blob: 96b3e7aa705c3ebe12f02b52083e80683973a966 [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>
David Howellscf9a2ae2006-08-29 19:05:54 +010035#include <linux/buffer_head.h>
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100037#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038
39/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060040 * Sleep at most 200ms at a time in balance_dirty_pages().
41 */
42#define MAX_PAUSE max(HZ/5, 1)
43
44/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060045 * Estimate write bandwidth at 200ms intervals.
46 */
47#define BANDWIDTH_INTERVAL max(HZ/5, 1)
48
Wu Fengguang6c14ae12011-03-02 16:04:18 -060049#define RATELIMIT_CALC_SHIFT 10
50
Wu Fengguange98be2d2010-08-29 11:22:30 -060051/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
53 * will look to see if it needs to force writeback or throttling.
54 */
55static long ratelimit_pages = 32;
56
Linus Torvalds1da177e2005-04-16 15:20:36 -070057/* 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 Gondwana195cf452008-02-04 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/*
133 * couple the period to the dirty_ratio:
134 *
135 * period/2 ~ roundup_pow_of_two(dirty limit)
136 */
137static int calc_period_shift(void)
138{
139 unsigned long dirty_total;
140
David Rientjes2da02992009-01-06 14:39:31 -0800141 if (vm_dirty_bytes)
142 dirty_total = vm_dirty_bytes / PAGE_SIZE;
143 else
144 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
145 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700146 return 2 + ilog2(dirty_total - 1);
147}
148
149/*
David Rientjes2da02992009-01-06 14:39:31 -0800150 * update the period when the dirty threshold changes.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700151 */
David Rientjes2da02992009-01-06 14:39:31 -0800152static void update_completion_period(void)
153{
154 int shift = calc_period_shift();
155 prop_change_shift(&vm_completions, shift);
Wu Fengguang9d823e82011-06-11 18:10:12 -0600156
157 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800158}
159
160int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700161 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800162 loff_t *ppos)
163{
164 int ret;
165
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700166 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800167 if (ret == 0 && write)
168 dirty_background_bytes = 0;
169 return ret;
170}
171
172int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700173 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800174 loff_t *ppos)
175{
176 int ret;
177
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700178 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800179 if (ret == 0 && write)
180 dirty_background_ratio = 0;
181 return ret;
182}
183
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700184int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700185 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700186 loff_t *ppos)
187{
188 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800189 int ret;
190
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700191 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700192 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
David Rientjes2da02992009-01-06 14:39:31 -0800193 update_completion_period();
194 vm_dirty_bytes = 0;
195 }
196 return ret;
197}
198
199
200int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700201 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800202 loff_t *ppos)
203{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800204 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800205 int ret;
206
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700207 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800208 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
209 update_completion_period();
210 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700211 }
212 return ret;
213}
214
215/*
216 * Increment the BDI's writeout completion count and the global writeout
217 * completion count. Called from test_clear_page_writeback().
218 */
219static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
220{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600221 __inc_bdi_stat(bdi, BDI_WRITTEN);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700222 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
223 bdi->max_prop_frac);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700224}
225
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700226void bdi_writeout_inc(struct backing_dev_info *bdi)
227{
228 unsigned long flags;
229
230 local_irq_save(flags);
231 __bdi_writeout_inc(bdi);
232 local_irq_restore(flags);
233}
234EXPORT_SYMBOL_GPL(bdi_writeout_inc);
235
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700236/*
237 * Obtain an accurate fraction of the BDI's portion.
238 */
239static void bdi_writeout_fraction(struct backing_dev_info *bdi,
240 long *numerator, long *denominator)
241{
Wu Fengguang3efaf0f2010-12-16 22:22:00 -0600242 prop_fraction_percpu(&vm_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700243 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700244}
245
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700246/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700247 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
248 * registered backing devices, which, for obvious reasons, can not
249 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700250 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700251static unsigned int bdi_min_ratio;
252
253int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
254{
255 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700256
Jens Axboecfc4ba52009-09-14 13:12:40 +0200257 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700258 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700259 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700260 } else {
261 min_ratio -= bdi->min_ratio;
262 if (bdi_min_ratio + min_ratio < 100) {
263 bdi_min_ratio += min_ratio;
264 bdi->min_ratio += min_ratio;
265 } else {
266 ret = -EINVAL;
267 }
268 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200269 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700270
271 return ret;
272}
273
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700274int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
275{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700276 int ret = 0;
277
278 if (max_ratio > 100)
279 return -EINVAL;
280
Jens Axboecfc4ba52009-09-14 13:12:40 +0200281 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700282 if (bdi->min_ratio > max_ratio) {
283 ret = -EINVAL;
284 } else {
285 bdi->max_ratio = max_ratio;
286 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
287 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200288 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700289
290 return ret;
291}
292EXPORT_SYMBOL(bdi_set_max_ratio);
293
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700294/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 * Work out the current dirty-memory clamping and background writeout
296 * thresholds.
297 *
298 * The main aim here is to lower them aggressively if there is a lot of mapped
299 * memory around. To avoid stressing page reclaim with lots of unreclaimable
300 * pages. It is better to clamp down on writers than to start swapping, and
301 * performing lots of scanning.
302 *
303 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
304 *
305 * We don't permit the clamping level to fall below 5% - that is getting rather
306 * excessive.
307 *
308 * We make sure that the background writeout level is below the adjusted
309 * clamping level.
310 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700311
312static unsigned long highmem_dirtyable_memory(unsigned long total)
313{
314#ifdef CONFIG_HIGHMEM
315 int node;
316 unsigned long x = 0;
317
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700318 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700319 struct zone *z =
320 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
321
Wu Fengguangadea02a2009-09-21 17:01:42 -0700322 x += zone_page_state(z, NR_FREE_PAGES) +
323 zone_reclaimable_pages(z);
Christoph Lameter1b424462007-05-06 14:48:59 -0700324 }
325 /*
326 * Make sure that the number of highmem pages is never larger
327 * than the number of the total dirtyable memory. This can only
328 * occur in very strange VM situations but we want to make sure
329 * that this does not occur.
330 */
331 return min(x, total);
332#else
333 return 0;
334#endif
335}
336
Steven Rostedt3eefae92008-05-12 21:21:04 +0200337/**
338 * determine_dirtyable_memory - amount of memory that may be used
339 *
340 * Returns the numebr of pages that can currently be freed and used
341 * by the kernel for direct mappings.
342 */
343unsigned long determine_dirtyable_memory(void)
Christoph Lameter1b424462007-05-06 14:48:59 -0700344{
345 unsigned long x;
346
Wu Fengguangadea02a2009-09-21 17:01:42 -0700347 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
Bron Gondwana195cf452008-02-04 22:29:20 -0800348
349 if (!vm_highmem_is_dirtyable)
350 x -= highmem_dirtyable_memory(x);
351
Christoph Lameter1b424462007-05-06 14:48:59 -0700352 return x + 1; /* Ensure that we never return 0 */
353}
354
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600355static unsigned long dirty_freerun_ceiling(unsigned long thresh,
356 unsigned long bg_thresh)
357{
358 return (thresh + bg_thresh) / 2;
359}
360
Wu Fengguangffd1f602011-06-19 22:18:42 -0600361static unsigned long hard_dirty_limit(unsigned long thresh)
362{
363 return max(thresh, global_dirty_limit);
364}
365
Randy Dunlap03ab4502010-08-14 13:05:17 -0700366/*
Wu Fengguang1babe182010-08-11 14:17:40 -0700367 * global_dirty_limits - background-writeback and dirty-throttling thresholds
368 *
369 * Calculate the dirty thresholds based on sysctl parameters
370 * - vm.dirty_background_ratio or vm.dirty_background_bytes
371 * - vm.dirty_ratio or vm.dirty_bytes
372 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Minchan Kimebd13732011-01-04 01:36:48 +0900373 * real-time tasks.
