blob: 206a68b1db1ab1b1f5a6ed6a611723f957e1e58b [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/fs-writeback.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * Contains all the functions related to writing back and waiting
7 * upon dirty inodes against superblocks, and writing back dirty
8 * pages against inodes. ie: data writeback. Writeout of the
9 * inode itself is not handled here.
10 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Split out of fs/inode.c
13 * Additions for address_space-based writeback
14 */
15
16#include <linux/kernel.h>
Paul Gortmaker630d9c42011-11-16 23:57:37 -050017#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090019#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#include <linux/sched.h>
21#include <linux/fs.h>
22#include <linux/mm.h>
Wu Fengguangbc31b862012-01-07 20:41:55 -060023#include <linux/pagemap.h>
Jens Axboe03ba3782009-09-09 09:08:54 +020024#include <linux/kthread.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/backing-dev.h>
Dave Chinner455b2862010-07-07 13:24:06 +100028#include <linux/tracepoint.h>
Al Viro719ea2f2013-09-29 11:24:49 -040029#include <linux/device.h>
Tejun Heo21c63212015-05-28 14:50:49 -040030#include <linux/memcontrol.h>
David Howells07f3f052006-09-30 20:52:18 +020031#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
Jens Axboed0bceac2009-05-18 08:20:32 +020033/*
Wu Fengguangbc31b862012-01-07 20:41:55 -060034 * 4MB minimal write chunk size
35 */
36#define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10))
37
Tejun Heocc395d72015-05-22 17:13:58 -040038struct wb_completion {
39 atomic_t cnt;
40};
41
Wu Fengguangbc31b862012-01-07 20:41:55 -060042/*
Jens Axboec4a77a62009-09-16 15:18:25 +020043 * Passed into wb_writeback(), essentially a subset of writeback_control
44 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +020045struct wb_writeback_work {
Jens Axboec4a77a62009-09-16 15:18:25 +020046 long nr_pages;
47 struct super_block *sb;
Jan Kara0dc83bd2014-02-21 11:19:04 +010048 unsigned long *older_than_this;
Jens Axboec4a77a62009-09-16 15:18:25 +020049 enum writeback_sync_modes sync_mode;
Wu Fengguang6e6938b2010-06-06 10:38:15 -060050 unsigned int tagged_writepages:1;
H Hartley Sweeten52957fe2010-04-01 20:36:30 -050051 unsigned int for_kupdate:1;
52 unsigned int range_cyclic:1;
53 unsigned int for_background:1;
Dave Chinner7747bd42013-07-02 22:38:35 +100054 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
Tejun Heoac7b19a2015-05-22 17:13:57 -040055 unsigned int auto_free:1; /* free on completion */
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 Heo21c63212015-05-28 14:50:49 -0400226void __inode_attach_wb(struct inode *inode, struct page *page)
227{
228 struct backing_dev_info *bdi = inode_to_bdi(inode);
229 struct bdi_writeback *wb = NULL;
230
231 if (inode_cgwb_enabled(inode)) {
232 struct cgroup_subsys_state *memcg_css;
233
234 if (page) {
235 memcg_css = mem_cgroup_css_from_page(page);
236 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
237 } else {
238 /* must pin memcg_css, see wb_get_create() */
239 memcg_css = task_get_css(current, memory_cgrp_id);
240 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
241 css_put(memcg_css);
242 }
243 }
244
245 if (!wb)
246 wb = &bdi->wb;
247
248 /*
249 * There may be multiple instances of this function racing to
250 * update the same inode. Use cmpxchg() to tell the winner.
251 */
252 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
253 wb_put(wb);
254}
255
Tejun Heo703c2702015-05-22 17:13:44 -0400256/**
Tejun Heo87e1d782015-05-28 14:50:52 -0400257 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
258 * @inode: inode of interest with i_lock held
259 *
260 * Returns @inode's wb with its list_lock held. @inode->i_lock must be
261 * held on entry and is released on return. The returned wb is guaranteed
262 * to stay @inode's associated wb until its list_lock is released.
263 */
264static struct bdi_writeback *
265locked_inode_to_wb_and_lock_list(struct inode *inode)
266 __releases(&inode->i_lock)
267 __acquires(&wb->list_lock)
268{
269 while (true) {
270 struct bdi_writeback *wb = inode_to_wb(inode);
271
272 /*
273 * inode_to_wb() association is protected by both
274 * @inode->i_lock and @wb->list_lock but list_lock nests
275 * outside i_lock. Drop i_lock and verify that the
276 * association hasn't changed after acquiring list_lock.
277 */
278 wb_get(wb);
279 spin_unlock(&inode->i_lock);
280 spin_lock(&wb->list_lock);
281 wb_put(wb); /* not gonna deref it anymore */
282
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400283 /* i_wb may have changed inbetween, can't use inode_to_wb() */
284 if (likely(wb == inode->i_wb))
Tejun Heo87e1d782015-05-28 14:50:52 -0400285 return wb; /* @inode already has ref */
286
287 spin_unlock(&wb->list_lock);
288 cpu_relax();
289 spin_lock(&inode->i_lock);
290 }
291}
292
293/**
294 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
295 * @inode: inode of interest
296 *
297 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
298 * on entry.
299 */
300static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
301 __acquires(&wb->list_lock)
302{
303 spin_lock(&inode->i_lock);
304 return locked_inode_to_wb_and_lock_list(inode);
305}
306
Tejun Heo682aa8e2015-05-28 14:50:53 -0400307struct inode_switch_wbs_context {
308 struct inode *inode;
309 struct bdi_writeback *new_wb;
310
311 struct rcu_head rcu_head;
312 struct work_struct work;
313};
314
315static void inode_switch_wbs_work_fn(struct work_struct *work)
316{
317 struct inode_switch_wbs_context *isw =
318 container_of(work, struct inode_switch_wbs_context, work);
319 struct inode *inode = isw->inode;
Tejun Heod10c8092015-05-28 14:50:56 -0400320 struct address_space *mapping = inode->i_mapping;
321 struct bdi_writeback *old_wb = inode->i_wb;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400322 struct bdi_writeback *new_wb = isw->new_wb;
Tejun Heod10c8092015-05-28 14:50:56 -0400323 struct radix_tree_iter iter;
324 bool switched = false;
325 void **slot;
Tejun Heo682aa8e2015-05-28 14:50:53 -0400326
327 /*
328 * By the time control reaches here, RCU grace period has passed
329 * since I_WB_SWITCH assertion and all wb stat update transactions
330 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
331 * synchronizing against mapping->tree_lock.
Tejun Heod10c8092015-05-28 14:50:56 -0400332 *
333 * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
334 * gives us exclusion against all wb related operations on @inode
335 * including IO list manipulations and stat updates.
Tejun Heo682aa8e2015-05-28 14:50:53 -0400336 */
Tejun Heod10c8092015-05-28 14:50:56 -0400337 if (old_wb < new_wb) {
338 spin_lock(&old_wb->list_lock);
339 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
340 } else {
341 spin_lock(&new_wb->list_lock);
342 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
343 }
Tejun Heo682aa8e2015-05-28 14:50:53 -0400344 spin_lock(&inode->i_lock);
Tejun Heod10c8092015-05-28 14:50:56 -0400345 spin_lock_irq(&mapping->tree_lock);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400346
Tejun Heod10c8092015-05-28 14:50:56 -0400347 /*
348 * Once I_FREEING is visible under i_lock, the eviction path owns
Dave Chinnerc7f54082015-03-04 14:07:22 -0500349 * the inode and we shouldn't modify ->i_io_list.
Tejun Heod10c8092015-05-28 14:50:56 -0400350 */
351 if (unlikely(inode->i_state & I_FREEING))
352 goto skip_switch;
353
354 /*
355 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
356 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
357 * pages actually under underwriteback.
358 */
359 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
360 PAGECACHE_TAG_DIRTY) {
361 struct page *page = radix_tree_deref_slot_protected(slot,
362 &mapping->tree_lock);
363 if (likely(page) && PageDirty(page)) {
364 __dec_wb_stat(old_wb, WB_RECLAIMABLE);
365 __inc_wb_stat(new_wb, WB_RECLAIMABLE);
366 }
367 }
368
369 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
370 PAGECACHE_TAG_WRITEBACK) {
371 struct page *page = radix_tree_deref_slot_protected(slot,
372 &mapping->tree_lock);
373 if (likely(page)) {
374 WARN_ON_ONCE(!PageWriteback(page));
375 __dec_wb_stat(old_wb, WB_WRITEBACK);
376 __inc_wb_stat(new_wb, WB_WRITEBACK);
377 }
378 }
379
380 wb_get(new_wb);
381
382 /*
383 * Transfer to @new_wb's IO list if necessary. The specific list
384 * @inode was on is ignored and the inode is put on ->b_dirty which
385 * is always correct including from ->b_dirty_time. The transfer
386 * preserves @inode->dirtied_when ordering.
387 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500388 if (!list_empty(&inode->i_io_list)) {
Tejun Heod10c8092015-05-28 14:50:56 -0400389 struct inode *pos;
390
Dave Chinnerc7f54082015-03-04 14:07:22 -0500391 inode_io_list_del_locked(inode, old_wb);
Tejun Heod10c8092015-05-28 14:50:56 -0400392 inode->i_wb = new_wb;
Dave Chinnerc7f54082015-03-04 14:07:22 -0500393 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
Tejun Heod10c8092015-05-28 14:50:56 -0400394 if (time_after_eq(inode->dirtied_when,
395 pos->dirtied_when))
396 break;
Dave Chinnerc7f54082015-03-04 14:07:22 -0500397 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
Tejun Heod10c8092015-05-28 14:50:56 -0400398 } else {
399 inode->i_wb = new_wb;
400 }
401
402 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400403 inode->i_wb_frn_winner = 0;
404 inode->i_wb_frn_avg_time = 0;
405 inode->i_wb_frn_history = 0;
Tejun Heod10c8092015-05-28 14:50:56 -0400406 switched = true;
407skip_switch:
Tejun Heo682aa8e2015-05-28 14:50:53 -0400408 /*
409 * Paired with load_acquire in unlocked_inode_to_wb_begin() and
410 * ensures that the new wb is visible if they see !I_WB_SWITCH.
411 */
412 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
413
Tejun Heod10c8092015-05-28 14:50:56 -0400414 spin_unlock_irq(&mapping->tree_lock);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400415 spin_unlock(&inode->i_lock);
Tejun Heod10c8092015-05-28 14:50:56 -0400416 spin_unlock(&new_wb->list_lock);
417 spin_unlock(&old_wb->list_lock);
418
419 if (switched) {
420 wb_wakeup(new_wb);
421 wb_put(old_wb);
422 }
423 wb_put(new_wb);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400424
425 iput(inode);
Tejun Heo682aa8e2015-05-28 14:50:53 -0400426 kfree(isw);
427}
428
429static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
430{
431 struct inode_switch_wbs_context *isw = container_of(rcu_head,
432 struct inode_switch_wbs_context, rcu_head);
433
434 /* needs to grab bh-unsafe locks, bounce to work item */
435 INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
436 schedule_work(&isw->work);
437}
438
439/**
440 * inode_switch_wbs - change the wb association of an inode
441 * @inode: target inode
442 * @new_wb_id: ID of the new wb
443 *
444 * Switch @inode's wb association to the wb identified by @new_wb_id. The
445 * switching is performed asynchronously and may fail silently.
446 */
447static void inode_switch_wbs(struct inode *inode, int new_wb_id)
448{
449 struct backing_dev_info *bdi = inode_to_bdi(inode);
450 struct cgroup_subsys_state *memcg_css;
451 struct inode_switch_wbs_context *isw;
452
453 /* noop if seems to be already in progress */
454 if (inode->i_state & I_WB_SWITCH)
455 return;
456
457 isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
458 if (!isw)
459 return;
460
461 /* find and pin the new wb */
462 rcu_read_lock();
463 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
464 if (memcg_css)
465 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
466 rcu_read_unlock();
467 if (!isw->new_wb)
468 goto out_free;
469
470 /* while holding I_WB_SWITCH, no one else can update the association */
471 spin_lock(&inode->i_lock);
472 if (inode->i_state & (I_WB_SWITCH | I_FREEING) ||
473 inode_to_wb(inode) == isw->new_wb) {
474 spin_unlock(&inode->i_lock);
475 goto out_free;
476 }
477 inode->i_state |= I_WB_SWITCH;
478 spin_unlock(&inode->i_lock);
479
480 ihold(inode);
481 isw->inode = inode;
482
483 /*
484 * In addition to synchronizing among switchers, I_WB_SWITCH tells
485 * the RCU protected stat update paths to grab the mapping's
486 * tree_lock so that stat transfer can synchronize against them.
487 * Let's continue after I_WB_SWITCH is guaranteed to be visible.
488 */
489 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
490 return;
491
492out_free:
493 if (isw->new_wb)
494 wb_put(isw->new_wb);
495 kfree(isw);
496}
497
Tejun Heo87e1d782015-05-28 14:50:52 -0400498/**
Tejun Heob16b1de2015-06-02 08:39:48 -0600499 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
500 * @wbc: writeback_control of interest
501 * @inode: target inode
502 *
503 * @inode is locked and about to be written back under the control of @wbc.
