blob: 6f4335238e33311de251a647fe725d06d5897060 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
Francois Camie1f8e872008-10-15 22:01:59 -070010 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Initial version
12 */
13
14#include <linux/kernel.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040015#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
Al Viroff01bb482011-09-16 02:31:11 -040035#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Jan Karaeb608e32012-05-24 18:59:11 +020037#include <linux/timer.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060038#include <linux/sched/rt.h>
Lisa Du6e543d52013-09-11 14:22:36 -070039#include <linux/mm_inline.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100040#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
Lisa Du6e543d52013-09-11 14:22:36 -070042#include "internal.h"
43
Linus Torvalds1da177e2005-04-16 15:20:36 -070044/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060045 * Sleep at most 200ms at a time in balance_dirty_pages().
46 */
47#define MAX_PAUSE max(HZ/5, 1)
48
49/*
Wu Fengguang5b9b3572011-12-06 13:17:17 -060050 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
52 */
53#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
54
55/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060056 * Estimate write bandwidth at 200ms intervals.
57 */
58#define BANDWIDTH_INTERVAL max(HZ/5, 1)
59
Wu Fengguang6c14ae12011-03-02 16:04:18 -060060#define RATELIMIT_CALC_SHIFT 10
61
Wu Fengguange98be2d2010-08-29 11:22:30 -060062/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
65 */
66static long ratelimit_pages = 32;
67
Linus Torvalds1da177e2005-04-16 15:20:36 -070068/* The following parameters are exported via /proc/sys/vm */
69
70/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020071 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080073int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
75/*
David Rientjes2da02992009-01-06 14:39:31 -080076 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
78 */
79unsigned long dirty_background_bytes;
80
81/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080082 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
84 */
85int vm_highmem_is_dirtyable;
86
87/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 * The generator of dirty data starts writeback at this percentage
89 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080090int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92/*
David Rientjes2da02992009-01-06 14:39:31 -080093 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
95 */
96unsigned long vm_dirty_bytes;
97
98/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070099 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Artem Bityutskiy91913a22012-03-21 22:33:00 -0400103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700106 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110/*
111 * Flag that makes the machine dump writes/reads and block dirtyings.
112 */
113int block_dump;
114
115/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118 */
119int laptop_mode;
120
121EXPORT_SYMBOL(laptop_mode);
122
123/* End of sysctl-exported parameters */
124
Wu Fengguangc42843f2011-03-02 15:54:09 -0600125unsigned long global_dirty_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700128 * Scale the writeback cache size proportional to the relative writeout speeds.
129 *
130 * We do this by keeping a floating proportion between BDIs, based on page
131 * writeback completions [end_page_writeback()]. Those devices that write out
132 * pages fastest will get the larger share, while the slower will get a smaller
133 * share.
134 *
135 * We use page writeout completions because we are interested in getting rid of
136 * dirty pages. Having them written out is the primary goal.
137 *
138 * We introduce a concept of time, a period over which we measure these events,
139 * because demand can/will vary over time. The length of this period itself is
140 * measured in page writeback completions.
141 *
142 */
Jan Karaeb608e32012-05-24 18:59:11 +0200143static struct fprop_global writeout_completions;
144
145static void writeout_period(unsigned long t);
146/* Timer for aging of writeout_completions */
147static struct timer_list writeout_period_timer =
148 TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
149static unsigned long writeout_period_time = 0;
150
151/*
152 * Length of period for aging writeout fractions of bdis. This is an
153 * arbitrarily chosen number. The longer the period, the slower fractions will
154 * reflect changes in current writeout rate.
155 */
156#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700157
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700158/*
Johannes Weinera756cf52012-01-10 15:07:49 -0800159 * In a memory zone, there is a certain amount of pages we consider
160 * available for the page cache, which is essentially the number of
161 * free and reclaimable pages, minus some zone reserves to protect
162 * lowmem and the ability to uphold the zone's watermarks without
163 * requiring writeback.
164 *
165 * This number of dirtyable pages is the base value of which the
166 * user-configurable dirty ratio is the effictive number of pages that
167 * are allowed to be actually dirtied. Per individual zone, or
168 * globally by using the sum of dirtyable pages over all zones.
169 *
170 * Because the user is allowed to specify the dirty limit globally as
171 * absolute number of bytes, calculating the per-zone dirty limit can
172 * require translating the configured limit into a percentage of
173 * global dirtyable memory first.
174 */
175
Johannes Weinera8045522014-01-29 14:05:39 -0800176/**
177 * zone_dirtyable_memory - number of dirtyable pages in a zone
178 * @zone: the zone
179 *
180 * Returns the zone's number of pages potentially available for dirty
181 * page cache. This is the base value for the per-zone dirty limits.
182 */
183static unsigned long zone_dirtyable_memory(struct zone *zone)
184{
185 unsigned long nr_pages;
186
187 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
188 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
189
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800190 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
191 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800192
193 return nr_pages;
194}
195
Johannes Weiner1edf2232012-01-10 15:06:57 -0800196static unsigned long highmem_dirtyable_memory(unsigned long total)
197{
198#ifdef CONFIG_HIGHMEM
199 int node;
200 unsigned long x = 0;
201
202 for_each_node_state(node, N_HIGH_MEMORY) {
Johannes Weinera8045522014-01-29 14:05:39 -0800203 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
Johannes Weiner1edf2232012-01-10 15:06:57 -0800204
Johannes Weinera8045522014-01-29 14:05:39 -0800205 x += zone_dirtyable_memory(z);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800206 }
207 /*
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800208 * Unreclaimable memory (kernel memory or anonymous memory
209 * without swap) can bring down the dirtyable pages below
210 * the zone's dirty balance reserve and the above calculation
211 * will underflow. However we still want to add in nodes
212 * which are below threshold (negative values) to get a more
213 * accurate calculation but make sure that the total never
214 * underflows.
215 */
216 if ((long)x < 0)
217 x = 0;
218
219 /*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800220 * Make sure that the number of highmem pages is never larger
221 * than the number of the total dirtyable memory. This can only
222 * occur in very strange VM situations but we want to make sure
223 * that this does not occur.