Wu Fengguang1babe182010-08-11 14:17:40 -0700374 */
Wu Fengguang16c40422010-08-11 14:17:39 -0700375void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376{
David Rientjes364aeb22009-01-06 14:39:29 -0800377 unsigned long background;
378 unsigned long dirty;
Minchan Kim240c8792011-01-13 15:46:27 -0800379 unsigned long uninitialized_var(available_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 struct task_struct *tsk;
381
Minchan Kim240c8792011-01-13 15:46:27 -0800382 if (!vm_dirty_bytes || !dirty_background_bytes)
383 available_memory = determine_dirtyable_memory();
384
David Rientjes2da02992009-01-06 14:39:31 -0800385 if (vm_dirty_bytes)
386 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
Wu Fengguang4cbec4c2010-10-26 14:21:45 -0700387 else
388 dirty = (vm_dirty_ratio * available_memory) / 100;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
David Rientjes2da02992009-01-06 14:39:31 -0800390 if (dirty_background_bytes)
391 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
392 else
393 background = (dirty_background_ratio * available_memory) / 100;
394
395 if (background >= dirty)
396 background = dirty / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 tsk = current;
398 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
399 background += background / 4;
400 dirty += dirty / 4;
401 }
402 *pbackground = background;
403 *pdirty = dirty;
Wu Fengguange1cbe232010-12-06 22:34:29 -0600404 trace_global_dirty_state(background, dirty);
Wu Fengguang16c40422010-08-11 14:17:39 -0700405}
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700406
Wu Fengguang6f718652011-03-02 17:14:34 -0600407/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700408 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600409 * @bdi: the backing_dev_info to query
410 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700411 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600412 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
413 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600414 *
415 * Note that balance_dirty_pages() will only seriously take it as a hard limit
416 * when sleeping max_pause per page is not enough to keep the dirty pages under
417 * control. For example, when the device is completely stalled due to some error
418 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
419 * In the other normal situations, it acts more gently by throttling the tasks
420 * more (rather than completely block them) when the bdi dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600421 *
422 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700423 * - starving fast devices
424 * - piling up dirty pages (that will take long time to sync) on slow devices
425 *
426 * The bdi's share of dirty limit will be adapting to its throughput and
427 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
428 */
429unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700430{
431 u64 bdi_dirty;
432 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700433
Wu Fengguang16c40422010-08-11 14:17:39 -0700434 /*
435 * Calculate this BDI's share of the dirty ratio.
436 */
437 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700438
Wu Fengguang16c40422010-08-11 14:17:39 -0700439 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
440 bdi_dirty *= numerator;
441 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700442
Wu Fengguang16c40422010-08-11 14:17:39 -0700443 bdi_dirty += (dirty * bdi->min_ratio) / 100;
444 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
445 bdi_dirty = dirty * bdi->max_ratio / 100;
446
447 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448}
449
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600450/*
451 * Dirty position control.
452 *
453 * (o) global/bdi setpoints
454 *
455 * We want the dirty pages be balanced around the global/bdi setpoints.
456 * When the number of dirty pages is higher/lower than the setpoint, the
457 * dirty position control ratio (and hence task dirty ratelimit) will be
458 * decreased/increased to bring the dirty pages back to the setpoint.
459 *
460 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
461 *
462 * if (dirty < setpoint) scale up pos_ratio
463 * if (dirty > setpoint) scale down pos_ratio
464 *
465 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
466 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
467 *
468 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
469 *
470 * (o) global control line
471 *
472 * ^ pos_ratio
473 * |
474 * | |<===== global dirty control scope ======>|
475 * 2.0 .............*
476 * | .*
477 * | . *
478 * | . *
479 * | . *
480 * | . *
481 * | . *
482 * 1.0 ................................*
483 * | . . *
484 * | . . *
485 * | . . *
486 * | . . *
487 * | . . *
488 * 0 +------------.------------------.----------------------*------------->
489 * freerun^ setpoint^ limit^ dirty pages
490 *
491 * (o) bdi control line
492 *
493 * ^ pos_ratio
494 * |
495 * | *
496 * | *
497 * | *
498 * | *
499 * | * |<=========== span ============>|
500 * 1.0 .......................*
501 * | . *
502 * | . *
503 * | . *
504 * | . *
505 * | . *
506 * | . *
507 * | . *
508 * | . *
509 * | . *
510 * | . *
511 * | . *
512 * 1/4 ...............................................* * * * * * * * * * * *
513 * | . .
514 * | . .
515 * | . .
516 * 0 +----------------------.-------------------------------.------------->
517 * bdi_setpoint^ x_intercept^
518 *
519 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
520 * be smoothly throttled down to normal if it starts high in situations like
521 * - start writing to a slow SD card and a fast disk at the same time. The SD
522 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
523 * - the bdi dirty thresh drops quickly due to change of JBOD workload
524 */
525static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
526 unsigned long thresh,
527 unsigned long bg_thresh,
528 unsigned long dirty,
529 unsigned long bdi_thresh,
530 unsigned long bdi_dirty)
531{
532 unsigned long write_bw = bdi->avg_write_bandwidth;
533 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
534 unsigned long limit = hard_dirty_limit(thresh);
535 unsigned long x_intercept;
536 unsigned long setpoint; /* dirty pages' target balance point */
537 unsigned long bdi_setpoint;
538 unsigned long span;
539 long long pos_ratio; /* for scaling up/down the rate limit */
540 long x;
541
542 if (unlikely(dirty >= limit))
543 return 0;
544
545 /*
546 * global setpoint
547 *
548 * setpoint - dirty 3
549 * f(dirty) := 1.0 + (----------------)
550 * limit - setpoint
551 *
552 * it's a 3rd order polynomial that subjects to
553 *
554 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
555 * (2) f(setpoint) = 1.0 => the balance point
556 * (3) f(limit) = 0 => the hard limit
557 * (4) df/dx <= 0 => negative feedback control
558 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
559 * => fast response on large errors; small oscillation near setpoint
560 */
561 setpoint = (freerun + limit) / 2;
562 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
563 limit - setpoint + 1);
564 pos_ratio = x;
565 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
566 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
567 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
568
569 /*
570 * We have computed basic pos_ratio above based on global situation. If
571 * the bdi is over/under its share of dirty pages, we want to scale
572 * pos_ratio further down/up. That is done by the following mechanism.
573 */
574
575 /*
576 * bdi setpoint
577 *
578 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
579 *
580 * x_intercept - bdi_dirty
581 * := --------------------------
582 * x_intercept - bdi_setpoint
583 *
584 * The main bdi control line is a linear function that subjects to
585 *
586 * (1) f(bdi_setpoint) = 1.0
587 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
588 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
589 *
590 * For single bdi case, the dirty pages are observed to fluctuate
591 * regularly within range
592 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
593 * for various filesystems, where (2) can yield in a reasonable 12.5%
594 * fluctuation range for pos_ratio.
595 *
596 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
597 * own size, so move the slope over accordingly and choose a slope that
598 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
599 */
600 if (unlikely(bdi_thresh > thresh))
601 bdi_thresh = thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600602 /*
603 * It's very possible that bdi_thresh is close to 0 not because the
604 * device is slow, but that it has remained inactive for long time.
605 * Honour such devices a reasonable good (hopefully IO efficient)
606 * threshold, so that the occasional writes won't be blocked and active
607 * writes can rampup the threshold quickly.
