blob: d4a6e91bd9e54ada7e2289573880084a6a691562 [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 Fengguangffd1f602011-06-19 22:18:42 -060040 * Sleep at most 200ms at a time in balance_dirty_pages().
41 */
42#define MAX_PAUSE max(HZ/5, 1)
43
44/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060045 * Estimate write bandwidth at 200ms intervals.
46 */
47#define BANDWIDTH_INTERVAL max(HZ/5, 1)
48
Wu Fengguang6c14ae12011-03-02 16:04:18 -060049#define RATELIMIT_CALC_SHIFT 10
50
Wu Fengguange98be2d2010-08-29 11:22:30 -060051/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
53 * will look to see if it needs to force writeback or throttling.
54 */
55static long ratelimit_pages = 32;
56
Linus Torvalds1da177e2005-04-16 15:20:36 -070057/*
58 * When balance_dirty_pages decides that the caller needs to perform some
59 * non-background writeback, this is how many pages it will attempt to write.
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080060 * It should be somewhat larger than dirtied pages to ensure that reasonably
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 * large amounts of I/O are submitted.
62 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080063static inline long sync_writeback_pages(unsigned long dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -070064{
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080065 if (dirtied < ratelimit_pages)
66 dirtied = ratelimit_pages;
67
68 return dirtied + dirtied / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -070069}
70
71/* The following parameters are exported via /proc/sys/vm */
72
73/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020074 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070075 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080076int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
78/*
David Rientjes2da02992009-01-06 14:39:31 -080079 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
80 * dirty_background_ratio * the amount of dirtyable memory
81 */
82unsigned long dirty_background_bytes;
83
84/*
Bron Gondwana195cf452008-02-04 22:29:20 -080085 * free highmem will not be subtracted from the total free memory
86 * for calculating free ratios if vm_highmem_is_dirtyable is true
87 */
88int vm_highmem_is_dirtyable;
89
90/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 * The generator of dirty data starts writeback at this percentage
92 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080093int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
95/*
David Rientjes2da02992009-01-06 14:39:31 -080096 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
97 * vm_dirty_ratio * the amount of dirtyable memory
98 */
99unsigned long vm_dirty_bytes;
100
101/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700102 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700104unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
106/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700107 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111/*
112 * Flag that makes the machine dump writes/reads and block dirtyings.
113 */
114int block_dump;
115
116/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800117 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 */
120int laptop_mode;
121
122EXPORT_SYMBOL(laptop_mode);
123
124/* End of sysctl-exported parameters */
125
Wu Fengguangc42843f2011-03-02 15:54:09 -0600126unsigned long global_dirty_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700129 * Scale the writeback cache size proportional to the relative writeout speeds.
130 *
131 * We do this by keeping a floating proportion between BDIs, based on page
132 * writeback completions [end_page_writeback()]. Those devices that write out
133 * pages fastest will get the larger share, while the slower will get a smaller
134 * share.
135 *
136 * We use page writeout completions because we are interested in getting rid of
137 * dirty pages. Having them written out is the primary goal.
138 *
139 * We introduce a concept of time, a period over which we measure these events,
140 * because demand can/will vary over time. The length of this period itself is
141 * measured in page writeback completions.
142 *
143 */
144static struct prop_descriptor vm_completions;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700145static struct prop_descriptor vm_dirties;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700146
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700147/*
148 * couple the period to the dirty_ratio:
149 *
150 * period/2 ~ roundup_pow_of_two(dirty limit)
151 */
152static int calc_period_shift(void)
153{
154 unsigned long dirty_total;
155
David Rientjes2da02992009-01-06 14:39:31 -0800156 if (vm_dirty_bytes)
157 dirty_total = vm_dirty_bytes / PAGE_SIZE;
158 else
159 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
160 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700161 return 2 + ilog2(dirty_total - 1);
162}
163
164/*
David Rientjes2da02992009-01-06 14:39:31 -0800165 * update the period when the dirty threshold changes.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700166 */
David Rientjes2da02992009-01-06 14:39:31 -0800167static void update_completion_period(void)
168{
169 int shift = calc_period_shift();
170 prop_change_shift(&vm_completions, shift);
171 prop_change_shift(&vm_dirties, shift);
172}
173
174int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700175 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800176 loff_t *ppos)
177{
178 int ret;
179
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700180 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800181 if (ret == 0 && write)
182 dirty_background_bytes = 0;
183 return ret;
184}
185
186int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700187 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800188 loff_t *ppos)
189{
190 int ret;
191
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700192 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800193 if (ret == 0 && write)
194 dirty_background_ratio = 0;
195 return ret;
196}
197
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700198int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700199 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700200 loff_t *ppos)
201{
202 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800203 int ret;
204
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700205 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700206 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
David Rientjes2da02992009-01-06 14:39:31 -0800207 update_completion_period();
208 vm_dirty_bytes = 0;
209 }
210 return ret;
211}
212
213
214int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700215 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800216 loff_t *ppos)
217{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800218 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800219 int ret;
220
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700221 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800222 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
223 update_completion_period();
224 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700225 }
226 return ret;
227}
228
229/*
230 * Increment the BDI's writeout completion count and the global writeout
231 * completion count. Called from test_clear_page_writeback().
232 */
233static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
234{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600235 __inc_bdi_stat(bdi, BDI_WRITTEN);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700236 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
237 bdi->max_prop_frac);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700238}
239
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700240void bdi_writeout_inc(struct backing_dev_info *bdi)
241{
242 unsigned long flags;
243
244 local_irq_save(flags);
245 __bdi_writeout_inc(bdi);
246 local_irq_restore(flags);
247}
248EXPORT_SYMBOL_GPL(bdi_writeout_inc);
249
Nick Piggin1cf6e7d2009-02-18 14:48:18 -0800250void task_dirty_inc(struct task_struct *tsk)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700251{
252 prop_inc_single(&vm_dirties, &tsk->dirties);
253}
254
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700255/*
256 * Obtain an accurate fraction of the BDI's portion.
257 */
258static void bdi_writeout_fraction(struct backing_dev_info *bdi,
259 long *numerator, long *denominator)
260{
Wu Fengguang3efaf0f2010-12-16 22:22:00 -0600261 prop_fraction_percpu(&vm_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700262 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700263}
264
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700265static inline void task_dirties_fraction(struct task_struct *tsk,
266 long *numerator, long *denominator)
267{
268 prop_fraction_single(&vm_dirties, &tsk->dirties,
269 numerator, denominator);
270}
271
272/*
Wu Fengguang1babe182010-08-11 14:17:40 -0700273 * task_dirty_limit - scale down dirty throttling threshold for one task
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700274 *
275 * task specific dirty limit:
276 *
277 * dirty -= (dirty/8) * p_{t}
Wu Fengguang1babe182010-08-11 14:17:40 -0700278 *
279 * To protect light/slow dirtying tasks from heavier/fast ones, we start
280 * throttling individual tasks before reaching the bdi dirty limit.
281 * Relatively low thresholds will be allocated to heavy dirtiers. So when
282 * dirty pages grow large, heavy dirtiers will be throttled first, which will
283 * effectively curb the growth of dirty pages. Light dirtiers with high enough
284 * dirty threshold may never get throttled.
