blob: 5cdd4f2b0c9d8b07188181badd2ecb3b76de4b75 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
Francois Camie1f8e872008-10-15 22:01:59 -070010 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Initial version
12 */
13
14#include <linux/kernel.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040015#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
Al Viroff01bb42011-09-16 02:31:11 -040035#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100037#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038
39/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060040 * Sleep at most 200ms at a time in balance_dirty_pages().
41 */
42#define MAX_PAUSE max(HZ/5, 1)
43
44/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060045 * Estimate write bandwidth at 200ms intervals.
46 */
47#define BANDWIDTH_INTERVAL max(HZ/5, 1)
48
Wu Fengguang6c14ae12011-03-02 16:04:18 -060049#define RATELIMIT_CALC_SHIFT 10
50
Wu Fengguange98be2d2010-08-29 11:22:30 -060051/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
53 * will look to see if it needs to force writeback or throttling.
54 */
55static long ratelimit_pages = 32;
56
Linus Torvalds1da177e2005-04-16 15:20:36 -070057/* The following parameters are exported via /proc/sys/vm */
58
59/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020060 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080062int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070063
64/*
David Rientjes2da02992009-01-06 14:39:31 -080065 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
66 * dirty_background_ratio * the amount of dirtyable memory
67 */
68unsigned long dirty_background_bytes;
69
70/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080071 * free highmem will not be subtracted from the total free memory
72 * for calculating free ratios if vm_highmem_is_dirtyable is true
73 */
74int vm_highmem_is_dirtyable;
75
76/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070077 * The generator of dirty data starts writeback at this percentage
78 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080079int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070080
81/*
David Rientjes2da02992009-01-06 14:39:31 -080082 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
83 * vm_dirty_ratio * the amount of dirtyable memory
84 */
85unsigned long vm_dirty_bytes;
86
87/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070088 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070090unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070093 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -070094 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070095unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
97/*
98 * Flag that makes the machine dump writes/reads and block dirtyings.
99 */
100int block_dump;
101
102/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800103 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
104 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105 */
106int laptop_mode;
107
108EXPORT_SYMBOL(laptop_mode);
109
110/* End of sysctl-exported parameters */
111
Wu Fengguangc42843f2011-03-02 15:54:09 -0600112unsigned long global_dirty_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700115 * Scale the writeback cache size proportional to the relative writeout speeds.
116 *
117 * We do this by keeping a floating proportion between BDIs, based on page
118 * writeback completions [end_page_writeback()]. Those devices that write out
119 * pages fastest will get the larger share, while the slower will get a smaller
120 * share.
121 *
122 * We use page writeout completions because we are interested in getting rid of
123 * dirty pages. Having them written out is the primary goal.
124 *
125 * We introduce a concept of time, a period over which we measure these events,
126 * because demand can/will vary over time. The length of this period itself is
127 * measured in page writeback completions.
128 *
129 */
130static struct prop_descriptor vm_completions;
131
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700132/*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800133 * Work out the current dirty-memory clamping and background writeout
134 * thresholds.
135 *
136 * The main aim here is to lower them aggressively if there is a lot of mapped
137 * memory around. To avoid stressing page reclaim with lots of unreclaimable
138 * pages. It is better to clamp down on writers than to start swapping, and
139 * performing lots of scanning.
140 *
141 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
142 *
143 * We don't permit the clamping level to fall below 5% - that is getting rather
144 * excessive.
145 *
146 * We make sure that the background writeout level is below the adjusted
147 * clamping level.
148 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800149
Johannes Weinera756cf52012-01-10 15:07:49 -0800150/*
151 * In a memory zone, there is a certain amount of pages we consider
152 * available for the page cache, which is essentially the number of
153 * free and reclaimable pages, minus some zone reserves to protect
154 * lowmem and the ability to uphold the zone's watermarks without
155 * requiring writeback.
156 *
157 * This number of dirtyable pages is the base value of which the
158 * user-configurable dirty ratio is the effictive number of pages that
159 * are allowed to be actually dirtied. Per individual zone, or
160 * globally by using the sum of dirtyable pages over all zones.
161 *
162 * Because the user is allowed to specify the dirty limit globally as
163 * absolute number of bytes, calculating the per-zone dirty limit can
164 * require translating the configured limit into a percentage of
165 * global dirtyable memory first.
166 */
167
Johannes Weiner1edf2232012-01-10 15:06:57 -0800168static unsigned long highmem_dirtyable_memory(unsigned long total)
169{
170#ifdef CONFIG_HIGHMEM
171 int node;
172 unsigned long x = 0;
173
174 for_each_node_state(node, N_HIGH_MEMORY) {
175 struct zone *z =
176 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
177
178 x += zone_page_state(z, NR_FREE_PAGES) +
Johannes Weinerab8fabd2012-01-10 15:07:42 -0800179 zone_reclaimable_pages(z) - z->dirty_balance_reserve;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800180 }
181 /*
182 * Make sure that the number of highmem pages is never larger
183 * than the number of the total dirtyable memory. This can only
184 * occur in very strange VM situations but we want to make sure
185 * that this does not occur.
186 */
187 return min(x, total);
188#else
189 return 0;
190#endif
191}
192
193/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800194 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800195 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800196 * Returns the global number of pages potentially available for dirty
197 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800198 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800199unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800200{
201 unsigned long x;
202
Johannes Weinerab8fabd2012-01-10 15:07:42 -0800203 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
204 dirty_balance_reserve;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800205
206 if (!vm_highmem_is_dirtyable)
207 x -= highmem_dirtyable_memory(x);
208
209 return x + 1; /* Ensure that we never return 0 */
210}
211
212/*
Johannes Weinerccafa282012-01-10 15:07:44 -0800213 * global_dirty_limits - background-writeback and dirty-throttling thresholds
214 *
215 * Calculate the dirty thresholds based on sysctl parameters
216 * - vm.dirty_background_ratio or vm.dirty_background_bytes
217 * - vm.dirty_ratio or vm.dirty_bytes
218 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
219 * real-time tasks.
220 */
221void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
222{
223 unsigned long background;
224 unsigned long dirty;
225 unsigned long uninitialized_var(available_memory);
226 struct task_struct *tsk;
227
228 if (!vm_dirty_bytes || !dirty_background_bytes)
229 available_memory = global_dirtyable_memory();
230
231 if (vm_dirty_bytes)
232 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
233 else
234 dirty = (vm_dirty_ratio * available_memory) / 100;
235
236 if (dirty_background_bytes)
237 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
238 else
239 background = (dirty_background_ratio * available_memory) / 100;
240
241 if (background >= dirty)
242 background = dirty / 2;
243 tsk = current;
244 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
245 background += background / 4;
246 dirty += dirty / 4;
247 }
248 *pbackground = background;
249 *pdirty = dirty;
250 trace_global_dirty_state(background, dirty);
251}
252
Johannes Weinera756cf52012-01-10 15:07:49 -0800253/**
254 * zone_dirtyable_memory - number of dirtyable pages in a zone
255 * @zone: the zone
256 *
257 * Returns the zone's number of pages potentially available for dirty
258 * page cache. This is the base value for the per-zone dirty limits.
259 */
260static unsigned long zone_dirtyable_memory(struct zone *zone)
261{
262 /*
263 * The effective global number of dirtyable pages may exclude
264 * highmem as a big-picture measure to keep the ratio between
265 * dirty memory and lowmem reasonable.
266 *
267 * But this function is purely about the individual zone and a
268 * highmem zone can hold its share of dirty pages, so we don't
269 * care about vm_highmem_is_dirtyable here.
270 */
271 return zone_page_state(zone, NR_FREE_PAGES) +
272 zone_reclaimable_pages(zone) -
273 zone->dirty_balance_reserve;
274}
275
276/**
277 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
278 * @zone: the zone
279 *
280 * Returns the maximum number of dirty pages allowed in a zone, based
281 * on the zone's dirtyable memory.
