blob: 66a0024becd9f6c87f5e437c92ce6852657638a9 [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>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040015#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#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>
Al Viroff01bb42011-09-16 02:31:11 -040035#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Jan Karaeb608e32012-05-24 18:59:11 +020037#include <linux/timer.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060038#include <linux/sched/rt.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100039#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060042 * Sleep at most 200ms at a time in balance_dirty_pages().
43 */
44#define MAX_PAUSE max(HZ/5, 1)
45
46/*
Wu Fengguang5b9b3572011-12-06 13:17:17 -060047 * Try to keep balance_dirty_pages() call intervals higher than this many pages
48 * by raising pause time to max_pause when falls below it.
49 */
50#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
51
52/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060053 * Estimate write bandwidth at 200ms intervals.
54 */
55#define BANDWIDTH_INTERVAL max(HZ/5, 1)
56
Wu Fengguang6c14ae12011-03-02 16:04:18 -060057#define RATELIMIT_CALC_SHIFT 10
58
Wu Fengguange98be2d2010-08-29 11:22:30 -060059/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070060 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
61 * will look to see if it needs to force writeback or throttling.
62 */
63static long ratelimit_pages = 32;
64
Linus Torvalds1da177e2005-04-16 15:20:36 -070065/* The following parameters are exported via /proc/sys/vm */
66
67/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020068 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080070int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070071
72/*
David Rientjes2da02992009-01-06 14:39:31 -080073 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
74 * dirty_background_ratio * the amount of dirtyable memory
75 */
76unsigned long dirty_background_bytes;
77
78/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080079 * free highmem will not be subtracted from the total free memory
80 * for calculating free ratios if vm_highmem_is_dirtyable is true
81 */
82int vm_highmem_is_dirtyable;
83
84/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070085 * The generator of dirty data starts writeback at this percentage
86 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080087int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
89/*
David Rientjes2da02992009-01-06 14:39:31 -080090 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
91 * vm_dirty_ratio * the amount of dirtyable memory
92 */
93unsigned long vm_dirty_bytes;
94
95/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070096 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -070097 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070098unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Artem Bityutskiy91913a22012-03-21 22:33:00 -0400100EXPORT_SYMBOL_GPL(dirty_writeback_interval);
101
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700103 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700105unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
107/*
108 * Flag that makes the machine dump writes/reads and block dirtyings.
109 */
110int block_dump;
111
112/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800113 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
114 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115 */
116int laptop_mode;
117
118EXPORT_SYMBOL(laptop_mode);
119
120/* End of sysctl-exported parameters */
121
Wu Fengguangc42843f2011-03-02 15:54:09 -0600122unsigned long global_dirty_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700125 * Scale the writeback cache size proportional to the relative writeout speeds.
126 *
127 * We do this by keeping a floating proportion between BDIs, based on page
128 * writeback completions [end_page_writeback()]. Those devices that write out
129 * pages fastest will get the larger share, while the slower will get a smaller
130 * share.
131 *
132 * We use page writeout completions because we are interested in getting rid of
133 * dirty pages. Having them written out is the primary goal.
134 *
135 * We introduce a concept of time, a period over which we measure these events,
136 * because demand can/will vary over time. The length of this period itself is
137 * measured in page writeback completions.
138 *
139 */
Jan Karaeb608e32012-05-24 18:59:11 +0200140static struct fprop_global writeout_completions;
141
142static void writeout_period(unsigned long t);
143/* Timer for aging of writeout_completions */
144static struct timer_list writeout_period_timer =
145 TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
146static unsigned long writeout_period_time = 0;
147
148/*
149 * Length of period for aging writeout fractions of bdis. This is an
150 * arbitrarily chosen number. The longer the period, the slower fractions will
151 * reflect changes in current writeout rate.
152 */
153#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700154
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700155/*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800156 * Work out the current dirty-memory clamping and background writeout
157 * thresholds.
158 *
159 * The main aim here is to lower them aggressively if there is a lot of mapped
160 * memory around. To avoid stressing page reclaim with lots of unreclaimable
161 * pages. It is better to clamp down on writers than to start swapping, and
162 * performing lots of scanning.
163 *
164 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
165 *
166 * We don't permit the clamping level to fall below 5% - that is getting rather
167 * excessive.
168 *
169 * We make sure that the background writeout level is below the adjusted
170 * clamping level.
171 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800172
Johannes Weinera756cf52012-01-10 15:07:49 -0800173/*
174 * In a memory zone, there is a certain amount of pages we consider
175 * available for the page cache, which is essentially the number of
176 * free and reclaimable pages, minus some zone reserves to protect
177 * lowmem and the ability to uphold the zone's watermarks without
178 * requiring writeback.
179 *
180 * This number of dirtyable pages is the base value of which the
181 * user-configurable dirty ratio is the effictive number of pages that
182 * are allowed to be actually dirtied. Per individual zone, or
183 * globally by using the sum of dirtyable pages over all zones.
184 *
185 * Because the user is allowed to specify the dirty limit globally as
186 * absolute number of bytes, calculating the per-zone dirty limit can
187 * require translating the configured limit into a percentage of
188 * global dirtyable memory first.
189 */
190
Johannes Weiner1edf2232012-01-10 15:06:57 -0800191static unsigned long highmem_dirtyable_memory(unsigned long total)
192{
193#ifdef CONFIG_HIGHMEM
194 int node;
195 unsigned long x = 0;
196
197 for_each_node_state(node, N_HIGH_MEMORY) {
198 struct zone *z =
199 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
200
201 x += zone_page_state(z, NR_FREE_PAGES) +
Johannes Weinerab8fabd2012-01-10 15:07:42 -0800202 zone_reclaimable_pages(z) - z->dirty_balance_reserve;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800203 }
204 /*
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800205 * Unreclaimable memory (kernel memory or anonymous memory
206 * without swap) can bring down the dirtyable pages below
207 * the zone's dirty balance reserve and the above calculation
208 * will underflow. However we still want to add in nodes
209 * which are below threshold (negative values) to get a more
210 * accurate calculation but make sure that the total never
211 * underflows.
212 */
213 if ((long)x < 0)
214 x = 0;
215
216 /*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800217 * Make sure that the number of highmem pages is never larger
218 * than the number of the total dirtyable memory. This can only
219 * occur in very strange VM situations but we want to make sure
220 * that this does not occur.
221 */
222 return min(x, total);
223#else
224 return 0;
225#endif
226}
227
228/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800229 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800230 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800231 * Returns the global number of pages potentially available for dirty
232 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800233 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700234static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800235{
236 unsigned long x;
237
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800238 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
239 x -= min(x, dirty_balance_reserve);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800240
241 if (!vm_highmem_is_dirtyable)
242 x -= highmem_dirtyable_memory(x);
243
244 return x + 1; /* Ensure that we never return 0 */
245}
246
247/*
Johannes Weinerccafa282012-01-10 15:07:44 -0800248 * global_dirty_limits - background-writeback and dirty-throttling thresholds
249 *
250 * Calculate the dirty thresholds based on sysctl parameters
251 * - vm.dirty_background_ratio or vm.dirty_background_bytes
252 * - vm.dirty_ratio or vm.dirty_bytes
253 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
254 * real-time tasks.
255 */
256void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
257{
258 unsigned long background;
259 unsigned long dirty;
260 unsigned long uninitialized_var(available_memory);
261 struct task_struct *tsk;
262
263 if (!vm_dirty_bytes || !dirty_background_bytes)
264 available_memory = global_dirtyable_memory();
265
266 if (vm_dirty_bytes)
267 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
268 else
269 dirty = (vm_dirty_ratio * available_memory) / 100;
270
271 if (dirty_background_bytes)
272 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
273 else
274 background = (dirty_background_ratio * available_memory) / 100;
275
276 if (background >= dirty)
277 background = dirty / 2;
278 tsk = current;
279 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
280 background += background / 4;
281 dirty += dirty / 4;
282 }
283 *pbackground = background;
284 *pdirty = dirty;
285 trace_global_dirty_state(background, dirty);
286}
287
Johannes Weinera756cf52012-01-10 15:07:49 -0800288/**
289 * zone_dirtyable_memory - number of dirtyable pages in a zone
290 * @zone: the zone
291 *
292 * Returns the zone's number of pages potentially available for dirty
293 * page cache. This is the base value for the per-zone dirty limits.
294 */
295static unsigned long zone_dirtyable_memory(struct zone *zone)
296{
297 /*
298 * The effective global number of dirtyable pages may exclude
299 * highmem as a big-picture measure to keep the ratio between
300 * dirty memory and lowmem reasonable.
301 *
302 * But this function is purely about the individual zone and a
303 * highmem zone can hold its share of dirty pages, so we don't
304 * care about vm_highmem_is_dirtyable here.