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700285 */
Jan Karabcff25f2011-07-01 13:31:25 -0600286#define TASK_LIMIT_FRACTION 8
Wu Fengguang16c40422010-08-11 14:17:39 -0700287static unsigned long task_dirty_limit(struct task_struct *tsk,
288 unsigned long bdi_dirty)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700289{
290 long numerator, denominator;
Wu Fengguang16c40422010-08-11 14:17:39 -0700291 unsigned long dirty = bdi_dirty;
Jan Karabcff25f2011-07-01 13:31:25 -0600292 u64 inv = dirty / TASK_LIMIT_FRACTION;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700293
294 task_dirties_fraction(tsk, &numerator, &denominator);
295 inv *= numerator;
296 do_div(inv, denominator);
297
298 dirty -= inv;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700299
Wu Fengguang16c40422010-08-11 14:17:39 -0700300 return max(dirty, bdi_dirty/2);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700301}
302
Jan Karabcff25f2011-07-01 13:31:25 -0600303/* Minimum limit for any task */
304static unsigned long task_min_dirty_limit(unsigned long bdi_dirty)
305{
306 return bdi_dirty - bdi_dirty / TASK_LIMIT_FRACTION;
307}
308
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700309/*
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700310 *
311 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700312static unsigned int bdi_min_ratio;
313
314int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
315{
316 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700317
Jens Axboecfc4ba52009-09-14 13:12:40 +0200318 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700319 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700320 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700321 } else {
322 min_ratio -= bdi->min_ratio;
323 if (bdi_min_ratio + min_ratio < 100) {
324 bdi_min_ratio += min_ratio;
325 bdi->min_ratio += min_ratio;
326 } else {
327 ret = -EINVAL;
328 }
329 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200330 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700331
332 return ret;
333}
334
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700335int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
336{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700337 int ret = 0;
338
339 if (max_ratio > 100)
340 return -EINVAL;
341
Jens Axboecfc4ba52009-09-14 13:12:40 +0200342 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700343 if (bdi->min_ratio > max_ratio) {
344 ret = -EINVAL;
345 } else {
346 bdi->max_ratio = max_ratio;
347 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200349 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700350
351 return ret;
352}
353EXPORT_SYMBOL(bdi_set_max_ratio);
354
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700355/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 * Work out the current dirty-memory clamping and background writeout
357 * thresholds.
358 *
359 * The main aim here is to lower them aggressively if there is a lot of mapped
360 * memory around. To avoid stressing page reclaim with lots of unreclaimable
361 * pages. It is better to clamp down on writers than to start swapping, and
362 * performing lots of scanning.
363 *
364 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
365 *
366 * We don't permit the clamping level to fall below 5% - that is getting rather
367 * excessive.
368 *
369 * We make sure that the background writeout level is below the adjusted
370 * clamping level.
371 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700372
373static unsigned long highmem_dirtyable_memory(unsigned long total)
374{
375#ifdef CONFIG_HIGHMEM
376 int node;
377 unsigned long x = 0;
378
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700379 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700380 struct zone *z =
381 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
382
Wu Fengguangadea02a2009-09-21 17:01:42 -0700383 x += zone_page_state(z, NR_FREE_PAGES) +
384 zone_reclaimable_pages(z);
Christoph Lameter1b424462007-05-06 14:48:59 -0700385 }
386 /*
387 * Make sure that the number of highmem pages is never larger
388 * than the number of the total dirtyable memory. This can only
389 * occur in very strange VM situations but we want to make sure
390 * that this does not occur.
391 */
392 return min(x, total);
393#else
394 return 0;
395#endif
396}
397
Steven Rostedt3eefae92008-05-12 21:21:04 +0200398/**
399 * determine_dirtyable_memory - amount of memory that may be used
400 *
401 * Returns the numebr of pages that can currently be freed and used
402 * by the kernel for direct mappings.
403 */
404unsigned long determine_dirtyable_memory(void)
Christoph Lameter1b424462007-05-06 14:48:59 -0700405{
406 unsigned long x;
407
Wu Fengguangadea02a2009-09-21 17:01:42 -0700408 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
Bron Gondwana195cf452008-02-04 22:29:20 -0800409
410 if (!vm_highmem_is_dirtyable)
411 x -= highmem_dirtyable_memory(x);
412
Christoph Lameter1b424462007-05-06 14:48:59 -0700413 return x + 1; /* Ensure that we never return 0 */
414}
415
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600416static unsigned long dirty_freerun_ceiling(unsigned long thresh,
417 unsigned long bg_thresh)
418{
419 return (thresh + bg_thresh) / 2;
420}
421
Wu Fengguangffd1f602011-06-19 22:18:42 -0600422static unsigned long hard_dirty_limit(unsigned long thresh)
423{
424 return max(thresh, global_dirty_limit);
425}
426
Randy Dunlap03ab4502010-08-14 13:05:17 -0700427/*
Wu Fengguang1babe182010-08-11 14:17:40 -0700428 * global_dirty_limits - background-writeback and dirty-throttling thresholds
429 *
430 * Calculate the dirty thresholds based on sysctl parameters
431 * - vm.dirty_background_ratio or vm.dirty_background_bytes
432 * - vm.dirty_ratio or vm.dirty_bytes
433 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Minchan Kimebd13732011-01-04 01:36:48 +0900434 * real-time tasks.
Wu Fengguang1babe182010-08-11 14:17:40 -0700435 */
Wu Fengguang16c40422010-08-11 14:17:39 -0700436void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
David Rientjes364aeb22009-01-06 14:39:29 -0800438 unsigned long background;
439 unsigned long dirty;
Minchan Kim240c8792011-01-13 15:46:27 -0800440 unsigned long uninitialized_var(available_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441 struct task_struct *tsk;
442
Minchan Kim240c8792011-01-13 15:46:27 -0800443 if (!vm_dirty_bytes || !dirty_background_bytes)
444 available_memory = determine_dirtyable_memory();
445
David Rientjes2da02992009-01-06 14:39:31 -0800446 if (vm_dirty_bytes)
447 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
Wu Fengguang4cbec4c2010-10-26 14:21:45 -0700448 else
449 dirty = (vm_dirty_ratio * available_memory) / 100;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
David Rientjes2da02992009-01-06 14:39:31 -0800451 if (dirty_background_bytes)
452 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
453 else
454 background = (dirty_background_ratio * available_memory) / 100;
455
456 if (background >= dirty)
457 background = dirty / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 tsk = current;
459 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
460 background += background / 4;
461 dirty += dirty / 4;
462 }
463 *pbackground = background;
464 *pdirty = dirty;
Wu Fengguange1cbe232010-12-06 22:34:29 -0600465 trace_global_dirty_state(background, dirty);
Wu Fengguang16c40422010-08-11 14:17:39 -0700466}
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700467
Wu Fengguang6f718652011-03-02 17:14:34 -0600468/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700469 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600470 * @bdi: the backing_dev_info to query
471 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700472 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600473 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
474 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
475 * And the "limit" in the name is not seriously taken as hard limit in
476 * balance_dirty_pages().
477 *
478 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700479 * - starving fast devices
480 * - piling up dirty pages (that will take long time to sync) on slow devices
481 *
482 * The bdi's share of dirty limit will be adapting to its throughput and
483 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
484 */
485unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700486{
487 u64 bdi_dirty;
488 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700489
Wu Fengguang16c40422010-08-11 14:17:39 -0700490 /*
491 * Calculate this BDI's share of the dirty ratio.
492 */
493 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700494
Wu Fengguang16c40422010-08-11 14:17:39 -0700495 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
496 bdi_dirty *= numerator;
497 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700498
Wu Fengguang16c40422010-08-11 14:17:39 -0700499 bdi_dirty += (dirty * bdi->min_ratio) / 100;
500 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
501 bdi_dirty = dirty * bdi->max_ratio / 100;
502
503 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504}
505
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600506/*
507 * Dirty position control.
508 *
509 * (o) global/bdi setpoints
510 *
511 * We want the dirty pages be balanced around the global/bdi setpoints.
512 * When the number of dirty pages is higher/lower than the setpoint, the
513 * dirty position control ratio (and hence task dirty ratelimit) will be
514 * decreased/increased to bring the dirty pages back to the setpoint.
