blob: 71252486bc6f1f161b87592ded6e8b7e6c0dc8ee [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 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/*
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 Gondwana195cf4532008-02-04 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.
414 * And the "limit" in the name is not seriously taken as hard limit in
415 * balance_dirty_pages().
416 *
417 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700418 * - starving fast devices
419 * - piling up dirty pages (that will take long time to sync) on slow devices
420 *
421 * The bdi's share of dirty limit will be adapting to its throughput and
422 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
423 */
424unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700425{
426 u64 bdi_dirty;
427 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700428
Wu Fengguang16c40422010-08-11 14:17:39 -0700429 /*
430 * Calculate this BDI's share of the dirty ratio.
431 */
432 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700433
Wu Fengguang16c40422010-08-11 14:17:39 -0700434 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
435 bdi_dirty *= numerator;
436 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700437
Wu Fengguang16c40422010-08-11 14:17:39 -0700438 bdi_dirty += (dirty * bdi->min_ratio) / 100;
439 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
440 bdi_dirty = dirty * bdi->max_ratio / 100;
441
442 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443}
444
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600445/*
446 * Dirty position control.
447 *
448 * (o) global/bdi setpoints
449 *
450 * We want the dirty pages be balanced around the global/bdi setpoints.
451 * When the number of dirty pages is higher/lower than the setpoint, the
452 * dirty position control ratio (and hence task dirty ratelimit) will be
453 * decreased/increased to bring the dirty pages back to the setpoint.
454 *
455 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
456 *
457 * if (dirty < setpoint) scale up pos_ratio
458 * if (dirty > setpoint) scale down pos_ratio
459 *
460 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
461 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
462 *
463 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
464 *
465 * (o) global control line
466 *
467 * ^ pos_ratio
468 * |
469 * | |<===== global dirty control scope ======>|
470 * 2.0 .............*
471 * | .*
472 * | . *
473 * | . *
474 * | . *
475 * | . *
476 * | . *
477 * 1.0 ................................*
478 * | . . *
479 * | . . *
480 * | . . *
481 * | . . *
482 * | . . *
483 * 0 +------------.------------------.----------------------*------------->
484 * freerun^ setpoint^ limit^ dirty pages
485 *
486 * (o) bdi control line
487 *
488 * ^ pos_ratio
489 * |
490 * | *
491 * | *
492 * | *
493 * | *
494 * | * |<=========== span ============>|
495 * 1.0 .......................*
496 * | . *
497 * | . *
498 * | . *
499 * | . *
500 * | . *
501 * | . *
502 * | . *
503 * | . *
504 * | . *
505 * | . *
506 * | . *
507 * 1/4 ...............................................* * * * * * * * * * * *
508 * | . .
509 * | . .
510 * | . .
511 * 0 +----------------------.-------------------------------.------------->
512 * bdi_setpoint^ x_intercept^
513 *
514 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
515 * be smoothly throttled down to normal if it starts high in situations like
516 * - start writing to a slow SD card and a fast disk at the same time. The SD
517 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
518 * - the bdi dirty thresh drops quickly due to change of JBOD workload
519 */
520static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
521 unsigned long thresh,
522 unsigned long bg_thresh,
523 unsigned long dirty,
524 unsigned long bdi_thresh,
525 unsigned long bdi_dirty)
526{
527 unsigned long write_bw = bdi->avg_write_bandwidth;
528 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
529 unsigned long limit = hard_dirty_limit(thresh);
530 unsigned long x_intercept;
531 unsigned long setpoint; /* dirty pages' target balance point */
532 unsigned long bdi_setpoint;
533 unsigned long span;
534 long long pos_ratio; /* for scaling up/down the rate limit */
535 long x;
536
537 if (unlikely(dirty >= limit))
538 return 0;
539
540 /*
541 * global setpoint
542 *
543 * setpoint - dirty 3
544 * f(dirty) := 1.0 + (----------------)
545 * limit - setpoint
546 *
547 * it's a 3rd order polynomial that subjects to
548 *
549 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
550 * (2) f(setpoint) = 1.0 => the balance point
551 * (3) f(limit) = 0 => the hard limit
552 * (4) df/dx <= 0 => negative feedback control
553 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
554 * => fast response on large errors; small oscillation near setpoint
555 */
556 setpoint = (freerun + limit) / 2;
557 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
558 limit - setpoint + 1);
559 pos_ratio = x;
560 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
561 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
562 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
563
564 /*
565 * We have computed basic pos_ratio above based on global situation. If
566 * the bdi is over/under its share of dirty pages, we want to scale
567 * pos_ratio further down/up. That is done by the following mechanism.
568 */
569
570 /*
571 * bdi setpoint
572 *
573 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
574 *
575 * x_intercept - bdi_dirty
576 * := --------------------------
577 * x_intercept - bdi_setpoint
578 *
579 * The main bdi control line is a linear function that subjects to
580 *
581 * (1) f(bdi_setpoint) = 1.0
582 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
583 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
584 *
585 * For single bdi case, the dirty pages are observed to fluctuate
586 * regularly within range
587 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
588 * for various filesystems, where (2) can yield in a reasonable 12.5%
589 * fluctuation range for pos_ratio.
590 *
591 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
592 * own size, so move the slope over accordingly and choose a slope that
593 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
594 */
595 if (unlikely(bdi_thresh > thresh))
596 bdi_thresh = thresh;
Wu Fengguang8927f662011-08-04 22:16:46 -0600597 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600598 /*
599 * scale global setpoint to bdi's:
600 * bdi_setpoint = setpoint * bdi_thresh / thresh
601 */
602 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
603 bdi_setpoint = setpoint * (u64)x >> 16;
604 /*
605 * Use span=(8*write_bw) in single bdi case as indicated by
606 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
607 *
608 * bdi_thresh thresh - bdi_thresh
609 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
610 * thresh thresh
611 */
612 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
613 x_intercept = bdi_setpoint + span;
614
615 if (bdi_dirty < x_intercept - span / 4) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600616 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
617 x_intercept - bdi_setpoint + 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600618 } else
619 pos_ratio /= 4;
620
Wu Fengguang8927f662011-08-04 22:16:46 -0600621 /*
622 * bdi reserve area, safeguard against dirty pool underrun and disk idle
623 * It may push the desired control point of global dirty pages higher
624 * than setpoint.
