blob: 5f3e1b46ace5820b24261f261ee8a6cd5e6155cc [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
Francois Camie1f8e872008-10-15 22:01:59 -070010 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
David Howellscf9a2ae2006-08-29 19:05:54 +010035#include <linux/buffer_head.h>
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100037#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038
39/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060040 * Estimate write bandwidth at 200ms intervals.
41 */
42#define BANDWIDTH_INTERVAL max(HZ/5, 1)
43
44/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070045 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
46 * will look to see if it needs to force writeback or throttling.
47 */
48static long ratelimit_pages = 32;
49
Linus Torvalds1da177e2005-04-16 15:20:36 -070050/*
51 * When balance_dirty_pages decides that the caller needs to perform some
52 * non-background writeback, this is how many pages it will attempt to write.
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080053 * It should be somewhat larger than dirtied pages to ensure that reasonably
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * large amounts of I/O are submitted.
55 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080056static inline long sync_writeback_pages(unsigned long dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -070057{
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080058 if (dirtied < ratelimit_pages)
59 dirtied = ratelimit_pages;
60
61 return dirtied + dirtied / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -070062}
63
64/* The following parameters are exported via /proc/sys/vm */
65
66/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020067 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070068 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080069int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070070
71/*
David Rientjes2da02992009-01-06 14:39:31 -080072 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
73 * dirty_background_ratio * the amount of dirtyable memory
74 */
75unsigned long dirty_background_bytes;
76
77/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080078 * free highmem will not be subtracted from the total free memory
79 * for calculating free ratios if vm_highmem_is_dirtyable is true
80 */
81int vm_highmem_is_dirtyable;
82
83/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 * The generator of dirty data starts writeback at this percentage
85 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080086int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88/*
David Rientjes2da02992009-01-06 14:39:31 -080089 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
90 * vm_dirty_ratio * the amount of dirtyable memory
91 */
92unsigned long vm_dirty_bytes;
93
94/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070095 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070097unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
99/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700100 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700102unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
104/*
105 * Flag that makes the machine dump writes/reads and block dirtyings.
106 */
107int block_dump;
108
109/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800110 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
111 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112 */
113int laptop_mode;
114
115EXPORT_SYMBOL(laptop_mode);
116
117/* End of sysctl-exported parameters */
118
119
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700121 * Scale the writeback cache size proportional to the relative writeout speeds.
122 *
123 * We do this by keeping a floating proportion between BDIs, based on page
124 * writeback completions [end_page_writeback()]. Those devices that write out
125 * pages fastest will get the larger share, while the slower will get a smaller
126 * share.
127 *
128 * We use page writeout completions because we are interested in getting rid of
129 * dirty pages. Having them written out is the primary goal.
130 *
131 * We introduce a concept of time, a period over which we measure these events,
132 * because demand can/will vary over time. The length of this period itself is
133 * measured in page writeback completions.
134 *
135 */
136static struct prop_descriptor vm_completions;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700137static struct prop_descriptor vm_dirties;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700138
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700139/*
140 * couple the period to the dirty_ratio:
141 *
142 * period/2 ~ roundup_pow_of_two(dirty limit)
143 */
144static int calc_period_shift(void)
145{
146 unsigned long dirty_total;
147
David Rientjes2da02992009-01-06 14:39:31 -0800148 if (vm_dirty_bytes)
149 dirty_total = vm_dirty_bytes / PAGE_SIZE;
150 else
151 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
152 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700153 return 2 + ilog2(dirty_total - 1);
154}
155
156/*
David Rientjes2da02992009-01-06 14:39:31 -0800157 * update the period when the dirty threshold changes.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700158 */
David Rientjes2da02992009-01-06 14:39:31 -0800159static void update_completion_period(void)
160{
161 int shift = calc_period_shift();
162 prop_change_shift(&vm_completions, shift);
163 prop_change_shift(&vm_dirties, shift);
164}
165
166int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700167 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800168 loff_t *ppos)
169{
170 int ret;
171
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700172 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800173 if (ret == 0 && write)
174 dirty_background_bytes = 0;
175 return ret;
176}
177
178int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700179 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800180 loff_t *ppos)
181{
182 int ret;
183
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700184 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800185 if (ret == 0 && write)
186 dirty_background_ratio = 0;
187 return ret;
188}
189
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700190int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700191 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700192 loff_t *ppos)
193{
194 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800195 int ret;
196
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700197 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700198 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
David Rientjes2da02992009-01-06 14:39:31 -0800199 update_completion_period();
200 vm_dirty_bytes = 0;
201 }
202 return ret;
203}
204
205
206int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700207 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800208 loff_t *ppos)
209{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800210 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800211 int ret;
212
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700213 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800214 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
215 update_completion_period();
216 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700217 }
218 return ret;
219}
220
221/*
222 * Increment the BDI's writeout completion count and the global writeout
223 * completion count. Called from test_clear_page_writeback().
224 */
225static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
226{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600227 __inc_bdi_stat(bdi, BDI_WRITTEN);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700228 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
229 bdi->max_prop_frac);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700230}
231
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700232void bdi_writeout_inc(struct backing_dev_info *bdi)
233{
234 unsigned long flags;
235
236 local_irq_save(flags);
237 __bdi_writeout_inc(bdi);
238 local_irq_restore(flags);
239}
240EXPORT_SYMBOL_GPL(bdi_writeout_inc);
241
Nick Piggin1cf6e7d2009-02-18 14:48:18 -0800242void task_dirty_inc(struct task_struct *tsk)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700243{
244 prop_inc_single(&vm_dirties, &tsk->dirties);
245}
246
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700247/*
248 * Obtain an accurate fraction of the BDI's portion.