305 */
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800306 unsigned long nr_pages = zone_page_state(zone, NR_FREE_PAGES) +
307 zone_reclaimable_pages(zone);
308
309 /* don't allow this to underflow */
310 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
311 return nr_pages;
Johannes Weinera756cf52012-01-10 15:07:49 -0800312}
313
314/**
315 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
316 * @zone: the zone
317 *
318 * Returns the maximum number of dirty pages allowed in a zone, based
319 * on the zone's dirtyable memory.
320 */
321static unsigned long zone_dirty_limit(struct zone *zone)
322{
323 unsigned long zone_memory = zone_dirtyable_memory(zone);
324 struct task_struct *tsk = current;
325 unsigned long dirty;
326
327 if (vm_dirty_bytes)
328 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
329 zone_memory / global_dirtyable_memory();
330 else
331 dirty = vm_dirty_ratio * zone_memory / 100;
332
333 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
334 dirty += dirty / 4;
335
336 return dirty;
337}
338
339/**
340 * zone_dirty_ok - tells whether a zone is within its dirty limits
341 * @zone: the zone to check
342 *
343 * Returns %true when the dirty pages in @zone are within the zone's
344 * dirty limit, %false if the limit is exceeded.
345 */
346bool zone_dirty_ok(struct zone *zone)
347{
348 unsigned long limit = zone_dirty_limit(zone);
349
350 return zone_page_state(zone, NR_FILE_DIRTY) +
351 zone_page_state(zone, NR_UNSTABLE_NFS) +
352 zone_page_state(zone, NR_WRITEBACK) <= limit;
353}
354
David Rientjes2da02992009-01-06 14:39:31 -0800355int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700356 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800357 loff_t *ppos)
358{
359 int ret;
360
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700361 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800362 if (ret == 0 && write)
363 dirty_background_bytes = 0;
364 return ret;
365}
366
367int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700368 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800369 loff_t *ppos)
370{
371 int ret;
372
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700373 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800374 if (ret == 0 && write)
375 dirty_background_ratio = 0;
376 return ret;
377}
378
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700379int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700380 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700381 loff_t *ppos)
382{
383 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800384 int ret;
385
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700386 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700387 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200388 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800389 vm_dirty_bytes = 0;
390 }
391 return ret;
392}
393
David Rientjes2da02992009-01-06 14:39:31 -0800394int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700395 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800396 loff_t *ppos)
397{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800398 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800399 int ret;
400
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700401 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800402 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200403 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800404 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700405 }
406 return ret;
407}
408
Jan Karaeb608e32012-05-24 18:59:11 +0200409static unsigned long wp_next_time(unsigned long cur_time)
410{
411 cur_time += VM_COMPLETIONS_PERIOD_LEN;
412 /* 0 has a special meaning... */
413 if (!cur_time)
414 return 1;
415 return cur_time;
416}
417
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700418/*
419 * Increment the BDI's writeout completion count and the global writeout
420 * completion count. Called from test_clear_page_writeback().
421 */
422static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
423{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600424 __inc_bdi_stat(bdi, BDI_WRITTEN);
Jan Karaeb608e32012-05-24 18:59:11 +0200425 __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
426 bdi->max_prop_frac);
427 /* First event after period switching was turned off? */
428 if (!unlikely(writeout_period_time)) {
429 /*
430 * We can race with other __bdi_writeout_inc calls here but
431 * it does not cause any harm since the resulting time when
432 * timer will fire and what is in writeout_period_time will be
433 * roughly the same.
434 */
435 writeout_period_time = wp_next_time(jiffies);
436 mod_timer(&writeout_period_timer, writeout_period_time);
437 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700438}
439
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700440void bdi_writeout_inc(struct backing_dev_info *bdi)
441{
442 unsigned long flags;
443
444 local_irq_save(flags);
445 __bdi_writeout_inc(bdi);
446 local_irq_restore(flags);
447}
448EXPORT_SYMBOL_GPL(bdi_writeout_inc);
449
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700450/*
451 * Obtain an accurate fraction of the BDI's portion.
452 */
453static void bdi_writeout_fraction(struct backing_dev_info *bdi,
454 long *numerator, long *denominator)
455{
Jan Karaeb608e32012-05-24 18:59:11 +0200456 fprop_fraction_percpu(&writeout_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700457 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700458}
459
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700460/*
Jan Karaeb608e32012-05-24 18:59:11 +0200461 * On idle system, we can be called long after we scheduled because we use
462 * deferred timers so count with missed periods.
463 */
464static void writeout_period(unsigned long t)
465{
466 int miss_periods = (jiffies - writeout_period_time) /
467 VM_COMPLETIONS_PERIOD_LEN;
468
469 if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
470 writeout_period_time = wp_next_time(writeout_period_time +
471 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
472 mod_timer(&writeout_period_timer, writeout_period_time);
473 } else {
474 /*
475 * Aging has zeroed all fractions. Stop wasting CPU on period
476 * updates.
477 */
478 writeout_period_time = 0;
479 }
480}
481
482/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700483 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
484 * registered backing devices, which, for obvious reasons, can not
485 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700486 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700487static unsigned int bdi_min_ratio;
488
489int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
490{
491 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700492
Jens Axboecfc4ba52009-09-14 13:12:40 +0200493 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700494 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700495 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700496 } else {
497 min_ratio -= bdi->min_ratio;
498 if (bdi_min_ratio + min_ratio < 100) {
499 bdi_min_ratio += min_ratio;
500 bdi->min_ratio += min_ratio;
501 } else {
502 ret = -EINVAL;
503 }
504 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200505 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700506
507 return ret;
508}
509
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700510int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
511{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700512 int ret = 0;
513
514 if (max_ratio > 100)
515 return -EINVAL;
516
Jens Axboecfc4ba52009-09-14 13:12:40 +0200517 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700518 if (bdi->min_ratio > max_ratio) {
519 ret = -EINVAL;
520 } else {
521 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200522 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700523 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200524 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700525
526 return ret;
527}
528EXPORT_SYMBOL(bdi_set_max_ratio);
529
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600530static unsigned long dirty_freerun_ceiling(unsigned long thresh,
531 unsigned long bg_thresh)
532{
533 return (thresh + bg_thresh) / 2;
534}
535
Wu Fengguangffd1f602011-06-19 22:18:42 -0600536static unsigned long hard_dirty_limit(unsigned long thresh)
537{
538 return max(thresh, global_dirty_limit);
539}
540
Wu Fengguang6f718652011-03-02 17:14:34 -0600541/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700542 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600543 * @bdi: the backing_dev_info to query
544 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700545 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600546 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
547 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600548 *
549 * Note that balance_dirty_pages() will only seriously take it as a hard limit
550 * when sleeping max_pause per page is not enough to keep the dirty pages under
551 * control. For example, when the device is completely stalled due to some error
552 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
553 * In the other normal situations, it acts more gently by throttling the tasks
554 * more (rather than completely block them) when the bdi dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600555 *
556 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700557 * - starving fast devices
558 * - piling up dirty pages (that will take long time to sync) on slow devices
559 *
560 * The bdi's share of dirty limit will be adapting to its throughput and
561 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
562 */
563unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700564{
565 u64 bdi_dirty;
566 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700567
Wu Fengguang16c40422010-08-11 14:17:39 -0700568 /*
569 * Calculate this BDI's share of the dirty ratio.
570 */
571 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700572
Wu Fengguang16c40422010-08-11 14:17:39 -0700573 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
574 bdi_dirty *= numerator;
575 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700576
Wu Fengguang16c40422010-08-11 14:17:39 -0700577 bdi_dirty += (dirty * bdi->min_ratio) / 100;
578 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
579 bdi_dirty = dirty * bdi->max_ratio / 100;
580
581 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582}
583
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600584/*
585 * Dirty position control.
586 *
587 * (o) global/bdi setpoints
588 *
589 * We want the dirty pages be balanced around the global/bdi setpoints.
590 * When the number of dirty pages is higher/lower than the setpoint, the
591 * dirty position control ratio (and hence task dirty ratelimit) will be
592 * decreased/increased to bring the dirty pages back to the setpoint.
