blob: f32f25092c66a977b68f57bc4546e587614d43bb [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>
15#include <linux/module.h>
16#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;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700131static struct prop_descriptor vm_dirties;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700132
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700133/*
134 * couple the period to the dirty_ratio:
135 *
136 * period/2 ~ roundup_pow_of_two(dirty limit)
137 */
138static int calc_period_shift(void)
139{
140 unsigned long dirty_total;
141
David Rientjes2da02992009-01-06 14:39:31 -0800142 if (vm_dirty_bytes)
143 dirty_total = vm_dirty_bytes / PAGE_SIZE;
144 else
145 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
146 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700147 return 2 + ilog2(dirty_total - 1);
148}
149
150/*
David Rientjes2da02992009-01-06 14:39:31 -0800151 * update the period when the dirty threshold changes.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700152 */
David Rientjes2da02992009-01-06 14:39:31 -0800153static void update_completion_period(void)
154{
155 int shift = calc_period_shift();
156 prop_change_shift(&vm_completions, shift);
157 prop_change_shift(&vm_dirties, shift);
Wu Fengguang9d823e82011-06-11 18:10:12 -0600158
159 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800160}
161
162int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700163 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800164 loff_t *ppos)
165{
166 int ret;
167
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700168 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800169 if (ret == 0 && write)
170 dirty_background_bytes = 0;
171 return ret;
172}
173
174int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700175 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800176 loff_t *ppos)
177{
178 int ret;
179
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700180 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800181 if (ret == 0 && write)
182 dirty_background_ratio = 0;
183 return ret;
184}
185
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700186int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700187 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700188 loff_t *ppos)
189{
190 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800191 int ret;
192
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700193 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700194 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
David Rientjes2da02992009-01-06 14:39:31 -0800195 update_completion_period();
196 vm_dirty_bytes = 0;
197 }
198 return ret;
199}
200
201
202int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700203 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800204 loff_t *ppos)
205{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800206 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800207 int ret;
208
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700209 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800210 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
211 update_completion_period();
212 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700213 }
214 return ret;
215}
216
217/*
218 * Increment the BDI's writeout completion count and the global writeout
219 * completion count. Called from test_clear_page_writeback().
220 */
221static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
222{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600223 __inc_bdi_stat(bdi, BDI_WRITTEN);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700224 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
225 bdi->max_prop_frac);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700226}
227
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700228void bdi_writeout_inc(struct backing_dev_info *bdi)
229{
230 unsigned long flags;
231
232 local_irq_save(flags);
233 __bdi_writeout_inc(bdi);
234 local_irq_restore(flags);
235}
236EXPORT_SYMBOL_GPL(bdi_writeout_inc);
237
Nick Piggin1cf6e7d2009-02-18 14:48:18 -0800238void task_dirty_inc(struct task_struct *tsk)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700239{
240 prop_inc_single(&vm_dirties, &tsk->dirties);
241}
242
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700243/*
244 * Obtain an accurate fraction of the BDI's portion.
245 */
246static void bdi_writeout_fraction(struct backing_dev_info *bdi,
247 long *numerator, long *denominator)
248{
Wu Fengguang3efaf0f2010-12-16 22:22:00 -0600249 prop_fraction_percpu(&vm_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700250 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700251}
252
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700253/*
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700254 *
255 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700256static unsigned int bdi_min_ratio;
257
258int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
259{
260 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700261
Jens Axboecfc4ba52009-09-14 13:12:40 +0200262 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700263 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700264 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700265 } else {
266 min_ratio -= bdi->min_ratio;
267 if (bdi_min_ratio + min_ratio < 100) {
268 bdi_min_ratio += min_ratio;
269 bdi->min_ratio += min_ratio;
270 } else {
271 ret = -EINVAL;
272 }
273 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200274 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700275
276 return ret;
277}
278
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700279int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
280{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700281 int ret = 0;
282
283 if (max_ratio > 100)
284 return -EINVAL;
285
Jens Axboecfc4ba52009-09-14 13:12:40 +0200286 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700287 if (bdi->min_ratio > max_ratio) {
288 ret = -EINVAL;
289 } else {
290 bdi->max_ratio = max_ratio;
291 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
292 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200293 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700294
295 return ret;
296}
297EXPORT_SYMBOL(bdi_set_max_ratio);
298
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700299/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 * Work out the current dirty-memory clamping and background writeout
301 * thresholds.
302 *
303 * The main aim here is to lower them aggressively if there is a lot of mapped
304 * memory around. To avoid stressing page reclaim with lots of unreclaimable
305 * pages. It is better to clamp down on writers than to start swapping, and
306 * performing lots of scanning.
307 *
308 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
309 *
310 * We don't permit the clamping level to fall below 5% - that is getting rather
311 * excessive.
312 *
313 * We make sure that the background writeout level is below the adjusted
314 * clamping level.
315 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700316
317static unsigned long highmem_dirtyable_memory(unsigned long total)
318{
319#ifdef CONFIG_HIGHMEM
320 int node;
321 unsigned long x = 0;
322
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700323 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700324 struct zone *z =
325 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
326
Wu Fengguangadea02a2009-09-21 17:01:42 -0700327 x += zone_page_state(z, NR_FREE_PAGES) +
328 zone_reclaimable_pages(z);
Christoph Lameter1b424462007-05-06 14:48:59 -0700329 }
330 /*
331 * Make sure that the number of highmem pages is never larger
332 * than the number of the total dirtyable memory. This can only
333 * occur in very strange VM situations but we want to make sure
334 * that this does not occur.
335 */
336 return min(x, total);
337#else
338 return 0;
339#endif
340}
341
Steven Rostedt3eefae92008-05-12 21:21:04 +0200342/**
343 * determine_dirtyable_memory - amount of memory that may be used
344 *
345 * Returns the numebr of pages that can currently be freed and used
346 * by the kernel for direct mappings.
