wbc_account_io - account IO issued during writeback
668 * @wbc: writeback_control of the writeback in progress
669 * @page: page being written out
670 * @bytes: number of bytes being written out
671 *
672 * @bytes from @page are about to written out during the writeback
673 * controlled by @wbc. Keep the book for foreign inode detection. See
674 * wbc_detach_inode().
675 */
676void wbc_account_io(struct writeback_control *wbc, struct page *page,
677 size_t bytes)
678{
679 int id;
680
681 /*
682 * pageout() path doesn't attach @wbc to the inode being written
683 * out. This is intentional as we don't want the function to block
684 * behind a slow cgroup. Ultimately, we want pageout() to kick off
685 * regular writeback instead of writing things out itself.
686 */
687 if (!wbc->wb)
688 return;
689
Tejun Heo2a814902015-05-28 14:50:51 -0400690 id = mem_cgroup_css_from_page(page)->id;
Tejun Heo2a814902015-05-28 14:50:51 -0400691
692 if (id == wbc->wb_id) {
693 wbc->wb_bytes += bytes;
694 return;
695 }
696
697 if (id == wbc->wb_lcand_id)
698 wbc->wb_lcand_bytes += bytes;
699
700 /* Boyer-Moore majority vote algorithm */
701 if (!wbc->wb_tcand_bytes)
702 wbc->wb_tcand_id = id;
703 if (id == wbc->wb_tcand_id)
704 wbc->wb_tcand_bytes += bytes;
705 else
706 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
707}
Tejun Heo5aa2a962015-07-23 14:27:09 -0400708EXPORT_SYMBOL_GPL(wbc_account_io);
Tejun Heo2a814902015-05-28 14:50:51 -0400709
710/**
Tejun Heo703c2702015-05-22 17:13:44 -0400711 * inode_congested - test whether an inode is congested
Tejun Heo60292bc2015-08-18 14:54:54 -0700712 * @inode: inode to test for congestion (may be NULL)
Tejun Heo703c2702015-05-22 17:13:44 -0400713 * @cong_bits: mask of WB_[a]sync_congested bits to test
714 *
715 * Tests whether @inode is congested. @cong_bits is the mask of congestion
716 * bits to test and the return value is the mask of set bits.
717 *
718 * If cgroup writeback is enabled for @inode, the congestion state is
719 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
720 * associated with @inode is congested; otherwise, the root wb's congestion
721 * state is used.
Tejun Heo60292bc2015-08-18 14:54:54 -0700722 *
723 * @inode is allowed to be NULL as this function is often called on
724 * mapping->host which is NULL for the swapper space.
Tejun Heo703c2702015-05-22 17:13:44 -0400725 */
726int inode_congested(struct inode *inode, int cong_bits)
727{
Tejun Heo5cb8b822015-05-28 14:50:54 -0400728 /*
729 * Once set, ->i_wb never becomes NULL while the inode is alive.
730 * Start transaction iff ->i_wb is visible.
731 */
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400732 if (inode && inode_to_wb_is_valid(inode)) {
Tejun Heo5cb8b822015-05-28 14:50:54 -0400733 struct bdi_writeback *wb;
734 bool locked, congested;
735
736 wb = unlocked_inode_to_wb_begin(inode, &locked);
737 congested = wb_congested(wb, cong_bits);
738 unlocked_inode_to_wb_end(inode, locked);
739 return congested;
Tejun Heo703c2702015-05-22 17:13:44 -0400740 }
741
742 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
743}
744EXPORT_SYMBOL_GPL(inode_congested);
745
Tejun Heof2b65122015-05-22 17:13:55 -0400746/**
747 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
748 * @wb: target bdi_writeback to split @nr_pages to
749 * @nr_pages: number of pages to write for the whole bdi
750 *
751 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
752 * relation to the total write bandwidth of all wb's w/ dirty inodes on
753 * @wb->bdi.
754 */
755static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
756{
757 unsigned long this_bw = wb->avg_write_bandwidth;
758 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
759
760 if (nr_pages == LONG_MAX)
761 return LONG_MAX;
762
763 /*
764 * This may be called on clean wb's and proportional distribution
765 * may not make sense, just use the original @nr_pages in those
766 * cases. In general, we wanna err on the side of writing more.
767 */
768 if (!tot_bw || this_bw >= tot_bw)
769 return nr_pages;
770 else
771 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
772}
773
Tejun Heodb125362015-05-22 17:14:01 -0400774/**
Tejun Heodb125362015-05-22 17:14:01 -0400775 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
776 * @bdi: target backing_dev_info
777 * @base_work: wb_writeback_work to issue
778 * @skip_if_busy: skip wb's which already have writeback in progress
779 *
780 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
781 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
782 * distributed to the busy wbs according to each wb's proportion in the
783 * total active write bandwidth of @bdi.
784 */
785static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
786 struct wb_writeback_work *base_work,
787 bool skip_if_busy)
788{
Tejun Heob8175252015-10-02 14:47:05 -0400789 struct bdi_writeback *last_wb = NULL;
Tejun Heob33e18f2015-10-27 14:19:39 +0900790 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
791 struct bdi_writeback, bdi_node);
Tejun Heodb125362015-05-22 17:14:01 -0400792
793 might_sleep();
Tejun Heodb125362015-05-22 17:14:01 -0400794restart:
795 rcu_read_lock();
Tejun Heob8175252015-10-02 14:47:05 -0400796 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
Tejun Heo8a1270c2015-08-18 14:54:53 -0700797 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
798 struct wb_writeback_work fallback_work;
799 struct wb_writeback_work *work;
800 long nr_pages;
801
Tejun Heob8175252015-10-02 14:47:05 -0400802 if (last_wb) {
803 wb_put(last_wb);
804 last_wb = NULL;
805 }
806
Tejun Heo006a0972015-08-25 14:11:52 -0400807 /* SYNC_ALL writes out I_DIRTY_TIME too */
808 if (!wb_has_dirty_io(wb) &&
809 (base_work->sync_mode == WB_SYNC_NONE ||
810 list_empty(&wb->b_dirty_time)))
811 continue;
812 if (skip_if_busy && writeback_in_progress(wb))
Tejun Heodb125362015-05-22 17:14:01 -0400813 continue;
814
Tejun Heo8a1270c2015-08-18 14:54:53 -0700815 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
816
817 work = kmalloc(sizeof(*work), GFP_ATOMIC);
818 if (work) {
819 *work = *base_work;
820 work->nr_pages = nr_pages;
821 work->auto_free = 1;
822 wb_queue_work(wb, work);
823 continue;
Tejun Heodb125362015-05-22 17:14:01 -0400824 }
Tejun Heo8a1270c2015-08-18 14:54:53 -0700825
826 /* alloc failed, execute synchronously using on-stack fallback */
827 work = &fallback_work;
828 *work = *base_work;
829 work->nr_pages = nr_pages;
830 work->auto_free = 0;
831 work->done = &fallback_work_done;
832
833 wb_queue_work(wb, work);
834
Tejun Heob8175252015-10-02 14:47:05 -0400835 /*
836 * Pin @wb so that it stays on @bdi->wb_list. This allows
837 * continuing iteration from @wb after dropping and
838 * regrabbing rcu read lock.
839 */
840 wb_get(wb);
841 last_wb = wb;
842
Tejun Heo8a1270c2015-08-18 14:54:53 -0700843 rcu_read_unlock();
844 wb_wait_for_completion(bdi, &fallback_work_done);
845 goto restart;
Tejun Heodb125362015-05-22 17:14:01 -0400846 }
847 rcu_read_unlock();
Tejun Heob8175252015-10-02 14:47:05 -0400848
849 if (last_wb)
850 wb_put(last_wb);
Tejun Heodb125362015-05-22 17:14:01 -0400851}
852
Tejun Heoa1a0e232016-02-29 18:28:53 -0500853/**
854 * cgroup_writeback_umount - flush inode wb switches for umount
855 *
856 * This function is called when a super_block is about to be destroyed and
857 * flushes in-flight inode wb switches. An inode wb switch goes through
858 * RCU and then workqueue, so the two need to be flushed in order to ensure
859 * that all previously scheduled switches are finished. As wb switches are
860 * rare occurrences and synchronize_rcu() can take a while, perform
861 * flushing iff wb switches are in flight.