249 */
250static void bdi_writeout_fraction(struct backing_dev_info *bdi,
251 long *numerator, long *denominator)
252{
Wu Fengguang3efaf0f2010-12-16 22:22:00 -0600253 prop_fraction_percpu(&vm_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700254 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700255}
256
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700257static inline void task_dirties_fraction(struct task_struct *tsk,
258 long *numerator, long *denominator)
259{
260 prop_fraction_single(&vm_dirties, &tsk->dirties,
261 numerator, denominator);
262}
263
264/*
Wu Fengguang1babe182010-08-11 14:17:40 -0700265 * task_dirty_limit - scale down dirty throttling threshold for one task
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700266 *
267 * task specific dirty limit:
268 *
269 * dirty -= (dirty/8) * p_{t}
Wu Fengguang1babe182010-08-11 14:17:40 -0700270 *
271 * To protect light/slow dirtying tasks from heavier/fast ones, we start
272 * throttling individual tasks before reaching the bdi dirty limit.
273 * Relatively low thresholds will be allocated to heavy dirtiers. So when
274 * dirty pages grow large, heavy dirtiers will be throttled first, which will
275 * effectively curb the growth of dirty pages. Light dirtiers with high enough
276 * dirty threshold may never get throttled.
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700277 */
Wu Fengguang16c40422010-08-11 14:17:39 -0700278static unsigned long task_dirty_limit(struct task_struct *tsk,
279 unsigned long bdi_dirty)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700280{
281 long numerator, denominator;
Wu Fengguang16c40422010-08-11 14:17:39 -0700282 unsigned long dirty = bdi_dirty;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700283 u64 inv = dirty >> 3;
284
285 task_dirties_fraction(tsk, &numerator, &denominator);
286 inv *= numerator;
287 do_div(inv, denominator);
288
289 dirty -= inv;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700290
Wu Fengguang16c40422010-08-11 14:17:39 -0700291 return max(dirty, bdi_dirty/2);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700292}
293
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700294/*
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700295 *
296 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700297static unsigned int bdi_min_ratio;
298
299int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
300{
301 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700302
Jens Axboecfc4ba52009-09-14 13:12:40 +0200303 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700304 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700305 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700306 } else {
307 min_ratio -= bdi->min_ratio;
308 if (bdi_min_ratio + min_ratio < 100) {
309 bdi_min_ratio += min_ratio;
310 bdi->min_ratio += min_ratio;
311 } else {
312 ret = -EINVAL;
313 }
314 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200315 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700316
317 return ret;
318}
319
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700320int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
321{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700322 int ret = 0;
323
324 if (max_ratio > 100)
325 return -EINVAL;
326
Jens Axboecfc4ba52009-09-14 13:12:40 +0200327 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700328 if (bdi->min_ratio > max_ratio) {
329 ret = -EINVAL;
330 } else {
331 bdi->max_ratio = max_ratio;
332 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
333 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200334 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700335
336 return ret;
337}
338EXPORT_SYMBOL(bdi_set_max_ratio);
339
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700340/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 * Work out the current dirty-memory clamping and background writeout
342 * thresholds.
343 *
344 * The main aim here is to lower them aggressively if there is a lot of mapped
345 * memory around. To avoid stressing page reclaim with lots of unreclaimable
346 * pages. It is better to clamp down on writers than to start swapping, and
347 * performing lots of scanning.
348 *
349 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
350 *
351 * We don't permit the clamping level to fall below 5% - that is getting rather
352 * excessive.
353 *
354 * We make sure that the background writeout level is below the adjusted
355 * clamping level.
356 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700357
358static unsigned long highmem_dirtyable_memory(unsigned long total)
359{
360#ifdef CONFIG_HIGHMEM
361 int node;
362 unsigned long x = 0;
363
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700364 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700365 struct zone *z =
366 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
367
Wu Fengguangadea02a2009-09-21 17:01:42 -0700368 x += zone_page_state(z, NR_FREE_PAGES) +
369 zone_reclaimable_pages(z);
Christoph Lameter1b424462007-05-06 14:48:59 -0700370 }
371 /*
372 * Make sure that the number of highmem pages is never larger
373 * than the number of the total dirtyable memory. This can only
374 * occur in very strange VM situations but we want to make sure
375 * that this does not occur.
376 */
377 return min(x, total);
378#else
379 return 0;
380#endif
381}
382
Steven Rostedt3eefae92008-05-12 21:21:04 +0200383/**
384 * determine_dirtyable_memory - amount of memory that may be used
385 *
386 * Returns the numebr of pages that can currently be freed and used
387 * by the kernel for direct mappings.
388 */
389unsigned long determine_dirtyable_memory(void)
Christoph Lameter1b424462007-05-06 14:48:59 -0700390{
391 unsigned long x;
392
Wu Fengguangadea02a2009-09-21 17:01:42 -0700393 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
Bron Gondwana195cf4532008-02-04 22:29:20 -0800394
395 if (!vm_highmem_is_dirtyable)
396 x -= highmem_dirtyable_memory(x);
397
Christoph Lameter1b424462007-05-06 14:48:59 -0700398 return x + 1; /* Ensure that we never return 0 */
399}
400
Randy Dunlap03ab4502010-08-14 13:05:17 -0700401/*
Wu Fengguang1babe182010-08-11 14:17:40 -0700402 * global_dirty_limits - background-writeback and dirty-throttling thresholds
403 *
404 * Calculate the dirty thresholds based on sysctl parameters
405 * - vm.dirty_background_ratio or vm.dirty_background_bytes
406 * - vm.dirty_ratio or vm.dirty_bytes
407 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Minchan Kimebd13732011-01-04 01:36:48 +0900408 * real-time tasks.