593 *
594 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
595 *
596 * if (dirty < setpoint) scale up pos_ratio
597 * if (dirty > setpoint) scale down pos_ratio
598 *
599 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
600 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
601 *
602 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
603 *
604 * (o) global control line
605 *
606 * ^ pos_ratio
607 * |
608 * | |<===== global dirty control scope ======>|
609 * 2.0 .............*
610 * | .*
611 * | . *
612 * | . *
613 * | . *
614 * | . *
615 * | . *
616 * 1.0 ................................*
617 * | . . *
618 * | . . *
619 * | . . *
620 * | . . *
621 * | . . *
622 * 0 +------------.------------------.----------------------*------------->
623 * freerun^ setpoint^ limit^ dirty pages
624 *
625 * (o) bdi control line
626 *
627 * ^ pos_ratio
628 * |
629 * | *
630 * | *
631 * | *
632 * | *
633 * | * |<=========== span ============>|
634 * 1.0 .......................*
635 * | . *
636 * | . *
637 * | . *
638 * | . *
639 * | . *
640 * | . *
641 * | . *
642 * | . *
643 * | . *
644 * | . *
645 * | . *
646 * 1/4 ...............................................* * * * * * * * * * * *
647 * | . .
648 * | . .
649 * | . .
650 * 0 +----------------------.-------------------------------.------------->
651 * bdi_setpoint^ x_intercept^
652 *
653 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
654 * be smoothly throttled down to normal if it starts high in situations like
655 * - start writing to a slow SD card and a fast disk at the same time. The SD
656 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
657 * - the bdi dirty thresh drops quickly due to change of JBOD workload
658 */
659static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
660 unsigned long thresh,
661 unsigned long bg_thresh,
662 unsigned long dirty,
663 unsigned long bdi_thresh,
664 unsigned long bdi_dirty)
665{
666 unsigned long write_bw = bdi->avg_write_bandwidth;
667 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
668 unsigned long limit = hard_dirty_limit(thresh);
669 unsigned long x_intercept;
670 unsigned long setpoint; /* dirty pages' target balance point */
671 unsigned long bdi_setpoint;
672 unsigned long span;
673 long long pos_ratio; /* for scaling up/down the rate limit */
674 long x;
675
676 if (unlikely(dirty >= limit))
677 return 0;
678
679 /*
680 * global setpoint
681 *
682 * setpoint - dirty 3
683 * f(dirty) := 1.0 + (----------------)
684 * limit - setpoint
685 *
686 * it's a 3rd order polynomial that subjects to
687 *
688 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
689 * (2) f(setpoint) = 1.0 => the balance point
690 * (3) f(limit) = 0 => the hard limit
691 * (4) df/dx <= 0 => negative feedback control
692 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
693 * => fast response on large errors; small oscillation near setpoint
694 */
695 setpoint = (freerun + limit) / 2;
696 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
697 limit - setpoint + 1);
698 pos_ratio = x;
699 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
700 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
701 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
702
703 /*
704 * We have computed basic pos_ratio above based on global situation. If
705 * the bdi is over/under its share of dirty pages, we want to scale
706 * pos_ratio further down/up. That is done by the following mechanism.
707 */
708
709 /*
710 * bdi setpoint
711 *
712 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
713 *
714 * x_intercept - bdi_dirty
715 * := --------------------------
716 * x_intercept - bdi_setpoint
717 *
718 * The main bdi control line is a linear function that subjects to
719 *
720 * (1) f(bdi_setpoint) = 1.0
721 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
722 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
723 *
724 * For single bdi case, the dirty pages are observed to fluctuate
725 * regularly within range
726 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
727 * for various filesystems, where (2) can yield in a reasonable 12.5%
728 * fluctuation range for pos_ratio.
729 *
730 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
731 * own size, so move the slope over accordingly and choose a slope that
732 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
733 */
734 if (unlikely(bdi_thresh > thresh))
735 bdi_thresh = thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600736 /*
737 * It's very possible that bdi_thresh is close to 0 not because the
738 * device is slow, but that it has remained inactive for long time.
739 * Honour such devices a reasonable good (hopefully IO efficient)
740 * threshold, so that the occasional writes won't be blocked and active
741 * writes can rampup the threshold quickly.
742 */
Wu Fengguang8927f662011-08-04 22:16:46 -0600743 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600744 /*
745 * scale global setpoint to bdi's:
746 * bdi_setpoint = setpoint * bdi_thresh / thresh
747 */
748 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
749 bdi_setpoint = setpoint * (u64)x >> 16;
750 /*
751 * Use span=(8*write_bw) in single bdi case as indicated by
752 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
753 *
754 * bdi_thresh thresh - bdi_thresh
755 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
756 * thresh thresh
757 */
758 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
759 x_intercept = bdi_setpoint + span;
760
761 if (bdi_dirty < x_intercept - span / 4) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600762 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
763 x_intercept - bdi_setpoint + 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600764 } else
765 pos_ratio /= 4;
766
Wu Fengguang8927f662011-08-04 22:16:46 -0600767 /*
768 * bdi reserve area, safeguard against dirty pool underrun and disk idle
769 * It may push the desired control point of global dirty pages higher
770 * than setpoint.
771 */
772 x_intercept = bdi_thresh / 2;
773 if (bdi_dirty < x_intercept) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600774 if (bdi_dirty > x_intercept / 8)
775 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
776 else
Wu Fengguang8927f662011-08-04 22:16:46 -0600777 pos_ratio *= 8;
778 }
779
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600780 return pos_ratio;
781}
782
Wu Fengguange98be2d2010-08-29 11:22:30 -0600783static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
784 unsigned long elapsed,
785 unsigned long written)
786{
787 const unsigned long period = roundup_pow_of_two(3 * HZ);
788 unsigned long avg = bdi->avg_write_bandwidth;
789 unsigned long old = bdi->write_bandwidth;
790 u64 bw;
791
792 /*
793 * bw = written * HZ / elapsed
794 *
795 * bw * elapsed + write_bandwidth * (period - elapsed)
796 * write_bandwidth = ---------------------------------------------------
797 * period
798 */
799 bw = written - bdi->written_stamp;
800 bw *= HZ;
801 if (unlikely(elapsed > period)) {
802 do_div(bw, elapsed);
803 avg = bw;
804 goto out;
805 }
806 bw += (u64)bdi->write_bandwidth * (period - elapsed);
807 bw >>= ilog2(period);
808
809 /*
810 * one more level of smoothing, for filtering out sudden spikes
811 */
812 if (avg > old && old >= (unsigned long)bw)
813 avg -= (avg - old) >> 3;
814
815 if (avg < old && old <= (unsigned long)bw)
816 avg += (old - avg) >> 3;
817
818out:
819 bdi->write_bandwidth = bw;
820 bdi->avg_write_bandwidth = avg;
821}
822
Wu Fengguangc42843f2011-03-02 15:54:09 -0600823/*
824 * The global dirtyable memory and dirty threshold could be suddenly knocked
825 * down by a large amount (eg. on the startup of KVM in a swapless system).
826 * This may throw the system into deep dirty exceeded state and throttle
827 * heavy/light dirtiers alike. To retain good responsiveness, maintain
828 * global_dirty_limit for tracking slowly down to the knocked down dirty
829 * threshold.
830 */
831static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
832{
833 unsigned long limit = global_dirty_limit;
834
835 /*
836 * Follow up in one step.
837 */
838 if (limit < thresh) {
839 limit = thresh;
840 goto update;
841 }
842
843 /*
844 * Follow down slowly. Use the higher one as the target, because thresh
845 * may drop below dirty. This is exactly the reason to introduce
846 * global_dirty_limit which is guaranteed to lie above the dirty pages.
847 */
848 thresh = max(thresh, dirty);
849 if (limit > thresh) {
850 limit -= (limit - thresh) >> 5;
851 goto update;
852 }
853 return;
854update:
855 global_dirty_limit = limit;
856}
857
858static void global_update_bandwidth(unsigned long thresh,
859 unsigned long dirty,
860 unsigned long now)
861{
862 static DEFINE_SPINLOCK(dirty_lock);
863 static unsigned long update_time;
864
865 /*
866 * check locklessly first to optimize away locking for the most time
867 */
868 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
869 return;
870
871 spin_lock(&dirty_lock);
872 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
873 update_dirty_limit(thresh, dirty);
874 update_time = now;
875 }
876 spin_unlock(&dirty_lock);
877}
878
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600879/*
880 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
881 *
882 * Normal bdi tasks will be curbed at or below it in long term.
883 * Obviously it should be around (write_bw / N) when there are N dd tasks.
884 */
885static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
886 unsigned long thresh,
887 unsigned long bg_thresh,
888 unsigned long dirty,
889 unsigned long bdi_thresh,
890 unsigned long bdi_dirty,
891 unsigned long dirtied,
892 unsigned long elapsed)
893{
Wu Fengguang73811312011-08-26 15:53:24 -0600894 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
895 unsigned long limit = hard_dirty_limit(thresh);
896 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600897 unsigned long write_bw = bdi->avg_write_bandwidth;
898 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
899 unsigned long dirty_rate;
900 unsigned long task_ratelimit;
901 unsigned long balanced_dirty_ratelimit;
902 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600903 unsigned long step;
904 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600905
906 /*
907 * The dirty rate will match the writeout rate in long term, except
908 * when dirty pages are truncated by userspace or re-dirtied by FS.