862 */
863void cgroup_writeback_umount(void)
864{
865 if (atomic_read(&isw_nr_in_flight)) {
866 synchronize_rcu();
867 flush_workqueue(isw_wq);
868 }
869}
870
871static int __init cgroup_writeback_init(void)
872{
873 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
874 if (!isw_wq)
875 return -ENOMEM;
876 return 0;
877}
878fs_initcall(cgroup_writeback_init);
879
Tejun Heof2b65122015-05-22 17:13:55 -0400880#else /* CONFIG_CGROUP_WRITEBACK */
881
Tejun Heo87e1d782015-05-28 14:50:52 -0400882static struct bdi_writeback *
883locked_inode_to_wb_and_lock_list(struct inode *inode)
884 __releases(&inode->i_lock)
885 __acquires(&wb->list_lock)
886{
887 struct bdi_writeback *wb = inode_to_wb(inode);
888
889 spin_unlock(&inode->i_lock);
890 spin_lock(&wb->list_lock);
891 return wb;
892}
893
894static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
895 __acquires(&wb->list_lock)
896{
897 struct bdi_writeback *wb = inode_to_wb(inode);
898
899 spin_lock(&wb->list_lock);
900 return wb;
901}
902
Tejun Heof2b65122015-05-22 17:13:55 -0400903static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
904{
905 return nr_pages;
906}
907
Tejun Heodb125362015-05-22 17:14:01 -0400908static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
909 struct wb_writeback_work *base_work,
910 bool skip_if_busy)
911{
912 might_sleep();
913
Tejun Heo006a0972015-08-25 14:11:52 -0400914 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
Tejun Heodb125362015-05-22 17:14:01 -0400915 base_work->auto_free = 0;
Tejun Heodb125362015-05-22 17:14:01 -0400916 wb_queue_work(&bdi->wb, base_work);
917 }
918}
919
Tejun Heo703c2702015-05-22 17:13:44 -0400920#endif /* CONFIG_CGROUP_WRITEBACK */
921
Tejun Heoc00ddad2015-05-22 17:13:51 -0400922void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
923 bool range_cyclic, enum wb_reason reason)
Jens Axboeb6e51312009-09-16 15:13:54 +0200924{
Tejun Heoc00ddad2015-05-22 17:13:51 -0400925 struct wb_writeback_work *work;
926
927 if (!wb_has_dirty_io(wb))
928 return;
929
930 /*
931 * This is WB_SYNC_NONE writeback, so if allocation fails just
932 * wakeup the thread for old dirty data writeback
933 */
934 work = kzalloc(sizeof(*work), GFP_ATOMIC);
935 if (!work) {
Tejun Heo5634cc22015-08-18 14:54:56 -0700936 trace_writeback_nowork(wb);
Tejun Heoc00ddad2015-05-22 17:13:51 -0400937 wb_wakeup(wb);
938 return;
939 }
940
941 work->sync_mode = WB_SYNC_NONE;
942 work->nr_pages = nr_pages;
943 work->range_cyclic = range_cyclic;
944 work->reason = reason;
Tejun Heoac7b19a2015-05-22 17:13:57 -0400945 work->auto_free = 1;
Tejun Heoc00ddad2015-05-22 17:13:51 -0400946
947 wb_queue_work(wb, work);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200948}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800949
Christoph Hellwigc5444192010-06-08 18:15:15 +0200950/**
Tejun Heo9ecf48662015-05-22 17:13:54 -0400951 * wb_start_background_writeback - start background writeback
952 * @wb: bdi_writback to write from
Christoph Hellwigc5444192010-06-08 18:15:15 +0200953 *
954 * Description:
Jan Kara65850272011-01-13 15:45:44 -0800955 * This makes sure WB_SYNC_NONE background writeback happens. When
Tejun Heo9ecf48662015-05-22 17:13:54 -0400956 * this function returns, it is only guaranteed that for given wb
Jan Kara65850272011-01-13 15:45:44 -0800957 * some IO is happening if we are over background dirty threshold.
958 * Caller need not hold sb s_umount semaphore.
Christoph Hellwigc5444192010-06-08 18:15:15 +0200959 */
Tejun Heo9ecf48662015-05-22 17:13:54 -0400960void wb_start_background_writeback(struct bdi_writeback *wb)
Christoph Hellwigc5444192010-06-08 18:15:15 +0200961{
Jan Kara65850272011-01-13 15:45:44 -0800962 /*
963 * We just wake up the flusher thread. It will perform background
964 * writeback as soon as there is no other work to do.
965 */
Tejun Heo5634cc22015-08-18 14:54:56 -0700966 trace_writeback_wake_background(wb);
Tejun Heo9ecf48662015-05-22 17:13:54 -0400967 wb_wakeup(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968}
969
970/*
Dave Chinnera66979a2011-03-22 22:23:41 +1100971 * Remove the inode from the writeback list it is on.
972 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500973void inode_io_list_del(struct inode *inode)
Dave Chinnera66979a2011-03-22 22:23:41 +1100974{
Tejun Heo87e1d782015-05-28 14:50:52 -0400975 struct bdi_writeback *wb;
Dave Chinnera66979a2011-03-22 22:23:41 +1100976
Tejun Heo87e1d782015-05-28 14:50:52 -0400977 wb = inode_to_wb_and_lock_list(inode);
Dave Chinnerc7f54082015-03-04 14:07:22 -0500978 inode_io_list_del_locked(inode, wb);
Tejun Heo52ebea72015-05-22 17:13:37 -0400979 spin_unlock(&wb->list_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600980}
Dave Chinnera66979a2011-03-22 22:23:41 +1100981
982/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700983 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
984 * furthest end of its superblock's dirty-inode list.
985 *
986 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200987 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700988 * the case then the inode must have been redirtied while it was being written
989 * out and we don't reset its dirtied_when.
990 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600991static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
Andrew Morton6610a0b2007-10-16 23:30:32 -0700992{
Jens Axboe03ba3782009-09-09 09:08:54 +0200993 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200994 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700995
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100996 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200997 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700998 inode->dirtied_when = jiffies;
999 }
Dave Chinnerc7f54082015-03-04 14:07:22 -05001000 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -07001001}
1002
1003/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001004 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001005 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001006static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001007{
Dave Chinnerc7f54082015-03-04 14:07:22 -05001008 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001009}
1010
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001011static void inode_sync_complete(struct inode *inode)
1012{
Jan Kara365b94ae2012-05-03 14:47:55 +02001013 inode->i_state &= ~I_SYNC;
Jan Kara4eff96d2012-11-26 16:29:51 -08001014 /* If inode is clean an unused, put it into LRU now... */
1015 inode_add_lru(inode);
Jan Kara365b94ae2012-05-03 14:47:55 +02001016 /* Waiters must see I_SYNC cleared before being woken up */
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001017 smp_mb();
1018 wake_up_bit(&inode->i_state, __I_SYNC);
1019}
1020
Jeff Laytond2caa3c52009-04-02 16:56:37 -07001021static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1022{
1023 bool ret = time_after(inode->dirtied_when, t);
1024#ifndef CONFIG_64BIT
1025 /*
1026 * For inodes being constantly redirtied, dirtied_when can get stuck.
1027 * It _appears_ to be in the future, but is actually in distant past.
1028 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +02001029 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -07001030 */
1031 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1032#endif
1033 return ret;
1034}
1035
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001036#define EXPIRE_DIRTY_ATIME 0x0001
1037
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001038/*
Wang Sheng-Hui0e2f2b22012-09-11 08:28:18 +08001039 * Move expired (dirtied before work->older_than_this) dirty inodes from
Jan Kara697e6fe2012-03-09 07:26:22 -08001040 * @delaying_queue to @dispatch_queue.
Fengguang Wu2c136572007-10-16 23:30:39 -07001041 */
Wu Fengguange84d0a42011-04-23 12:27:27 -06001042static int move_expired_inodes(struct list_head *delaying_queue,
Fengguang Wu2c136572007-10-16 23:30:39 -07001043 struct list_head *dispatch_queue,
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001044 int flags,
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001045 struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -07001046{
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001047 unsigned long *older_than_this = NULL;
1048 unsigned long expire_time;
Shaohua Li5c034492009-09-24 14:42:33 +02001049 LIST_HEAD(tmp);
1050 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +02001051 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +02001052 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +02001053 int do_sb_sort = 0;
Wu Fengguange84d0a42011-04-23 12:27:27 -06001054 int moved = 0;
Shaohua Li5c034492009-09-24 14:42:33 +02001055
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001056 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1057 older_than_this = work->older_than_this;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001058 else if (!work->for_sync) {
1059 expire_time = jiffies - (dirtytime_expire_interval * HZ);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001060 older_than_this = &expire_time;
1061 }
Fengguang Wu2c136572007-10-16 23:30:39 -07001062 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001063 inode = wb_inode(delaying_queue->prev);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001064 if (older_than_this &&
1065 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -07001066 break;
Dave Chinnerc7f54082015-03-04 14:07:22 -05001067 list_move(&inode->i_io_list, &tmp);
Jan Karaa8855992013-07-09 22:36:45 +08001068 moved++;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001069 if (flags & EXPIRE_DIRTY_ATIME)
1070 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
Jan Karaa8855992013-07-09 22:36:45 +08001071 if (sb_is_blkdev_sb(inode->i_sb))
1072 continue;
Jens Axboecf137302009-09-24 15:12:57 +02001073 if (sb && sb != inode->i_sb)
1074 do_sb_sort = 1;
1075 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001076 }
1077
Jens Axboecf137302009-09-24 15:12:57 +02001078 /* just one sb in list, splice to dispatch_queue and we're done */
1079 if (!do_sb_sort) {
1080 list_splice(&tmp, dispatch_queue);
Wu Fengguange84d0a42011-04-23 12:27:27 -06001081 goto out;
Jens Axboecf137302009-09-24 15:12:57 +02001082 }
1083
Shaohua Li5c034492009-09-24 14:42:33 +02001084 /* Move inodes from one superblock together */
1085 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001086 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001087 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001088 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +02001089 if (inode->i_sb == sb)
Dave Chinnerc7f54082015-03-04 14:07:22 -05001090 list_move(&inode->i_io_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +02001091 }
Fengguang Wu2c136572007-10-16 23:30:39 -07001092 }
Wu Fengguange84d0a42011-04-23 12:27:27 -06001093out:
1094 return moved;
Fengguang Wu2c136572007-10-16 23:30:39 -07001095}
1096
1097/*
1098 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001099 * Before
1100 * newly dirtied b_dirty b_io b_more_io
1101 * =============> gf edc BA
1102 * After
1103 * newly dirtied b_dirty b_io b_more_io
1104 * =============> g fBAedc
1105 * |
1106 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -07001107 */
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001108static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -07001109{
Wu Fengguange84d0a42011-04-23 12:27:27 -06001110 int moved;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001111
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001112 assert_spin_locked(&wb->list_lock);
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001113 list_splice_init(&wb->b_more_io, &wb->b_io);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001114 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1115 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1116 EXPIRE_DIRTY_ATIME, work);
Tejun Heod6c10f12015-05-22 17:13:45 -04001117 if (moved)
1118 wb_io_lists_populated(wb);
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001119 trace_writeback_queue_io(wb, work, moved);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001120}
1121
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001122static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001123{
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001124 int ret;
1125
1126 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1127 trace_writeback_write_inode_start(inode, wbc);
1128 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1129 trace_writeback_write_inode(inode, wbc);
1130 return ret;
1131 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001132 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -07001133}
1134
1135/*
Jan Kara169ebd92012-05-03 14:48:03 +02001136 * Wait for writeback on an inode to complete. Called with i_lock held.