224 */
225 return min(x, total);
226#else
227 return 0;
228#endif
229}
230
231/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800232 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800233 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800234 * Returns the global number of pages potentially available for dirty
235 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800236 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700237static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800238{
239 unsigned long x;
240
Johannes Weinera8045522014-01-29 14:05:39 -0800241 x = global_page_state(NR_FREE_PAGES);
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800242 x -= min(x, dirty_balance_reserve);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800243
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800244 x += global_page_state(NR_INACTIVE_FILE);
245 x += global_page_state(NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800246
Johannes Weiner1edf2232012-01-10 15:06:57 -0800247 if (!vm_highmem_is_dirtyable)
248 x -= highmem_dirtyable_memory(x);
249
250 return x + 1; /* Ensure that we never return 0 */
251}
252
253/*
Johannes Weinerccafa282012-01-10 15:07:44 -0800254 * global_dirty_limits - background-writeback and dirty-throttling thresholds
255 *
256 * Calculate the dirty thresholds based on sysctl parameters
257 * - vm.dirty_background_ratio or vm.dirty_background_bytes
258 * - vm.dirty_ratio or vm.dirty_bytes
259 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
260 * real-time tasks.
261 */
262void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
263{
David Rientjes9ef0a0f2014-08-06 16:07:31 -0700264 const unsigned long available_memory = global_dirtyable_memory();
Johannes Weinerccafa282012-01-10 15:07:44 -0800265 unsigned long background;
266 unsigned long dirty;
Johannes Weinerccafa282012-01-10 15:07:44 -0800267 struct task_struct *tsk;
268
Johannes Weinerccafa282012-01-10 15:07:44 -0800269 if (vm_dirty_bytes)
270 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
271 else
272 dirty = (vm_dirty_ratio * available_memory) / 100;
273
274 if (dirty_background_bytes)
275 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
276 else
277 background = (dirty_background_ratio * available_memory) / 100;
278
279 if (background >= dirty)
280 background = dirty / 2;
281 tsk = current;
282 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
283 background += background / 4;
284 dirty += dirty / 4;
285 }
286 *pbackground = background;
287 *pdirty = dirty;
288 trace_global_dirty_state(background, dirty);
289}
290
Johannes Weinera756cf52012-01-10 15:07:49 -0800291/**
Johannes Weinera756cf52012-01-10 15:07:49 -0800292 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
293 * @zone: the zone
294 *
295 * Returns the maximum number of dirty pages allowed in a zone, based
296 * on the zone's dirtyable memory.
297 */
298static unsigned long zone_dirty_limit(struct zone *zone)
299{
300 unsigned long zone_memory = zone_dirtyable_memory(zone);
301 struct task_struct *tsk = current;
302 unsigned long dirty;
303
304 if (vm_dirty_bytes)
305 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
306 zone_memory / global_dirtyable_memory();
307 else
308 dirty = vm_dirty_ratio * zone_memory / 100;
309
310 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
311 dirty += dirty / 4;
312
313 return dirty;
314}
315
316/**
317 * zone_dirty_ok - tells whether a zone is within its dirty limits
318 * @zone: the zone to check
319 *
320 * Returns %true when the dirty pages in @zone are within the zone's
321 * dirty limit, %false if the limit is exceeded.
322 */
323bool zone_dirty_ok(struct zone *zone)
324{
325 unsigned long limit = zone_dirty_limit(zone);
326
327 return zone_page_state(zone, NR_FILE_DIRTY) +
328 zone_page_state(zone, NR_UNSTABLE_NFS) +
329 zone_page_state(zone, NR_WRITEBACK) <= limit;
330}
331
David Rientjes2da02992009-01-06 14:39:31 -0800332int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700333 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800334 loff_t *ppos)
335{
336 int ret;
337
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700338 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800339 if (ret == 0 && write)
340 dirty_background_bytes = 0;
341 return ret;
342}
343
344int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700345 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800346 loff_t *ppos)
347{
348 int ret;
349
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700350 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800351 if (ret == 0 && write)
352 dirty_background_ratio = 0;
353 return ret;
354}
355
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700356int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700357 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700358 loff_t *ppos)
359{
360 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800361 int ret;
362
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700363 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700364 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200365 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800366 vm_dirty_bytes = 0;
367 }
368 return ret;
369}
370
David Rientjes2da02992009-01-06 14:39:31 -0800371int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700372 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800373 loff_t *ppos)
374{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800375 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800376 int ret;
377
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700378 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800379 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200380 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800381 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700382 }
383 return ret;
384}
385
Jan Karaeb608e32012-05-24 18:59:11 +0200386static unsigned long wp_next_time(unsigned long cur_time)
387{
388 cur_time += VM_COMPLETIONS_PERIOD_LEN;
389 /* 0 has a special meaning... */
390 if (!cur_time)
391 return 1;
392 return cur_time;
393}
394
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700395/*
396 * Increment the BDI's writeout completion count and the global writeout
397 * completion count. Called from test_clear_page_writeback().
398 */
399static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
400{
Jan Karaf7d2b1e2010-12-08 22:44:24 -0600401 __inc_bdi_stat(bdi, BDI_WRITTEN);
Jan Karaeb608e32012-05-24 18:59:11 +0200402 __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
403 bdi->max_prop_frac);
404 /* First event after period switching was turned off? */
405 if (!unlikely(writeout_period_time)) {
406 /*
407 * We can race with other __bdi_writeout_inc calls here but
408 * it does not cause any harm since the resulting time when
409 * timer will fire and what is in writeout_period_time will be
410 * roughly the same.
411 */
412 writeout_period_time = wp_next_time(jiffies);
413 mod_timer(&writeout_period_timer, writeout_period_time);
414 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700415}
416
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700417void bdi_writeout_inc(struct backing_dev_info *bdi)
418{
419 unsigned long flags;
420
421 local_irq_save(flags);
422 __bdi_writeout_inc(bdi);
423 local_irq_restore(flags);
424}
425EXPORT_SYMBOL_GPL(bdi_writeout_inc);
426
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700427/*
428 * Obtain an accurate fraction of the BDI's portion.
429 */
430static void bdi_writeout_fraction(struct backing_dev_info *bdi,
431 long *numerator, long *denominator)
432{
Jan Karaeb608e32012-05-24 18:59:11 +0200433 fprop_fraction_percpu(&writeout_completions, &bdi->completions,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700434 numerator, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700435}
436
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700437/*
Jan Karaeb608e32012-05-24 18:59:11 +0200438 * On idle system, we can be called long after we scheduled because we use
439 * deferred timers so count with missed periods.