625 */
626 x_intercept = bdi_thresh / 2;
627 if (bdi_dirty < x_intercept) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600628 if (bdi_dirty > x_intercept / 8)
629 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
630 else
Wu Fengguang8927f662011-08-04 22:16:46 -0600631 pos_ratio *= 8;
632 }
633
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600634 return pos_ratio;
635}
636
Wu Fengguange98be2d2010-08-29 11:22:30 -0600637static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
638 unsigned long elapsed,
639 unsigned long written)
640{
641 const unsigned long period = roundup_pow_of_two(3 * HZ);
642 unsigned long avg = bdi->avg_write_bandwidth;
643 unsigned long old = bdi->write_bandwidth;
644 u64 bw;
645
646 /*
647 * bw = written * HZ / elapsed
648 *
649 * bw * elapsed + write_bandwidth * (period - elapsed)
650 * write_bandwidth = ---------------------------------------------------
651 * period
652 */
653 bw = written - bdi->written_stamp;
654 bw *= HZ;
655 if (unlikely(elapsed > period)) {
656 do_div(bw, elapsed);
657 avg = bw;
658 goto out;
659 }
660 bw += (u64)bdi->write_bandwidth * (period - elapsed);
661 bw >>= ilog2(period);
662
663 /*
664 * one more level of smoothing, for filtering out sudden spikes
665 */
666 if (avg > old && old >= (unsigned long)bw)
667 avg -= (avg - old) >> 3;
668
669 if (avg < old && old <= (unsigned long)bw)
670 avg += (old - avg) >> 3;
671
672out:
673 bdi->write_bandwidth = bw;
674 bdi->avg_write_bandwidth = avg;
675}
676
Wu Fengguangc42843f2011-03-02 15:54:09 -0600677/*
678 * The global dirtyable memory and dirty threshold could be suddenly knocked
679 * down by a large amount (eg. on the startup of KVM in a swapless system).
680 * This may throw the system into deep dirty exceeded state and throttle
681 * heavy/light dirtiers alike. To retain good responsiveness, maintain
682 * global_dirty_limit for tracking slowly down to the knocked down dirty
683 * threshold.
684 */
685static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
686{
687 unsigned long limit = global_dirty_limit;
688
689 /*
690 * Follow up in one step.
691 */
692 if (limit < thresh) {
693 limit = thresh;
694 goto update;
695 }
696
697 /*
698 * Follow down slowly. Use the higher one as the target, because thresh
699 * may drop below dirty. This is exactly the reason to introduce
700 * global_dirty_limit which is guaranteed to lie above the dirty pages.
701 */
702 thresh = max(thresh, dirty);
703 if (limit > thresh) {
704 limit -= (limit - thresh) >> 5;
705 goto update;
706 }
707 return;
708update:
709 global_dirty_limit = limit;
710}
711
712static void global_update_bandwidth(unsigned long thresh,
713 unsigned long dirty,
714 unsigned long now)
715{
716 static DEFINE_SPINLOCK(dirty_lock);
717 static unsigned long update_time;
718
719 /*
720 * check locklessly first to optimize away locking for the most time
721 */
722 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
723 return;
724
725 spin_lock(&dirty_lock);
726 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
727 update_dirty_limit(thresh, dirty);
728 update_time = now;
729 }
730 spin_unlock(&dirty_lock);
731}
732
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600733/*
734 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
735 *
736 * Normal bdi tasks will be curbed at or below it in long term.
737 * Obviously it should be around (write_bw / N) when there are N dd tasks.
738 */
739static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
740 unsigned long thresh,
741 unsigned long bg_thresh,
742 unsigned long dirty,
743 unsigned long bdi_thresh,
744 unsigned long bdi_dirty,
745 unsigned long dirtied,
746 unsigned long elapsed)
747{
Wu Fengguang73811312011-08-26 15:53:24 -0600748 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
749 unsigned long limit = hard_dirty_limit(thresh);
750 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600751 unsigned long write_bw = bdi->avg_write_bandwidth;
752 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
753 unsigned long dirty_rate;
754 unsigned long task_ratelimit;
755 unsigned long balanced_dirty_ratelimit;
756 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600757 unsigned long step;
758 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600759
760 /*
761 * The dirty rate will match the writeout rate in long term, except
762 * when dirty pages are truncated by userspace or re-dirtied by FS.
763 */
764 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
765
766 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
767 bdi_thresh, bdi_dirty);
768 /*
769 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
770 */
771 task_ratelimit = (u64)dirty_ratelimit *
772 pos_ratio >> RATELIMIT_CALC_SHIFT;
773 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
774
775 /*
776 * A linear estimation of the "balanced" throttle rate. The theory is,
777 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
778 * dirty_rate will be measured to be (N * task_ratelimit). So the below
779 * formula will yield the balanced rate limit (write_bw / N).
780 *
781 * Note that the expanded form is not a pure rate feedback:
782 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
783 * but also takes pos_ratio into account:
784 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
785 *
786 * (1) is not realistic because pos_ratio also takes part in balancing
787 * the dirty rate. Consider the state
788 * pos_ratio = 0.5 (3)
789 * rate = 2 * (write_bw / N) (4)
790 * If (1) is used, it will stuck in that state! Because each dd will
791 * be throttled at
792 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
793 * yielding
794 * dirty_rate = N * task_ratelimit = write_bw (6)
795 * put (6) into (1) we get
796 * rate_(i+1) = rate_(i) (7)
797 *
798 * So we end up using (2) to always keep
799 * rate_(i+1) ~= (write_bw / N) (8)
800 * regardless of the value of pos_ratio. As long as (8) is satisfied,
801 * pos_ratio is able to drive itself to 1.0, which is not only where
802 * the dirty count meet the setpoint, but also where the slope of
803 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
804 */
805 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
806 dirty_rate | 1);
807
Wu Fengguang73811312011-08-26 15:53:24 -0600808 /*
809 * We could safely do this and return immediately:
810 *
811 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
812 *
813 * However to get a more stable dirty_ratelimit, the below elaborated
814 * code makes use of task_ratelimit to filter out sigular points and
815 * limit the step size.
816 *
817 * The below code essentially only uses the relative value of
818 *
819 * task_ratelimit - dirty_ratelimit
820 * = (pos_ratio - 1) * dirty_ratelimit
821 *
822 * which reflects the direction and size of dirty position error.