909 */
910 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
911
912 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
913 bdi_thresh, bdi_dirty);
914 /*
915 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
916 */
917 task_ratelimit = (u64)dirty_ratelimit *
918 pos_ratio >> RATELIMIT_CALC_SHIFT;
919 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
920
921 /*
922 * A linear estimation of the "balanced" throttle rate. The theory is,
923 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
924 * dirty_rate will be measured to be (N * task_ratelimit). So the below
925 * formula will yield the balanced rate limit (write_bw / N).
926 *
927 * Note that the expanded form is not a pure rate feedback:
928 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
929 * but also takes pos_ratio into account:
930 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
931 *
932 * (1) is not realistic because pos_ratio also takes part in balancing
933 * the dirty rate. Consider the state
934 * pos_ratio = 0.5 (3)
935 * rate = 2 * (write_bw / N) (4)
936 * If (1) is used, it will stuck in that state! Because each dd will
937 * be throttled at
938 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
939 * yielding
940 * dirty_rate = N * task_ratelimit = write_bw (6)
941 * put (6) into (1) we get
942 * rate_(i+1) = rate_(i) (7)
943 *
944 * So we end up using (2) to always keep
945 * rate_(i+1) ~= (write_bw / N) (8)
946 * regardless of the value of pos_ratio. As long as (8) is satisfied,
947 * pos_ratio is able to drive itself to 1.0, which is not only where
948 * the dirty count meet the setpoint, but also where the slope of
949 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
950 */
951 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
952 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -0600953 /*
954 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
955 */
956 if (unlikely(balanced_dirty_ratelimit > write_bw))
957 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600958
Wu Fengguang73811312011-08-26 15:53:24 -0600959 /*
960 * We could safely do this and return immediately:
961 *
962 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
963 *
964 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +0800965 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -0600966 * limit the step size.
967 *
968 * The below code essentially only uses the relative value of
969 *
970 * task_ratelimit - dirty_ratelimit
971 * = (pos_ratio - 1) * dirty_ratelimit
972 *
973 * which reflects the direction and size of dirty position error.
974 */
975
976 /*
977 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
978 * task_ratelimit is on the same side of dirty_ratelimit, too.
979 * For example, when
980 * - dirty_ratelimit > balanced_dirty_ratelimit
981 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
982 * lowering dirty_ratelimit will help meet both the position and rate
983 * control targets. Otherwise, don't update dirty_ratelimit if it will
984 * only help meet the rate target. After all, what the users ultimately
985 * feel and care are stable dirty rate and small position error.
986 *
987 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +0800988 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -0600989 * keeps jumping around randomly and can even leap far away at times
990 * due to the small 200ms estimation period of dirty_rate (we want to
991 * keep that period small to reduce time lags).
992 */
993 step = 0;
994 if (dirty < setpoint) {
995 x = min(bdi->balanced_dirty_ratelimit,
996 min(balanced_dirty_ratelimit, task_ratelimit));
997 if (dirty_ratelimit < x)
998 step = x - dirty_ratelimit;
999 } else {
1000 x = max(bdi->balanced_dirty_ratelimit,
1001 max(balanced_dirty_ratelimit, task_ratelimit));
1002 if (dirty_ratelimit > x)
1003 step = dirty_ratelimit - x;
1004 }
1005
1006 /*
1007 * Don't pursue 100% rate matching. It's impossible since the balanced
1008 * rate itself is constantly fluctuating. So decrease the track speed
1009 * when it gets close to the target. Helps eliminate pointless tremors.
1010 */
1011 step >>= dirty_ratelimit / (2 * step + 1);
1012 /*
1013 * Limit the tracking speed to avoid overshooting.
1014 */
1015 step = (step + 7) / 8;
1016
1017 if (dirty_ratelimit < balanced_dirty_ratelimit)
1018 dirty_ratelimit += step;
1019 else
1020 dirty_ratelimit -= step;
1021
1022 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1023 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001024
1025 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001026}
1027
Wu Fengguange98be2d2010-08-29 11:22:30 -06001028void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001029 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001030 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001031 unsigned long dirty,
1032 unsigned long bdi_thresh,
1033 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -06001034 unsigned long start_time)
1035{
1036 unsigned long now = jiffies;
1037 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001038 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001039 unsigned long written;
1040
1041 /*
1042 * rate-limit, only update once every 200ms.
1043 */
1044 if (elapsed < BANDWIDTH_INTERVAL)
1045 return;
1046
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001047 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001048 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1049
1050 /*
1051 * Skip quiet periods when disk bandwidth is under-utilized.
1052 * (at least 1s idle time between two flusher runs)
1053 */
1054 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1055 goto snapshot;
1056
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001057 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -06001058 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001059 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1060 bdi_thresh, bdi_dirty,
1061 dirtied, elapsed);
1062 }
Wu Fengguange98be2d2010-08-29 11:22:30 -06001063 bdi_update_write_bandwidth(bdi, elapsed, written);
1064
1065snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001066 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001067 bdi->written_stamp = written;
1068 bdi->bw_time_stamp = now;
1069}
1070
1071static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001072 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001073 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001074 unsigned long dirty,
1075 unsigned long bdi_thresh,
1076 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -06001077 unsigned long start_time)
1078{
1079 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1080 return;
1081 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001082 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1083 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001084 spin_unlock(&bdi->wb.list_lock);
1085}
1086
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001088 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001089 * will look to see if it needs to start dirty throttling.
1090 *
1091 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1092 * global_page_state() too often. So scale it near-sqrt to the safety margin
1093 * (the number of pages we may dirty without exceeding the dirty limits).
1094 */
1095static unsigned long dirty_poll_interval(unsigned long dirty,
1096 unsigned long thresh)
1097{
1098 if (thresh > dirty)
1099 return 1UL << (ilog2(thresh - dirty) >> 1);
1100
1101 return 1;
1102}
1103
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001104static long bdi_max_pause(struct backing_dev_info *bdi,
1105 unsigned long bdi_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001106{
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001107 long bw = bdi->avg_write_bandwidth;
1108 long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001109
1110 /*
1111 * Limit pause time for small memory systems. If sleeping for too long
1112 * time, a small pool of dirty/writeback pages may go empty and disk go
1113 * idle.
1114 *
1115 * 8 serves as the safety ratio.
1116 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001117 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1118 t++;
1119
1120 return min_t(long, t, MAX_PAUSE);
1121}
1122
1123static long bdi_min_pause(struct backing_dev_info *bdi,
1124 long max_pause,
1125 unsigned long task_ratelimit,
1126 unsigned long dirty_ratelimit,
1127 int *nr_dirtied_pause)
1128{
1129 long hi = ilog2(bdi->avg_write_bandwidth);
1130 long lo = ilog2(bdi->dirty_ratelimit);
1131 long t; /* target pause */
1132 long pause; /* estimated next pause */
1133 int pages; /* target nr_dirtied_pause */
1134
1135 /* target for 10ms pause on 1-dd case */
1136 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001137
1138 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001139 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1140 * overheads.
1141 *
1142 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001143 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001144 if (hi > lo)
1145 t += (hi - lo) * (10 * HZ) / 1024;
1146
1147 /*
1148 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1149 * on the much more stable dirty_ratelimit. However the next pause time
1150 * will be computed based on task_ratelimit and the two rate limits may
1151 * depart considerably at some time. Especially if task_ratelimit goes
1152 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1153 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1154 * result task_ratelimit won't be executed faithfully, which could
1155 * eventually bring down dirty_ratelimit.
1156 *
1157 * We apply two rules to fix it up:
1158 * 1) try to estimate the next pause time and if necessary, use a lower
1159 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1160 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1161 * 2) limit the target pause time to max_pause/2, so that the normal
1162 * small fluctuations of task_ratelimit won't trigger rule (1) and
1163 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1164 */
1165 t = min(t, 1 + max_pause / 2);
1166 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1167
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001168 /*
1169 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1170 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1171 * When the 16 consecutive reads are often interrupted by some dirty
1172 * throttling pause during the async writes, cfq will go into idles
1173 * (deadline is fine). So push nr_dirtied_pause as high as possible
1174 * until reaches DIRTY_POLL_THRESH=32 pages.
1175 */
1176 if (pages < DIRTY_POLL_THRESH) {
1177 t = max_pause;
1178 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1179 if (pages > DIRTY_POLL_THRESH) {
1180 pages = DIRTY_POLL_THRESH;
1181 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1182 }
1183 }
1184
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001185 pause = HZ * pages / (task_ratelimit + 1);
1186 if (pause > max_pause) {
1187 t = max_pause;
1188 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1189 }
1190
1191 *nr_dirtied_pause = pages;
1192 /*
1193 * The minimal pause time will normally be half the target pause time.