1137 * Caller must make sure inode cannot go away when we drop i_lock.
Christoph Hellwig01c03192009-06-08 13:35:40 +02001138 */
Jan Kara169ebd92012-05-03 14:48:03 +02001139static void __inode_wait_for_writeback(struct inode *inode)
1140 __releases(inode->i_lock)
1141 __acquires(inode->i_lock)
Christoph Hellwig01c03192009-06-08 13:35:40 +02001142{
1143 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1144 wait_queue_head_t *wqh;
1145
1146 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +11001147 while (inode->i_state & I_SYNC) {
1148 spin_unlock(&inode->i_lock);
NeilBrown74316202014-07-07 15:16:04 +10001149 __wait_on_bit(wqh, &wq, bit_wait,
1150 TASK_UNINTERRUPTIBLE);
Dave Chinner250df6e2011-03-22 22:23:36 +11001151 spin_lock(&inode->i_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -07001152 }
Christoph Hellwig01c03192009-06-08 13:35:40 +02001153}
1154
1155/*
Jan Kara169ebd92012-05-03 14:48:03 +02001156 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1157 */
1158void inode_wait_for_writeback(struct inode *inode)
1159{
1160 spin_lock(&inode->i_lock);
1161 __inode_wait_for_writeback(inode);
1162 spin_unlock(&inode->i_lock);
1163}
1164
1165/*
1166 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1167 * held and drops it. It is aimed for callers not holding any inode reference
1168 * so once i_lock is dropped, inode can go away.
1169 */
1170static void inode_sleep_on_writeback(struct inode *inode)
1171 __releases(inode->i_lock)
1172{
1173 DEFINE_WAIT(wait);
1174 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1175 int sleep;
1176
1177 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1178 sleep = inode->i_state & I_SYNC;
1179 spin_unlock(&inode->i_lock);
1180 if (sleep)
1181 schedule();
1182 finish_wait(wqh, &wait);
1183}
1184
1185/*
Jan Karaccb26b52012-05-03 14:47:58 +02001186 * Find proper writeback list for the inode depending on its current state and
1187 * possibly also change of its state while we were doing writeback. Here we
1188 * handle things such as livelock prevention or fairness of writeback among
1189 * inodes. This function can be called only by flusher thread - noone else
1190 * processes all inodes in writeback lists and requeueing inodes behind flusher
1191 * thread's back can have unexpected consequences.
1192 */
1193static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1194 struct writeback_control *wbc)
1195{
1196 if (inode->i_state & I_FREEING)
1197 return;
1198
1199 /*
1200 * Sync livelock prevention. Each inode is tagged and synced in one
1201 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1202 * the dirty time to prevent enqueue and sync it again.
1203 */
1204 if ((inode->i_state & I_DIRTY) &&
1205 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1206 inode->dirtied_when = jiffies;
1207
Jan Kara4f8ad652012-05-03 14:48:00 +02001208 if (wbc->pages_skipped) {
1209 /*
1210 * writeback is not making progress due to locked
1211 * buffers. Skip this inode for now.
1212 */
1213 redirty_tail(inode, wb);
1214 return;
1215 }
1216
Jan Karaccb26b52012-05-03 14:47:58 +02001217 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1218 /*
1219 * We didn't write back all the pages. nfs_writepages()
1220 * sometimes bales out without doing anything.
1221 */
1222 if (wbc->nr_to_write <= 0) {
1223 /* Slice used up. Queue for next turn. */
1224 requeue_io(inode, wb);
1225 } else {
1226 /*
1227 * Writeback blocked by something other than
1228 * congestion. Delay the inode for some time to
1229 * avoid spinning on the CPU (100% iowait)
1230 * retrying writeback of the dirty page/inode
1231 * that cannot be performed immediately.
1232 */
1233 redirty_tail(inode, wb);
1234 }
1235 } else if (inode->i_state & I_DIRTY) {
1236 /*
1237 * Filesystems can dirty the inode during writeback operations,
1238 * such as delayed allocation during submission or metadata
1239 * updates after data IO completion.
1240 */
1241 redirty_tail(inode, wb);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001242 } else if (inode->i_state & I_DIRTY_TIME) {
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001243 inode->dirtied_when = jiffies;
Dave Chinnerc7f54082015-03-04 14:07:22 -05001244 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
Jan Karaccb26b52012-05-03 14:47:58 +02001245 } else {
1246 /* The inode is clean. Remove from writeback lists. */
Dave Chinnerc7f54082015-03-04 14:07:22 -05001247 inode_io_list_del_locked(inode, wb);
Jan Karaccb26b52012-05-03 14:47:58 +02001248 }
1249}
1250
1251/*
Jan Kara4f8ad652012-05-03 14:48:00 +02001252 * Write out an inode and its dirty pages. Do not update the writeback list
1253 * linkage. That is left to the caller. The caller is also responsible for
1254 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 */
1256static int
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001257__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 struct address_space *mapping = inode->i_mapping;
Wu Fengguang251d6a42010-12-01 17:33:37 -06001260 long nr_to_write = wbc->nr_to_write;
Christoph Hellwig01c03192009-06-08 13:35:40 +02001261 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 int ret;
1263
Jan Kara4f8ad652012-05-03 14:48:00 +02001264 WARN_ON(!(inode->i_state & I_SYNC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001266 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1267
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 ret = do_writepages(mapping, wbc);
1269
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001270 /*
1271 * Make sure to wait on the data before writing out the metadata.
1272 * This is important for filesystems that modify metadata on data
Dave Chinner7747bd42013-07-02 22:38:35 +10001273 * I/O completion. We don't do it for sync(2) writeback because it has a
1274 * separate, external IO completion path and ->sync_fs for guaranteeing
1275 * inode metadata is written back correctly.
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001276 */
Dave Chinner7747bd42013-07-02 22:38:35 +10001277 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001278 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 if (ret == 0)
1280 ret = err;
1281 }
1282
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001283 /*
1284 * Some filesystems may redirty the inode during the writeback
1285 * due to delalloc, clear dirty metadata flags right before
1286 * write_inode()
1287 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001288 spin_lock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001289
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001290 dirty = inode->i_state & I_DIRTY;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001291 if (inode->i_state & I_DIRTY_TIME) {
1292 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1293 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1294 unlikely(time_after(jiffies,
1295 (inode->dirtied_time_when +
1296 dirtytime_expire_interval * HZ)))) {
1297 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1298 trace_writeback_lazytime(inode);
1299 }
1300 } else
1301 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001302 inode->i_state &= ~dirty;
Tejun Heo9c6ac782014-10-24 15:38:21 -04001303
1304 /*
1305 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1306 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1307 * either they see the I_DIRTY bits cleared or we see the dirtied
1308 * inode.
1309 *
1310 * I_DIRTY_PAGES is always cleared together above even if @mapping
1311 * still has dirty pages. The flag is reinstated after smp_mb() if
1312 * necessary. This guarantees that either __mark_inode_dirty()
1313 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1314 */
1315 smp_mb();
1316
1317 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1318 inode->i_state |= I_DIRTY_PAGES;
1319
Dave Chinner250df6e2011-03-22 22:23:36 +11001320 spin_unlock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001321
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001322 if (dirty & I_DIRTY_TIME)
1323 mark_inode_dirty_sync(inode);
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001324 /* Don't write the inode if only I_DIRTY_PAGES was set */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001325 if (dirty & ~I_DIRTY_PAGES) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001326 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 if (ret == 0)
1328 ret = err;
1329 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001330 trace_writeback_single_inode(inode, wbc, nr_to_write);
1331 return ret;
1332}
1333
1334/*
1335 * Write out an inode's dirty pages. Either the caller has an active reference
1336 * on the inode or the inode has I_WILL_FREE set.
1337 *
1338 * This function is designed to be called for writing back one inode which
1339 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1340 * and does more profound writeback list handling in writeback_sb_inodes().
1341 */
Tejun Heoaaf25592016-03-18 13:52:04 -04001342static int writeback_single_inode(struct inode *inode,
1343 struct writeback_control *wbc)
Jan Kara4f8ad652012-05-03 14:48:00 +02001344{
Tejun Heoaaf25592016-03-18 13:52:04 -04001345 struct bdi_writeback *wb;
Jan Kara4f8ad652012-05-03 14:48:00 +02001346 int ret = 0;
1347
1348 spin_lock(&inode->i_lock);
1349 if (!atomic_read(&inode->i_count))
1350 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1351 else
1352 WARN_ON(inode->i_state & I_WILL_FREE);
1353
1354 if (inode->i_state & I_SYNC) {
1355 if (wbc->sync_mode != WB_SYNC_ALL)
1356 goto out;
1357 /*
Jan Kara169ebd92012-05-03 14:48:03 +02001358 * It's a data-integrity sync. We must wait. Since callers hold
1359 * inode reference or inode has I_WILL_FREE set, it cannot go
1360 * away under us.