440 */
441static void writeout_period(unsigned long t)
442{
443 int miss_periods = (jiffies - writeout_period_time) /
444 VM_COMPLETIONS_PERIOD_LEN;
445
446 if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
447 writeout_period_time = wp_next_time(writeout_period_time +
448 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
449 mod_timer(&writeout_period_timer, writeout_period_time);
450 } else {
451 /*
452 * Aging has zeroed all fractions. Stop wasting CPU on period
453 * updates.
454 */
455 writeout_period_time = 0;
456 }
457}
458
459/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700460 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
461 * registered backing devices, which, for obvious reasons, can not
462 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700463 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700464static unsigned int bdi_min_ratio;
465
466int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
467{
468 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700469
Jens Axboecfc4ba52009-09-14 13:12:40 +0200470 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700471 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700472 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700473 } else {
474 min_ratio -= bdi->min_ratio;
475 if (bdi_min_ratio + min_ratio < 100) {
476 bdi_min_ratio += min_ratio;
477 bdi->min_ratio += min_ratio;
478 } else {
479 ret = -EINVAL;
480 }
481 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200482 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700483
484 return ret;
485}
486
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700487int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
488{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700489 int ret = 0;
490
491 if (max_ratio > 100)
492 return -EINVAL;
493
Jens Axboecfc4ba52009-09-14 13:12:40 +0200494 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700495 if (bdi->min_ratio > max_ratio) {
496 ret = -EINVAL;
497 } else {
498 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200499 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700500 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200501 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700502
503 return ret;
504}
505EXPORT_SYMBOL(bdi_set_max_ratio);
506
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600507static unsigned long dirty_freerun_ceiling(unsigned long thresh,
508 unsigned long bg_thresh)
509{
510 return (thresh + bg_thresh) / 2;
511}
512
Wu Fengguangffd1f602011-06-19 22:18:42 -0600513static unsigned long hard_dirty_limit(unsigned long thresh)
514{
515 return max(thresh, global_dirty_limit);
516}
517
Wu Fengguang6f718652011-03-02 17:14:34 -0600518/**
Wu Fengguang1babe182010-08-11 14:17:40 -0700519 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
Wu Fengguang6f718652011-03-02 17:14:34 -0600520 * @bdi: the backing_dev_info to query
521 * @dirty: global dirty limit in pages
Wu Fengguang1babe182010-08-11 14:17:40 -0700522 *
Wu Fengguang6f718652011-03-02 17:14:34 -0600523 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
524 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600525 *
526 * Note that balance_dirty_pages() will only seriously take it as a hard limit
527 * when sleeping max_pause per page is not enough to keep the dirty pages under
528 * control. For example, when the device is completely stalled due to some error
529 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
530 * In the other normal situations, it acts more gently by throttling the tasks
531 * more (rather than completely block them) when the bdi dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600532 *
533 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700534 * - starving fast devices
535 * - piling up dirty pages (that will take long time to sync) on slow devices
536 *
537 * The bdi's share of dirty limit will be adapting to its throughput and
538 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
539 */
540unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
Wu Fengguang16c40422010-08-11 14:17:39 -0700541{
542 u64 bdi_dirty;
543 long numerator, denominator;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700544
Wu Fengguang16c40422010-08-11 14:17:39 -0700545 /*
546 * Calculate this BDI's share of the dirty ratio.
547 */
548 bdi_writeout_fraction(bdi, &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700549
Wu Fengguang16c40422010-08-11 14:17:39 -0700550 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
551 bdi_dirty *= numerator;
552 do_div(bdi_dirty, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700553
Wu Fengguang16c40422010-08-11 14:17:39 -0700554 bdi_dirty += (dirty * bdi->min_ratio) / 100;
555 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
556 bdi_dirty = dirty * bdi->max_ratio / 100;
557
558 return bdi_dirty;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600561/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700562 * setpoint - dirty 3
563 * f(dirty) := 1.0 + (----------------)
564 * limit - setpoint
565 *
566 * it's a 3rd order polynomial that subjects to
567 *
568 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
569 * (2) f(setpoint) = 1.0 => the balance point
570 * (3) f(limit) = 0 => the hard limit
571 * (4) df/dx <= 0 => negative feedback control
572 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
573 * => fast response on large errors; small oscillation near setpoint
574 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700575static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700576 unsigned long dirty,
577 unsigned long limit)
578{
579 long long pos_ratio;
580 long x;
581
Rik van Rield5c9fde2014-05-06 12:50:01 -0700582 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700583 limit - setpoint + 1);
584 pos_ratio = x;
585 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
586 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
587 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
588
589 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
590}
591
592/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600593 * Dirty position control.
594 *
595 * (o) global/bdi setpoints
596 *
597 * We want the dirty pages be balanced around the global/bdi setpoints.
598 * When the number of dirty pages is higher/lower than the setpoint, the
599 * dirty position control ratio (and hence task dirty ratelimit) will be
600 * decreased/increased to bring the dirty pages back to the setpoint.
601 *
602 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
603 *
604 * if (dirty < setpoint) scale up pos_ratio
605 * if (dirty > setpoint) scale down pos_ratio
606 *
607 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
608 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
609 *
610 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
611 *
612 * (o) global control line
613 *
614 * ^ pos_ratio
615 * |
616 * | |<===== global dirty control scope ======>|
617 * 2.0 .............*
618 * | .*
619 * | . *
620 * | . *
621 * | . *
622 * | . *
623 * | . *
624 * 1.0 ................................*
625 * | . . *
626 * | . . *
627 * | . . *
628 * | . . *
629 * | . . *
630 * 0 +------------.------------------.----------------------*------------->
631 * freerun^ setpoint^ limit^ dirty pages
632 *
633 * (o) bdi control line
634 *
635 * ^ pos_ratio
636 * |
637 * | *
638 * | *
639 * | *
640 * | *
641 * | * |<=========== span ============>|
642 * 1.0 .......................*
643 * | . *
644 * | . *
645 * | . *
646 * | . *
647 * | . *
648 * | . *
649 * | . *
650 * | . *
651 * | . *
652 * | . *
653 * | . *
654 * 1/4 ...............................................* * * * * * * * * * * *
655 * | . .
656 * | . .
657 * | . .