823 */
824
825 /*
826 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
827 * task_ratelimit is on the same side of dirty_ratelimit, too.
828 * For example, when
829 * - dirty_ratelimit > balanced_dirty_ratelimit
830 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
831 * lowering dirty_ratelimit will help meet both the position and rate
832 * control targets. Otherwise, don't update dirty_ratelimit if it will
833 * only help meet the rate target. After all, what the users ultimately
834 * feel and care are stable dirty rate and small position error.
835 *
836 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
837 * and filter out the sigular points of balanced_dirty_ratelimit. Which
838 * keeps jumping around randomly and can even leap far away at times
839 * due to the small 200ms estimation period of dirty_rate (we want to
840 * keep that period small to reduce time lags).
841 */
842 step = 0;
843 if (dirty < setpoint) {
844 x = min(bdi->balanced_dirty_ratelimit,
845 min(balanced_dirty_ratelimit, task_ratelimit));
846 if (dirty_ratelimit < x)
847 step = x - dirty_ratelimit;
848 } else {
849 x = max(bdi->balanced_dirty_ratelimit,
850 max(balanced_dirty_ratelimit, task_ratelimit));
851 if (dirty_ratelimit > x)
852 step = dirty_ratelimit - x;
853 }
854
855 /*
856 * Don't pursue 100% rate matching. It's impossible since the balanced
857 * rate itself is constantly fluctuating. So decrease the track speed
858 * when it gets close to the target. Helps eliminate pointless tremors.
859 */
860 step >>= dirty_ratelimit / (2 * step + 1);
861 /*
862 * Limit the tracking speed to avoid overshooting.
863 */
864 step = (step + 7) / 8;
865
866 if (dirty_ratelimit < balanced_dirty_ratelimit)
867 dirty_ratelimit += step;
868 else
869 dirty_ratelimit -= step;
870
871 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
872 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -0600873
874 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600875}
876
Wu Fengguange98be2d2010-08-29 11:22:30 -0600877void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600878 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600879 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600880 unsigned long dirty,
881 unsigned long bdi_thresh,
882 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600883 unsigned long start_time)
884{
885 unsigned long now = jiffies;
886 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600887 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600888 unsigned long written;
889
890 /*
891 * rate-limit, only update once every 200ms.
892 */
893 if (elapsed < BANDWIDTH_INTERVAL)
894 return;
895
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600896 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600897 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
898
899 /*
900 * Skip quiet periods when disk bandwidth is under-utilized.
901 * (at least 1s idle time between two flusher runs)
902 */
903 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
904 goto snapshot;
905
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600906 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -0600907 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600908 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
909 bdi_thresh, bdi_dirty,
910 dirtied, elapsed);
911 }
Wu Fengguange98be2d2010-08-29 11:22:30 -0600912 bdi_update_write_bandwidth(bdi, elapsed, written);
913
914snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600915 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600916 bdi->written_stamp = written;
917 bdi->bw_time_stamp = now;
918}
919
920static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600921 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600922 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600923 unsigned long dirty,
924 unsigned long bdi_thresh,
925 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600926 unsigned long start_time)
927{
928 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
929 return;
930 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600931 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
932 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600933 spin_unlock(&bdi->wb.list_lock);
934}
935
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936/*
Wu Fengguang9d823e82011-06-11 18:10:12 -0600937 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
938 * will look to see if it needs to start dirty throttling.
939 *
940 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
941 * global_page_state() too often. So scale it near-sqrt to the safety margin
942 * (the number of pages we may dirty without exceeding the dirty limits).
943 */
944static unsigned long dirty_poll_interval(unsigned long dirty,
945 unsigned long thresh)
946{
947 if (thresh > dirty)
948 return 1UL << (ilog2(thresh - dirty) >> 1);
949
950 return 1;
951}
952
Wu Fengguangc8462cc2011-06-11 19:21:43 -0600953static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
954 unsigned long bdi_dirty)
955{
956 unsigned long bw = bdi->avg_write_bandwidth;
957 unsigned long hi = ilog2(bw);
958 unsigned long lo = ilog2(bdi->dirty_ratelimit);
959 unsigned long t;
960
961 /* target for 20ms max pause on 1-dd case */
962 t = HZ / 50;
963
964 /*
965 * Scale up pause time for concurrent dirtiers in order to reduce CPU
966 * overheads.
967 *
968 * (N * 20ms) on 2^N concurrent tasks.
969 */
970 if (hi > lo)
971 t += (hi - lo) * (20 * HZ) / 1024;
972
973 /*
974 * Limit pause time for small memory systems. If sleeping for too long
975 * time, a small pool of dirty/writeback pages may go empty and disk go
976 * idle.
977 *
978 * 8 serves as the safety ratio.
979 */
980 if (bdi_dirty)
981 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
982
983 /*
984 * The pause time will be settled within range (max_pause/4, max_pause).
985 * Apply a minimal value of 4 to get a non-zero max_pause/4.
986 */
987 return clamp_val(t, 4, MAX_PAUSE);
988}
989
Wu Fengguang9d823e82011-06-11 18:10:12 -0600990/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 * balance_dirty_pages() must be called by processes which are generating dirty
992 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -0600993 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +0200994 * If we're over `background_thresh' then the writeback threads are woken to
995 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800997static void balance_dirty_pages(struct address_space *mapping,
Wu Fengguang143dfe82010-08-27 18:45:12 -0600998 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999{
Wu Fengguang143dfe82010-08-27 18:45:12 -06001000 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1001 unsigned long bdi_reclaimable;
Wu Fengguang77627412010-09-12 13:34:05 -06001002 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
1003 unsigned long bdi_dirty;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001004 unsigned long freerun;
David Rientjes364aeb22009-01-06 14:39:29 -08001005 unsigned long background_thresh;
1006 unsigned long dirty_thresh;
1007 unsigned long bdi_thresh;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001008 long pause = 0;
Wu Fengguang50657fc2011-10-11 17:06:33 -06001009 long uninitialized_var(max_pause);
Wu Fengguange50e3722010-08-11 14:17:37 -07001010 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001011 unsigned long task_ratelimit;
Wu Fengguang50657fc2011-10-11 17:06:33 -06001012 unsigned long uninitialized_var(dirty_ratelimit);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001013 unsigned long pos_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001015 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016
1017 for (;;) {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001018 /*
1019 * Unstable writes are a feature of certain networked
1020 * filesystems (i.e. NFS) in which data may have been
1021 * written to the server's write cache, but has not yet
1022 * been flushed to permanent storage.