282 */
283static unsigned long zone_dirty_limit(struct zone *zone)
284{
285 unsigned long zone_memory = zone_dirtyable_memory(zone);
286 struct task_struct *tsk = current;
287 unsigned long dirty;
288
289 if (vm_dirty_bytes)
290 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
291 zone_memory / global_dirtyable_memory();
292 else
293 dirty = vm_dirty_ratio * zone_memory / 100;
294
295 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
296 dirty += dirty / 4;
297
298 return dirty;
299}
300
301/**
302 * zone_dirty_ok - tells whether a zone is within its dirty limits
303 * @zone: the zone to check
304 *
305 * Returns %true when the dirty pages in @zone are within the zone's
306 * dirty limit, %false if the limit is exceeded.
307 */
308bool zone_dirty_ok(struct zone *zone)
309{
310 unsigned long limit = zone_dirty_limit(zone);
311
312 return zone_page_state(zone, NR_FILE_DIRTY) +
313 zone_page_state(zone, NR_UNSTABLE_NFS) +
314 zone_page_state(zone, NR_WRITEBACK) <= limit;
315}
316
Johannes Weinerccafa282012-01-10 15:07:44 -0800317/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700318 * couple the period to the dirty_ratio:
319 *
320 * period/2 ~ roundup_pow_of_two(dirty limit)
321 */
322static int calc_period_shift(void)
323{
324 unsigned long dirty_total;
325
David Rientjes2da02992009-01-06 14:39:31 -0800326 if (vm_dirty_bytes)
327 dirty_total = vm_dirty_bytes / PAGE_SIZE;
328 else
Johannes Weinerccafa282012-01-10 15:07:44 -0800329 dirty_total = (vm_dirty_ratio * global_dirtyable_memory()) /
David Rientjes2da02992009-01-06 14:39:31 -0800330 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700331 return 2 + ilog2(dirty_total - 1);
332}
333
334/*
David Rientjes2da02992009-01-06 14:39:31 -0800335 * update the period when the dirty threshold changes.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700336 */
David Rientjes2da02992009-01-06 14:39:31 -0800337static void update_completion_period(void)
338{
339 int shift = calc_period_shift();
340 prop_change_shift(&vm_completions, shift);
Wu Fengguang9d823e82011-06-11 18:10:12 -0600341
342 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800343}
344
345int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700346 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800347 loff_t *ppos)
348{
349 int ret;
350
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700351 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800352 if (ret == 0 && write)
353 dirty_background_bytes = 0;
354 return ret;
355}
356
357int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700358 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800359 loff_t *ppos)
360{
361 int ret;
362
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700363 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800364 if (ret == 0 && write)
365 dirty_background_ratio = 0;
366 return ret;
367}
368
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700369int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700370 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700371 loff_t *ppos)
372{
373 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800374 int ret;
375
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700376 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700377 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
David Rientjes2da02992009-01-06 14:39:31 -0800378 update_completion_period();
379 vm_dirty_bytes = 0;
380 }
381 return ret;
382}
383
David Rientjes2da02992009-01-06 14:39:31 -0800384int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700385 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800386 loff_t *ppos)
387{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800388 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800389 int ret;
390
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700391 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800392 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
393 update_completion_period();
394 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700395 }
396 return ret;
397}
398
399/*
400 * Increment the BDI's writeout completion count and the global writeout
401 * completion count. Called from test_clear_page_writeback().
402 */
403static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
404{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600405 __inc_bdi_stat(bdi, BDI_WRITTEN);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700406 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
407 bdi->max_prop_frac);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700408}
409
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700410void bdi_writeout_inc(struct backing_dev_info *bdi)
411{
412 unsigned long flags;
413
414 local_irq_save(flags);
415 __bdi_writeout_inc(bdi);
416 local_irq_restore(flags);
417}
418EXPORT_SYMBOL_GPL(bdi_writeout_inc);
419
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700420/*
421 * Obtain an accurate fraction of the BDI's portion.
422 */
423static void bdi_writeout_fraction(struct backing_dev_info *bdi,
424 long *numerator, long *denominator)
425{
Wu Fengguang3efaf0f2010-12-16 22:22:00 -0600426 prop_fraction_percpu(&vm_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700427 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700428}
429
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700430/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700431 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
432 * registered backing devices, which, for obvious reasons, can not
433 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700434 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700435static unsigned int bdi_min_ratio;
436
437int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
438{
439 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700440
Jens Axboecfc4ba52009-09-14 13:12:40 +0200441 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700442 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700443 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700444 } else {
445 min_ratio -= bdi->min_ratio;
446 if (bdi_min_ratio + min_ratio < 100) {
447 bdi_min_ratio += min_ratio;
448 bdi->min_ratio += min_ratio;
449 } else {
450 ret = -EINVAL;
451 }
452 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200453 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700454
455 return ret;
456}
457
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700458int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
459{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700460 int ret = 0;
461
462 if (max_ratio > 100)
463 return -EINVAL;
464
Jens Axboecfc4ba52009-09-14 13:12:40 +0200465 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700466 if (bdi->min_ratio > max_ratio) {
467 ret = -EINVAL;
468 } else {
469 bdi->max_ratio = max_ratio;
470 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
471 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200472 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700473
474 return ret;
475}
476EXPORT_SYMBOL(bdi_set_max_ratio);
477
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600478static unsigned long dirty_freerun_ceiling(unsigned long thresh,
479 unsigned long bg_thresh)
480{
481 return (thresh + bg_thresh) / 2;
482}
483
Wu Fengguangffd1f602011-06-19 22:18:42 -0600484static unsigned long hard_dirty_limit(unsigned long thresh)
485{
486 return max(thresh, global_dirty_limit);
487}
488
Wu Fengguang6f718652011-03-02 17:14:34 -0600489/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700490 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600491 * @bdi: the backing_dev_info to query
492 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700493 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600494 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
495 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600496 *
497 * Note that balance_dirty_pages() will only seriously take it as a hard limit
498 * when sleeping max_pause per page is not enough to keep the dirty pages under
499 * control. For example, when the device is completely stalled due to some error
500 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
501 * In the other normal situations, it acts more gently by throttling the tasks
502 * more (rather than completely block them) when the bdi dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600503 *
504 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700505 * - starving fast devices
506 * - piling up dirty pages (that will take long time to sync) on slow devices
507 *
508 * The bdi's share of dirty limit will be adapting to its throughput and
509 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
510 */
511unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700512{
513 u64 bdi_dirty;
514 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700515
Wu Fengguang16c40422010-08-11 14:17:39 -0700516 /*
517 * Calculate this BDI's share of the dirty ratio.
518 */
519 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700520
Wu Fengguang16c40422010-08-11 14:17:39 -0700521 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
522 bdi_dirty *= numerator;
523 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700524
Wu Fengguang16c40422010-08-11 14:17:39 -0700525 bdi_dirty += (dirty * bdi->min_ratio) / 100;
526 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
527 bdi_dirty = dirty * bdi->max_ratio / 100;
528
529 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530}
531
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600532/*
533 * Dirty position control.
534 *
535 * (o) global/bdi setpoints
536 *
537 * We want the dirty pages be balanced around the global/bdi setpoints.
538 * When the number of dirty pages is higher/lower than the setpoint, the
539 * dirty position control ratio (and hence task dirty ratelimit) will be
540 * decreased/increased to bring the dirty pages back to the setpoint.
541 *
542 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
543 *
544 * if (dirty < setpoint) scale up pos_ratio
545 * if (dirty > setpoint) scale down pos_ratio
546 *
547 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
548 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
549 *
550 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
551 *
552 * (o) global control line
553 *
554 * ^ pos_ratio
555 * |
556 * | |<===== global dirty control scope ======>|
557 * 2.0 .............*
558 * | .*
559 * | . *
560 * | . *
561 * | . *
562 * | . *
563 * | . *
564 * 1.0 ................................*
565 * | . . *
566 * | . . *
567 * | . . *
568 * | . . *
569 * | . . *
570 * 0 +------------.------------------.----------------------*------------->
571 * freerun^ setpoint^ limit^ dirty pages
572 *
573 * (o) bdi control line
574 *
575 * ^ pos_ratio
576 * |
577 * | *
578 * | *
579 * | *
580 * | *
581 * | * |<=========== span ============>|
582 * 1.0 .......................*
583 * | . *
584 * | . *
585 * | . *
586 * | . *
587 * | . *
588 * | . *
589 * | . *
590 * | . *
591 * | . *
592 * | . *
593 * | . *
594 * 1/4 ...............................................* * * * * * * * * * * *
595 * | . .