515 *
516 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
517 *
518 * if (dirty < setpoint) scale up pos_ratio
519 * if (dirty > setpoint) scale down pos_ratio
520 *
521 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
522 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
523 *
524 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
525 *
526 * (o) global control line
527 *
528 * ^ pos_ratio
529 * |
530 * | |<===== global dirty control scope ======>|
531 * 2.0 .............*
532 * | .*
533 * | . *
534 * | . *
535 * | . *
536 * | . *
537 * | . *
538 * 1.0 ................................*
539 * | . . *
540 * | . . *
541 * | . . *
542 * | . . *
543 * | . . *
544 * 0 +------------.------------------.----------------------*------------->
545 * freerun^ setpoint^ limit^ dirty pages
546 *
547 * (o) bdi control line
548 *
549 * ^ pos_ratio
550 * |
551 * | *
552 * | *
553 * | *
554 * | *
555 * | * |<=========== span ============>|
556 * 1.0 .......................*
557 * | . *
558 * | . *
559 * | . *
560 * | . *
561 * | . *
562 * | . *
563 * | . *
564 * | . *
565 * | . *
566 * | . *
567 * | . *
568 * 1/4 ...............................................* * * * * * * * * * * *
569 * | . .
570 * | . .
571 * | . .
572 * 0 +----------------------.-------------------------------.------------->
573 * bdi_setpoint^ x_intercept^
574 *
575 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
576 * be smoothly throttled down to normal if it starts high in situations like
577 * - start writing to a slow SD card and a fast disk at the same time. The SD
578 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
579 * - the bdi dirty thresh drops quickly due to change of JBOD workload
580 */
581static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
582 unsigned long thresh,
583 unsigned long bg_thresh,
584 unsigned long dirty,
585 unsigned long bdi_thresh,
586 unsigned long bdi_dirty)
587{
588 unsigned long write_bw = bdi->avg_write_bandwidth;
589 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
590 unsigned long limit = hard_dirty_limit(thresh);
591 unsigned long x_intercept;
592 unsigned long setpoint; /* dirty pages' target balance point */
593 unsigned long bdi_setpoint;
594 unsigned long span;
595 long long pos_ratio; /* for scaling up/down the rate limit */
596 long x;
597
598 if (unlikely(dirty >= limit))
599 return 0;
600
601 /*
602 * global setpoint
603 *
604 * setpoint - dirty 3
605 * f(dirty) := 1.0 + (----------------)
606 * limit - setpoint
607 *
608 * it's a 3rd order polynomial that subjects to
609 *
610 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
611 * (2) f(setpoint) = 1.0 => the balance point
612 * (3) f(limit) = 0 => the hard limit
613 * (4) df/dx <= 0 => negative feedback control
614 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
615 * => fast response on large errors; small oscillation near setpoint
616 */
617 setpoint = (freerun + limit) / 2;
618 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
619 limit - setpoint + 1);
620 pos_ratio = x;
621 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
622 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
623 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
624
625 /*
626 * We have computed basic pos_ratio above based on global situation. If
627 * the bdi is over/under its share of dirty pages, we want to scale
628 * pos_ratio further down/up. That is done by the following mechanism.
629 */
630
631 /*
632 * bdi setpoint
633 *
634 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
635 *
636 * x_intercept - bdi_dirty
637 * := --------------------------
638 * x_intercept - bdi_setpoint
639 *
640 * The main bdi control line is a linear function that subjects to
641 *
642 * (1) f(bdi_setpoint) = 1.0
643 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
644 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
645 *
646 * For single bdi case, the dirty pages are observed to fluctuate
647 * regularly within range
648 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
649 * for various filesystems, where (2) can yield in a reasonable 12.5%
650 * fluctuation range for pos_ratio.
651 *
652 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
653 * own size, so move the slope over accordingly and choose a slope that
654 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
655 */
656 if (unlikely(bdi_thresh > thresh))
657 bdi_thresh = thresh;
658 /*
659 * scale global setpoint to bdi's:
660 * bdi_setpoint = setpoint * bdi_thresh / thresh
661 */
662 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
663 bdi_setpoint = setpoint * (u64)x >> 16;
664 /*
665 * Use span=(8*write_bw) in single bdi case as indicated by
666 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
667 *
668 * bdi_thresh thresh - bdi_thresh
669 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
670 * thresh thresh
671 */
672 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
673 x_intercept = bdi_setpoint + span;
674
675 if (bdi_dirty < x_intercept - span / 4) {
676 pos_ratio *= x_intercept - bdi_dirty;
677 do_div(pos_ratio, x_intercept - bdi_setpoint + 1);
678 } else
679 pos_ratio /= 4;
680
681 return pos_ratio;
682}
683
Wu Fengguange98be2d2010-08-29 11:22:30 -0600684static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
685 unsigned long elapsed,
686 unsigned long written)
687{
688 const unsigned long period = roundup_pow_of_two(3 * HZ);
689 unsigned long avg = bdi->avg_write_bandwidth;
690 unsigned long old = bdi->write_bandwidth;
691 u64 bw;
692
693 /*
694 * bw = written * HZ / elapsed
695 *
696 * bw * elapsed + write_bandwidth * (period - elapsed)
697 * write_bandwidth = ---------------------------------------------------
698 * period
699 */
700 bw = written - bdi->written_stamp;
701 bw *= HZ;
702 if (unlikely(elapsed > period)) {
703 do_div(bw, elapsed);
704 avg = bw;
705 goto out;
706 }
707 bw += (u64)bdi->write_bandwidth * (period - elapsed);
708 bw >>= ilog2(period);
709
710 /*
711 * one more level of smoothing, for filtering out sudden spikes
712 */
713 if (avg > old && old >= (unsigned long)bw)
714 avg -= (avg - old) >> 3;
715
716 if (avg < old && old <= (unsigned long)bw)
717 avg += (old - avg) >> 3;
718
719out:
720 bdi->write_bandwidth = bw;
721 bdi->avg_write_bandwidth = avg;
722}
723
Wu Fengguangc42843f2011-03-02 15:54:09 -0600724/*
725 * The global dirtyable memory and dirty threshold could be suddenly knocked
726 * down by a large amount (eg. on the startup of KVM in a swapless system).
727 * This may throw the system into deep dirty exceeded state and throttle
728 * heavy/light dirtiers alike. To retain good responsiveness, maintain
729 * global_dirty_limit for tracking slowly down to the knocked down dirty
730 * threshold.
731 */
732static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
733{
734 unsigned long limit = global_dirty_limit;
735
736 /*
737 * Follow up in one step.
738 */
739 if (limit < thresh) {
740 limit = thresh;
741 goto update;
742 }
743
744 /*
745 * Follow down slowly. Use the higher one as the target, because thresh
746 * may drop below dirty. This is exactly the reason to introduce
747 * global_dirty_limit which is guaranteed to lie above the dirty pages.
748 */
749 thresh = max(thresh, dirty);
750 if (limit > thresh) {
751 limit -= (limit - thresh) >> 5;
752 goto update;
753 }
754 return;
755update:
756 global_dirty_limit = limit;
757}
758
759static void global_update_bandwidth(unsigned long thresh,
760 unsigned long dirty,
761 unsigned long now)
762{
763 static DEFINE_SPINLOCK(dirty_lock);
764 static unsigned long update_time;
765
766 /*
767 * check locklessly first to optimize away locking for the most time
768 */
769 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
770 return;
771
772 spin_lock(&dirty_lock);
773 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
774 update_dirty_limit(thresh, dirty);
775 update_time = now;
776 }
777 spin_unlock(&dirty_lock);
778}
779
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600780/*
781 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
782 *
783 * Normal bdi tasks will be curbed at or below it in long term.
784 * Obviously it should be around (write_bw / N) when there are N dd tasks.