504 * Record @inode's writeback context into @wbc and unlock the i_lock. On
505 * writeback completion, wbc_detach_inode() should be called. This is used
506 * to track the cgroup writeback context.
507 */
508void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
509 struct inode *inode)
510{
Tejun Heodd73e4b2015-06-16 18:48:30 -0400511 if (!inode_cgwb_enabled(inode)) {
512 spin_unlock(&inode->i_lock);
513 return;
514 }
515
Tejun Heob16b1de2015-06-02 08:39:48 -0600516 wbc->wb = inode_to_wb(inode);
Tejun Heo2a814902015-05-28 14:50:51 -0400517 wbc->inode = inode;
518
519 wbc->wb_id = wbc->wb->memcg_css->id;
520 wbc->wb_lcand_id = inode->i_wb_frn_winner;
521 wbc->wb_tcand_id = 0;
522 wbc->wb_bytes = 0;
523 wbc->wb_lcand_bytes = 0;
524 wbc->wb_tcand_bytes = 0;
525
Tejun Heob16b1de2015-06-02 08:39:48 -0600526 wb_get(wbc->wb);
527 spin_unlock(&inode->i_lock);
Tejun Heoe8a7abf2015-05-28 14:50:57 -0400528
529 /*
530 * A dying wb indicates that the memcg-blkcg mapping has changed
531 * and a new wb is already serving the memcg. Switch immediately.
532 */
533 if (unlikely(wb_dying(wbc->wb)))
534 inode_switch_wbs(inode, wbc->wb_id);
Tejun Heob16b1de2015-06-02 08:39:48 -0600535}
536
537/**
Tejun Heo2a814902015-05-28 14:50:51 -0400538 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
539 * @wbc: writeback_control of the just finished writeback
Tejun Heob16b1de2015-06-02 08:39:48 -0600540 *
541 * To be called after a writeback attempt of an inode finishes and undoes
542 * wbc_attach_and_unlock_inode(). Can be called under any context.
Tejun Heo2a814902015-05-28 14:50:51 -0400543 *
544 * As concurrent write sharing of an inode is expected to be very rare and
545 * memcg only tracks page ownership on first-use basis severely confining
546 * the usefulness of such sharing, cgroup writeback tracks ownership
547 * per-inode. While the support for concurrent write sharing of an inode
548 * is deemed unnecessary, an inode being written to by different cgroups at
549 * different points in time is a lot more common, and, more importantly,
550 * charging only by first-use can too readily lead to grossly incorrect
551 * behaviors (single foreign page can lead to gigabytes of writeback to be
552 * incorrectly attributed).
553 *
554 * To resolve this issue, cgroup writeback detects the majority dirtier of
555 * an inode and transfers the ownership to it. To avoid unnnecessary
556 * oscillation, the detection mechanism keeps track of history and gives
557 * out the switch verdict only if the foreign usage pattern is stable over
558 * a certain amount of time and/or writeback attempts.
559 *
560 * On each writeback attempt, @wbc tries to detect the majority writer
561 * using Boyer-Moore majority vote algorithm. In addition to the byte
562 * count from the majority voting, it also counts the bytes written for the
563 * current wb and the last round's winner wb (max of last round's current
564 * wb, the winner from two rounds ago, and the last round's majority
565 * candidate). Keeping track of the historical winner helps the algorithm
566 * to semi-reliably detect the most active writer even when it's not the
567 * absolute majority.
568 *
569 * Once the winner of the round is determined, whether the winner is
570 * foreign or not and how much IO time the round consumed is recorded in
571 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
572 * over a certain threshold, the switch verdict is given.
Tejun Heob16b1de2015-06-02 08:39:48 -0600573 */
574void wbc_detach_inode(struct writeback_control *wbc)
575{
Tejun Heo2a814902015-05-28 14:50:51 -0400576 struct bdi_writeback *wb = wbc->wb;
577 struct inode *inode = wbc->inode;
Tejun Heodd73e4b2015-06-16 18:48:30 -0400578 unsigned long avg_time, max_bytes, max_time;
579 u16 history;
Tejun Heo2a814902015-05-28 14:50:51 -0400580 int max_id;
581
Tejun Heodd73e4b2015-06-16 18:48:30 -0400582 if (!wb)
583 return;
584
585 history = inode->i_wb_frn_history;
586 avg_time = inode->i_wb_frn_avg_time;
587
Tejun Heo2a814902015-05-28 14:50:51 -0400588 /* pick the winner of this round */
589 if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
590 wbc->wb_bytes >= wbc->wb_tcand_bytes) {
591 max_id = wbc->wb_id;
592 max_bytes = wbc->wb_bytes;
593 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
594 max_id = wbc->wb_lcand_id;
595 max_bytes = wbc->wb_lcand_bytes;
596 } else {
597 max_id = wbc->wb_tcand_id;
598 max_bytes = wbc->wb_tcand_bytes;
599 }
600
601 /*
602 * Calculate the amount of IO time the winner consumed and fold it
603 * into the running average kept per inode. If the consumed IO
604 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
605 * deciding whether to switch or not. This is to prevent one-off
606 * small dirtiers from skewing the verdict.
607 */
608 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
609 wb->avg_write_bandwidth);
610 if (avg_time)
611 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
612 (avg_time >> WB_FRN_TIME_AVG_SHIFT);
613 else
614 avg_time = max_time; /* immediate catch up on first run */
615
616 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
617 int slots;
618
619 /*
620 * The switch verdict is reached if foreign wb's consume
621 * more than a certain proportion of IO time in a
622 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
623 * history mask where each bit represents one sixteenth of
624 * the period. Determine the number of slots to shift into
625 * history from @max_time.
626 */
627 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
628 (unsigned long)WB_FRN_HIST_MAX_SLOTS);
629 history <<= slots;
630 if (wbc->wb_id != max_id)
631 history |= (1U << slots) - 1;
632
633 /*
634 * Switch if the current wb isn't the consistent winner.
635 * If there are multiple closely competing dirtiers, the
636 * inode may switch across them repeatedly over time, which
637 * is okay. The main goal is avoiding keeping an inode on
638 * the wrong wb for an extended period of time.
639 */
Tejun Heo682aa8e2015-05-28 14:50:53 -0400640 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
641 inode_switch_wbs(inode, max_id);
Tejun Heo2a814902015-05-28 14:50:51 -0400642 }
643
644 /*
645 * Multiple instances of this function may race to update the
646 * following fields but we don't mind occassional inaccuracies.
647 */
648 inode->i_wb_frn_winner = max_id;
649 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
650 inode->i_wb_frn_history = history;
651
Tejun Heob16b1de2015-06-02 08:39:48 -0600652 wb_put(wbc->wb);
653 wbc->wb = NULL;
654}
655
656/**
Tejun Heo2a814902015-05-28 14:50:51 -0400657 * wbc_account_io - account IO issued during writeback
658 * @wbc: writeback_control of the writeback in progress
659 * @page: page being written out
660 * @bytes: number of bytes being written out
661 *
662 * @bytes from @page are about to written out during the writeback
663 * controlled by @wbc. Keep the book for foreign inode detection. See
664 * wbc_detach_inode().
665 */
666void wbc_account_io(struct writeback_control *wbc, struct page *page,
667 size_t bytes)
668{
669 int id;
670
671 /*
672 * pageout() path doesn't attach @wbc to the inode being written
673 * out. This is intentional as we don't want the function to block
674 * behind a slow cgroup. Ultimately, we want pageout() to kick off
675 * regular writeback instead of writing things out itself.
676 */
677 if (!wbc->wb)
678 return;
679
680 rcu_read_lock();
681 id = mem_cgroup_css_from_page(page)->id;
682 rcu_read_unlock();
683
684 if (id == wbc->wb_id) {
685 wbc->wb_bytes += bytes;
686 return;
687 }
688
689 if (id == wbc->wb_lcand_id)
690 wbc->wb_lcand_bytes += bytes;
691
692 /* Boyer-Moore majority vote algorithm */
693 if (!wbc->wb_tcand_bytes)
694 wbc->wb_tcand_id = id;
695 if (id == wbc->wb_tcand_id)
696 wbc->wb_tcand_bytes += bytes;
697 else
698 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
699}
Tejun Heo5aa2a962015-07-23 14:27:09 -0400700EXPORT_SYMBOL_GPL(wbc_account_io);
Tejun Heo2a814902015-05-28 14:50:51 -0400701
702/**
Tejun Heo703c2702015-05-22 17:13:44 -0400703 * inode_congested - test whether an inode is congested
Tejun Heo60292bc2015-08-18 14:54:54 -0700704 * @inode: inode to test for congestion (may be NULL)
Tejun Heo703c2702015-05-22 17:13:44 -0400705 * @cong_bits: mask of WB_[a]sync_congested bits to test
706 *
707 * Tests whether @inode is congested. @cong_bits is the mask of congestion
708 * bits to test and the return value is the mask of set bits.
709 *
710 * If cgroup writeback is enabled for @inode, the congestion state is
711 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
712 * associated with @inode is congested; otherwise, the root wb's congestion
713 * state is used.
Tejun Heo60292bc2015-08-18 14:54:54 -0700714 *
715 * @inode is allowed to be NULL as this function is often called on
716 * mapping->host which is NULL for the swapper space.
Tejun Heo703c2702015-05-22 17:13:44 -0400717 */
718int inode_congested(struct inode *inode, int cong_bits)
719{
Tejun Heo5cb8b822015-05-28 14:50:54 -0400720 /*
721 * Once set, ->i_wb never becomes NULL while the inode is alive.
722 * Start transaction iff ->i_wb is visible.
723 */
Tejun Heoaaa2cac2015-05-28 14:50:55 -0400724 if (inode && inode_to_wb_is_valid(inode)) {
Tejun Heo5cb8b822015-05-28 14:50:54 -0400725 struct bdi_writeback *wb;
726 bool locked, congested;
727
728 wb = unlocked_inode_to_wb_begin(inode, &locked);
729 congested = wb_congested(wb, cong_bits);
730 unlocked_inode_to_wb_end(inode, locked);
731 return congested;
Tejun Heo703c2702015-05-22 17:13:44 -0400732 }
733
734 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
735}
736EXPORT_SYMBOL_GPL(inode_congested);
737
Tejun Heof2b65122015-05-22 17:13:55 -0400738/**
739 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
740 * @wb: target bdi_writeback to split @nr_pages to
741 * @nr_pages: number of pages to write for the whole bdi
742 *
743 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
744 * relation to the total write bandwidth of all wb's w/ dirty inodes on
745 * @wb->bdi.
746 */
747static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
748{
749 unsigned long this_bw = wb->avg_write_bandwidth;
750 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
751
752 if (nr_pages == LONG_MAX)
753 return LONG_MAX;
754
755 /*
756 * This may be called on clean wb's and proportional distribution
757 * may not make sense, just use the original @nr_pages in those
758 * cases. In general, we wanna err on the side of writing more.
759 */
760 if (!tot_bw || this_bw >= tot_bw)
761 return nr_pages;
762 else
763 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
764}
765
Tejun Heodb125362015-05-22 17:14:01 -0400766/**
Tejun Heodb125362015-05-22 17:14:01 -0400767 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
768 * @bdi: target backing_dev_info
769 * @base_work: wb_writeback_work to issue
770 * @skip_if_busy: skip wb's which already have writeback in progress
771 *
772 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
773 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
774 * distributed to the busy wbs according to each wb's proportion in the
775 * total active write bandwidth of @bdi.
776 */
777static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
778 struct wb_writeback_work *base_work,
779 bool skip_if_busy)
780{
Tejun Heob8175252015-10-02 14:47:05 -0400781 struct bdi_writeback *last_wb = NULL;
Tejun Heob33e18f2015-10-27 14:19:39 +0900782 struct bdi_writeback *wb = list_entry(&bdi->wb_list,
783 struct bdi_writeback, bdi_node);
Tejun Heodb125362015-05-22 17:14:01 -0400784
785 might_sleep();
Tejun Heodb125362015-05-22 17:14:01 -0400786restart:
787 rcu_read_lock();
Tejun Heob8175252015-10-02 14:47:05 -0400788 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
Tejun Heo8a1270c2015-08-18 14:54:53 -0700789 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
790 struct wb_writeback_work fallback_work;
791 struct wb_writeback_work *work;
792 long nr_pages;
793
Tejun Heob8175252015-10-02 14:47:05 -0400794 if (last_wb) {
795 wb_put(last_wb);
796 last_wb = NULL;
797 }
798
Tejun Heo006a0972015-08-25 14:11:52 -0400799 /* SYNC_ALL writes out I_DIRTY_TIME too */
800 if (!wb_has_dirty_io(wb) &&
801 (base_work->sync_mode == WB_SYNC_NONE ||
802 list_empty(&wb->b_dirty_time)))
803 continue;
804 if (skip_if_busy && writeback_in_progress(wb))
Tejun Heodb125362015-05-22 17:14:01 -0400805 continue;
806
Tejun Heo8a1270c2015-08-18 14:54:53 -0700807 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
808
809 work = kmalloc(sizeof(*work), GFP_ATOMIC);
810 if (work) {
811 *work = *base_work;
812 work->nr_pages = nr_pages;
813 work->auto_free = 1;
814 wb_queue_work(wb, work);
815 continue;
Tejun Heodb125362015-05-22 17:14:01 -0400816 }
Tejun Heo8a1270c2015-08-18 14:54:53 -0700817
818 /* alloc failed, execute synchronously using on-stack fallback */
819 work = &fallback_work;
820 *work = *base_work;
821 work->nr_pages = nr_pages;
822 work->auto_free = 0;
823 work->done = &fallback_work_done;
824
825 wb_queue_work(wb, work);
826
Tejun Heob8175252015-10-02 14:47:05 -0400827 /*
828 * Pin @wb so that it stays on @bdi->wb_list. This allows
829 * continuing iteration from @wb after dropping and
830 * regrabbing rcu read lock.
