blob: 6f4271224493a01699f7ffdc29afa54b7ffd94ac [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 Viroff01bb482011-09-16 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>
Dave Chinner028c2dd2010-07-07 13:24:07 +100038#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039
40/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060041 * Sleep at most 200ms at a time in balance_dirty_pages().
42 */
43#define MAX_PAUSE max(HZ/5, 1)
44
45/*
Wu Fengguang5b9b3572011-12-06 13:17:17 -060046 * Try to keep balance_dirty_pages() call intervals higher than this many pages
47 * by raising pause time to max_pause when falls below it.
48 */
49#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
50
51/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060052 * Estimate write bandwidth at 200ms intervals.
53 */
54#define BANDWIDTH_INTERVAL max(HZ/5, 1)
55
Wu Fengguang6c14ae12011-03-02 16:04:18 -060056#define RATELIMIT_CALC_SHIFT 10
57
Wu Fengguange98be2d2010-08-29 11:22:30 -060058/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
60 * will look to see if it needs to force writeback or throttling.
61 */
62static long ratelimit_pages = 32;
63
Linus Torvalds1da177e2005-04-16 15:20:36 -070064/* The following parameters are exported via /proc/sys/vm */
65
66/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020067 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070068 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080069int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070070
71/*
David Rientjes2da02992009-01-06 14:39:31 -080072 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
73 * dirty_background_ratio * the amount of dirtyable memory
74 */
75unsigned long dirty_background_bytes;
76
77/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080078 * free highmem will not be subtracted from the total free memory
79 * for calculating free ratios if vm_highmem_is_dirtyable is true
80 */
81int vm_highmem_is_dirtyable;
82
83/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 * The generator of dirty data starts writeback at this percentage
85 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080086int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88/*
David Rientjes2da02992009-01-06 14:39:31 -080089 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
90 * vm_dirty_ratio * the amount of dirtyable memory
91 */
92unsigned long vm_dirty_bytes;
93
94/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070095 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070097unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
Artem Bityutskiy91913a22012-03-21 22:33:00 -040099EXPORT_SYMBOL_GPL(dirty_writeback_interval);
100
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700102 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700104unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
106/*
107 * Flag that makes the machine dump writes/reads and block dirtyings.
108 */
109int block_dump;
110
111/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800112 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
113 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 */
115int laptop_mode;
116
117EXPORT_SYMBOL(laptop_mode);
118
119/* End of sysctl-exported parameters */
120
Wu Fengguangc42843f2011-03-02 15:54:09 -0600121unsigned long global_dirty_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700124 * Scale the writeback cache size proportional to the relative writeout speeds.
125 *
126 * We do this by keeping a floating proportion between BDIs, based on page
127 * writeback completions [end_page_writeback()]. Those devices that write out
128 * pages fastest will get the larger share, while the slower will get a smaller
129 * share.
130 *
131 * We use page writeout completions because we are interested in getting rid of
132 * dirty pages. Having them written out is the primary goal.
133 *
134 * We introduce a concept of time, a period over which we measure these events,
135 * because demand can/will vary over time. The length of this period itself is
136 * measured in page writeback completions.
137 *
138 */
Jan Karaeb608e32012-05-24 18:59:11 +0200139static struct fprop_global writeout_completions;
140
141static void writeout_period(unsigned long t);
142/* Timer for aging of writeout_completions */
143static struct timer_list writeout_period_timer =
144 TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
145static unsigned long writeout_period_time = 0;
146
147/*
148 * Length of period for aging writeout fractions of bdis. This is an
149 * arbitrarily chosen number. The longer the period, the slower fractions will
150 * reflect changes in current writeout rate.
151 */
152#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700153
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700154/*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800155 * Work out the current dirty-memory clamping and background writeout
156 * thresholds.
157 *
158 * The main aim here is to lower them aggressively if there is a lot of mapped
159 * memory around. To avoid stressing page reclaim with lots of unreclaimable
160 * pages. It is better to clamp down on writers than to start swapping, and
161 * performing lots of scanning.
162 *
163 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
164 *
165 * We don't permit the clamping level to fall below 5% - that is getting rather
166 * excessive.
167 *
168 * We make sure that the background writeout level is below the adjusted
169 * clamping level.
170 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800171
Johannes Weinera756cf52012-01-10 15:07:49 -0800172/*
173 * In a memory zone, there is a certain amount of pages we consider
174 * available for the page cache, which is essentially the number of
175 * free and reclaimable pages, minus some zone reserves to protect
176 * lowmem and the ability to uphold the zone's watermarks without
177 * requiring writeback.
178 *
179 * This number of dirtyable pages is the base value of which the
180 * user-configurable dirty ratio is the effictive number of pages that
181 * are allowed to be actually dirtied. Per individual zone, or
182 * globally by using the sum of dirtyable pages over all zones.
183 *
184 * Because the user is allowed to specify the dirty limit globally as
185 * absolute number of bytes, calculating the per-zone dirty limit can
186 * require translating the configured limit into a percentage of
187 * global dirtyable memory first.
188 */
189
Johannes Weiner1edf2232012-01-10 15:06:57 -0800190static unsigned long highmem_dirtyable_memory(unsigned long total)
191{
192#ifdef CONFIG_HIGHMEM
193 int node;
194 unsigned long x = 0;
195
196 for_each_node_state(node, N_HIGH_MEMORY) {
197 struct zone *z =
198 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
199
200 x += zone_page_state(z, NR_FREE_PAGES) +
Johannes Weinerab8fabd2012-01-10 15:07:42 -0800201 zone_reclaimable_pages(z) - z->dirty_balance_reserve;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800202 }
203 /*
204 * Make sure that the number of highmem pages is never larger
205 * than the number of the total dirtyable memory. This can only
206 * occur in very strange VM situations but we want to make sure
207 * that this does not occur.
208 */
209 return min(x, total);
210#else
211 return 0;
212#endif
213}
214
215/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800216 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800217 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800218 * Returns the global number of pages potentially available for dirty
219 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800220 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700221static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800222{
223 unsigned long x;
224
Johannes Weinerab8fabd2012-01-10 15:07:42 -0800225 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
226 dirty_balance_reserve;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800227
228 if (!vm_highmem_is_dirtyable)
229 x -= highmem_dirtyable_memory(x);
230
231 return x + 1; /* Ensure that we never return 0 */
232}
233
234/*
Johannes Weinerccafa282012-01-10 15:07:44 -0800235 * global_dirty_limits - background-writeback and dirty-throttling thresholds
236 *
237 * Calculate the dirty thresholds based on sysctl parameters
238 * - vm.dirty_background_ratio or vm.dirty_background_bytes
239 * - vm.dirty_ratio or vm.dirty_bytes
240 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
241 * real-time tasks.
242 */
243void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
244{
245 unsigned long background;
246 unsigned long dirty;
247 unsigned long uninitialized_var(available_memory);
248 struct task_struct *tsk;
249
250 if (!vm_dirty_bytes || !dirty_background_bytes)
251 available_memory = global_dirtyable_memory();
252
253 if (vm_dirty_bytes)
254 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
255 else
256 dirty = (vm_dirty_ratio * available_memory) / 100;
257
258 if (dirty_background_bytes)
259 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
260 else
261 background = (dirty_background_ratio * available_memory) / 100;
262
263 if (background >= dirty)
264 background = dirty / 2;
265 tsk = current;
266 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
267 background += background / 4;
268 dirty += dirty / 4;
269 }
270 *pbackground = background;
271 *pdirty = dirty;
272 trace_global_dirty_state(background, dirty);
273}
274
Johannes Weinera756cf52012-01-10 15:07:49 -0800275/**
276 * zone_dirtyable_memory - number of dirtyable pages in a zone
277 * @zone: the zone
278 *
279 * Returns the zone's number of pages potentially available for dirty
280 * page cache. This is the base value for the per-zone dirty limits.
281 */
282static unsigned long zone_dirtyable_memory(struct zone *zone)
283{
284 /*
285 * The effective global number of dirtyable pages may exclude
286 * highmem as a big-picture measure to keep the ratio between
287 * dirty memory and lowmem reasonable.
288 *
289 * But this function is purely about the individual zone and a
290 * highmem zone can hold its share of dirty pages, so we don't
291 * care about vm_highmem_is_dirtyable here.
292 */
293 return zone_page_state(zone, NR_FREE_PAGES) +
294 zone_reclaimable_pages(zone) -
295 zone->dirty_balance_reserve;
296}
297
298/**
299 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
300 * @zone: the zone
301 *
302 * Returns the maximum number of dirty pages allowed in a zone, based
303 * on the zone's dirtyable memory.