608 */
Wu Fengguang8927f662011-08-04 22:16:46 -0600609 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600610 /*
611 * scale global setpoint to bdi's:
612 * bdi_setpoint = setpoint * bdi_thresh / thresh
613 */
614 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
615 bdi_setpoint = setpoint * (u64)x >> 16;
616 /*
617 * Use span=(8*write_bw) in single bdi case as indicated by
618 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
619 *
620 * bdi_thresh thresh - bdi_thresh
621 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
622 * thresh thresh
623 */
624 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
625 x_intercept = bdi_setpoint + span;
626
627 if (bdi_dirty < x_intercept - span / 4) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600628 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
629 x_intercept - bdi_setpoint + 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600630 } else
631 pos_ratio /= 4;
632
Wu Fengguang8927f662011-08-04 22:16:46 -0600633 /*
634 * bdi reserve area, safeguard against dirty pool underrun and disk idle
635 * It may push the desired control point of global dirty pages higher
636 * than setpoint.
637 */
638 x_intercept = bdi_thresh / 2;
639 if (bdi_dirty < x_intercept) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600640 if (bdi_dirty > x_intercept / 8)
641 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
642 else
Wu Fengguang8927f662011-08-04 22:16:46 -0600643 pos_ratio *= 8;
644 }
645
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600646 return pos_ratio;
647}
648
Wu Fengguange98be2d2010-08-29 11:22:30 -0600649static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
650 unsigned long elapsed,
651 unsigned long written)
652{
653 const unsigned long period = roundup_pow_of_two(3 * HZ);
654 unsigned long avg = bdi->avg_write_bandwidth;
655 unsigned long old = bdi->write_bandwidth;
656 u64 bw;
657
658 /*
659 * bw = written * HZ / elapsed
660 *
661 * bw * elapsed + write_bandwidth * (period - elapsed)
662 * write_bandwidth = ---------------------------------------------------
663 * period
664 */
665 bw = written - bdi->written_stamp;
666 bw *= HZ;
667 if (unlikely(elapsed > period)) {
668 do_div(bw, elapsed);
669 avg = bw;
670 goto out;
671 }
672 bw += (u64)bdi->write_bandwidth * (period - elapsed);
673 bw >>= ilog2(period);
674
675 /*
676 * one more level of smoothing, for filtering out sudden spikes
677 */
678 if (avg > old && old >= (unsigned long)bw)
679 avg -= (avg - old) >> 3;
680
681 if (avg < old && old <= (unsigned long)bw)
682 avg += (old - avg) >> 3;
683
684out:
685 bdi->write_bandwidth = bw;
686 bdi->avg_write_bandwidth = avg;
687}
688
Wu Fengguangc42843f2011-03-02 15:54:09 -0600689/*
690 * The global dirtyable memory and dirty threshold could be suddenly knocked
691 * down by a large amount (eg. on the startup of KVM in a swapless system).
692 * This may throw the system into deep dirty exceeded state and throttle
693 * heavy/light dirtiers alike. To retain good responsiveness, maintain
694 * global_dirty_limit for tracking slowly down to the knocked down dirty
695 * threshold.
696 */
697static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
698{
699 unsigned long limit = global_dirty_limit;
700
701 /*
702 * Follow up in one step.
703 */
704 if (limit < thresh) {
705 limit = thresh;
706 goto update;
707 }
708
709 /*
710 * Follow down slowly. Use the higher one as the target, because thresh
711 * may drop below dirty. This is exactly the reason to introduce
712 * global_dirty_limit which is guaranteed to lie above the dirty pages.
713 */
714 thresh = max(thresh, dirty);
715 if (limit > thresh) {
716 limit -= (limit - thresh) >> 5;
717 goto update;
718 }
719 return;
720update:
721 global_dirty_limit = limit;
722}
723
724static void global_update_bandwidth(unsigned long thresh,
725 unsigned long dirty,
726 unsigned long now)
727{
728 static DEFINE_SPINLOCK(dirty_lock);
729 static unsigned long update_time;
730
731 /*
732 * check locklessly first to optimize away locking for the most time
733 */
734 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
735 return;
736
737 spin_lock(&dirty_lock);
738 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
739 update_dirty_limit(thresh, dirty);
740 update_time = now;
741 }
742 spin_unlock(&dirty_lock);
743}
744
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600745/*
746 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
747 *
748 * Normal bdi tasks will be curbed at or below it in long term.
749 * Obviously it should be around (write_bw / N) when there are N dd tasks.
750 */
751static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
752 unsigned long thresh,
753 unsigned long bg_thresh,
754 unsigned long dirty,
755 unsigned long bdi_thresh,
756 unsigned long bdi_dirty,
757 unsigned long dirtied,
758 unsigned long elapsed)
759{
Wu Fengguang73811312011-08-26 15:53:24 -0600760 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
761 unsigned long limit = hard_dirty_limit(thresh);
762 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600763 unsigned long write_bw = bdi->avg_write_bandwidth;
764 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
765 unsigned long dirty_rate;
766 unsigned long task_ratelimit;
767 unsigned long balanced_dirty_ratelimit;
768 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600769 unsigned long step;
770 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600771
772 /*
773 * The dirty rate will match the writeout rate in long term, except
774 * when dirty pages are truncated by userspace or re-dirtied by FS.
775 */
776 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
777
778 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
779 bdi_thresh, bdi_dirty);
780 /*
781 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
782 */
783 task_ratelimit = (u64)dirty_ratelimit *
784 pos_ratio >> RATELIMIT_CALC_SHIFT;
785 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
786
787 /*
788 * A linear estimation of the "balanced" throttle rate. The theory is,
789 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
790 * dirty_rate will be measured to be (N * task_ratelimit). So the below
791 * formula will yield the balanced rate limit (write_bw / N).
792 *
793 * Note that the expanded form is not a pure rate feedback:
794 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
795 * but also takes pos_ratio into account:
796 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
797 *
798 * (1) is not realistic because pos_ratio also takes part in balancing
799 * the dirty rate. Consider the state
800 * pos_ratio = 0.5 (3)
801 * rate = 2 * (write_bw / N) (4)
802 * If (1) is used, it will stuck in that state! Because each dd will
803 * be throttled at
804 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
805 * yielding
806 * dirty_rate = N * task_ratelimit = write_bw (6)
807 * put (6) into (1) we get
808 * rate_(i+1) = rate_(i) (7)
809 *
810 * So we end up using (2) to always keep
811 * rate_(i+1) ~= (write_bw / N) (8)
812 * regardless of the value of pos_ratio. As long as (8) is satisfied,
813 * pos_ratio is able to drive itself to 1.0, which is not only where
814 * the dirty count meet the setpoint, but also where the slope of
815 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
816 */
817 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
818 dirty_rate | 1);
819
Wu Fengguang73811312011-08-26 15:53:24 -0600820 /*
821 * We could safely do this and return immediately:
822 *
823 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
824 *
825 * However to get a more stable dirty_ratelimit, the below elaborated
826 * code makes use of task_ratelimit to filter out sigular points and
827 * limit the step size.
828 *
829 * The below code essentially only uses the relative value of
830 *
831 * task_ratelimit - dirty_ratelimit
832 * = (pos_ratio - 1) * dirty_ratelimit
833 *
834 * which reflects the direction and size of dirty position error.
835 */
836
837 /*
838 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
839 * task_ratelimit is on the same side of dirty_ratelimit, too.
840 * For example, when
841 * - dirty_ratelimit > balanced_dirty_ratelimit
842 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
843 * lowering dirty_ratelimit will help meet both the position and rate
844 * control targets. Otherwise, don't update dirty_ratelimit if it will
845 * only help meet the rate target. After all, what the users ultimately
846 * feel and care are stable dirty rate and small position error.
847 *
848 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
849 * and filter out the sigular points of balanced_dirty_ratelimit. Which
850 * keeps jumping around randomly and can even leap far away at times
851 * due to the small 200ms estimation period of dirty_rate (we want to
852 * keep that period small to reduce time lags).