831 */
832 wb_get(wb);
833 last_wb = wb;
834
Tejun Heo8a1270c2015-08-18 14:54:53 -0700835 rcu_read_unlock();
836 wb_wait_for_completion(bdi, &fallback_work_done);
837 goto restart;
Tejun Heodb125362015-05-22 17:14:01 -0400838 }
839 rcu_read_unlock();
Tejun Heob8175252015-10-02 14:47:05 -0400840
841 if (last_wb)
842 wb_put(last_wb);
Tejun Heodb125362015-05-22 17:14:01 -0400843}
844
Tejun Heof2b65122015-05-22 17:13:55 -0400845#else /* CONFIG_CGROUP_WRITEBACK */
846
Tejun Heo87e1d782015-05-28 14:50:52 -0400847static struct bdi_writeback *
848locked_inode_to_wb_and_lock_list(struct inode *inode)
849 __releases(&inode->i_lock)
850 __acquires(&wb->list_lock)
851{
852 struct bdi_writeback *wb = inode_to_wb(inode);
853
854 spin_unlock(&inode->i_lock);
855 spin_lock(&wb->list_lock);
856 return wb;
857}
858
859static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
860 __acquires(&wb->list_lock)
861{
862 struct bdi_writeback *wb = inode_to_wb(inode);
863
864 spin_lock(&wb->list_lock);
865 return wb;
866}
867
Tejun Heof2b65122015-05-22 17:13:55 -0400868static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
869{
870 return nr_pages;
871}
872
Tejun Heodb125362015-05-22 17:14:01 -0400873static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
874 struct wb_writeback_work *base_work,
875 bool skip_if_busy)
876{
877 might_sleep();
878
Tejun Heo006a0972015-08-25 14:11:52 -0400879 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
Tejun Heodb125362015-05-22 17:14:01 -0400880 base_work->auto_free = 0;
Tejun Heodb125362015-05-22 17:14:01 -0400881 wb_queue_work(&bdi->wb, base_work);
882 }
883}
884
Tejun Heo703c2702015-05-22 17:13:44 -0400885#endif /* CONFIG_CGROUP_WRITEBACK */
886
Tejun Heoc00ddad2015-05-22 17:13:51 -0400887void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
888 bool range_cyclic, enum wb_reason reason)
Jens Axboeb6e51312009-09-16 15:13:54 +0200889{
Tejun Heoc00ddad2015-05-22 17:13:51 -0400890 struct wb_writeback_work *work;
891
892 if (!wb_has_dirty_io(wb))
893 return;
894
895 /*
896 * This is WB_SYNC_NONE writeback, so if allocation fails just
897 * wakeup the thread for old dirty data writeback
898 */
899 work = kzalloc(sizeof(*work), GFP_ATOMIC);
900 if (!work) {
Tejun Heo5634cc22015-08-18 14:54:56 -0700901 trace_writeback_nowork(wb);
Tejun Heoc00ddad2015-05-22 17:13:51 -0400902 wb_wakeup(wb);
903 return;
904 }
905
906 work->sync_mode = WB_SYNC_NONE;
907 work->nr_pages = nr_pages;
908 work->range_cyclic = range_cyclic;
909 work->reason = reason;
Tejun Heoac7b19a2015-05-22 17:13:57 -0400910 work->auto_free = 1;
Tejun Heoc00ddad2015-05-22 17:13:51 -0400911
912 wb_queue_work(wb, work);
Christoph Hellwigc5444192010-06-08 18:15:15 +0200913}
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800914
Christoph Hellwigc5444192010-06-08 18:15:15 +0200915/**
Tejun Heo9ecf48662015-05-22 17:13:54 -0400916 * wb_start_background_writeback - start background writeback
917 * @wb: bdi_writback to write from
Christoph Hellwigc5444192010-06-08 18:15:15 +0200918 *
919 * Description:
Jan Kara65850272011-01-13 15:45:44 -0800920 * This makes sure WB_SYNC_NONE background writeback happens. When
Tejun Heo9ecf48662015-05-22 17:13:54 -0400921 * this function returns, it is only guaranteed that for given wb
Jan Kara65850272011-01-13 15:45:44 -0800922 * some IO is happening if we are over background dirty threshold.
923 * Caller need not hold sb s_umount semaphore.
Christoph Hellwigc5444192010-06-08 18:15:15 +0200924 */
Tejun Heo9ecf48662015-05-22 17:13:54 -0400925void wb_start_background_writeback(struct bdi_writeback *wb)
Christoph Hellwigc5444192010-06-08 18:15:15 +0200926{
Jan Kara65850272011-01-13 15:45:44 -0800927 /*
928 * We just wake up the flusher thread. It will perform background
929 * writeback as soon as there is no other work to do.
930 */
Tejun Heo5634cc22015-08-18 14:54:56 -0700931 trace_writeback_wake_background(wb);
Tejun Heo9ecf48662015-05-22 17:13:54 -0400932 wb_wakeup(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933}
934
935/*
Dave Chinnera66979a2011-03-22 22:23:41 +1100936 * Remove the inode from the writeback list it is on.
937 */
Dave Chinnerc7f54082015-03-04 14:07:22 -0500938void inode_io_list_del(struct inode *inode)
Dave Chinnera66979a2011-03-22 22:23:41 +1100939{
Tejun Heo87e1d782015-05-28 14:50:52 -0400940 struct bdi_writeback *wb;
Dave Chinnera66979a2011-03-22 22:23:41 +1100941
Tejun Heo87e1d782015-05-28 14:50:52 -0400942 wb = inode_to_wb_and_lock_list(inode);
Dave Chinnerc7f54082015-03-04 14:07:22 -0500943 inode_io_list_del_locked(inode, wb);
Tejun Heo52ebea72015-05-22 17:13:37 -0400944 spin_unlock(&wb->list_lock);
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600945}
Dave Chinnera66979a2011-03-22 22:23:41 +1100946
947/*
Andrew Morton6610a0b2007-10-16 23:30:32 -0700948 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
949 * furthest end of its superblock's dirty-inode list.
950 *
951 * Before stamping the inode's ->dirtied_when, we check to see whether it is
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200952 * already the most-recently-dirtied inode on the b_dirty list. If that is
Andrew Morton6610a0b2007-10-16 23:30:32 -0700953 * the case then the inode must have been redirtied while it was being written
954 * out and we don't reset its dirtied_when.
955 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600956static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
Andrew Morton6610a0b2007-10-16 23:30:32 -0700957{
Jens Axboe03ba3782009-09-09 09:08:54 +0200958 if (!list_empty(&wb->b_dirty)) {
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200959 struct inode *tail;
Andrew Morton6610a0b2007-10-16 23:30:32 -0700960
Nick Piggin7ccf19a2010-10-21 11:49:30 +1100961 tail = wb_inode(wb->b_dirty.next);
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200962 if (time_before(inode->dirtied_when, tail->dirtied_when))
Andrew Morton6610a0b2007-10-16 23:30:32 -0700963 inode->dirtied_when = jiffies;
964 }
Dave Chinnerc7f54082015-03-04 14:07:22 -0500965 inode_io_list_move_locked(inode, wb, &wb->b_dirty);
Andrew Morton6610a0b2007-10-16 23:30:32 -0700966}
967
968/*
Jens Axboe66f3b8e2009-09-02 09:19:46 +0200969 * requeue inode for re-scanning after bdi->b_io list is exhausted.
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700970 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -0600971static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700972{
Dave Chinnerc7f54082015-03-04 14:07:22 -0500973 inode_io_list_move_locked(inode, wb, &wb->b_more_io);
Andrew Mortonc986d1e2007-10-16 23:30:34 -0700974}
975
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700976static void inode_sync_complete(struct inode *inode)
977{
Jan Kara365b94ae2012-05-03 14:47:55 +0200978 inode->i_state &= ~I_SYNC;
Jan Kara4eff96d2012-11-26 16:29:51 -0800979 /* If inode is clean an unused, put it into LRU now... */
980 inode_add_lru(inode);
Jan Kara365b94ae2012-05-03 14:47:55 +0200981 /* Waiters must see I_SYNC cleared before being woken up */
Joern Engel1c0eeaf2007-10-16 23:30:44 -0700982 smp_mb();
983 wake_up_bit(&inode->i_state, __I_SYNC);
984}
985
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700986static bool inode_dirtied_after(struct inode *inode, unsigned long t)
987{
988 bool ret = time_after(inode->dirtied_when, t);
989#ifndef CONFIG_64BIT
990 /*
991 * For inodes being constantly redirtied, dirtied_when can get stuck.
992 * It _appears_ to be in the future, but is actually in distant past.
993 * This test is necessary to prevent such wrapped-around relative times
Jens Axboe5b0830c2009-09-23 19:37:09 +0200994 * from permanently stopping the whole bdi writeback.
Jeff Laytond2caa3c52009-04-02 16:56:37 -0700995 */
996 ret = ret && time_before_eq(inode->dirtied_when, jiffies);
997#endif
998 return ret;
999}
1000
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001001#define EXPIRE_DIRTY_ATIME 0x0001
1002
Andrew Mortonc986d1e2007-10-16 23:30:34 -07001003/*
Wang Sheng-Hui0e2f2b22012-09-11 08:28:18 +08001004 * Move expired (dirtied before work->older_than_this) dirty inodes from
Jan Kara697e6fe2012-03-09 07:26:22 -08001005 * @delaying_queue to @dispatch_queue.
Fengguang Wu2c136572007-10-16 23:30:39 -07001006 */
Wu Fengguange84d0a42011-04-23 12:27:27 -06001007static int move_expired_inodes(struct list_head *delaying_queue,
Fengguang Wu2c136572007-10-16 23:30:39 -07001008 struct list_head *dispatch_queue,
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001009 int flags,
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001010 struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -07001011{
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001012 unsigned long *older_than_this = NULL;
1013 unsigned long expire_time;
Shaohua Li5c034492009-09-24 14:42:33 +02001014 LIST_HEAD(tmp);
1015 struct list_head *pos, *node;
Jens Axboecf137302009-09-24 15:12:57 +02001016 struct super_block *sb = NULL;
Shaohua Li5c034492009-09-24 14:42:33 +02001017 struct inode *inode;
Jens Axboecf137302009-09-24 15:12:57 +02001018 int do_sb_sort = 0;
Wu Fengguange84d0a42011-04-23 12:27:27 -06001019 int moved = 0;
Shaohua Li5c034492009-09-24 14:42:33 +02001020
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001021 if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1022 older_than_this = work->older_than_this;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001023 else if (!work->for_sync) {
1024 expire_time = jiffies - (dirtytime_expire_interval * HZ);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001025 older_than_this = &expire_time;
1026 }
Fengguang Wu2c136572007-10-16 23:30:39 -07001027 while (!list_empty(delaying_queue)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001028 inode = wb_inode(delaying_queue->prev);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001029 if (older_than_this &&
1030 inode_dirtied_after(inode, *older_than_this))
Fengguang Wu2c136572007-10-16 23:30:39 -07001031 break;
Dave Chinnerc7f54082015-03-04 14:07:22 -05001032 list_move(&inode->i_io_list, &tmp);
Jan Karaa8855992013-07-09 22:36:45 +08001033 moved++;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001034 if (flags & EXPIRE_DIRTY_ATIME)
1035 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
Jan Karaa8855992013-07-09 22:36:45 +08001036 if (sb_is_blkdev_sb(inode->i_sb))
1037 continue;
Jens Axboecf137302009-09-24 15:12:57 +02001038 if (sb && sb != inode->i_sb)
1039 do_sb_sort = 1;
1040 sb = inode->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001041 }
1042
Jens Axboecf137302009-09-24 15:12:57 +02001043 /* just one sb in list, splice to dispatch_queue and we're done */
1044 if (!do_sb_sort) {
1045 list_splice(&tmp, dispatch_queue);
Wu Fengguange84d0a42011-04-23 12:27:27 -06001046 goto out;
Jens Axboecf137302009-09-24 15:12:57 +02001047 }
1048
Shaohua Li5c034492009-09-24 14:42:33 +02001049 /* Move inodes from one superblock together */
1050 while (!list_empty(&tmp)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001051 sb = wb_inode(tmp.prev)->i_sb;
Shaohua Li5c034492009-09-24 14:42:33 +02001052 list_for_each_prev_safe(pos, node, &tmp) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001053 inode = wb_inode(pos);
Shaohua Li5c034492009-09-24 14:42:33 +02001054 if (inode->i_sb == sb)
Dave Chinnerc7f54082015-03-04 14:07:22 -05001055 list_move(&inode->i_io_list, dispatch_queue);
Shaohua Li5c034492009-09-24 14:42:33 +02001056 }
Fengguang Wu2c136572007-10-16 23:30:39 -07001057 }
Wu Fengguange84d0a42011-04-23 12:27:27 -06001058out:
1059 return moved;
Fengguang Wu2c136572007-10-16 23:30:39 -07001060}
1061
1062/*
1063 * Queue all expired dirty inodes for io, eldest first.