304 */
305static unsigned long zone_dirty_limit(struct zone *zone)
306{
307 unsigned long zone_memory = zone_dirtyable_memory(zone);
308 struct task_struct *tsk = current;
309 unsigned long dirty;
310
311 if (vm_dirty_bytes)
312 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
313 zone_memory / global_dirtyable_memory();
314 else
315 dirty = vm_dirty_ratio * zone_memory / 100;
316
317 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
318 dirty += dirty / 4;
319
320 return dirty;
321}
322
323/**
324 * zone_dirty_ok - tells whether a zone is within its dirty limits
325 * @zone: the zone to check
326 *
327 * Returns %true when the dirty pages in @zone are within the zone's
328 * dirty limit, %false if the limit is exceeded.
329 */
330bool zone_dirty_ok(struct zone *zone)
331{
332 unsigned long limit = zone_dirty_limit(zone);
333
334 return zone_page_state(zone, NR_FILE_DIRTY) +
335 zone_page_state(zone, NR_UNSTABLE_NFS) +
336 zone_page_state(zone, NR_WRITEBACK) <= limit;
337}
338
David Rientjes2da02992009-01-06 14:39:31 -0800339int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700340 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800341 loff_t *ppos)
342{
343 int ret;
344
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700345 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800346 if (ret == 0 && write)
347 dirty_background_bytes = 0;
348 return ret;
349}
350
351int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700352 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800353 loff_t *ppos)
354{
355 int ret;
356
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700357 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800358 if (ret == 0 && write)
359 dirty_background_ratio = 0;
360 return ret;
361}
362
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700363int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700364 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700365 loff_t *ppos)
366{
367 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800368 int ret;
369
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700370 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700371 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200372 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800373 vm_dirty_bytes = 0;
374 }
375 return ret;
376}
377
David Rientjes2da02992009-01-06 14:39:31 -0800378int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700379 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800380 loff_t *ppos)
381{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800382 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800383 int ret;
384
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700385 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800386 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200387 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800388 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700389 }
390 return ret;
391}
392
Jan Karaeb608e32012-05-24 18:59:11 +0200393static unsigned long wp_next_time(unsigned long cur_time)
394{
395 cur_time += VM_COMPLETIONS_PERIOD_LEN;
396 /* 0 has a special meaning... */
397 if (!cur_time)
398 return 1;
399 return cur_time;
400}
401
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700402/*
403 * Increment the BDI's writeout completion count and the global writeout
404 * completion count. Called from test_clear_page_writeback().
405 */
406static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
407{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600408 __inc_bdi_stat(bdi, BDI_WRITTEN);
Jan Karaeb608e32012-05-24 18:59:11 +0200409 __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
410 bdi->max_prop_frac);
411 /* First event after period switching was turned off? */
412 if (!unlikely(writeout_period_time)) {
413 /*
414 * We can race with other __bdi_writeout_inc calls here but
415 * it does not cause any harm since the resulting time when
416 * timer will fire and what is in writeout_period_time will be
417 * roughly the same.
418 */
419 writeout_period_time = wp_next_time(jiffies);
420 mod_timer(&writeout_period_timer, writeout_period_time);
421 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700422}
423
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700424void bdi_writeout_inc(struct backing_dev_info *bdi)
425{
426 unsigned long flags;
427
428 local_irq_save(flags);
429 __bdi_writeout_inc(bdi);
430 local_irq_restore(flags);
431}
432EXPORT_SYMBOL_GPL(bdi_writeout_inc);
433
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700434/*
435 * Obtain an accurate fraction of the BDI's portion.
436 */
437static void bdi_writeout_fraction(struct backing_dev_info *bdi,
438 long *numerator, long *denominator)
439{
Jan Karaeb608e32012-05-24 18:59:11 +0200440 fprop_fraction_percpu(&writeout_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700441 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700442}
443
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700444/*
Jan Karaeb608e32012-05-24 18:59:11 +0200445 * On idle system, we can be called long after we scheduled because we use
446 * deferred timers so count with missed periods.
447 */
448static void writeout_period(unsigned long t)
449{
450 int miss_periods = (jiffies - writeout_period_time) /
451 VM_COMPLETIONS_PERIOD_LEN;
452
453 if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
454 writeout_period_time = wp_next_time(writeout_period_time +
455 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
456 mod_timer(&writeout_period_timer, writeout_period_time);
457 } else {
458 /*
459 * Aging has zeroed all fractions. Stop wasting CPU on period
460 * updates.
461 */
462 writeout_period_time = 0;
463 }
464}
465
466/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700467 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
468 * registered backing devices, which, for obvious reasons, can not
469 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700470 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700471static unsigned int bdi_min_ratio;
472
473int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
474{
475 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700476
Jens Axboecfc4ba52009-09-14 13:12:40 +0200477 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700478 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700479 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700480 } else {
481 min_ratio -= bdi->min_ratio;
482 if (bdi_min_ratio + min_ratio < 100) {
483 bdi_min_ratio += min_ratio;
484 bdi->min_ratio += min_ratio;
485 } else {
486 ret = -EINVAL;
487 }
488 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200489 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700490
491 return ret;
492}
493
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700494int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
495{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700496 int ret = 0;
497
498 if (max_ratio > 100)
499 return -EINVAL;
500
Jens Axboecfc4ba52009-09-14 13:12:40 +0200501 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700502 if (bdi->min_ratio > max_ratio) {
503 ret = -EINVAL;
504 } else {
505 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200506 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700507 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200508 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700509
510 return ret;
511}
512EXPORT_SYMBOL(bdi_set_max_ratio);
513
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600514static unsigned long dirty_freerun_ceiling(unsigned long thresh,
515 unsigned long bg_thresh)
516{
517 return (thresh + bg_thresh) / 2;
518}
519
Wu Fengguangffd1f602011-06-19 22:18:42 -0600520static unsigned long hard_dirty_limit(unsigned long thresh)
521{
522 return max(thresh, global_dirty_limit);
523}
524
Wu Fengguang6f718652011-03-02 17:14:34 -0600525/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700526 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600527 * @bdi: the backing_dev_info to query
528 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700529 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600530 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
531 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600532 *
533 * Note that balance_dirty_pages() will only seriously take it as a hard limit
534 * when sleeping max_pause per page is not enough to keep the dirty pages under
535 * control. For example, when the device is completely stalled due to some error
536 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
537 * In the other normal situations, it acts more gently by throttling the tasks
538 * more (rather than completely block them) when the bdi dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600539 *
540 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700541 * - starving fast devices
542 * - piling up dirty pages (that will take long time to sync) on slow devices
543 *
544 * The bdi's share of dirty limit will be adapting to its throughput and
545 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
546 */
547unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700548{
549 u64 bdi_dirty;
550 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700551
Wu Fengguang16c40422010-08-11 14:17:39 -0700552 /*
553 * Calculate this BDI's share of the dirty ratio.
554 */
555 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700556
Wu Fengguang16c40422010-08-11 14:17:39 -0700557 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
558 bdi_dirty *= numerator;
559 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700560
Wu Fengguang16c40422010-08-11 14:17:39 -0700561 bdi_dirty += (dirty * bdi->min_ratio) / 100;
562 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
563 bdi_dirty = dirty * bdi->max_ratio / 100;
564
565 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566}
567
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600568/*
569 * Dirty position control.
570 *
571 * (o) global/bdi setpoints
572 *
573 * We want the dirty pages be balanced around the global/bdi setpoints.
574 * When the number of dirty pages is higher/lower than the setpoint, the
575 * dirty position control ratio (and hence task dirty ratelimit) will be
576 * decreased/increased to bring the dirty pages back to the setpoint.
577 *
578 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
579 *
580 * if (dirty < setpoint) scale up pos_ratio
581 * if (dirty > setpoint) scale down pos_ratio
582 *
583 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
584 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
585 *
586 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
587 *
588 * (o) global control line
589 *
590 * ^ pos_ratio
591 * |
592 * | |<===== global dirty control scope ======>|
593 * 2.0 .............*
594 * | .*
595 * | . *
596 * | . *
597 * | . *
598 * | . *
599 * | . *
600 * 1.0 ................................*
601 * | . . *
602 * | . . *
603 * | . . *
604 * | . . *
605 * | . . *
606 * 0 +------------.------------------.----------------------*------------->
607 * freerun^ setpoint^ limit^ dirty pages
608 *
609 * (o) bdi control line
610 *
611 * ^ pos_ratio
612 * |
613 * | *
614 * | *
615 * | *
616 * | *
617 * | * |<=========== span ============>|
618 * 1.0 .......................*
619 * | . *
620 * | . *
621 * | . *
622 * | . *
623 * | . *
624 * | . *
625 * | . *
626 * | . *
627 * | . *
628 * | . *
629 * | . *
630 * 1/4 ...............................................* * * * * * * * * * * *
631 * | . .