596 * | . .
597 * | . .
598 * 0 +----------------------.-------------------------------.------------->
599 * bdi_setpoint^ x_intercept^
600 *
601 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
602 * be smoothly throttled down to normal if it starts high in situations like
603 * - start writing to a slow SD card and a fast disk at the same time. The SD
604 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
605 * - the bdi dirty thresh drops quickly due to change of JBOD workload
606 */
607static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
608 unsigned long thresh,
609 unsigned long bg_thresh,
610 unsigned long dirty,
611 unsigned long bdi_thresh,
612 unsigned long bdi_dirty)
613{
614 unsigned long write_bw = bdi->avg_write_bandwidth;
615 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
616 unsigned long limit = hard_dirty_limit(thresh);
617 unsigned long x_intercept;
618 unsigned long setpoint; /* dirty pages' target balance point */
619 unsigned long bdi_setpoint;
620 unsigned long span;
621 long long pos_ratio; /* for scaling up/down the rate limit */
622 long x;
623
624 if (unlikely(dirty >= limit))
625 return 0;
626
627 /*
628 * global setpoint
629 *
630 * setpoint - dirty 3
631 * f(dirty) := 1.0 + (----------------)
632 * limit - setpoint
633 *
634 * it's a 3rd order polynomial that subjects to
635 *
636 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
637 * (2) f(setpoint) = 1.0 => the balance point
638 * (3) f(limit) = 0 => the hard limit
639 * (4) df/dx <= 0 => negative feedback control
640 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
641 * => fast response on large errors; small oscillation near setpoint
642 */
643 setpoint = (freerun + limit) / 2;
644 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
645 limit - setpoint + 1);
646 pos_ratio = x;
647 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
648 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
649 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
650
651 /*
652 * We have computed basic pos_ratio above based on global situation. If
653 * the bdi is over/under its share of dirty pages, we want to scale
654 * pos_ratio further down/up. That is done by the following mechanism.
655 */
656
657 /*
658 * bdi setpoint
659 *
660 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
661 *
662 * x_intercept - bdi_dirty
663 * := --------------------------
664 * x_intercept - bdi_setpoint
665 *
666 * The main bdi control line is a linear function that subjects to
667 *
668 * (1) f(bdi_setpoint) = 1.0
669 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
670 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
671 *
672 * For single bdi case, the dirty pages are observed to fluctuate
673 * regularly within range
674 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
675 * for various filesystems, where (2) can yield in a reasonable 12.5%
676 * fluctuation range for pos_ratio.
677 *
678 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
679 * own size, so move the slope over accordingly and choose a slope that
680 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
681 */
682 if (unlikely(bdi_thresh > thresh))
683 bdi_thresh = thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600684 /*
685 * It's very possible that bdi_thresh is close to 0 not because the
686 * device is slow, but that it has remained inactive for long time.
687 * Honour such devices a reasonable good (hopefully IO efficient)
688 * threshold, so that the occasional writes won't be blocked and active
689 * writes can rampup the threshold quickly.
690 */
Wu Fengguang8927f662011-08-04 22:16:46 -0600691 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600692 /*
693 * scale global setpoint to bdi's:
694 * bdi_setpoint = setpoint * bdi_thresh / thresh
695 */
696 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
697 bdi_setpoint = setpoint * (u64)x >> 16;
698 /*
699 * Use span=(8*write_bw) in single bdi case as indicated by
700 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
701 *
702 * bdi_thresh thresh - bdi_thresh
703 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
704 * thresh thresh
705 */
706 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
707 x_intercept = bdi_setpoint + span;
708
709 if (bdi_dirty < x_intercept - span / 4) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600710 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
711 x_intercept - bdi_setpoint + 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600712 } else
713 pos_ratio /= 4;
714
Wu Fengguang8927f662011-08-04 22:16:46 -0600715 /*
716 * bdi reserve area, safeguard against dirty pool underrun and disk idle
717 * It may push the desired control point of global dirty pages higher
718 * than setpoint.
719 */
720 x_intercept = bdi_thresh / 2;
721 if (bdi_dirty < x_intercept) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600722 if (bdi_dirty > x_intercept / 8)
723 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
724 else
Wu Fengguang8927f662011-08-04 22:16:46 -0600725 pos_ratio *= 8;
726 }
727
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600728 return pos_ratio;
729}
730
Wu Fengguange98be2d2010-08-29 11:22:30 -0600731static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
732 unsigned long elapsed,
733 unsigned long written)
734{
735 const unsigned long period = roundup_pow_of_two(3 * HZ);
736 unsigned long avg = bdi->avg_write_bandwidth;
737 unsigned long old = bdi->write_bandwidth;
738 u64 bw;
739
740 /*
741 * bw = written * HZ / elapsed
742 *
743 * bw * elapsed + write_bandwidth * (period - elapsed)
744 * write_bandwidth = ---------------------------------------------------
745 * period
746 */
747 bw = written - bdi->written_stamp;
748 bw *= HZ;
749 if (unlikely(elapsed > period)) {
750 do_div(bw, elapsed);
751 avg = bw;
752 goto out;
753 }
754 bw += (u64)bdi->write_bandwidth * (period - elapsed);
755 bw >>= ilog2(period);
756
757 /*
758 * one more level of smoothing, for filtering out sudden spikes
759 */
760 if (avg > old && old >= (unsigned long)bw)
761 avg -= (avg - old) >> 3;
762
763 if (avg < old && old <= (unsigned long)bw)
764 avg += (old - avg) >> 3;
765
766out:
767 bdi->write_bandwidth = bw;
768 bdi->avg_write_bandwidth = avg;
769}
770
Wu Fengguangc42843f2011-03-02 15:54:09 -0600771/*
772 * The global dirtyable memory and dirty threshold could be suddenly knocked
773 * down by a large amount (eg. on the startup of KVM in a swapless system).
774 * This may throw the system into deep dirty exceeded state and throttle
775 * heavy/light dirtiers alike. To retain good responsiveness, maintain
776 * global_dirty_limit for tracking slowly down to the knocked down dirty
777 * threshold.
778 */
779static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
780{
781 unsigned long limit = global_dirty_limit;
782
783 /*
784 * Follow up in one step.
785 */
786 if (limit < thresh) {
787 limit = thresh;
788 goto update;
789 }
790
791 /*
792 * Follow down slowly. Use the higher one as the target, because thresh
793 * may drop below dirty. This is exactly the reason to introduce
794 * global_dirty_limit which is guaranteed to lie above the dirty pages.
795 */
796 thresh = max(thresh, dirty);
797 if (limit > thresh) {
798 limit -= (limit - thresh) >> 5;
799 goto update;
800 }
801 return;
802update:
803 global_dirty_limit = limit;
804}
805
806static void global_update_bandwidth(unsigned long thresh,
807 unsigned long dirty,
808 unsigned long now)
809{
810 static DEFINE_SPINLOCK(dirty_lock);
811 static unsigned long update_time;
812
813 /*
814 * check locklessly first to optimize away locking for the most time
815 */
816 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
817 return;
818
819 spin_lock(&dirty_lock);
820 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
821 update_dirty_limit(thresh, dirty);
822 update_time = now;
823 }
824 spin_unlock(&dirty_lock);
825}
826
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600827/*
828 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
829 *
830 * Normal bdi tasks will be curbed at or below it in long term.
831 * Obviously it should be around (write_bw / N) when there are N dd tasks.
832 */
833static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
834 unsigned long thresh,
835 unsigned long bg_thresh,
836 unsigned long dirty,
837 unsigned long bdi_thresh,
838 unsigned long bdi_dirty,
839 unsigned long dirtied,
840 unsigned long elapsed)
841{
Wu Fengguang73811312011-08-26 15:53:24 -0600842 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
843 unsigned long limit = hard_dirty_limit(thresh);
844 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600845 unsigned long write_bw = bdi->avg_write_bandwidth;
846 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
847 unsigned long dirty_rate;
848 unsigned long task_ratelimit;
849 unsigned long balanced_dirty_ratelimit;
850 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600851 unsigned long step;
852 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600853
854 /*
855 * The dirty rate will match the writeout rate in long term, except
856 * when dirty pages are truncated by userspace or re-dirtied by FS.