Jan Kara4f8ad652012-05-03 14:48:00 +02001361 */
Jan Kara169ebd92012-05-03 14:48:03 +02001362 __inode_wait_for_writeback(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001363 }
1364 WARN_ON(inode->i_state & I_SYNC);
1365 /*
Jan Karaf9b0e052013-12-14 04:21:26 +08001366 * Skip inode if it is clean and we have no outstanding writeback in
1367 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1368 * function since flusher thread may be doing for example sync in
1369 * parallel and if we move the inode, it could get skipped. So here we
1370 * make sure inode is on some writeback list and leave it there unless
1371 * we have completely cleaned the inode.
Jan Kara4f8ad652012-05-03 14:48:00 +02001372 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001373 if (!(inode->i_state & I_DIRTY_ALL) &&
Jan Karaf9b0e052013-12-14 04:21:26 +08001374 (wbc->sync_mode != WB_SYNC_ALL ||
1375 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
Jan Kara4f8ad652012-05-03 14:48:00 +02001376 goto out;
1377 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001378 wbc_attach_and_unlock_inode(wbc, inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001379
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001380 ret = __writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381
Tejun Heob16b1de2015-06-02 08:39:48 -06001382 wbc_detach_inode(wbc);
Tejun Heoaaf25592016-03-18 13:52:04 -04001383
1384 wb = inode_to_wb_and_lock_list(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11001385 spin_lock(&inode->i_lock);
Jan Kara4f8ad652012-05-03 14:48:00 +02001386 /*
1387 * If inode is clean, remove it from writeback lists. Otherwise don't
1388 * touch it. See comment above for explanation.
1389 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001390 if (!(inode->i_state & I_DIRTY_ALL))
Dave Chinnerc7f54082015-03-04 14:07:22 -05001391 inode_io_list_del_locked(inode, wb);
Jan Kara4f8ad652012-05-03 14:48:00 +02001392 spin_unlock(&wb->list_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001393 inode_sync_complete(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001394out:
1395 spin_unlock(&inode->i_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 return ret;
1397}
1398
Tejun Heoa88a3412015-05-22 17:13:28 -04001399static long writeback_chunk_size(struct bdi_writeback *wb,
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001400 struct wb_writeback_work *work)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001401{
1402 long pages;
1403
1404 /*
1405 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1406 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1407 * here avoids calling into writeback_inodes_wb() more than once.
1408 *
1409 * The intended call sequence for WB_SYNC_ALL writeback is:
1410 *
1411 * wb_writeback()
1412 * writeback_sb_inodes() <== called only once
1413 * write_cache_pages() <== called once for each inode
1414 * (quickly) tag currently dirty pages
1415 * (maybe slowly) sync all tagged pages
1416 */
1417 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1418 pages = LONG_MAX;
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001419 else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001420 pages = min(wb->avg_write_bandwidth / 2,
Tejun Heodcc25ae2015-05-22 18:23:22 -04001421 global_wb_domain.dirty_limit / DIRTY_SCOPE);
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001422 pages = min(pages, work->nr_pages);
1423 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1424 MIN_WRITEBACK_PAGES);
1425 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001426
1427 return pages;
1428}
1429
Jens Axboe03ba3782009-09-09 09:08:54 +02001430/*
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001431 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001432 *
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001433 * Return the number of pages and/or inodes written.
Linus Torvalds0ba13fd2015-09-11 13:26:39 -07001434 *
1435 * NOTE! This is called with wb->list_lock held, and will
1436 * unlock and relock that for each inode it ends up doing
1437 * IO for.
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001438 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001439static long writeback_sb_inodes(struct super_block *sb,
1440 struct bdi_writeback *wb,
1441 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001442{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001443 struct writeback_control wbc = {
1444 .sync_mode = work->sync_mode,
1445 .tagged_writepages = work->tagged_writepages,
1446 .for_kupdate = work->for_kupdate,
1447 .for_background = work->for_background,
Dave Chinner7747bd42013-07-02 22:38:35 +10001448 .for_sync = work->for_sync,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001449 .range_cyclic = work->range_cyclic,
1450 .range_start = 0,
1451 .range_end = LLONG_MAX,
1452 };
1453 unsigned long start_time = jiffies;
1454 long write_chunk;
1455 long wrote = 0; /* count both pages and inodes */
1456
Jens Axboe03ba3782009-09-09 09:08:54 +02001457 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001458 struct inode *inode = wb_inode(wb->b_io.prev);
Tejun Heoaaf25592016-03-18 13:52:04 -04001459 struct bdi_writeback *tmp_wb;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001460
1461 if (inode->i_sb != sb) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001462 if (work->sb) {
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001463 /*
1464 * We only want to write back data for this
1465 * superblock, move all inodes not belonging
1466 * to it back onto the dirty list.
1467 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001468 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001469 continue;
1470 }
1471
1472 /*
1473 * The inode belongs to a different superblock.
1474 * Bounce back to the caller to unpin this and
1475 * pin the next superblock.
1476 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001477 break;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001478 }
1479
Christoph Hellwig9843b762010-10-24 19:40:46 +02001480 /*
Wanpeng Li331cbde2012-06-09 11:10:55 +08001481 * Don't bother with new inodes or inodes being freed, first
1482 * kind does not need periodic writeout yet, and for the latter
Christoph Hellwig9843b762010-10-24 19:40:46 +02001483 * kind writeout is handled by the freer.
1484 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001485 spin_lock(&inode->i_lock);
Christoph Hellwig9843b762010-10-24 19:40:46 +02001486 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Dave Chinner250df6e2011-03-22 22:23:36 +11001487 spin_unlock(&inode->i_lock);
Wu Fengguangfcc5c222011-07-11 23:08:50 -07001488 redirty_tail(inode, wb);
Nick Piggin7ef0d732009-03-12 14:31:38 -07001489 continue;
1490 }
Jan Karacc1676d2012-05-03 14:47:56 +02001491 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1492 /*
1493 * If this inode is locked for writeback and we are not
1494 * doing writeback-for-data-integrity, move it to
1495 * b_more_io so that writeback can proceed with the
1496 * other inodes on s_io.
1497 *
1498 * We'll have another go at writing back this inode
1499 * when we completed a full scan of b_io.
1500 */
1501 spin_unlock(&inode->i_lock);
1502 requeue_io(inode, wb);
1503 trace_writeback_sb_inodes_requeue(inode);
1504 continue;
1505 }
Jan Karaf0d07b72012-05-03 14:47:59 +02001506 spin_unlock(&wb->list_lock);
1507
Jan Kara4f8ad652012-05-03 14:48:00 +02001508 /*
1509 * We already requeued the inode if it had I_SYNC set and we
1510 * are doing WB_SYNC_NONE writeback. So this catches only the
1511 * WB_SYNC_ALL case.
1512 */
Jan Kara169ebd92012-05-03 14:48:03 +02001513 if (inode->i_state & I_SYNC) {
1514 /* Wait for I_SYNC. This function drops i_lock... */
1515 inode_sleep_on_writeback(inode);
1516 /* Inode may be gone, start again */
Jan Karaead188f2012-06-08 17:07:36 +02001517 spin_lock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001518 continue;
1519 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001520 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001521 wbc_attach_and_unlock_inode(&wbc, inode);
Jan Kara169ebd92012-05-03 14:48:03 +02001522
Tejun Heoa88a3412015-05-22 17:13:28 -04001523 write_chunk = writeback_chunk_size(wb, work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001524 wbc.nr_to_write = write_chunk;
1525 wbc.pages_skipped = 0;
Dave Chinner250df6e2011-03-22 22:23:36 +11001526
Jan Kara169ebd92012-05-03 14:48:03 +02001527 /*
1528 * We use I_SYNC to pin the inode in memory. While it is set
1529 * evict_inode() will wait so the inode cannot be freed.
1530 */
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001531 __writeback_single_inode(inode, &wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001532
Tejun Heob16b1de2015-06-02 08:39:48 -06001533 wbc_detach_inode(&wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001534 work->nr_pages -= write_chunk - wbc.nr_to_write;
1535 wrote += write_chunk - wbc.nr_to_write;
Chris Mason590dca32015-09-18 13:35:08 -04001536
1537 if (need_resched()) {
1538 /*
1539 * We're trying to balance between building up a nice
1540 * long list of IOs to improve our merge rate, and
1541 * getting those IOs out quickly for anyone throttling
1542 * in balance_dirty_pages(). cond_resched() doesn't
1543 * unplug, so get our IOs out the door before we
1544 * give up the CPU.
1545 */
1546 blk_flush_plug(current);
1547 cond_resched();
1548 }
1549
Tejun Heoaaf25592016-03-18 13:52:04 -04001550 /*
1551 * Requeue @inode if still dirty. Be careful as @inode may
1552 * have been switched to another wb in the meantime.
1553 */
1554 tmp_wb = inode_to_wb_and_lock_list(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001555 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001556 if (!(inode->i_state & I_DIRTY_ALL))
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001557 wrote++;
Tejun Heoaaf25592016-03-18 13:52:04 -04001558 requeue_inode(inode, tmp_wb, &wbc);
Jan Kara4f8ad652012-05-03 14:48:00 +02001559 inode_sync_complete(inode);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001560 spin_unlock(&inode->i_lock);
Chris Mason590dca32015-09-18 13:35:08 -04001561
Tejun Heoaaf25592016-03-18 13:52:04 -04001562 if (unlikely(tmp_wb != wb)) {
1563 spin_unlock(&tmp_wb->list_lock);
1564 spin_lock(&wb->list_lock);
1565 }
1566
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001567 /*
1568 * bail out to wb_writeback() often enough to check
1569 * background threshold and other termination conditions.