632 * | . .
633 * | . .
634 * 0 +----------------------.-------------------------------.------------->
635 * bdi_setpoint^ x_intercept^
636 *
637 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
638 * be smoothly throttled down to normal if it starts high in situations like
639 * - start writing to a slow SD card and a fast disk at the same time. The SD
640 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
641 * - the bdi dirty thresh drops quickly due to change of JBOD workload
642 */
643static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
644 unsigned long thresh,
645 unsigned long bg_thresh,
646 unsigned long dirty,
647 unsigned long bdi_thresh,
648 unsigned long bdi_dirty)
649{
650 unsigned long write_bw = bdi->avg_write_bandwidth;
651 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
652 unsigned long limit = hard_dirty_limit(thresh);
653 unsigned long x_intercept;
654 unsigned long setpoint; /* dirty pages' target balance point */
655 unsigned long bdi_setpoint;
656 unsigned long span;
657 long long pos_ratio; /* for scaling up/down the rate limit */
658 long x;
659
660 if (unlikely(dirty >= limit))
661 return 0;
662
663 /*
664 * global setpoint
665 *
666 * setpoint - dirty 3
667 * f(dirty) := 1.0 + (----------------)
668 * limit - setpoint
669 *
670 * it's a 3rd order polynomial that subjects to
671 *
672 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
673 * (2) f(setpoint) = 1.0 => the balance point
674 * (3) f(limit) = 0 => the hard limit
675 * (4) df/dx <= 0 => negative feedback control
676 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
677 * => fast response on large errors; small oscillation near setpoint
678 */
679 setpoint = (freerun + limit) / 2;
680 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
681 limit - setpoint + 1);
682 pos_ratio = x;
683 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
684 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
685 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
686
687 /*
688 * We have computed basic pos_ratio above based on global situation. If
689 * the bdi is over/under its share of dirty pages, we want to scale
690 * pos_ratio further down/up. That is done by the following mechanism.
691 */
692
693 /*
694 * bdi setpoint
695 *
696 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
697 *
698 * x_intercept - bdi_dirty
699 * := --------------------------
700 * x_intercept - bdi_setpoint
701 *
702 * The main bdi control line is a linear function that subjects to
703 *
704 * (1) f(bdi_setpoint) = 1.0
705 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
706 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
707 *
708 * For single bdi case, the dirty pages are observed to fluctuate
709 * regularly within range
710 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
711 * for various filesystems, where (2) can yield in a reasonable 12.5%
712 * fluctuation range for pos_ratio.
713 *
714 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
715 * own size, so move the slope over accordingly and choose a slope that
716 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
717 */
718 if (unlikely(bdi_thresh > thresh))
719 bdi_thresh = thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600720 /*
721 * It's very possible that bdi_thresh is close to 0 not because the
722 * device is slow, but that it has remained inactive for long time.
723 * Honour such devices a reasonable good (hopefully IO efficient)
724 * threshold, so that the occasional writes won't be blocked and active
725 * writes can rampup the threshold quickly.
726 */
Wu Fengguang8927f662011-08-04 22:16:46 -0600727 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600728 /*
729 * scale global setpoint to bdi's:
730 * bdi_setpoint = setpoint * bdi_thresh / thresh
731 */
732 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
733 bdi_setpoint = setpoint * (u64)x >> 16;
734 /*
735 * Use span=(8*write_bw) in single bdi case as indicated by
736 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
737 *
738 * bdi_thresh thresh - bdi_thresh
739 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
740 * thresh thresh
741 */
742 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
743 x_intercept = bdi_setpoint + span;
744
745 if (bdi_dirty < x_intercept - span / 4) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600746 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
747 x_intercept - bdi_setpoint + 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600748 } else
749 pos_ratio /= 4;
750
Wu Fengguang8927f662011-08-04 22:16:46 -0600751 /*
752 * bdi reserve area, safeguard against dirty pool underrun and disk idle
753 * It may push the desired control point of global dirty pages higher
754 * than setpoint.
755 */
756 x_intercept = bdi_thresh / 2;
757 if (bdi_dirty < x_intercept) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600758 if (bdi_dirty > x_intercept / 8)
759 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
760 else
Wu Fengguang8927f662011-08-04 22:16:46 -0600761 pos_ratio *= 8;
762 }
763
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600764 return pos_ratio;
765}
766
Wu Fengguange98be2d2010-08-29 11:22:30 -0600767static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
768 unsigned long elapsed,
769 unsigned long written)
770{
771 const unsigned long period = roundup_pow_of_two(3 * HZ);
772 unsigned long avg = bdi->avg_write_bandwidth;
773 unsigned long old = bdi->write_bandwidth;
774 u64 bw;
775
776 /*
777 * bw = written * HZ / elapsed
778 *
779 * bw * elapsed + write_bandwidth * (period - elapsed)
780 * write_bandwidth = ---------------------------------------------------
781 * period
782 */
783 bw = written - bdi->written_stamp;
784 bw *= HZ;
785 if (unlikely(elapsed > period)) {
786 do_div(bw, elapsed);
787 avg = bw;
788 goto out;
789 }
790 bw += (u64)bdi->write_bandwidth * (period - elapsed);
791 bw >>= ilog2(period);
792
793 /*
794 * one more level of smoothing, for filtering out sudden spikes
795 */
796 if (avg > old && old >= (unsigned long)bw)
797 avg -= (avg - old) >> 3;
798
799 if (avg < old && old <= (unsigned long)bw)
800 avg += (old - avg) >> 3;
801
802out:
803 bdi->write_bandwidth = bw;
804 bdi->avg_write_bandwidth = avg;
805}
806
Wu Fengguangc42843f2011-03-02 15:54:09 -0600807/*
808 * The global dirtyable memory and dirty threshold could be suddenly knocked
809 * down by a large amount (eg. on the startup of KVM in a swapless system).
810 * This may throw the system into deep dirty exceeded state and throttle
811 * heavy/light dirtiers alike. To retain good responsiveness, maintain
812 * global_dirty_limit for tracking slowly down to the knocked down dirty
813 * threshold.
814 */
815static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
816{
817 unsigned long limit = global_dirty_limit;
818
819 /*
820 * Follow up in one step.
821 */
822 if (limit < thresh) {
823 limit = thresh;
824 goto update;
825 }
826
827 /*
828 * Follow down slowly. Use the higher one as the target, because thresh
829 * may drop below dirty. This is exactly the reason to introduce
830 * global_dirty_limit which is guaranteed to lie above the dirty pages.
831 */
832 thresh = max(thresh, dirty);
833 if (limit > thresh) {
834 limit -= (limit - thresh) >> 5;
835 goto update;
836 }
837 return;
838update:
839 global_dirty_limit = limit;
840}
841
842static void global_update_bandwidth(unsigned long thresh,
843 unsigned long dirty,
844 unsigned long now)
845{
846 static DEFINE_SPINLOCK(dirty_lock);
847 static unsigned long update_time;
848
849 /*
850 * check locklessly first to optimize away locking for the most time
851 */
852 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
853 return;
854
855 spin_lock(&dirty_lock);
856 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
857 update_dirty_limit(thresh, dirty);
858 update_time = now;
859 }
860 spin_unlock(&dirty_lock);
861}
862
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600863/*
864 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
865 *
866 * Normal bdi tasks will be curbed at or below it in long term.
867 * Obviously it should be around (write_bw / N) when there are N dd tasks.
868 */
869static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
870 unsigned long thresh,
871 unsigned long bg_thresh,
872 unsigned long dirty,
873 unsigned long bdi_thresh,
874 unsigned long bdi_dirty,
875 unsigned long dirtied,
876 unsigned long elapsed)
877{
Wu Fengguang73811312011-08-26 15:53:24 -0600878 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
879 unsigned long limit = hard_dirty_limit(thresh);
880 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600881 unsigned long write_bw = bdi->avg_write_bandwidth;
882 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
883 unsigned long dirty_rate;
884 unsigned long task_ratelimit;
885 unsigned long balanced_dirty_ratelimit;
886 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600887 unsigned long step;
888 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600889
890 /*
891 * The dirty rate will match the writeout rate in long term, except
892 * when dirty pages are truncated by userspace or re-dirtied by FS.
893 */
894 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
895
896 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
897 bdi_thresh, bdi_dirty);
898 /*
899 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
900 */
901 task_ratelimit = (u64)dirty_ratelimit *
902 pos_ratio >> RATELIMIT_CALC_SHIFT;
903 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
904
905 /*
906 * A linear estimation of the "balanced" throttle rate. The theory is,
907 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
908 * dirty_rate will be measured to be (N * task_ratelimit). So the below
909 * formula will yield the balanced rate limit (write_bw / N).