853 */
854 step = 0;
855 if (dirty < setpoint) {
856 x = min(bdi->balanced_dirty_ratelimit,
857 min(balanced_dirty_ratelimit, task_ratelimit));
858 if (dirty_ratelimit < x)
859 step = x - dirty_ratelimit;
860 } else {
861 x = max(bdi->balanced_dirty_ratelimit,
862 max(balanced_dirty_ratelimit, task_ratelimit));
863 if (dirty_ratelimit > x)
864 step = dirty_ratelimit - x;
865 }
866
867 /*
868 * Don't pursue 100% rate matching. It's impossible since the balanced
869 * rate itself is constantly fluctuating. So decrease the track speed
870 * when it gets close to the target. Helps eliminate pointless tremors.
871 */
872 step >>= dirty_ratelimit / (2 * step + 1);
873 /*
874 * Limit the tracking speed to avoid overshooting.
875 */
876 step = (step + 7) / 8;
877
878 if (dirty_ratelimit < balanced_dirty_ratelimit)
879 dirty_ratelimit += step;
880 else
881 dirty_ratelimit -= step;
882
883 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
884 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -0600885
886 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600887}
888
Wu Fengguange98be2d2010-08-29 11:22:30 -0600889void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600890 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600891 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600892 unsigned long dirty,
893 unsigned long bdi_thresh,
894 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600895 unsigned long start_time)
896{
897 unsigned long now = jiffies;
898 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600899 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600900 unsigned long written;
901
902 /*
903 * rate-limit, only update once every 200ms.
904 */
905 if (elapsed < BANDWIDTH_INTERVAL)
906 return;
907
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600908 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600909 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
910
911 /*
912 * Skip quiet periods when disk bandwidth is under-utilized.
913 * (at least 1s idle time between two flusher runs)
914 */
915 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
916 goto snapshot;
917
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600918 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -0600919 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600920 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
921 bdi_thresh, bdi_dirty,
922 dirtied, elapsed);
923 }
Wu Fengguange98be2d2010-08-29 11:22:30 -0600924 bdi_update_write_bandwidth(bdi, elapsed, written);
925
926snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600927 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600928 bdi->written_stamp = written;
929 bdi->bw_time_stamp = now;
930}
931
932static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600933 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600934 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600935 unsigned long dirty,
936 unsigned long bdi_thresh,
937 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600938 unsigned long start_time)
939{
940 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
941 return;
942 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600943 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
944 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600945 spin_unlock(&bdi->wb.list_lock);
946}
947
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948/*
Wu Fengguang9d823e82011-06-11 18:10:12 -0600949 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
950 * will look to see if it needs to start dirty throttling.
951 *
952 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
953 * global_page_state() too often. So scale it near-sqrt to the safety margin
954 * (the number of pages we may dirty without exceeding the dirty limits).
955 */
956static unsigned long dirty_poll_interval(unsigned long dirty,
957 unsigned long thresh)
958{
959 if (thresh > dirty)
960 return 1UL << (ilog2(thresh - dirty) >> 1);
961
962 return 1;
963}
964
Wu Fengguangc8462cc2011-06-11 19:21:43 -0600965static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
966 unsigned long bdi_dirty)
967{
968 unsigned long bw = bdi->avg_write_bandwidth;
969 unsigned long hi = ilog2(bw);
970 unsigned long lo = ilog2(bdi->dirty_ratelimit);
971 unsigned long t;
972
973 /* target for 20ms max pause on 1-dd case */
974 t = HZ / 50;
975
976 /*
977 * Scale up pause time for concurrent dirtiers in order to reduce CPU
978 * overheads.
979 *
980 * (N * 20ms) on 2^N concurrent tasks.
981 */
982 if (hi > lo)
983 t += (hi - lo) * (20 * HZ) / 1024;
984
985 /*
986 * Limit pause time for small memory systems. If sleeping for too long
987 * time, a small pool of dirty/writeback pages may go empty and disk go
988 * idle.
989 *
990 * 8 serves as the safety ratio.
991 */
Wu Fengguang82e230a2011-12-02 18:21:51 -0600992 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
Wu Fengguangc8462cc2011-06-11 19:21:43 -0600993
994 /*
995 * The pause time will be settled within range (max_pause/4, max_pause).
996 * Apply a minimal value of 4 to get a non-zero max_pause/4.
997 */
998 return clamp_val(t, 4, MAX_PAUSE);
999}
1000
Wu Fengguang9d823e82011-06-11 18:10:12 -06001001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 * balance_dirty_pages() must be called by processes which are generating dirty
1003 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001004 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001005 * If we're over `background_thresh' then the writeback threads are woken to
1006 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001008static void balance_dirty_pages(struct address_space *mapping,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001009 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010{
Wu Fengguang143dfe82010-08-27 18:45:12 -06001011 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1012 unsigned long bdi_reclaimable;
Wu Fengguang77627412010-09-12 13:34:05 -06001013 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1014 unsigned long bdi_dirty;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001015 unsigned long freerun;
David Rientjes364aeb22009-01-06 14:39:29 -08001016 unsigned long background_thresh;
1017 unsigned long dirty_thresh;
1018 unsigned long bdi_thresh;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001019 long pause = 0;
Wu Fengguang50657fc2011-10-11 17:06:33 -06001020 long uninitialized_var(max_pause);
Wu Fengguange50e3722010-08-11 14:17:37 -07001021 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001022 unsigned long task_ratelimit;
Wu Fengguang50657fc2011-10-11 17:06:33 -06001023 unsigned long uninitialized_var(dirty_ratelimit);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001024 unsigned long pos_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001026 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027
1028 for (;;) {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001029 /*
1030 * Unstable writes are a feature of certain networked
1031 * filesystems (i.e. NFS) in which data may have been
1032 * written to the server's write cache, but has not yet
1033 * been flushed to permanent storage.
1034 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001035 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1036 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -06001037 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001038
Wu Fengguang16c40422010-08-11 14:17:39 -07001039 global_dirty_limits(&background_thresh, &dirty_thresh);
1040
1041 /*
1042 * Throttle it only when the background writeback cannot
1043 * catch-up. This avoids (excessively) small writeouts
1044 * when the bdi limits are ramping up.
1045 */
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001046 freerun = dirty_freerun_ceiling(dirty_thresh,
1047 background_thresh);
1048 if (nr_dirty <= freerun)
Wu Fengguang16c40422010-08-11 14:17:39 -07001049 break;
1050
Wu Fengguang143dfe82010-08-27 18:45:12 -06001051 if (unlikely(!writeback_in_progress(bdi)))
1052 bdi_start_background_writeback(bdi);
1053
1054 /*
1055 * bdi_thresh is not treated as some limiting factor as
1056 * dirty_thresh, due to reasons
1057 * - in JBOD setup, bdi_thresh can fluctuate a lot
1058 * - in a system with HDD and USB key, the USB key may somehow
1059 * go into state (bdi_dirty >> bdi_thresh) either because
1060 * bdi_dirty starts high, or because bdi_thresh drops low.
1061 * In this case we don't want to hard throttle the USB key
1062 * dirtiers for 100 seconds until bdi_dirty drops under
1063 * bdi_thresh. Instead the auxiliary bdi control line in
1064 * bdi_position_ratio() will let the dirtier task progress
1065 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1066 */
Wu Fengguang16c40422010-08-11 14:17:39 -07001067 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001068
Wu Fengguange50e3722010-08-11 14:17:37 -07001069 /*
1070 * In order to avoid the stacked BDI deadlock we need
1071 * to ensure we accurately count the 'dirty' pages when
1072 * the threshold is low.
1073 *
1074 * Otherwise it would be possible to get thresh+n pages
1075 * reported dirty, even though there are thresh-m pages
1076 * actually dirty; with m+n sitting in the percpu
1077 * deltas.