658 * 0 +----------------------.-------------------------------.------------->
659 * bdi_setpoint^ x_intercept^
660 *
661 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
662 * be smoothly throttled down to normal if it starts high in situations like
663 * - start writing to a slow SD card and a fast disk at the same time. The SD
664 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
665 * - the bdi dirty thresh drops quickly due to change of JBOD workload
666 */
667static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
668 unsigned long thresh,
669 unsigned long bg_thresh,
670 unsigned long dirty,
671 unsigned long bdi_thresh,
672 unsigned long bdi_dirty)
673{
674 unsigned long write_bw = bdi->avg_write_bandwidth;
675 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
676 unsigned long limit = hard_dirty_limit(thresh);
677 unsigned long x_intercept;
678 unsigned long setpoint; /* dirty pages' target balance point */
679 unsigned long bdi_setpoint;
680 unsigned long span;
681 long long pos_ratio; /* for scaling up/down the rate limit */
682 long x;
683
684 if (unlikely(dirty >= limit))
685 return 0;
686
687 /*
688 * global setpoint
689 *
Maxim Patlasov5a537482013-09-11 14:22:46 -0700690 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600691 */
692 setpoint = (freerun + limit) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700693 pos_ratio = pos_ratio_polynom(setpoint, dirty, limit);
694
695 /*
696 * The strictlimit feature is a tool preventing mistrusted filesystems
697 * from growing a large number of dirty pages before throttling. For
698 * such filesystems balance_dirty_pages always checks bdi counters
699 * against bdi limits. Even if global "nr_dirty" is under "freerun".
700 * This is especially important for fuse which sets bdi->max_ratio to
701 * 1% by default. Without strictlimit feature, fuse writeback may
702 * consume arbitrary amount of RAM because it is accounted in
703 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
704 *
705 * Here, in bdi_position_ratio(), we calculate pos_ratio based on
706 * two values: bdi_dirty and bdi_thresh. Let's consider an example:
707 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
708 * limits are set by default to 10% and 20% (background and throttle).
709 * Then bdi_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
710 * bdi_dirty_limit(bdi, bg_thresh) is about ~4K pages. bdi_setpoint is
711 * about ~6K pages (as the average of background and throttle bdi
712 * limits). The 3rd order polynomial will provide positive feedback if
713 * bdi_dirty is under bdi_setpoint and vice versa.
714 *
715 * Note, that we cannot use global counters in these calculations
716 * because we want to throttle process writing to a strictlimit BDI
717 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
718 * in the example above).
719 */
720 if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
721 long long bdi_pos_ratio;
722 unsigned long bdi_bg_thresh;
723
724 if (bdi_dirty < 8)
725 return min_t(long long, pos_ratio * 2,
726 2 << RATELIMIT_CALC_SHIFT);
727
728 if (bdi_dirty >= bdi_thresh)
729 return 0;
730
731 bdi_bg_thresh = div_u64((u64)bdi_thresh * bg_thresh, thresh);
732 bdi_setpoint = dirty_freerun_ceiling(bdi_thresh,
733 bdi_bg_thresh);
734
735 if (bdi_setpoint == 0 || bdi_setpoint == bdi_thresh)
736 return 0;
737
738 bdi_pos_ratio = pos_ratio_polynom(bdi_setpoint, bdi_dirty,
739 bdi_thresh);
740
741 /*
742 * Typically, for strictlimit case, bdi_setpoint << setpoint
743 * and pos_ratio >> bdi_pos_ratio. In the other words global
744 * state ("dirty") is not limiting factor and we have to
745 * make decision based on bdi counters. But there is an
746 * important case when global pos_ratio should get precedence:
747 * global limits are exceeded (e.g. due to activities on other
748 * BDIs) while given strictlimit BDI is below limit.
749 *
750 * "pos_ratio * bdi_pos_ratio" would work for the case above,
751 * but it would look too non-natural for the case of all
752 * activity in the system coming from a single strictlimit BDI
753 * with bdi->max_ratio == 100%.
754 *
755 * Note that min() below somewhat changes the dynamics of the
756 * control system. Normally, pos_ratio value can be well over 3
757 * (when globally we are at freerun and bdi is well below bdi
758 * setpoint). Now the maximum pos_ratio in the same situation
759 * is 2. We might want to tweak this if we observe the control
760 * system is too slow to adapt.
761 */
762 return min(pos_ratio, bdi_pos_ratio);
763 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600764
765 /*
766 * We have computed basic pos_ratio above based on global situation. If
767 * the bdi is over/under its share of dirty pages, we want to scale
768 * pos_ratio further down/up. That is done by the following mechanism.
769 */
770
771 /*
772 * bdi setpoint
773 *
774 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
775 *
776 * x_intercept - bdi_dirty
777 * := --------------------------
778 * x_intercept - bdi_setpoint
779 *
780 * The main bdi control line is a linear function that subjects to
781 *
782 * (1) f(bdi_setpoint) = 1.0
783 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
784 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
785 *
786 * For single bdi case, the dirty pages are observed to fluctuate
787 * regularly within range
788 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
789 * for various filesystems, where (2) can yield in a reasonable 12.5%
790 * fluctuation range for pos_ratio.
791 *
792 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
793 * own size, so move the slope over accordingly and choose a slope that
794 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
795 */
796 if (unlikely(bdi_thresh > thresh))
797 bdi_thresh = thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600798 /*
799 * It's very possible that bdi_thresh is close to 0 not because the
800 * device is slow, but that it has remained inactive for long time.
801 * Honour such devices a reasonable good (hopefully IO efficient)
802 * threshold, so that the occasional writes won't be blocked and active
803 * writes can rampup the threshold quickly.
804 */
Wu Fengguang8927f662011-08-04 22:16:46 -0600805 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600806 /*
807 * scale global setpoint to bdi's:
808 * bdi_setpoint = setpoint * bdi_thresh / thresh
809 */
810 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
811 bdi_setpoint = setpoint * (u64)x >> 16;
812 /*
813 * Use span=(8*write_bw) in single bdi case as indicated by
814 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
815 *
816 * bdi_thresh thresh - bdi_thresh
817 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
818 * thresh thresh
819 */
820 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
821 x_intercept = bdi_setpoint + span;
822
823 if (bdi_dirty < x_intercept - span / 4) {
Rik van Rield5c9fde2014-05-06 12:50:01 -0700824 pos_ratio = div64_u64(pos_ratio * (x_intercept - bdi_dirty),
Wu Fengguang50657fc2011-10-11 17:06:33 -0600825 x_intercept - bdi_setpoint + 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600826 } else
827 pos_ratio /= 4;
828
Wu Fengguang8927f662011-08-04 22:16:46 -0600829 /*
830 * bdi reserve area, safeguard against dirty pool underrun and disk idle
831 * It may push the desired control point of global dirty pages higher
832 * than setpoint.