1023 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001024 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1025 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -06001026 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001027
Wu Fengguang16c40422010-08-11 14:17:39 -07001028 global_dirty_limits(&background_thresh, &dirty_thresh);
1029
1030 /*
1031 * Throttle it only when the background writeback cannot
1032 * catch-up. This avoids (excessively) small writeouts
1033 * when the bdi limits are ramping up.
1034 */
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001035 freerun = dirty_freerun_ceiling(dirty_thresh,
1036 background_thresh);
1037 if (nr_dirty <= freerun)
Wu Fengguang16c40422010-08-11 14:17:39 -07001038 break;
1039
Wu Fengguang143dfe82010-08-27 18:45:12 -06001040 if (unlikely(!writeback_in_progress(bdi)))
1041 bdi_start_background_writeback(bdi);
1042
1043 /*
1044 * bdi_thresh is not treated as some limiting factor as
1045 * dirty_thresh, due to reasons
1046 * - in JBOD setup, bdi_thresh can fluctuate a lot
1047 * - in a system with HDD and USB key, the USB key may somehow
1048 * go into state (bdi_dirty >> bdi_thresh) either because
1049 * bdi_dirty starts high, or because bdi_thresh drops low.
1050 * In this case we don't want to hard throttle the USB key
1051 * dirtiers for 100 seconds until bdi_dirty drops under
1052 * bdi_thresh. Instead the auxiliary bdi control line in
1053 * bdi_position_ratio() will let the dirtier task progress
1054 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1055 */
Wu Fengguang16c40422010-08-11 14:17:39 -07001056 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001057
Wu Fengguange50e3722010-08-11 14:17:37 -07001058 /*
1059 * In order to avoid the stacked BDI deadlock we need
1060 * to ensure we accurately count the 'dirty' pages when
1061 * the threshold is low.
1062 *
1063 * Otherwise it would be possible to get thresh+n pages
1064 * reported dirty, even though there are thresh-m pages
1065 * actually dirty; with m+n sitting in the percpu
1066 * deltas.
1067 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001068 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1069 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1070 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001071 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001072 } else {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001073 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1074 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001075 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001076 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001077
Wu Fengguang143dfe82010-08-27 18:45:12 -06001078 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
Wu Fengguang77627412010-09-12 13:34:05 -06001079 (nr_dirty > dirty_thresh);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001080 if (dirty_exceeded && !bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001081 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001083 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1084 nr_dirty, bdi_thresh, bdi_dirty,
1085 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001086
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001087 max_pause = bdi_max_pause(bdi, bdi_dirty);
1088
Wu Fengguang143dfe82010-08-27 18:45:12 -06001089 dirty_ratelimit = bdi->dirty_ratelimit;
1090 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1091 background_thresh, nr_dirty,
1092 bdi_thresh, bdi_dirty);
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001093 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1094 RATELIMIT_CALC_SHIFT;
1095 if (unlikely(task_ratelimit == 0)) {
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001096 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001097 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 }
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001099 pause = HZ * pages_dirtied / task_ratelimit;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001100 if (unlikely(pause <= 0)) {
Wu Fengguangece13ac2010-08-29 23:33:20 -06001101 trace_balance_dirty_pages(bdi,
1102 dirty_thresh,
1103 background_thresh,
1104 nr_dirty,
1105 bdi_thresh,
1106 bdi_dirty,
1107 dirty_ratelimit,
1108 task_ratelimit,
1109 pages_dirtied,
1110 pause,
1111 start_time);
Wu Fengguang57fc9782011-06-11 19:32:32 -06001112 pause = 1; /* avoid resetting nr_dirtied_pause below */
1113 break;
1114 }
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001115 pause = min(pause, max_pause);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001116
1117pause:
Wu Fengguangece13ac2010-08-29 23:33:20 -06001118 trace_balance_dirty_pages(bdi,
1119 dirty_thresh,
1120 background_thresh,
1121 nr_dirty,
1122 bdi_thresh,
1123 bdi_dirty,
1124 dirty_ratelimit,
1125 task_ratelimit,
1126 pages_dirtied,
1127 pause,
1128 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001129 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001130 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001131
Wu Fengguangffd1f602011-06-19 22:18:42 -06001132 /*
Wu Fengguang1df64712011-11-13 19:47:32 -06001133 * This is typically equal to (nr_dirty < dirty_thresh) and can
1134 * also keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001135 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001136 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001137 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001138
1139 if (fatal_signal_pending(current))
1140 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 }
1142
Wu Fengguang143dfe82010-08-27 18:45:12 -06001143 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001144 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145
Wu Fengguang9d823e82011-06-11 18:10:12 -06001146 current->nr_dirtied = 0;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001147 if (pause == 0) { /* in freerun area */
1148 current->nr_dirtied_pause =
1149 dirty_poll_interval(nr_dirty, dirty_thresh);
1150 } else if (pause <= max_pause / 4 &&
1151 pages_dirtied >= current->nr_dirtied_pause) {
1152 current->nr_dirtied_pause = clamp_val(
1153 dirty_ratelimit * (max_pause / 2) / HZ,
1154 pages_dirtied + pages_dirtied / 8,
1155 pages_dirtied * 4);
1156 } else if (pause >= max_pause) {
1157 current->nr_dirtied_pause = 1 | clamp_val(
1158 dirty_ratelimit * (max_pause / 2) / HZ,
1159 pages_dirtied / 4,
1160 pages_dirtied - pages_dirtied / 8);
1161 }
Wu Fengguang9d823e82011-06-11 18:10:12 -06001162
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001164 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165
1166 /*
1167 * In laptop mode, we wait until hitting the higher threshold before
1168 * starting background writeout, and then write out all the way down
1169 * to the lower threshold. So slow writers cause minimal disk activity.
1170 *
1171 * In normal mode, we start background writeout at the lower
1172 * background_thresh, to keep the amount of dirty memory low.