1194 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001195 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001196}
1197
Wu Fengguang9d823e82011-06-11 18:10:12 -06001198/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 * balance_dirty_pages() must be called by processes which are generating dirty
1200 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001201 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001202 * If we're over `background_thresh' then the writeback threads are woken to
1203 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001205static void balance_dirty_pages(struct address_space *mapping,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001206 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207{
Wu Fengguang143dfe82010-08-27 18:45:12 -06001208 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1209 unsigned long bdi_reclaimable;
Wu Fengguang77627412010-09-12 13:34:05 -06001210 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1211 unsigned long bdi_dirty;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001212 unsigned long freerun;
David Rientjes364aeb22009-01-06 14:39:29 -08001213 unsigned long background_thresh;
1214 unsigned long dirty_thresh;
1215 unsigned long bdi_thresh;
Wu Fengguang83712352011-06-11 19:25:42 -06001216 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001217 long pause;
1218 long max_pause;
1219 long min_pause;
1220 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001221 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001222 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001223 unsigned long dirty_ratelimit;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001224 unsigned long pos_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001226 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227
1228 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001229 unsigned long now = jiffies;
1230
Wu Fengguang143dfe82010-08-27 18:45:12 -06001231 /*
1232 * Unstable writes are a feature of certain networked
1233 * filesystems (i.e. NFS) in which data may have been
1234 * written to the server's write cache, but has not yet
1235 * been flushed to permanent storage.
1236 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001237 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1238 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -06001239 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001240
Wu Fengguang16c40422010-08-11 14:17:39 -07001241 global_dirty_limits(&background_thresh, &dirty_thresh);
1242
1243 /*
1244 * Throttle it only when the background writeback cannot
1245 * catch-up. This avoids (excessively) small writeouts
1246 * when the bdi limits are ramping up.
1247 */
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001248 freerun = dirty_freerun_ceiling(dirty_thresh,
1249 background_thresh);
Wu Fengguang83712352011-06-11 19:25:42 -06001250 if (nr_dirty <= freerun) {
1251 current->dirty_paused_when = now;
1252 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001253 current->nr_dirtied_pause =
1254 dirty_poll_interval(nr_dirty, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001255 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001256 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001257
Wu Fengguang143dfe82010-08-27 18:45:12 -06001258 if (unlikely(!writeback_in_progress(bdi)))
1259 bdi_start_background_writeback(bdi);
1260
1261 /*
1262 * bdi_thresh is not treated as some limiting factor as
1263 * dirty_thresh, due to reasons
1264 * - in JBOD setup, bdi_thresh can fluctuate a lot
1265 * - in a system with HDD and USB key, the USB key may somehow
1266 * go into state (bdi_dirty >> bdi_thresh) either because
1267 * bdi_dirty starts high, or because bdi_thresh drops low.
1268 * In this case we don't want to hard throttle the USB key
1269 * dirtiers for 100 seconds until bdi_dirty drops under
1270 * bdi_thresh. Instead the auxiliary bdi control line in
1271 * bdi_position_ratio() will let the dirtier task progress
1272 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1273 */
Wu Fengguang16c40422010-08-11 14:17:39 -07001274 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001275
Wu Fengguange50e3722010-08-11 14:17:37 -07001276 /*
1277 * In order to avoid the stacked BDI deadlock we need
1278 * to ensure we accurately count the 'dirty' pages when
1279 * the threshold is low.
1280 *
1281 * Otherwise it would be possible to get thresh+n pages
1282 * reported dirty, even though there are thresh-m pages
1283 * actually dirty; with m+n sitting in the percpu
1284 * deltas.
1285 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001286 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1287 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1288 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001289 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001290 } else {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001291 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1292 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001293 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001294 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001295
Wu Fengguang82791942011-12-03 21:26:01 -06001296 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
Wu Fengguang77627412010-09-12 13:34:05 -06001297 (nr_dirty > dirty_thresh);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001298 if (dirty_exceeded && !bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001299 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001301 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1302 nr_dirty, bdi_thresh, bdi_dirty,
1303 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001304
Wu Fengguang143dfe82010-08-27 18:45:12 -06001305 dirty_ratelimit = bdi->dirty_ratelimit;
1306 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1307 background_thresh, nr_dirty,
1308 bdi_thresh, bdi_dirty);
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001309 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1310 RATELIMIT_CALC_SHIFT;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001311 max_pause = bdi_max_pause(bdi, bdi_dirty);
1312 min_pause = bdi_min_pause(bdi, max_pause,
1313 task_ratelimit, dirty_ratelimit,
1314 &nr_dirtied_pause);
1315
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001316 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001317 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001318 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001319 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 }
Wu Fengguang83712352011-06-11 19:25:42 -06001321 period = HZ * pages_dirtied / task_ratelimit;
1322 pause = period;
1323 if (current->dirty_paused_when)
1324 pause -= now - current->dirty_paused_when;
1325 /*
1326 * For less than 1s think time (ext3/4 may block the dirtier
1327 * for up to 800ms from time to time on 1-HDD; so does xfs,
1328 * however at much less frequency), try to compensate it in
1329 * future periods by updating the virtual time; otherwise just
1330 * do a reset, as it may be a light dirtier.
1331 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001332 if (pause < min_pause) {
Wu Fengguangece13ac2010-08-29 23:33:20 -06001333 trace_balance_dirty_pages(bdi,
1334 dirty_thresh,
1335 background_thresh,
1336 nr_dirty,
1337 bdi_thresh,
1338 bdi_dirty,
1339 dirty_ratelimit,
1340 task_ratelimit,
1341 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001342 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001343 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001344 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001345 if (pause < -HZ) {
1346 current->dirty_paused_when = now;
1347 current->nr_dirtied = 0;
1348 } else if (period) {
1349 current->dirty_paused_when += period;
1350 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001351 } else if (current->nr_dirtied_pause <= pages_dirtied)
1352 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001353 break;
1354 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001355 if (unlikely(pause > max_pause)) {
1356 /* for occasional dropped task_ratelimit */
1357 now += min(pause - max_pause, max_pause);
1358 pause = max_pause;
1359 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001360
1361pause:
Wu Fengguangece13ac2010-08-29 23:33:20 -06001362 trace_balance_dirty_pages(bdi,
1363 dirty_thresh,
1364 background_thresh,
1365 nr_dirty,
1366 bdi_thresh,
1367 bdi_dirty,
1368 dirty_ratelimit,
1369 task_ratelimit,
1370 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001371 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001372 pause,
1373 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001374 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001375 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001376
Wu Fengguang83712352011-06-11 19:25:42 -06001377 current->dirty_paused_when = now + pause;
1378 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001379 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001380
Wu Fengguangffd1f602011-06-19 22:18:42 -06001381 /*
Wu Fengguang1df64712011-11-13 19:47:32 -06001382 * This is typically equal to (nr_dirty < dirty_thresh) and can
1383 * also keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001384 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001385 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001386 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001387
Wu Fengguangc5c63432011-12-02 10:21:33 -06001388 /*
1389 * In the case of an unresponding NFS server and the NFS dirty
1390 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1391 * to go through, so that tasks on them still remain responsive.
1392 *
1393 * In theory 1 page is enough to keep the comsumer-producer
1394 * pipe going: the flusher cleans 1 page => the task dirties 1
1395 * more page. However bdi_dirty has accounting errors. So use
1396 * the larger and more IO friendly bdi_stat_error.
1397 */
1398 if (bdi_dirty <= bdi_stat_error(bdi))
1399 break;
1400
Jan Kara499d05e2011-11-16 19:34:48 +08001401 if (fatal_signal_pending(current))
1402 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 }
1404
Wu Fengguang143dfe82010-08-27 18:45:12 -06001405 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001406 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
1408 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001409 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410
1411 /*
1412 * In laptop mode, we wait until hitting the higher threshold before
1413 * starting background writeout, and then write out all the way down
1414 * to the lower threshold. So slow writers cause minimal disk activity.
1415 *
1416 * In normal mode, we start background writeout at the lower
1417 * background_thresh, to keep the amount of dirty memory low.
1418 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001419 if (laptop_mode)
1420 return;
1421
1422 if (nr_reclaimable > background_thresh)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001423 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424}
1425
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001426void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001427{
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001428 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001429 struct address_space *mapping = page_mapping(page);
1430
1431 if (mapping)
1432 balance_dirty_pages_ratelimited(mapping);
1433 }
1434}
1435
Wu Fengguang9d823e82011-06-11 18:10:12 -06001436static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001437
Wu Fengguang54848d72011-04-05 13:21:19 -06001438/*
1439 * Normal tasks are throttled by
1440 * loop {
1441 * dirty tsk->nr_dirtied_pause pages;
1442 * take a snap in balance_dirty_pages();
1443 * }
1444 * However there is a worst case. If every task exit immediately when dirtied
1445 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1446 * called to throttle the page dirties. The solution is to save the not yet
1447 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1448 * randomly into the running tasks. This works well for the above worst case,
1449 * as the new task will pick up and accumulate the old task's leaked dirty
1450 * count and eventually get throttled.