833 */
834 x_intercept = bdi_thresh / 2;
835 if (bdi_dirty < x_intercept) {
Wu Fengguang50657fc2011-10-11 17:06:33 -0600836 if (bdi_dirty > x_intercept / 8)
837 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
838 else
Wu Fengguang8927f662011-08-04 22:16:46 -0600839 pos_ratio *= 8;
840 }
841
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600842 return pos_ratio;
843}
844
Wu Fengguange98be2d2010-08-29 11:22:30 -0600845static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
846 unsigned long elapsed,
847 unsigned long written)
848{
849 const unsigned long period = roundup_pow_of_two(3 * HZ);
850 unsigned long avg = bdi->avg_write_bandwidth;
851 unsigned long old = bdi->write_bandwidth;
852 u64 bw;
853
854 /*
855 * bw = written * HZ / elapsed
856 *
857 * bw * elapsed + write_bandwidth * (period - elapsed)
858 * write_bandwidth = ---------------------------------------------------
859 * period
860 */
861 bw = written - bdi->written_stamp;
862 bw *= HZ;
863 if (unlikely(elapsed > period)) {
864 do_div(bw, elapsed);
865 avg = bw;
866 goto out;
867 }
868 bw += (u64)bdi->write_bandwidth * (period - elapsed);
869 bw >>= ilog2(period);
870
871 /*
872 * one more level of smoothing, for filtering out sudden spikes
873 */
874 if (avg > old && old >= (unsigned long)bw)
875 avg -= (avg - old) >> 3;
876
877 if (avg < old && old <= (unsigned long)bw)
878 avg += (old - avg) >> 3;
879
880out:
881 bdi->write_bandwidth = bw;
882 bdi->avg_write_bandwidth = avg;
883}
884
Wu Fengguangc42843f2011-03-02 15:54:09 -0600885/*
886 * The global dirtyable memory and dirty threshold could be suddenly knocked
887 * down by a large amount (eg. on the startup of KVM in a swapless system).
888 * This may throw the system into deep dirty exceeded state and throttle
889 * heavy/light dirtiers alike. To retain good responsiveness, maintain
890 * global_dirty_limit for tracking slowly down to the knocked down dirty
891 * threshold.
892 */
893static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
894{
895 unsigned long limit = global_dirty_limit;
896
897 /*
898 * Follow up in one step.
899 */
900 if (limit < thresh) {
901 limit = thresh;
902 goto update;
903 }
904
905 /*
906 * Follow down slowly. Use the higher one as the target, because thresh
907 * may drop below dirty. This is exactly the reason to introduce
908 * global_dirty_limit which is guaranteed to lie above the dirty pages.
909 */
910 thresh = max(thresh, dirty);
911 if (limit > thresh) {
912 limit -= (limit - thresh) >> 5;
913 goto update;
914 }
915 return;
916update:
917 global_dirty_limit = limit;
918}
919
920static void global_update_bandwidth(unsigned long thresh,
921 unsigned long dirty,
922 unsigned long now)
923{
924 static DEFINE_SPINLOCK(dirty_lock);
925 static unsigned long update_time;
926
927 /*
928 * check locklessly first to optimize away locking for the most time
929 */
930 if (time_before(now, update_time + BANDWIDTH_INTERVAL))
931 return;
932
933 spin_lock(&dirty_lock);
934 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
935 update_dirty_limit(thresh, dirty);
936 update_time = now;
937 }
938 spin_unlock(&dirty_lock);
939}
940
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600941/*
942 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
943 *
944 * Normal bdi tasks will be curbed at or below it in long term.
945 * Obviously it should be around (write_bw / N) when there are N dd tasks.
946 */
947static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
948 unsigned long thresh,
949 unsigned long bg_thresh,
950 unsigned long dirty,
951 unsigned long bdi_thresh,
952 unsigned long bdi_dirty,
953 unsigned long dirtied,
954 unsigned long elapsed)
955{
Wu Fengguang73811312011-08-26 15:53:24 -0600956 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
957 unsigned long limit = hard_dirty_limit(thresh);
958 unsigned long setpoint = (freerun + limit) / 2;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600959 unsigned long write_bw = bdi->avg_write_bandwidth;
960 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
961 unsigned long dirty_rate;
962 unsigned long task_ratelimit;
963 unsigned long balanced_dirty_ratelimit;
964 unsigned long pos_ratio;
Wu Fengguang73811312011-08-26 15:53:24 -0600965 unsigned long step;
966 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -0600967
968 /*
969 * The dirty rate will match the writeout rate in long term, except
970 * when dirty pages are truncated by userspace or re-dirtied by FS.
971 */
972 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
973
974 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
975 bdi_thresh, bdi_dirty);
976 /*
977 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
978 */
979 task_ratelimit = (u64)dirty_ratelimit *
980 pos_ratio >> RATELIMIT_CALC_SHIFT;
981 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
982
983 /*
984 * A linear estimation of the "balanced" throttle rate. The theory is,
985 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
986 * dirty_rate will be measured to be (N * task_ratelimit). So the below
987 * formula will yield the balanced rate limit (write_bw / N).
988 *
989 * Note that the expanded form is not a pure rate feedback:
990 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
991 * but also takes pos_ratio into account:
992 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
993 *
994 * (1) is not realistic because pos_ratio also takes part in balancing
995 * the dirty rate. Consider the state
996 * pos_ratio = 0.5 (3)
997 * rate = 2 * (write_bw / N) (4)
998 * If (1) is used, it will stuck in that state! Because each dd will
999 * be throttled at
1000 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1001 * yielding
1002 * dirty_rate = N * task_ratelimit = write_bw (6)
1003 * put (6) into (1) we get
1004 * rate_(i+1) = rate_(i) (7)
1005 *
1006 * So we end up using (2) to always keep
1007 * rate_(i+1) ~= (write_bw / N) (8)
1008 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1009 * pos_ratio is able to drive itself to 1.0, which is not only where
1010 * the dirty count meet the setpoint, but also where the slope of
1011 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1012 */
1013 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1014 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001015 /*
1016 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1017 */
1018 if (unlikely(balanced_dirty_ratelimit > write_bw))
1019 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001020
Wu Fengguang73811312011-08-26 15:53:24 -06001021 /*
1022 * We could safely do this and return immediately:
1023 *
1024 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
1025 *
1026 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001027 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001028 * limit the step size.
1029 *
1030 * The below code essentially only uses the relative value of
1031 *
1032 * task_ratelimit - dirty_ratelimit
1033 * = (pos_ratio - 1) * dirty_ratelimit
1034 *
1035 * which reflects the direction and size of dirty position error.