1173 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001174 if (laptop_mode)
1175 return;
1176
1177 if (nr_reclaimable > background_thresh)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001178 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179}
1180
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001181void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001182{
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001183 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001184 struct address_space *mapping = page_mapping(page);
1185
1186 if (mapping)
1187 balance_dirty_pages_ratelimited(mapping);
1188 }
1189}
1190
Wu Fengguang9d823e82011-06-11 18:10:12 -06001191static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001192
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001194 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001195 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +02001196 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 *
1198 * Processes which are dirtying memory should call in here once for each page
1199 * which was newly dirtied. The function will periodically check the system's
1200 * dirty state and will initiate writeback if needed.
1201 *
1202 * On really big machines, get_writeback_state is expensive, so try to avoid
1203 * calling it too often (ratelimiting). But once we're over the dirty memory
1204 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1205 * from overshooting the limit by (ratelimit_pages) each.
1206 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001207void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1208 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001210 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001211 int ratelimit;
1212 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213
Wu Fengguang36715ce2011-06-11 17:53:57 -06001214 if (!bdi_cap_account_dirty(bdi))
1215 return;
1216
Wu Fengguang9d823e82011-06-11 18:10:12 -06001217 ratelimit = current->nr_dirtied_pause;
1218 if (bdi->dirty_exceeded)
1219 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220
Wu Fengguang9d823e82011-06-11 18:10:12 -06001221 current->nr_dirtied += nr_pages_dirtied;
1222
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001223 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001224 /*
1225 * This prevents one CPU to accumulate too many dirtied pages without
1226 * calling into balance_dirty_pages(), which can happen when there are
1227 * 1000+ tasks, all of them start dirtying pages at exactly the same
1228 * time, hence all honoured too large initial task->nr_dirtied_pause.
1229 */
Tejun Heo245b2e72009-06-24 15:13:48 +09001230 p = &__get_cpu_var(bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001231 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001232 *p = 0;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001233 else {
1234 *p += nr_pages_dirtied;
1235 if (unlikely(*p >= ratelimit_pages)) {
1236 *p = 0;
1237 ratelimit = 0;
1238 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001240 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001241
1242 if (unlikely(current->nr_dirtied >= ratelimit))
1243 balance_dirty_pages(mapping, current->nr_dirtied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244}
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001245EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246
Andrew Morton232ea4d2007-02-28 20:13:21 -08001247void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248{
David Rientjes364aeb22009-01-06 14:39:29 -08001249 unsigned long background_thresh;
1250 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251
1252 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001253 global_dirty_limits(&background_thresh, &dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254
1255 /*
1256 * Boost the allowable dirty threshold a bit for page
1257 * allocators so they don't get DoS'ed by heavy writers
1258 */
1259 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1260
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001261 if (global_page_state(NR_UNSTABLE_NFS) +
1262 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1263 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001264 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001265
1266 /*
1267 * The caller might hold locks which can prevent IO completion
1268 * or progress in the filesystem. So we cannot just sit here
1269 * waiting for IO to complete.
1270 */
1271 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1272 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 }
1274}
1275
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1278 */
1279int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001280 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001282 proc_dointvec(table, write, buffer, length, ppos);
Jens Axboe64231042010-05-21 20:00:35 +02001283 bdi_arm_supers_timer();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 return 0;
1285}
1286
Jens Axboec2c49862010-05-20 09:18:47 +02001287#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001288void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289{
Matthew Garrett31373d02010-04-06 14:25:14 +02001290 struct request_queue *q = (struct request_queue *)data;
1291 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1292 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293
Matthew Garrett31373d02010-04-06 14:25:14 +02001294 /*
1295 * We want to write everything out, not just down to the dirty
1296 * threshold
1297 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001298 if (bdi_has_dirty_io(&q->backing_dev_info))
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001299 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1300 WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301}
1302
1303/*
1304 * We've spun up the disk and we're in laptop mode: schedule writeback
1305 * of all dirty data a few seconds from now. If the flush is already scheduled
1306 * then push it back - the user is still using the disk.
1307 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001308void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309{
Matthew Garrett31373d02010-04-06 14:25:14 +02001310 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311}
1312
1313/*
1314 * We're in laptop mode and we've just synced. The sync's writes will have
1315 * caused another writeback to be scheduled by laptop_io_completion.
1316 * Nothing needs to be written back anymore, so we unschedule the writeback.
1317 */
1318void laptop_sync_completion(void)
1319{
Matthew Garrett31373d02010-04-06 14:25:14 +02001320 struct backing_dev_info *bdi;
1321
1322 rcu_read_lock();
1323
1324 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1325 del_timer(&bdi->laptop_mode_wb_timer);
1326
1327 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328}
Jens Axboec2c49862010-05-20 09:18:47 +02001329#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330
1331/*
1332 * If ratelimit_pages is too high then we can get into dirty-data overload
1333 * if a large number of processes all perform writes at the same time.
1334 * If it is too low then SMP machines will call the (expensive)
1335 * get_writeback_state too often.
1336 *
1337 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1338 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06001339 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 */
1341
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001342void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343{
Wu Fengguang9d823e82011-06-11 18:10:12 -06001344 unsigned long background_thresh;
1345 unsigned long dirty_thresh;
1346 global_dirty_limits(&background_thresh, &dirty_thresh);
1347 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 if (ratelimit_pages < 16)
1349 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350}
1351
Chandra Seetharaman26c21432006-06-27 02:54:10 -07001352static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1354{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001355 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -08001356 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357}
1358
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001359static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 .notifier_call = ratelimit_handler,
1361 .next = NULL,
1362};
1363
1364/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001365 * Called early on to tune the page writeback dirty limits.
1366 *
1367 * We used to scale dirty pages according to how total memory
1368 * related to pages that could be allocated for buffers (by
1369 * comparing nr_free_buffer_pages() to vm_total_pages.
1370 *
1371 * However, that was when we used "dirty_ratio" to scale with
1372 * all memory, and we don't do that any more. "dirty_ratio"
1373 * is now applied to total non-HIGHPAGE memory (by subtracting
1374 * totalhigh_pages from vm_total_pages), and as such we can't
1375 * get into the old insane situation any more where we had
1376 * large amounts of dirty pages compared to a small amount of
1377 * non-HIGHMEM memory.