1078 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001079 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1080 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1081 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001082 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001083 } else {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001084 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1085 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001086 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001087 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001088
Wu Fengguang143dfe82010-08-27 18:45:12 -06001089 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
Wu Fengguang77627412010-09-12 13:34:05 -06001090 (nr_dirty > dirty_thresh);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001091 if (dirty_exceeded && !bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001092 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001094 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1095 nr_dirty, bdi_thresh, bdi_dirty,
1096 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001097
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001098 max_pause = bdi_max_pause(bdi, bdi_dirty);
1099
Wu Fengguang143dfe82010-08-27 18:45:12 -06001100 dirty_ratelimit = bdi->dirty_ratelimit;
1101 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1102 background_thresh, nr_dirty,
1103 bdi_thresh, bdi_dirty);
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001104 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1105 RATELIMIT_CALC_SHIFT;
1106 if (unlikely(task_ratelimit == 0)) {
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001107 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001108 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 }
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001110 pause = HZ * pages_dirtied / task_ratelimit;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001111 if (unlikely(pause <= 0)) {
Wu Fengguangece13ac2010-08-29 23:33:20 -06001112 trace_balance_dirty_pages(bdi,
1113 dirty_thresh,
1114 background_thresh,
1115 nr_dirty,
1116 bdi_thresh,
1117 bdi_dirty,
1118 dirty_ratelimit,
1119 task_ratelimit,
1120 pages_dirtied,
1121 pause,
1122 start_time);
Wu Fengguang57fc9782011-06-11 19:32:32 -06001123 pause = 1; /* avoid resetting nr_dirtied_pause below */
1124 break;
1125 }
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001126 pause = min(pause, max_pause);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001127
1128pause:
Wu Fengguangece13ac2010-08-29 23:33:20 -06001129 trace_balance_dirty_pages(bdi,
1130 dirty_thresh,
1131 background_thresh,
1132 nr_dirty,
1133 bdi_thresh,
1134 bdi_dirty,
1135 dirty_ratelimit,
1136 task_ratelimit,
1137 pages_dirtied,
1138 pause,
1139 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001140 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001141 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001142
Wu Fengguangffd1f602011-06-19 22:18:42 -06001143 /*
Wu Fengguang1df64712011-11-13 19:47:32 -06001144 * This is typically equal to (nr_dirty < dirty_thresh) and can
1145 * also keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001146 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001147 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001148 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001149
Wu Fengguangc5c63432011-12-02 10:21:33 -06001150 /*
1151 * In the case of an unresponding NFS server and the NFS dirty
1152 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1153 * to go through, so that tasks on them still remain responsive.
1154 *
1155 * In theory 1 page is enough to keep the comsumer-producer
1156 * pipe going: the flusher cleans 1 page => the task dirties 1
1157 * more page. However bdi_dirty has accounting errors. So use
1158 * the larger and more IO friendly bdi_stat_error.
1159 */
1160 if (bdi_dirty <= bdi_stat_error(bdi))
1161 break;
1162
Jan Kara499d05e2011-11-16 19:34:48 +08001163 if (fatal_signal_pending(current))
1164 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 }
1166
Wu Fengguang143dfe82010-08-27 18:45:12 -06001167 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001168 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
Wu Fengguang9d823e82011-06-11 18:10:12 -06001170 current->nr_dirtied = 0;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001171 if (pause == 0) { /* in freerun area */
1172 current->nr_dirtied_pause =
1173 dirty_poll_interval(nr_dirty, dirty_thresh);
1174 } else if (pause <= max_pause / 4 &&
1175 pages_dirtied >= current->nr_dirtied_pause) {
1176 current->nr_dirtied_pause = clamp_val(
1177 dirty_ratelimit * (max_pause / 2) / HZ,
1178 pages_dirtied + pages_dirtied / 8,
1179 pages_dirtied * 4);
1180 } else if (pause >= max_pause) {
1181 current->nr_dirtied_pause = 1 | clamp_val(
1182 dirty_ratelimit * (max_pause / 2) / HZ,
1183 pages_dirtied / 4,
1184 pages_dirtied - pages_dirtied / 8);
1185 }
Wu Fengguang9d823e82011-06-11 18:10:12 -06001186
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001188 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189
1190 /*
1191 * In laptop mode, we wait until hitting the higher threshold before
1192 * starting background writeout, and then write out all the way down
1193 * to the lower threshold. So slow writers cause minimal disk activity.
1194 *
1195 * In normal mode, we start background writeout at the lower
1196 * background_thresh, to keep the amount of dirty memory low.
1197 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001198 if (laptop_mode)
1199 return;
1200
1201 if (nr_reclaimable > background_thresh)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001202 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203}
1204
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001205void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001206{
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001207 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001208 struct address_space *mapping = page_mapping(page);
1209
1210 if (mapping)
1211 balance_dirty_pages_ratelimited(mapping);
1212 }
1213}
1214
Wu Fengguang9d823e82011-06-11 18:10:12 -06001215static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001216
Wu Fengguang54848d72011-04-05 13:21:19 -06001217/*
1218 * Normal tasks are throttled by
1219 * loop {
1220 * dirty tsk->nr_dirtied_pause pages;
1221 * take a snap in balance_dirty_pages();
1222 * }
1223 * However there is a worst case. If every task exit immediately when dirtied
1224 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1225 * called to throttle the page dirties. The solution is to save the not yet
1226 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1227 * randomly into the running tasks. This works well for the above worst case,
1228 * as the new task will pick up and accumulate the old task's leaked dirty
1229 * count and eventually get throttled.
1230 */
1231DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1232
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001234 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001235 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +02001236 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 *
1238 * Processes which are dirtying memory should call in here once for each page
1239 * which was newly dirtied. The function will periodically check the system's
1240 * dirty state and will initiate writeback if needed.
1241 *
1242 * On really big machines, get_writeback_state is expensive, so try to avoid
1243 * calling it too often (ratelimiting). But once we're over the dirty memory
1244 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1245 * from overshooting the limit by (ratelimit_pages) each.
1246 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001247void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1248 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001250 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001251 int ratelimit;
1252 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
Wu Fengguang36715ce2011-06-11 17:53:57 -06001254 if (!bdi_cap_account_dirty(bdi))
1255 return;
1256
Wu Fengguang9d823e82011-06-11 18:10:12 -06001257 ratelimit = current->nr_dirtied_pause;
1258 if (bdi->dirty_exceeded)
1259 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001261 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001262 /*
1263 * This prevents one CPU to accumulate too many dirtied pages without
1264 * calling into balance_dirty_pages(), which can happen when there are
1265 * 1000+ tasks, all of them start dirtying pages at exactly the same
1266 * time, hence all honoured too large initial task->nr_dirtied_pause.
1267 */
Tejun Heo245b2e72009-06-24 15:13:48 +09001268 p = &__get_cpu_var(bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001269 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001270 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001271 else if (unlikely(*p >= ratelimit_pages)) {
1272 *p = 0;
1273 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001275 /*
1276 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1277 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1278 * the dirty throttling and livelock other long-run dirtiers.
1279 */
1280 p = &__get_cpu_var(dirty_throttle_leaks);
1281 if (*p > 0 && current->nr_dirtied < ratelimit) {
1282 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1283 *p -= nr_pages_dirtied;
1284 current->nr_dirtied += nr_pages_dirtied;
1285 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001286 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001287
1288 if (unlikely(current->nr_dirtied >= ratelimit))
1289 balance_dirty_pages(mapping, current->nr_dirtied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290}
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001291EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292
Andrew Morton232ea4d2007-02-28 20:13:21 -08001293void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294{
David Rientjes364aeb22009-01-06 14:39:29 -08001295 unsigned long background_thresh;
1296 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297
1298 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001299 global_dirty_limits(&background_thresh, &dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300
1301 /*
1302 * Boost the allowable dirty threshold a bit for page
1303 * allocators so they don't get DoS'ed by heavy writers
1304 */
1305 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1306
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001307 if (global_page_state(NR_UNSTABLE_NFS) +
1308 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1309 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001310 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001311
1312 /*
1313 * The caller might hold locks which can prevent IO completion
1314 * or progress in the filesystem. So we cannot just sit here
1315 * waiting for IO to complete.