1451 */
1452DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1453
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001455 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001456 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 *
1458 * Processes which are dirtying memory should call in here once for each page
1459 * which was newly dirtied. The function will periodically check the system's
1460 * dirty state and will initiate writeback if needed.
1461 *
1462 * On really big machines, get_writeback_state is expensive, so try to avoid
1463 * calling it too often (ratelimiting). But once we're over the dirty memory
1464 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1465 * from overshooting the limit by (ratelimit_pages) each.
1466 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001467void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001469 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001470 int ratelimit;
1471 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
Wu Fengguang36715ce2011-06-11 17:53:57 -06001473 if (!bdi_cap_account_dirty(bdi))
1474 return;
1475
Wu Fengguang9d823e82011-06-11 18:10:12 -06001476 ratelimit = current->nr_dirtied_pause;
1477 if (bdi->dirty_exceeded)
1478 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001480 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001481 /*
1482 * This prevents one CPU to accumulate too many dirtied pages without
1483 * calling into balance_dirty_pages(), which can happen when there are
1484 * 1000+ tasks, all of them start dirtying pages at exactly the same
1485 * time, hence all honoured too large initial task->nr_dirtied_pause.
1486 */
Tejun Heo245b2e72009-06-24 15:13:48 +09001487 p = &__get_cpu_var(bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001488 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001489 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001490 else if (unlikely(*p >= ratelimit_pages)) {
1491 *p = 0;
1492 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001494 /*
1495 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1496 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1497 * the dirty throttling and livelock other long-run dirtiers.
1498 */
1499 p = &__get_cpu_var(dirty_throttle_leaks);
1500 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001501 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001502 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1503 *p -= nr_pages_dirtied;
1504 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001506 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001507
1508 if (unlikely(current->nr_dirtied >= ratelimit))
1509 balance_dirty_pages(mapping, current->nr_dirtied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001511EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512
Andrew Morton232ea4d2007-02-28 20:13:21 -08001513void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514{
David Rientjes364aeb22009-01-06 14:39:29 -08001515 unsigned long background_thresh;
1516 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517
1518 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001519 global_dirty_limits(&background_thresh, &dirty_thresh);
Fengguang Wu47a13332012-03-21 16:34:09 -07001520 dirty_thresh = hard_dirty_limit(dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521
1522 /*
1523 * Boost the allowable dirty threshold a bit for page
1524 * allocators so they don't get DoS'ed by heavy writers
1525 */
1526 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1527
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001528 if (global_page_state(NR_UNSTABLE_NFS) +
1529 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1530 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001531 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001532
1533 /*
1534 * The caller might hold locks which can prevent IO completion
1535 * or progress in the filesystem. So we cannot just sit here
1536 * waiting for IO to complete.
1537 */
1538 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1539 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 }
1541}
1542
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1545 */
1546int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001547 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001549 proc_dointvec(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 return 0;
1551}
1552
Jens Axboec2c49862010-05-20 09:18:47 +02001553#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001554void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555{
Matthew Garrett31373d02010-04-06 14:25:14 +02001556 struct request_queue *q = (struct request_queue *)data;
1557 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1558 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559
Matthew Garrett31373d02010-04-06 14:25:14 +02001560 /*
1561 * We want to write everything out, not just down to the dirty
1562 * threshold
1563 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001564 if (bdi_has_dirty_io(&q->backing_dev_info))
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001565 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1566 WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567}
1568
1569/*
1570 * We've spun up the disk and we're in laptop mode: schedule writeback
1571 * of all dirty data a few seconds from now. If the flush is already scheduled
1572 * then push it back - the user is still using the disk.
1573 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001574void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575{
Matthew Garrett31373d02010-04-06 14:25:14 +02001576 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577}
1578
1579/*
1580 * We're in laptop mode and we've just synced. The sync's writes will have
1581 * caused another writeback to be scheduled by laptop_io_completion.
1582 * Nothing needs to be written back anymore, so we unschedule the writeback.
1583 */
1584void laptop_sync_completion(void)
1585{
Matthew Garrett31373d02010-04-06 14:25:14 +02001586 struct backing_dev_info *bdi;
1587
1588 rcu_read_lock();
1589
1590 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1591 del_timer(&bdi->laptop_mode_wb_timer);
1592
1593 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594}
Jens Axboec2c49862010-05-20 09:18:47 +02001595#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596
1597/*
1598 * If ratelimit_pages is too high then we can get into dirty-data overload
1599 * if a large number of processes all perform writes at the same time.
1600 * If it is too low then SMP machines will call the (expensive)
1601 * get_writeback_state too often.
1602 *
1603 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1604 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06001605 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 */
1607
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001608void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609{
Wu Fengguang9d823e82011-06-11 18:10:12 -06001610 unsigned long background_thresh;
1611 unsigned long dirty_thresh;
1612 global_dirty_limits(&background_thresh, &dirty_thresh);
Fengguang Wu68809c72012-05-06 13:21:42 +08001613 global_dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001614 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 if (ratelimit_pages < 16)
1616 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617}
1618
Chandra Seetharaman26c21432006-06-27 02:54:10 -07001619static int __cpuinit
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08001620ratelimit_handler(struct notifier_block *self, unsigned long action,
1621 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622{
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08001623
1624 switch (action & ~CPU_TASKS_FROZEN) {
1625 case CPU_ONLINE:
1626 case CPU_DEAD:
1627 writeback_set_ratelimit();
1628 return NOTIFY_OK;
1629 default:
1630 return NOTIFY_DONE;
1631 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632}
1633
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001634static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 .notifier_call = ratelimit_handler,
1636 .next = NULL,
1637};
1638
1639/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001640 * Called early on to tune the page writeback dirty limits.
1641 *
1642 * We used to scale dirty pages according to how total memory
1643 * related to pages that could be allocated for buffers (by
1644 * comparing nr_free_buffer_pages() to vm_total_pages.
1645 *
1646 * However, that was when we used "dirty_ratio" to scale with
1647 * all memory, and we don't do that any more. "dirty_ratio"
1648 * is now applied to total non-HIGHPAGE memory (by subtracting
1649 * totalhigh_pages from vm_total_pages), and as such we can't
1650 * get into the old insane situation any more where we had
1651 * large amounts of dirty pages compared to a small amount of
1652 * non-HIGHMEM memory.
1653 *
1654 * But we might still want to scale the dirty_ratio by how
1655 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 */
1657void __init page_writeback_init(void)
1658{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001659 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001661
Jan Karaeb608e32012-05-24 18:59:11 +02001662 fprop_global_init(&writeout_completions);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663}
1664
David Howells811d7362006-08-29 19:06:09 +01001665/**
Jan Karaf446daae2010-08-09 17:19:12 -07001666 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1667 * @mapping: address space structure to write
1668 * @start: starting page index
1669 * @end: ending page index (inclusive)
1670 *
1671 * This function scans the page range from @start to @end (inclusive) and tags
1672 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1673 * that write_cache_pages (or whoever calls this function) will then use
1674 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1675 * used to avoid livelocking of writeback by a process steadily creating new
1676 * dirty pages in the file (thus it is important for this function to be quick
1677 * so that it can tag pages faster than a dirtying process can create them).