857 */
858 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
859
860 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
861 bdi_thresh, bdi_dirty);
862 /*
863 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
864 */
865 task_ratelimit = (u64)dirty_ratelimit *
866 pos_ratio >> RATELIMIT_CALC_SHIFT;
867 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
868
869 /*
870 * A linear estimation of the "balanced" throttle rate. The theory is,
871 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
872 * dirty_rate will be measured to be (N * task_ratelimit). So the below
873 * formula will yield the balanced rate limit (write_bw / N).
874 *
875 * Note that the expanded form is not a pure rate feedback:
876 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
877 * but also takes pos_ratio into account:
878 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
879 *
880 * (1) is not realistic because pos_ratio also takes part in balancing
881 * the dirty rate. Consider the state
882 * pos_ratio = 0.5 (3)
883 * rate = 2 * (write_bw / N) (4)
884 * If (1) is used, it will stuck in that state! Because each dd will
885 * be throttled at
886 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
887 * yielding
888 * dirty_rate = N * task_ratelimit = write_bw (6)
889 * put (6) into (1) we get
890 * rate_(i+1) = rate_(i) (7)
891 *
892 * So we end up using (2) to always keep
893 * rate_(i+1) ~= (write_bw / N) (8)
894 * regardless of the value of pos_ratio. As long as (8) is satisfied,
895 * pos_ratio is able to drive itself to 1.0, which is not only where
896 * the dirty count meet the setpoint, but also where the slope of
897 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
898 */
899 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
900 dirty_rate | 1);
901
Wu Fengguang73811312011-08-26 15:53:24 -0600902 /*
903 * We could safely do this and return immediately:
904 *
905 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
906 *
907 * However to get a more stable dirty_ratelimit, the below elaborated
908 * code makes use of task_ratelimit to filter out sigular points and
909 * limit the step size.
910 *
911 * The below code essentially only uses the relative value of
912 *
913 * task_ratelimit - dirty_ratelimit
914 * = (pos_ratio - 1) * dirty_ratelimit
915 *
916 * which reflects the direction and size of dirty position error.
917 */
918
919 /*
920 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
921 * task_ratelimit is on the same side of dirty_ratelimit, too.
922 * For example, when
923 * - dirty_ratelimit > balanced_dirty_ratelimit
924 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
925 * lowering dirty_ratelimit will help meet both the position and rate
926 * control targets. Otherwise, don't update dirty_ratelimit if it will
927 * only help meet the rate target. After all, what the users ultimately
928 * feel and care are stable dirty rate and small position error.
929 *
930 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
931 * and filter out the sigular points of balanced_dirty_ratelimit. Which
932 * keeps jumping around randomly and can even leap far away at times
933 * due to the small 200ms estimation period of dirty_rate (we want to
934 * keep that period small to reduce time lags).
935 */
936 step = 0;
937 if (dirty < setpoint) {
938 x = min(bdi->balanced_dirty_ratelimit,
939 min(balanced_dirty_ratelimit, task_ratelimit));
940 if (dirty_ratelimit < x)
941 step = x - dirty_ratelimit;
942 } else {
943 x = max(bdi->balanced_dirty_ratelimit,
944 max(balanced_dirty_ratelimit, task_ratelimit));
945 if (dirty_ratelimit > x)
946 step = dirty_ratelimit - x;
947 }
948
949 /*
950 * Don't pursue 100% rate matching. It's impossible since the balanced
951 * rate itself is constantly fluctuating. So decrease the track speed
952 * when it gets close to the target. Helps eliminate pointless tremors.
953 */
954 step >>= dirty_ratelimit / (2 * step + 1);
955 /*
956 * Limit the tracking speed to avoid overshooting.
957 */
958 step = (step + 7) / 8;
959
960 if (dirty_ratelimit < balanced_dirty_ratelimit)
961 dirty_ratelimit += step;
962 else
963 dirty_ratelimit -= step;
964
965 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
966 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -0600967
968 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600969}
970
Wu Fengguange98be2d2010-08-29 11:22:30 -0600971void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600972 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600973 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600974 unsigned long dirty,
975 unsigned long bdi_thresh,
976 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600977 unsigned long start_time)
978{
979 unsigned long now = jiffies;
980 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600981 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600982 unsigned long written;
983
984 /*
985 * rate-limit, only update once every 200ms.
986 */
987 if (elapsed < BANDWIDTH_INTERVAL)
988 return;
989
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600990 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600991 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
992
993 /*
994 * Skip quiet periods when disk bandwidth is under-utilized.
995 * (at least 1s idle time between two flusher runs)
996 */
997 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
998 goto snapshot;
999
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001000 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -06001001 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001002 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1003 bdi_thresh, bdi_dirty,
1004 dirtied, elapsed);
1005 }
Wu Fengguange98be2d2010-08-29 11:22:30 -06001006 bdi_update_write_bandwidth(bdi, elapsed, written);
1007
1008snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001009 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001010 bdi->written_stamp = written;
1011 bdi->bw_time_stamp = now;
1012}
1013
1014static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001015 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001016 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001017 unsigned long dirty,
1018 unsigned long bdi_thresh,
1019 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -06001020 unsigned long start_time)
1021{
1022 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1023 return;
1024 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001025 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1026 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001027 spin_unlock(&bdi->wb.list_lock);
1028}
1029
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030/*
Wu Fengguang9d823e82011-06-11 18:10:12 -06001031 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1032 * will look to see if it needs to start dirty throttling.
1033 *
1034 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1035 * global_page_state() too often. So scale it near-sqrt to the safety margin
1036 * (the number of pages we may dirty without exceeding the dirty limits).
1037 */
1038static unsigned long dirty_poll_interval(unsigned long dirty,
1039 unsigned long thresh)
1040{
1041 if (thresh > dirty)
1042 return 1UL << (ilog2(thresh - dirty) >> 1);
1043
1044 return 1;
1045}
1046
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001047static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1048 unsigned long bdi_dirty)
1049{
1050 unsigned long bw = bdi->avg_write_bandwidth;
1051 unsigned long hi = ilog2(bw);
1052 unsigned long lo = ilog2(bdi->dirty_ratelimit);
1053 unsigned long t;
1054
1055 /* target for 20ms max pause on 1-dd case */
1056 t = HZ / 50;
1057
1058 /*
1059 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1060 * overheads.
1061 *
1062 * (N * 20ms) on 2^N concurrent tasks.
1063 */
1064 if (hi > lo)
1065 t += (hi - lo) * (20 * HZ) / 1024;
1066
1067 /*
1068 * Limit pause time for small memory systems. If sleeping for too long
1069 * time, a small pool of dirty/writeback pages may go empty and disk go
1070 * idle.
1071 *
1072 * 8 serves as the safety ratio.
1073 */
Wu Fengguang82e230a2011-12-02 18:21:51 -06001074 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001075
1076 /*
1077 * The pause time will be settled within range (max_pause/4, max_pause).
1078 * Apply a minimal value of 4 to get a non-zero max_pause/4.
1079 */
1080 return clamp_val(t, 4, MAX_PAUSE);
1081}
1082
Wu Fengguang9d823e82011-06-11 18:10:12 -06001083/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084 * balance_dirty_pages() must be called by processes which are generating dirty
1085 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001086 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001087 * If we're over `background_thresh' then the writeback threads are woken to
1088 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001090static void balance_dirty_pages(struct address_space *mapping,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001091 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092{
Wu Fengguang143dfe82010-08-27 18:45:12 -06001093 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1094 unsigned long bdi_reclaimable;
Wu Fengguang77627412010-09-12 13:34:05 -06001095 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1096 unsigned long bdi_dirty;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001097 unsigned long freerun;
David Rientjes364aeb22009-01-06 14:39:29 -08001098 unsigned long background_thresh;
1099 unsigned long dirty_thresh;
1100 unsigned long bdi_thresh;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001101 long pause = 0;
Wu Fengguang50657fc2011-10-11 17:06:33 -06001102 long uninitialized_var(max_pause);
Wu Fengguange50e3722010-08-11 14:17:37 -07001103 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001104 unsigned long task_ratelimit;
Wu Fengguang50657fc2011-10-11 17:06:33 -06001105 unsigned long uninitialized_var(dirty_ratelimit);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001106 unsigned long pos_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001108 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109
1110 for (;;) {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001111 /*
1112 * Unstable writes are a feature of certain networked
1113 * filesystems (i.e. NFS) in which data may have been
1114 * written to the server's write cache, but has not yet
1115 * been flushed to permanent storage.