1316 */
1317 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1318 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 }
1320}
1321
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1324 */
1325int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001326 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001328 proc_dointvec(table, write, buffer, length, ppos);
Jens Axboe64231042010-05-21 20:00:35 +02001329 bdi_arm_supers_timer();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 return 0;
1331}
1332
Jens Axboec2c49862010-05-20 09:18:47 +02001333#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001334void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335{
Matthew Garrett31373d02010-04-06 14:25:14 +02001336 struct request_queue *q = (struct request_queue *)data;
1337 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1338 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339
Matthew Garrett31373d02010-04-06 14:25:14 +02001340 /*
1341 * We want to write everything out, not just down to the dirty
1342 * threshold
1343 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001344 if (bdi_has_dirty_io(&q->backing_dev_info))
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001345 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1346 WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347}
1348
1349/*
1350 * We've spun up the disk and we're in laptop mode: schedule writeback
1351 * of all dirty data a few seconds from now. If the flush is already scheduled
1352 * then push it back - the user is still using the disk.
1353 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001354void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355{
Matthew Garrett31373d02010-04-06 14:25:14 +02001356 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357}
1358
1359/*
1360 * We're in laptop mode and we've just synced. The sync's writes will have
1361 * caused another writeback to be scheduled by laptop_io_completion.
1362 * Nothing needs to be written back anymore, so we unschedule the writeback.
1363 */
1364void laptop_sync_completion(void)
1365{
Matthew Garrett31373d02010-04-06 14:25:14 +02001366 struct backing_dev_info *bdi;
1367
1368 rcu_read_lock();
1369
1370 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1371 del_timer(&bdi->laptop_mode_wb_timer);
1372
1373 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374}
Jens Axboec2c49862010-05-20 09:18:47 +02001375#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
1377/*
1378 * If ratelimit_pages is too high then we can get into dirty-data overload
1379 * if a large number of processes all perform writes at the same time.
1380 * If it is too low then SMP machines will call the (expensive)
1381 * get_writeback_state too often.
1382 *
1383 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1384 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06001385 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 */
1387
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001388void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389{
Wu Fengguang9d823e82011-06-11 18:10:12 -06001390 unsigned long background_thresh;
1391 unsigned long dirty_thresh;
1392 global_dirty_limits(&background_thresh, &dirty_thresh);
1393 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 if (ratelimit_pages < 16)
1395 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396}
1397
Chandra Seetharaman26c21432006-06-27 02:54:10 -07001398static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1400{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001401 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -08001402 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403}
1404
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001405static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 .notifier_call = ratelimit_handler,
1407 .next = NULL,
1408};
1409
1410/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001411 * Called early on to tune the page writeback dirty limits.
1412 *
1413 * We used to scale dirty pages according to how total memory
1414 * related to pages that could be allocated for buffers (by
1415 * comparing nr_free_buffer_pages() to vm_total_pages.
1416 *
1417 * However, that was when we used "dirty_ratio" to scale with
1418 * all memory, and we don't do that any more. "dirty_ratio"
1419 * is now applied to total non-HIGHPAGE memory (by subtracting
1420 * totalhigh_pages from vm_total_pages), and as such we can't
1421 * get into the old insane situation any more where we had
1422 * large amounts of dirty pages compared to a small amount of
1423 * non-HIGHMEM memory.
1424 *
1425 * But we might still want to scale the dirty_ratio by how
1426 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 */
1428void __init page_writeback_init(void)
1429{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001430 int shift;
1431
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001432 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001434
1435 shift = calc_period_shift();
1436 prop_descriptor_init(&vm_completions, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437}
1438
David Howells811d7362006-08-29 19:06:09 +01001439/**
Jan Karaf446daa2010-08-09 17:19:12 -07001440 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1441 * @mapping: address space structure to write
1442 * @start: starting page index
1443 * @end: ending page index (inclusive)
1444 *
1445 * This function scans the page range from @start to @end (inclusive) and tags
1446 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1447 * that write_cache_pages (or whoever calls this function) will then use
1448 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1449 * used to avoid livelocking of writeback by a process steadily creating new
1450 * dirty pages in the file (thus it is important for this function to be quick
1451 * so that it can tag pages faster than a dirtying process can create them).
1452 */
1453/*
1454 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1455 */
Jan Karaf446daa2010-08-09 17:19:12 -07001456void tag_pages_for_writeback(struct address_space *mapping,
1457 pgoff_t start, pgoff_t end)
1458{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001459#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daa2010-08-09 17:19:12 -07001460 unsigned long tagged;
1461
1462 do {
1463 spin_lock_irq(&mapping->tree_lock);
1464 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1465 &start, end, WRITEBACK_TAG_BATCH,
1466 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1467 spin_unlock_irq(&mapping->tree_lock);
1468 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1469 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001470 /* We check 'start' to handle wrapping when end == ~0UL */
1471 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daa2010-08-09 17:19:12 -07001472}
1473EXPORT_SYMBOL(tag_pages_for_writeback);
1474
1475/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001476 * 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 +01001477 * @mapping: address space structure to write
1478 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001479 * @writepage: function called for each page
1480 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001481 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001482 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001483 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1484 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1485 * and msync() need to guarantee that all the data which was dirty at the time
1486 * the call was made get new I/O started against them. If wbc->sync_mode is
1487 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1488 * existing IO to complete.
Jan Karaf446daa2010-08-09 17:19:12 -07001489 *
1490 * To avoid livelocks (when other process dirties new pages), we first tag
1491 * pages which should be written back with TOWRITE tag and only then start
1492 * writing them. For data-integrity sync we have to be careful so that we do
1493 * not miss some pages (e.g., because some other process has cleared TOWRITE
1494 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1495 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001496 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001497int write_cache_pages(struct address_space *mapping,
1498 struct writeback_control *wbc, writepage_t writepage,
1499 void *data)
David Howells811d7362006-08-29 19:06:09 +01001500{
David Howells811d7362006-08-29 19:06:09 +01001501 int ret = 0;
1502 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001503 struct pagevec pvec;
1504 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001505 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001506 pgoff_t index;
1507 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001508 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001509 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001510 int range_whole = 0;
Jan Karaf446daa2010-08-09 17:19:12 -07001511 int tag;
David Howells811d7362006-08-29 19:06:09 +01001512
David Howells811d7362006-08-29 19:06:09 +01001513 pagevec_init(&pvec, 0);
1514 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001515 writeback_index = mapping->writeback_index; /* prev offset */
1516 index = writeback_index;
1517 if (index == 0)
1518 cycled = 1;
1519 else
1520 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001521 end = -1;
1522 } else {
1523 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1524 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1525 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1526 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001527 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001528 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001529 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daa2010-08-09 17:19:12 -07001530 tag = PAGECACHE_TAG_TOWRITE;
1531 else
1532 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001533retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001534 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daa2010-08-09 17:19:12 -07001535 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001536 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001537 while (!done && (index <= end)) {
1538 int i;
1539
Jan Karaf446daa2010-08-09 17:19:12 -07001540 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001541 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1542 if (nr_pages == 0)
1543 break;
David Howells811d7362006-08-29 19:06:09 +01001544
David Howells811d7362006-08-29 19:06:09 +01001545 for (i = 0; i < nr_pages; i++) {
1546 struct page *page = pvec.pages[i];
1547
Nick Piggind5482cd2009-01-06 14:39:11 -08001548 /*
1549 * At this point, the page may be truncated or
1550 * invalidated (changing page->mapping to NULL), or
1551 * even swizzled back from swapper_space to tmpfs file
1552 * mapping. However, page->index will not change
1553 * because we have a reference on the page.