347 */
348unsigned long determine_dirtyable_memory(void)
Christoph Lameter1b424462007-05-06 14:48:59 -0700349{
350 unsigned long x;
351
Wu Fengguangadea02a2009-09-21 17:01:42 -0700352 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
Bron Gondwana195cf452008-02-04 22:29:20 -0800353
354 if (!vm_highmem_is_dirtyable)
355 x -= highmem_dirtyable_memory(x);
356
Christoph Lameter1b424462007-05-06 14:48:59 -0700357 return x + 1; /* Ensure that we never return 0 */
358}
359
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600360static unsigned long dirty_freerun_ceiling(unsigned long thresh,
361 unsigned long bg_thresh)
362{
363 return (thresh + bg_thresh) / 2;
364}
365
Wu Fengguangffd1f602011-06-19 22:18:42 -0600366static unsigned long hard_dirty_limit(unsigned long thresh)
367{
368 return max(thresh, global_dirty_limit);
369}
370
Randy Dunlap03ab4502010-08-14 13:05:17 -0700371/*
Wu Fengguang1babe182010-08-11 14:17:40 -0700372 * global_dirty_limits - background-writeback and dirty-throttling thresholds
373 *
374 * Calculate the dirty thresholds based on sysctl parameters
375 * - vm.dirty_background_ratio or vm.dirty_background_bytes
376 * - vm.dirty_ratio or vm.dirty_bytes
377 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Minchan Kimebd13732011-01-04 01:36:48 +0900378 * real-time tasks.
Wu Fengguang1babe182010-08-11 14:17:40 -0700379 */
Wu Fengguang16c40422010-08-11 14:17:39 -0700380void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381{
David Rientjes364aeb22009-01-06 14:39:29 -0800382 unsigned long background;
383 unsigned long dirty;
Minchan Kim240c8792011-01-13 15:46:27 -0800384 unsigned long uninitialized_var(available_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 struct task_struct *tsk;
386
Minchan Kim240c8792011-01-13 15:46:27 -0800387 if (!vm_dirty_bytes || !dirty_background_bytes)
388 available_memory = determine_dirtyable_memory();
389
David Rientjes2da02992009-01-06 14:39:31 -0800390 if (vm_dirty_bytes)
391 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
Wu Fengguang4cbec4c2010-10-26 14:21:45 -0700392 else
393 dirty = (vm_dirty_ratio * available_memory) / 100;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
David Rientjes2da02992009-01-06 14:39:31 -0800395 if (dirty_background_bytes)
396 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
397 else
398 background = (dirty_background_ratio * available_memory) / 100;
399
400 if (background >= dirty)
401 background = dirty / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 tsk = current;
403 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
404 background += background / 4;
405 dirty += dirty / 4;
406 }
407 *pbackground = background;
408 *pdirty = dirty;
Wu Fengguange1cbe232010-12-06 22:34:29 -0600409 trace_global_dirty_state(background, dirty);
Wu Fengguang16c40422010-08-11 14:17:39 -0700410}
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700411
Wu Fengguang6f718652011-03-02 17:14:34 -0600412/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700413 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600414 * @bdi: the backing_dev_info to query
415 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700416 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600417 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
418 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
419 * And the "limit" in the name is not seriously taken as hard limit in
420 * balance_dirty_pages().
421 *
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;
602 /*
603 * scale global setpoint to bdi's:
604 * bdi_setpoint = setpoint * bdi_thresh / thresh
605 */
606 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
607 bdi_setpoint = setpoint * (u64)x >> 16;
608 /*
609 * Use span=(8*write_bw) in single bdi case as indicated by
610 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
611 *
612 * bdi_thresh thresh - bdi_thresh
613 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
614 * thresh thresh
615 */
616 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
617 x_intercept = bdi_setpoint + span;
618
619 if (bdi_dirty < x_intercept - span / 4) {
620 pos_ratio *= x_intercept - bdi_dirty;
621 do_div(pos_ratio, x_intercept - bdi_setpoint + 1);
622 } else
623 pos_ratio /= 4;
624
625 return pos_ratio;
626}
627
Wu Fengguange98be2d2010-08-29 11:22:30 -0600628static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
629 unsigned long elapsed,
630 unsigned long written)
631{
632 const unsigned long period = roundup_pow_of_two(3 * HZ);
633 unsigned long avg = bdi->avg_write_bandwidth;
634 unsigned long old = bdi->write_bandwidth;
635 u64 bw;
636
637 /*
638 * bw = written * HZ / elapsed
639 *
640 * bw * elapsed + write_bandwidth * (period - elapsed)
641 * write_bandwidth = ---------------------------------------------------
642 * period
643 */
644 bw = written - bdi->written_stamp;
645 bw *= HZ;
646 if (unlikely(elapsed > period)) {
647 do_div(bw, elapsed);
648 avg = bw;
649 goto out;
650 }
651 bw += (u64)bdi->write_bandwidth * (period - elapsed);
652 bw >>= ilog2(period);
653
654 /*
655 * one more level of smoothing, for filtering out sudden spikes
656 */
657 if (avg > old && old >= (unsigned long)bw)
658 avg -= (avg - old) >> 3;
659
660 if (avg < old && old <= (unsigned long)bw)
661 avg += (old - avg) >> 3;
662
663out:
664 bdi->write_bandwidth = bw;
665 bdi->avg_write_bandwidth = avg;
666}
667
Wu Fengguangc42843f2011-03-02 15:54:09 -0600668/*
669 * The global dirtyable memory and dirty threshold could be suddenly knocked
670 * down by a large amount (eg. on the startup of KVM in a swapless system).
671 * This may throw the system into deep dirty exceeded state and throttle
672 * heavy/light dirtiers alike. To retain good responsiveness, maintain
673 * global_dirty_limit for tracking slowly down to the knocked down dirty
674 * threshold.
675 */
676static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
677{
678 unsigned long limit = global_dirty_limit;
679
680 /*
681 * Follow up in one step.
682 */
683 if (limit < thresh) {
684 limit = thresh;
685 goto update;
686 }
687
688 /*
689 * Follow down slowly. Use the higher one as the target, because thresh
690 * may drop below dirty. This is exactly the reason to introduce
691 * global_dirty_limit which is guaranteed to lie above the dirty pages.
692 */
693 thresh = max(thresh, dirty);
694 if (limit > thresh) {
695 limit -= (limit - thresh) >> 5;
696 goto update;
697 }
698 return;
699update:
700 global_dirty_limit = limit;
701}
702
703static void global_update_bandwidth(unsigned long thresh,
704 unsigned long dirty,
705 unsigned long now)
706{
707 static DEFINE_SPINLOCK(dirty_lock);
708 static unsigned long update_time;
709
710 /*
711 * check locklessly first to optimize away locking for the most time
712 */
713 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
714 return;
715
716 spin_lock(&dirty_lock);
717 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
718 update_dirty_limit(thresh, dirty);
719 update_time = now;
720 }
721 spin_unlock(&dirty_lock);
722}
723
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600724/*
725 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
726 *
727 * Normal bdi tasks will be curbed at or below it in long term.