1570 */
1571 if (wrote) {
1572 if (time_is_before_jiffies(start_time + HZ / 10UL))
1573 break;
1574 if (work->nr_pages <= 0)
1575 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -08001576 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001578 return wrote;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001579}
Nick Piggin38f21972009-01-06 14:40:25 -08001580
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001581static long __writeback_inodes_wb(struct bdi_writeback *wb,
1582 struct wb_writeback_work *work)
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001583{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001584 unsigned long start_time = jiffies;
1585 long wrote = 0;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001586
1587 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001588 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001589 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001590
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001591 if (!trylock_super(sb)) {
Wu Fengguang0e995812011-07-29 22:14:35 -06001592 /*
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001593 * trylock_super() may fail consistently due to
Wu Fengguang0e995812011-07-29 22:14:35 -06001594 * s_umount being grabbed by someone else. Don't use
1595 * requeue_io() to avoid busy retrying the inode/sb.
1596 */
1597 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001598 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001599 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001600 wrote += writeback_sb_inodes(sb, wb, work);
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001601 up_read(&sb->s_umount);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001602
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001603 /* refer to the same tests at the end of writeback_sb_inodes */
1604 if (wrote) {
1605 if (time_is_before_jiffies(start_time + HZ / 10UL))
1606 break;
1607 if (work->nr_pages <= 0)
1608 break;
1609 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001610 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001611 /* Leave any unwritten inodes on b_io */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001612 return wrote;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001613}
1614
Wanpeng Li7d9f0732013-09-11 14:22:40 -07001615static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001616 enum wb_reason reason)
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001617{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001618 struct wb_writeback_work work = {
1619 .nr_pages = nr_pages,
1620 .sync_mode = WB_SYNC_NONE,
1621 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001622 .reason = reason,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001623 };
Linus Torvalds505a6662015-09-11 13:37:19 -07001624 struct blk_plug plug;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001625
Linus Torvalds505a6662015-09-11 13:37:19 -07001626 blk_start_plug(&plug);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001627 spin_lock(&wb->list_lock);
Wu Fengguang424b3512010-07-21 20:11:53 -06001628 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001629 queue_io(wb, &work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001630 __writeback_inodes_wb(wb, &work);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001631 spin_unlock(&wb->list_lock);
Linus Torvalds505a6662015-09-11 13:37:19 -07001632 blk_finish_plug(&plug);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001633
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001634 return nr_pages - work.nr_pages;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001635}
1636
Jens Axboe03ba3782009-09-09 09:08:54 +02001637/*
1638 * Explicit flushing or periodic writeback of "old" data.
1639 *
1640 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1641 * dirtying-time in the inode's address_space. So this periodic writeback code
1642 * just walks the superblock inode list, writing back any inodes which are
1643 * older than a specific point in time.
1644 *
1645 * Try to run once per dirty_writeback_interval. But if a writeback event
1646 * takes longer than a dirty_writeback_interval interval, then leave a
1647 * one-second gap.
1648 *
1649 * older_than_this takes precedence over nr_to_write. So we'll only write back
1650 * all dirty pages if they are all attached to "old" mappings.
1651 */
Jens Axboec4a77a62009-09-16 15:18:25 +02001652static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001653 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001654{
Wu Fengguange98be2d2010-08-29 11:22:30 -06001655 unsigned long wb_start = jiffies;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001656 long nr_pages = work->nr_pages;
Jan Kara0dc83bd2014-02-21 11:19:04 +01001657 unsigned long oldest_jif;
Jan Karaa5989bd2009-09-16 19:22:48 +02001658 struct inode *inode;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001659 long progress;
Linus Torvalds505a6662015-09-11 13:37:19 -07001660 struct blk_plug plug;
Jens Axboe03ba3782009-09-09 09:08:54 +02001661
Jan Kara0dc83bd2014-02-21 11:19:04 +01001662 oldest_jif = jiffies;
1663 work->older_than_this = &oldest_jif;
Jens Axboe03ba3782009-09-09 09:08:54 +02001664
Linus Torvalds505a6662015-09-11 13:37:19 -07001665 blk_start_plug(&plug);
Wu Fengguange8dfc302011-04-21 12:06:32 -06001666 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001667 for (;;) {
1668 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001669 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +02001670 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001671 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +02001672 break;
1673
1674 /*
Jan Karaaa373cf2011-01-13 15:45:47 -08001675 * Background writeout and kupdate-style writeback may
1676 * run forever. Stop them if there is other work to do
1677 * so that e.g. sync can proceed. They'll be restarted
1678 * after the other works are all done.
1679 */
1680 if ((work->for_background || work->for_kupdate) &&
Tejun Heof0054bb2015-05-22 17:13:30 -04001681 !list_empty(&wb->work_list))
Jan Karaaa373cf2011-01-13 15:45:47 -08001682 break;
1683
1684 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001685 * For background writeout, stop when we are below the
1686 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +02001687 */
Tejun Heoaa661bb2015-05-22 18:23:31 -04001688 if (work->for_background && !wb_over_bg_thresh(wb))
Jens Axboe03ba3782009-09-09 09:08:54 +02001689 break;
1690
Jan Kara1bc36b62011-10-19 11:44:41 +02001691 /*
1692 * Kupdate and background works are special and we want to
1693 * include all inodes that need writing. Livelock avoidance is
1694 * handled by these works yielding to any other work so we are
1695 * safe.
1696 */
Wu Fengguangba9aa832010-07-21 20:32:30 -06001697 if (work->for_kupdate) {
Jan Kara0dc83bd2014-02-21 11:19:04 +01001698 oldest_jif = jiffies -
Wu Fengguangba9aa832010-07-21 20:32:30 -06001699 msecs_to_jiffies(dirty_expire_interval * 10);
Jan Kara1bc36b62011-10-19 11:44:41 +02001700 } else if (work->for_background)
Jan Kara0dc83bd2014-02-21 11:19:04 +01001701 oldest_jif = jiffies;
Dave Chinner028c2dd2010-07-07 13:24:07 +10001702
Tejun Heo5634cc22015-08-18 14:54:56 -07001703 trace_writeback_start(wb, work);
Wu Fengguange8dfc302011-04-21 12:06:32 -06001704 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001705 queue_io(wb, work);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001706 if (work->sb)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001707 progress = writeback_sb_inodes(work->sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001708 else
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001709 progress = __writeback_inodes_wb(wb, work);
Tejun Heo5634cc22015-08-18 14:54:56 -07001710 trace_writeback_written(wb, work);
Dave Chinner028c2dd2010-07-07 13:24:07 +10001711
Wu Fengguange98be2d2010-08-29 11:22:30 -06001712 wb_update_bandwidth(wb, wb_start);
Jens Axboe03ba3782009-09-09 09:08:54 +02001713
1714 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +02001715 * Did we write something? Try for more
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001716 *
1717 * Dirty inodes are moved to b_io for writeback in batches.
1718 * The completion of the current batch does not necessarily
1719 * mean the overall work is done. So we keep looping as long
1720 * as made some progress on cleaning pages or inodes.
Jens Axboe71fd05a2009-09-23 19:32:26 +02001721 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001722 if (progress)
Jens Axboe71fd05a2009-09-23 19:32:26 +02001723 continue;
1724 /*
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001725 * No more inodes for IO, bail
Jens Axboe71fd05a2009-09-23 19:32:26 +02001726 */
Wu Fengguangb7a24412010-07-21 22:19:51 -06001727 if (list_empty(&wb->b_more_io))
Jens Axboe03ba3782009-09-09 09:08:54 +02001728 break;
1729 /*
Jens Axboe8010c3b2009-09-15 20:04:57 +02001730 * Nothing written. Wait for some inode to
1731 * become available for writeback. Otherwise
1732 * we'll just busyloop.
Jens Axboe03ba3782009-09-09 09:08:54 +02001733 */
Jens Axboe03ba3782009-09-09 09:08:54 +02001734 if (!list_empty(&wb->b_more_io)) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001735 trace_writeback_wait(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001736 inode = wb_inode(wb->b_more_io.prev);
Dave Chinner250df6e2011-03-22 22:23:36 +11001737 spin_lock(&inode->i_lock);
Jan Karaf0d07b72012-05-03 14:47:59 +02001738 spin_unlock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001739 /* This function drops i_lock... */
1740 inode_sleep_on_writeback(inode);
Jan Karaf0d07b72012-05-03 14:47:59 +02001741 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001742 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001743 }
Wu Fengguange8dfc302011-04-21 12:06:32 -06001744 spin_unlock(&wb->list_lock);
Linus Torvalds505a6662015-09-11 13:37:19 -07001745 blk_finish_plug(&plug);
Jens Axboe03ba3782009-09-09 09:08:54 +02001746
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001747 return nr_pages - work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +02001748}
1749
1750/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001751 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +02001752 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001753static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001754{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001755 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +02001756
Tejun Heof0054bb2015-05-22 17:13:30 -04001757 spin_lock_bh(&wb->work_lock);
1758 if (!list_empty(&wb->work_list)) {
1759 work = list_entry(wb->work_list.next,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001760 struct wb_writeback_work, list);
1761 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +02001762 }
Tejun Heof0054bb2015-05-22 17:13:30 -04001763 spin_unlock_bh(&wb->work_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001764 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +02001765}
1766
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001767/*
1768 * Add in the number of potentially dirty inodes, because each inode
1769 * write can dirty pagecache in the underlying blockdev.