910 *
911 * Note that the expanded form is not a pure rate feedback:
912 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
913 * but also takes pos_ratio into account:
914 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
915 *
916 * (1) is not realistic because pos_ratio also takes part in balancing
917 * the dirty rate. Consider the state
918 * pos_ratio = 0.5 (3)
919 * rate = 2 * (write_bw / N) (4)
920 * If (1) is used, it will stuck in that state! Because each dd will
921 * be throttled at
922 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
923 * yielding
924 * dirty_rate = N * task_ratelimit = write_bw (6)
925 * put (6) into (1) we get
926 * rate_(i+1) = rate_(i) (7)
927 *
928 * So we end up using (2) to always keep
929 * rate_(i+1) ~= (write_bw / N) (8)
930 * regardless of the value of pos_ratio. As long as (8) is satisfied,
931 * pos_ratio is able to drive itself to 1.0, which is not only where
932 * the dirty count meet the setpoint, but also where the slope of
933 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
934 */
935 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
936 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -0600937 /*
938 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
939 */
940 if (unlikely(balanced_dirty_ratelimit > write_bw))
941 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600942
Wu Fengguang73811312011-08-26 15:53:24 -0600943 /*
944 * We could safely do this and return immediately:
945 *
946 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
947 *
948 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +0800949 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -0600950 * limit the step size.
951 *
952 * The below code essentially only uses the relative value of
953 *
954 * task_ratelimit - dirty_ratelimit
955 * = (pos_ratio - 1) * dirty_ratelimit
956 *
957 * which reflects the direction and size of dirty position error.
958 */
959
960 /*
961 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
962 * task_ratelimit is on the same side of dirty_ratelimit, too.
963 * For example, when
964 * - dirty_ratelimit > balanced_dirty_ratelimit
965 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
966 * lowering dirty_ratelimit will help meet both the position and rate
967 * control targets. Otherwise, don't update dirty_ratelimit if it will
968 * only help meet the rate target. After all, what the users ultimately
969 * feel and care are stable dirty rate and small position error.
970 *
971 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +0800972 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -0600973 * keeps jumping around randomly and can even leap far away at times
974 * due to the small 200ms estimation period of dirty_rate (we want to
975 * keep that period small to reduce time lags).
976 */
977 step = 0;
978 if (dirty < setpoint) {
979 x = min(bdi->balanced_dirty_ratelimit,
980 min(balanced_dirty_ratelimit, task_ratelimit));
981 if (dirty_ratelimit < x)
982 step = x - dirty_ratelimit;
983 } else {
984 x = max(bdi->balanced_dirty_ratelimit,
985 max(balanced_dirty_ratelimit, task_ratelimit));
986 if (dirty_ratelimit > x)
987 step = dirty_ratelimit - x;
988 }
989
990 /*
991 * Don't pursue 100% rate matching. It's impossible since the balanced
992 * rate itself is constantly fluctuating. So decrease the track speed
993 * when it gets close to the target. Helps eliminate pointless tremors.
994 */
995 step >>= dirty_ratelimit / (2 * step + 1);
996 /*
997 * Limit the tracking speed to avoid overshooting.
998 */
999 step = (step + 7) / 8;
1000
1001 if (dirty_ratelimit < balanced_dirty_ratelimit)
1002 dirty_ratelimit += step;
1003 else
1004 dirty_ratelimit -= step;
1005
1006 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1007 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001008
1009 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001010}
1011
Wu Fengguange98be2d2010-08-29 11:22:30 -06001012void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001013 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001014 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001015 unsigned long dirty,
1016 unsigned long bdi_thresh,
1017 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -06001018 unsigned long start_time)
1019{
1020 unsigned long now = jiffies;
1021 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001022 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001023 unsigned long written;
1024
1025 /*
1026 * rate-limit, only update once every 200ms.
1027 */
1028 if (elapsed < BANDWIDTH_INTERVAL)
1029 return;
1030
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001031 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001032 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1033
1034 /*
1035 * Skip quiet periods when disk bandwidth is under-utilized.
1036 * (at least 1s idle time between two flusher runs)
1037 */
1038 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1039 goto snapshot;
1040
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001041 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -06001042 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001043 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1044 bdi_thresh, bdi_dirty,
1045 dirtied, elapsed);
1046 }
Wu Fengguange98be2d2010-08-29 11:22:30 -06001047 bdi_update_write_bandwidth(bdi, elapsed, written);
1048
1049snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001050 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001051 bdi->written_stamp = written;
1052 bdi->bw_time_stamp = now;
1053}
1054
1055static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001056 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001057 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001058 unsigned long dirty,
1059 unsigned long bdi_thresh,
1060 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -06001061 unsigned long start_time)
1062{
1063 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1064 return;
1065 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001066 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1067 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001068 spin_unlock(&bdi->wb.list_lock);
1069}
1070
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001072 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001073 * will look to see if it needs to start dirty throttling.
1074 *
1075 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1076 * global_page_state() too often. So scale it near-sqrt to the safety margin
1077 * (the number of pages we may dirty without exceeding the dirty limits).
1078 */
1079static unsigned long dirty_poll_interval(unsigned long dirty,
1080 unsigned long thresh)
1081{
1082 if (thresh > dirty)
1083 return 1UL << (ilog2(thresh - dirty) >> 1);
1084
1085 return 1;
1086}
1087
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001088static long bdi_max_pause(struct backing_dev_info *bdi,
1089 unsigned long bdi_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001090{
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001091 long bw = bdi->avg_write_bandwidth;
1092 long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001093
1094 /*
1095 * Limit pause time for small memory systems. If sleeping for too long
1096 * time, a small pool of dirty/writeback pages may go empty and disk go
1097 * idle.
1098 *
1099 * 8 serves as the safety ratio.
1100 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001101 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1102 t++;
1103
1104 return min_t(long, t, MAX_PAUSE);
1105}
1106
1107static long bdi_min_pause(struct backing_dev_info *bdi,
1108 long max_pause,
1109 unsigned long task_ratelimit,
1110 unsigned long dirty_ratelimit,
1111 int *nr_dirtied_pause)
1112{
1113 long hi = ilog2(bdi->avg_write_bandwidth);
1114 long lo = ilog2(bdi->dirty_ratelimit);
1115 long t; /* target pause */
1116 long pause; /* estimated next pause */
1117 int pages; /* target nr_dirtied_pause */
1118
1119 /* target for 10ms pause on 1-dd case */
1120 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001121
1122 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001123 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1124 * overheads.
1125 *
1126 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001127 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001128 if (hi > lo)
1129 t += (hi - lo) * (10 * HZ) / 1024;
1130
1131 /*
1132 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1133 * on the much more stable dirty_ratelimit. However the next pause time
1134 * will be computed based on task_ratelimit and the two rate limits may
1135 * depart considerably at some time. Especially if task_ratelimit goes
1136 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1137 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1138 * result task_ratelimit won't be executed faithfully, which could
1139 * eventually bring down dirty_ratelimit.
1140 *
1141 * We apply two rules to fix it up:
1142 * 1) try to estimate the next pause time and if necessary, use a lower
1143 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1144 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1145 * 2) limit the target pause time to max_pause/2, so that the normal
1146 * small fluctuations of task_ratelimit won't trigger rule (1) and
1147 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1148 */
1149 t = min(t, 1 + max_pause / 2);
1150 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1151
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001152 /*
1153 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1154 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1155 * When the 16 consecutive reads are often interrupted by some dirty
1156 * throttling pause during the async writes, cfq will go into idles
1157 * (deadline is fine). So push nr_dirtied_pause as high as possible
1158 * until reaches DIRTY_POLL_THRESH=32 pages.
1159 */
1160 if (pages < DIRTY_POLL_THRESH) {
1161 t = max_pause;
1162 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1163 if (pages > DIRTY_POLL_THRESH) {
1164 pages = DIRTY_POLL_THRESH;
1165 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1166 }
1167 }
1168
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001169 pause = HZ * pages / (task_ratelimit + 1);
1170 if (pause > max_pause) {
1171 t = max_pause;
1172 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1173 }
1174
1175 *nr_dirtied_pause = pages;
1176 /*
1177 * The minimal pause time will normally be half the target pause time.