Wu Fengguang1babe182010-08-11 14:17:40 -0700409 */
Wu Fengguang16c40422010-08-11 14:17:39 -0700410void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411{
David Rientjes364aeb22009-01-06 14:39:29 -0800412 unsigned long background;
413 unsigned long dirty;
Minchan Kim240c8792011-01-13 15:46:27 -0800414 unsigned long uninitialized_var(available_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 struct task_struct *tsk;
416
Minchan Kim240c8792011-01-13 15:46:27 -0800417 if (!vm_dirty_bytes || !dirty_background_bytes)
418 available_memory = determine_dirtyable_memory();
419
David Rientjes2da02992009-01-06 14:39:31 -0800420 if (vm_dirty_bytes)
421 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
Wu Fengguang4cbec4c2010-10-26 14:21:45 -0700422 else
423 dirty = (vm_dirty_ratio * available_memory) / 100;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424
David Rientjes2da02992009-01-06 14:39:31 -0800425 if (dirty_background_bytes)
426 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
427 else
428 background = (dirty_background_ratio * available_memory) / 100;
429
430 if (background >= dirty)
431 background = dirty / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 tsk = current;
433 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
434 background += background / 4;
435 dirty += dirty / 4;
436 }
437 *pbackground = background;
438 *pdirty = dirty;
Wu Fengguang16c40422010-08-11 14:17:39 -0700439}
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700440
Wu Fengguang6f718652011-03-02 17:14:34 -0600441/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700442 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600443 * @bdi: the backing_dev_info to query
444 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700445 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600446 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
447 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
448 * And the "limit" in the name is not seriously taken as hard limit in
449 * balance_dirty_pages().
450 *
451 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700452 * - starving fast devices
453 * - piling up dirty pages (that will take long time to sync) on slow devices
454 *
455 * The bdi's share of dirty limit will be adapting to its throughput and
456 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
457 */
458unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700459{
460 u64 bdi_dirty;
461 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700462
Wu Fengguang16c40422010-08-11 14:17:39 -0700463 /*
464 * Calculate this BDI's share of the dirty ratio.
465 */
466 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700467
Wu Fengguang16c40422010-08-11 14:17:39 -0700468 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
469 bdi_dirty *= numerator;
470 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700471
Wu Fengguang16c40422010-08-11 14:17:39 -0700472 bdi_dirty += (dirty * bdi->min_ratio) / 100;
473 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
474 bdi_dirty = dirty * bdi->max_ratio / 100;
475
476 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477}
478
Wu Fengguange98be2d2010-08-29 11:22:30 -0600479static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
480 unsigned long elapsed,
481 unsigned long written)
482{
483 const unsigned long period = roundup_pow_of_two(3 * HZ);
484 unsigned long avg = bdi->avg_write_bandwidth;
485 unsigned long old = bdi->write_bandwidth;
486 u64 bw;
487
488 /*
489 * bw = written * HZ / elapsed
490 *
491 * bw * elapsed + write_bandwidth * (period - elapsed)
492 * write_bandwidth = ---------------------------------------------------
493 * period
494 */
495 bw = written - bdi->written_stamp;
496 bw *= HZ;
497 if (unlikely(elapsed > period)) {
498 do_div(bw, elapsed);
499 avg = bw;
500 goto out;
501 }
502 bw += (u64)bdi->write_bandwidth * (period - elapsed);
503 bw >>= ilog2(period);
504
505 /*
506 * one more level of smoothing, for filtering out sudden spikes
507 */
508 if (avg > old && old >= (unsigned long)bw)
509 avg -= (avg - old) >> 3;
510
511 if (avg < old && old <= (unsigned long)bw)
512 avg += (old - avg) >> 3;
513
514out:
515 bdi->write_bandwidth = bw;
516 bdi->avg_write_bandwidth = avg;
517}
518
519void __bdi_update_bandwidth(struct backing_dev_info *bdi,
520 unsigned long start_time)
521{
522 unsigned long now = jiffies;
523 unsigned long elapsed = now - bdi->bw_time_stamp;
524 unsigned long written;
525
526 /*
527 * rate-limit, only update once every 200ms.
528 */
529 if (elapsed < BANDWIDTH_INTERVAL)
530 return;
531
532 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
533
534 /*
535 * Skip quiet periods when disk bandwidth is under-utilized.
536 * (at least 1s idle time between two flusher runs)
537 */
538 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
539 goto snapshot;
540
541 bdi_update_write_bandwidth(bdi, elapsed, written);
542
543snapshot:
544 bdi->written_stamp = written;
545 bdi->bw_time_stamp = now;
546}
547
548static void bdi_update_bandwidth(struct backing_dev_info *bdi,
549 unsigned long start_time)
550{
551 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
552 return;
553 spin_lock(&bdi->wb.list_lock);
554 __bdi_update_bandwidth(bdi, start_time);
555 spin_unlock(&bdi->wb.list_lock);
556}
557
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558/*
559 * balance_dirty_pages() must be called by processes which are generating dirty
560 * data. It looks at the number of dirty pages in the machine and will force
561 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
Jens Axboe5b0830c2009-09-23 19:37:09 +0200562 * If we're over `background_thresh' then the writeback threads are woken to
563 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800565static void balance_dirty_pages(struct address_space *mapping,
566 unsigned long write_chunk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
Wu Fengguang77627412010-09-12 13:34:05 -0600568 unsigned long nr_reclaimable, bdi_nr_reclaimable;
569 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
570 unsigned long bdi_dirty;
David Rientjes364aeb22009-01-06 14:39:29 -0800571 unsigned long background_thresh;
572 unsigned long dirty_thresh;
573 unsigned long bdi_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 unsigned long pages_written = 0;
Jens Axboe87c6a9b2009-09-17 19:59:14 +0200575 unsigned long pause = 1;
Wu Fengguange50e3722010-08-11 14:17:37 -0700576 bool dirty_exceeded = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600578 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579
580 for (;;) {
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800581 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
582 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -0600583 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800584
Wu Fengguang16c40422010-08-11 14:17:39 -0700585 global_dirty_limits(&background_thresh, &dirty_thresh);
586
587 /*
588 * Throttle it only when the background writeback cannot
589 * catch-up. This avoids (excessively) small writeouts
590 * when the bdi limits are ramping up.
591 */
Wu Fengguang77627412010-09-12 13:34:05 -0600592 if (nr_dirty <= (background_thresh + dirty_thresh) / 2)
Wu Fengguang16c40422010-08-11 14:17:39 -0700593 break;
594
595 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
596 bdi_thresh = task_dirty_limit(current, bdi_thresh);
597
Wu Fengguange50e3722010-08-11 14:17:37 -0700598 /*
599 * In order to avoid the stacked BDI deadlock we need
600 * to ensure we accurately count the 'dirty' pages when
601 * the threshold is low.