1770 */
1771static unsigned long get_nr_dirty_pages(void)
1772{
1773 return global_page_state(NR_FILE_DIRTY) +
1774 global_page_state(NR_UNSTABLE_NFS) +
1775 get_nr_dirty_inodes();
1776}
1777
Jan Kara65850272011-01-13 15:45:44 -08001778static long wb_check_background_flush(struct bdi_writeback *wb)
1779{
Tejun Heoaa661bb2015-05-22 18:23:31 -04001780 if (wb_over_bg_thresh(wb)) {
Jan Kara65850272011-01-13 15:45:44 -08001781
1782 struct wb_writeback_work work = {
1783 .nr_pages = LONG_MAX,
1784 .sync_mode = WB_SYNC_NONE,
1785 .for_background = 1,
1786 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001787 .reason = WB_REASON_BACKGROUND,
Jan Kara65850272011-01-13 15:45:44 -08001788 };
1789
1790 return wb_writeback(wb, &work);
1791 }
1792
1793 return 0;
1794}
1795
Jens Axboe03ba3782009-09-09 09:08:54 +02001796static long wb_check_old_data_flush(struct bdi_writeback *wb)
1797{
1798 unsigned long expired;
1799 long nr_pages;
1800
Jens Axboe69b62d02010-05-17 12:51:03 +02001801 /*
1802 * When set to zero, disable periodic writeback
1803 */
1804 if (!dirty_writeback_interval)
1805 return 0;
1806
Jens Axboe03ba3782009-09-09 09:08:54 +02001807 expired = wb->last_old_flush +
1808 msecs_to_jiffies(dirty_writeback_interval * 10);
1809 if (time_before(jiffies, expired))
1810 return 0;
1811
1812 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001813 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +02001814
Jens Axboec4a77a62009-09-16 15:18:25 +02001815 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001816 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +02001817 .nr_pages = nr_pages,
1818 .sync_mode = WB_SYNC_NONE,
1819 .for_kupdate = 1,
1820 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001821 .reason = WB_REASON_PERIODIC,
Jens Axboec4a77a62009-09-16 15:18:25 +02001822 };
1823
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001824 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +02001825 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001826
1827 return 0;
1828}
1829
1830/*
1831 * Retrieve work items and do the writeback they describe
1832 */
Wanpeng Li25d130b2013-07-08 16:00:14 -07001833static long wb_do_writeback(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001834{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001835 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +02001836 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +02001837
Tejun Heo44522262015-05-22 17:13:26 -04001838 set_bit(WB_writeback_running, &wb->state);
Tejun Heof0054bb2015-05-22 17:13:30 -04001839 while ((work = get_next_work_item(wb)) != NULL) {
Tejun Heocc395d72015-05-22 17:13:58 -04001840 struct wb_completion *done = work->done;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001841
Tejun Heo5634cc22015-08-18 14:54:56 -07001842 trace_writeback_exec(wb, work);
Dave Chinner455b2862010-07-07 13:24:06 +10001843
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001844 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001845
Tejun Heo8a1270c2015-08-18 14:54:53 -07001846 if (work->auto_free)
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001847 kfree(work);
Tejun Heocc395d72015-05-22 17:13:58 -04001848 if (done && atomic_dec_and_test(&done->cnt))
1849 wake_up_all(&wb->bdi->wb_waitq);
Jens Axboe03ba3782009-09-09 09:08:54 +02001850 }
1851
1852 /*
1853 * Check for periodic writeback, kupdated() style
1854 */
1855 wrote += wb_check_old_data_flush(wb);
Jan Kara65850272011-01-13 15:45:44 -08001856 wrote += wb_check_background_flush(wb);
Tejun Heo44522262015-05-22 17:13:26 -04001857 clear_bit(WB_writeback_running, &wb->state);
Jens Axboe03ba3782009-09-09 09:08:54 +02001858
1859 return wrote;
1860}
1861
1862/*
1863 * Handle writeback of dirty data for the device backed by this bdi. Also
Tejun Heo839a8e82013-04-01 19:08:06 -07001864 * reschedules periodically and does kupdated style flushing.
Jens Axboe03ba3782009-09-09 09:08:54 +02001865 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001866void wb_workfn(struct work_struct *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001867{
Tejun Heo839a8e82013-04-01 19:08:06 -07001868 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1869 struct bdi_writeback, dwork);
Jens Axboe03ba3782009-09-09 09:08:54 +02001870 long pages_written;
1871
Tejun Heof0054bb2015-05-22 17:13:30 -04001872 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
Peter Zijlstra766f9162010-10-26 14:22:45 -07001873 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +02001874
Tejun Heo839a8e82013-04-01 19:08:06 -07001875 if (likely(!current_is_workqueue_rescuer() ||
Tejun Heo44522262015-05-22 17:13:26 -04001876 !test_bit(WB_registered, &wb->state))) {
Artem Bityutskiy64677162010-07-25 14:29:22 +03001877 /*
Tejun Heof0054bb2015-05-22 17:13:30 -04001878 * The normal path. Keep writing back @wb until its
Tejun Heo839a8e82013-04-01 19:08:06 -07001879 * work_list is empty. Note that this path is also taken
Tejun Heof0054bb2015-05-22 17:13:30 -04001880 * if @wb is shutting down even when we're running off the
Tejun Heo839a8e82013-04-01 19:08:06 -07001881 * rescuer as work_list needs to be drained.
Artem Bityutskiy64677162010-07-25 14:29:22 +03001882 */
Tejun Heo839a8e82013-04-01 19:08:06 -07001883 do {
Wanpeng Li25d130b2013-07-08 16:00:14 -07001884 pages_written = wb_do_writeback(wb);
Tejun Heo839a8e82013-04-01 19:08:06 -07001885 trace_writeback_pages_written(pages_written);
Tejun Heof0054bb2015-05-22 17:13:30 -04001886 } while (!list_empty(&wb->work_list));
Tejun Heo839a8e82013-04-01 19:08:06 -07001887 } else {
1888 /*
1889 * bdi_wq can't get enough workers and we're running off
1890 * the emergency worker. Don't hog it. Hopefully, 1024 is
1891 * enough for efficient IO.
1892 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001893 pages_written = writeback_inodes_wb(wb, 1024,
Tejun Heo839a8e82013-04-01 19:08:06 -07001894 WB_REASON_FORKER_THREAD);
Dave Chinner455b2862010-07-07 13:24:06 +10001895 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +02001896 }
1897
Tejun Heof0054bb2015-05-22 17:13:30 -04001898 if (!list_empty(&wb->work_list))
Derek Basehore6ca738d2014-04-03 14:46:22 -07001899 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1900 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Tejun Heof0054bb2015-05-22 17:13:30 -04001901 wb_wakeup_delayed(wb);
Dave Chinner455b2862010-07-07 13:24:06 +10001902
Tejun Heo839a8e82013-04-01 19:08:06 -07001903 current->flags &= ~PF_SWAPWRITE;
Jens Axboe03ba3782009-09-09 09:08:54 +02001904}
1905
1906/*
Jens Axboe03ba3782009-09-09 09:08:54 +02001907 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1908 * the whole world.
1909 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001910void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
Jens Axboe03ba3782009-09-09 09:08:54 +02001911{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001912 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001913
Jan Kara47df3dd2013-09-11 14:22:22 -07001914 if (!nr_pages)
1915 nr_pages = get_nr_dirty_pages();
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001916
1917 rcu_read_lock();
Tejun Heof2b65122015-05-22 17:13:55 -04001918 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1919 struct bdi_writeback *wb;
Tejun Heof2b65122015-05-22 17:13:55 -04001920
1921 if (!bdi_has_dirty_io(bdi))
1922 continue;
1923
Tejun Heob8175252015-10-02 14:47:05 -04001924 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
Tejun Heof2b65122015-05-22 17:13:55 -04001925 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1926 false, reason);
1927 }
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001928 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02001929}
1930
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001931/*
1932 * Wake up bdi's periodically to make sure dirtytime inodes gets
1933 * written back periodically. We deliberately do *not* check the
1934 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1935 * kernel to be constantly waking up once there are any dirtytime
1936 * inodes on the system. So instead we define a separate delayed work
1937 * function which gets called much more rarely. (By default, only
1938 * once every 12 hours.)
1939 *
1940 * If there is any other write activity going on in the file system,
1941 * this function won't be necessary. But if the only thing that has
1942 * happened on the file system is a dirtytime inode caused by an atime
1943 * update, we need this infrastructure below to make sure that inode
1944 * eventually gets pushed out to disk.
1945 */
1946static void wakeup_dirtytime_writeback(struct work_struct *w);
1947static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1948
1949static void wakeup_dirtytime_writeback(struct work_struct *w)
1950{
1951 struct backing_dev_info *bdi;
1952
1953 rcu_read_lock();
1954 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Tejun Heo001fe6f2015-05-22 17:13:56 -04001955 struct bdi_writeback *wb;
Tejun Heo001fe6f2015-05-22 17:13:56 -04001956
Tejun Heob8175252015-10-02 14:47:05 -04001957 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
Tejun Heo6fdf8602015-09-29 12:47:51 -04001958 if (!list_empty(&wb->b_dirty_time))
1959 wb_wakeup(wb);
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001960 }
1961 rcu_read_unlock();
1962 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1963}
1964
1965static int __init start_dirtytime_writeback(void)
1966{
1967 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1968 return 0;
1969}
1970__initcall(start_dirtytime_writeback);
1971
Theodore Ts'o1efff912015-03-17 12:23:32 -04001972int dirtytime_interval_handler(struct ctl_table *table, int write,
1973 void __user *buffer, size_t *lenp, loff_t *ppos)
1974{
1975 int ret;
1976
1977 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1978 if (ret == 0 && write)
1979 mod_delayed_work(system_wq, &dirtytime_work, 0);
1980 return ret;
1981}
1982
Jens Axboe03ba3782009-09-09 09:08:54 +02001983static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1984{
1985 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1986 struct dentry *dentry;
1987 const char *name = "?";
1988
1989 dentry = d_find_alias(inode);
1990 if (dentry) {
1991 spin_lock(&dentry->d_lock);
1992 name = (const char *) dentry->d_name.name;
1993 }
1994 printk(KERN_DEBUG
1995 "%s(%d): dirtied inode %lu (%s) on %s\n",
1996 current->comm, task_pid_nr(current), inode->i_ino,
1997 name, inode->i_sb->s_id);
1998 if (dentry) {
1999 spin_unlock(&dentry->d_lock);
2000 dput(dentry);
2001 }
2002 }
2003}
2004
2005/**
2006 * __mark_inode_dirty - internal function
2007 * @inode: inode to mark
2008 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2009 * Mark an inode as dirty. Callers should use mark_inode_dirty or
2010 * mark_inode_dirty_sync.