1378 *
1379 * But we might still want to scale the dirty_ratio by how
1380 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 */
1382void __init page_writeback_init(void)
1383{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001384 int shift;
1385
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001386 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001388
1389 shift = calc_period_shift();
1390 prop_descriptor_init(&vm_completions, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391}
1392
David Howells811d7362006-08-29 19:06:09 +01001393/**
Jan Karaf446daae2010-08-09 17:19:12 -07001394 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1395 * @mapping: address space structure to write
1396 * @start: starting page index
1397 * @end: ending page index (inclusive)
1398 *
1399 * This function scans the page range from @start to @end (inclusive) and tags
1400 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1401 * that write_cache_pages (or whoever calls this function) will then use
1402 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1403 * used to avoid livelocking of writeback by a process steadily creating new
1404 * dirty pages in the file (thus it is important for this function to be quick
1405 * so that it can tag pages faster than a dirtying process can create them).
1406 */
1407/*
1408 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1409 */
Jan Karaf446daae2010-08-09 17:19:12 -07001410void tag_pages_for_writeback(struct address_space *mapping,
1411 pgoff_t start, pgoff_t end)
1412{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001413#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07001414 unsigned long tagged;
1415
1416 do {
1417 spin_lock_irq(&mapping->tree_lock);
1418 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1419 &start, end, WRITEBACK_TAG_BATCH,
1420 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1421 spin_unlock_irq(&mapping->tree_lock);
1422 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1423 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001424 /* We check 'start' to handle wrapping when end == ~0UL */
1425 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07001426}
1427EXPORT_SYMBOL(tag_pages_for_writeback);
1428
1429/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001430 * 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 +01001431 * @mapping: address space structure to write
1432 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001433 * @writepage: function called for each page
1434 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001435 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001436 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001437 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1438 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1439 * and msync() need to guarantee that all the data which was dirty at the time
1440 * the call was made get new I/O started against them. If wbc->sync_mode is
1441 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1442 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07001443 *
1444 * To avoid livelocks (when other process dirties new pages), we first tag
1445 * pages which should be written back with TOWRITE tag and only then start
1446 * writing them. For data-integrity sync we have to be careful so that we do
1447 * not miss some pages (e.g., because some other process has cleared TOWRITE
1448 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1449 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001450 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001451int write_cache_pages(struct address_space *mapping,
1452 struct writeback_control *wbc, writepage_t writepage,
1453 void *data)
David Howells811d7362006-08-29 19:06:09 +01001454{
David Howells811d7362006-08-29 19:06:09 +01001455 int ret = 0;
1456 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001457 struct pagevec pvec;
1458 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001459 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001460 pgoff_t index;
1461 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001462 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001463 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001464 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07001465 int tag;
David Howells811d7362006-08-29 19:06:09 +01001466
David Howells811d7362006-08-29 19:06:09 +01001467 pagevec_init(&pvec, 0);
1468 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001469 writeback_index = mapping->writeback_index; /* prev offset */
1470 index = writeback_index;
1471 if (index == 0)
1472 cycled = 1;
1473 else
1474 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001475 end = -1;
1476 } else {
1477 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1478 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1479 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1480 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001481 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001482 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001483 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001484 tag = PAGECACHE_TAG_TOWRITE;
1485 else
1486 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001487retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001488 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001489 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001490 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001491 while (!done && (index <= end)) {
1492 int i;
1493
Jan Karaf446daae2010-08-09 17:19:12 -07001494 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001495 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1496 if (nr_pages == 0)
1497 break;
David Howells811d7362006-08-29 19:06:09 +01001498
David Howells811d7362006-08-29 19:06:09 +01001499 for (i = 0; i < nr_pages; i++) {
1500 struct page *page = pvec.pages[i];
1501
Nick Piggind5482cd2009-01-06 14:39:11 -08001502 /*
1503 * At this point, the page may be truncated or
1504 * invalidated (changing page->mapping to NULL), or
1505 * even swizzled back from swapper_space to tmpfs file
1506 * mapping. However, page->index will not change
1507 * because we have a reference on the page.
1508 */
1509 if (page->index > end) {
1510 /*
1511 * can't be range_cyclic (1st pass) because
1512 * end == -1 in that case.
1513 */
1514 done = 1;
1515 break;
1516 }
1517
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001518 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001519
David Howells811d7362006-08-29 19:06:09 +01001520 lock_page(page);
1521
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001522 /*
1523 * Page truncated or invalidated. We can freely skip it
1524 * then, even for data integrity operations: the page
1525 * has disappeared concurrently, so there could be no
1526 * real expectation of this data interity operation
1527 * even if there is now a new, dirty page at the same
1528 * pagecache address.
1529 */
David Howells811d7362006-08-29 19:06:09 +01001530 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001531continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001532 unlock_page(page);
1533 continue;
1534 }
1535
Nick Piggin515f4a02009-01-06 14:39:10 -08001536 if (!PageDirty(page)) {
1537 /* someone wrote it for us */
1538 goto continue_unlock;
1539 }
David Howells811d7362006-08-29 19:06:09 +01001540
Nick Piggin515f4a02009-01-06 14:39:10 -08001541 if (PageWriteback(page)) {
1542 if (wbc->sync_mode != WB_SYNC_NONE)
1543 wait_on_page_writeback(page);
1544 else
1545 goto continue_unlock;
1546 }
1547
1548 BUG_ON(PageWriteback(page));
1549 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001550 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001551
Dave Chinner9e094382010-07-07 13:24:08 +10001552 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001553 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001554 if (unlikely(ret)) {
1555 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1556 unlock_page(page);
1557 ret = 0;
1558 } else {
1559 /*
1560 * done_index is set past this page,
1561 * so media errors will not choke
1562 * background writeout for the entire
1563 * file. This has consequences for
1564 * range_cyclic semantics (ie. it may
1565 * not be suitable for data integrity
1566 * writeout).
1567 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001568 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001569 done = 1;
1570 break;
1571 }
Dave Chinner0b564922010-06-09 10:37:18 +10001572 }
David Howells811d7362006-08-29 19:06:09 +01001573
Dave Chinner546a1922010-08-24 11:44:34 +10001574 /*
1575 * We stop writing back only if we are not doing
1576 * integrity sync. In case of integrity sync we have to
1577 * keep going until we have written all the pages
1578 * we tagged for writeback prior to entering this loop.