1178 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001179 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001180}
1181
Wu Fengguang9d823e82011-06-11 18:10:12 -06001182/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 * balance_dirty_pages() must be called by processes which are generating dirty
1184 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001185 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001186 * If we're over `background_thresh' then the writeback threads are woken to
1187 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001189static void balance_dirty_pages(struct address_space *mapping,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001190 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191{
Wu Fengguang143dfe82010-08-27 18:45:12 -06001192 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1193 unsigned long bdi_reclaimable;
Wu Fengguang77627412010-09-12 13:34:05 -06001194 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1195 unsigned long bdi_dirty;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001196 unsigned long freerun;
David Rientjes364aeb22009-01-06 14:39:29 -08001197 unsigned long background_thresh;
1198 unsigned long dirty_thresh;
1199 unsigned long bdi_thresh;
Wu Fengguang83712352011-06-11 19:25:42 -06001200 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001201 long pause;
1202 long max_pause;
1203 long min_pause;
1204 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001205 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001206 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001207 unsigned long dirty_ratelimit;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001208 unsigned long pos_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001210 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211
1212 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001213 unsigned long now = jiffies;
1214
Wu Fengguang143dfe82010-08-27 18:45:12 -06001215 /*
1216 * Unstable writes are a feature of certain networked
1217 * filesystems (i.e. NFS) in which data may have been
1218 * written to the server's write cache, but has not yet
1219 * been flushed to permanent storage.
1220 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001221 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1222 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -06001223 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001224
Wu Fengguang16c40422010-08-11 14:17:39 -07001225 global_dirty_limits(&background_thresh, &dirty_thresh);
1226
1227 /*
1228 * Throttle it only when the background writeback cannot
1229 * catch-up. This avoids (excessively) small writeouts
1230 * when the bdi limits are ramping up.
1231 */
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001232 freerun = dirty_freerun_ceiling(dirty_thresh,
1233 background_thresh);
Wu Fengguang83712352011-06-11 19:25:42 -06001234 if (nr_dirty <= freerun) {
1235 current->dirty_paused_when = now;
1236 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001237 current->nr_dirtied_pause =
1238 dirty_poll_interval(nr_dirty, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001239 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001240 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001241
Wu Fengguang143dfe82010-08-27 18:45:12 -06001242 if (unlikely(!writeback_in_progress(bdi)))
1243 bdi_start_background_writeback(bdi);
1244
1245 /*
1246 * bdi_thresh is not treated as some limiting factor as
1247 * dirty_thresh, due to reasons
1248 * - in JBOD setup, bdi_thresh can fluctuate a lot
1249 * - in a system with HDD and USB key, the USB key may somehow
1250 * go into state (bdi_dirty >> bdi_thresh) either because
1251 * bdi_dirty starts high, or because bdi_thresh drops low.
1252 * In this case we don't want to hard throttle the USB key
1253 * dirtiers for 100 seconds until bdi_dirty drops under
1254 * bdi_thresh. Instead the auxiliary bdi control line in
1255 * bdi_position_ratio() will let the dirtier task progress
1256 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1257 */
Wu Fengguang16c40422010-08-11 14:17:39 -07001258 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001259
Wu Fengguange50e3722010-08-11 14:17:37 -07001260 /*
1261 * In order to avoid the stacked BDI deadlock we need
1262 * to ensure we accurately count the 'dirty' pages when
1263 * the threshold is low.
1264 *
1265 * Otherwise it would be possible to get thresh+n pages
1266 * reported dirty, even though there are thresh-m pages
1267 * actually dirty; with m+n sitting in the percpu
1268 * deltas.
1269 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001270 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1271 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1272 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001273 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001274 } else {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001275 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1276 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001277 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001278 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001279
Wu Fengguang82791942011-12-03 21:26:01 -06001280 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
Wu Fengguang77627412010-09-12 13:34:05 -06001281 (nr_dirty > dirty_thresh);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001282 if (dirty_exceeded && !bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001283 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001285 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1286 nr_dirty, bdi_thresh, bdi_dirty,
1287 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001288
Wu Fengguang143dfe82010-08-27 18:45:12 -06001289 dirty_ratelimit = bdi->dirty_ratelimit;
1290 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1291 background_thresh, nr_dirty,
1292 bdi_thresh, bdi_dirty);
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001293 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1294 RATELIMIT_CALC_SHIFT;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001295 max_pause = bdi_max_pause(bdi, bdi_dirty);
1296 min_pause = bdi_min_pause(bdi, max_pause,
1297 task_ratelimit, dirty_ratelimit,
1298 &nr_dirtied_pause);
1299
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001300 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001301 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001302 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001303 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 }
Wu Fengguang83712352011-06-11 19:25:42 -06001305 period = HZ * pages_dirtied / task_ratelimit;
1306 pause = period;
1307 if (current->dirty_paused_when)
1308 pause -= now - current->dirty_paused_when;
1309 /*
1310 * For less than 1s think time (ext3/4 may block the dirtier
1311 * for up to 800ms from time to time on 1-HDD; so does xfs,
1312 * however at much less frequency), try to compensate it in
1313 * future periods by updating the virtual time; otherwise just
1314 * do a reset, as it may be a light dirtier.
1315 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001316 if (pause < min_pause) {
Wu Fengguangece13ac2010-08-29 23:33:20 -06001317 trace_balance_dirty_pages(bdi,
1318 dirty_thresh,
1319 background_thresh,
1320 nr_dirty,
1321 bdi_thresh,
1322 bdi_dirty,
1323 dirty_ratelimit,
1324 task_ratelimit,
1325 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001326 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001327 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001328 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001329 if (pause < -HZ) {
1330 current->dirty_paused_when = now;
1331 current->nr_dirtied = 0;
1332 } else if (period) {
1333 current->dirty_paused_when += period;
1334 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001335 } else if (current->nr_dirtied_pause <= pages_dirtied)
1336 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001337 break;
1338 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001339 if (unlikely(pause > max_pause)) {
1340 /* for occasional dropped task_ratelimit */
1341 now += min(pause - max_pause, max_pause);
1342 pause = max_pause;
1343 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001344
1345pause:
Wu Fengguangece13ac2010-08-29 23:33:20 -06001346 trace_balance_dirty_pages(bdi,
1347 dirty_thresh,
1348 background_thresh,
1349 nr_dirty,
1350 bdi_thresh,
1351 bdi_dirty,
1352 dirty_ratelimit,
1353 task_ratelimit,
1354 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001355 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001356 pause,
1357 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001358 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001359 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001360
Wu Fengguang83712352011-06-11 19:25:42 -06001361 current->dirty_paused_when = now + pause;
1362 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001363 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001364
Wu Fengguangffd1f602011-06-19 22:18:42 -06001365 /*
Wu Fengguang1df64712011-11-13 19:47:32 -06001366 * This is typically equal to (nr_dirty < dirty_thresh) and can
1367 * also keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001368 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001369 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001370 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001371
Wu Fengguangc5c63432011-12-02 10:21:33 -06001372 /*
1373 * In the case of an unresponding NFS server and the NFS dirty
1374 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1375 * to go through, so that tasks on them still remain responsive.
1376 *
1377 * In theory 1 page is enough to keep the comsumer-producer
1378 * pipe going: the flusher cleans 1 page => the task dirties 1
1379 * more page. However bdi_dirty has accounting errors. So use
1380 * the larger and more IO friendly bdi_stat_error.
1381 */
1382 if (bdi_dirty <= bdi_stat_error(bdi))
1383 break;
1384
Jan Kara499d05e2011-11-16 19:34:48 +08001385 if (fatal_signal_pending(current))
1386 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 }
1388
Wu Fengguang143dfe82010-08-27 18:45:12 -06001389 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001390 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
1392 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001393 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
1395 /*
1396 * In laptop mode, we wait until hitting the higher threshold before
1397 * starting background writeout, and then write out all the way down
1398 * to the lower threshold. So slow writers cause minimal disk activity.
1399 *
1400 * In normal mode, we start background writeout at the lower
1401 * background_thresh, to keep the amount of dirty memory low.
1402 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001403 if (laptop_mode)
1404 return;
1405
1406 if (nr_reclaimable > background_thresh)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001407 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408}
1409
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001410void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001411{
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001412 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001413 struct address_space *mapping = page_mapping(page);
1414
1415 if (mapping)
1416 balance_dirty_pages_ratelimited(mapping);
1417 }
1418}
1419
Wu Fengguang9d823e82011-06-11 18:10:12 -06001420static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001421
Wu Fengguang54848d72011-04-05 13:21:19 -06001422/*
1423 * Normal tasks are throttled by
1424 * loop {
1425 * dirty tsk->nr_dirtied_pause pages;
1426 * take a snap in balance_dirty_pages();
1427 * }
1428 * However there is a worst case. If every task exit immediately when dirtied
1429 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1430 * called to throttle the page dirties. The solution is to save the not yet
1431 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1432 * randomly into the running tasks. This works well for the above worst case,
1433 * as the new task will pick up and accumulate the old task's leaked dirty
1434 * count and eventually get throttled.