1116 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001117 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1118 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -06001119 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001120
Wu Fengguang16c40422010-08-11 14:17:39 -07001121 global_dirty_limits(&background_thresh, &dirty_thresh);
1122
1123 /*
1124 * Throttle it only when the background writeback cannot
1125 * catch-up. This avoids (excessively) small writeouts
1126 * when the bdi limits are ramping up.
1127 */
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001128 freerun = dirty_freerun_ceiling(dirty_thresh,
1129 background_thresh);
1130 if (nr_dirty <= freerun)
Wu Fengguang16c40422010-08-11 14:17:39 -07001131 break;
1132
Wu Fengguang143dfe82010-08-27 18:45:12 -06001133 if (unlikely(!writeback_in_progress(bdi)))
1134 bdi_start_background_writeback(bdi);
1135
1136 /*
1137 * bdi_thresh is not treated as some limiting factor as
1138 * dirty_thresh, due to reasons
1139 * - in JBOD setup, bdi_thresh can fluctuate a lot
1140 * - in a system with HDD and USB key, the USB key may somehow
1141 * go into state (bdi_dirty >> bdi_thresh) either because
1142 * bdi_dirty starts high, or because bdi_thresh drops low.
1143 * In this case we don't want to hard throttle the USB key
1144 * dirtiers for 100 seconds until bdi_dirty drops under
1145 * bdi_thresh. Instead the auxiliary bdi control line in
1146 * bdi_position_ratio() will let the dirtier task progress
1147 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1148 */
Wu Fengguang16c40422010-08-11 14:17:39 -07001149 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001150
Wu Fengguange50e3722010-08-11 14:17:37 -07001151 /*
1152 * In order to avoid the stacked BDI deadlock we need
1153 * to ensure we accurately count the 'dirty' pages when
1154 * the threshold is low.
1155 *
1156 * Otherwise it would be possible to get thresh+n pages
1157 * reported dirty, even though there are thresh-m pages
1158 * actually dirty; with m+n sitting in the percpu
1159 * deltas.
1160 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001161 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1162 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1163 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001164 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001165 } else {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001166 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1167 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001168 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001169 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001170
Wu Fengguang143dfe82010-08-27 18:45:12 -06001171 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
Wu Fengguang77627412010-09-12 13:34:05 -06001172 (nr_dirty > dirty_thresh);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001173 if (dirty_exceeded && !bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001174 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001176 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1177 nr_dirty, bdi_thresh, bdi_dirty,
1178 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001179
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001180 max_pause = bdi_max_pause(bdi, bdi_dirty);
1181
Wu Fengguang143dfe82010-08-27 18:45:12 -06001182 dirty_ratelimit = bdi->dirty_ratelimit;
1183 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1184 background_thresh, nr_dirty,
1185 bdi_thresh, bdi_dirty);
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001186 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1187 RATELIMIT_CALC_SHIFT;
1188 if (unlikely(task_ratelimit == 0)) {
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001189 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001190 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 }
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001192 pause = HZ * pages_dirtied / task_ratelimit;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001193 if (unlikely(pause <= 0)) {
Wu Fengguangece13ac2010-08-29 23:33:20 -06001194 trace_balance_dirty_pages(bdi,
1195 dirty_thresh,
1196 background_thresh,
1197 nr_dirty,
1198 bdi_thresh,
1199 bdi_dirty,
1200 dirty_ratelimit,
1201 task_ratelimit,
1202 pages_dirtied,
1203 pause,
1204 start_time);
Wu Fengguang57fc9782011-06-11 19:32:32 -06001205 pause = 1; /* avoid resetting nr_dirtied_pause below */
1206 break;
1207 }
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001208 pause = min(pause, max_pause);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001209
1210pause:
Wu Fengguangece13ac2010-08-29 23:33:20 -06001211 trace_balance_dirty_pages(bdi,
1212 dirty_thresh,
1213 background_thresh,
1214 nr_dirty,
1215 bdi_thresh,
1216 bdi_dirty,
1217 dirty_ratelimit,
1218 task_ratelimit,
1219 pages_dirtied,
1220 pause,
1221 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001222 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001223 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001224
Wu Fengguangffd1f602011-06-19 22:18:42 -06001225 /*
Wu Fengguang1df64712011-11-13 19:47:32 -06001226 * This is typically equal to (nr_dirty < dirty_thresh) and can
1227 * also keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001228 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001229 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001230 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001231
Wu Fengguangc5c63432011-12-02 10:21:33 -06001232 /*
1233 * In the case of an unresponding NFS server and the NFS dirty
1234 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1235 * to go through, so that tasks on them still remain responsive.
1236 *
1237 * In theory 1 page is enough to keep the comsumer-producer
1238 * pipe going: the flusher cleans 1 page => the task dirties 1
1239 * more page. However bdi_dirty has accounting errors. So use
1240 * the larger and more IO friendly bdi_stat_error.
1241 */
1242 if (bdi_dirty <= bdi_stat_error(bdi))
1243 break;
1244
Jan Kara499d05e2011-11-16 19:34:48 +08001245 if (fatal_signal_pending(current))
1246 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 }
1248
Wu Fengguang143dfe82010-08-27 18:45:12 -06001249 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001250 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251
Wu Fengguang9d823e82011-06-11 18:10:12 -06001252 current->nr_dirtied = 0;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001253 if (pause == 0) { /* in freerun area */
1254 current->nr_dirtied_pause =
1255 dirty_poll_interval(nr_dirty, dirty_thresh);
1256 } else if (pause <= max_pause / 4 &&
1257 pages_dirtied >= current->nr_dirtied_pause) {
1258 current->nr_dirtied_pause = clamp_val(
1259 dirty_ratelimit * (max_pause / 2) / HZ,
1260 pages_dirtied + pages_dirtied / 8,
1261 pages_dirtied * 4);
1262 } else if (pause >= max_pause) {
1263 current->nr_dirtied_pause = 1 | clamp_val(
1264 dirty_ratelimit * (max_pause / 2) / HZ,
1265 pages_dirtied / 4,
1266 pages_dirtied - pages_dirtied / 8);
1267 }
Wu Fengguang9d823e82011-06-11 18:10:12 -06001268
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001270 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271
1272 /*
1273 * In laptop mode, we wait until hitting the higher threshold before
1274 * starting background writeout, and then write out all the way down
1275 * to the lower threshold. So slow writers cause minimal disk activity.
1276 *
1277 * In normal mode, we start background writeout at the lower
1278 * background_thresh, to keep the amount of dirty memory low.
1279 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001280 if (laptop_mode)
1281 return;
1282
1283 if (nr_reclaimable > background_thresh)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001284 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285}
1286
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001287void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001288{
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001289 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001290 struct address_space *mapping = page_mapping(page);
1291
1292 if (mapping)
1293 balance_dirty_pages_ratelimited(mapping);
1294 }
1295}
1296
Wu Fengguang9d823e82011-06-11 18:10:12 -06001297static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001298
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001300 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001301 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +02001302 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303 *
1304 * Processes which are dirtying memory should call in here once for each page
1305 * which was newly dirtied. The function will periodically check the system's
1306 * dirty state and will initiate writeback if needed.
1307 *
1308 * On really big machines, get_writeback_state is expensive, so try to avoid
1309 * calling it too often (ratelimiting). But once we're over the dirty memory
1310 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1311 * from overshooting the limit by (ratelimit_pages) each.