1579 */
1580 if (--wbc->nr_to_write <= 0 &&
1581 wbc->sync_mode == WB_SYNC_NONE) {
1582 done = 1;
1583 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001584 }
David Howells811d7362006-08-29 19:06:09 +01001585 }
1586 pagevec_release(&pvec);
1587 cond_resched();
1588 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001589 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001590 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001591 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001592 * We hit the last page and there is more work to be done: wrap
1593 * back to the start of the file
1594 */
Nick Piggin31a12662009-01-06 14:39:04 -08001595 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001596 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001597 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001598 goto retry;
1599 }
Dave Chinner0b564922010-06-09 10:37:18 +10001600 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1601 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001602
David Howells811d7362006-08-29 19:06:09 +01001603 return ret;
1604}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001605EXPORT_SYMBOL(write_cache_pages);
1606
1607/*
1608 * Function used by generic_writepages to call the real writepage
1609 * function and set the mapping flags on error
1610 */
1611static int __writepage(struct page *page, struct writeback_control *wbc,
1612 void *data)
1613{
1614 struct address_space *mapping = data;
1615 int ret = mapping->a_ops->writepage(page, wbc);
1616 mapping_set_error(mapping, ret);
1617 return ret;
1618}
1619
1620/**
1621 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1622 * @mapping: address space structure to write
1623 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1624 *
1625 * This is a library function, which implements the writepages()
1626 * address_space_operation.
1627 */
1628int generic_writepages(struct address_space *mapping,
1629 struct writeback_control *wbc)
1630{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001631 struct blk_plug plug;
1632 int ret;
1633
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001634 /* deal with chardevs and other special file */
1635 if (!mapping->a_ops->writepage)
1636 return 0;
1637
Shaohua Li9b6096a2011-03-17 10:47:06 +01001638 blk_start_plug(&plug);
1639 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1640 blk_finish_plug(&plug);
1641 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001642}
David Howells811d7362006-08-29 19:06:09 +01001643
1644EXPORT_SYMBOL(generic_writepages);
1645
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1647{
Andrew Morton22905f72005-11-16 15:07:01 -08001648 int ret;
1649
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 if (wbc->nr_to_write <= 0)
1651 return 0;
1652 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001653 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001654 else
1655 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001656 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657}
1658
1659/**
1660 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001661 * @page: the page to write
1662 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 *
1664 * The page must be locked by the caller and will be unlocked upon return.
1665 *
1666 * write_one_page() returns a negative error code if I/O failed.
1667 */
1668int write_one_page(struct page *page, int wait)
1669{
1670 struct address_space *mapping = page->mapping;
1671 int ret = 0;
1672 struct writeback_control wbc = {
1673 .sync_mode = WB_SYNC_ALL,
1674 .nr_to_write = 1,
1675 };
1676
1677 BUG_ON(!PageLocked(page));
1678
1679 if (wait)
1680 wait_on_page_writeback(page);
1681
1682 if (clear_page_dirty_for_io(page)) {
1683 page_cache_get(page);
1684 ret = mapping->a_ops->writepage(page, &wbc);
1685 if (ret == 0 && wait) {
1686 wait_on_page_writeback(page);
1687 if (PageError(page))
1688 ret = -EIO;
1689 }
1690 page_cache_release(page);
1691 } else {
1692 unlock_page(page);
1693 }
1694 return ret;
1695}
1696EXPORT_SYMBOL(write_one_page);
1697
1698/*
Ken Chen76719322007-02-10 01:43:15 -08001699 * For address_spaces which do not use buffers nor write back.
1700 */
1701int __set_page_dirty_no_writeback(struct page *page)
1702{
1703 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001704 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001705 return 0;
1706}
1707
1708/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001709 * Helper function for set_page_dirty family.
1710 * NOTE: This relies on being atomic wrt interrupts.
1711 */
1712void account_page_dirtied(struct page *page, struct address_space *mapping)
1713{
1714 if (mapping_cap_account_dirty(mapping)) {
1715 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001716 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001717 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06001718 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001719 task_io_account_write(PAGE_CACHE_SIZE);
1720 }
1721}
Michael Rubin679ceac2010-08-20 02:31:26 -07001722EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001723
1724/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001725 * Helper function for set_page_writeback family.
1726 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1727 * wrt interrupts.
1728 */
1729void account_page_writeback(struct page *page)
1730{
1731 inc_zone_page_state(page, NR_WRITEBACK);
1732}
1733EXPORT_SYMBOL(account_page_writeback);
1734
1735/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 * For address_spaces which do not use buffers. Just tag the page as dirty in
1737 * its radix tree.
1738 *
1739 * This is also used when a single buffer is being dirtied: we want to set the
1740 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1741 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1742 *
1743 * Most callers have locked the page, which pins the address_space in memory.
1744 * But zap_pte_range() does not lock the page, however in that case the
1745 * mapping is pinned by the vma's ->vm_file reference.
1746 *
1747 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001748 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 */
1750int __set_page_dirty_nobuffers(struct page *page)
1751{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 if (!TestSetPageDirty(page)) {
1753 struct address_space *mapping = page_mapping(page);
1754 struct address_space *mapping2;
1755
Andrew Morton8c085402006-12-10 02:19:24 -08001756 if (!mapping)
1757 return 1;
1758
Nick Piggin19fd6232008-07-25 19:45:32 -07001759 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001760 mapping2 = page_mapping(page);
1761 if (mapping2) { /* Race with truncate? */
1762 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001763 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001764 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001765 radix_tree_tag_set(&mapping->page_tree,
1766 page_index(page), PAGECACHE_TAG_DIRTY);
1767 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001768 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001769 if (mapping->host) {
1770 /* !PageAnon && !swapper_space */
1771 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001773 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001775 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776}
1777EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1778
1779/*
1780 * When a writepage implementation decides that it doesn't want to write this
1781 * page for some reason, it should redirty the locked page via
1782 * redirty_page_for_writepage() and it should then unlock the page and return 0
1783 */
1784int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1785{
1786 wbc->pages_skipped++;
1787 return __set_page_dirty_nobuffers(page);
1788}
1789EXPORT_SYMBOL(redirty_page_for_writepage);
1790
1791/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02001792 * Dirty a page.
1793 *
1794 * For pages with a mapping this should be done under the page lock
1795 * for the benefit of asynchronous memory errors who prefer a consistent
1796 * dirty state. This rule can be broken in some special cases,
1797 * but should be better not to.