785 */
786static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
787 unsigned long thresh,
788 unsigned long bg_thresh,
789 unsigned long dirty,
790 unsigned long bdi_thresh,
791 unsigned long bdi_dirty,
792 unsigned long dirtied,
793 unsigned long elapsed)
794{
Wu Fengguang73811312011-08-26 15:53:24 -0600795 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
796 unsigned long limit = hard_dirty_limit(thresh);
797 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600798 unsigned long write_bw = bdi->avg_write_bandwidth;
799 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
800 unsigned long dirty_rate;
801 unsigned long task_ratelimit;
802 unsigned long balanced_dirty_ratelimit;
803 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600804 unsigned long step;
805 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600806
807 /*
808 * The dirty rate will match the writeout rate in long term, except
809 * when dirty pages are truncated by userspace or re-dirtied by FS.
810 */
811 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
812
813 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
814 bdi_thresh, bdi_dirty);
815 /*
816 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
817 */
818 task_ratelimit = (u64)dirty_ratelimit *
819 pos_ratio >> RATELIMIT_CALC_SHIFT;
820 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
821
822 /*
823 * A linear estimation of the "balanced" throttle rate. The theory is,
824 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
825 * dirty_rate will be measured to be (N * task_ratelimit). So the below
826 * formula will yield the balanced rate limit (write_bw / N).
827 *
828 * Note that the expanded form is not a pure rate feedback:
829 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
830 * but also takes pos_ratio into account:
831 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
832 *
833 * (1) is not realistic because pos_ratio also takes part in balancing
834 * the dirty rate. Consider the state
835 * pos_ratio = 0.5 (3)
836 * rate = 2 * (write_bw / N) (4)
837 * If (1) is used, it will stuck in that state! Because each dd will
838 * be throttled at
839 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
840 * yielding
841 * dirty_rate = N * task_ratelimit = write_bw (6)
842 * put (6) into (1) we get
843 * rate_(i+1) = rate_(i) (7)
844 *
845 * So we end up using (2) to always keep
846 * rate_(i+1) ~= (write_bw / N) (8)
847 * regardless of the value of pos_ratio. As long as (8) is satisfied,
848 * pos_ratio is able to drive itself to 1.0, which is not only where
849 * the dirty count meet the setpoint, but also where the slope of
850 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
851 */
852 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
853 dirty_rate | 1);
854
Wu Fengguang73811312011-08-26 15:53:24 -0600855 /*
856 * We could safely do this and return immediately:
857 *
858 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
859 *
860 * However to get a more stable dirty_ratelimit, the below elaborated
861 * code makes use of task_ratelimit to filter out sigular points and
862 * limit the step size.
863 *
864 * The below code essentially only uses the relative value of
865 *
866 * task_ratelimit - dirty_ratelimit
867 * = (pos_ratio - 1) * dirty_ratelimit
868 *
869 * which reflects the direction and size of dirty position error.
870 */
871
872 /*
873 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
874 * task_ratelimit is on the same side of dirty_ratelimit, too.
875 * For example, when
876 * - dirty_ratelimit > balanced_dirty_ratelimit
877 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
878 * lowering dirty_ratelimit will help meet both the position and rate
879 * control targets. Otherwise, don't update dirty_ratelimit if it will
880 * only help meet the rate target. After all, what the users ultimately
881 * feel and care are stable dirty rate and small position error.
882 *
883 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
884 * and filter out the sigular points of balanced_dirty_ratelimit. Which
885 * keeps jumping around randomly and can even leap far away at times
886 * due to the small 200ms estimation period of dirty_rate (we want to
887 * keep that period small to reduce time lags).
888 */
889 step = 0;
890 if (dirty < setpoint) {
891 x = min(bdi->balanced_dirty_ratelimit,
892 min(balanced_dirty_ratelimit, task_ratelimit));
893 if (dirty_ratelimit < x)
894 step = x - dirty_ratelimit;
895 } else {
896 x = max(bdi->balanced_dirty_ratelimit,
897 max(balanced_dirty_ratelimit, task_ratelimit));
898 if (dirty_ratelimit > x)
899 step = dirty_ratelimit - x;
900 }
901
902 /*
903 * Don't pursue 100% rate matching. It's impossible since the balanced
904 * rate itself is constantly fluctuating. So decrease the track speed
905 * when it gets close to the target. Helps eliminate pointless tremors.
906 */
907 step >>= dirty_ratelimit / (2 * step + 1);
908 /*
909 * Limit the tracking speed to avoid overshooting.
910 */
911 step = (step + 7) / 8;
912
913 if (dirty_ratelimit < balanced_dirty_ratelimit)
914 dirty_ratelimit += step;
915 else
916 dirty_ratelimit -= step;
917
918 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
919 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600920}
921
Wu Fengguange98be2d2010-08-29 11:22:30 -0600922void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600923 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600924 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600925 unsigned long dirty,
926 unsigned long bdi_thresh,
927 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600928 unsigned long start_time)
929{
930 unsigned long now = jiffies;
931 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600932 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600933 unsigned long written;
934
935 /*
936 * rate-limit, only update once every 200ms.
937 */
938 if (elapsed < BANDWIDTH_INTERVAL)
939 return;
940
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600941 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600942 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
943
944 /*
945 * Skip quiet periods when disk bandwidth is under-utilized.
946 * (at least 1s idle time between two flusher runs)
947 */
948 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
949 goto snapshot;
950
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600951 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -0600952 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600953 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
954 bdi_thresh, bdi_dirty,
955 dirtied, elapsed);
956 }
Wu Fengguange98be2d2010-08-29 11:22:30 -0600957 bdi_update_write_bandwidth(bdi, elapsed, written);
958
959snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600960 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600961 bdi->written_stamp = written;
962 bdi->bw_time_stamp = now;
963}
964
965static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600966 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600967 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600968 unsigned long dirty,
969 unsigned long bdi_thresh,
970 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600971 unsigned long start_time)
972{
973 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
974 return;
975 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600976 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
977 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600978 spin_unlock(&bdi->wb.list_lock);
979}
980
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981/*
982 * balance_dirty_pages() must be called by processes which are generating dirty
983 * data. It looks at the number of dirty pages in the machine and will force
984 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
Jens Axboe5b0830c2009-09-23 19:37:09 +0200985 * If we're over `background_thresh' then the writeback threads are woken to
986 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800988static void balance_dirty_pages(struct address_space *mapping,
989 unsigned long write_chunk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990{
Wu Fengguang77627412010-09-12 13:34:05 -0600991 unsigned long nr_reclaimable, bdi_nr_reclaimable;
992 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
993 unsigned long bdi_dirty;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600994 unsigned long freerun;
David Rientjes364aeb22009-01-06 14:39:29 -0800995 unsigned long background_thresh;
996 unsigned long dirty_thresh;
997 unsigned long bdi_thresh;
Jan Karabcff25f2011-07-01 13:31:25 -0600998 unsigned long task_bdi_thresh;
999 unsigned long min_task_bdi_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 unsigned long pages_written = 0;
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001001 unsigned long pause = 1;
Wu Fengguange50e3722010-08-11 14:17:37 -07001002 bool dirty_exceeded = false;
Jan Karabcff25f2011-07-01 13:31:25 -06001003 bool clear_dirty_exceeded = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001005 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006
1007 for (;;) {
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001008 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1009 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -06001010 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001011
Wu Fengguang16c40422010-08-11 14:17:39 -07001012 global_dirty_limits(&background_thresh, &dirty_thresh);
1013
1014 /*
1015 * Throttle it only when the background writeback cannot
1016 * catch-up. This avoids (excessively) small writeouts
1017 * when the bdi limits are ramping up.