728 * Obviously it should be around (write_bw / N) when there are N dd tasks.
729 */
730static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
731 unsigned long thresh,
732 unsigned long bg_thresh,
733 unsigned long dirty,
734 unsigned long bdi_thresh,
735 unsigned long bdi_dirty,
736 unsigned long dirtied,
737 unsigned long elapsed)
738{
Wu Fengguang73811312011-08-26 15:53:24 -0600739 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
740 unsigned long limit = hard_dirty_limit(thresh);
741 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600742 unsigned long write_bw = bdi->avg_write_bandwidth;
743 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
744 unsigned long dirty_rate;
745 unsigned long task_ratelimit;
746 unsigned long balanced_dirty_ratelimit;
747 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600748 unsigned long step;
749 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600750
751 /*
752 * The dirty rate will match the writeout rate in long term, except
753 * when dirty pages are truncated by userspace or re-dirtied by FS.
754 */
755 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
756
757 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
758 bdi_thresh, bdi_dirty);
759 /*
760 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
761 */
762 task_ratelimit = (u64)dirty_ratelimit *
763 pos_ratio >> RATELIMIT_CALC_SHIFT;
764 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
765
766 /*
767 * A linear estimation of the "balanced" throttle rate. The theory is,
768 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
769 * dirty_rate will be measured to be (N * task_ratelimit). So the below
770 * formula will yield the balanced rate limit (write_bw / N).
771 *
772 * Note that the expanded form is not a pure rate feedback:
773 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
774 * but also takes pos_ratio into account:
775 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
776 *
777 * (1) is not realistic because pos_ratio also takes part in balancing
778 * the dirty rate. Consider the state
779 * pos_ratio = 0.5 (3)
780 * rate = 2 * (write_bw / N) (4)
781 * If (1) is used, it will stuck in that state! Because each dd will
782 * be throttled at
783 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
784 * yielding
785 * dirty_rate = N * task_ratelimit = write_bw (6)
786 * put (6) into (1) we get
787 * rate_(i+1) = rate_(i) (7)
788 *
789 * So we end up using (2) to always keep
790 * rate_(i+1) ~= (write_bw / N) (8)
791 * regardless of the value of pos_ratio. As long as (8) is satisfied,
792 * pos_ratio is able to drive itself to 1.0, which is not only where
793 * the dirty count meet the setpoint, but also where the slope of
794 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
795 */
796 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
797 dirty_rate | 1);
798
Wu Fengguang73811312011-08-26 15:53:24 -0600799 /*
800 * We could safely do this and return immediately:
801 *
802 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
803 *
804 * However to get a more stable dirty_ratelimit, the below elaborated
805 * code makes use of task_ratelimit to filter out sigular points and
806 * limit the step size.
807 *
808 * The below code essentially only uses the relative value of
809 *
810 * task_ratelimit - dirty_ratelimit
811 * = (pos_ratio - 1) * dirty_ratelimit
812 *
813 * which reflects the direction and size of dirty position error.
814 */
815
816 /*
817 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
818 * task_ratelimit is on the same side of dirty_ratelimit, too.
819 * For example, when
820 * - dirty_ratelimit > balanced_dirty_ratelimit
821 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
822 * lowering dirty_ratelimit will help meet both the position and rate
823 * control targets. Otherwise, don't update dirty_ratelimit if it will
824 * only help meet the rate target. After all, what the users ultimately
825 * feel and care are stable dirty rate and small position error.
826 *
827 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
828 * and filter out the sigular points of balanced_dirty_ratelimit. Which
829 * keeps jumping around randomly and can even leap far away at times
830 * due to the small 200ms estimation period of dirty_rate (we want to
831 * keep that period small to reduce time lags).
832 */
833 step = 0;
834 if (dirty < setpoint) {
835 x = min(bdi->balanced_dirty_ratelimit,
836 min(balanced_dirty_ratelimit, task_ratelimit));
837 if (dirty_ratelimit < x)
838 step = x - dirty_ratelimit;
839 } else {
840 x = max(bdi->balanced_dirty_ratelimit,
841 max(balanced_dirty_ratelimit, task_ratelimit));
842 if (dirty_ratelimit > x)
843 step = dirty_ratelimit - x;
844 }
845
846 /*
847 * Don't pursue 100% rate matching. It's impossible since the balanced
848 * rate itself is constantly fluctuating. So decrease the track speed
849 * when it gets close to the target. Helps eliminate pointless tremors.
850 */
851 step >>= dirty_ratelimit / (2 * step + 1);
852 /*
853 * Limit the tracking speed to avoid overshooting.
854 */
855 step = (step + 7) / 8;
856
857 if (dirty_ratelimit < balanced_dirty_ratelimit)
858 dirty_ratelimit += step;
859 else
860 dirty_ratelimit -= step;
861
862 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
863 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600864}
865
Wu Fengguange98be2d2010-08-29 11:22:30 -0600866void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600867 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600868 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600869 unsigned long dirty,
870 unsigned long bdi_thresh,
871 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600872 unsigned long start_time)
873{
874 unsigned long now = jiffies;
875 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600876 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600877 unsigned long written;
878
879 /*
880 * rate-limit, only update once every 200ms.
881 */
882 if (elapsed < BANDWIDTH_INTERVAL)
883 return;
884
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600885 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600886 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
887
888 /*
889 * Skip quiet periods when disk bandwidth is under-utilized.
890 * (at least 1s idle time between two flusher runs)
891 */
892 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
893 goto snapshot;
894
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600895 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -0600896 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600897 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
898 bdi_thresh, bdi_dirty,
899 dirtied, elapsed);
900 }
Wu Fengguange98be2d2010-08-29 11:22:30 -0600901 bdi_update_write_bandwidth(bdi, elapsed, written);
902
903snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600904 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600905 bdi->written_stamp = written;
906 bdi->bw_time_stamp = now;
907}
908
909static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600910 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600911 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -0600912 unsigned long dirty,
913 unsigned long bdi_thresh,
914 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -0600915 unsigned long start_time)
916{
917 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
918 return;
919 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -0600920 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
921 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -0600922 spin_unlock(&bdi->wb.list_lock);
923}
924
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925/*
Wu Fengguang9d823e82011-06-11 18:10:12 -0600926 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
927 * will look to see if it needs to start dirty throttling.
928 *
929 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
930 * global_page_state() too often. So scale it near-sqrt to the safety margin
931 * (the number of pages we may dirty without exceeding the dirty limits).