1312 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001313void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1314 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001316 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001317 int ratelimit;
1318 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319
Wu Fengguang36715ce2011-06-11 17:53:57 -06001320 if (!bdi_cap_account_dirty(bdi))
1321 return;
1322
Wu Fengguang9d823e82011-06-11 18:10:12 -06001323 ratelimit = current->nr_dirtied_pause;
1324 if (bdi->dirty_exceeded)
1325 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326
Wu Fengguang9d823e82011-06-11 18:10:12 -06001327 current->nr_dirtied += nr_pages_dirtied;
1328
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001329 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001330 /*
1331 * This prevents one CPU to accumulate too many dirtied pages without
1332 * calling into balance_dirty_pages(), which can happen when there are
1333 * 1000+ tasks, all of them start dirtying pages at exactly the same
1334 * time, hence all honoured too large initial task->nr_dirtied_pause.
1335 */
Tejun Heo245b2e72009-06-24 15:13:48 +09001336 p = &__get_cpu_var(bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001337 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001338 *p = 0;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001339 else {
1340 *p += nr_pages_dirtied;
1341 if (unlikely(*p >= ratelimit_pages)) {
1342 *p = 0;
1343 ratelimit = 0;
1344 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001346 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001347
1348 if (unlikely(current->nr_dirtied >= ratelimit))
1349 balance_dirty_pages(mapping, current->nr_dirtied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350}
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001351EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352
Andrew Morton232ea4d2007-02-28 20:13:21 -08001353void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354{
David Rientjes364aeb22009-01-06 14:39:29 -08001355 unsigned long background_thresh;
1356 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357
1358 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001359 global_dirty_limits(&background_thresh, &dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360
1361 /*
1362 * Boost the allowable dirty threshold a bit for page
1363 * allocators so they don't get DoS'ed by heavy writers
1364 */
1365 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1366
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001367 if (global_page_state(NR_UNSTABLE_NFS) +
1368 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1369 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001370 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001371
1372 /*
1373 * The caller might hold locks which can prevent IO completion
1374 * or progress in the filesystem. So we cannot just sit here
1375 * waiting for IO to complete.
1376 */
1377 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1378 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 }
1380}
1381
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1384 */
1385int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001386 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001388 proc_dointvec(table, write, buffer, length, ppos);
Jens Axboe64231042010-05-21 20:00:35 +02001389 bdi_arm_supers_timer();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 return 0;
1391}
1392
Jens Axboec2c49862010-05-20 09:18:47 +02001393#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001394void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395{
Matthew Garrett31373d02010-04-06 14:25:14 +02001396 struct request_queue *q = (struct request_queue *)data;
1397 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1398 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399
Matthew Garrett31373d02010-04-06 14:25:14 +02001400 /*
1401 * We want to write everything out, not just down to the dirty
1402 * threshold
1403 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001404 if (bdi_has_dirty_io(&q->backing_dev_info))
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001405 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1406 WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407}
1408
1409/*
1410 * We've spun up the disk and we're in laptop mode: schedule writeback
1411 * of all dirty data a few seconds from now. If the flush is already scheduled
1412 * then push it back - the user is still using the disk.
1413 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001414void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415{
Matthew Garrett31373d02010-04-06 14:25:14 +02001416 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417}
1418
1419/*
1420 * We're in laptop mode and we've just synced. The sync's writes will have
1421 * caused another writeback to be scheduled by laptop_io_completion.
1422 * Nothing needs to be written back anymore, so we unschedule the writeback.
1423 */
1424void laptop_sync_completion(void)
1425{
Matthew Garrett31373d02010-04-06 14:25:14 +02001426 struct backing_dev_info *bdi;
1427
1428 rcu_read_lock();
1429
1430 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1431 del_timer(&bdi->laptop_mode_wb_timer);
1432
1433 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434}
Jens Axboec2c49862010-05-20 09:18:47 +02001435#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436
1437/*
1438 * If ratelimit_pages is too high then we can get into dirty-data overload
1439 * if a large number of processes all perform writes at the same time.
1440 * If it is too low then SMP machines will call the (expensive)
1441 * get_writeback_state too often.
1442 *
1443 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1444 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06001445 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446 */
1447
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001448void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449{
Wu Fengguang9d823e82011-06-11 18:10:12 -06001450 unsigned long background_thresh;
1451 unsigned long dirty_thresh;
1452 global_dirty_limits(&background_thresh, &dirty_thresh);
1453 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 if (ratelimit_pages < 16)
1455 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456}
1457
Chandra Seetharaman26c21432006-06-27 02:54:10 -07001458static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1460{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001461 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -08001462 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463}
1464
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001465static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 .notifier_call = ratelimit_handler,
1467 .next = NULL,
1468};
1469
1470/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001471 * Called early on to tune the page writeback dirty limits.
1472 *
1473 * We used to scale dirty pages according to how total memory
1474 * related to pages that could be allocated for buffers (by
1475 * comparing nr_free_buffer_pages() to vm_total_pages.
1476 *
1477 * However, that was when we used "dirty_ratio" to scale with
1478 * all memory, and we don't do that any more. "dirty_ratio"
1479 * is now applied to total non-HIGHPAGE memory (by subtracting
1480 * totalhigh_pages from vm_total_pages), and as such we can't
1481 * get into the old insane situation any more where we had
1482 * large amounts of dirty pages compared to a small amount of
1483 * non-HIGHMEM memory.
1484 *
1485 * But we might still want to scale the dirty_ratio by how
1486 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 */
1488void __init page_writeback_init(void)
1489{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001490 int shift;
1491
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001492 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001494
1495 shift = calc_period_shift();
1496 prop_descriptor_init(&vm_completions, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497}
1498
David Howells811d7362006-08-29 19:06:09 +01001499/**
Jan Karaf446daae2010-08-09 17:19:12 -07001500 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1501 * @mapping: address space structure to write
1502 * @start: starting page index
1503 * @end: ending page index (inclusive)
1504 *
1505 * This function scans the page range from @start to @end (inclusive) and tags
1506 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1507 * that write_cache_pages (or whoever calls this function) will then use
1508 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1509 * used to avoid livelocking of writeback by a process steadily creating new
1510 * dirty pages in the file (thus it is important for this function to be quick
1511 * so that it can tag pages faster than a dirtying process can create them).
1512 */
1513/*
1514 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1515 */
Jan Karaf446daae2010-08-09 17:19:12 -07001516void tag_pages_for_writeback(struct address_space *mapping,
1517 pgoff_t start, pgoff_t end)
1518{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001519#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07001520 unsigned long tagged;
1521
1522 do {
1523 spin_lock_irq(&mapping->tree_lock);
1524 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1525 &start, end, WRITEBACK_TAG_BATCH,
1526 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1527 spin_unlock_irq(&mapping->tree_lock);
1528 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1529 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001530 /* We check 'start' to handle wrapping when end == ~0UL */
1531 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07001532}
1533EXPORT_SYMBOL(tag_pages_for_writeback);
1534
1535/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001536 * 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 +01001537 * @mapping: address space structure to write
1538 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001539 * @writepage: function called for each page
1540 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001541 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001542 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001543 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1544 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1545 * and msync() need to guarantee that all the data which was dirty at the time
1546 * the call was made get new I/O started against them. If wbc->sync_mode is
1547 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1548 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07001549 *
1550 * To avoid livelocks (when other process dirties new pages), we first tag
1551 * pages which should be written back with TOWRITE tag and only then start
1552 * writing them. For data-integrity sync we have to be careful so that we do
1553 * not miss some pages (e.g., because some other process has cleared TOWRITE
1554 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1555 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001556 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001557int write_cache_pages(struct address_space *mapping,
1558 struct writeback_control *wbc, writepage_t writepage,
1559 void *data)
David Howells811d7362006-08-29 19:06:09 +01001560{
David Howells811d7362006-08-29 19:06:09 +01001561 int ret = 0;
1562 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001563 struct pagevec pvec;
1564 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001565 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001566 pgoff_t index;
1567 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001568 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001569 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001570 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07001571 int tag;
David Howells811d7362006-08-29 19:06:09 +01001572
David Howells811d7362006-08-29 19:06:09 +01001573 pagevec_init(&pvec, 0);
1574 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001575 writeback_index = mapping->writeback_index; /* prev offset */
1576 index = writeback_index;
1577 if (index == 0)
1578 cycled = 1;
1579 else
1580 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001581 end = -1;
1582 } else {
1583 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1584 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1585 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1586 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001587 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001588 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001589 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001590 tag = PAGECACHE_TAG_TOWRITE;
1591 else
1592 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001593retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001594 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001595 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001596 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001597 while (!done && (index <= end)) {
1598 int i;
1599
Jan Karaf446daae2010-08-09 17:19:12 -07001600 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001601 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1602 if (nr_pages == 0)
1603 break;
David Howells811d7362006-08-29 19:06:09 +01001604
David Howells811d7362006-08-29 19:06:09 +01001605 for (i = 0; i < nr_pages; i++) {
1606 struct page *page = pvec.pages[i];
1607
Nick Piggind5482cd2009-01-06 14:39:11 -08001608 /*
1609 * At this point, the page may be truncated or
1610 * invalidated (changing page->mapping to NULL), or
1611 * even swizzled back from swapper_space to tmpfs file
1612 * mapping. However, page->index will not change
1613 * because we have a reference on the page.