1435 */
1436DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1437
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001439 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001440 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 *
1442 * Processes which are dirtying memory should call in here once for each page
1443 * which was newly dirtied. The function will periodically check the system's
1444 * dirty state and will initiate writeback if needed.
1445 *
1446 * On really big machines, get_writeback_state is expensive, so try to avoid
1447 * calling it too often (ratelimiting). But once we're over the dirty memory
1448 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1449 * from overshooting the limit by (ratelimit_pages) each.
1450 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001451void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001453 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001454 int ratelimit;
1455 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456
Wu Fengguang36715ce2011-06-11 17:53:57 -06001457 if (!bdi_cap_account_dirty(bdi))
1458 return;
1459
Wu Fengguang9d823e82011-06-11 18:10:12 -06001460 ratelimit = current->nr_dirtied_pause;
1461 if (bdi->dirty_exceeded)
1462 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001464 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001465 /*
1466 * This prevents one CPU to accumulate too many dirtied pages without
1467 * calling into balance_dirty_pages(), which can happen when there are
1468 * 1000+ tasks, all of them start dirtying pages at exactly the same
1469 * time, hence all honoured too large initial task->nr_dirtied_pause.
1470 */
Tejun Heo245b2e72009-06-24 15:13:48 +09001471 p = &__get_cpu_var(bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001472 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001473 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001474 else if (unlikely(*p >= ratelimit_pages)) {
1475 *p = 0;
1476 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001478 /*
1479 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1480 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1481 * the dirty throttling and livelock other long-run dirtiers.
1482 */
1483 p = &__get_cpu_var(dirty_throttle_leaks);
1484 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001485 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001486 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1487 *p -= nr_pages_dirtied;
1488 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001490 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001491
1492 if (unlikely(current->nr_dirtied >= ratelimit))
1493 balance_dirty_pages(mapping, current->nr_dirtied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001495EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
Andrew Morton232ea4d2007-02-28 20:13:21 -08001497void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498{
David Rientjes364aeb22009-01-06 14:39:29 -08001499 unsigned long background_thresh;
1500 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501
1502 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001503 global_dirty_limits(&background_thresh, &dirty_thresh);
Fengguang Wu47a13332012-03-21 16:34:09 -07001504 dirty_thresh = hard_dirty_limit(dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
1506 /*
1507 * Boost the allowable dirty threshold a bit for page
1508 * allocators so they don't get DoS'ed by heavy writers
1509 */
1510 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1511
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001512 if (global_page_state(NR_UNSTABLE_NFS) +
1513 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1514 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001515 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001516
1517 /*
1518 * The caller might hold locks which can prevent IO completion
1519 * or progress in the filesystem. So we cannot just sit here
1520 * waiting for IO to complete.
1521 */
1522 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1523 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 }
1525}
1526
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1529 */
1530int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001531 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001533 proc_dointvec(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 return 0;
1535}
1536
Jens Axboec2c49862010-05-20 09:18:47 +02001537#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001538void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539{
Matthew Garrett31373d02010-04-06 14:25:14 +02001540 struct request_queue *q = (struct request_queue *)data;
1541 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1542 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543
Matthew Garrett31373d02010-04-06 14:25:14 +02001544 /*
1545 * We want to write everything out, not just down to the dirty
1546 * threshold
1547 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001548 if (bdi_has_dirty_io(&q->backing_dev_info))
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001549 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1550 WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551}
1552
1553/*
1554 * We've spun up the disk and we're in laptop mode: schedule writeback
1555 * of all dirty data a few seconds from now. If the flush is already scheduled
1556 * then push it back - the user is still using the disk.
1557 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001558void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559{
Matthew Garrett31373d02010-04-06 14:25:14 +02001560 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561}
1562
1563/*
1564 * We're in laptop mode and we've just synced. The sync's writes will have
1565 * caused another writeback to be scheduled by laptop_io_completion.
1566 * Nothing needs to be written back anymore, so we unschedule the writeback.
1567 */
1568void laptop_sync_completion(void)
1569{
Matthew Garrett31373d02010-04-06 14:25:14 +02001570 struct backing_dev_info *bdi;
1571
1572 rcu_read_lock();
1573
1574 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1575 del_timer(&bdi->laptop_mode_wb_timer);
1576
1577 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578}
Jens Axboec2c49862010-05-20 09:18:47 +02001579#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580
1581/*
1582 * If ratelimit_pages is too high then we can get into dirty-data overload
1583 * if a large number of processes all perform writes at the same time.
1584 * If it is too low then SMP machines will call the (expensive)
1585 * get_writeback_state too often.
1586 *
1587 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1588 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06001589 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 */
1591
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001592void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593{
Wu Fengguang9d823e82011-06-11 18:10:12 -06001594 unsigned long background_thresh;
1595 unsigned long dirty_thresh;
1596 global_dirty_limits(&background_thresh, &dirty_thresh);
Fengguang Wu68809c72012-05-06 13:21:42 +08001597 global_dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001598 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 if (ratelimit_pages < 16)
1600 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601}
1602
Chandra Seetharaman26c21432006-06-27 02:54:10 -07001603static int __cpuinit
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08001604ratelimit_handler(struct notifier_block *self, unsigned long action,
1605 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606{
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08001607
1608 switch (action & ~CPU_TASKS_FROZEN) {
1609 case CPU_ONLINE:
1610 case CPU_DEAD:
1611 writeback_set_ratelimit();
1612 return NOTIFY_OK;
1613 default:
1614 return NOTIFY_DONE;
1615 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616}
1617
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001618static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 .notifier_call = ratelimit_handler,
1620 .next = NULL,
1621};
1622
1623/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001624 * Called early on to tune the page writeback dirty limits.
1625 *
1626 * We used to scale dirty pages according to how total memory
1627 * related to pages that could be allocated for buffers (by
1628 * comparing nr_free_buffer_pages() to vm_total_pages.
1629 *
1630 * However, that was when we used "dirty_ratio" to scale with
1631 * all memory, and we don't do that any more. "dirty_ratio"
1632 * is now applied to total non-HIGHPAGE memory (by subtracting
1633 * totalhigh_pages from vm_total_pages), and as such we can't
1634 * get into the old insane situation any more where we had
1635 * large amounts of dirty pages compared to a small amount of
1636 * non-HIGHMEM memory.
1637 *
1638 * But we might still want to scale the dirty_ratio by how
1639 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 */
1641void __init page_writeback_init(void)
1642{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001643 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001645
Jan Karaeb608e32012-05-24 18:59:11 +02001646 fprop_global_init(&writeout_completions);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647}
1648
David Howells811d7362006-08-29 19:06:09 +01001649/**
Jan Karaf446daae2010-08-09 17:19:12 -07001650 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1651 * @mapping: address space structure to write
1652 * @start: starting page index
1653 * @end: ending page index (inclusive)
1654 *
1655 * This function scans the page range from @start to @end (inclusive) and tags
1656 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1657 * that write_cache_pages (or whoever calls this function) will then use
1658 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1659 * used to avoid livelocking of writeback by a process steadily creating new
1660 * dirty pages in the file (thus it is important for this function to be quick
1661 * so that it can tag pages faster than a dirtying process can create them).