602 *
603 * Otherwise it would be possible to get thresh+n pages
604 * reported dirty, even though there are thresh-m pages
605 * actually dirty; with m+n sitting in the percpu
606 * deltas.
607 */
608 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
609 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
Wu Fengguang77627412010-09-12 13:34:05 -0600610 bdi_dirty = bdi_nr_reclaimable +
611 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -0700612 } else {
613 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
Wu Fengguang77627412010-09-12 13:34:05 -0600614 bdi_dirty = bdi_nr_reclaimable +
615 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -0700616 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800617
Wu Fengguange50e3722010-08-11 14:17:37 -0700618 /*
619 * The bdi thresh is somehow "soft" limit derived from the
620 * global "hard" limit. The former helps to prevent heavy IO
621 * bdi or process from holding back light ones; The latter is
622 * the last resort safeguard.
623 */
Wu Fengguang77627412010-09-12 13:34:05 -0600624 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
625 (nr_dirty > dirty_thresh);
Wu Fengguange50e3722010-08-11 14:17:37 -0700626
627 if (!dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700628 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700630 if (!bdi->dirty_exceeded)
631 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
Wu Fengguange98be2d2010-08-29 11:22:30 -0600633 bdi_update_bandwidth(bdi, start_time);
634
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
636 * Unstable writes are a feature of certain networked
637 * filesystems (i.e. NFS) in which data may have been
638 * written to the server's write cache, but has not yet
639 * been flushed to permanent storage.
Richard Kennedyd7831a02009-06-30 11:41:35 -0700640 * Only move pages to writeback if this bdi is over its
641 * threshold otherwise wait until the disk writes catch
642 * up.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600644 trace_balance_dirty_start(bdi);
Richard Kennedyd7831a02009-06-30 11:41:35 -0700645 if (bdi_nr_reclaimable > bdi_thresh) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600646 pages_written += writeback_inodes_wb(&bdi->wb,
647 write_chunk);
648 trace_balance_dirty_written(bdi, pages_written);
Wu Fengguange50e3722010-08-11 14:17:37 -0700649 if (pages_written >= write_chunk)
650 break; /* We've done our duty */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 }
Wu Fengguangd153ba62010-12-21 17:24:21 -0800652 __set_current_state(TASK_UNINTERRUPTIBLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +0200653 io_schedule_timeout(pause);
Wu Fengguangd46db3d2011-05-04 19:54:37 -0600654 trace_balance_dirty_wait(bdi);
Jens Axboe87c6a9b2009-09-17 19:59:14 +0200655
656 /*
657 * Increase the delay for each loop, up to our previous
658 * default of taking a 100ms nap.
659 */
660 pause <<= 1;
661 if (pause > HZ / 10)
662 pause = HZ / 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 }
664
Wu Fengguange50e3722010-08-11 14:17:37 -0700665 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700666 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
668 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +0200669 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670
671 /*
672 * In laptop mode, we wait until hitting the higher threshold before
673 * starting background writeout, and then write out all the way down
674 * to the lower threshold. So slow writers cause minimal disk activity.
675 *
676 * In normal mode, we start background writeout at the lower
677 * background_thresh, to keep the amount of dirty memory low.
678 */
679 if ((laptop_mode && pages_written) ||
Wu Fengguange50e3722010-08-11 14:17:37 -0700680 (!laptop_mode && (nr_reclaimable > background_thresh)))
Christoph Hellwigc5444192010-06-08 18:15:15 +0200681 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682}
683
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200684void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700685{
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200686 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700687 struct address_space *mapping = page_mapping(page);
688
689 if (mapping)
690 balance_dirty_pages_ratelimited(mapping);
691 }
692}
693
Tejun Heo245b2e72009-06-24 15:13:48 +0900694static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
695
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800697 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -0700698 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +0200699 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 *
701 * Processes which are dirtying memory should call in here once for each page
702 * which was newly dirtied. The function will periodically check the system's
703 * dirty state and will initiate writeback if needed.
704 *
705 * On really big machines, get_writeback_state is expensive, so try to avoid
706 * calling it too often (ratelimiting). But once we're over the dirty memory
707 * limit we decrease the ratelimiting by a lot, to prevent individual processes
708 * from overshooting the limit by (ratelimit_pages) each.
709 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800710void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
711 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712{
Wu Fengguang36715ce2011-06-11 17:53:57 -0600713 struct backing_dev_info *bdi = mapping->backing_dev_info;
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800714 unsigned long ratelimit;
715 unsigned long *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716
Wu Fengguang36715ce2011-06-11 17:53:57 -0600717 if (!bdi_cap_account_dirty(bdi))
718 return;
719
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 ratelimit = ratelimit_pages;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700721 if (mapping->backing_dev_info->dirty_exceeded)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722 ratelimit = 8;
723
724 /*
725 * Check the rate limiting. Also, we do not want to throttle real-time
726 * tasks in balance_dirty_pages(). Period.
727 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800728 preempt_disable();
Tejun Heo245b2e72009-06-24 15:13:48 +0900729 p = &__get_cpu_var(bdp_ratelimits);
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800730 *p += nr_pages_dirtied;
731 if (unlikely(*p >= ratelimit)) {
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800732 ratelimit = sync_writeback_pages(*p);
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800733 *p = 0;
734 preempt_enable();
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800735 balance_dirty_pages(mapping, ratelimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 return;
737 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800738 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739}
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800740EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741
Andrew Morton232ea4d2007-02-28 20:13:21 -0800742void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743{
David Rientjes364aeb22009-01-06 14:39:29 -0800744 unsigned long background_thresh;
745 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746
747 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -0700748 global_dirty_limits(&background_thresh, &dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749
750 /*
751 * Boost the allowable dirty threshold a bit for page
752 * allocators so they don't get DoS'ed by heavy writers
753 */
754 dirty_thresh += dirty_thresh / 10; /* wheeee... */
755
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700756 if (global_page_state(NR_UNSTABLE_NFS) +
757 global_page_state(NR_WRITEBACK) <= dirty_thresh)
758 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +0200759 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -0700760
761 /*
762 * The caller might hold locks which can prevent IO completion
763 * or progress in the filesystem. So we cannot just sit here
764 * waiting for IO to complete.