1036 */
1037
1038 /*
1039 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1040 * task_ratelimit is on the same side of dirty_ratelimit, too.
1041 * For example, when
1042 * - dirty_ratelimit > balanced_dirty_ratelimit
1043 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1044 * lowering dirty_ratelimit will help meet both the position and rate
1045 * control targets. Otherwise, don't update dirty_ratelimit if it will
1046 * only help meet the rate target. After all, what the users ultimately
1047 * feel and care are stable dirty rate and small position error.
1048 *
1049 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001050 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001051 * keeps jumping around randomly and can even leap far away at times
1052 * due to the small 200ms estimation period of dirty_rate (we want to
1053 * keep that period small to reduce time lags).
1054 */
1055 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001056
1057 /*
1058 * For strictlimit case, calculations above were based on bdi counters
1059 * and limits (starting from pos_ratio = bdi_position_ratio() and up to
1060 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1061 * Hence, to calculate "step" properly, we have to use bdi_dirty as
1062 * "dirty" and bdi_setpoint as "setpoint".
1063 *
1064 * We rampup dirty_ratelimit forcibly if bdi_dirty is low because
1065 * it's possible that bdi_thresh is close to zero due to inactivity
1066 * of backing device (see the implementation of bdi_dirty_limit()).
1067 */
1068 if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1069 dirty = bdi_dirty;
1070 if (bdi_dirty < 8)
1071 setpoint = bdi_dirty + 1;
1072 else
1073 setpoint = (bdi_thresh +
1074 bdi_dirty_limit(bdi, bg_thresh)) / 2;
1075 }
1076
Wu Fengguang73811312011-08-26 15:53:24 -06001077 if (dirty < setpoint) {
Mark Rustad7c809962014-10-09 15:28:15 -07001078 x = min3(bdi->balanced_dirty_ratelimit,
1079 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001080 if (dirty_ratelimit < x)
1081 step = x - dirty_ratelimit;
1082 } else {
Mark Rustad7c809962014-10-09 15:28:15 -07001083 x = max3(bdi->balanced_dirty_ratelimit,
1084 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001085 if (dirty_ratelimit > x)
1086 step = dirty_ratelimit - x;
1087 }
1088
1089 /*
1090 * Don't pursue 100% rate matching. It's impossible since the balanced
1091 * rate itself is constantly fluctuating. So decrease the track speed
1092 * when it gets close to the target. Helps eliminate pointless tremors.
1093 */
1094 step >>= dirty_ratelimit / (2 * step + 1);
1095 /*
1096 * Limit the tracking speed to avoid overshooting.
1097 */
1098 step = (step + 7) / 8;
1099
1100 if (dirty_ratelimit < balanced_dirty_ratelimit)
1101 dirty_ratelimit += step;
1102 else
1103 dirty_ratelimit -= step;
1104
1105 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1106 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001107
1108 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001109}
1110
Wu Fengguange98be2d2010-08-29 11:22:30 -06001111void __bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001112 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001113 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001114 unsigned long dirty,
1115 unsigned long bdi_thresh,
1116 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -06001117 unsigned long start_time)
1118{
1119 unsigned long now = jiffies;
1120 unsigned long elapsed = now - bdi->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001121 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001122 unsigned long written;
1123
1124 /*
1125 * rate-limit, only update once every 200ms.
1126 */
1127 if (elapsed < BANDWIDTH_INTERVAL)
1128 return;
1129
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001130 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001131 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1132
1133 /*
1134 * Skip quiet periods when disk bandwidth is under-utilized.
1135 * (at least 1s idle time between two flusher runs)
1136 */
1137 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1138 goto snapshot;
1139
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001140 if (thresh) {
Wu Fengguangc42843f2011-03-02 15:54:09 -06001141 global_update_bandwidth(thresh, dirty, now);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001142 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1143 bdi_thresh, bdi_dirty,
1144 dirtied, elapsed);
1145 }
Wu Fengguange98be2d2010-08-29 11:22:30 -06001146 bdi_update_write_bandwidth(bdi, elapsed, written);
1147
1148snapshot:
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001149 bdi->dirtied_stamp = dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001150 bdi->written_stamp = written;
1151 bdi->bw_time_stamp = now;
1152}
1153
1154static void bdi_update_bandwidth(struct backing_dev_info *bdi,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001155 unsigned long thresh,
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001156 unsigned long bg_thresh,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001157 unsigned long dirty,
1158 unsigned long bdi_thresh,
1159 unsigned long bdi_dirty,
Wu Fengguange98be2d2010-08-29 11:22:30 -06001160 unsigned long start_time)
1161{
1162 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1163 return;
1164 spin_lock(&bdi->wb.list_lock);
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001165 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1166 bdi_thresh, bdi_dirty, start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001167 spin_unlock(&bdi->wb.list_lock);
1168}
1169
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001171 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001172 * will look to see if it needs to start dirty throttling.
1173 *
1174 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1175 * global_page_state() too often. So scale it near-sqrt to the safety margin
1176 * (the number of pages we may dirty without exceeding the dirty limits).
1177 */
1178static unsigned long dirty_poll_interval(unsigned long dirty,
1179 unsigned long thresh)
1180{
1181 if (thresh > dirty)
1182 return 1UL << (ilog2(thresh - dirty) >> 1);
1183
1184 return 1;
1185}
1186
Fengguang Wue3b6c652013-10-16 13:47:03 -07001187static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1188 unsigned long bdi_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001189{
Fengguang Wue3b6c652013-10-16 13:47:03 -07001190 unsigned long bw = bdi->avg_write_bandwidth;
1191 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001192
1193 /*
1194 * Limit pause time for small memory systems. If sleeping for too long
1195 * time, a small pool of dirty/writeback pages may go empty and disk go
1196 * idle.
1197 *
1198 * 8 serves as the safety ratio.
1199 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001200 t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1201 t++;
1202
Fengguang Wue3b6c652013-10-16 13:47:03 -07001203 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001204}
1205
1206static long bdi_min_pause(struct backing_dev_info *bdi,
1207 long max_pause,
1208 unsigned long task_ratelimit,
1209 unsigned long dirty_ratelimit,
1210 int *nr_dirtied_pause)
1211{
1212 long hi = ilog2(bdi->avg_write_bandwidth);
1213 long lo = ilog2(bdi->dirty_ratelimit);
1214 long t; /* target pause */
1215 long pause; /* estimated next pause */
1216 int pages; /* target nr_dirtied_pause */
1217
1218 /* target for 10ms pause on 1-dd case */
1219 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001220
1221 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001222 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1223 * overheads.