932 */
933static unsigned long dirty_poll_interval(unsigned long dirty,
934 unsigned long thresh)
935{
936 if (thresh > dirty)
937 return 1UL << (ilog2(thresh - dirty) >> 1);
938
939 return 1;
940}
941
942/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 * balance_dirty_pages() must be called by processes which are generating dirty
944 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -0600945 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +0200946 * If we're over `background_thresh' then the writeback threads are woken to
947 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800949static void balance_dirty_pages(struct address_space *mapping,
Wu Fengguang143dfe82010-08-27 18:45:12 -0600950 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951{
Wu Fengguang143dfe82010-08-27 18:45:12 -0600952 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
953 unsigned long bdi_reclaimable;
Wu Fengguang77627412010-09-12 13:34:05 -0600954 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
955 unsigned long bdi_dirty;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600956 unsigned long freerun;
David Rientjes364aeb22009-01-06 14:39:29 -0800957 unsigned long background_thresh;
958 unsigned long dirty_thresh;
959 unsigned long bdi_thresh;
Wu Fengguang143dfe82010-08-27 18:45:12 -0600960 long pause = 0;
Wu Fengguange50e3722010-08-11 14:17:37 -0700961 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -0600962 unsigned long task_ratelimit;
963 unsigned long dirty_ratelimit;
964 unsigned long pos_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguange98be2d2010-08-29 11:22:30 -0600966 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967
968 for (;;) {
Wu Fengguang143dfe82010-08-27 18:45:12 -0600969 /*
970 * Unstable writes are a feature of certain networked
971 * filesystems (i.e. NFS) in which data may have been
972 * written to the server's write cache, but has not yet
973 * been flushed to permanent storage.
974 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800975 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
976 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -0600977 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800978
Wu Fengguang16c40422010-08-11 14:17:39 -0700979 global_dirty_limits(&background_thresh, &dirty_thresh);
980
981 /*
982 * Throttle it only when the background writeback cannot
983 * catch-up. This avoids (excessively) small writeouts
984 * when the bdi limits are ramping up.
985 */
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600986 freerun = dirty_freerun_ceiling(dirty_thresh,
987 background_thresh);
988 if (nr_dirty <= freerun)
Wu Fengguang16c40422010-08-11 14:17:39 -0700989 break;
990
Wu Fengguang143dfe82010-08-27 18:45:12 -0600991 if (unlikely(!writeback_in_progress(bdi)))
992 bdi_start_background_writeback(bdi);
993
994 /*
995 * bdi_thresh is not treated as some limiting factor as
996 * dirty_thresh, due to reasons
997 * - in JBOD setup, bdi_thresh can fluctuate a lot
998 * - in a system with HDD and USB key, the USB key may somehow
999 * go into state (bdi_dirty >> bdi_thresh) either because
1000 * bdi_dirty starts high, or because bdi_thresh drops low.
1001 * In this case we don't want to hard throttle the USB key
1002 * dirtiers for 100 seconds until bdi_dirty drops under
1003 * bdi_thresh. Instead the auxiliary bdi control line in
1004 * bdi_position_ratio() will let the dirtier task progress
1005 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1006 */
Wu Fengguang16c40422010-08-11 14:17:39 -07001007 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001008
Wu Fengguange50e3722010-08-11 14:17:37 -07001009 /*
1010 * In order to avoid the stacked BDI deadlock we need
1011 * to ensure we accurately count the 'dirty' pages when
1012 * the threshold is low.
1013 *
1014 * Otherwise it would be possible to get thresh+n pages
1015 * reported dirty, even though there are thresh-m pages
1016 * actually dirty; with m+n sitting in the percpu
1017 * deltas.
1018 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001019 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1020 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1021 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001022 bdi_stat_sum(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001023 } else {
Wu Fengguang143dfe82010-08-27 18:45:12 -06001024 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1025 bdi_dirty = bdi_reclaimable +
Wu Fengguang77627412010-09-12 13:34:05 -06001026 bdi_stat(bdi, BDI_WRITEBACK);
Wu Fengguange50e3722010-08-11 14:17:37 -07001027 }
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001028
Wu Fengguang143dfe82010-08-27 18:45:12 -06001029 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
Wu Fengguang77627412010-09-12 13:34:05 -06001030 (nr_dirty > dirty_thresh);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001031 if (dirty_exceeded && !bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001032 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001033
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001034 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1035 nr_dirty, bdi_thresh, bdi_dirty,
1036 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001037
Wu Fengguang143dfe82010-08-27 18:45:12 -06001038 dirty_ratelimit = bdi->dirty_ratelimit;
1039 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1040 background_thresh, nr_dirty,
1041 bdi_thresh, bdi_dirty);
1042 if (unlikely(pos_ratio == 0)) {
1043 pause = MAX_PAUSE;
1044 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001046 task_ratelimit = (u64)dirty_ratelimit *
1047 pos_ratio >> RATELIMIT_CALC_SHIFT;
1048 pause = (HZ * pages_dirtied) / (task_ratelimit | 1);
1049 pause = min_t(long, pause, MAX_PAUSE);
1050
1051pause:
Wu Fengguangd153ba62010-12-21 17:24:21 -08001052 __set_current_state(TASK_UNINTERRUPTIBLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001053 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001054
Wu Fengguangffd1f602011-06-19 22:18:42 -06001055 dirty_thresh = hard_dirty_limit(dirty_thresh);
1056 /*
1057 * max-pause area. If dirty exceeded but still within this
1058 * area, no need to sleep for more than 200ms: (a) 8 pages per
1059 * 200ms is typically more than enough to curb heavy dirtiers;
1060 * (b) the pause time limit makes the dirtiers more responsive.
1061 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001062 if (nr_dirty < dirty_thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001063 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 }
1065
Wu Fengguang143dfe82010-08-27 18:45:12 -06001066 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001067 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
Wu Fengguang9d823e82011-06-11 18:10:12 -06001069 current->nr_dirtied = 0;
1070 current->nr_dirtied_pause = dirty_poll_interval(nr_dirty, dirty_thresh);
1071
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001073 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074
1075 /*
1076 * In laptop mode, we wait until hitting the higher threshold before
1077 * starting background writeout, and then write out all the way down
1078 * to the lower threshold. So slow writers cause minimal disk activity.
1079 *
1080 * In normal mode, we start background writeout at the lower
1081 * background_thresh, to keep the amount of dirty memory low.