765 */
766 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
767 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 }
769}
770
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
773 */
774int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700775 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700777 proc_dointvec(table, write, buffer, length, ppos);
Jens Axboe64231042010-05-21 20:00:35 +0200778 bdi_arm_supers_timer();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 return 0;
780}
781
Jens Axboec2c49862010-05-20 09:18:47 +0200782#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +0200783void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784{
Matthew Garrett31373d02010-04-06 14:25:14 +0200785 struct request_queue *q = (struct request_queue *)data;
786 int nr_pages = global_page_state(NR_FILE_DIRTY) +
787 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
Matthew Garrett31373d02010-04-06 14:25:14 +0200789 /*
790 * We want to write everything out, not just down to the dirty
791 * threshold
792 */
Matthew Garrett31373d02010-04-06 14:25:14 +0200793 if (bdi_has_dirty_io(&q->backing_dev_info))
Christoph Hellwigc5444192010-06-08 18:15:15 +0200794 bdi_start_writeback(&q->backing_dev_info, nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795}
796
797/*
798 * We've spun up the disk and we're in laptop mode: schedule writeback
799 * of all dirty data a few seconds from now. If the flush is already scheduled
800 * then push it back - the user is still using the disk.
801 */
Matthew Garrett31373d02010-04-06 14:25:14 +0200802void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803{
Matthew Garrett31373d02010-04-06 14:25:14 +0200804 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805}
806
807/*
808 * We're in laptop mode and we've just synced. The sync's writes will have
809 * caused another writeback to be scheduled by laptop_io_completion.
810 * Nothing needs to be written back anymore, so we unschedule the writeback.
811 */
812void laptop_sync_completion(void)
813{
Matthew Garrett31373d02010-04-06 14:25:14 +0200814 struct backing_dev_info *bdi;
815
816 rcu_read_lock();
817
818 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
819 del_timer(&bdi->laptop_mode_wb_timer);
820
821 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822}
Jens Axboec2c49862010-05-20 09:18:47 +0200823#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824
825/*
826 * If ratelimit_pages is too high then we can get into dirty-data overload
827 * if a large number of processes all perform writes at the same time.
828 * If it is too low then SMP machines will call the (expensive)
829 * get_writeback_state too often.
830 *
831 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
832 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
833 * thresholds before writeback cuts in.
834 *
835 * But the limit should not be set too high. Because it also controls the
836 * amount of memory which the balance_dirty_pages() caller has to write back.
837 * If this is too large then the caller will block on the IO queue all the
838 * time. So limit it to four megabytes - the balance_dirty_pages() caller
839 * will write six megabyte chunks, max.
840 */
841
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700842void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843{
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700844 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 if (ratelimit_pages < 16)
846 ratelimit_pages = 16;
847 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
848 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
849}
850
Chandra Seetharaman26c21432006-06-27 02:54:10 -0700851static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
853{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700854 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -0800855 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856}
857
Chandra Seetharaman74b85f32006-06-27 02:54:09 -0700858static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 .notifier_call = ratelimit_handler,
860 .next = NULL,
861};
862
863/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800864 * Called early on to tune the page writeback dirty limits.
865 *
866 * We used to scale dirty pages according to how total memory
867 * related to pages that could be allocated for buffers (by
868 * comparing nr_free_buffer_pages() to vm_total_pages.
869 *
870 * However, that was when we used "dirty_ratio" to scale with
871 * all memory, and we don't do that any more. "dirty_ratio"
872 * is now applied to total non-HIGHPAGE memory (by subtracting
873 * totalhigh_pages from vm_total_pages), and as such we can't
874 * get into the old insane situation any more where we had
875 * large amounts of dirty pages compared to a small amount of
876 * non-HIGHMEM memory.
877 *
878 * But we might still want to scale the dirty_ratio by how
879 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 */
881void __init page_writeback_init(void)
882{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700883 int shift;
884
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700885 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700887
888 shift = calc_period_shift();
889 prop_descriptor_init(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700890 prop_descriptor_init(&vm_dirties, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891}
892
David Howells811d7362006-08-29 19:06:09 +0100893/**
Jan Karaf446daae2010-08-09 17:19:12 -0700894 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
895 * @mapping: address space structure to write
896 * @start: starting page index
897 * @end: ending page index (inclusive)
898 *
899 * This function scans the page range from @start to @end (inclusive) and tags
900 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
901 * that write_cache_pages (or whoever calls this function) will then use
902 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
903 * used to avoid livelocking of writeback by a process steadily creating new
904 * dirty pages in the file (thus it is important for this function to be quick
905 * so that it can tag pages faster than a dirtying process can create them).