1662 */
1663/*
1664 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1665 */
Jan Karaf446daae2010-08-09 17:19:12 -07001666void tag_pages_for_writeback(struct address_space *mapping,
1667 pgoff_t start, pgoff_t end)
1668{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001669#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07001670 unsigned long tagged;
1671
1672 do {
1673 spin_lock_irq(&mapping->tree_lock);
1674 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1675 &start, end, WRITEBACK_TAG_BATCH,
1676 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1677 spin_unlock_irq(&mapping->tree_lock);
1678 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1679 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001680 /* We check 'start' to handle wrapping when end == ~0UL */
1681 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07001682}
1683EXPORT_SYMBOL(tag_pages_for_writeback);
1684
1685/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001686 * 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 +01001687 * @mapping: address space structure to write
1688 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001689 * @writepage: function called for each page
1690 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001691 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001692 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001693 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1694 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1695 * and msync() need to guarantee that all the data which was dirty at the time
1696 * the call was made get new I/O started against them. If wbc->sync_mode is
1697 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1698 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07001699 *
1700 * To avoid livelocks (when other process dirties new pages), we first tag
1701 * pages which should be written back with TOWRITE tag and only then start
1702 * writing them. For data-integrity sync we have to be careful so that we do
1703 * not miss some pages (e.g., because some other process has cleared TOWRITE
1704 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1705 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001706 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001707int write_cache_pages(struct address_space *mapping,
1708 struct writeback_control *wbc, writepage_t writepage,
1709 void *data)
David Howells811d7362006-08-29 19:06:09 +01001710{
David Howells811d7362006-08-29 19:06:09 +01001711 int ret = 0;
1712 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001713 struct pagevec pvec;
1714 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001715 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001716 pgoff_t index;
1717 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001718 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001719 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001720 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07001721 int tag;
David Howells811d7362006-08-29 19:06:09 +01001722
David Howells811d7362006-08-29 19:06:09 +01001723 pagevec_init(&pvec, 0);
1724 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001725 writeback_index = mapping->writeback_index; /* prev offset */
1726 index = writeback_index;
1727 if (index == 0)
1728 cycled = 1;
1729 else
1730 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001731 end = -1;
1732 } else {
1733 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1734 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1735 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1736 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001737 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001738 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001739 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001740 tag = PAGECACHE_TAG_TOWRITE;
1741 else
1742 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001743retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001744 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001745 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001746 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001747 while (!done && (index <= end)) {
1748 int i;
1749
Jan Karaf446daae2010-08-09 17:19:12 -07001750 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001751 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1752 if (nr_pages == 0)
1753 break;
David Howells811d7362006-08-29 19:06:09 +01001754
David Howells811d7362006-08-29 19:06:09 +01001755 for (i = 0; i < nr_pages; i++) {
1756 struct page *page = pvec.pages[i];
1757
Nick Piggind5482cd2009-01-06 14:39:11 -08001758 /*
1759 * At this point, the page may be truncated or
1760 * invalidated (changing page->mapping to NULL), or
1761 * even swizzled back from swapper_space to tmpfs file
1762 * mapping. However, page->index will not change
1763 * because we have a reference on the page.
1764 */
1765 if (page->index > end) {
1766 /*
1767 * can't be range_cyclic (1st pass) because
1768 * end == -1 in that case.
1769 */
1770 done = 1;
1771 break;
1772 }
1773
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001774 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001775
David Howells811d7362006-08-29 19:06:09 +01001776 lock_page(page);
1777
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001778 /*
1779 * Page truncated or invalidated. We can freely skip it
1780 * then, even for data integrity operations: the page
1781 * has disappeared concurrently, so there could be no
1782 * real expectation of this data interity operation
1783 * even if there is now a new, dirty page at the same
1784 * pagecache address.
1785 */
David Howells811d7362006-08-29 19:06:09 +01001786 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001787continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001788 unlock_page(page);
1789 continue;
1790 }
1791
Nick Piggin515f4a02009-01-06 14:39:10 -08001792 if (!PageDirty(page)) {
1793 /* someone wrote it for us */
1794 goto continue_unlock;
1795 }
David Howells811d7362006-08-29 19:06:09 +01001796
Nick Piggin515f4a02009-01-06 14:39:10 -08001797 if (PageWriteback(page)) {
1798 if (wbc->sync_mode != WB_SYNC_NONE)
1799 wait_on_page_writeback(page);
1800 else
1801 goto continue_unlock;
1802 }
1803
1804 BUG_ON(PageWriteback(page));
1805 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001806 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001807
Dave Chinner9e094382010-07-07 13:24:08 +10001808 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001809 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001810 if (unlikely(ret)) {
1811 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1812 unlock_page(page);
1813 ret = 0;
1814 } else {
1815 /*
1816 * done_index is set past this page,
1817 * so media errors will not choke
1818 * background writeout for the entire
1819 * file. This has consequences for
1820 * range_cyclic semantics (ie. it may
1821 * not be suitable for data integrity
1822 * writeout).
1823 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001824 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001825 done = 1;
1826 break;
1827 }
Dave Chinner0b564922010-06-09 10:37:18 +10001828 }
David Howells811d7362006-08-29 19:06:09 +01001829
Dave Chinner546a1922010-08-24 11:44:34 +10001830 /*
1831 * We stop writing back only if we are not doing
1832 * integrity sync. In case of integrity sync we have to
1833 * keep going until we have written all the pages
1834 * we tagged for writeback prior to entering this loop.
1835 */
1836 if (--wbc->nr_to_write <= 0 &&
1837 wbc->sync_mode == WB_SYNC_NONE) {
1838 done = 1;
1839 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001840 }
David Howells811d7362006-08-29 19:06:09 +01001841 }
1842 pagevec_release(&pvec);
1843 cond_resched();
1844 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001845 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001846 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001847 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001848 * We hit the last page and there is more work to be done: wrap
1849 * back to the start of the file
1850 */
Nick Piggin31a12662009-01-06 14:39:04 -08001851 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001852 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001853 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001854 goto retry;
1855 }
Dave Chinner0b564922010-06-09 10:37:18 +10001856 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1857 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001858
David Howells811d7362006-08-29 19:06:09 +01001859 return ret;
1860}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001861EXPORT_SYMBOL(write_cache_pages);
1862
1863/*
1864 * Function used by generic_writepages to call the real writepage
1865 * function and set the mapping flags on error
1866 */
1867static int __writepage(struct page *page, struct writeback_control *wbc,
1868 void *data)
1869{
1870 struct address_space *mapping = data;
1871 int ret = mapping->a_ops->writepage(page, wbc);
1872 mapping_set_error(mapping, ret);
1873 return ret;
1874}
1875
1876/**
1877 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1878 * @mapping: address space structure to write
1879 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1880 *
1881 * This is a library function, which implements the writepages()
1882 * address_space_operation.
1883 */
1884int generic_writepages(struct address_space *mapping,
1885 struct writeback_control *wbc)
1886{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001887 struct blk_plug plug;
1888 int ret;
1889
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001890 /* deal with chardevs and other special file */
1891 if (!mapping->a_ops->writepage)
1892 return 0;
1893
Shaohua Li9b6096a2011-03-17 10:47:06 +01001894 blk_start_plug(&plug);
1895 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1896 blk_finish_plug(&plug);
1897 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001898}
David Howells811d7362006-08-29 19:06:09 +01001899
1900EXPORT_SYMBOL(generic_writepages);
1901
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1903{
Andrew Morton22905f72005-11-16 15:07:01 -08001904 int ret;
1905
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 if (wbc->nr_to_write <= 0)
1907 return 0;
1908 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001909 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001910 else
1911 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001912 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913}
1914
1915/**
1916 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001917 * @page: the page to write
1918 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 *
1920 * The page must be locked by the caller and will be unlocked upon return.
1921 *
1922 * write_one_page() returns a negative error code if I/O failed.
1923 */
1924int write_one_page(struct page *page, int wait)
1925{
1926 struct address_space *mapping = page->mapping;
1927 int ret = 0;
1928 struct writeback_control wbc = {
1929 .sync_mode = WB_SYNC_ALL,
1930 .nr_to_write = 1,
1931 };
1932
1933 BUG_ON(!PageLocked(page));
1934
1935 if (wait)
1936 wait_on_page_writeback(page);
1937
1938 if (clear_page_dirty_for_io(page)) {
1939 page_cache_get(page);
1940 ret = mapping->a_ops->writepage(page, &wbc);
1941 if (ret == 0 && wait) {
1942 wait_on_page_writeback(page);
1943 if (PageError(page))
1944 ret = -EIO;
1945 }
1946 page_cache_release(page);
1947 } else {
1948 unlock_page(page);
1949 }
1950 return ret;
1951}
1952EXPORT_SYMBOL(write_one_page);
1953
1954/*
Ken Chen76719322007-02-10 01:43:15 -08001955 * For address_spaces which do not use buffers nor write back.
1956 */
1957int __set_page_dirty_no_writeback(struct page *page)
1958{
1959 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001960 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001961 return 0;
1962}
1963
1964/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001965 * Helper function for set_page_dirty family.
1966 * NOTE: This relies on being atomic wrt interrupts.
1967 */
1968void account_page_dirtied(struct page *page, struct address_space *mapping)
1969{
1970 if (mapping_cap_account_dirty(mapping)) {
1971 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001972 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001973 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06001974 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001975 task_io_account_write(PAGE_CACHE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001976 current->nr_dirtied++;
1977 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001978 }
1979}
Michael Rubin679ceac2010-08-20 02:31:26 -07001980EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001981
1982/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001983 * Helper function for set_page_writeback family.
1984 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1985 * wrt interrupts.
1986 */
1987void account_page_writeback(struct page *page)
1988{
1989 inc_zone_page_state(page, NR_WRITEBACK);
1990}
1991EXPORT_SYMBOL(account_page_writeback);
1992
1993/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 * For address_spaces which do not use buffers. Just tag the page as dirty in
1995 * its radix tree.