1018 */
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001019 freerun = dirty_freerun_ceiling(dirty_thresh,
1020 background_thresh);
1021 if (nr_dirty <= freerun)
Wu Fengguang16c40422010-08-11 14:17:39 -07001022 break;
1023
1024 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Jan Karabcff25f2011-07-01 13:31:25 -06001025 min_task_bdi_thresh = task_min_dirty_limit(bdi_thresh);
1026 task_bdi_thresh = task_dirty_limit(current, bdi_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001027
Wu Fengguange50e3722010-08-11 14:17:37 -07001028 /*
1029 * In order to avoid the stacked BDI deadlock we need
1030 * to ensure we accurately count the 'dirty' pages when
1031 * the threshold is low.
1032 *
1033 * Otherwise it would be possible to get thresh+n pages
1034 * reported dirty, even though there are thresh-m pages
1035 * actually dirty; with m+n sitting in the percpu
1036 * deltas.
1037 */
Jan Karabcff25f2011-07-01 13:31:25 -06001038 if (task_bdi_thresh < 2 * bdi_stat_error(bdi)) {
Wu Fengguange50e3722010-08-11 14:17:37 -07001039 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
Wu Fengguang77627412010-09-12 13:34:05 -06001040 bdi_dirty = bdi_nr_reclaimable +
1041 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001042 } else {
1043 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
Wu Fengguang77627412010-09-12 13:34:05 -06001044 bdi_dirty = bdi_nr_reclaimable +
1045 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001046 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001047
Wu Fengguange50e3722010-08-11 14:17:37 -07001048 /*
1049 * The bdi thresh is somehow "soft" limit derived from the
1050 * global "hard" limit. The former helps to prevent heavy IO
1051 * bdi or process from holding back light ones; The latter is
1052 * the last resort safeguard.
1053 */
Jan Karabcff25f2011-07-01 13:31:25 -06001054 dirty_exceeded = (bdi_dirty > task_bdi_thresh) ||
Wu Fengguang77627412010-09-12 13:34:05 -06001055 (nr_dirty > dirty_thresh);
Jan Karabcff25f2011-07-01 13:31:25 -06001056 clear_dirty_exceeded = (bdi_dirty <= min_task_bdi_thresh) &&
1057 (nr_dirty <= dirty_thresh);
Wu Fengguange50e3722010-08-11 14:17:37 -07001058
1059 if (!dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001060 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001062 if (!bdi->dirty_exceeded)
1063 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001065 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1066 nr_dirty, bdi_thresh, bdi_dirty,
1067 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001068
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
1070 * Unstable writes are a feature of certain networked
1071 * filesystems (i.e. NFS) in which data may have been
1072 * written to the server's write cache, but has not yet
1073 * been flushed to permanent storage.
Richard Kennedyd7831a02009-06-30 11:41:35 -07001074 * Only move pages to writeback if this bdi is over its
1075 * threshold otherwise wait until the disk writes catch
1076 * up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 */
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001078 trace_balance_dirty_start(bdi);
Jan Karabcff25f2011-07-01 13:31:25 -06001079 if (bdi_nr_reclaimable > task_bdi_thresh) {
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001080 pages_written += writeback_inodes_wb(&bdi->wb,
1081 write_chunk);
1082 trace_balance_dirty_written(bdi, pages_written);
Wu Fengguange50e3722010-08-11 14:17:37 -07001083 if (pages_written >= write_chunk)
1084 break; /* We've done our duty */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085 }
Wu Fengguangd153ba62010-12-21 17:24:21 -08001086 __set_current_state(TASK_UNINTERRUPTIBLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001087 io_schedule_timeout(pause);
Wu Fengguangd46db3d2011-05-04 19:54:37 -06001088 trace_balance_dirty_wait(bdi);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001089
Wu Fengguangffd1f602011-06-19 22:18:42 -06001090 dirty_thresh = hard_dirty_limit(dirty_thresh);
1091 /*
1092 * max-pause area. If dirty exceeded but still within this
1093 * area, no need to sleep for more than 200ms: (a) 8 pages per
1094 * 200ms is typically more than enough to curb heavy dirtiers;
1095 * (b) the pause time limit makes the dirtiers more responsive.
1096 */
Wu Fengguangbb082292011-08-16 13:37:14 -06001097 if (nr_dirty < dirty_thresh &&
1098 bdi_dirty < (task_bdi_thresh + bdi_thresh) / 2 &&
Wu Fengguangffd1f602011-06-19 22:18:42 -06001099 time_after(jiffies, start_time + MAX_PAUSE))
1100 break;
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001101
1102 /*
1103 * Increase the delay for each loop, up to our previous
1104 * default of taking a 100ms nap.
1105 */
1106 pause <<= 1;
1107 if (pause > HZ / 10)
1108 pause = HZ / 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 }
1110
Jan Karabcff25f2011-07-01 13:31:25 -06001111 /* Clear dirty_exceeded flag only when no task can exceed the limit */
1112 if (clear_dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001113 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114
1115 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001116 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117
1118 /*
1119 * In laptop mode, we wait until hitting the higher threshold before
1120 * starting background writeout, and then write out all the way down
1121 * to the lower threshold. So slow writers cause minimal disk activity.
1122 *
1123 * In normal mode, we start background writeout at the lower
1124 * background_thresh, to keep the amount of dirty memory low.
1125 */
1126 if ((laptop_mode && pages_written) ||
Wu Fengguange50e3722010-08-11 14:17:37 -07001127 (!laptop_mode && (nr_reclaimable > background_thresh)))
Christoph Hellwigc5444192010-06-08 18:15:15 +02001128 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129}
1130
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001131void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001132{
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001133 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001134 struct address_space *mapping = page_mapping(page);
1135
1136 if (mapping)
1137 balance_dirty_pages_ratelimited(mapping);
1138 }
1139}
1140
Tejun Heo245b2e72009-06-24 15:13:48 +09001141static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
1142
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001144 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001145 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +02001146 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 *
1148 * Processes which are dirtying memory should call in here once for each page
1149 * which was newly dirtied. The function will periodically check the system's
1150 * dirty state and will initiate writeback if needed.
1151 *
1152 * On really big machines, get_writeback_state is expensive, so try to avoid
1153 * calling it too often (ratelimiting). But once we're over the dirty memory
1154 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1155 * from overshooting the limit by (ratelimit_pages) each.
1156 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001157void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1158 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001160 struct backing_dev_info *bdi = mapping->backing_dev_info;
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001161 unsigned long ratelimit;
1162 unsigned long *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163
Wu Fengguang36715ce2011-06-11 17:53:57 -06001164 if (!bdi_cap_account_dirty(bdi))
1165 return;
1166
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 ratelimit = ratelimit_pages;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001168 if (mapping->backing_dev_info->dirty_exceeded)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 ratelimit = 8;
1170
1171 /*
1172 * Check the rate limiting. Also, we do not want to throttle real-time
1173 * tasks in balance_dirty_pages(). Period.
1174 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001175 preempt_disable();
Tejun Heo245b2e72009-06-24 15:13:48 +09001176 p = &__get_cpu_var(bdp_ratelimits);
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001177 *p += nr_pages_dirtied;
1178 if (unlikely(*p >= ratelimit)) {
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001179 ratelimit = sync_writeback_pages(*p);
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001180 *p = 0;
1181 preempt_enable();
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001182 balance_dirty_pages(mapping, ratelimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 return;
1184 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001185 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186}
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001187EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
Andrew Morton232ea4d2007-02-28 20:13:21 -08001189void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190{
David Rientjes364aeb22009-01-06 14:39:29 -08001191 unsigned long background_thresh;
1192 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193
1194 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001195 global_dirty_limits(&background_thresh, &dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196
1197 /*
1198 * Boost the allowable dirty threshold a bit for page
1199 * allocators so they don't get DoS'ed by heavy writers
1200 */
1201 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1202
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001203 if (global_page_state(NR_UNSTABLE_NFS) +
1204 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1205 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001206 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001207
1208 /*
1209 * The caller might hold locks which can prevent IO completion
1210 * or progress in the filesystem. So we cannot just sit here
1211 * waiting for IO to complete.