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001064 * Before
1065 * newly dirtied b_dirty b_io b_more_io
1066 * =============> gf edc BA
1067 * After
1068 * newly dirtied b_dirty b_io b_more_io
1069 * =============> g fBAedc
1070 * |
1071 * +--> dequeue for IO
Fengguang Wu2c136572007-10-16 23:30:39 -07001072 */
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001073static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
Fengguang Wu2c136572007-10-16 23:30:39 -07001074{
Wu Fengguange84d0a42011-04-23 12:27:27 -06001075 int moved;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001076
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001077 assert_spin_locked(&wb->list_lock);
Wu Fengguang4ea879b2010-08-11 14:17:42 -07001078 list_splice_init(&wb->b_more_io, &wb->b_io);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001079 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1080 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1081 EXPIRE_DIRTY_ATIME, work);
Tejun Heod6c10f12015-05-22 17:13:45 -04001082 if (moved)
1083 wb_io_lists_populated(wb);
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001084 trace_writeback_queue_io(wb, work, moved);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001085}
1086
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001087static int write_inode(struct inode *inode, struct writeback_control *wbc)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001088{
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001089 int ret;
1090
1091 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1092 trace_writeback_write_inode_start(inode, wbc);
1093 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1094 trace_writeback_write_inode(inode, wbc);
1095 return ret;
1096 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001097 return 0;
Fengguang Wu2c136572007-10-16 23:30:39 -07001098}
1099
1100/*
Jan Kara169ebd92012-05-03 14:48:03 +02001101 * Wait for writeback on an inode to complete. Called with i_lock held.
1102 * Caller must make sure inode cannot go away when we drop i_lock.
Christoph Hellwig01c03192009-06-08 13:35:40 +02001103 */
Jan Kara169ebd92012-05-03 14:48:03 +02001104static void __inode_wait_for_writeback(struct inode *inode)
1105 __releases(inode->i_lock)
1106 __acquires(inode->i_lock)
Christoph Hellwig01c03192009-06-08 13:35:40 +02001107{
1108 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1109 wait_queue_head_t *wqh;
1110
1111 wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
Dave Chinner250df6e2011-03-22 22:23:36 +11001112 while (inode->i_state & I_SYNC) {
1113 spin_unlock(&inode->i_lock);
NeilBrown74316202014-07-07 15:16:04 +10001114 __wait_on_bit(wqh, &wq, bit_wait,
1115 TASK_UNINTERRUPTIBLE);
Dave Chinner250df6e2011-03-22 22:23:36 +11001116 spin_lock(&inode->i_lock);
Richard Kennedy58a9d3d82010-05-24 14:32:38 -07001117 }
Christoph Hellwig01c03192009-06-08 13:35:40 +02001118}
1119
1120/*
Jan Kara169ebd92012-05-03 14:48:03 +02001121 * Wait for writeback on an inode to complete. Caller must have inode pinned.
1122 */
1123void inode_wait_for_writeback(struct inode *inode)
1124{
1125 spin_lock(&inode->i_lock);
1126 __inode_wait_for_writeback(inode);
1127 spin_unlock(&inode->i_lock);
1128}
1129
1130/*
1131 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1132 * held and drops it. It is aimed for callers not holding any inode reference
1133 * so once i_lock is dropped, inode can go away.
1134 */
1135static void inode_sleep_on_writeback(struct inode *inode)
1136 __releases(inode->i_lock)
1137{
1138 DEFINE_WAIT(wait);
1139 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1140 int sleep;
1141
1142 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1143 sleep = inode->i_state & I_SYNC;
1144 spin_unlock(&inode->i_lock);
1145 if (sleep)
1146 schedule();
1147 finish_wait(wqh, &wait);
1148}
1149
1150/*
Jan Karaccb26b52012-05-03 14:47:58 +02001151 * Find proper writeback list for the inode depending on its current state and
1152 * possibly also change of its state while we were doing writeback. Here we
1153 * handle things such as livelock prevention or fairness of writeback among
1154 * inodes. This function can be called only by flusher thread - noone else
1155 * processes all inodes in writeback lists and requeueing inodes behind flusher
1156 * thread's back can have unexpected consequences.
1157 */
1158static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1159 struct writeback_control *wbc)
1160{
1161 if (inode->i_state & I_FREEING)
1162 return;
1163
1164 /*
1165 * Sync livelock prevention. Each inode is tagged and synced in one
1166 * shot. If still dirty, it will be redirty_tail()'ed below. Update
1167 * the dirty time to prevent enqueue and sync it again.
1168 */
1169 if ((inode->i_state & I_DIRTY) &&
1170 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1171 inode->dirtied_when = jiffies;
1172
Jan Kara4f8ad652012-05-03 14:48:00 +02001173 if (wbc->pages_skipped) {
1174 /*
1175 * writeback is not making progress due to locked
1176 * buffers. Skip this inode for now.
1177 */
1178 redirty_tail(inode, wb);
1179 return;
1180 }
1181
Jan Karaccb26b52012-05-03 14:47:58 +02001182 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1183 /*
1184 * We didn't write back all the pages. nfs_writepages()
1185 * sometimes bales out without doing anything.
1186 */
1187 if (wbc->nr_to_write <= 0) {
1188 /* Slice used up. Queue for next turn. */
1189 requeue_io(inode, wb);
1190 } else {
1191 /*
1192 * Writeback blocked by something other than
1193 * congestion. Delay the inode for some time to
1194 * avoid spinning on the CPU (100% iowait)
1195 * retrying writeback of the dirty page/inode
1196 * that cannot be performed immediately.
1197 */
1198 redirty_tail(inode, wb);
1199 }
1200 } else if (inode->i_state & I_DIRTY) {
1201 /*
1202 * Filesystems can dirty the inode during writeback operations,
1203 * such as delayed allocation during submission or metadata
1204 * updates after data IO completion.
1205 */
1206 redirty_tail(inode, wb);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001207 } else if (inode->i_state & I_DIRTY_TIME) {
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001208 inode->dirtied_when = jiffies;
Dave Chinnerc7f54082015-03-04 14:07:22 -05001209 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
Jan Karaccb26b52012-05-03 14:47:58 +02001210 } else {
1211 /* The inode is clean. Remove from writeback lists. */
Dave Chinnerc7f54082015-03-04 14:07:22 -05001212 inode_io_list_del_locked(inode, wb);
Jan Karaccb26b52012-05-03 14:47:58 +02001213 }
1214}
1215
1216/*
Jan Kara4f8ad652012-05-03 14:48:00 +02001217 * Write out an inode and its dirty pages. Do not update the writeback list
1218 * linkage. That is left to the caller. The caller is also responsible for
1219 * setting I_SYNC flag and calling inode_sync_complete() to clear it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 */
1221static int
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001222__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 struct address_space *mapping = inode->i_mapping;
Wu Fengguang251d6a42010-12-01 17:33:37 -06001225 long nr_to_write = wbc->nr_to_write;
Christoph Hellwig01c03192009-06-08 13:35:40 +02001226 unsigned dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 int ret;
1228
Jan Kara4f8ad652012-05-03 14:48:00 +02001229 WARN_ON(!(inode->i_state & I_SYNC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001231 trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1232
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 ret = do_writepages(mapping, wbc);
1234
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001235 /*
1236 * Make sure to wait on the data before writing out the metadata.
1237 * This is important for filesystems that modify metadata on data
Dave Chinner7747bd42013-07-02 22:38:35 +10001238 * I/O completion. We don't do it for sync(2) writeback because it has a
1239 * separate, external IO completion path and ->sync_fs for guaranteeing
1240 * inode metadata is written back correctly.
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001241 */
Dave Chinner7747bd42013-07-02 22:38:35 +10001242 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001243 int err = filemap_fdatawait(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244 if (ret == 0)
1245 ret = err;
1246 }
1247
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001248 /*
1249 * Some filesystems may redirty the inode during the writeback
1250 * due to delalloc, clear dirty metadata flags right before
1251 * write_inode()
1252 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001253 spin_lock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001254
Dmitry Monakhov5547e8a2010-05-07 13:35:44 +04001255 dirty = inode->i_state & I_DIRTY;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001256 if (inode->i_state & I_DIRTY_TIME) {
1257 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1258 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1259 unlikely(time_after(jiffies,
1260 (inode->dirtied_time_when +
1261 dirtytime_expire_interval * HZ)))) {
1262 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1263 trace_writeback_lazytime(inode);
1264 }
1265 } else
1266 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001267 inode->i_state &= ~dirty;
Tejun Heo9c6ac782014-10-24 15:38:21 -04001268
1269 /*
1270 * Paired with smp_mb() in __mark_inode_dirty(). This allows
1271 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1272 * either they see the I_DIRTY bits cleared or we see the dirtied
1273 * inode.
1274 *
1275 * I_DIRTY_PAGES is always cleared together above even if @mapping
1276 * still has dirty pages. The flag is reinstated after smp_mb() if
1277 * necessary. This guarantees that either __mark_inode_dirty()
1278 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1279 */
1280 smp_mb();
1281
1282 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1283 inode->i_state |= I_DIRTY_PAGES;
1284
Dave Chinner250df6e2011-03-22 22:23:36 +11001285 spin_unlock(&inode->i_lock);
Tejun Heo9c6ac782014-10-24 15:38:21 -04001286
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001287 if (dirty & I_DIRTY_TIME)
1288 mark_inode_dirty_sync(inode);
Christoph Hellwig26821ed2010-03-05 09:21:21 +01001289 /* Don't write the inode if only I_DIRTY_PAGES was set */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001290 if (dirty & ~I_DIRTY_PAGES) {
Christoph Hellwiga9185b42010-03-05 09:21:37 +01001291 int err = write_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 if (ret == 0)
1293 ret = err;
1294 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001295 trace_writeback_single_inode(inode, wbc, nr_to_write);
1296 return ret;
1297}
1298
1299/*
1300 * Write out an inode's dirty pages. Either the caller has an active reference
1301 * on the inode or the inode has I_WILL_FREE set.
1302 *
1303 * This function is designed to be called for writing back one inode which
1304 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1305 * and does more profound writeback list handling in writeback_sb_inodes().
1306 */
1307static int
1308writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
1309 struct writeback_control *wbc)
1310{
1311 int ret = 0;
1312
1313 spin_lock(&inode->i_lock);
1314 if (!atomic_read(&inode->i_count))
1315 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1316 else
1317 WARN_ON(inode->i_state & I_WILL_FREE);
1318
1319 if (inode->i_state & I_SYNC) {
1320 if (wbc->sync_mode != WB_SYNC_ALL)
1321 goto out;
1322 /*
Jan Kara169ebd92012-05-03 14:48:03 +02001323 * It's a data-integrity sync. We must wait. Since callers hold
1324 * inode reference or inode has I_WILL_FREE set, it cannot go
1325 * away under us.
Jan Kara4f8ad652012-05-03 14:48:00 +02001326 */
Jan Kara169ebd92012-05-03 14:48:03 +02001327 __inode_wait_for_writeback(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001328 }
1329 WARN_ON(inode->i_state & I_SYNC);
1330 /*
Jan Karaf9b0e052013-12-14 04:21:26 +08001331 * Skip inode if it is clean and we have no outstanding writeback in
1332 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1333 * function since flusher thread may be doing for example sync in
1334 * parallel and if we move the inode, it could get skipped. So here we
1335 * make sure inode is on some writeback list and leave it there unless
1336 * we have completely cleaned the inode.
Jan Kara4f8ad652012-05-03 14:48:00 +02001337 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001338 if (!(inode->i_state & I_DIRTY_ALL) &&
Jan Karaf9b0e052013-12-14 04:21:26 +08001339 (wbc->sync_mode != WB_SYNC_ALL ||
1340 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
Jan Kara4f8ad652012-05-03 14:48:00 +02001341 goto out;
1342 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001343 wbc_attach_and_unlock_inode(wbc, inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001344
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001345 ret = __writeback_single_inode(inode, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346
Tejun Heob16b1de2015-06-02 08:39:48 -06001347 wbc_detach_inode(wbc);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001348 spin_lock(&wb->list_lock);
Dave Chinner250df6e2011-03-22 22:23:36 +11001349 spin_lock(&inode->i_lock);
Jan Kara4f8ad652012-05-03 14:48:00 +02001350 /*
1351 * If inode is clean, remove it from writeback lists. Otherwise don't
1352 * touch it. See comment above for explanation.