1798 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 * If the mapping doesn't provide a set_page_dirty a_op, then
1800 * just fall through and assume that it wants buffer_heads.
1801 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001802int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803{
1804 struct address_space *mapping = page_mapping(page);
1805
1806 if (likely(mapping)) {
1807 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07001808 /*
1809 * readahead/lru_deactivate_page could remain
1810 * PG_readahead/PG_reclaim due to race with end_page_writeback
1811 * About readahead, if the page is written, the flags would be
1812 * reset. So no problem.
1813 * About lru_deactivate_page, if the page is redirty, the flag
1814 * will be reset. So no problem. but if the page is used by readahead
1815 * it will confuse readahead and make it restart the size rampup
1816 * process. But it's a trivial problem.
1817 */
1818 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02001819#ifdef CONFIG_BLOCK
1820 if (!spd)
1821 spd = __set_page_dirty_buffers;
1822#endif
1823 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001825 if (!PageDirty(page)) {
1826 if (!TestSetPageDirty(page))
1827 return 1;
1828 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 return 0;
1830}
1831EXPORT_SYMBOL(set_page_dirty);
1832
1833/*
1834 * set_page_dirty() is racy if the caller has no reference against
1835 * page->mapping->host, and if the page is unlocked. This is because another
1836 * CPU could truncate the page off the mapping and then free the mapping.
1837 *
1838 * Usually, the page _is_ locked, or the caller is a user-space process which
1839 * holds a reference on the inode by having an open file.
1840 *
1841 * In other cases, the page should be locked before running set_page_dirty().
1842 */
1843int set_page_dirty_lock(struct page *page)
1844{
1845 int ret;
1846
Jens Axboe7eaceac2011-03-10 08:52:07 +01001847 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848 ret = set_page_dirty(page);
1849 unlock_page(page);
1850 return ret;
1851}
1852EXPORT_SYMBOL(set_page_dirty_lock);
1853
1854/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 * Clear a page's dirty flag, while caring for dirty memory accounting.
1856 * Returns true if the page was previously dirty.
1857 *
1858 * This is for preparing to put the page under writeout. We leave the page
1859 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1860 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1861 * implementation will run either set_page_writeback() or set_page_dirty(),
1862 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1863 * back into sync.
1864 *
1865 * This incoherency between the page's dirty flag and radix-tree tag is
1866 * unfortunate, but it only exists while the page is locked.
1867 */
1868int clear_page_dirty_for_io(struct page *page)
1869{
1870 struct address_space *mapping = page_mapping(page);
1871
Nick Piggin79352892007-07-19 01:47:22 -07001872 BUG_ON(!PageLocked(page));
1873
Linus Torvalds7658cc22006-12-29 10:00:58 -08001874 if (mapping && mapping_cap_account_dirty(mapping)) {
1875 /*
1876 * Yes, Virginia, this is indeed insane.
1877 *
1878 * We use this sequence to make sure that
1879 * (a) we account for dirty stats properly
1880 * (b) we tell the low-level filesystem to
1881 * mark the whole page dirty if it was
1882 * dirty in a pagetable. Only to then
1883 * (c) clean the page again and return 1 to
1884 * cause the writeback.
1885 *
1886 * This way we avoid all nasty races with the
1887 * dirty bit in multiple places and clearing
1888 * them concurrently from different threads.
1889 *
1890 * Note! Normally the "set_page_dirty(page)"
1891 * has no effect on the actual dirty bit - since
1892 * that will already usually be set. But we
1893 * need the side effects, and it can help us
1894 * avoid races.
1895 *
1896 * We basically use the page "master dirty bit"
1897 * as a serialization point for all the different
1898 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001899 */
1900 if (page_mkclean(page))
1901 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001902 /*
1903 * We carefully synchronise fault handlers against
1904 * installing a dirty pte and marking the page dirty
1905 * at this point. We do this by having them hold the
1906 * page lock at some point after installing their
1907 * pte, but before marking the page dirty.
1908 * Pages are always locked coming in here, so we get
1909 * the desired exclusion. See mm/memory.c:do_wp_page()
1910 * for more comments.
1911 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001912 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001913 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001914 dec_bdi_stat(mapping->backing_dev_info,
1915 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001916 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001918 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001920 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001922EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923
1924int test_clear_page_writeback(struct page *page)
1925{
1926 struct address_space *mapping = page_mapping(page);
1927 int ret;
1928
1929 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001930 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 unsigned long flags;
1932
Nick Piggin19fd6232008-07-25 19:45:32 -07001933 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001935 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 radix_tree_tag_clear(&mapping->page_tree,
1937 page_index(page),
1938 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001939 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001940 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001941 __bdi_writeout_inc(bdi);
1942 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001943 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001944 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945 } else {
1946 ret = TestClearPageWriteback(page);
1947 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07001948 if (ret) {
Andrew Mortond688abf2007-07-19 01:49:17 -07001949 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07001950 inc_zone_page_state(page, NR_WRITTEN);
1951 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 return ret;
1953}
1954
1955int test_set_page_writeback(struct page *page)
1956{
1957 struct address_space *mapping = page_mapping(page);
1958 int ret;
1959
1960 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001961 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 unsigned long flags;
1963
Nick Piggin19fd6232008-07-25 19:45:32 -07001964 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001966 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967 radix_tree_tag_set(&mapping->page_tree,
1968 page_index(page),
1969 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001970 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001971 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1972 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 if (!PageDirty(page))
1974 radix_tree_tag_clear(&mapping->page_tree,
1975 page_index(page),
1976 PAGECACHE_TAG_DIRTY);
Jan Karaf446daae2010-08-09 17:19:12 -07001977 radix_tree_tag_clear(&mapping->page_tree,
1978 page_index(page),
1979 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07001980 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 } else {
1982 ret = TestSetPageWriteback(page);
1983 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001984 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07001985 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 return ret;
1987
1988}
1989EXPORT_SYMBOL(test_set_page_writeback);
1990
1991/*
Nick Piggin00128182007-10-16 01:24:40 -07001992 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 * passed tag.
1994 */
1995int mapping_tagged(struct address_space *mapping, int tag)
1996{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07001997 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998}
1999EXPORT_SYMBOL(mapping_tagged);