1614 */
1615 if (page->index > end) {
1616 /*
1617 * can't be range_cyclic (1st pass) because
1618 * end == -1 in that case.
1619 */
1620 done = 1;
1621 break;
1622 }
1623
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001624 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001625
David Howells811d7362006-08-29 19:06:09 +01001626 lock_page(page);
1627
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001628 /*
1629 * Page truncated or invalidated. We can freely skip it
1630 * then, even for data integrity operations: the page
1631 * has disappeared concurrently, so there could be no
1632 * real expectation of this data interity operation
1633 * even if there is now a new, dirty page at the same
1634 * pagecache address.
1635 */
David Howells811d7362006-08-29 19:06:09 +01001636 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001637continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001638 unlock_page(page);
1639 continue;
1640 }
1641
Nick Piggin515f4a02009-01-06 14:39:10 -08001642 if (!PageDirty(page)) {
1643 /* someone wrote it for us */
1644 goto continue_unlock;
1645 }
David Howells811d7362006-08-29 19:06:09 +01001646
Nick Piggin515f4a02009-01-06 14:39:10 -08001647 if (PageWriteback(page)) {
1648 if (wbc->sync_mode != WB_SYNC_NONE)
1649 wait_on_page_writeback(page);
1650 else
1651 goto continue_unlock;
1652 }
1653
1654 BUG_ON(PageWriteback(page));
1655 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001656 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001657
Dave Chinner9e094382010-07-07 13:24:08 +10001658 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001659 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001660 if (unlikely(ret)) {
1661 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1662 unlock_page(page);
1663 ret = 0;
1664 } else {
1665 /*
1666 * done_index is set past this page,
1667 * so media errors will not choke
1668 * background writeout for the entire
1669 * file. This has consequences for
1670 * range_cyclic semantics (ie. it may
1671 * not be suitable for data integrity
1672 * writeout).
1673 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001674 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001675 done = 1;
1676 break;
1677 }
Dave Chinner0b564922010-06-09 10:37:18 +10001678 }
David Howells811d7362006-08-29 19:06:09 +01001679
Dave Chinner546a1922010-08-24 11:44:34 +10001680 /*
1681 * We stop writing back only if we are not doing
1682 * integrity sync. In case of integrity sync we have to
1683 * keep going until we have written all the pages
1684 * we tagged for writeback prior to entering this loop.
1685 */
1686 if (--wbc->nr_to_write <= 0 &&
1687 wbc->sync_mode == WB_SYNC_NONE) {
1688 done = 1;
1689 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001690 }
David Howells811d7362006-08-29 19:06:09 +01001691 }
1692 pagevec_release(&pvec);
1693 cond_resched();
1694 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001695 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001696 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001697 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001698 * We hit the last page and there is more work to be done: wrap
1699 * back to the start of the file
1700 */
Nick Piggin31a12662009-01-06 14:39:04 -08001701 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001702 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001703 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001704 goto retry;
1705 }
Dave Chinner0b564922010-06-09 10:37:18 +10001706 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1707 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001708
David Howells811d7362006-08-29 19:06:09 +01001709 return ret;
1710}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001711EXPORT_SYMBOL(write_cache_pages);
1712
1713/*
1714 * Function used by generic_writepages to call the real writepage
1715 * function and set the mapping flags on error
1716 */
1717static int __writepage(struct page *page, struct writeback_control *wbc,
1718 void *data)
1719{
1720 struct address_space *mapping = data;
1721 int ret = mapping->a_ops->writepage(page, wbc);
1722 mapping_set_error(mapping, ret);
1723 return ret;
1724}
1725
1726/**
1727 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1728 * @mapping: address space structure to write
1729 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1730 *
1731 * This is a library function, which implements the writepages()
1732 * address_space_operation.
1733 */
1734int generic_writepages(struct address_space *mapping,
1735 struct writeback_control *wbc)
1736{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001737 struct blk_plug plug;
1738 int ret;
1739
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001740 /* deal with chardevs and other special file */
1741 if (!mapping->a_ops->writepage)
1742 return 0;
1743
Shaohua Li9b6096a2011-03-17 10:47:06 +01001744 blk_start_plug(&plug);
1745 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1746 blk_finish_plug(&plug);
1747 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001748}
David Howells811d7362006-08-29 19:06:09 +01001749
1750EXPORT_SYMBOL(generic_writepages);
1751
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1753{
Andrew Morton22905f72005-11-16 15:07:01 -08001754 int ret;
1755
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 if (wbc->nr_to_write <= 0)
1757 return 0;
1758 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001759 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001760 else
1761 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001762 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763}
1764
1765/**
1766 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001767 * @page: the page to write
1768 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 *
1770 * The page must be locked by the caller and will be unlocked upon return.
1771 *
1772 * write_one_page() returns a negative error code if I/O failed.
1773 */
1774int write_one_page(struct page *page, int wait)
1775{
1776 struct address_space *mapping = page->mapping;
1777 int ret = 0;
1778 struct writeback_control wbc = {
1779 .sync_mode = WB_SYNC_ALL,
1780 .nr_to_write = 1,
1781 };
1782
1783 BUG_ON(!PageLocked(page));
1784
1785 if (wait)
1786 wait_on_page_writeback(page);
1787
1788 if (clear_page_dirty_for_io(page)) {
1789 page_cache_get(page);
1790 ret = mapping->a_ops->writepage(page, &wbc);
1791 if (ret == 0 && wait) {
1792 wait_on_page_writeback(page);
1793 if (PageError(page))
1794 ret = -EIO;
1795 }
1796 page_cache_release(page);
1797 } else {
1798 unlock_page(page);
1799 }
1800 return ret;
1801}
1802EXPORT_SYMBOL(write_one_page);
1803
1804/*
Ken Chen76719322007-02-10 01:43:15 -08001805 * For address_spaces which do not use buffers nor write back.
1806 */
1807int __set_page_dirty_no_writeback(struct page *page)
1808{
1809 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001810 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001811 return 0;
1812}
1813
1814/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001815 * Helper function for set_page_dirty family.
1816 * NOTE: This relies on being atomic wrt interrupts.
1817 */
1818void account_page_dirtied(struct page *page, struct address_space *mapping)
1819{
1820 if (mapping_cap_account_dirty(mapping)) {
1821 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001822 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001823 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06001824 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001825 task_io_account_write(PAGE_CACHE_SIZE);
1826 }
1827}
Michael Rubin679ceac2010-08-20 02:31:26 -07001828EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001829
1830/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001831 * Helper function for set_page_writeback family.
1832 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1833 * wrt interrupts.
1834 */
1835void account_page_writeback(struct page *page)
1836{
1837 inc_zone_page_state(page, NR_WRITEBACK);
1838}
1839EXPORT_SYMBOL(account_page_writeback);
1840
1841/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 * For address_spaces which do not use buffers. Just tag the page as dirty in
1843 * its radix tree.