1353 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001354 if (!(inode->i_state & I_DIRTY_ALL))
Dave Chinnerc7f54082015-03-04 14:07:22 -05001355 inode_io_list_del_locked(inode, wb);
Jan Kara4f8ad652012-05-03 14:48:00 +02001356 spin_unlock(&wb->list_lock);
Joern Engel1c0eeaf2007-10-16 23:30:44 -07001357 inode_sync_complete(inode);
Jan Kara4f8ad652012-05-03 14:48:00 +02001358out:
1359 spin_unlock(&inode->i_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 return ret;
1361}
1362
Tejun Heoa88a3412015-05-22 17:13:28 -04001363static long writeback_chunk_size(struct bdi_writeback *wb,
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001364 struct wb_writeback_work *work)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001365{
1366 long pages;
1367
1368 /*
1369 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1370 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1371 * here avoids calling into writeback_inodes_wb() more than once.
1372 *
1373 * The intended call sequence for WB_SYNC_ALL writeback is:
1374 *
1375 * wb_writeback()
1376 * writeback_sb_inodes() <== called only once
1377 * write_cache_pages() <== called once for each inode
1378 * (quickly) tag currently dirty pages
1379 * (maybe slowly) sync all tagged pages
1380 */
1381 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1382 pages = LONG_MAX;
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001383 else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001384 pages = min(wb->avg_write_bandwidth / 2,
Tejun Heodcc25ae2015-05-22 18:23:22 -04001385 global_wb_domain.dirty_limit / DIRTY_SCOPE);
Wu Fengguang1a12d8b2010-08-29 13:28:09 -06001386 pages = min(pages, work->nr_pages);
1387 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1388 MIN_WRITEBACK_PAGES);
1389 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001390
1391 return pages;
1392}
1393
Jens Axboe03ba3782009-09-09 09:08:54 +02001394/*
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001395 * Write a portion of b_io inodes which belong to @sb.
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001396 *
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001397 * Return the number of pages and/or inodes written.
Linus Torvalds0ba13fd2015-09-11 13:26:39 -07001398 *
1399 * NOTE! This is called with wb->list_lock held, and will
1400 * unlock and relock that for each inode it ends up doing
1401 * IO for.
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001402 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001403static long writeback_sb_inodes(struct super_block *sb,
1404 struct bdi_writeback *wb,
1405 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001406{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001407 struct writeback_control wbc = {
1408 .sync_mode = work->sync_mode,
1409 .tagged_writepages = work->tagged_writepages,
1410 .for_kupdate = work->for_kupdate,
1411 .for_background = work->for_background,
Dave Chinner7747bd42013-07-02 22:38:35 +10001412 .for_sync = work->for_sync,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001413 .range_cyclic = work->range_cyclic,
1414 .range_start = 0,
1415 .range_end = LLONG_MAX,
1416 };
1417 unsigned long start_time = jiffies;
1418 long write_chunk;
1419 long wrote = 0; /* count both pages and inodes */
1420
Jens Axboe03ba3782009-09-09 09:08:54 +02001421 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001422 struct inode *inode = wb_inode(wb->b_io.prev);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001423
1424 if (inode->i_sb != sb) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001425 if (work->sb) {
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001426 /*
1427 * We only want to write back data for this
1428 * superblock, move all inodes not belonging
1429 * to it back onto the dirty list.
1430 */
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001431 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001432 continue;
1433 }
1434
1435 /*
1436 * The inode belongs to a different superblock.
1437 * Bounce back to the caller to unpin this and
1438 * pin the next superblock.
1439 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001440 break;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001441 }
1442
Christoph Hellwig9843b762010-10-24 19:40:46 +02001443 /*
Wanpeng Li331cbde2012-06-09 11:10:55 +08001444 * Don't bother with new inodes or inodes being freed, first
1445 * kind does not need periodic writeout yet, and for the latter
Christoph Hellwig9843b762010-10-24 19:40:46 +02001446 * kind writeout is handled by the freer.
1447 */
Dave Chinner250df6e2011-03-22 22:23:36 +11001448 spin_lock(&inode->i_lock);
Christoph Hellwig9843b762010-10-24 19:40:46 +02001449 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
Dave Chinner250df6e2011-03-22 22:23:36 +11001450 spin_unlock(&inode->i_lock);
Wu Fengguangfcc5c222011-07-11 23:08:50 -07001451 redirty_tail(inode, wb);
Nick Piggin7ef0d732009-03-12 14:31:38 -07001452 continue;
1453 }
Jan Karacc1676d2012-05-03 14:47:56 +02001454 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1455 /*
1456 * If this inode is locked for writeback and we are not
1457 * doing writeback-for-data-integrity, move it to
1458 * b_more_io so that writeback can proceed with the
1459 * other inodes on s_io.
1460 *
1461 * We'll have another go at writing back this inode
1462 * when we completed a full scan of b_io.
1463 */
1464 spin_unlock(&inode->i_lock);
1465 requeue_io(inode, wb);
1466 trace_writeback_sb_inodes_requeue(inode);
1467 continue;
1468 }
Jan Karaf0d07b72012-05-03 14:47:59 +02001469 spin_unlock(&wb->list_lock);
1470
Jan Kara4f8ad652012-05-03 14:48:00 +02001471 /*
1472 * We already requeued the inode if it had I_SYNC set and we
1473 * are doing WB_SYNC_NONE writeback. So this catches only the
1474 * WB_SYNC_ALL case.
1475 */
Jan Kara169ebd92012-05-03 14:48:03 +02001476 if (inode->i_state & I_SYNC) {
1477 /* Wait for I_SYNC. This function drops i_lock... */
1478 inode_sleep_on_writeback(inode);
1479 /* Inode may be gone, start again */
Jan Karaead188f2012-06-08 17:07:36 +02001480 spin_lock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001481 continue;
1482 }
Jan Kara4f8ad652012-05-03 14:48:00 +02001483 inode->i_state |= I_SYNC;
Tejun Heob16b1de2015-06-02 08:39:48 -06001484 wbc_attach_and_unlock_inode(&wbc, inode);
Jan Kara169ebd92012-05-03 14:48:03 +02001485
Tejun Heoa88a3412015-05-22 17:13:28 -04001486 write_chunk = writeback_chunk_size(wb, work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001487 wbc.nr_to_write = write_chunk;
1488 wbc.pages_skipped = 0;
Dave Chinner250df6e2011-03-22 22:23:36 +11001489
Jan Kara169ebd92012-05-03 14:48:03 +02001490 /*
1491 * We use I_SYNC to pin the inode in memory. While it is set
1492 * evict_inode() will wait so the inode cannot be freed.
1493 */
Yan Hongcd8ed2a2012-10-08 16:33:45 -07001494 __writeback_single_inode(inode, &wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001495
Tejun Heob16b1de2015-06-02 08:39:48 -06001496 wbc_detach_inode(&wbc);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001497 work->nr_pages -= write_chunk - wbc.nr_to_write;
1498 wrote += write_chunk - wbc.nr_to_write;
Chris Mason590dca32015-09-18 13:35:08 -04001499
1500 if (need_resched()) {
1501 /*
1502 * We're trying to balance between building up a nice
1503 * long list of IOs to improve our merge rate, and
1504 * getting those IOs out quickly for anyone throttling
1505 * in balance_dirty_pages(). cond_resched() doesn't
1506 * unplug, so get our IOs out the door before we
1507 * give up the CPU.
1508 */
1509 blk_flush_plug(current);
1510 cond_resched();
1511 }
1512
1513
Jan Kara4f8ad652012-05-03 14:48:00 +02001514 spin_lock(&wb->list_lock);
1515 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001516 if (!(inode->i_state & I_DIRTY_ALL))
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001517 wrote++;
Jan Kara4f8ad652012-05-03 14:48:00 +02001518 requeue_inode(inode, wb, &wbc);
1519 inode_sync_complete(inode);
Dave Chinner0f1b1fd2011-03-22 22:23:43 +11001520 spin_unlock(&inode->i_lock);
Chris Mason590dca32015-09-18 13:35:08 -04001521
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001522 /*
1523 * bail out to wb_writeback() often enough to check
1524 * background threshold and other termination conditions.
1525 */
1526 if (wrote) {
1527 if (time_is_before_jiffies(start_time + HZ / 10UL))
1528 break;
1529 if (work->nr_pages <= 0)
1530 break;
Fengguang Wu8bc3be22008-02-04 22:29:36 -08001531 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001533 return wrote;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001534}
Nick Piggin38f21972009-01-06 14:40:25 -08001535
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001536static long __writeback_inodes_wb(struct bdi_writeback *wb,
1537 struct wb_writeback_work *work)
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001538{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001539 unsigned long start_time = jiffies;
1540 long wrote = 0;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001541
1542 while (!list_empty(&wb->b_io)) {
Nick Piggin7ccf19a2010-10-21 11:49:30 +11001543 struct inode *inode = wb_inode(wb->b_io.prev);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001544 struct super_block *sb = inode->i_sb;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001545
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001546 if (!trylock_super(sb)) {
Wu Fengguang0e9958162011-07-29 22:14:35 -06001547 /*
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001548 * trylock_super() may fail consistently due to
Wu Fengguang0e9958162011-07-29 22:14:35 -06001549 * s_umount being grabbed by someone else. Don't use
1550 * requeue_io() to avoid busy retrying the inode/sb.
1551 */
1552 redirty_tail(inode, wb);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001553 continue;
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001554 }
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001555 wrote += writeback_sb_inodes(sb, wb, work);
Konstantin Khlebnikoveb6ef3d2015-02-19 20:19:35 +03001556 up_read(&sb->s_umount);
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001557
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001558 /* refer to the same tests at the end of writeback_sb_inodes */
1559 if (wrote) {
1560 if (time_is_before_jiffies(start_time + HZ / 10UL))
1561 break;
1562 if (work->nr_pages <= 0)
1563 break;
1564 }
Edward Shishkinf11c9c52010-03-11 14:09:47 -08001565 }
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001566 /* Leave any unwritten inodes on b_io */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001567 return wrote;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001568}
1569
Wanpeng Li7d9f0732013-09-11 14:22:40 -07001570static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001571 enum wb_reason reason)
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001572{
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001573 struct wb_writeback_work work = {
1574 .nr_pages = nr_pages,
1575 .sync_mode = WB_SYNC_NONE,
1576 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001577 .reason = reason,
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001578 };
Linus Torvalds505a6662015-09-11 13:37:19 -07001579 struct blk_plug plug;
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001580
Linus Torvalds505a6662015-09-11 13:37:19 -07001581 blk_start_plug(&plug);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001582 spin_lock(&wb->list_lock);
Wu Fengguang424b3512010-07-21 20:11:53 -06001583 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001584 queue_io(wb, &work);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001585 __writeback_inodes_wb(wb, &work);
Christoph Hellwigf758eea2011-04-21 18:19:44 -06001586 spin_unlock(&wb->list_lock);
Linus Torvalds505a6662015-09-11 13:37:19 -07001587 blk_finish_plug(&plug);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001588
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001589 return nr_pages - work.nr_pages;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02001590}
1591
Jens Axboe03ba3782009-09-09 09:08:54 +02001592/*
1593 * Explicit flushing or periodic writeback of "old" data.
1594 *
1595 * Define "old": the first time one of an inode's pages is dirtied, we mark the
1596 * dirtying-time in the inode's address_space. So this periodic writeback code
1597 * just walks the superblock inode list, writing back any inodes which are
1598 * older than a specific point in time.
1599 *
1600 * Try to run once per dirty_writeback_interval. But if a writeback event
1601 * takes longer than a dirty_writeback_interval interval, then leave a
1602 * one-second gap.
1603 *
1604 * older_than_this takes precedence over nr_to_write. So we'll only write back
1605 * all dirty pages if they are all attached to "old" mappings.
1606 */
Jens Axboec4a77a62009-09-16 15:18:25 +02001607static long wb_writeback(struct bdi_writeback *wb,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001608 struct wb_writeback_work *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001609{
Wu Fengguange98be2d2010-08-29 11:22:30 -06001610 unsigned long wb_start = jiffies;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001611 long nr_pages = work->nr_pages;
Jan Kara0dc83bd2014-02-21 11:19:04 +01001612 unsigned long oldest_jif;
Jan Karaa5989bd2009-09-16 19:22:48 +02001613 struct inode *inode;
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001614 long progress;
Linus Torvalds505a6662015-09-11 13:37:19 -07001615 struct blk_plug plug;
Jens Axboe03ba3782009-09-09 09:08:54 +02001616
Jan Kara0dc83bd2014-02-21 11:19:04 +01001617 oldest_jif = jiffies;
1618 work->older_than_this = &oldest_jif;
Jens Axboe03ba3782009-09-09 09:08:54 +02001619
Linus Torvalds505a6662015-09-11 13:37:19 -07001620 blk_start_plug(&plug);
Wu Fengguange8dfc302011-04-21 12:06:32 -06001621 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001622 for (;;) {
1623 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001624 * Stop writeback when nr_pages has been consumed
Jens Axboe03ba3782009-09-09 09:08:54 +02001625 */
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001626 if (work->nr_pages <= 0)
Jens Axboe03ba3782009-09-09 09:08:54 +02001627 break;
1628
1629 /*
Jan Karaaa373cf2011-01-13 15:45:47 -08001630 * Background writeout and kupdate-style writeback may
1631 * run forever. Stop them if there is other work to do
1632 * so that e.g. sync can proceed. They'll be restarted
1633 * after the other works are all done.