1678 */
1679/*
1680 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1681 */
Jan Karaf446daae2010-08-09 17:19:12 -07001682void tag_pages_for_writeback(struct address_space *mapping,
1683 pgoff_t start, pgoff_t end)
1684{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001685#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07001686 unsigned long tagged;
1687
1688 do {
1689 spin_lock_irq(&mapping->tree_lock);
1690 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1691 &start, end, WRITEBACK_TAG_BATCH,
1692 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1693 spin_unlock_irq(&mapping->tree_lock);
1694 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1695 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001696 /* We check 'start' to handle wrapping when end == ~0UL */
1697 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07001698}
1699EXPORT_SYMBOL(tag_pages_for_writeback);
1700
1701/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001702 * 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 +01001703 * @mapping: address space structure to write
1704 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001705 * @writepage: function called for each page
1706 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001707 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001708 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001709 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1710 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1711 * and msync() need to guarantee that all the data which was dirty at the time
1712 * the call was made get new I/O started against them. If wbc->sync_mode is
1713 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1714 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07001715 *
1716 * To avoid livelocks (when other process dirties new pages), we first tag
1717 * pages which should be written back with TOWRITE tag and only then start
1718 * writing them. For data-integrity sync we have to be careful so that we do
1719 * not miss some pages (e.g., because some other process has cleared TOWRITE
1720 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1721 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001722 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001723int write_cache_pages(struct address_space *mapping,
1724 struct writeback_control *wbc, writepage_t writepage,
1725 void *data)
David Howells811d7362006-08-29 19:06:09 +01001726{
David Howells811d7362006-08-29 19:06:09 +01001727 int ret = 0;
1728 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001729 struct pagevec pvec;
1730 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001731 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001732 pgoff_t index;
1733 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001734 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001735 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001736 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07001737 int tag;
David Howells811d7362006-08-29 19:06:09 +01001738
David Howells811d7362006-08-29 19:06:09 +01001739 pagevec_init(&pvec, 0);
1740 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001741 writeback_index = mapping->writeback_index; /* prev offset */
1742 index = writeback_index;
1743 if (index == 0)
1744 cycled = 1;
1745 else
1746 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001747 end = -1;
1748 } else {
1749 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1750 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1751 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1752 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001753 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001754 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001755 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001756 tag = PAGECACHE_TAG_TOWRITE;
1757 else
1758 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001759retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001760 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001761 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001762 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001763 while (!done && (index <= end)) {
1764 int i;
1765
Jan Karaf446daae2010-08-09 17:19:12 -07001766 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001767 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1768 if (nr_pages == 0)
1769 break;
David Howells811d7362006-08-29 19:06:09 +01001770
David Howells811d7362006-08-29 19:06:09 +01001771 for (i = 0; i < nr_pages; i++) {
1772 struct page *page = pvec.pages[i];
1773
Nick Piggind5482cd2009-01-06 14:39:11 -08001774 /*
1775 * At this point, the page may be truncated or
1776 * invalidated (changing page->mapping to NULL), or
1777 * even swizzled back from swapper_space to tmpfs file
1778 * mapping. However, page->index will not change
1779 * because we have a reference on the page.
1780 */
1781 if (page->index > end) {
1782 /*
1783 * can't be range_cyclic (1st pass) because
1784 * end == -1 in that case.
1785 */
1786 done = 1;
1787 break;
1788 }
1789
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001790 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001791
David Howells811d7362006-08-29 19:06:09 +01001792 lock_page(page);
1793
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001794 /*
1795 * Page truncated or invalidated. We can freely skip it
1796 * then, even for data integrity operations: the page
1797 * has disappeared concurrently, so there could be no
1798 * real expectation of this data interity operation
1799 * even if there is now a new, dirty page at the same
1800 * pagecache address.
1801 */
David Howells811d7362006-08-29 19:06:09 +01001802 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001803continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001804 unlock_page(page);
1805 continue;
1806 }
1807
Nick Piggin515f4a02009-01-06 14:39:10 -08001808 if (!PageDirty(page)) {
1809 /* someone wrote it for us */
1810 goto continue_unlock;
1811 }
David Howells811d7362006-08-29 19:06:09 +01001812
Nick Piggin515f4a02009-01-06 14:39:10 -08001813 if (PageWriteback(page)) {
1814 if (wbc->sync_mode != WB_SYNC_NONE)
1815 wait_on_page_writeback(page);
1816 else
1817 goto continue_unlock;
1818 }
1819
1820 BUG_ON(PageWriteback(page));
1821 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001822 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001823
Dave Chinner9e094382010-07-07 13:24:08 +10001824 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001825 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001826 if (unlikely(ret)) {
1827 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1828 unlock_page(page);
1829 ret = 0;
1830 } else {
1831 /*
1832 * done_index is set past this page,
1833 * so media errors will not choke
1834 * background writeout for the entire
1835 * file. This has consequences for
1836 * range_cyclic semantics (ie. it may
1837 * not be suitable for data integrity
1838 * writeout).
1839 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001840 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001841 done = 1;
1842 break;
1843 }
Dave Chinner0b564922010-06-09 10:37:18 +10001844 }
David Howells811d7362006-08-29 19:06:09 +01001845
Dave Chinner546a1922010-08-24 11:44:34 +10001846 /*
1847 * We stop writing back only if we are not doing
1848 * integrity sync. In case of integrity sync we have to
1849 * keep going until we have written all the pages
1850 * we tagged for writeback prior to entering this loop.
1851 */
1852 if (--wbc->nr_to_write <= 0 &&
1853 wbc->sync_mode == WB_SYNC_NONE) {
1854 done = 1;
1855 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001856 }
David Howells811d7362006-08-29 19:06:09 +01001857 }
1858 pagevec_release(&pvec);
1859 cond_resched();
1860 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001861 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001862 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001863 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001864 * We hit the last page and there is more work to be done: wrap
1865 * back to the start of the file
1866 */
Nick Piggin31a12662009-01-06 14:39:04 -08001867 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001868 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001869 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001870 goto retry;
1871 }
Dave Chinner0b564922010-06-09 10:37:18 +10001872 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1873 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001874
David Howells811d7362006-08-29 19:06:09 +01001875 return ret;
1876}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001877EXPORT_SYMBOL(write_cache_pages);
1878
1879/*
1880 * Function used by generic_writepages to call the real writepage
1881 * function and set the mapping flags on error
1882 */
1883static int __writepage(struct page *page, struct writeback_control *wbc,
1884 void *data)
1885{
1886 struct address_space *mapping = data;
1887 int ret = mapping->a_ops->writepage(page, wbc);
1888 mapping_set_error(mapping, ret);
1889 return ret;
1890}
1891
1892/**
1893 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1894 * @mapping: address space structure to write
1895 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1896 *
1897 * This is a library function, which implements the writepages()
1898 * address_space_operation.
1899 */
1900int generic_writepages(struct address_space *mapping,
1901 struct writeback_control *wbc)
1902{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001903 struct blk_plug plug;
1904 int ret;
1905
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001906 /* deal with chardevs and other special file */
1907 if (!mapping->a_ops->writepage)
1908 return 0;
1909
Shaohua Li9b6096a2011-03-17 10:47:06 +01001910 blk_start_plug(&plug);
1911 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1912 blk_finish_plug(&plug);
1913 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001914}
David Howells811d7362006-08-29 19:06:09 +01001915
1916EXPORT_SYMBOL(generic_writepages);
1917
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1919{
Andrew Morton22905f72005-11-16 15:07:01 -08001920 int ret;
1921
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 if (wbc->nr_to_write <= 0)
1923 return 0;
1924 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001925 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001926 else
1927 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001928 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929}
1930
1931/**
1932 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001933 * @page: the page to write
1934 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 *
1936 * The page must be locked by the caller and will be unlocked upon return.
1937 *
1938 * write_one_page() returns a negative error code if I/O failed.
1939 */
1940int write_one_page(struct page *page, int wait)
1941{
1942 struct address_space *mapping = page->mapping;
1943 int ret = 0;
1944 struct writeback_control wbc = {
1945 .sync_mode = WB_SYNC_ALL,
1946 .nr_to_write = 1,
1947 };
1948
1949 BUG_ON(!PageLocked(page));
1950
1951 if (wait)
1952 wait_on_page_writeback(page);
1953
1954 if (clear_page_dirty_for_io(page)) {
1955 page_cache_get(page);
1956 ret = mapping->a_ops->writepage(page, &wbc);
1957 if (ret == 0 && wait) {
1958 wait_on_page_writeback(page);
1959 if (PageError(page))
1960 ret = -EIO;
1961 }
1962 page_cache_release(page);
1963 } else {
1964 unlock_page(page);
1965 }
1966 return ret;
1967}
1968EXPORT_SYMBOL(write_one_page);
1969
1970/*
Ken Chen76719322007-02-10 01:43:15 -08001971 * For address_spaces which do not use buffers nor write back.
1972 */
1973int __set_page_dirty_no_writeback(struct page *page)
1974{
1975 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001976 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001977 return 0;
1978}
1979
1980/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001981 * Helper function for set_page_dirty family.
1982 * NOTE: This relies on being atomic wrt interrupts.
1983 */
1984void account_page_dirtied(struct page *page, struct address_space *mapping)
1985{
1986 if (mapping_cap_account_dirty(mapping)) {
1987 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001988 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001989 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06001990 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001991 task_io_account_write(PAGE_CACHE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001992 current->nr_dirtied++;
1993 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001994 }
1995}
Michael Rubin679ceac2010-08-20 02:31:26 -07001996EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001997
1998/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001999 * Helper function for set_page_writeback family.
2000 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2001 * wrt interrupts.
2002 */
2003void account_page_writeback(struct page *page)
2004{
2005 inc_zone_page_state(page, NR_WRITEBACK);
2006}
2007EXPORT_SYMBOL(account_page_writeback);
2008
2009/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 * For address_spaces which do not use buffers. Just tag the page as dirty in
2011 * its radix tree.