1224 *
1225 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001226 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001227 if (hi > lo)
1228 t += (hi - lo) * (10 * HZ) / 1024;
1229
1230 /*
1231 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1232 * on the much more stable dirty_ratelimit. However the next pause time
1233 * will be computed based on task_ratelimit and the two rate limits may
1234 * depart considerably at some time. Especially if task_ratelimit goes
1235 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1236 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1237 * result task_ratelimit won't be executed faithfully, which could
1238 * eventually bring down dirty_ratelimit.
1239 *
1240 * We apply two rules to fix it up:
1241 * 1) try to estimate the next pause time and if necessary, use a lower
1242 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1243 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1244 * 2) limit the target pause time to max_pause/2, so that the normal
1245 * small fluctuations of task_ratelimit won't trigger rule (1) and
1246 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1247 */
1248 t = min(t, 1 + max_pause / 2);
1249 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1250
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001251 /*
1252 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1253 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1254 * When the 16 consecutive reads are often interrupted by some dirty
1255 * throttling pause during the async writes, cfq will go into idles
1256 * (deadline is fine). So push nr_dirtied_pause as high as possible
1257 * until reaches DIRTY_POLL_THRESH=32 pages.
1258 */
1259 if (pages < DIRTY_POLL_THRESH) {
1260 t = max_pause;
1261 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1262 if (pages > DIRTY_POLL_THRESH) {
1263 pages = DIRTY_POLL_THRESH;
1264 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1265 }
1266 }
1267
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001268 pause = HZ * pages / (task_ratelimit + 1);
1269 if (pause > max_pause) {
1270 t = max_pause;
1271 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1272 }
1273
1274 *nr_dirtied_pause = pages;
1275 /*
1276 * The minimal pause time will normally be half the target pause time.
1277 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001278 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001279}
1280
Maxim Patlasov5a537482013-09-11 14:22:46 -07001281static inline void bdi_dirty_limits(struct backing_dev_info *bdi,
1282 unsigned long dirty_thresh,
1283 unsigned long background_thresh,
1284 unsigned long *bdi_dirty,
1285 unsigned long *bdi_thresh,
1286 unsigned long *bdi_bg_thresh)
1287{
1288 unsigned long bdi_reclaimable;
1289
1290 /*
1291 * bdi_thresh is not treated as some limiting factor as
1292 * dirty_thresh, due to reasons
1293 * - in JBOD setup, bdi_thresh can fluctuate a lot
1294 * - in a system with HDD and USB key, the USB key may somehow
1295 * go into state (bdi_dirty >> bdi_thresh) either because
1296 * bdi_dirty starts high, or because bdi_thresh drops low.
1297 * In this case we don't want to hard throttle the USB key
1298 * dirtiers for 100 seconds until bdi_dirty drops under
1299 * bdi_thresh. Instead the auxiliary bdi control line in
1300 * bdi_position_ratio() will let the dirtier task progress
1301 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1302 */
1303 *bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1304
1305 if (bdi_bg_thresh)
Maxim Patlasovf6789592014-07-30 16:08:21 -07001306 *bdi_bg_thresh = dirty_thresh ? div_u64((u64)*bdi_thresh *
1307 background_thresh,
1308 dirty_thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001309
1310 /*
1311 * In order to avoid the stacked BDI deadlock we need
1312 * to ensure we accurately count the 'dirty' pages when
1313 * the threshold is low.
1314 *
1315 * Otherwise it would be possible to get thresh+n pages
1316 * reported dirty, even though there are thresh-m pages
1317 * actually dirty; with m+n sitting in the percpu
1318 * deltas.
1319 */
1320 if (*bdi_thresh < 2 * bdi_stat_error(bdi)) {
1321 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1322 *bdi_dirty = bdi_reclaimable +
1323 bdi_stat_sum(bdi, BDI_WRITEBACK);
1324 } else {
1325 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1326 *bdi_dirty = bdi_reclaimable +
1327 bdi_stat(bdi, BDI_WRITEBACK);
1328 }
1329}
1330
Wu Fengguang9d823e82011-06-11 18:10:12 -06001331/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 * balance_dirty_pages() must be called by processes which are generating dirty
1333 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001334 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001335 * If we're over `background_thresh' then the writeback threads are woken to
1336 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001338static void balance_dirty_pages(struct address_space *mapping,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001339 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340{
Wu Fengguang143dfe82010-08-27 18:45:12 -06001341 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
Wu Fengguang77627412010-09-12 13:34:05 -06001342 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
David Rientjes364aeb22009-01-06 14:39:29 -08001343 unsigned long background_thresh;
1344 unsigned long dirty_thresh;
Wu Fengguang83712352011-06-11 19:25:42 -06001345 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001346 long pause;
1347 long max_pause;
1348 long min_pause;
1349 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001350 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001351 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001352 unsigned long dirty_ratelimit;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001353 unsigned long pos_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 struct backing_dev_info *bdi = mapping->backing_dev_info;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001355 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001356 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357
1358 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001359 unsigned long now = jiffies;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001360 unsigned long uninitialized_var(bdi_thresh);
1361 unsigned long thresh;
1362 unsigned long uninitialized_var(bdi_dirty);
1363 unsigned long dirty;
1364 unsigned long bg_thresh;
Wu Fengguang83712352011-06-11 19:25:42 -06001365
Wu Fengguang143dfe82010-08-27 18:45:12 -06001366 /*
1367 * Unstable writes are a feature of certain networked
1368 * filesystems (i.e. NFS) in which data may have been
1369 * written to the server's write cache, but has not yet
1370 * been flushed to permanent storage.
1371 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001372 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1373 global_page_state(NR_UNSTABLE_NFS);
Wu Fengguang77627412010-09-12 13:34:05 -06001374 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001375
Wu Fengguang16c40422010-08-11 14:17:39 -07001376 global_dirty_limits(&background_thresh, &dirty_thresh);
1377
Maxim Patlasov5a537482013-09-11 14:22:46 -07001378 if (unlikely(strictlimit)) {
1379 bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1380 &bdi_dirty, &bdi_thresh, &bg_thresh);
1381
1382 dirty = bdi_dirty;
1383 thresh = bdi_thresh;
1384 } else {
1385 dirty = nr_dirty;
1386 thresh = dirty_thresh;
1387 bg_thresh = background_thresh;
1388 }
1389
Wu Fengguang16c40422010-08-11 14:17:39 -07001390 /*
1391 * Throttle it only when the background writeback cannot
1392 * catch-up. This avoids (excessively) small writeouts
Maxim Patlasov5a537482013-09-11 14:22:46 -07001393 * when the bdi limits are ramping up in case of !strictlimit.