1634 */
1635 if ((work->for_background || work->for_kupdate) &&
Tejun Heof0054bb2015-05-22 17:13:30 -04001636 !list_empty(&wb->work_list))
Jan Karaaa373cf2011-01-13 15:45:47 -08001637 break;
1638
1639 /*
Wu Fengguangd3ddec72009-09-23 20:33:40 +08001640 * For background writeout, stop when we are below the
1641 * background dirty threshold
Jens Axboe03ba3782009-09-09 09:08:54 +02001642 */
Tejun Heoaa661bb2015-05-22 18:23:31 -04001643 if (work->for_background && !wb_over_bg_thresh(wb))
Jens Axboe03ba3782009-09-09 09:08:54 +02001644 break;
1645
Jan Kara1bc36b62011-10-19 11:44:41 +02001646 /*
1647 * Kupdate and background works are special and we want to
1648 * include all inodes that need writing. Livelock avoidance is
1649 * handled by these works yielding to any other work so we are
1650 * safe.
1651 */
Wu Fengguangba9aa832010-07-21 20:32:30 -06001652 if (work->for_kupdate) {
Jan Kara0dc83bd2014-02-21 11:19:04 +01001653 oldest_jif = jiffies -
Wu Fengguangba9aa832010-07-21 20:32:30 -06001654 msecs_to_jiffies(dirty_expire_interval * 10);
Jan Kara1bc36b62011-10-19 11:44:41 +02001655 } else if (work->for_background)
Jan Kara0dc83bd2014-02-21 11:19:04 +01001656 oldest_jif = jiffies;
Dave Chinner028c2dd2010-07-07 13:24:07 +10001657
Tejun Heo5634cc22015-08-18 14:54:56 -07001658 trace_writeback_start(wb, work);
Wu Fengguange8dfc302011-04-21 12:06:32 -06001659 if (list_empty(&wb->b_io))
Curt Wohlgemuthad4e38d2011-10-07 21:51:56 -06001660 queue_io(wb, work);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001661 if (work->sb)
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001662 progress = writeback_sb_inodes(work->sb, wb, work);
Christoph Hellwigedadfb12010-06-10 12:07:54 +02001663 else
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001664 progress = __writeback_inodes_wb(wb, work);
Tejun Heo5634cc22015-08-18 14:54:56 -07001665 trace_writeback_written(wb, work);
Dave Chinner028c2dd2010-07-07 13:24:07 +10001666
Wu Fengguange98be2d2010-08-29 11:22:30 -06001667 wb_update_bandwidth(wb, wb_start);
Jens Axboe03ba3782009-09-09 09:08:54 +02001668
1669 /*
Jens Axboe71fd05a2009-09-23 19:32:26 +02001670 * Did we write something? Try for more
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001671 *
1672 * Dirty inodes are moved to b_io for writeback in batches.
1673 * The completion of the current batch does not necessarily
1674 * mean the overall work is done. So we keep looping as long
1675 * as made some progress on cleaning pages or inodes.
Jens Axboe71fd05a2009-09-23 19:32:26 +02001676 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001677 if (progress)
Jens Axboe71fd05a2009-09-23 19:32:26 +02001678 continue;
1679 /*
Wu Fengguange6fb6da2010-07-22 10:23:44 -06001680 * No more inodes for IO, bail
Jens Axboe71fd05a2009-09-23 19:32:26 +02001681 */
Wu Fengguangb7a24412010-07-21 22:19:51 -06001682 if (list_empty(&wb->b_more_io))
Jens Axboe03ba3782009-09-09 09:08:54 +02001683 break;
1684 /*
Jens Axboe8010c3b2009-09-15 20:04:57 +02001685 * Nothing written. Wait for some inode to
1686 * become available for writeback. Otherwise
1687 * we'll just busyloop.
Jens Axboe03ba3782009-09-09 09:08:54 +02001688 */
Jens Axboe03ba3782009-09-09 09:08:54 +02001689 if (!list_empty(&wb->b_more_io)) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001690 trace_writeback_wait(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001691 inode = wb_inode(wb->b_more_io.prev);
Dave Chinner250df6e2011-03-22 22:23:36 +11001692 spin_lock(&inode->i_lock);
Jan Karaf0d07b72012-05-03 14:47:59 +02001693 spin_unlock(&wb->list_lock);
Jan Kara169ebd92012-05-03 14:48:03 +02001694 /* This function drops i_lock... */
1695 inode_sleep_on_writeback(inode);
Jan Karaf0d07b72012-05-03 14:47:59 +02001696 spin_lock(&wb->list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02001697 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001698 }
Wu Fengguange8dfc302011-04-21 12:06:32 -06001699 spin_unlock(&wb->list_lock);
Linus Torvalds505a6662015-09-11 13:37:19 -07001700 blk_finish_plug(&plug);
Jens Axboe03ba3782009-09-09 09:08:54 +02001701
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001702 return nr_pages - work->nr_pages;
Jens Axboe03ba3782009-09-09 09:08:54 +02001703}
1704
1705/*
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001706 * Return the next wb_writeback_work struct that hasn't been processed yet.
Jens Axboe03ba3782009-09-09 09:08:54 +02001707 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001708static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001709{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001710 struct wb_writeback_work *work = NULL;
Jens Axboe03ba3782009-09-09 09:08:54 +02001711
Tejun Heof0054bb2015-05-22 17:13:30 -04001712 spin_lock_bh(&wb->work_lock);
1713 if (!list_empty(&wb->work_list)) {
1714 work = list_entry(wb->work_list.next,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001715 struct wb_writeback_work, list);
1716 list_del_init(&work->list);
Jens Axboe03ba3782009-09-09 09:08:54 +02001717 }
Tejun Heof0054bb2015-05-22 17:13:30 -04001718 spin_unlock_bh(&wb->work_lock);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001719 return work;
Jens Axboe03ba3782009-09-09 09:08:54 +02001720}
1721
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001722/*
1723 * Add in the number of potentially dirty inodes, because each inode
1724 * write can dirty pagecache in the underlying blockdev.
1725 */
1726static unsigned long get_nr_dirty_pages(void)
1727{
1728 return global_page_state(NR_FILE_DIRTY) +
1729 global_page_state(NR_UNSTABLE_NFS) +
1730 get_nr_dirty_inodes();
1731}
1732
Jan Kara65850272011-01-13 15:45:44 -08001733static long wb_check_background_flush(struct bdi_writeback *wb)
1734{
Tejun Heoaa661bb2015-05-22 18:23:31 -04001735 if (wb_over_bg_thresh(wb)) {
Jan Kara65850272011-01-13 15:45:44 -08001736
1737 struct wb_writeback_work work = {
1738 .nr_pages = LONG_MAX,
1739 .sync_mode = WB_SYNC_NONE,
1740 .for_background = 1,
1741 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001742 .reason = WB_REASON_BACKGROUND,
Jan Kara65850272011-01-13 15:45:44 -08001743 };
1744
1745 return wb_writeback(wb, &work);
1746 }
1747
1748 return 0;
1749}
1750
Jens Axboe03ba3782009-09-09 09:08:54 +02001751static long wb_check_old_data_flush(struct bdi_writeback *wb)
1752{
1753 unsigned long expired;
1754 long nr_pages;
1755
Jens Axboe69b62d02010-05-17 12:51:03 +02001756 /*
1757 * When set to zero, disable periodic writeback
1758 */
1759 if (!dirty_writeback_interval)
1760 return 0;
1761
Jens Axboe03ba3782009-09-09 09:08:54 +02001762 expired = wb->last_old_flush +
1763 msecs_to_jiffies(dirty_writeback_interval * 10);
1764 if (time_before(jiffies, expired))
1765 return 0;
1766
1767 wb->last_old_flush = jiffies;
Linus Torvaldscdf01dd2010-10-30 08:55:52 -07001768 nr_pages = get_nr_dirty_pages();
Jens Axboe03ba3782009-09-09 09:08:54 +02001769
Jens Axboec4a77a62009-09-16 15:18:25 +02001770 if (nr_pages) {
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001771 struct wb_writeback_work work = {
Jens Axboec4a77a62009-09-16 15:18:25 +02001772 .nr_pages = nr_pages,
1773 .sync_mode = WB_SYNC_NONE,
1774 .for_kupdate = 1,
1775 .range_cyclic = 1,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001776 .reason = WB_REASON_PERIODIC,
Jens Axboec4a77a62009-09-16 15:18:25 +02001777 };
1778
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001779 return wb_writeback(wb, &work);
Jens Axboec4a77a62009-09-16 15:18:25 +02001780 }
Jens Axboe03ba3782009-09-09 09:08:54 +02001781
1782 return 0;
1783}
1784
1785/*
1786 * Retrieve work items and do the writeback they describe
1787 */
Wanpeng Li25d130b2013-07-08 16:00:14 -07001788static long wb_do_writeback(struct bdi_writeback *wb)
Jens Axboe03ba3782009-09-09 09:08:54 +02001789{
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001790 struct wb_writeback_work *work;
Jens Axboec4a77a62009-09-16 15:18:25 +02001791 long wrote = 0;
Jens Axboe03ba3782009-09-09 09:08:54 +02001792
Tejun Heo44522262015-05-22 17:13:26 -04001793 set_bit(WB_writeback_running, &wb->state);
Tejun Heof0054bb2015-05-22 17:13:30 -04001794 while ((work = get_next_work_item(wb)) != NULL) {
Tejun Heocc395d72015-05-22 17:13:58 -04001795 struct wb_completion *done = work->done;
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001796
Tejun Heo5634cc22015-08-18 14:54:56 -07001797 trace_writeback_exec(wb, work);
Dave Chinner455b2862010-07-07 13:24:06 +10001798
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001799 wrote += wb_writeback(wb, work);
Jens Axboe03ba3782009-09-09 09:08:54 +02001800
Tejun Heo8a1270c2015-08-18 14:54:53 -07001801 if (work->auto_free)
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02001802 kfree(work);
Tejun Heocc395d72015-05-22 17:13:58 -04001803 if (done && atomic_dec_and_test(&done->cnt))
1804 wake_up_all(&wb->bdi->wb_waitq);
Jens Axboe03ba3782009-09-09 09:08:54 +02001805 }
1806
1807 /*
1808 * Check for periodic writeback, kupdated() style
1809 */
1810 wrote += wb_check_old_data_flush(wb);
Jan Kara65850272011-01-13 15:45:44 -08001811 wrote += wb_check_background_flush(wb);
Tejun Heo44522262015-05-22 17:13:26 -04001812 clear_bit(WB_writeback_running, &wb->state);
Jens Axboe03ba3782009-09-09 09:08:54 +02001813
1814 return wrote;
1815}
1816
1817/*
1818 * Handle writeback of dirty data for the device backed by this bdi. Also
Tejun Heo839a8e82013-04-01 19:08:06 -07001819 * reschedules periodically and does kupdated style flushing.
Jens Axboe03ba3782009-09-09 09:08:54 +02001820 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001821void wb_workfn(struct work_struct *work)
Jens Axboe03ba3782009-09-09 09:08:54 +02001822{
Tejun Heo839a8e82013-04-01 19:08:06 -07001823 struct bdi_writeback *wb = container_of(to_delayed_work(work),
1824 struct bdi_writeback, dwork);
Jens Axboe03ba3782009-09-09 09:08:54 +02001825 long pages_written;
1826
Tejun Heof0054bb2015-05-22 17:13:30 -04001827 set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
Peter Zijlstra766f9162010-10-26 14:22:45 -07001828 current->flags |= PF_SWAPWRITE;
Christoph Hellwig08243902010-06-19 23:08:22 +02001829
Tejun Heo839a8e82013-04-01 19:08:06 -07001830 if (likely(!current_is_workqueue_rescuer() ||
Tejun Heo44522262015-05-22 17:13:26 -04001831 !test_bit(WB_registered, &wb->state))) {
Artem Bityutskiy64677162010-07-25 14:29:22 +03001832 /*
Tejun Heof0054bb2015-05-22 17:13:30 -04001833 * The normal path. Keep writing back @wb until its
Tejun Heo839a8e82013-04-01 19:08:06 -07001834 * work_list is empty. Note that this path is also taken
Tejun Heof0054bb2015-05-22 17:13:30 -04001835 * if @wb is shutting down even when we're running off the
Tejun Heo839a8e82013-04-01 19:08:06 -07001836 * rescuer as work_list needs to be drained.
Artem Bityutskiy64677162010-07-25 14:29:22 +03001837 */
Tejun Heo839a8e82013-04-01 19:08:06 -07001838 do {
Wanpeng Li25d130b2013-07-08 16:00:14 -07001839 pages_written = wb_do_writeback(wb);
Tejun Heo839a8e82013-04-01 19:08:06 -07001840 trace_writeback_pages_written(pages_written);
Tejun Heof0054bb2015-05-22 17:13:30 -04001841 } while (!list_empty(&wb->work_list));
Tejun Heo839a8e82013-04-01 19:08:06 -07001842 } else {
1843 /*
1844 * bdi_wq can't get enough workers and we're running off
1845 * the emergency worker. Don't hog it. Hopefully, 1024 is
1846 * enough for efficient IO.