1554 */
1555 if (page->index > end) {
1556 /*
1557 * can't be range_cyclic (1st pass) because
1558 * end == -1 in that case.
1559 */
1560 done = 1;
1561 break;
1562 }
1563
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001564 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001565
David Howells811d7362006-08-29 19:06:09 +01001566 lock_page(page);
1567
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001568 /*
1569 * Page truncated or invalidated. We can freely skip it
1570 * then, even for data integrity operations: the page
1571 * has disappeared concurrently, so there could be no
1572 * real expectation of this data interity operation
1573 * even if there is now a new, dirty page at the same
1574 * pagecache address.
1575 */
David Howells811d7362006-08-29 19:06:09 +01001576 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001577continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001578 unlock_page(page);
1579 continue;
1580 }
1581
Nick Piggin515f4a02009-01-06 14:39:10 -08001582 if (!PageDirty(page)) {
1583 /* someone wrote it for us */
1584 goto continue_unlock;
1585 }
David Howells811d7362006-08-29 19:06:09 +01001586
Nick Piggin515f4a02009-01-06 14:39:10 -08001587 if (PageWriteback(page)) {
1588 if (wbc->sync_mode != WB_SYNC_NONE)
1589 wait_on_page_writeback(page);
1590 else
1591 goto continue_unlock;
1592 }
1593
1594 BUG_ON(PageWriteback(page));
1595 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001596 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001597
Dave Chinner9e094382010-07-07 13:24:08 +10001598 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001599 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001600 if (unlikely(ret)) {
1601 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1602 unlock_page(page);
1603 ret = 0;
1604 } else {
1605 /*
1606 * done_index is set past this page,
1607 * so media errors will not choke
1608 * background writeout for the entire
1609 * file. This has consequences for
1610 * range_cyclic semantics (ie. it may
1611 * not be suitable for data integrity
1612 * writeout).
1613 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001614 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001615 done = 1;
1616 break;
1617 }
Dave Chinner0b564922010-06-09 10:37:18 +10001618 }
David Howells811d7362006-08-29 19:06:09 +01001619
Dave Chinner546a1922010-08-24 11:44:34 +10001620 /*
1621 * We stop writing back only if we are not doing
1622 * integrity sync. In case of integrity sync we have to
1623 * keep going until we have written all the pages
1624 * we tagged for writeback prior to entering this loop.
1625 */
1626 if (--wbc->nr_to_write <= 0 &&
1627 wbc->sync_mode == WB_SYNC_NONE) {
1628 done = 1;
1629 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001630 }
David Howells811d7362006-08-29 19:06:09 +01001631 }
1632 pagevec_release(&pvec);
1633 cond_resched();
1634 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001635 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001636 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001637 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001638 * We hit the last page and there is more work to be done: wrap
1639 * back to the start of the file
1640 */
Nick Piggin31a12662009-01-06 14:39:04 -08001641 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001642 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001643 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001644 goto retry;
1645 }
Dave Chinner0b564922010-06-09 10:37:18 +10001646 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1647 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001648
David Howells811d7362006-08-29 19:06:09 +01001649 return ret;
1650}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001651EXPORT_SYMBOL(write_cache_pages);
1652
1653/*
1654 * Function used by generic_writepages to call the real writepage
1655 * function and set the mapping flags on error
1656 */
1657static int __writepage(struct page *page, struct writeback_control *wbc,
1658 void *data)
1659{
1660 struct address_space *mapping = data;
1661 int ret = mapping->a_ops->writepage(page, wbc);
1662 mapping_set_error(mapping, ret);
1663 return ret;
1664}
1665
1666/**
1667 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1668 * @mapping: address space structure to write
1669 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1670 *
1671 * This is a library function, which implements the writepages()
1672 * address_space_operation.
1673 */
1674int generic_writepages(struct address_space *mapping,
1675 struct writeback_control *wbc)
1676{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001677 struct blk_plug plug;
1678 int ret;
1679
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001680 /* deal with chardevs and other special file */
1681 if (!mapping->a_ops->writepage)
1682 return 0;
1683
Shaohua Li9b6096a2011-03-17 10:47:06 +01001684 blk_start_plug(&plug);
1685 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1686 blk_finish_plug(&plug);
1687 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001688}
David Howells811d7362006-08-29 19:06:09 +01001689
1690EXPORT_SYMBOL(generic_writepages);
1691
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1693{
Andrew Morton22905f72005-11-16 15:07:01 -08001694 int ret;
1695
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 if (wbc->nr_to_write <= 0)
1697 return 0;
1698 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001699 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001700 else
1701 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001702 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703}
1704
1705/**
1706 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001707 * @page: the page to write
1708 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 *
1710 * The page must be locked by the caller and will be unlocked upon return.
1711 *
1712 * write_one_page() returns a negative error code if I/O failed.
1713 */
1714int write_one_page(struct page *page, int wait)
1715{
1716 struct address_space *mapping = page->mapping;
1717 int ret = 0;
1718 struct writeback_control wbc = {
1719 .sync_mode = WB_SYNC_ALL,
1720 .nr_to_write = 1,
1721 };
1722
1723 BUG_ON(!PageLocked(page));
1724
1725 if (wait)
1726 wait_on_page_writeback(page);
1727
1728 if (clear_page_dirty_for_io(page)) {
1729 page_cache_get(page);
1730 ret = mapping->a_ops->writepage(page, &wbc);
1731 if (ret == 0 && wait) {
1732 wait_on_page_writeback(page);
1733 if (PageError(page))
1734 ret = -EIO;
1735 }
1736 page_cache_release(page);
1737 } else {
1738 unlock_page(page);
1739 }
1740 return ret;
1741}
1742EXPORT_SYMBOL(write_one_page);
1743
1744/*
Ken Chen76719322007-02-10 01:43:15 -08001745 * For address_spaces which do not use buffers nor write back.
1746 */
1747int __set_page_dirty_no_writeback(struct page *page)
1748{
1749 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001750 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001751 return 0;
1752}
1753
1754/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001755 * Helper function for set_page_dirty family.
1756 * NOTE: This relies on being atomic wrt interrupts.
1757 */
1758void account_page_dirtied(struct page *page, struct address_space *mapping)
1759{
1760 if (mapping_cap_account_dirty(mapping)) {
1761 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001762 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001763 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06001764 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001765 task_io_account_write(PAGE_CACHE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001766 current->nr_dirtied++;
1767 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001768 }
1769}
Michael Rubin679ceac2010-08-20 02:31:26 -07001770EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001771
1772/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001773 * Helper function for set_page_writeback family.
1774 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1775 * wrt interrupts.
1776 */
1777void account_page_writeback(struct page *page)
1778{
1779 inc_zone_page_state(page, NR_WRITEBACK);
1780}
1781EXPORT_SYMBOL(account_page_writeback);
1782
1783/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 * For address_spaces which do not use buffers. Just tag the page as dirty in
1785 * its radix tree.
1786 *
1787 * This is also used when a single buffer is being dirtied: we want to set the
1788 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1789 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1790 *
1791 * Most callers have locked the page, which pins the address_space in memory.
1792 * But zap_pte_range() does not lock the page, however in that case the
1793 * mapping is pinned by the vma's ->vm_file reference.
1794 *
1795 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001796 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 */
1798int __set_page_dirty_nobuffers(struct page *page)
1799{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 if (!TestSetPageDirty(page)) {
1801 struct address_space *mapping = page_mapping(page);
1802 struct address_space *mapping2;
1803
Andrew Morton8c085402006-12-10 02:19:24 -08001804 if (!mapping)
1805 return 1;
1806
Nick Piggin19fd6232008-07-25 19:45:32 -07001807 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001808 mapping2 = page_mapping(page);
1809 if (mapping2) { /* Race with truncate? */
1810 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001811 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001812 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001813 radix_tree_tag_set(&mapping->page_tree,
1814 page_index(page), PAGECACHE_TAG_DIRTY);
1815 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001816 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001817 if (mapping->host) {
1818 /* !PageAnon && !swapper_space */
1819 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001821 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001823 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824}
1825EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1826
1827/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06001828 * Call this whenever redirtying a page, to de-account the dirty counters
1829 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
1830 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
1831 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
1832 * control.