1082 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001083 if (laptop_mode)
1084 return;
1085
1086 if (nr_reclaimable > background_thresh)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001087 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088}
1089
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001090void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001091{
Peter Zijlstraa200ee12007-10-08 18:54:37 +02001092 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -07001093 struct address_space *mapping = page_mapping(page);
1094
1095 if (mapping)
1096 balance_dirty_pages_ratelimited(mapping);
1097 }
1098}
1099
Wu Fengguang9d823e82011-06-11 18:10:12 -06001100static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001101
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001103 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001104 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +02001105 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 *
1107 * Processes which are dirtying memory should call in here once for each page
1108 * which was newly dirtied. The function will periodically check the system's
1109 * dirty state and will initiate writeback if needed.
1110 *
1111 * On really big machines, get_writeback_state is expensive, so try to avoid
1112 * calling it too often (ratelimiting). But once we're over the dirty memory
1113 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1114 * from overshooting the limit by (ratelimit_pages) each.
1115 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001116void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1117 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001119 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001120 int ratelimit;
1121 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122
Wu Fengguang36715ce2011-06-11 17:53:57 -06001123 if (!bdi_cap_account_dirty(bdi))
1124 return;
1125
Wu Fengguang9d823e82011-06-11 18:10:12 -06001126 ratelimit = current->nr_dirtied_pause;
1127 if (bdi->dirty_exceeded)
1128 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129
Wu Fengguang9d823e82011-06-11 18:10:12 -06001130 current->nr_dirtied += nr_pages_dirtied;
1131
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001132 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001133 /*
1134 * This prevents one CPU to accumulate too many dirtied pages without
1135 * calling into balance_dirty_pages(), which can happen when there are
1136 * 1000+ tasks, all of them start dirtying pages at exactly the same
1137 * time, hence all honoured too large initial task->nr_dirtied_pause.
1138 */
Tejun Heo245b2e72009-06-24 15:13:48 +09001139 p = &__get_cpu_var(bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001140 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001141 *p = 0;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001142 else {
1143 *p += nr_pages_dirtied;
1144 if (unlikely(*p >= ratelimit_pages)) {
1145 *p = 0;
1146 ratelimit = 0;
1147 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001149 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001150
1151 if (unlikely(current->nr_dirtied >= ratelimit))
1152 balance_dirty_pages(mapping, current->nr_dirtied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153}
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001154EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Andrew Morton232ea4d2007-02-28 20:13:21 -08001156void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157{
David Rientjes364aeb22009-01-06 14:39:29 -08001158 unsigned long background_thresh;
1159 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160
1161 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001162 global_dirty_limits(&background_thresh, &dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163
1164 /*
1165 * Boost the allowable dirty threshold a bit for page
1166 * allocators so they don't get DoS'ed by heavy writers
1167 */
1168 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1169
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001170 if (global_page_state(NR_UNSTABLE_NFS) +
1171 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1172 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001173 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001174
1175 /*
1176 * The caller might hold locks which can prevent IO completion
1177 * or progress in the filesystem. So we cannot just sit here
1178 * waiting for IO to complete.
1179 */
1180 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1181 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 }
1183}
1184
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1187 */
1188int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001189 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001191 proc_dointvec(table, write, buffer, length, ppos);
Jens Axboe64231042010-05-21 20:00:35 +02001192 bdi_arm_supers_timer();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 return 0;
1194}
1195
Jens Axboec2c49862010-05-20 09:18:47 +02001196#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001197void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198{
Matthew Garrett31373d02010-04-06 14:25:14 +02001199 struct request_queue *q = (struct request_queue *)data;
1200 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1201 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202
Matthew Garrett31373d02010-04-06 14:25:14 +02001203 /*
1204 * We want to write everything out, not just down to the dirty
1205 * threshold
1206 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001207 if (bdi_has_dirty_io(&q->backing_dev_info))
Christoph Hellwigc5444192010-06-08 18:15:15 +02001208 bdi_start_writeback(&q->backing_dev_info, nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209}
1210
1211/*
1212 * We've spun up the disk and we're in laptop mode: schedule writeback
1213 * of all dirty data a few seconds from now. If the flush is already scheduled
1214 * then push it back - the user is still using the disk.
1215 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001216void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217{
Matthew Garrett31373d02010-04-06 14:25:14 +02001218 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219}
1220
1221/*
1222 * We're in laptop mode and we've just synced. The sync's writes will have
1223 * caused another writeback to be scheduled by laptop_io_completion.
1224 * Nothing needs to be written back anymore, so we unschedule the writeback.
1225 */
1226void laptop_sync_completion(void)
1227{
Matthew Garrett31373d02010-04-06 14:25:14 +02001228 struct backing_dev_info *bdi;
1229
1230 rcu_read_lock();
1231
1232 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1233 del_timer(&bdi->laptop_mode_wb_timer);
1234
1235 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236}
Jens Axboec2c49862010-05-20 09:18:47 +02001237#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238
1239/*
1240 * If ratelimit_pages is too high then we can get into dirty-data overload
1241 * if a large number of processes all perform writes at the same time.
1242 * If it is too low then SMP machines will call the (expensive)
1243 * get_writeback_state too often.
1244 *
1245 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1246 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06001247 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 */
1249
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001250void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251{
Wu Fengguang9d823e82011-06-11 18:10:12 -06001252 unsigned long background_thresh;
1253 unsigned long dirty_thresh;
1254 global_dirty_limits(&background_thresh, &dirty_thresh);
1255 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 if (ratelimit_pages < 16)
1257 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258}
1259
Chandra Seetharaman26c21432006-06-27 02:54:10 -07001260static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1262{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001263 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -08001264 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265}
1266
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001267static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 .notifier_call = ratelimit_handler,
1269 .next = NULL,
1270};
1271
1272/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001273 * Called early on to tune the page writeback dirty limits.
1274 *
1275 * We used to scale dirty pages according to how total memory
1276 * related to pages that could be allocated for buffers (by
1277 * comparing nr_free_buffer_pages() to vm_total_pages.
1278 *
1279 * However, that was when we used "dirty_ratio" to scale with
1280 * all memory, and we don't do that any more. "dirty_ratio"
1281 * is now applied to total non-HIGHPAGE memory (by subtracting
1282 * totalhigh_pages from vm_total_pages), and as such we can't
1283 * get into the old insane situation any more where we had
1284 * large amounts of dirty pages compared to a small amount of
1285 * non-HIGHMEM memory.