2012 *
2013 * This is also used when a single buffer is being dirtied: we want to set the
2014 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2015 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2016 *
2017 * Most callers have locked the page, which pins the address_space in memory.
2018 * But zap_pte_range() does not lock the page, however in that case the
2019 * mapping is pinned by the vma's ->vm_file reference.
2020 *
2021 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02002022 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 */
2024int __set_page_dirty_nobuffers(struct page *page)
2025{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 if (!TestSetPageDirty(page)) {
2027 struct address_space *mapping = page_mapping(page);
2028 struct address_space *mapping2;
2029
Andrew Morton8c085402006-12-10 02:19:24 -08002030 if (!mapping)
2031 return 1;
2032
Nick Piggin19fd6232008-07-25 19:45:32 -07002033 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08002034 mapping2 = page_mapping(page);
2035 if (mapping2) { /* Race with truncate? */
2036 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07002037 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002038 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08002039 radix_tree_tag_set(&mapping->page_tree,
2040 page_index(page), PAGECACHE_TAG_DIRTY);
2041 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002042 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08002043 if (mapping->host) {
2044 /* !PageAnon && !swapper_space */
2045 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002047 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002049 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050}
2051EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2052
2053/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002054 * Call this whenever redirtying a page, to de-account the dirty counters
2055 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2056 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2057 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2058 * control.
2059 */
2060void account_page_redirty(struct page *page)
2061{
2062 struct address_space *mapping = page->mapping;
2063 if (mapping && mapping_cap_account_dirty(mapping)) {
2064 current->nr_dirtied--;
2065 dec_zone_page_state(page, NR_DIRTIED);
2066 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2067 }
2068}
2069EXPORT_SYMBOL(account_page_redirty);
2070
2071/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 * When a writepage implementation decides that it doesn't want to write this
2073 * page for some reason, it should redirty the locked page via
2074 * redirty_page_for_writepage() and it should then unlock the page and return 0
2075 */
2076int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2077{
2078 wbc->pages_skipped++;
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002079 account_page_redirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 return __set_page_dirty_nobuffers(page);
2081}
2082EXPORT_SYMBOL(redirty_page_for_writepage);
2083
2084/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002085 * Dirty a page.
2086 *
2087 * For pages with a mapping this should be done under the page lock
2088 * for the benefit of asynchronous memory errors who prefer a consistent
2089 * dirty state. This rule can be broken in some special cases,
2090 * but should be better not to.
2091 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 * If the mapping doesn't provide a set_page_dirty a_op, then
2093 * just fall through and assume that it wants buffer_heads.
2094 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002095int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096{
2097 struct address_space *mapping = page_mapping(page);
2098
2099 if (likely(mapping)) {
2100 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002101 /*
2102 * readahead/lru_deactivate_page could remain
2103 * PG_readahead/PG_reclaim due to race with end_page_writeback
2104 * About readahead, if the page is written, the flags would be
2105 * reset. So no problem.
2106 * About lru_deactivate_page, if the page is redirty, the flag
2107 * will be reset. So no problem. but if the page is used by readahead
2108 * it will confuse readahead and make it restart the size rampup
2109 * process. But it's a trivial problem.
2110 */
2111 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002112#ifdef CONFIG_BLOCK
2113 if (!spd)
2114 spd = __set_page_dirty_buffers;
2115#endif
2116 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002118 if (!PageDirty(page)) {
2119 if (!TestSetPageDirty(page))
2120 return 1;
2121 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 return 0;
2123}
2124EXPORT_SYMBOL(set_page_dirty);
2125
2126/*
2127 * set_page_dirty() is racy if the caller has no reference against
2128 * page->mapping->host, and if the page is unlocked. This is because another
2129 * CPU could truncate the page off the mapping and then free the mapping.
2130 *
2131 * Usually, the page _is_ locked, or the caller is a user-space process which
2132 * holds a reference on the inode by having an open file.
2133 *
2134 * In other cases, the page should be locked before running set_page_dirty().
2135 */
2136int set_page_dirty_lock(struct page *page)
2137{
2138 int ret;
2139
Jens Axboe7eaceac2011-03-10 08:52:07 +01002140 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 ret = set_page_dirty(page);
2142 unlock_page(page);
2143 return ret;
2144}
2145EXPORT_SYMBOL(set_page_dirty_lock);
2146
2147/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 * Clear a page's dirty flag, while caring for dirty memory accounting.
2149 * Returns true if the page was previously dirty.
2150 *
2151 * This is for preparing to put the page under writeout. We leave the page
2152 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2153 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2154 * implementation will run either set_page_writeback() or set_page_dirty(),
2155 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2156 * back into sync.
2157 *
2158 * This incoherency between the page's dirty flag and radix-tree tag is
2159 * unfortunate, but it only exists while the page is locked.
2160 */
2161int clear_page_dirty_for_io(struct page *page)
2162{
2163 struct address_space *mapping = page_mapping(page);
2164
Nick Piggin79352892007-07-19 01:47:22 -07002165 BUG_ON(!PageLocked(page));
2166
Linus Torvalds7658cc22006-12-29 10:00:58 -08002167 if (mapping && mapping_cap_account_dirty(mapping)) {
2168 /*
2169 * Yes, Virginia, this is indeed insane.
2170 *
2171 * We use this sequence to make sure that
2172 * (a) we account for dirty stats properly
2173 * (b) we tell the low-level filesystem to
2174 * mark the whole page dirty if it was
2175 * dirty in a pagetable. Only to then
2176 * (c) clean the page again and return 1 to
2177 * cause the writeback.
2178 *
2179 * This way we avoid all nasty races with the
2180 * dirty bit in multiple places and clearing
2181 * them concurrently from different threads.
2182 *
2183 * Note! Normally the "set_page_dirty(page)"
2184 * has no effect on the actual dirty bit - since
2185 * that will already usually be set. But we
2186 * need the side effects, and it can help us
2187 * avoid races.
2188 *
2189 * We basically use the page "master dirty bit"
2190 * as a serialization point for all the different
2191 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002192 */
2193 if (page_mkclean(page))
2194 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002195 /*
2196 * We carefully synchronise fault handlers against
2197 * installing a dirty pte and marking the page dirty
2198 * at this point. We do this by having them hold the
2199 * page lock at some point after installing their
2200 * pte, but before marking the page dirty.
2201 * Pages are always locked coming in here, so we get
2202 * the desired exclusion. See mm/memory.c:do_wp_page()
2203 * for more comments.
2204 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08002205 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08002206 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07002207 dec_bdi_stat(mapping->backing_dev_info,
2208 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002209 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002211 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002213 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002215EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216
2217int test_clear_page_writeback(struct page *page)
2218{
2219 struct address_space *mapping = page_mapping(page);
2220 int ret;
2221
2222 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002223 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 unsigned long flags;
2225
Nick Piggin19fd6232008-07-25 19:45:32 -07002226 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002228 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 radix_tree_tag_clear(&mapping->page_tree,
2230 page_index(page),
2231 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002232 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002233 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002234 __bdi_writeout_inc(bdi);
2235 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002236 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002237 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 } else {
2239 ret = TestClearPageWriteback(page);
2240 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002241 if (ret) {
Andrew Mortond688abf2007-07-19 01:49:17 -07002242 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002243 inc_zone_page_state(page, NR_WRITTEN);
2244 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 return ret;
2246}
2247
2248int test_set_page_writeback(struct page *page)
2249{
2250 struct address_space *mapping = page_mapping(page);
2251 int ret;
2252
2253 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002254 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 unsigned long flags;
2256
Nick Piggin19fd6232008-07-25 19:45:32 -07002257 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002259 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 radix_tree_tag_set(&mapping->page_tree,
2261 page_index(page),
2262 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002263 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002264 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2265 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 if (!PageDirty(page))
2267 radix_tree_tag_clear(&mapping->page_tree,
2268 page_index(page),
2269 PAGECACHE_TAG_DIRTY);
Jan Karaf446daae2010-08-09 17:19:12 -07002270 radix_tree_tag_clear(&mapping->page_tree,
2271 page_index(page),
2272 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002273 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 } else {
2275 ret = TestSetPageWriteback(page);
2276 }
Andrew Mortond688abf2007-07-19 01:49:17 -07002277 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07002278 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 return ret;
2280
2281}
2282EXPORT_SYMBOL(test_set_page_writeback);
2283
2284/*
Nick Piggin00128182007-10-16 01:24:40 -07002285 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 * passed tag.
2287 */
2288int mapping_tagged(struct address_space *mapping, int tag)
2289{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002290 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291}
2292EXPORT_SYMBOL(mapping_tagged);