906 */
907/*
908 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
909 */
Jan Karaf446daae2010-08-09 17:19:12 -0700910void tag_pages_for_writeback(struct address_space *mapping,
911 pgoff_t start, pgoff_t end)
912{
Randy Dunlap3c111a02010-08-11 14:17:30 -0700913#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -0700914 unsigned long tagged;
915
916 do {
917 spin_lock_irq(&mapping->tree_lock);
918 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
919 &start, end, WRITEBACK_TAG_BATCH,
920 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
921 spin_unlock_irq(&mapping->tree_lock);
922 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
923 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -0700924 /* We check 'start' to handle wrapping when end == ~0UL */
925 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -0700926}
927EXPORT_SYMBOL(tag_pages_for_writeback);
928
929/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700930 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
David Howells811d7362006-08-29 19:06:09 +0100931 * @mapping: address space structure to write
932 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700933 * @writepage: function called for each page
934 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +0100935 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700936 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +0100937 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
938 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
939 * and msync() need to guarantee that all the data which was dirty at the time
940 * the call was made get new I/O started against them. If wbc->sync_mode is
941 * WB_SYNC_ALL then we were called for data integrity and we must wait for
942 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -0700943 *
944 * To avoid livelocks (when other process dirties new pages), we first tag
945 * pages which should be written back with TOWRITE tag and only then start
946 * writing them. For data-integrity sync we have to be careful so that we do
947 * not miss some pages (e.g., because some other process has cleared TOWRITE
948 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
949 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +0100950 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700951int write_cache_pages(struct address_space *mapping,
952 struct writeback_control *wbc, writepage_t writepage,
953 void *data)
David Howells811d7362006-08-29 19:06:09 +0100954{
David Howells811d7362006-08-29 19:06:09 +0100955 int ret = 0;
956 int done = 0;
David Howells811d7362006-08-29 19:06:09 +0100957 struct pagevec pvec;
958 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -0800959 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +0100960 pgoff_t index;
961 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -0800962 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -0800963 int cycled;
David Howells811d7362006-08-29 19:06:09 +0100964 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -0700965 int tag;
David Howells811d7362006-08-29 19:06:09 +0100966
David Howells811d7362006-08-29 19:06:09 +0100967 pagevec_init(&pvec, 0);
968 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -0800969 writeback_index = mapping->writeback_index; /* prev offset */
970 index = writeback_index;
971 if (index == 0)
972 cycled = 1;
973 else
974 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +0100975 end = -1;
976 } else {
977 index = wbc->range_start >> PAGE_CACHE_SHIFT;
978 end = wbc->range_end >> PAGE_CACHE_SHIFT;
979 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
980 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -0800981 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +0100982 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -0600983 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -0700984 tag = PAGECACHE_TAG_TOWRITE;
985 else
986 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +0100987retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -0600988 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -0700989 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -0800990 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800991 while (!done && (index <= end)) {
992 int i;
993
Jan Karaf446daae2010-08-09 17:19:12 -0700994 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800995 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
996 if (nr_pages == 0)
997 break;
David Howells811d7362006-08-29 19:06:09 +0100998
David Howells811d7362006-08-29 19:06:09 +0100999 for (i = 0; i < nr_pages; i++) {
1000 struct page *page = pvec.pages[i];
1001
Nick Piggind5482cd2009-01-06 14:39:11 -08001002 /*
1003 * At this point, the page may be truncated or
1004 * invalidated (changing page->mapping to NULL), or
1005 * even swizzled back from swapper_space to tmpfs file
1006 * mapping. However, page->index will not change
1007 * because we have a reference on the page.
1008 */
1009 if (page->index > end) {
1010 /*
1011 * can't be range_cyclic (1st pass) because
1012 * end == -1 in that case.
1013 */
1014 done = 1;
1015 break;
1016 }
1017
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001018 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001019
David Howells811d7362006-08-29 19:06:09 +01001020 lock_page(page);
1021
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001022 /*
1023 * Page truncated or invalidated. We can freely skip it
1024 * then, even for data integrity operations: the page
1025 * has disappeared concurrently, so there could be no
1026 * real expectation of this data interity operation
1027 * even if there is now a new, dirty page at the same
1028 * pagecache address.
1029 */
David Howells811d7362006-08-29 19:06:09 +01001030 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001031continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001032 unlock_page(page);
1033 continue;
1034 }
1035
Nick Piggin515f4a02009-01-06 14:39:10 -08001036 if (!PageDirty(page)) {
1037 /* someone wrote it for us */
1038 goto continue_unlock;
1039 }
David Howells811d7362006-08-29 19:06:09 +01001040
Nick Piggin515f4a02009-01-06 14:39:10 -08001041 if (PageWriteback(page)) {
1042 if (wbc->sync_mode != WB_SYNC_NONE)
1043 wait_on_page_writeback(page);
1044 else
1045 goto continue_unlock;
1046 }
1047
1048 BUG_ON(PageWriteback(page));
1049 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001050 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001051
Dave Chinner9e094382010-07-07 13:24:08 +10001052 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001053 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001054 if (unlikely(ret)) {
1055 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1056 unlock_page(page);
1057 ret = 0;
1058 } else {
1059 /*
1060 * done_index is set past this page,
1061 * so media errors will not choke
1062 * background writeout for the entire
1063 * file. This has consequences for
1064 * range_cyclic semantics (ie. it may
1065 * not be suitable for data integrity
1066 * writeout).
1067 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001068 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001069 done = 1;
1070 break;
1071 }
Dave Chinner0b564922010-06-09 10:37:18 +10001072 }
David Howells811d7362006-08-29 19:06:09 +01001073
Dave Chinner546a1922010-08-24 11:44:34 +10001074 /*
1075 * We stop writing back only if we are not doing
1076 * integrity sync. In case of integrity sync we have to
1077 * keep going until we have written all the pages
1078 * we tagged for writeback prior to entering this loop.
1079 */
1080 if (--wbc->nr_to_write <= 0 &&
1081 wbc->sync_mode == WB_SYNC_NONE) {
1082 done = 1;
1083 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001084 }
David Howells811d7362006-08-29 19:06:09 +01001085 }
1086 pagevec_release(&pvec);
1087 cond_resched();
1088 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001089 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001090 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001091 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001092 * We hit the last page and there is more work to be done: wrap
1093 * back to the start of the file
1094 */
Nick Piggin31a12662009-01-06 14:39:04 -08001095 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001096 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001097 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001098 goto retry;
1099 }
Dave Chinner0b564922010-06-09 10:37:18 +10001100 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1101 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001102
David Howells811d7362006-08-29 19:06:09 +01001103 return ret;
1104}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001105EXPORT_SYMBOL(write_cache_pages);
1106
1107/*
1108 * Function used by generic_writepages to call the real writepage
1109 * function and set the mapping flags on error
1110 */
1111static int __writepage(struct page *page, struct writeback_control *wbc,
1112 void *data)
1113{
1114 struct address_space *mapping = data;
1115 int ret = mapping->a_ops->writepage(page, wbc);
1116 mapping_set_error(mapping, ret);
1117 return ret;
1118}
1119
1120/**
1121 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1122 * @mapping: address space structure to write
1123 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1124 *
1125 * This is a library function, which implements the writepages()
1126 * address_space_operation.