1212 */
1213 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1214 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 }
1216}
1217
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1220 */
1221int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001222 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001224 proc_dointvec(table, write, buffer, length, ppos);
Jens Axboe64231042010-05-21 20:00:35 +02001225 bdi_arm_supers_timer();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 return 0;
1227}
1228
Jens Axboec2c49862010-05-20 09:18:47 +02001229#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001230void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231{
Matthew Garrett31373d02010-04-06 14:25:14 +02001232 struct request_queue *q = (struct request_queue *)data;
1233 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1234 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235
Matthew Garrett31373d02010-04-06 14:25:14 +02001236 /*
1237 * We want to write everything out, not just down to the dirty
1238 * threshold
1239 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001240 if (bdi_has_dirty_io(&q->backing_dev_info))
Christoph Hellwigc5444192010-06-08 18:15:15 +02001241 bdi_start_writeback(&q->backing_dev_info, nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242}
1243
1244/*
1245 * We've spun up the disk and we're in laptop mode: schedule writeback
1246 * of all dirty data a few seconds from now. If the flush is already scheduled
1247 * then push it back - the user is still using the disk.
1248 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001249void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250{
Matthew Garrett31373d02010-04-06 14:25:14 +02001251 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252}
1253
1254/*
1255 * We're in laptop mode and we've just synced. The sync's writes will have
1256 * caused another writeback to be scheduled by laptop_io_completion.
1257 * Nothing needs to be written back anymore, so we unschedule the writeback.
1258 */
1259void laptop_sync_completion(void)
1260{
Matthew Garrett31373d02010-04-06 14:25:14 +02001261 struct backing_dev_info *bdi;
1262
1263 rcu_read_lock();
1264
1265 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1266 del_timer(&bdi->laptop_mode_wb_timer);
1267
1268 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269}
Jens Axboec2c49862010-05-20 09:18:47 +02001270#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271
1272/*
1273 * If ratelimit_pages is too high then we can get into dirty-data overload
1274 * if a large number of processes all perform writes at the same time.
1275 * If it is too low then SMP machines will call the (expensive)
1276 * get_writeback_state too often.
1277 *
1278 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1279 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1280 * thresholds before writeback cuts in.
1281 *
1282 * But the limit should not be set too high. Because it also controls the
1283 * amount of memory which the balance_dirty_pages() caller has to write back.
1284 * If this is too large then the caller will block on the IO queue all the
1285 * time. So limit it to four megabytes - the balance_dirty_pages() caller
1286 * will write six megabyte chunks, max.
1287 */
1288
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001289void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290{
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -07001291 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 if (ratelimit_pages < 16)
1293 ratelimit_pages = 16;
1294 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
1295 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
1296}
1297
Chandra Seetharaman26c21432006-06-27 02:54:10 -07001298static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1300{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001301 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -08001302 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303}
1304
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001305static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 .notifier_call = ratelimit_handler,
1307 .next = NULL,
1308};
1309
1310/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001311 * Called early on to tune the page writeback dirty limits.
1312 *
1313 * We used to scale dirty pages according to how total memory
1314 * related to pages that could be allocated for buffers (by
1315 * comparing nr_free_buffer_pages() to vm_total_pages.
1316 *
1317 * However, that was when we used "dirty_ratio" to scale with
1318 * all memory, and we don't do that any more. "dirty_ratio"
1319 * is now applied to total non-HIGHPAGE memory (by subtracting
1320 * totalhigh_pages from vm_total_pages), and as such we can't
1321 * get into the old insane situation any more where we had
1322 * large amounts of dirty pages compared to a small amount of
1323 * non-HIGHMEM memory.
1324 *
1325 * But we might still want to scale the dirty_ratio by how
1326 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 */
1328void __init page_writeback_init(void)
1329{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001330 int shift;
1331
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001332 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001334
1335 shift = calc_period_shift();
1336 prop_descriptor_init(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -07001337 prop_descriptor_init(&vm_dirties, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338}
1339
David Howells811d7362006-08-29 19:06:09 +01001340/**
Jan Karaf446daa2010-08-09 17:19:12 -07001341 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1342 * @mapping: address space structure to write
1343 * @start: starting page index
1344 * @end: ending page index (inclusive)
1345 *
1346 * This function scans the page range from @start to @end (inclusive) and tags
1347 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1348 * that write_cache_pages (or whoever calls this function) will then use
1349 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1350 * used to avoid livelocking of writeback by a process steadily creating new
1351 * dirty pages in the file (thus it is important for this function to be quick
1352 * so that it can tag pages faster than a dirtying process can create them).
1353 */
1354/*
1355 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1356 */
Jan Karaf446daa2010-08-09 17:19:12 -07001357void tag_pages_for_writeback(struct address_space *mapping,
1358 pgoff_t start, pgoff_t end)
1359{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001360#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daa2010-08-09 17:19:12 -07001361 unsigned long tagged;
1362
1363 do {
1364 spin_lock_irq(&mapping->tree_lock);
1365 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1366 &start, end, WRITEBACK_TAG_BATCH,
1367 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1368 spin_unlock_irq(&mapping->tree_lock);
1369 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1370 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001371 /* We check 'start' to handle wrapping when end == ~0UL */
1372 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daa2010-08-09 17:19:12 -07001373}
1374EXPORT_SYMBOL(tag_pages_for_writeback);
1375
1376/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001377 * 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 +01001378 * @mapping: address space structure to write
1379 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001380 * @writepage: function called for each page
1381 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001382 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001383 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001384 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1385 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1386 * and msync() need to guarantee that all the data which was dirty at the time
1387 * the call was made get new I/O started against them. If wbc->sync_mode is
1388 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1389 * existing IO to complete.
Jan Karaf446daa2010-08-09 17:19:12 -07001390 *
1391 * To avoid livelocks (when other process dirties new pages), we first tag
1392 * pages which should be written back with TOWRITE tag and only then start
1393 * writing them. For data-integrity sync we have to be careful so that we do
1394 * not miss some pages (e.g., because some other process has cleared TOWRITE
1395 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1396 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001397 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001398int write_cache_pages(struct address_space *mapping,
1399 struct writeback_control *wbc, writepage_t writepage,
1400 void *data)
David Howells811d7362006-08-29 19:06:09 +01001401{
David Howells811d7362006-08-29 19:06:09 +01001402 int ret = 0;
1403 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001404 struct pagevec pvec;
1405 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001406 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001407 pgoff_t index;
1408 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001409 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001410 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001411 int range_whole = 0;
Jan Karaf446daa2010-08-09 17:19:12 -07001412 int tag;
David Howells811d7362006-08-29 19:06:09 +01001413
David Howells811d7362006-08-29 19:06:09 +01001414 pagevec_init(&pvec, 0);
1415 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001416 writeback_index = mapping->writeback_index; /* prev offset */
1417 index = writeback_index;
1418 if (index == 0)
1419 cycled = 1;
1420 else
1421 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001422 end = -1;
1423 } else {
1424 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1425 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1426 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1427 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001428 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001429 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001430 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daa2010-08-09 17:19:12 -07001431 tag = PAGECACHE_TAG_TOWRITE;
1432 else
1433 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001434retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001435 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daa2010-08-09 17:19:12 -07001436 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001437 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001438 while (!done && (index <= end)) {
1439 int i;
1440
Jan Karaf446daa2010-08-09 17:19:12 -07001441 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001442 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1443 if (nr_pages == 0)
1444 break;
David Howells811d7362006-08-29 19:06:09 +01001445
David Howells811d7362006-08-29 19:06:09 +01001446 for (i = 0; i < nr_pages; i++) {
1447 struct page *page = pvec.pages[i];
1448
Nick Piggind5482cd2009-01-06 14:39:11 -08001449 /*
1450 * At this point, the page may be truncated or
1451 * invalidated (changing page->mapping to NULL), or
1452 * even swizzled back from swapper_space to tmpfs file
1453 * mapping. However, page->index will not change
1454 * because we have a reference on the page.