1844 *
1845 * This is also used when a single buffer is being dirtied: we want to set the
1846 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1847 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1848 *
1849 * Most callers have locked the page, which pins the address_space in memory.
1850 * But zap_pte_range() does not lock the page, however in that case the
1851 * mapping is pinned by the vma's ->vm_file reference.
1852 *
1853 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001854 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 */
1856int __set_page_dirty_nobuffers(struct page *page)
1857{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 if (!TestSetPageDirty(page)) {
1859 struct address_space *mapping = page_mapping(page);
1860 struct address_space *mapping2;
1861
Andrew Morton8c085402006-12-10 02:19:24 -08001862 if (!mapping)
1863 return 1;
1864
Nick Piggin19fd6232008-07-25 19:45:32 -07001865 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001866 mapping2 = page_mapping(page);
1867 if (mapping2) { /* Race with truncate? */
1868 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001869 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001870 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001871 radix_tree_tag_set(&mapping->page_tree,
1872 page_index(page), PAGECACHE_TAG_DIRTY);
1873 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001874 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001875 if (mapping->host) {
1876 /* !PageAnon && !swapper_space */
1877 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001879 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001881 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882}
1883EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1884
1885/*
1886 * When a writepage implementation decides that it doesn't want to write this
1887 * page for some reason, it should redirty the locked page via
1888 * redirty_page_for_writepage() and it should then unlock the page and return 0
1889 */
1890int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1891{
1892 wbc->pages_skipped++;
1893 return __set_page_dirty_nobuffers(page);
1894}
1895EXPORT_SYMBOL(redirty_page_for_writepage);
1896
1897/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02001898 * Dirty a page.
1899 *
1900 * For pages with a mapping this should be done under the page lock
1901 * for the benefit of asynchronous memory errors who prefer a consistent
1902 * dirty state. This rule can be broken in some special cases,
1903 * but should be better not to.
1904 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 * If the mapping doesn't provide a set_page_dirty a_op, then
1906 * just fall through and assume that it wants buffer_heads.
1907 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001908int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909{
1910 struct address_space *mapping = page_mapping(page);
1911
1912 if (likely(mapping)) {
1913 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07001914 /*
1915 * readahead/lru_deactivate_page could remain
1916 * PG_readahead/PG_reclaim due to race with end_page_writeback
1917 * About readahead, if the page is written, the flags would be
1918 * reset. So no problem.
1919 * About lru_deactivate_page, if the page is redirty, the flag
1920 * will be reset. So no problem. but if the page is used by readahead
1921 * it will confuse readahead and make it restart the size rampup
1922 * process. But it's a trivial problem.
1923 */
1924 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02001925#ifdef CONFIG_BLOCK
1926 if (!spd)
1927 spd = __set_page_dirty_buffers;
1928#endif
1929 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001931 if (!PageDirty(page)) {
1932 if (!TestSetPageDirty(page))
1933 return 1;
1934 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 return 0;
1936}
1937EXPORT_SYMBOL(set_page_dirty);
1938
1939/*
1940 * set_page_dirty() is racy if the caller has no reference against
1941 * page->mapping->host, and if the page is unlocked. This is because another
1942 * CPU could truncate the page off the mapping and then free the mapping.
1943 *
1944 * Usually, the page _is_ locked, or the caller is a user-space process which
1945 * holds a reference on the inode by having an open file.
1946 *
1947 * In other cases, the page should be locked before running set_page_dirty().
1948 */
1949int set_page_dirty_lock(struct page *page)
1950{
1951 int ret;
1952
Jens Axboe7eaceac2011-03-10 08:52:07 +01001953 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 ret = set_page_dirty(page);
1955 unlock_page(page);
1956 return ret;
1957}
1958EXPORT_SYMBOL(set_page_dirty_lock);
1959
1960/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 * Clear a page's dirty flag, while caring for dirty memory accounting.
1962 * Returns true if the page was previously dirty.
1963 *
1964 * This is for preparing to put the page under writeout. We leave the page
1965 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1966 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1967 * implementation will run either set_page_writeback() or set_page_dirty(),
1968 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1969 * back into sync.
1970 *
1971 * This incoherency between the page's dirty flag and radix-tree tag is
1972 * unfortunate, but it only exists while the page is locked.
1973 */
1974int clear_page_dirty_for_io(struct page *page)
1975{
1976 struct address_space *mapping = page_mapping(page);
1977
Nick Piggin79352892007-07-19 01:47:22 -07001978 BUG_ON(!PageLocked(page));
1979
Linus Torvalds7658cc22006-12-29 10:00:58 -08001980 if (mapping && mapping_cap_account_dirty(mapping)) {
1981 /*
1982 * Yes, Virginia, this is indeed insane.
1983 *
1984 * We use this sequence to make sure that
1985 * (a) we account for dirty stats properly
1986 * (b) we tell the low-level filesystem to
1987 * mark the whole page dirty if it was
1988 * dirty in a pagetable. Only to then
1989 * (c) clean the page again and return 1 to
1990 * cause the writeback.
1991 *
1992 * This way we avoid all nasty races with the
1993 * dirty bit in multiple places and clearing
1994 * them concurrently from different threads.
1995 *
1996 * Note! Normally the "set_page_dirty(page)"
1997 * has no effect on the actual dirty bit - since
1998 * that will already usually be set. But we
1999 * need the side effects, and it can help us
2000 * avoid races.
2001 *
2002 * We basically use the page "master dirty bit"
2003 * as a serialization point for all the different
2004 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002005 */
2006 if (page_mkclean(page))
2007 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002008 /*
2009 * We carefully synchronise fault handlers against
2010 * installing a dirty pte and marking the page dirty
2011 * at this point. We do this by having them hold the
2012 * page lock at some point after installing their
2013 * pte, but before marking the page dirty.
2014 * Pages are always locked coming in here, so we get
2015 * the desired exclusion. See mm/memory.c:do_wp_page()
2016 * for more comments.
2017 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08002018 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08002019 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07002020 dec_bdi_stat(mapping->backing_dev_info,
2021 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002022 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002024 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002026 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002028EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029
2030int test_clear_page_writeback(struct page *page)
2031{
2032 struct address_space *mapping = page_mapping(page);
2033 int ret;
2034
2035 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002036 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 unsigned long flags;
2038
Nick Piggin19fd6232008-07-25 19:45:32 -07002039 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002041 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 radix_tree_tag_clear(&mapping->page_tree,
2043 page_index(page),
2044 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002045 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002046 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002047 __bdi_writeout_inc(bdi);
2048 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002049 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002050 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 } else {
2052 ret = TestClearPageWriteback(page);
2053 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002054 if (ret) {
Andrew Mortond688abf2007-07-19 01:49:17 -07002055 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002056 inc_zone_page_state(page, NR_WRITTEN);
2057 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 return ret;
2059}
2060
2061int test_set_page_writeback(struct page *page)
2062{
2063 struct address_space *mapping = page_mapping(page);
2064 int ret;
2065
2066 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002067 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 unsigned long flags;
2069
Nick Piggin19fd6232008-07-25 19:45:32 -07002070 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002072 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 radix_tree_tag_set(&mapping->page_tree,
2074 page_index(page),
2075 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002076 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002077 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2078 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 if (!PageDirty(page))
2080 radix_tree_tag_clear(&mapping->page_tree,
2081 page_index(page),
2082 PAGECACHE_TAG_DIRTY);
Jan Karaf446daae2010-08-09 17:19:12 -07002083 radix_tree_tag_clear(&mapping->page_tree,
2084 page_index(page),
2085 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002086 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 } else {
2088 ret = TestSetPageWriteback(page);
2089 }
Andrew Mortond688abf2007-07-19 01:49:17 -07002090 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07002091 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 return ret;
2093
2094}
2095EXPORT_SYMBOL(test_set_page_writeback);
2096
2097/*
Nick Piggin00128182007-10-16 01:24:40 -07002098 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 * passed tag.
2100 */
2101int mapping_tagged(struct address_space *mapping, int tag)
2102{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002103 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104}
2105EXPORT_SYMBOL(mapping_tagged);