1996 *
1997 * This is also used when a single buffer is being dirtied: we want to set the
1998 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1999 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2000 *
2001 * Most callers have locked the page, which pins the address_space in memory.
2002 * But zap_pte_range() does not lock the page, however in that case the
2003 * mapping is pinned by the vma's ->vm_file reference.
2004 *
2005 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02002006 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007 */
2008int __set_page_dirty_nobuffers(struct page *page)
2009{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 if (!TestSetPageDirty(page)) {
2011 struct address_space *mapping = page_mapping(page);
2012 struct address_space *mapping2;
2013
Andrew Morton8c085402006-12-10 02:19:24 -08002014 if (!mapping)
2015 return 1;
2016
Nick Piggin19fd6232008-07-25 19:45:32 -07002017 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08002018 mapping2 = page_mapping(page);
2019 if (mapping2) { /* Race with truncate? */
2020 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07002021 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002022 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08002023 radix_tree_tag_set(&mapping->page_tree,
2024 page_index(page), PAGECACHE_TAG_DIRTY);
2025 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002026 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08002027 if (mapping->host) {
2028 /* !PageAnon && !swapper_space */
2029 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002031 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002033 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034}
2035EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2036
2037/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002038 * Call this whenever redirtying a page, to de-account the dirty counters
2039 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2040 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2041 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2042 * control.
2043 */
2044void account_page_redirty(struct page *page)
2045{
2046 struct address_space *mapping = page->mapping;
2047 if (mapping && mapping_cap_account_dirty(mapping)) {
2048 current->nr_dirtied--;
2049 dec_zone_page_state(page, NR_DIRTIED);
2050 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2051 }
2052}
2053EXPORT_SYMBOL(account_page_redirty);
2054
2055/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 * When a writepage implementation decides that it doesn't want to write this
2057 * page for some reason, it should redirty the locked page via
2058 * redirty_page_for_writepage() and it should then unlock the page and return 0
2059 */
2060int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2061{
2062 wbc->pages_skipped++;
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002063 account_page_redirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 return __set_page_dirty_nobuffers(page);
2065}
2066EXPORT_SYMBOL(redirty_page_for_writepage);
2067
2068/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002069 * Dirty a page.
2070 *
2071 * For pages with a mapping this should be done under the page lock
2072 * for the benefit of asynchronous memory errors who prefer a consistent
2073 * dirty state. This rule can be broken in some special cases,
2074 * but should be better not to.
2075 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 * If the mapping doesn't provide a set_page_dirty a_op, then
2077 * just fall through and assume that it wants buffer_heads.
2078 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002079int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080{
2081 struct address_space *mapping = page_mapping(page);
2082
2083 if (likely(mapping)) {
2084 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002085 /*
2086 * readahead/lru_deactivate_page could remain
2087 * PG_readahead/PG_reclaim due to race with end_page_writeback
2088 * About readahead, if the page is written, the flags would be
2089 * reset. So no problem.
2090 * About lru_deactivate_page, if the page is redirty, the flag
2091 * will be reset. So no problem. but if the page is used by readahead
2092 * it will confuse readahead and make it restart the size rampup
2093 * process. But it's a trivial problem.
2094 */
2095 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002096#ifdef CONFIG_BLOCK
2097 if (!spd)
2098 spd = __set_page_dirty_buffers;
2099#endif
2100 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002102 if (!PageDirty(page)) {
2103 if (!TestSetPageDirty(page))
2104 return 1;
2105 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 return 0;
2107}
2108EXPORT_SYMBOL(set_page_dirty);
2109
2110/*
2111 * set_page_dirty() is racy if the caller has no reference against
2112 * page->mapping->host, and if the page is unlocked. This is because another
2113 * CPU could truncate the page off the mapping and then free the mapping.
2114 *
2115 * Usually, the page _is_ locked, or the caller is a user-space process which
2116 * holds a reference on the inode by having an open file.
2117 *
2118 * In other cases, the page should be locked before running set_page_dirty().
2119 */
2120int set_page_dirty_lock(struct page *page)
2121{
2122 int ret;
2123
Jens Axboe7eaceac2011-03-10 08:52:07 +01002124 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 ret = set_page_dirty(page);
2126 unlock_page(page);
2127 return ret;
2128}
2129EXPORT_SYMBOL(set_page_dirty_lock);
2130
2131/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 * Clear a page's dirty flag, while caring for dirty memory accounting.
2133 * Returns true if the page was previously dirty.
2134 *
2135 * This is for preparing to put the page under writeout. We leave the page
2136 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2137 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2138 * implementation will run either set_page_writeback() or set_page_dirty(),
2139 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2140 * back into sync.
2141 *
2142 * This incoherency between the page's dirty flag and radix-tree tag is
2143 * unfortunate, but it only exists while the page is locked.
2144 */
2145int clear_page_dirty_for_io(struct page *page)
2146{
2147 struct address_space *mapping = page_mapping(page);
2148
Nick Piggin79352892007-07-19 01:47:22 -07002149 BUG_ON(!PageLocked(page));
2150
Linus Torvalds7658cc22006-12-29 10:00:58 -08002151 if (mapping && mapping_cap_account_dirty(mapping)) {
2152 /*
2153 * Yes, Virginia, this is indeed insane.
2154 *
2155 * We use this sequence to make sure that
2156 * (a) we account for dirty stats properly
2157 * (b) we tell the low-level filesystem to
2158 * mark the whole page dirty if it was
2159 * dirty in a pagetable. Only to then
2160 * (c) clean the page again and return 1 to
2161 * cause the writeback.
2162 *
2163 * This way we avoid all nasty races with the
2164 * dirty bit in multiple places and clearing
2165 * them concurrently from different threads.
2166 *
2167 * Note! Normally the "set_page_dirty(page)"
2168 * has no effect on the actual dirty bit - since
2169 * that will already usually be set. But we
2170 * need the side effects, and it can help us
2171 * avoid races.
2172 *
2173 * We basically use the page "master dirty bit"
2174 * as a serialization point for all the different
2175 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002176 */
2177 if (page_mkclean(page))
2178 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002179 /*
2180 * We carefully synchronise fault handlers against
2181 * installing a dirty pte and marking the page dirty
2182 * at this point. We do this by having them hold the
2183 * page lock at some point after installing their
2184 * pte, but before marking the page dirty.
2185 * Pages are always locked coming in here, so we get
2186 * the desired exclusion. See mm/memory.c:do_wp_page()
2187 * for more comments.
2188 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08002189 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08002190 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07002191 dec_bdi_stat(mapping->backing_dev_info,
2192 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002193 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002195 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002197 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002199EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200
2201int test_clear_page_writeback(struct page *page)
2202{
2203 struct address_space *mapping = page_mapping(page);
2204 int ret;
2205
2206 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002207 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 unsigned long flags;
2209
Nick Piggin19fd6232008-07-25 19:45:32 -07002210 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002212 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 radix_tree_tag_clear(&mapping->page_tree,
2214 page_index(page),
2215 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002216 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002217 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002218 __bdi_writeout_inc(bdi);
2219 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002220 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002221 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 } else {
2223 ret = TestClearPageWriteback(page);
2224 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002225 if (ret) {
Andrew Mortond688abf2007-07-19 01:49:17 -07002226 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002227 inc_zone_page_state(page, NR_WRITTEN);
2228 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 return ret;
2230}
2231
2232int test_set_page_writeback(struct page *page)
2233{
2234 struct address_space *mapping = page_mapping(page);
2235 int ret;
2236
2237 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002238 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 unsigned long flags;
2240
Nick Piggin19fd6232008-07-25 19:45:32 -07002241 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002243 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 radix_tree_tag_set(&mapping->page_tree,
2245 page_index(page),
2246 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002247 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002248 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2249 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 if (!PageDirty(page))
2251 radix_tree_tag_clear(&mapping->page_tree,
2252 page_index(page),
2253 PAGECACHE_TAG_DIRTY);
Jan Karaf446daae2010-08-09 17:19:12 -07002254 radix_tree_tag_clear(&mapping->page_tree,
2255 page_index(page),
2256 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002257 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 } else {
2259 ret = TestSetPageWriteback(page);
2260 }
Andrew Mortond688abf2007-07-19 01:49:17 -07002261 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07002262 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 return ret;
2264
2265}
2266EXPORT_SYMBOL(test_set_page_writeback);
2267
2268/*
Nick Piggin00128182007-10-16 01:24:40 -07002269 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 * passed tag.
2271 */
2272int mapping_tagged(struct address_space *mapping, int tag)
2273{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002274 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275}
2276EXPORT_SYMBOL(mapping_tagged);