2011 *
2012 * Put the inode on the super block's dirty list.
2013 *
2014 * CAREFUL! We mark it dirty unconditionally, but move it onto the
2015 * dirty list only if it is hashed or if it refers to a blockdev.
2016 * If it was not hashed, it will never be added to the dirty list
2017 * even if it is later hashed, as it will have been marked dirty already.
2018 *
2019 * In short, make sure you hash any inodes _before_ you start marking
2020 * them dirty.
2021 *
Jens Axboe03ba3782009-09-09 09:08:54 +02002022 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2023 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
2024 * the kernel-internal blockdev inode represents the dirtying time of the
2025 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
2026 * page->mapping->host, so the page-dirtying time is recorded in the internal
2027 * blockdev inode.
2028 */
2029void __mark_inode_dirty(struct inode *inode, int flags)
2030{
Randy Dunlapdbce03b2015-11-09 14:58:00 -08002031#define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
Jens Axboe03ba3782009-09-09 09:08:54 +02002032 struct super_block *sb = inode->i_sb;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002033 int dirtytime;
2034
2035 trace_writeback_mark_inode_dirty(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02002036
2037 /*
2038 * Don't do this for I_DIRTY_PAGES - that doesn't actually
2039 * dirty the inode itself
2040 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002041 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002042 trace_writeback_dirty_inode_start(inode, flags);
2043
Jens Axboe03ba3782009-09-09 09:08:54 +02002044 if (sb->s_op->dirty_inode)
Christoph Hellwigaa385722011-05-27 06:53:02 -04002045 sb->s_op->dirty_inode(inode, flags);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002046
2047 trace_writeback_dirty_inode(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02002048 }
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002049 if (flags & I_DIRTY_INODE)
2050 flags &= ~I_DIRTY_TIME;
2051 dirtytime = flags & I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02002052
2053 /*
Tejun Heo9c6ac782014-10-24 15:38:21 -04002054 * Paired with smp_mb() in __writeback_single_inode() for the
2055 * following lockless i_state test. See there for details.
Jens Axboe03ba3782009-09-09 09:08:54 +02002056 */
2057 smp_mb();
2058
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002059 if (((inode->i_state & flags) == flags) ||
2060 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
Jens Axboe03ba3782009-09-09 09:08:54 +02002061 return;
2062
2063 if (unlikely(block_dump))
2064 block_dump___mark_inode_dirty(inode);
2065
Dave Chinner250df6e2011-03-22 22:23:36 +11002066 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002067 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2068 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002069 if ((inode->i_state & flags) != flags) {
2070 const int was_dirty = inode->i_state & I_DIRTY;
2071
Tejun Heo52ebea72015-05-22 17:13:37 -04002072 inode_attach_wb(inode, NULL);
2073
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002074 if (flags & I_DIRTY_INODE)
2075 inode->i_state &= ~I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02002076 inode->i_state |= flags;
2077
2078 /*
2079 * If the inode is being synced, just update its dirty state.
2080 * The unlocker will place the inode on the appropriate
2081 * superblock list, based upon its state.
2082 */
2083 if (inode->i_state & I_SYNC)
Dave Chinner250df6e2011-03-22 22:23:36 +11002084 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002085
2086 /*
2087 * Only add valid (hashed) inodes to the superblock's
2088 * dirty list. Add blockdev inodes as well.
2089 */
2090 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382cb2010-10-23 15:19:20 -04002091 if (inode_unhashed(inode))
Dave Chinner250df6e2011-03-22 22:23:36 +11002092 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002093 }
Al Viroa4ffdde2010-06-02 17:38:30 -04002094 if (inode->i_state & I_FREEING)
Dave Chinner250df6e2011-03-22 22:23:36 +11002095 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002096
2097 /*
2098 * If the inode was already on b_dirty/b_io/b_more_io, don't
2099 * reposition it (that would break b_dirty time-ordering).
2100 */
2101 if (!was_dirty) {
Tejun Heo87e1d782015-05-28 14:50:52 -04002102 struct bdi_writeback *wb;
Tejun Heod6c10f12015-05-22 17:13:45 -04002103 struct list_head *dirty_list;
Dave Chinnera66979a2011-03-22 22:23:41 +11002104 bool wakeup_bdi = false;
Jens Axboe500b0672009-09-09 09:10:25 +02002105
Tejun Heo87e1d782015-05-28 14:50:52 -04002106 wb = locked_inode_to_wb_and_lock_list(inode);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002107
Tejun Heo07472592015-05-22 17:14:02 -04002108 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2109 !test_bit(WB_registered, &wb->state),
2110 "bdi-%s not registered\n", wb->bdi->name);
Jens Axboe03ba3782009-09-09 09:08:54 +02002111
2112 inode->dirtied_when = jiffies;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002113 if (dirtytime)
2114 inode->dirtied_time_when = jiffies;
Tejun Heod6c10f12015-05-22 17:13:45 -04002115
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002116 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
Tejun Heo07472592015-05-22 17:14:02 -04002117 dirty_list = &wb->b_dirty;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002118 else
Tejun Heo07472592015-05-22 17:14:02 -04002119 dirty_list = &wb->b_dirty_time;
Tejun Heod6c10f12015-05-22 17:13:45 -04002120
Dave Chinnerc7f54082015-03-04 14:07:22 -05002121 wakeup_bdi = inode_io_list_move_locked(inode, wb,
Tejun Heod6c10f12015-05-22 17:13:45 -04002122 dirty_list);
2123
Tejun Heo07472592015-05-22 17:14:02 -04002124 spin_unlock(&wb->list_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002125 trace_writeback_dirty_inode_enqueue(inode);
Dave Chinnera66979a2011-03-22 22:23:41 +11002126
Tejun Heod6c10f12015-05-22 17:13:45 -04002127 /*
2128 * If this is the first dirty inode for this bdi,
2129 * we have to wake-up the corresponding bdi thread
2130 * to make sure background write-back happens
2131 * later.
2132 */
Tejun Heo07472592015-05-22 17:14:02 -04002133 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2134 wb_wakeup_delayed(wb);
Dave Chinnera66979a2011-03-22 22:23:41 +11002135 return;
Jens Axboe03ba3782009-09-09 09:08:54 +02002136 }
2137 }
Dave Chinner250df6e2011-03-22 22:23:36 +11002138out_unlock_inode:
2139 spin_unlock(&inode->i_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002140
Randy Dunlapdbce03b2015-11-09 14:58:00 -08002141#undef I_DIRTY_INODE
Jens Axboe03ba3782009-09-09 09:08:54 +02002142}
2143EXPORT_SYMBOL(__mark_inode_dirty);
2144
Dave Chinnere97fedb2015-03-04 13:40:00 -05002145/*
2146 * The @s_sync_lock is used to serialise concurrent sync operations
2147 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2148 * Concurrent callers will block on the s_sync_lock rather than doing contending
2149 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2150 * has been issued up to the time this function is enter is guaranteed to be
2151 * completed by the time we have gained the lock and waited for all IO that is
2152 * in progress regardless of the order callers are granted the lock.
2153 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002154static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002155{
Jens Axboe03ba3782009-09-09 09:08:54 +02002156 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002157
Jens Axboe03ba3782009-09-09 09:08:54 +02002158 /*
2159 * We need to be protected against the filesystem going from
2160 * r/o to r/w or vice versa.
2161 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002162 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002163
Dave Chinnere97fedb2015-03-04 13:40:00 -05002164 mutex_lock(&sb->s_sync_lock);
Dave Chinner74278da2015-03-04 12:37:22 -05002165 spin_lock(&sb->s_inode_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002166
2167 /*
2168 * Data integrity sync. Must wait for all pages under writeback,
2169 * because there may have been pages dirtied before our sync
2170 * call, but which had writeout started before we write it out.
2171 * In which case, the inode may not be on the dirty list, but
2172 * we still have to wait for that writeout.
2173 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002174 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Dave Chinner250df6e2011-03-22 22:23:36 +11002175 struct address_space *mapping = inode->i_mapping;
Jens Axboe03ba3782009-09-09 09:08:54 +02002176
Dave Chinner250df6e2011-03-22 22:23:36 +11002177 spin_lock(&inode->i_lock);
2178 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2179 (mapping->nrpages == 0)) {
2180 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002181 continue;
Dave Chinner250df6e2011-03-22 22:23:36 +11002182 }
Jens Axboe03ba3782009-09-09 09:08:54 +02002183 __iget(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11002184 spin_unlock(&inode->i_lock);
Dave Chinner74278da2015-03-04 12:37:22 -05002185 spin_unlock(&sb->s_inode_list_lock);
Dave Chinner55fa6092011-03-22 22:23:40 +11002186
Jens Axboe03ba3782009-09-09 09:08:54 +02002187 /*
Dave Chinner55fa6092011-03-22 22:23:40 +11002188 * We hold a reference to 'inode' so it couldn't have been
2189 * removed from s_inodes list while we dropped the
Dave Chinner74278da2015-03-04 12:37:22 -05002190 * s_inode_list_lock. We cannot iput the inode now as we can
Dave Chinner55fa6092011-03-22 22:23:40 +11002191 * be holding the last reference and we cannot iput it under
Dave Chinner74278da2015-03-04 12:37:22 -05002192 * s_inode_list_lock. So we keep the reference and iput it
Dave Chinner55fa6092011-03-22 22:23:40 +11002193 * later.