1455 */
1456 if (page->index > end) {
1457 /*
1458 * can't be range_cyclic (1st pass) because
1459 * end == -1 in that case.
1460 */
1461 done = 1;
1462 break;
1463 }
1464
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001465 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001466
David Howells811d7362006-08-29 19:06:09 +01001467 lock_page(page);
1468
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001469 /*
1470 * Page truncated or invalidated. We can freely skip it
1471 * then, even for data integrity operations: the page
1472 * has disappeared concurrently, so there could be no
1473 * real expectation of this data interity operation
1474 * even if there is now a new, dirty page at the same
1475 * pagecache address.
1476 */
David Howells811d7362006-08-29 19:06:09 +01001477 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001478continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001479 unlock_page(page);
1480 continue;
1481 }
1482
Nick Piggin515f4a02009-01-06 14:39:10 -08001483 if (!PageDirty(page)) {
1484 /* someone wrote it for us */
1485 goto continue_unlock;
1486 }
David Howells811d7362006-08-29 19:06:09 +01001487
Nick Piggin515f4a02009-01-06 14:39:10 -08001488 if (PageWriteback(page)) {
1489 if (wbc->sync_mode != WB_SYNC_NONE)
1490 wait_on_page_writeback(page);
1491 else
1492 goto continue_unlock;
1493 }
1494
1495 BUG_ON(PageWriteback(page));
1496 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001497 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001498
Dave Chinner9e094382010-07-07 13:24:08 +10001499 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001500 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001501 if (unlikely(ret)) {
1502 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1503 unlock_page(page);
1504 ret = 0;
1505 } else {
1506 /*
1507 * done_index is set past this page,
1508 * so media errors will not choke
1509 * background writeout for the entire
1510 * file. This has consequences for
1511 * range_cyclic semantics (ie. it may
1512 * not be suitable for data integrity
1513 * writeout).
1514 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001515 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001516 done = 1;
1517 break;
1518 }
Dave Chinner0b564922010-06-09 10:37:18 +10001519 }
David Howells811d7362006-08-29 19:06:09 +01001520
Dave Chinner546a1922010-08-24 11:44:34 +10001521 /*
1522 * We stop writing back only if we are not doing
1523 * integrity sync. In case of integrity sync we have to
1524 * keep going until we have written all the pages
1525 * we tagged for writeback prior to entering this loop.
1526 */
1527 if (--wbc->nr_to_write <= 0 &&
1528 wbc->sync_mode == WB_SYNC_NONE) {
1529 done = 1;
1530 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001531 }
David Howells811d7362006-08-29 19:06:09 +01001532 }
1533 pagevec_release(&pvec);
1534 cond_resched();
1535 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001536 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001537 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001538 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001539 * We hit the last page and there is more work to be done: wrap
1540 * back to the start of the file
1541 */
Nick Piggin31a12662009-01-06 14:39:04 -08001542 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001543 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001544 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001545 goto retry;
1546 }
Dave Chinner0b564922010-06-09 10:37:18 +10001547 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1548 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001549
David Howells811d7362006-08-29 19:06:09 +01001550 return ret;
1551}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001552EXPORT_SYMBOL(write_cache_pages);
1553
1554/*
1555 * Function used by generic_writepages to call the real writepage
1556 * function and set the mapping flags on error
1557 */
1558static int __writepage(struct page *page, struct writeback_control *wbc,
1559 void *data)
1560{
1561 struct address_space *mapping = data;
1562 int ret = mapping->a_ops->writepage(page, wbc);
1563 mapping_set_error(mapping, ret);
1564 return ret;
1565}
1566
1567/**
1568 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1569 * @mapping: address space structure to write
1570 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1571 *
1572 * This is a library function, which implements the writepages()
1573 * address_space_operation.
1574 */
1575int generic_writepages(struct address_space *mapping,
1576 struct writeback_control *wbc)
1577{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001578 struct blk_plug plug;
1579 int ret;
1580
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001581 /* deal with chardevs and other special file */
1582 if (!mapping->a_ops->writepage)
1583 return 0;
1584
Shaohua Li9b6096a2011-03-17 10:47:06 +01001585 blk_start_plug(&plug);
1586 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1587 blk_finish_plug(&plug);
1588 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001589}
David Howells811d7362006-08-29 19:06:09 +01001590
1591EXPORT_SYMBOL(generic_writepages);
1592
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1594{
Andrew Morton22905f72005-11-16 15:07:01 -08001595 int ret;
1596
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 if (wbc->nr_to_write <= 0)
1598 return 0;
1599 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001600 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001601 else
1602 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001603 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604}
1605
1606/**
1607 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001608 * @page: the page to write
1609 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 *
1611 * The page must be locked by the caller and will be unlocked upon return.
1612 *
1613 * write_one_page() returns a negative error code if I/O failed.
1614 */
1615int write_one_page(struct page *page, int wait)
1616{
1617 struct address_space *mapping = page->mapping;
1618 int ret = 0;
1619 struct writeback_control wbc = {
1620 .sync_mode = WB_SYNC_ALL,
1621 .nr_to_write = 1,
1622 };
1623
1624 BUG_ON(!PageLocked(page));
1625
1626 if (wait)
1627 wait_on_page_writeback(page);
1628
1629 if (clear_page_dirty_for_io(page)) {
1630 page_cache_get(page);
1631 ret = mapping->a_ops->writepage(page, &wbc);
1632 if (ret == 0 && wait) {
1633 wait_on_page_writeback(page);
1634 if (PageError(page))
1635 ret = -EIO;
1636 }
1637 page_cache_release(page);
1638 } else {
1639 unlock_page(page);
1640 }
1641 return ret;
1642}
1643EXPORT_SYMBOL(write_one_page);
1644
1645/*
Ken Chen76719322007-02-10 01:43:15 -08001646 * For address_spaces which do not use buffers nor write back.
1647 */
1648int __set_page_dirty_no_writeback(struct page *page)
1649{
1650 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001651 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001652 return 0;
1653}
1654
1655/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001656 * Helper function for set_page_dirty family.
1657 * NOTE: This relies on being atomic wrt interrupts.
1658 */
1659void account_page_dirtied(struct page *page, struct address_space *mapping)
1660{
1661 if (mapping_cap_account_dirty(mapping)) {
1662 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001663 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001664 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06001665 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001666 task_dirty_inc(current);
1667 task_io_account_write(PAGE_CACHE_SIZE);
1668 }
1669}
Michael Rubin679ceac2010-08-20 02:31:26 -07001670EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001671
1672/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001673 * Helper function for set_page_writeback family.
1674 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1675 * wrt interrupts.
1676 */
1677void account_page_writeback(struct page *page)
1678{
1679 inc_zone_page_state(page, NR_WRITEBACK);
1680}
1681EXPORT_SYMBOL(account_page_writeback);
1682
1683/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 * For address_spaces which do not use buffers. Just tag the page as dirty in
1685 * its radix tree.
1686 *
1687 * This is also used when a single buffer is being dirtied: we want to set the
1688 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1689 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1690 *
1691 * Most callers have locked the page, which pins the address_space in memory.
1692 * But zap_pte_range() does not lock the page, however in that case the
1693 * mapping is pinned by the vma's ->vm_file reference.