1847 */
Tejun Heof0054bb2015-05-22 17:13:30 -04001848 pages_written = writeback_inodes_wb(wb, 1024,
Tejun Heo839a8e82013-04-01 19:08:06 -07001849 WB_REASON_FORKER_THREAD);
Dave Chinner455b2862010-07-07 13:24:06 +10001850 trace_writeback_pages_written(pages_written);
Jens Axboe03ba3782009-09-09 09:08:54 +02001851 }
1852
Tejun Heof0054bb2015-05-22 17:13:30 -04001853 if (!list_empty(&wb->work_list))
Derek Basehore6ca738d2014-04-03 14:46:22 -07001854 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1855 else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
Tejun Heof0054bb2015-05-22 17:13:30 -04001856 wb_wakeup_delayed(wb);
Dave Chinner455b2862010-07-07 13:24:06 +10001857
Tejun Heo839a8e82013-04-01 19:08:06 -07001858 current->flags &= ~PF_SWAPWRITE;
Jens Axboe03ba3782009-09-09 09:08:54 +02001859}
1860
1861/*
Jens Axboe03ba3782009-09-09 09:08:54 +02001862 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
1863 * the whole world.
1864 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001865void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
Jens Axboe03ba3782009-09-09 09:08:54 +02001866{
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001867 struct backing_dev_info *bdi;
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001868
Jan Kara47df3dd2013-09-11 14:22:22 -07001869 if (!nr_pages)
1870 nr_pages = get_nr_dirty_pages();
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001871
1872 rcu_read_lock();
Tejun Heof2b65122015-05-22 17:13:55 -04001873 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1874 struct bdi_writeback *wb;
Tejun Heof2b65122015-05-22 17:13:55 -04001875
1876 if (!bdi_has_dirty_io(bdi))
1877 continue;
1878
Tejun Heob8175252015-10-02 14:47:05 -04001879 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
Tejun Heof2b65122015-05-22 17:13:55 -04001880 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1881 false, reason);
1882 }
Christoph Hellwigb8c2f342010-06-08 18:15:07 +02001883 rcu_read_unlock();
Jens Axboe03ba3782009-09-09 09:08:54 +02001884}
1885
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001886/*
1887 * Wake up bdi's periodically to make sure dirtytime inodes gets
1888 * written back periodically. We deliberately do *not* check the
1889 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1890 * kernel to be constantly waking up once there are any dirtytime
1891 * inodes on the system. So instead we define a separate delayed work
1892 * function which gets called much more rarely. (By default, only
1893 * once every 12 hours.)
1894 *
1895 * If there is any other write activity going on in the file system,
1896 * this function won't be necessary. But if the only thing that has
1897 * happened on the file system is a dirtytime inode caused by an atime
1898 * update, we need this infrastructure below to make sure that inode
1899 * eventually gets pushed out to disk.
1900 */
1901static void wakeup_dirtytime_writeback(struct work_struct *w);
1902static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1903
1904static void wakeup_dirtytime_writeback(struct work_struct *w)
1905{
1906 struct backing_dev_info *bdi;
1907
1908 rcu_read_lock();
1909 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
Tejun Heo001fe6f2015-05-22 17:13:56 -04001910 struct bdi_writeback *wb;
Tejun Heo001fe6f2015-05-22 17:13:56 -04001911
Tejun Heob8175252015-10-02 14:47:05 -04001912 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
Tejun Heo6fdf8602015-09-29 12:47:51 -04001913 if (!list_empty(&wb->b_dirty_time))
1914 wb_wakeup(wb);
Theodore Ts'oa2f48702015-03-17 12:23:19 -04001915 }
1916 rcu_read_unlock();
1917 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1918}
1919
1920static int __init start_dirtytime_writeback(void)
1921{
1922 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1923 return 0;
1924}
1925__initcall(start_dirtytime_writeback);
1926
Theodore Ts'o1efff912015-03-17 12:23:32 -04001927int dirtytime_interval_handler(struct ctl_table *table, int write,
1928 void __user *buffer, size_t *lenp, loff_t *ppos)
1929{
1930 int ret;
1931
1932 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1933 if (ret == 0 && write)
1934 mod_delayed_work(system_wq, &dirtytime_work, 0);
1935 return ret;
1936}
1937
Jens Axboe03ba3782009-09-09 09:08:54 +02001938static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1939{
1940 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1941 struct dentry *dentry;
1942 const char *name = "?";
1943
1944 dentry = d_find_alias(inode);
1945 if (dentry) {
1946 spin_lock(&dentry->d_lock);
1947 name = (const char *) dentry->d_name.name;
1948 }
1949 printk(KERN_DEBUG
1950 "%s(%d): dirtied inode %lu (%s) on %s\n",
1951 current->comm, task_pid_nr(current), inode->i_ino,
1952 name, inode->i_sb->s_id);
1953 if (dentry) {
1954 spin_unlock(&dentry->d_lock);
1955 dput(dentry);
1956 }
1957 }
1958}
1959
1960/**
1961 * __mark_inode_dirty - internal function
1962 * @inode: inode to mark
1963 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1964 * Mark an inode as dirty. Callers should use mark_inode_dirty or
1965 * mark_inode_dirty_sync.
1966 *
1967 * Put the inode on the super block's dirty list.
1968 *
1969 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1970 * dirty list only if it is hashed or if it refers to a blockdev.
1971 * If it was not hashed, it will never be added to the dirty list
1972 * even if it is later hashed, as it will have been marked dirty already.
1973 *
1974 * In short, make sure you hash any inodes _before_ you start marking
1975 * them dirty.
1976 *
Jens Axboe03ba3782009-09-09 09:08:54 +02001977 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1978 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
1979 * the kernel-internal blockdev inode represents the dirtying time of the
1980 * blockdev's pages. This is why for I_DIRTY_PAGES we always use
1981 * page->mapping->host, so the page-dirtying time is recorded in the internal
1982 * blockdev inode.
1983 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001984#define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
Jens Axboe03ba3782009-09-09 09:08:54 +02001985void __mark_inode_dirty(struct inode *inode, int flags)
1986{
1987 struct super_block *sb = inode->i_sb;
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001988 int dirtytime;
1989
1990 trace_writeback_mark_inode_dirty(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02001991
1992 /*
1993 * Don't do this for I_DIRTY_PAGES - that doesn't actually
1994 * dirty the inode itself
1995 */
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05001996 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08001997 trace_writeback_dirty_inode_start(inode, flags);
1998
Jens Axboe03ba3782009-09-09 09:08:54 +02001999 if (sb->s_op->dirty_inode)
Christoph Hellwigaa385722011-05-27 06:53:02 -04002000 sb->s_op->dirty_inode(inode, flags);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002001
2002 trace_writeback_dirty_inode(inode, flags);
Jens Axboe03ba3782009-09-09 09:08:54 +02002003 }
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002004 if (flags & I_DIRTY_INODE)
2005 flags &= ~I_DIRTY_TIME;
2006 dirtytime = flags & I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02002007
2008 /*
Tejun Heo9c6ac782014-10-24 15:38:21 -04002009 * Paired with smp_mb() in __writeback_single_inode() for the
2010 * following lockless i_state test. See there for details.
Jens Axboe03ba3782009-09-09 09:08:54 +02002011 */
2012 smp_mb();
2013
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002014 if (((inode->i_state & flags) == flags) ||
2015 (dirtytime && (inode->i_state & I_DIRTY_INODE)))
Jens Axboe03ba3782009-09-09 09:08:54 +02002016 return;
2017
2018 if (unlikely(block_dump))
2019 block_dump___mark_inode_dirty(inode);
2020
Dave Chinner250df6e2011-03-22 22:23:36 +11002021 spin_lock(&inode->i_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002022 if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2023 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002024 if ((inode->i_state & flags) != flags) {
2025 const int was_dirty = inode->i_state & I_DIRTY;
2026
Tejun Heo52ebea72015-05-22 17:13:37 -04002027 inode_attach_wb(inode, NULL);
2028
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002029 if (flags & I_DIRTY_INODE)
2030 inode->i_state &= ~I_DIRTY_TIME;
Jens Axboe03ba3782009-09-09 09:08:54 +02002031 inode->i_state |= flags;
2032
2033 /*
2034 * If the inode is being synced, just update its dirty state.
2035 * The unlocker will place the inode on the appropriate
2036 * superblock list, based upon its state.
2037 */
2038 if (inode->i_state & I_SYNC)
Dave Chinner250df6e2011-03-22 22:23:36 +11002039 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002040
2041 /*
2042 * Only add valid (hashed) inodes to the superblock's
2043 * dirty list. Add blockdev inodes as well.
2044 */
2045 if (!S_ISBLK(inode->i_mode)) {
Al Viro1d3382cb2010-10-23 15:19:20 -04002046 if (inode_unhashed(inode))
Dave Chinner250df6e2011-03-22 22:23:36 +11002047 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002048 }
Al Viroa4ffdde2010-06-02 17:38:30 -04002049 if (inode->i_state & I_FREEING)
Dave Chinner250df6e2011-03-22 22:23:36 +11002050 goto out_unlock_inode;
Jens Axboe03ba3782009-09-09 09:08:54 +02002051
2052 /*
2053 * If the inode was already on b_dirty/b_io/b_more_io, don't
2054 * reposition it (that would break b_dirty time-ordering).
2055 */
2056 if (!was_dirty) {
Tejun Heo87e1d782015-05-28 14:50:52 -04002057 struct bdi_writeback *wb;
Tejun Heod6c10f12015-05-22 17:13:45 -04002058 struct list_head *dirty_list;
Dave Chinnera66979a2011-03-22 22:23:41 +11002059 bool wakeup_bdi = false;
Jens Axboe500b0672009-09-09 09:10:25 +02002060
Tejun Heo87e1d782015-05-28 14:50:52 -04002061 wb = locked_inode_to_wb_and_lock_list(inode);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002062
Tejun Heo07472592015-05-22 17:14:02 -04002063 WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2064 !test_bit(WB_registered, &wb->state),
2065 "bdi-%s not registered\n", wb->bdi->name);
Jens Axboe03ba3782009-09-09 09:08:54 +02002066
2067 inode->dirtied_when = jiffies;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002068 if (dirtytime)
2069 inode->dirtied_time_when = jiffies;
Tejun Heod6c10f12015-05-22 17:13:45 -04002070
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002071 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
Tejun Heo07472592015-05-22 17:14:02 -04002072 dirty_list = &wb->b_dirty;
Theodore Ts'oa2f48702015-03-17 12:23:19 -04002073 else
Tejun Heo07472592015-05-22 17:14:02 -04002074 dirty_list = &wb->b_dirty_time;
Tejun Heod6c10f12015-05-22 17:13:45 -04002075
Dave Chinnerc7f54082015-03-04 14:07:22 -05002076 wakeup_bdi = inode_io_list_move_locked(inode, wb,
Tejun Heod6c10f12015-05-22 17:13:45 -04002077 dirty_list);
2078
Tejun Heo07472592015-05-22 17:14:02 -04002079 spin_unlock(&wb->list_lock);
Theodore Ts'o0ae45f62015-02-02 00:37:00 -05002080 trace_writeback_dirty_inode_enqueue(inode);
Dave Chinnera66979a2011-03-22 22:23:41 +11002081
Tejun Heod6c10f12015-05-22 17:13:45 -04002082 /*
2083 * If this is the first dirty inode for this bdi,
2084 * we have to wake-up the corresponding bdi thread
2085 * to make sure background write-back happens
2086 * later.
2087 */
Tejun Heo07472592015-05-22 17:14:02 -04002088 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2089 wb_wakeup_delayed(wb);
Dave Chinnera66979a2011-03-22 22:23:41 +11002090 return;
Jens Axboe03ba3782009-09-09 09:08:54 +02002091 }
2092 }
Dave Chinner250df6e2011-03-22 22:23:36 +11002093out_unlock_inode:
2094 spin_unlock(&inode->i_lock);
Artem Bityutskiy253c34e2010-07-25 14:29:21 +03002095
Jens Axboe03ba3782009-09-09 09:08:54 +02002096}
2097EXPORT_SYMBOL(__mark_inode_dirty);
2098
Dave Chinnere97fedb2015-03-04 13:40:00 -05002099/*
2100 * The @s_sync_lock is used to serialise concurrent sync operations
2101 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2102 * Concurrent callers will block on the s_sync_lock rather than doing contending
2103 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2104 * has been issued up to the time this function is enter is guaranteed to be
2105 * completed by the time we have gained the lock and waited for all IO that is
2106 * in progress regardless of the order callers are granted the lock.
2107 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002108static void wait_sb_inodes(struct super_block *sb)
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002109{
Jens Axboe03ba3782009-09-09 09:08:54 +02002110 struct inode *inode, *old_inode = NULL;
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002111
Jens Axboe03ba3782009-09-09 09:08:54 +02002112 /*
2113 * We need to be protected against the filesystem going from
2114 * r/o to r/w or vice versa.