1127 */
1128int generic_writepages(struct address_space *mapping,
1129 struct writeback_control *wbc)
1130{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001131 struct blk_plug plug;
1132 int ret;
1133
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001134 /* deal with chardevs and other special file */
1135 if (!mapping->a_ops->writepage)
1136 return 0;
1137
Shaohua Li9b6096a2011-03-17 10:47:06 +01001138 blk_start_plug(&plug);
1139 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1140 blk_finish_plug(&plug);
1141 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001142}
David Howells811d7362006-08-29 19:06:09 +01001143
1144EXPORT_SYMBOL(generic_writepages);
1145
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1147{
Andrew Morton22905f72005-11-16 15:07:01 -08001148 int ret;
1149
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 if (wbc->nr_to_write <= 0)
1151 return 0;
1152 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001153 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001154 else
1155 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001156 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157}
1158
1159/**
1160 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001161 * @page: the page to write
1162 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 *
1164 * The page must be locked by the caller and will be unlocked upon return.
1165 *
1166 * write_one_page() returns a negative error code if I/O failed.
1167 */
1168int write_one_page(struct page *page, int wait)
1169{
1170 struct address_space *mapping = page->mapping;
1171 int ret = 0;
1172 struct writeback_control wbc = {
1173 .sync_mode = WB_SYNC_ALL,
1174 .nr_to_write = 1,
1175 };
1176
1177 BUG_ON(!PageLocked(page));
1178
1179 if (wait)
1180 wait_on_page_writeback(page);
1181
1182 if (clear_page_dirty_for_io(page)) {
1183 page_cache_get(page);
1184 ret = mapping->a_ops->writepage(page, &wbc);
1185 if (ret == 0 && wait) {
1186 wait_on_page_writeback(page);
1187 if (PageError(page))
1188 ret = -EIO;
1189 }
1190 page_cache_release(page);
1191 } else {
1192 unlock_page(page);
1193 }
1194 return ret;
1195}
1196EXPORT_SYMBOL(write_one_page);
1197
1198/*
Ken Chen76719322007-02-10 01:43:15 -08001199 * For address_spaces which do not use buffers nor write back.
1200 */
1201int __set_page_dirty_no_writeback(struct page *page)
1202{
1203 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001204 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001205 return 0;
1206}
1207
1208/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001209 * Helper function for set_page_dirty family.
1210 * NOTE: This relies on being atomic wrt interrupts.
1211 */
1212void account_page_dirtied(struct page *page, struct address_space *mapping)
1213{
1214 if (mapping_cap_account_dirty(mapping)) {
1215 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001216 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001217 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1218 task_dirty_inc(current);
1219 task_io_account_write(PAGE_CACHE_SIZE);
1220 }
1221}
Michael Rubin679ceac2010-08-20 02:31:26 -07001222EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001223
1224/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001225 * Helper function for set_page_writeback family.
1226 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1227 * wrt interrupts.
1228 */
1229void account_page_writeback(struct page *page)
1230{
1231 inc_zone_page_state(page, NR_WRITEBACK);
Michael Rubinea941f02010-10-26 14:21:35 -07001232 inc_zone_page_state(page, NR_WRITTEN);
Michael Rubinf629d1c2010-10-26 14:21:33 -07001233}
1234EXPORT_SYMBOL(account_page_writeback);
1235
1236/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 * For address_spaces which do not use buffers. Just tag the page as dirty in
1238 * its radix tree.
1239 *
1240 * This is also used when a single buffer is being dirtied: we want to set the
1241 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1242 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1243 *
1244 * Most callers have locked the page, which pins the address_space in memory.
1245 * But zap_pte_range() does not lock the page, however in that case the
1246 * mapping is pinned by the vma's ->vm_file reference.
1247 *
1248 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001249 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 */
1251int __set_page_dirty_nobuffers(struct page *page)
1252{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 if (!TestSetPageDirty(page)) {
1254 struct address_space *mapping = page_mapping(page);
1255 struct address_space *mapping2;
1256
Andrew Morton8c085402006-12-10 02:19:24 -08001257 if (!mapping)
1258 return 1;
1259
Nick Piggin19fd6232008-07-25 19:45:32 -07001260 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001261 mapping2 = page_mapping(page);
1262 if (mapping2) { /* Race with truncate? */
1263 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001264 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001265 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001266 radix_tree_tag_set(&mapping->page_tree,
1267 page_index(page), PAGECACHE_TAG_DIRTY);
1268 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001269 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001270 if (mapping->host) {
1271 /* !PageAnon && !swapper_space */
1272 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001274 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001276 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277}
1278EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1279
1280/*
1281 * When a writepage implementation decides that it doesn't want to write this
1282 * page for some reason, it should redirty the locked page via
1283 * redirty_page_for_writepage() and it should then unlock the page and return 0
1284 */
1285int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1286{
1287 wbc->pages_skipped++;
1288 return __set_page_dirty_nobuffers(page);
1289}
1290EXPORT_SYMBOL(redirty_page_for_writepage);
1291
1292/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02001293 * Dirty a page.
1294 *
1295 * For pages with a mapping this should be done under the page lock
1296 * for the benefit of asynchronous memory errors who prefer a consistent
1297 * dirty state. This rule can be broken in some special cases,
1298 * but should be better not to.
1299 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 * If the mapping doesn't provide a set_page_dirty a_op, then
1301 * just fall through and assume that it wants buffer_heads.