1286 *
1287 * But we might still want to scale the dirty_ratio by how
1288 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 */
1290void __init page_writeback_init(void)
1291{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001292 int shift;
1293
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001294 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001295 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001296
1297 shift = calc_period_shift();
1298 prop_descriptor_init(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -07001299 prop_descriptor_init(&vm_dirties, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300}
1301
David Howells811d7362006-08-29 19:06:09 +01001302/**
Jan Karaf446daa2010-08-09 17:19:12 -07001303 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1304 * @mapping: address space structure to write
1305 * @start: starting page index
1306 * @end: ending page index (inclusive)
1307 *
1308 * This function scans the page range from @start to @end (inclusive) and tags
1309 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1310 * that write_cache_pages (or whoever calls this function) will then use
1311 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1312 * used to avoid livelocking of writeback by a process steadily creating new
1313 * dirty pages in the file (thus it is important for this function to be quick
1314 * so that it can tag pages faster than a dirtying process can create them).
1315 */
1316/*
1317 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1318 */
Jan Karaf446daa2010-08-09 17:19:12 -07001319void tag_pages_for_writeback(struct address_space *mapping,
1320 pgoff_t start, pgoff_t end)
1321{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001322#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daa2010-08-09 17:19:12 -07001323 unsigned long tagged;
1324
1325 do {
1326 spin_lock_irq(&mapping->tree_lock);
1327 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1328 &start, end, WRITEBACK_TAG_BATCH,
1329 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1330 spin_unlock_irq(&mapping->tree_lock);
1331 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1332 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001333 /* We check 'start' to handle wrapping when end == ~0UL */
1334 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daa2010-08-09 17:19:12 -07001335}
1336EXPORT_SYMBOL(tag_pages_for_writeback);
1337
1338/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001339 * 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 +01001340 * @mapping: address space structure to write
1341 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001342 * @writepage: function called for each page
1343 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001344 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001345 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001346 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1347 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1348 * and msync() need to guarantee that all the data which was dirty at the time
1349 * the call was made get new I/O started against them. If wbc->sync_mode is
1350 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1351 * existing IO to complete.
Jan Karaf446daa2010-08-09 17:19:12 -07001352 *
1353 * To avoid livelocks (when other process dirties new pages), we first tag
1354 * pages which should be written back with TOWRITE tag and only then start
1355 * writing them. For data-integrity sync we have to be careful so that we do
1356 * not miss some pages (e.g., because some other process has cleared TOWRITE
1357 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1358 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001359 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001360int write_cache_pages(struct address_space *mapping,
1361 struct writeback_control *wbc, writepage_t writepage,
1362 void *data)
David Howells811d7362006-08-29 19:06:09 +01001363{
David Howells811d7362006-08-29 19:06:09 +01001364 int ret = 0;
1365 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001366 struct pagevec pvec;
1367 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001368 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001369 pgoff_t index;
1370 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001371 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001372 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001373 int range_whole = 0;
Jan Karaf446daa2010-08-09 17:19:12 -07001374 int tag;
David Howells811d7362006-08-29 19:06:09 +01001375
David Howells811d7362006-08-29 19:06:09 +01001376 pagevec_init(&pvec, 0);
1377 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001378 writeback_index = mapping->writeback_index; /* prev offset */
1379 index = writeback_index;
1380 if (index == 0)
1381 cycled = 1;
1382 else
1383 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001384 end = -1;
1385 } else {
1386 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1387 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1388 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1389 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001390 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001391 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001392 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daa2010-08-09 17:19:12 -07001393 tag = PAGECACHE_TAG_TOWRITE;
1394 else
1395 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001396retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001397 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daa2010-08-09 17:19:12 -07001398 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001399 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001400 while (!done && (index <= end)) {
1401 int i;
1402
Jan Karaf446daa2010-08-09 17:19:12 -07001403 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001404 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1405 if (nr_pages == 0)
1406 break;
David Howells811d7362006-08-29 19:06:09 +01001407
David Howells811d7362006-08-29 19:06:09 +01001408 for (i = 0; i < nr_pages; i++) {
1409 struct page *page = pvec.pages[i];
1410
Nick Piggind5482cd2009-01-06 14:39:11 -08001411 /*
1412 * At this point, the page may be truncated or
1413 * invalidated (changing page->mapping to NULL), or
1414 * even swizzled back from swapper_space to tmpfs file
1415 * mapping. However, page->index will not change
1416 * because we have a reference on the page.
1417 */
1418 if (page->index > end) {
1419 /*
1420 * can't be range_cyclic (1st pass) because
1421 * end == -1 in that case.
1422 */
1423 done = 1;
1424 break;
1425 }
1426
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001427 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001428
David Howells811d7362006-08-29 19:06:09 +01001429 lock_page(page);
1430
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001431 /*
1432 * Page truncated or invalidated. We can freely skip it
1433 * then, even for data integrity operations: the page
1434 * has disappeared concurrently, so there could be no
1435 * real expectation of this data interity operation
1436 * even if there is now a new, dirty page at the same
1437 * pagecache address.
1438 */
David Howells811d7362006-08-29 19:06:09 +01001439 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001440continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001441 unlock_page(page);
1442 continue;
1443 }
1444
Nick Piggin515f4a02009-01-06 14:39:10 -08001445 if (!PageDirty(page)) {
1446 /* someone wrote it for us */
1447 goto continue_unlock;
1448 }
David Howells811d7362006-08-29 19:06:09 +01001449
Nick Piggin515f4a02009-01-06 14:39:10 -08001450 if (PageWriteback(page)) {
1451 if (wbc->sync_mode != WB_SYNC_NONE)
1452 wait_on_page_writeback(page);
1453 else
1454 goto continue_unlock;
1455 }
1456
1457 BUG_ON(PageWriteback(page));
1458 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001459 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001460
Dave Chinner9e094382010-07-07 13:24:08 +10001461 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001462 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001463 if (unlikely(ret)) {
1464 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1465 unlock_page(page);
1466 ret = 0;
1467 } else {
1468 /*
1469 * done_index is set past this page,
1470 * so media errors will not choke
1471 * background writeout for the entire
1472 * file. This has consequences for
1473 * range_cyclic semantics (ie. it may
1474 * not be suitable for data integrity
1475 * writeout).
1476 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001477 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001478 done = 1;
1479 break;
1480 }
Dave Chinner0b564922010-06-09 10:37:18 +10001481 }
David Howells811d7362006-08-29 19:06:09 +01001482
Dave Chinner546a1922010-08-24 11:44:34 +10001483 /*
1484 * We stop writing back only if we are not doing
1485 * integrity sync. In case of integrity sync we have to
1486 * keep going until we have written all the pages
1487 * we tagged for writeback prior to entering this loop.