1833 */
1834void account_page_redirty(struct page *page)
1835{
1836 struct address_space *mapping = page->mapping;
1837 if (mapping && mapping_cap_account_dirty(mapping)) {
1838 current->nr_dirtied--;
1839 dec_zone_page_state(page, NR_DIRTIED);
1840 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1841 }
1842}
1843EXPORT_SYMBOL(account_page_redirty);
1844
1845/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 * When a writepage implementation decides that it doesn't want to write this
1847 * page for some reason, it should redirty the locked page via
1848 * redirty_page_for_writepage() and it should then unlock the page and return 0
1849 */
1850int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1851{
1852 wbc->pages_skipped++;
Wu Fengguang2f800fb2011-08-08 15:22:00 -06001853 account_page_redirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 return __set_page_dirty_nobuffers(page);
1855}
1856EXPORT_SYMBOL(redirty_page_for_writepage);
1857
1858/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02001859 * Dirty a page.
1860 *
1861 * For pages with a mapping this should be done under the page lock
1862 * for the benefit of asynchronous memory errors who prefer a consistent
1863 * dirty state. This rule can be broken in some special cases,
1864 * but should be better not to.
1865 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 * If the mapping doesn't provide a set_page_dirty a_op, then
1867 * just fall through and assume that it wants buffer_heads.
1868 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001869int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870{
1871 struct address_space *mapping = page_mapping(page);
1872
1873 if (likely(mapping)) {
1874 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07001875 /*
1876 * readahead/lru_deactivate_page could remain
1877 * PG_readahead/PG_reclaim due to race with end_page_writeback
1878 * About readahead, if the page is written, the flags would be
1879 * reset. So no problem.
1880 * About lru_deactivate_page, if the page is redirty, the flag
1881 * will be reset. So no problem. but if the page is used by readahead
1882 * it will confuse readahead and make it restart the size rampup
1883 * process. But it's a trivial problem.
1884 */
1885 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02001886#ifdef CONFIG_BLOCK
1887 if (!spd)
1888 spd = __set_page_dirty_buffers;
1889#endif
1890 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001892 if (!PageDirty(page)) {
1893 if (!TestSetPageDirty(page))
1894 return 1;
1895 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 return 0;
1897}
1898EXPORT_SYMBOL(set_page_dirty);
1899
1900/*
1901 * set_page_dirty() is racy if the caller has no reference against
1902 * page->mapping->host, and if the page is unlocked. This is because another
1903 * CPU could truncate the page off the mapping and then free the mapping.
1904 *
1905 * Usually, the page _is_ locked, or the caller is a user-space process which
1906 * holds a reference on the inode by having an open file.
1907 *
1908 * In other cases, the page should be locked before running set_page_dirty().
1909 */
1910int set_page_dirty_lock(struct page *page)
1911{
1912 int ret;
1913
Jens Axboe7eaceac2011-03-10 08:52:07 +01001914 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 ret = set_page_dirty(page);
1916 unlock_page(page);
1917 return ret;
1918}
1919EXPORT_SYMBOL(set_page_dirty_lock);
1920
1921/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 * Clear a page's dirty flag, while caring for dirty memory accounting.
1923 * Returns true if the page was previously dirty.
1924 *
1925 * This is for preparing to put the page under writeout. We leave the page
1926 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1927 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1928 * implementation will run either set_page_writeback() or set_page_dirty(),
1929 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1930 * back into sync.
1931 *
1932 * This incoherency between the page's dirty flag and radix-tree tag is
1933 * unfortunate, but it only exists while the page is locked.
1934 */
1935int clear_page_dirty_for_io(struct page *page)
1936{
1937 struct address_space *mapping = page_mapping(page);
1938
Nick Piggin79352892007-07-19 01:47:22 -07001939 BUG_ON(!PageLocked(page));
1940
Linus Torvalds7658cc22006-12-29 10:00:58 -08001941 if (mapping && mapping_cap_account_dirty(mapping)) {
1942 /*
1943 * Yes, Virginia, this is indeed insane.
1944 *
1945 * We use this sequence to make sure that
1946 * (a) we account for dirty stats properly
1947 * (b) we tell the low-level filesystem to
1948 * mark the whole page dirty if it was
1949 * dirty in a pagetable. Only to then
1950 * (c) clean the page again and return 1 to
1951 * cause the writeback.
1952 *
1953 * This way we avoid all nasty races with the
1954 * dirty bit in multiple places and clearing
1955 * them concurrently from different threads.
1956 *
1957 * Note! Normally the "set_page_dirty(page)"
1958 * has no effect on the actual dirty bit - since
1959 * that will already usually be set. But we
1960 * need the side effects, and it can help us
1961 * avoid races.
1962 *
1963 * We basically use the page "master dirty bit"
1964 * as a serialization point for all the different
1965 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001966 */
1967 if (page_mkclean(page))
1968 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001969 /*
1970 * We carefully synchronise fault handlers against
1971 * installing a dirty pte and marking the page dirty
1972 * at this point. We do this by having them hold the
1973 * page lock at some point after installing their
1974 * pte, but before marking the page dirty.
1975 * Pages are always locked coming in here, so we get
1976 * the desired exclusion. See mm/memory.c:do_wp_page()
1977 * for more comments.
1978 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001979 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001980 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001981 dec_bdi_stat(mapping->backing_dev_info,
1982 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001983 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001985 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001987 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001989EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990
1991int test_clear_page_writeback(struct page *page)
1992{
1993 struct address_space *mapping = page_mapping(page);
1994 int ret;
1995
1996 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001997 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 unsigned long flags;
1999
Nick Piggin19fd6232008-07-25 19:45:32 -07002000 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002002 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 radix_tree_tag_clear(&mapping->page_tree,
2004 page_index(page),
2005 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002006 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002007 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002008 __bdi_writeout_inc(bdi);
2009 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002010 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002011 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012 } else {
2013 ret = TestClearPageWriteback(page);
2014 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002015 if (ret) {
Andrew Mortond688abf2007-07-19 01:49:17 -07002016 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002017 inc_zone_page_state(page, NR_WRITTEN);
2018 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 return ret;
2020}
2021
2022int test_set_page_writeback(struct page *page)
2023{
2024 struct address_space *mapping = page_mapping(page);
2025 int ret;
2026
2027 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002028 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 unsigned long flags;
2030
Nick Piggin19fd6232008-07-25 19:45:32 -07002031 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002033 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 radix_tree_tag_set(&mapping->page_tree,
2035 page_index(page),
2036 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002037 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002038 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2039 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 if (!PageDirty(page))
2041 radix_tree_tag_clear(&mapping->page_tree,
2042 page_index(page),
2043 PAGECACHE_TAG_DIRTY);
Jan Karaf446daa2010-08-09 17:19:12 -07002044 radix_tree_tag_clear(&mapping->page_tree,
2045 page_index(page),
2046 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002047 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 } else {
2049 ret = TestSetPageWriteback(page);
2050 }
Andrew Mortond688abf2007-07-19 01:49:17 -07002051 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07002052 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 return ret;
2054
2055}
2056EXPORT_SYMBOL(test_set_page_writeback);
2057
2058/*
Nick Piggin00128182007-10-16 01:24:40 -07002059 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 * passed tag.
2061 */
2062int mapping_tagged(struct address_space *mapping, int tag)
2063{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002064 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065}
2066EXPORT_SYMBOL(mapping_tagged);