1694 *
1695 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001696 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 */
1698int __set_page_dirty_nobuffers(struct page *page)
1699{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 if (!TestSetPageDirty(page)) {
1701 struct address_space *mapping = page_mapping(page);
1702 struct address_space *mapping2;
1703
Andrew Morton8c085402006-12-10 02:19:24 -08001704 if (!mapping)
1705 return 1;
1706
Nick Piggin19fd6232008-07-25 19:45:32 -07001707 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001708 mapping2 = page_mapping(page);
1709 if (mapping2) { /* Race with truncate? */
1710 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001711 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001712 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001713 radix_tree_tag_set(&mapping->page_tree,
1714 page_index(page), PAGECACHE_TAG_DIRTY);
1715 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001716 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001717 if (mapping->host) {
1718 /* !PageAnon && !swapper_space */
1719 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001721 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001723 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724}
1725EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1726
1727/*
1728 * When a writepage implementation decides that it doesn't want to write this
1729 * page for some reason, it should redirty the locked page via
1730 * redirty_page_for_writepage() and it should then unlock the page and return 0
1731 */
1732int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1733{
1734 wbc->pages_skipped++;
1735 return __set_page_dirty_nobuffers(page);
1736}
1737EXPORT_SYMBOL(redirty_page_for_writepage);
1738
1739/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02001740 * Dirty a page.
1741 *
1742 * For pages with a mapping this should be done under the page lock
1743 * for the benefit of asynchronous memory errors who prefer a consistent
1744 * dirty state. This rule can be broken in some special cases,
1745 * but should be better not to.
1746 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 * If the mapping doesn't provide a set_page_dirty a_op, then
1748 * just fall through and assume that it wants buffer_heads.
1749 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001750int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751{
1752 struct address_space *mapping = page_mapping(page);
1753
1754 if (likely(mapping)) {
1755 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07001756 /*
1757 * readahead/lru_deactivate_page could remain
1758 * PG_readahead/PG_reclaim due to race with end_page_writeback
1759 * About readahead, if the page is written, the flags would be
1760 * reset. So no problem.
1761 * About lru_deactivate_page, if the page is redirty, the flag
1762 * will be reset. So no problem. but if the page is used by readahead
1763 * it will confuse readahead and make it restart the size rampup
1764 * process. But it's a trivial problem.
1765 */
1766 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02001767#ifdef CONFIG_BLOCK
1768 if (!spd)
1769 spd = __set_page_dirty_buffers;
1770#endif
1771 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001773 if (!PageDirty(page)) {
1774 if (!TestSetPageDirty(page))
1775 return 1;
1776 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 return 0;
1778}
1779EXPORT_SYMBOL(set_page_dirty);
1780
1781/*
1782 * set_page_dirty() is racy if the caller has no reference against
1783 * page->mapping->host, and if the page is unlocked. This is because another
1784 * CPU could truncate the page off the mapping and then free the mapping.
1785 *
1786 * Usually, the page _is_ locked, or the caller is a user-space process which
1787 * holds a reference on the inode by having an open file.
1788 *
1789 * In other cases, the page should be locked before running set_page_dirty().
1790 */
1791int set_page_dirty_lock(struct page *page)
1792{
1793 int ret;
1794
Jens Axboe7eaceac2011-03-10 08:52:07 +01001795 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 ret = set_page_dirty(page);
1797 unlock_page(page);
1798 return ret;
1799}
1800EXPORT_SYMBOL(set_page_dirty_lock);
1801
1802/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 * Clear a page's dirty flag, while caring for dirty memory accounting.
1804 * Returns true if the page was previously dirty.
1805 *
1806 * This is for preparing to put the page under writeout. We leave the page
1807 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1808 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1809 * implementation will run either set_page_writeback() or set_page_dirty(),
1810 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1811 * back into sync.
1812 *
1813 * This incoherency between the page's dirty flag and radix-tree tag is
1814 * unfortunate, but it only exists while the page is locked.
1815 */
1816int clear_page_dirty_for_io(struct page *page)
1817{
1818 struct address_space *mapping = page_mapping(page);
1819
Nick Piggin79352892007-07-19 01:47:22 -07001820 BUG_ON(!PageLocked(page));
1821
Linus Torvalds7658cc22006-12-29 10:00:58 -08001822 if (mapping && mapping_cap_account_dirty(mapping)) {
1823 /*
1824 * Yes, Virginia, this is indeed insane.
1825 *
1826 * We use this sequence to make sure that
1827 * (a) we account for dirty stats properly
1828 * (b) we tell the low-level filesystem to
1829 * mark the whole page dirty if it was
1830 * dirty in a pagetable. Only to then
1831 * (c) clean the page again and return 1 to
1832 * cause the writeback.
1833 *
1834 * This way we avoid all nasty races with the
1835 * dirty bit in multiple places and clearing
1836 * them concurrently from different threads.
1837 *
1838 * Note! Normally the "set_page_dirty(page)"
1839 * has no effect on the actual dirty bit - since
1840 * that will already usually be set. But we
1841 * need the side effects, and it can help us
1842 * avoid races.
1843 *
1844 * We basically use the page "master dirty bit"
1845 * as a serialization point for all the different
1846 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001847 */
1848 if (page_mkclean(page))
1849 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001850 /*
1851 * We carefully synchronise fault handlers against
1852 * installing a dirty pte and marking the page dirty
1853 * at this point. We do this by having them hold the
1854 * page lock at some point after installing their
1855 * pte, but before marking the page dirty.
1856 * Pages are always locked coming in here, so we get
1857 * the desired exclusion. See mm/memory.c:do_wp_page()
1858 * for more comments.
1859 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001860 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001861 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001862 dec_bdi_stat(mapping->backing_dev_info,
1863 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001864 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001866 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001868 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001870EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871
1872int test_clear_page_writeback(struct page *page)
1873{
1874 struct address_space *mapping = page_mapping(page);
1875 int ret;
1876
1877 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001878 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 unsigned long flags;
1880
Nick Piggin19fd6232008-07-25 19:45:32 -07001881 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001883 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884 radix_tree_tag_clear(&mapping->page_tree,
1885 page_index(page),
1886 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001887 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001888 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001889 __bdi_writeout_inc(bdi);
1890 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001891 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001892 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 } else {
1894 ret = TestClearPageWriteback(page);
1895 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07001896 if (ret) {
Andrew Mortond688abf2007-07-19 01:49:17 -07001897 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07001898 inc_zone_page_state(page, NR_WRITTEN);
1899 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 return ret;
1901}
1902
1903int test_set_page_writeback(struct page *page)
1904{
1905 struct address_space *mapping = page_mapping(page);
1906 int ret;
1907
1908 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001909 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 unsigned long flags;
1911
Nick Piggin19fd6232008-07-25 19:45:32 -07001912 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001914 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 radix_tree_tag_set(&mapping->page_tree,
1916 page_index(page),
1917 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001918 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001919 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 if (!PageDirty(page))
1922 radix_tree_tag_clear(&mapping->page_tree,
1923 page_index(page),
1924 PAGECACHE_TAG_DIRTY);
Jan Karaf446daa2010-08-09 17:19:12 -07001925 radix_tree_tag_clear(&mapping->page_tree,
1926 page_index(page),
1927 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07001928 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 } else {
1930 ret = TestSetPageWriteback(page);
1931 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001932 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07001933 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 return ret;
1935
1936}
1937EXPORT_SYMBOL(test_set_page_writeback);
1938
1939/*
Nick Piggin00128182007-10-16 01:24:40 -07001940 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 * passed tag.
1942 */
1943int mapping_tagged(struct address_space *mapping, int tag)
1944{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07001945 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946}
1947EXPORT_SYMBOL(mapping_tagged);