1488 */
1489 if (--wbc->nr_to_write <= 0 &&
1490 wbc->sync_mode == WB_SYNC_NONE) {
1491 done = 1;
1492 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001493 }
David Howells811d7362006-08-29 19:06:09 +01001494 }
1495 pagevec_release(&pvec);
1496 cond_resched();
1497 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001498 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001499 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001500 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001501 * We hit the last page and there is more work to be done: wrap
1502 * back to the start of the file
1503 */
Nick Piggin31a12662009-01-06 14:39:04 -08001504 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001505 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001506 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001507 goto retry;
1508 }
Dave Chinner0b564922010-06-09 10:37:18 +10001509 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1510 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001511
David Howells811d7362006-08-29 19:06:09 +01001512 return ret;
1513}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001514EXPORT_SYMBOL(write_cache_pages);
1515
1516/*
1517 * Function used by generic_writepages to call the real writepage
1518 * function and set the mapping flags on error
1519 */
1520static int __writepage(struct page *page, struct writeback_control *wbc,
1521 void *data)
1522{
1523 struct address_space *mapping = data;
1524 int ret = mapping->a_ops->writepage(page, wbc);
1525 mapping_set_error(mapping, ret);
1526 return ret;
1527}
1528
1529/**
1530 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1531 * @mapping: address space structure to write
1532 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1533 *
1534 * This is a library function, which implements the writepages()
1535 * address_space_operation.
1536 */
1537int generic_writepages(struct address_space *mapping,
1538 struct writeback_control *wbc)
1539{
Shaohua Li9b6096a2011-03-17 10:47:06 +01001540 struct blk_plug plug;
1541 int ret;
1542
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001543 /* deal with chardevs and other special file */
1544 if (!mapping->a_ops->writepage)
1545 return 0;
1546
Shaohua Li9b6096a2011-03-17 10:47:06 +01001547 blk_start_plug(&plug);
1548 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1549 blk_finish_plug(&plug);
1550 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001551}
David Howells811d7362006-08-29 19:06:09 +01001552
1553EXPORT_SYMBOL(generic_writepages);
1554
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1556{
Andrew Morton22905f72005-11-16 15:07:01 -08001557 int ret;
1558
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 if (wbc->nr_to_write <= 0)
1560 return 0;
1561 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001562 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001563 else
1564 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001565 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566}
1567
1568/**
1569 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001570 * @page: the page to write
1571 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 *
1573 * The page must be locked by the caller and will be unlocked upon return.
1574 *
1575 * write_one_page() returns a negative error code if I/O failed.
1576 */
1577int write_one_page(struct page *page, int wait)
1578{
1579 struct address_space *mapping = page->mapping;
1580 int ret = 0;
1581 struct writeback_control wbc = {
1582 .sync_mode = WB_SYNC_ALL,
1583 .nr_to_write = 1,
1584 };
1585
1586 BUG_ON(!PageLocked(page));
1587
1588 if (wait)
1589 wait_on_page_writeback(page);
1590
1591 if (clear_page_dirty_for_io(page)) {
1592 page_cache_get(page);
1593 ret = mapping->a_ops->writepage(page, &wbc);
1594 if (ret == 0 && wait) {
1595 wait_on_page_writeback(page);
1596 if (PageError(page))
1597 ret = -EIO;
1598 }
1599 page_cache_release(page);
1600 } else {
1601 unlock_page(page);
1602 }
1603 return ret;
1604}
1605EXPORT_SYMBOL(write_one_page);
1606
1607/*
Ken Chen76719322007-02-10 01:43:15 -08001608 * For address_spaces which do not use buffers nor write back.
1609 */
1610int __set_page_dirty_no_writeback(struct page *page)
1611{
1612 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08001613 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08001614 return 0;
1615}
1616
1617/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001618 * Helper function for set_page_dirty family.
1619 * NOTE: This relies on being atomic wrt interrupts.
1620 */
1621void account_page_dirtied(struct page *page, struct address_space *mapping)
1622{
1623 if (mapping_cap_account_dirty(mapping)) {
1624 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07001625 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001626 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06001627 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001628 task_dirty_inc(current);
1629 task_io_account_write(PAGE_CACHE_SIZE);
1630 }
1631}
Michael Rubin679ceac2010-08-20 02:31:26 -07001632EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001633
1634/*
Michael Rubinf629d1c2010-10-26 14:21:33 -07001635 * Helper function for set_page_writeback family.
1636 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1637 * wrt interrupts.
1638 */
1639void account_page_writeback(struct page *page)
1640{
1641 inc_zone_page_state(page, NR_WRITEBACK);
1642}
1643EXPORT_SYMBOL(account_page_writeback);
1644
1645/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 * For address_spaces which do not use buffers. Just tag the page as dirty in
1647 * its radix tree.
1648 *
1649 * This is also used when a single buffer is being dirtied: we want to set the
1650 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1651 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1652 *
1653 * Most callers have locked the page, which pins the address_space in memory.
1654 * But zap_pte_range() does not lock the page, however in that case the
1655 * mapping is pinned by the vma's ->vm_file reference.
1656 *
1657 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001658 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 */
1660int __set_page_dirty_nobuffers(struct page *page)
1661{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 if (!TestSetPageDirty(page)) {
1663 struct address_space *mapping = page_mapping(page);
1664 struct address_space *mapping2;
1665
Andrew Morton8c085402006-12-10 02:19:24 -08001666 if (!mapping)
1667 return 1;
1668
Nick Piggin19fd6232008-07-25 19:45:32 -07001669 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001670 mapping2 = page_mapping(page);
1671 if (mapping2) { /* Race with truncate? */
1672 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001673 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001674 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001675 radix_tree_tag_set(&mapping->page_tree,
1676 page_index(page), PAGECACHE_TAG_DIRTY);
1677 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001678 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001679 if (mapping->host) {
1680 /* !PageAnon && !swapper_space */
1681 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001683 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001685 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686}
1687EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1688
1689/*
1690 * When a writepage implementation decides that it doesn't want to write this
1691 * page for some reason, it should redirty the locked page via
1692 * redirty_page_for_writepage() and it should then unlock the page and return 0
1693 */
1694int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1695{
1696 wbc->pages_skipped++;
1697 return __set_page_dirty_nobuffers(page);
1698}
1699EXPORT_SYMBOL(redirty_page_for_writepage);
1700
1701/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02001702 * Dirty a page.