2115 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002116 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002117
Dave Chinnere97fedb2015-03-04 13:40:00 -05002118 mutex_lock(&sb->s_sync_lock);
Dave Chinner74278da2015-03-04 12:37:22 -05002119 spin_lock(&sb->s_inode_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002120
2121 /*
2122 * Data integrity sync. Must wait for all pages under writeback,
2123 * because there may have been pages dirtied before our sync
2124 * call, but which had writeout started before we write it out.
2125 * In which case, the inode may not be on the dirty list, but
2126 * we still have to wait for that writeout.
2127 */
Jens Axboeb6e51312009-09-16 15:13:54 +02002128 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
Dave Chinner250df6e2011-03-22 22:23:36 +11002129 struct address_space *mapping = inode->i_mapping;
Jens Axboe03ba3782009-09-09 09:08:54 +02002130
Dave Chinner250df6e2011-03-22 22:23:36 +11002131 spin_lock(&inode->i_lock);
2132 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2133 (mapping->nrpages == 0)) {
2134 spin_unlock(&inode->i_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002135 continue;
Dave Chinner250df6e2011-03-22 22:23:36 +11002136 }
Jens Axboe03ba3782009-09-09 09:08:54 +02002137 __iget(inode);
Dave Chinner250df6e2011-03-22 22:23:36 +11002138 spin_unlock(&inode->i_lock);
Dave Chinner74278da2015-03-04 12:37:22 -05002139 spin_unlock(&sb->s_inode_list_lock);
Dave Chinner55fa6092011-03-22 22:23:40 +11002140
Jens Axboe03ba3782009-09-09 09:08:54 +02002141 /*
Dave Chinner55fa6092011-03-22 22:23:40 +11002142 * We hold a reference to 'inode' so it couldn't have been
2143 * removed from s_inodes list while we dropped the
Dave Chinner74278da2015-03-04 12:37:22 -05002144 * s_inode_list_lock. We cannot iput the inode now as we can
Dave Chinner55fa6092011-03-22 22:23:40 +11002145 * be holding the last reference and we cannot iput it under
Dave Chinner74278da2015-03-04 12:37:22 -05002146 * s_inode_list_lock. So we keep the reference and iput it
Dave Chinner55fa6092011-03-22 22:23:40 +11002147 * later.
Jens Axboe03ba3782009-09-09 09:08:54 +02002148 */
2149 iput(old_inode);
2150 old_inode = inode;
2151
Junichi Nomuraaa750fd2015-11-05 18:47:23 -08002152 /*
2153 * We keep the error status of individual mapping so that
2154 * applications can catch the writeback error using fsync(2).
2155 * See filemap_fdatawait_keep_errors() for details.
2156 */
2157 filemap_fdatawait_keep_errors(mapping);
Jens Axboe03ba3782009-09-09 09:08:54 +02002158
2159 cond_resched();
Nick Piggin38f21972009-01-06 14:40:25 -08002160
Dave Chinner74278da2015-03-04 12:37:22 -05002161 spin_lock(&sb->s_inode_list_lock);
Jens Axboe66f3b8e2009-09-02 09:19:46 +02002162 }
Dave Chinner74278da2015-03-04 12:37:22 -05002163 spin_unlock(&sb->s_inode_list_lock);
Jens Axboe03ba3782009-09-09 09:08:54 +02002164 iput(old_inode);
Dave Chinnere97fedb2015-03-04 13:40:00 -05002165 mutex_unlock(&sb->s_sync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166}
2167
Tejun Heof30a7d02015-05-22 17:14:00 -04002168static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2169 enum wb_reason reason, bool skip_if_busy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170{
Tejun Heocc395d72015-05-22 17:13:58 -04002171 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002172 struct wb_writeback_work work = {
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002173 .sb = sb,
2174 .sync_mode = WB_SYNC_NONE,
2175 .tagged_writepages = 1,
2176 .done = &done,
2177 .nr_pages = nr,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002178 .reason = reason,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002179 };
Tejun Heoe7972912015-05-22 17:13:48 -04002180 struct backing_dev_info *bdi = sb->s_bdi;
Jens Axboe0e3c9a22010-06-01 11:08:43 +02002181
Tejun Heoe7972912015-05-22 17:13:48 -04002182 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002183 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002184 WARN_ON(!rwsem_is_locked(&sb->s_umount));
Tejun Heof30a7d02015-05-22 17:14:00 -04002185
Tejun Heodb125362015-05-22 17:14:01 -04002186 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
Tejun Heocc395d72015-05-22 17:13:58 -04002187 wb_wait_for_completion(bdi, &done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188}
Tejun Heof30a7d02015-05-22 17:14:00 -04002189
2190/**
2191 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
2192 * @sb: the superblock
2193 * @nr: the number of pages to write
2194 * @reason: reason why some writeback work initiated
2195 *
2196 * Start writeback on some inodes on this super_block. No guarantees are made
2197 * on how many (if any) will be written, and this function does not wait
2198 * for IO completion of submitted IO.
2199 */
2200void writeback_inodes_sb_nr(struct super_block *sb,
2201 unsigned long nr,
2202 enum wb_reason reason)
2203{
2204 __writeback_inodes_sb_nr(sb, nr, reason, false);
2205}
Chris Mason3259f8b2010-10-29 11:16:17 -04002206EXPORT_SYMBOL(writeback_inodes_sb_nr);
2207
2208/**
2209 * writeback_inodes_sb - writeback dirty inodes from given super_block
2210 * @sb: the superblock
Marcos Paulo de Souza786228a2011-11-23 20:56:45 +08002211 * @reason: reason why some writeback work was initiated
Chris Mason3259f8b2010-10-29 11:16:17 -04002212 *
2213 * Start writeback on some inodes on this super_block. No guarantees are made
2214 * on how many (if any) will be written, and this function does not wait
2215 * for IO completion of submitted IO.
2216 */
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002217void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002218{
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002219 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
Chris Mason3259f8b2010-10-29 11:16:17 -04002220}
Jens Axboed8a85592009-09-02 12:34:32 +02002221EXPORT_SYMBOL(writeback_inodes_sb);
2222
2223/**
Miao Xie10ee27a2013-01-10 13:47:57 +08002224 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
Chris Mason3259f8b2010-10-29 11:16:17 -04002225 * @sb: the superblock
2226 * @nr: the number of pages to write
Miao Xie10ee27a2013-01-10 13:47:57 +08002227 * @reason: the reason of writeback
Chris Mason3259f8b2010-10-29 11:16:17 -04002228 *
Miao Xie10ee27a2013-01-10 13:47:57 +08002229 * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
Chris Mason3259f8b2010-10-29 11:16:17 -04002230 * Returns 1 if writeback was started, 0 if not.
2231 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002232bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2233 enum wb_reason reason)
Chris Mason3259f8b2010-10-29 11:16:17 -04002234{
Miao Xie10ee27a2013-01-10 13:47:57 +08002235 if (!down_read_trylock(&sb->s_umount))
Tejun Heof30a7d02015-05-22 17:14:00 -04002236 return false;
Miao Xie10ee27a2013-01-10 13:47:57 +08002237
Tejun Heof30a7d02015-05-22 17:14:00 -04002238 __writeback_inodes_sb_nr(sb, nr, reason, true);
Miao Xie10ee27a2013-01-10 13:47:57 +08002239 up_read(&sb->s_umount);
Tejun Heof30a7d02015-05-22 17:14:00 -04002240 return true;
Chris Mason3259f8b2010-10-29 11:16:17 -04002241}
Miao Xie10ee27a2013-01-10 13:47:57 +08002242EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2243
2244/**
2245 * try_to_writeback_inodes_sb - try to start writeback if none underway
2246 * @sb: the superblock
2247 * @reason: reason why some writeback work was initiated
2248 *
2249 * Implement by try_to_writeback_inodes_sb_nr()
2250 * Returns 1 if writeback was started, 0 if not.
2251 */
Tejun Heof30a7d02015-05-22 17:14:00 -04002252bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
Miao Xie10ee27a2013-01-10 13:47:57 +08002253{
2254 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2255}
2256EXPORT_SYMBOL(try_to_writeback_inodes_sb);
Chris Mason3259f8b2010-10-29 11:16:17 -04002257
2258/**
Jens Axboed8a85592009-09-02 12:34:32 +02002259 * sync_inodes_sb - sync sb inode pages
Jan Kara0dc83bd2014-02-21 11:19:04 +01002260 * @sb: the superblock
Jens Axboed8a85592009-09-02 12:34:32 +02002261 *
2262 * This function writes and waits on any dirty inode belonging to this
Jan Kara0dc83bd2014-02-21 11:19:04 +01002263 * super_block.
Jens Axboed8a85592009-09-02 12:34:32 +02002264 */
Jan Kara0dc83bd2014-02-21 11:19:04 +01002265void sync_inodes_sb(struct super_block *sb)
Jens Axboed8a85592009-09-02 12:34:32 +02002266{
Tejun Heocc395d72015-05-22 17:13:58 -04002267 DEFINE_WB_COMPLETION_ONSTACK(done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002268 struct wb_writeback_work work = {
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002269 .sb = sb,
2270 .sync_mode = WB_SYNC_ALL,
2271 .nr_pages = LONG_MAX,
2272 .range_cyclic = 0,
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002273 .done = &done,
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002274 .reason = WB_REASON_SYNC,
Dave Chinner7747bd42013-07-02 22:38:35 +10002275 .for_sync = 1,
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002276 };
Tejun Heoe7972912015-05-22 17:13:48 -04002277 struct backing_dev_info *bdi = sb->s_bdi;
Christoph Hellwig3c4d7162010-06-08 18:14:43 +02002278
Tejun Heo006a0972015-08-25 14:11:52 -04002279 /*
2280 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2281 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2282 * bdi_has_dirty() need to be written out too.
2283 */
2284 if (bdi == &noop_backing_dev_info)
Jan Kara6eedc702012-07-03 16:45:27 +02002285 return;
Christoph Hellwigcf37e972010-06-08 18:14:51 +02002286 WARN_ON(!rwsem_is_locked(&sb->s_umount));
2287
Tejun Heodb125362015-05-22 17:14:01 -04002288 bdi_split_work_to_wbs(bdi, &work, false);
Tejun Heocc395d72015-05-22 17:13:58 -04002289 wb_wait_for_completion(bdi, &done);
Christoph Hellwig83ba7b02010-07-06 08:59:53 +02002290
Jens Axboeb6e51312009-09-16 15:13:54 +02002291 wait_sb_inodes(sb);
Jens Axboed8a85592009-09-02 12:34:32 +02002292}
2293EXPORT_SYMBOL(sync_inodes_sb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295/**
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002296 * write_inode_now - write an inode to disk
2297 * @inode: inode to write to disk
2298 * @sync: whether the write should be synchronous or not
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 *
Andrea Arcangeli7f04c262005-10-30 15:03:05 -08002300 * This function commits an inode to disk immediately if it is dirty. This is
2301 * primarily needed by knfsd.
2302 *
2303 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305int write_inode_now(struct inode *inode, int sync)
2306{
Christoph Hellwigf758eea2011-04-21 18:19:44 -06002307 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 struct writeback_control wbc = {
2309 .nr_to_write = LONG_MAX,
Mike Galbraith18914b12008-02-08 04:20:23 -08002310 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -07002311 .range_start = 0,
2312 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 };
2314
2315 if (!mapping_cap_writeback_dirty(inode->i_mapping))
Andrew Morton49364ce2005-11-07 00:59:15 -08002316 wbc.nr_to_write = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317
2318 might_sleep();
Jan Kara4f8ad652012-05-03 14:48:00 +02002319 return writeback_single_inode(inode, wb, &wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320}
2321EXPORT_SYMBOL(write_inode_now);
2322
2323/**
2324 * sync_inode - write an inode and its pages to disk.
2325 * @inode: the inode to sync
2326 * @wbc: controls the writeback mode
2327 *
2328 * sync_inode() will write an inode and its pages to disk. It will also
2329 * correctly update the inode on its superblock's dirty inode lists and will
2330 * update inode->i_state.
2331 *
2332 * The caller must have a ref on the inode.
2333 */
2334int sync_inode(struct inode *inode, struct writeback_control *wbc)
2335{
Jan Kara4f8ad652012-05-03 14:48:00 +02002336 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337}
2338EXPORT_SYMBOL(sync_inode);
Christoph Hellwigc37650162010-10-06 10:48:20 +02002339
2340/**
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002341 * sync_inode_metadata - write an inode to disk
Christoph Hellwigc37650162010-10-06 10:48:20 +02002342 * @inode: the inode to sync
2343 * @wait: wait for I/O to complete.
2344 *
Andrew Mortonc691b9d2011-01-13 15:45:48 -08002345 * Write an inode to disk and adjust its dirty state after completion.
Christoph Hellwigc37650162010-10-06 10:48:20 +02002346 *
2347 * Note: only writes the actual inode, no associated data or other metadata.
2348 */
2349int sync_inode_metadata(struct inode *inode, int wait)
2350{
2351 struct writeback_control wbc = {
2352 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2353 .nr_to_write = 0, /* metadata-only */
2354 };
2355
2356 return sync_inode(inode, &wbc);
2357}
2358EXPORT_SYMBOL(sync_inode_metadata);