1302 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001303int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304{
1305 struct address_space *mapping = page_mapping(page);
1306
1307 if (likely(mapping)) {
1308 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07001309 /*
1310 * readahead/lru_deactivate_page could remain
1311 * PG_readahead/PG_reclaim due to race with end_page_writeback
1312 * About readahead, if the page is written, the flags would be
1313 * reset. So no problem.
1314 * About lru_deactivate_page, if the page is redirty, the flag
1315 * will be reset. So no problem. but if the page is used by readahead
1316 * it will confuse readahead and make it restart the size rampup
1317 * process. But it's a trivial problem.
1318 */
1319 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02001320#ifdef CONFIG_BLOCK
1321 if (!spd)
1322 spd = __set_page_dirty_buffers;
1323#endif
1324 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001326 if (!PageDirty(page)) {
1327 if (!TestSetPageDirty(page))
1328 return 1;
1329 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 return 0;
1331}
1332EXPORT_SYMBOL(set_page_dirty);
1333
1334/*
1335 * set_page_dirty() is racy if the caller has no reference against
1336 * page->mapping->host, and if the page is unlocked. This is because another
1337 * CPU could truncate the page off the mapping and then free the mapping.
1338 *
1339 * Usually, the page _is_ locked, or the caller is a user-space process which
1340 * holds a reference on the inode by having an open file.
1341 *
1342 * In other cases, the page should be locked before running set_page_dirty().
1343 */
1344int set_page_dirty_lock(struct page *page)
1345{
1346 int ret;
1347
Jens Axboe7eaceac2011-03-10 08:52:07 +01001348 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 ret = set_page_dirty(page);
1350 unlock_page(page);
1351 return ret;
1352}
1353EXPORT_SYMBOL(set_page_dirty_lock);
1354
1355/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 * Clear a page's dirty flag, while caring for dirty memory accounting.
1357 * Returns true if the page was previously dirty.
1358 *
1359 * This is for preparing to put the page under writeout. We leave the page
1360 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1361 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1362 * implementation will run either set_page_writeback() or set_page_dirty(),
1363 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1364 * back into sync.
1365 *
1366 * This incoherency between the page's dirty flag and radix-tree tag is
1367 * unfortunate, but it only exists while the page is locked.
1368 */
1369int clear_page_dirty_for_io(struct page *page)
1370{
1371 struct address_space *mapping = page_mapping(page);
1372
Nick Piggin79352892007-07-19 01:47:22 -07001373 BUG_ON(!PageLocked(page));
1374
Linus Torvalds7658cc22006-12-29 10:00:58 -08001375 if (mapping && mapping_cap_account_dirty(mapping)) {
1376 /*
1377 * Yes, Virginia, this is indeed insane.
1378 *
1379 * We use this sequence to make sure that
1380 * (a) we account for dirty stats properly
1381 * (b) we tell the low-level filesystem to
1382 * mark the whole page dirty if it was
1383 * dirty in a pagetable. Only to then
1384 * (c) clean the page again and return 1 to
1385 * cause the writeback.
1386 *
1387 * This way we avoid all nasty races with the
1388 * dirty bit in multiple places and clearing
1389 * them concurrently from different threads.
1390 *
1391 * Note! Normally the "set_page_dirty(page)"
1392 * has no effect on the actual dirty bit - since
1393 * that will already usually be set. But we
1394 * need the side effects, and it can help us
1395 * avoid races.
1396 *
1397 * We basically use the page "master dirty bit"
1398 * as a serialization point for all the different
1399 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001400 */
1401 if (page_mkclean(page))
1402 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001403 /*
1404 * We carefully synchronise fault handlers against
1405 * installing a dirty pte and marking the page dirty
1406 * at this point. We do this by having them hold the
1407 * page lock at some point after installing their
1408 * pte, but before marking the page dirty.
1409 * Pages are always locked coming in here, so we get
1410 * the desired exclusion. See mm/memory.c:do_wp_page()
1411 * for more comments.
1412 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001413 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001414 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001415 dec_bdi_stat(mapping->backing_dev_info,
1416 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001417 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001419 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001421 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001423EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424
1425int test_clear_page_writeback(struct page *page)
1426{
1427 struct address_space *mapping = page_mapping(page);
1428 int ret;
1429
1430 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001431 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 unsigned long flags;
1433
Nick Piggin19fd6232008-07-25 19:45:32 -07001434 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001436 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 radix_tree_tag_clear(&mapping->page_tree,
1438 page_index(page),
1439 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001440 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001441 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001442 __bdi_writeout_inc(bdi);
1443 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001444 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001445 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446 } else {
1447 ret = TestClearPageWriteback(page);
1448 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001449 if (ret)
1450 dec_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 return ret;
1452}
1453
1454int test_set_page_writeback(struct page *page)
1455{
1456 struct address_space *mapping = page_mapping(page);
1457 int ret;
1458
1459 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001460 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 unsigned long flags;
1462
Nick Piggin19fd6232008-07-25 19:45:32 -07001463 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001465 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 radix_tree_tag_set(&mapping->page_tree,
1467 page_index(page),
1468 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001469 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001470 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1471 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 if (!PageDirty(page))
1473 radix_tree_tag_clear(&mapping->page_tree,
1474 page_index(page),
1475 PAGECACHE_TAG_DIRTY);
Jan Karaf446daae2010-08-09 17:19:12 -07001476 radix_tree_tag_clear(&mapping->page_tree,
1477 page_index(page),
1478 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07001479 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 } else {
1481 ret = TestSetPageWriteback(page);
1482 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001483 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07001484 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 return ret;
1486
1487}
1488EXPORT_SYMBOL(test_set_page_writeback);
1489
1490/*
Nick Piggin00128182007-10-16 01:24:40 -07001491 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 * passed tag.
1493 */
1494int mapping_tagged(struct address_space *mapping, int tag)
1495{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 int ret;
Nick Piggin00128182007-10-16 01:24:40 -07001497 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498 ret = radix_tree_tagged(&mapping->page_tree, tag);
Nick Piggin00128182007-10-16 01:24:40 -07001499 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 return ret;
1501}
1502EXPORT_SYMBOL(mapping_tagged);