1703 *
1704 * For pages with a mapping this should be done under the page lock
1705 * for the benefit of asynchronous memory errors who prefer a consistent
1706 * dirty state. This rule can be broken in some special cases,
1707 * but should be better not to.
1708 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 * If the mapping doesn't provide a set_page_dirty a_op, then
1710 * just fall through and assume that it wants buffer_heads.
1711 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001712int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713{
1714 struct address_space *mapping = page_mapping(page);
1715
1716 if (likely(mapping)) {
1717 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07001718 /*
1719 * readahead/lru_deactivate_page could remain
1720 * PG_readahead/PG_reclaim due to race with end_page_writeback
1721 * About readahead, if the page is written, the flags would be
1722 * reset. So no problem.
1723 * About lru_deactivate_page, if the page is redirty, the flag
1724 * will be reset. So no problem. but if the page is used by readahead
1725 * it will confuse readahead and make it restart the size rampup
1726 * process. But it's a trivial problem.
1727 */
1728 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02001729#ifdef CONFIG_BLOCK
1730 if (!spd)
1731 spd = __set_page_dirty_buffers;
1732#endif
1733 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001735 if (!PageDirty(page)) {
1736 if (!TestSetPageDirty(page))
1737 return 1;
1738 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 return 0;
1740}
1741EXPORT_SYMBOL(set_page_dirty);
1742
1743/*
1744 * set_page_dirty() is racy if the caller has no reference against
1745 * page->mapping->host, and if the page is unlocked. This is because another
1746 * CPU could truncate the page off the mapping and then free the mapping.
1747 *
1748 * Usually, the page _is_ locked, or the caller is a user-space process which
1749 * holds a reference on the inode by having an open file.
1750 *
1751 * In other cases, the page should be locked before running set_page_dirty().
1752 */
1753int set_page_dirty_lock(struct page *page)
1754{
1755 int ret;
1756
Jens Axboe7eaceac2011-03-10 08:52:07 +01001757 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 ret = set_page_dirty(page);
1759 unlock_page(page);
1760 return ret;
1761}
1762EXPORT_SYMBOL(set_page_dirty_lock);
1763
1764/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 * Clear a page's dirty flag, while caring for dirty memory accounting.
1766 * Returns true if the page was previously dirty.
1767 *
1768 * This is for preparing to put the page under writeout. We leave the page
1769 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1770 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1771 * implementation will run either set_page_writeback() or set_page_dirty(),
1772 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1773 * back into sync.
1774 *
1775 * This incoherency between the page's dirty flag and radix-tree tag is
1776 * unfortunate, but it only exists while the page is locked.
1777 */
1778int clear_page_dirty_for_io(struct page *page)
1779{
1780 struct address_space *mapping = page_mapping(page);
1781
Nick Piggin79352892007-07-19 01:47:22 -07001782 BUG_ON(!PageLocked(page));
1783
Linus Torvalds7658cc22006-12-29 10:00:58 -08001784 if (mapping && mapping_cap_account_dirty(mapping)) {
1785 /*
1786 * Yes, Virginia, this is indeed insane.
1787 *
1788 * We use this sequence to make sure that
1789 * (a) we account for dirty stats properly
1790 * (b) we tell the low-level filesystem to
1791 * mark the whole page dirty if it was
1792 * dirty in a pagetable. Only to then
1793 * (c) clean the page again and return 1 to
1794 * cause the writeback.
1795 *
1796 * This way we avoid all nasty races with the
1797 * dirty bit in multiple places and clearing
1798 * them concurrently from different threads.
1799 *
1800 * Note! Normally the "set_page_dirty(page)"
1801 * has no effect on the actual dirty bit - since
1802 * that will already usually be set. But we
1803 * need the side effects, and it can help us
1804 * avoid races.
1805 *
1806 * We basically use the page "master dirty bit"
1807 * as a serialization point for all the different
1808 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001809 */
1810 if (page_mkclean(page))
1811 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001812 /*
1813 * We carefully synchronise fault handlers against
1814 * installing a dirty pte and marking the page dirty
1815 * at this point. We do this by having them hold the
1816 * page lock at some point after installing their
1817 * pte, but before marking the page dirty.
1818 * Pages are always locked coming in here, so we get
1819 * the desired exclusion. See mm/memory.c:do_wp_page()
1820 * for more comments.
1821 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001822 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001823 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001824 dec_bdi_stat(mapping->backing_dev_info,
1825 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001826 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001828 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001830 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001832EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833
1834int test_clear_page_writeback(struct page *page)
1835{
1836 struct address_space *mapping = page_mapping(page);
1837 int ret;
1838
1839 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001840 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841 unsigned long flags;
1842
Nick Piggin19fd6232008-07-25 19:45:32 -07001843 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001845 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 radix_tree_tag_clear(&mapping->page_tree,
1847 page_index(page),
1848 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001849 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001850 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001851 __bdi_writeout_inc(bdi);
1852 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001853 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001854 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 } else {
1856 ret = TestClearPageWriteback(page);
1857 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07001858 if (ret) {
Andrew Mortond688abf2007-07-19 01:49:17 -07001859 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07001860 inc_zone_page_state(page, NR_WRITTEN);
1861 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 return ret;
1863}
1864
1865int test_set_page_writeback(struct page *page)
1866{
1867 struct address_space *mapping = page_mapping(page);
1868 int ret;
1869
1870 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001871 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 unsigned long flags;
1873
Nick Piggin19fd6232008-07-25 19:45:32 -07001874 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001876 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 radix_tree_tag_set(&mapping->page_tree,
1878 page_index(page),
1879 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001880 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001881 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1882 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 if (!PageDirty(page))
1884 radix_tree_tag_clear(&mapping->page_tree,
1885 page_index(page),
1886 PAGECACHE_TAG_DIRTY);
Jan Karaf446daa2010-08-09 17:19:12 -07001887 radix_tree_tag_clear(&mapping->page_tree,
1888 page_index(page),
1889 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07001890 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 } else {
1892 ret = TestSetPageWriteback(page);
1893 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001894 if (!ret)
Michael Rubinf629d1c2010-10-26 14:21:33 -07001895 account_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 return ret;
1897
1898}
1899EXPORT_SYMBOL(test_set_page_writeback);
1900
1901/*
Nick Piggin00128182007-10-16 01:24:40 -07001902 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 * passed tag.
1904 */
1905int mapping_tagged(struct address_space *mapping, int tag)
1906{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07001907 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908}
1909EXPORT_SYMBOL(mapping_tagged);