Jens Axboe03ba3782009-09-09 09:08:54 +02002194 */
2195 iput(old_inode);
2196 old_inode = inode;
2197
Junichi Nomuraaa750fd2015-11-05 18:47:23 -08002198 /*
2199 * We keep the error status of individual mapping so that
2200 * applications can catch the writeback error using fsync(2).
2201 * See filemap_fdatawait_keep_errors() for details.
2202 */
2203 filemap_fdatawait_keep_errors(mapping);
Jens Axboe03ba3782009-09-09 09:08:54 +02002204
2205 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08002206
Dave Chinner74278da2015-03-04 12:37:22 -05002207 spin_lock(&sb->s_inode_list_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002208 }
Dave Chinner74278da2015-03-04 12:37:22 -05002209 spin_unlock(&sb->s_inode_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002210 iput(old_inode);
Dave Chinnere97fedb2015-03-04 13:40:00 -05002211 mutex_unlock(&sb->s_sync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212}
2213
Tejun Heof30a7d02015-05-22 17:14:00 -04002214static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2215 enum wb_reason reason, bool skip_if_busy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216{
Tejun Heocc395d72015-05-22 17:13:58 -04002217 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002218 struct wb_writeback_work work = {
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002219 .sb = sb,
2220 .sync_mode = WB_SYNC_NONE,
2221 .tagged_writepages = 1,
2222 .done = &done,
2223 .nr_pages = nr,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002224 .reason = reason,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002225 };
Tejun Heoe7972912015-05-22 17:13:48 -04002226 struct backing_dev_info *bdi = sb->s_bdi;
Jens Axboe0e3c9a22010-06-01 11:08:43 +02002227
Tejun Heoe7972912015-05-22 17:13:48 -04002228 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002229 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002230 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Tejun Heof30a7d02015-05-22 17:14:00 -04002231
Tejun Heodb125362015-05-22 17:14:01 -04002232 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
Tejun Heocc395d72015-05-22 17:13:58 -04002233 wb_wait_for_completion(bdi, &done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234}
Tejun Heof30a7d02015-05-22 17:14:00 -04002235
2236/**
2237 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2238 * @sb: the superblock
2239 * @nr: the number of pages to write
2240 * @reason: reason why some writeback work initiated
2241 *
2242 * Start writeback on some inodes on this super_block. No guarantees are made
2243 * on how many (if any) will be written, and this function does not wait
2244 * for IO completion of submitted IO.
2245 */
2246void writeback_inodes_sb_nr(struct super_block *sb,
2247 unsigned long nr,
2248 enum wb_reason reason)
2249{
2250 __writeback_inodes_sb_nr(sb, nr, reason, false);
2251}
Chris Mason3259f8b2010-10-29 11:16:17 -04002252EXPORT_SYMBOL(writeback_inodes_sb_nr);
2253
2254/**
2255 * writeback_inodes_sb - writeback dirty inodes from given super_block
2256 * @sb: the superblock
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +08002257 * @reason: reason why some writeback work was initiated
Chris Mason3259f8b2010-10-29 11:16:17 -04002258 *
2259 * Start writeback on some inodes on this super_block. No guarantees are made
2260 * on how many (if any) will be written, and this function does not wait
2261 * for IO completion of submitted IO.
2262 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002263void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002264{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002265 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04002266}
Jens Axboed8a85592009-09-02 12:34:32 +02002267EXPORT_SYMBOL(writeback_inodes_sb);
2268
2269/**
Miao Xie10ee27a2013-01-10 13:47:57 +08002270 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
Chris Mason3259f8b2010-10-29 11:16:17 -04002271 * @sb: the superblock
2272 * @nr: the number of pages to write
Miao Xie10ee27a2013-01-10 13:47:57 +08002273 * @reason: the reason of writeback
Chris Mason3259f8b2010-10-29 11:16:17 -04002274 *
Miao Xie10ee27a2013-01-10 13:47:57 +08002275 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
Chris Mason3259f8b2010-10-29 11:16:17 -04002276 * Returns 1 if writeback was started, 0 if not.
2277 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002278bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2279 enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002280{
Miao Xie10ee27a2013-01-10 13:47:57 +08002281 if (!down_read_trylock(&sb->s_umount))
Tejun Heof30a7d02015-05-22 17:14:00 -04002282 return false;
Miao Xie10ee27a2013-01-10 13:47:57 +08002283
Tejun Heof30a7d02015-05-22 17:14:00 -04002284 __writeback_inodes_sb_nr(sb, nr, reason, true);
Miao Xie10ee27a2013-01-10 13:47:57 +08002285 up_read(&sb->s_umount);
Tejun Heof30a7d02015-05-22 17:14:00 -04002286 return true;
Chris Mason3259f8b2010-10-29 11:16:17 -04002287}
Miao Xie10ee27a2013-01-10 13:47:57 +08002288EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2289
2290/**
2291 * try_to_writeback_inodes_sb - try to start writeback if none underway
2292 * @sb: the superblock
2293 * @reason: reason why some writeback work was initiated
2294 *
2295 * Implement by try_to_writeback_inodes_sb_nr()
2296 * Returns 1 if writeback was started, 0 if not.
2297 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002298bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Miao Xie10ee27a2013-01-10 13:47:57 +08002299{
2300 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2301}
2302EXPORT_SYMBOL(try_to_writeback_inodes_sb);
Chris Mason3259f8b2010-10-29 11:16:17 -04002303
2304/**
Jens Axboed8a85592009-09-02 12:34:32 +02002305 * sync_inodes_sb - sync sb inode pages
Jan Kara0dc83bd2014-02-21 11:19:04 +01002306 * @sb: the superblock
Jens Axboed8a85592009-09-02 12:34:32 +02002307 *
2308 * This function writes and waits on any dirty inode belonging to this
Jan Kara0dc83bd2014-02-21 11:19:04 +01002309 * super_block.
Jens Axboed8a85592009-09-02 12:34:32 +02002310 */
Jan Kara0dc83bd2014-02-21 11:19:04 +01002311void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02002312{
Tejun Heocc395d72015-05-22 17:13:58 -04002313 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002314 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002315 .sb = sb,
2316 .sync_mode = WB_SYNC_ALL,
2317 .nr_pages = LONG_MAX,
2318 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002319 .done = &done,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002320 .reason = WB_REASON_SYNC,
Dave Chinner7747bd42013-07-02 22:38:35 +10002321 .for_sync = 1,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002322 };
Tejun Heoe7972912015-05-22 17:13:48 -04002323 struct backing_dev_info *bdi = sb->s_bdi;
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002324
Tejun Heo006a0972015-08-25 14:11:52 -04002325 /*
2326 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2327 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2328 * bdi_has_dirty() need to be written out too.
2329 */
2330 if (bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002331 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002332 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2333
Tejun Heodb125362015-05-22 17:14:01 -04002334 bdi_split_work_to_wbs(bdi, &work, false);
Tejun Heocc395d72015-05-22 17:13:58 -04002335 wb_wait_for_completion(bdi, &done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002336
Jens Axboeb6e51312009-09-16 15:13:54 +02002337 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02002338}
2339EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002342 * write_inode_now - write an inode to disk
2343 * @inode: inode to write to disk
2344 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002346 * This function commits an inode to disk immediately if it is dirty. This is
2347 * primarily needed by knfsd.
2348 *
2349 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351int write_inode_now(struct inode *inode, int sync)
2352{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 struct writeback_control wbc = {
2354 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08002355 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07002356 .range_start = 0,
2357 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 };
2359
2360 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08002361 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362
2363 might_sleep();
Tejun Heoaaf25592016-03-18 13:52:04 -04002364 return writeback_single_inode(inode, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365}
2366EXPORT_SYMBOL(write_inode_now);
2367
2368/**
2369 * sync_inode - write an inode and its pages to disk.
2370 * @inode: the inode to sync
2371 * @wbc: controls the writeback mode
2372 *
2373 * sync_inode() will write an inode and its pages to disk. It will also
2374 * correctly update the inode on its superblock's dirty inode lists and will
2375 * update inode->i_state.
2376 *
2377 * The caller must have a ref on the inode.
2378 */
2379int sync_inode(struct inode *inode, struct writeback_control *wbc)
2380{
Tejun Heoaaf25592016-03-18 13:52:04 -04002381 return writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382}
2383EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc37650162010-10-06 10:48:20 +02002384
2385/**
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002386 * sync_inode_metadata - write an inode to disk
Christoph Hellwigc37650162010-10-06 10:48:20 +02002387 * @inode: the inode to sync
2388 * @wait: wait for I/O to complete.
2389 *
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002390 * Write an inode to disk and adjust its dirty state after completion.
Christoph Hellwigc37650162010-10-06 10:48:20 +02002391 *
2392 * Note: only writes the actual inode, no associated data or other metadata.
2393 */
2394int sync_inode_metadata(struct inode *inode, int wait)
2395{
2396 struct writeback_control wbc = {
2397 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2398 .nr_to_write = 0, /* metadata-only */
2399 };
2400
2401 return sync_inode(inode, &wbc);
2402}
2403EXPORT_SYMBOL(sync_inode_metadata);