1394 *
1395 * In strictlimit case make decision based on the bdi counters
1396 * and limits. Small writeouts when the bdi limits are ramping
1397 * up are the price we consciously pay for strictlimit-ing.
Wu Fengguang16c40422010-08-11 14:17:39 -07001398 */
Maxim Patlasov5a537482013-09-11 14:22:46 -07001399 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001400 current->dirty_paused_when = now;
1401 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001402 current->nr_dirtied_pause =
Maxim Patlasov5a537482013-09-11 14:22:46 -07001403 dirty_poll_interval(dirty, thresh);
Wu Fengguang16c40422010-08-11 14:17:39 -07001404 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001405 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001406
Wu Fengguang143dfe82010-08-27 18:45:12 -06001407 if (unlikely(!writeback_in_progress(bdi)))
1408 bdi_start_background_writeback(bdi);
1409
Maxim Patlasov5a537482013-09-11 14:22:46 -07001410 if (!strictlimit)
1411 bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1412 &bdi_dirty, &bdi_thresh, NULL);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001413
Wu Fengguang82791942011-12-03 21:26:01 -06001414 dirty_exceeded = (bdi_dirty > bdi_thresh) &&
Maxim Patlasov5a537482013-09-11 14:22:46 -07001415 ((nr_dirty > dirty_thresh) || strictlimit);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001416 if (dirty_exceeded && !bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001417 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
Wu Fengguangaf6a3112011-10-03 20:46:17 -06001419 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1420 nr_dirty, bdi_thresh, bdi_dirty,
1421 start_time);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001422
Wu Fengguang143dfe82010-08-27 18:45:12 -06001423 dirty_ratelimit = bdi->dirty_ratelimit;
1424 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1425 background_thresh, nr_dirty,
1426 bdi_thresh, bdi_dirty);
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001427 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1428 RATELIMIT_CALC_SHIFT;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001429 max_pause = bdi_max_pause(bdi, bdi_dirty);
1430 min_pause = bdi_min_pause(bdi, max_pause,
1431 task_ratelimit, dirty_ratelimit,
1432 &nr_dirtied_pause);
1433
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001434 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001435 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001436 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001437 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 }
Wu Fengguang83712352011-06-11 19:25:42 -06001439 period = HZ * pages_dirtied / task_ratelimit;
1440 pause = period;
1441 if (current->dirty_paused_when)
1442 pause -= now - current->dirty_paused_when;
1443 /*
1444 * For less than 1s think time (ext3/4 may block the dirtier
1445 * for up to 800ms from time to time on 1-HDD; so does xfs,
1446 * however at much less frequency), try to compensate it in
1447 * future periods by updating the virtual time; otherwise just
1448 * do a reset, as it may be a light dirtier.
1449 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001450 if (pause < min_pause) {
Wu Fengguangece13ac2010-08-29 23:33:20 -06001451 trace_balance_dirty_pages(bdi,
1452 dirty_thresh,
1453 background_thresh,
1454 nr_dirty,
1455 bdi_thresh,
1456 bdi_dirty,
1457 dirty_ratelimit,
1458 task_ratelimit,
1459 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001460 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001461 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001462 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001463 if (pause < -HZ) {
1464 current->dirty_paused_when = now;
1465 current->nr_dirtied = 0;
1466 } else if (period) {
1467 current->dirty_paused_when += period;
1468 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001469 } else if (current->nr_dirtied_pause <= pages_dirtied)
1470 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001471 break;
1472 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001473 if (unlikely(pause > max_pause)) {
1474 /* for occasional dropped task_ratelimit */
1475 now += min(pause - max_pause, max_pause);
1476 pause = max_pause;
1477 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001478
1479pause:
Wu Fengguangece13ac2010-08-29 23:33:20 -06001480 trace_balance_dirty_pages(bdi,
1481 dirty_thresh,
1482 background_thresh,
1483 nr_dirty,
1484 bdi_thresh,
1485 bdi_dirty,
1486 dirty_ratelimit,
1487 task_ratelimit,
1488 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001489 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001490 pause,
1491 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001492 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001493 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001494
Wu Fengguang83712352011-06-11 19:25:42 -06001495 current->dirty_paused_when = now + pause;
1496 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001497 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001498
Wu Fengguangffd1f602011-06-19 22:18:42 -06001499 /*
Wu Fengguang1df64712011-11-13 19:47:32 -06001500 * This is typically equal to (nr_dirty < dirty_thresh) and can
1501 * also keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001502 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001503 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001504 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001505
Wu Fengguangc5c63432011-12-02 10:21:33 -06001506 /*
1507 * In the case of an unresponding NFS server and the NFS dirty
1508 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1509 * to go through, so that tasks on them still remain responsive.
1510 *
1511 * In theory 1 page is enough to keep the comsumer-producer
1512 * pipe going: the flusher cleans 1 page => the task dirties 1
1513 * more page. However bdi_dirty has accounting errors. So use
1514 * the larger and more IO friendly bdi_stat_error.
1515 */
1516 if (bdi_dirty <= bdi_stat_error(bdi))
1517 break;
1518
Jan Kara499d05e2011-11-16 19:34:48 +08001519 if (fatal_signal_pending(current))
1520 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 }
1522
Wu Fengguang143dfe82010-08-27 18:45:12 -06001523 if (!dirty_exceeded && bdi->dirty_exceeded)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001524 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525
1526 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001527 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528
1529 /*
1530 * In laptop mode, we wait until hitting the higher threshold before
1531 * starting background writeout, and then write out all the way down
1532 * to the lower threshold. So slow writers cause minimal disk activity.
1533 *
1534 * In normal mode, we start background writeout at the lower
1535 * background_thresh, to keep the amount of dirty memory low.
1536 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001537 if (laptop_mode)
1538 return;
1539
1540 if (nr_reclaimable > background_thresh)
Christoph Hellwigc5444192010-06-08 18:15:15 +02001541 bdi_start_background_writeback(bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542}
1543
Wu Fengguang9d823e82011-06-11 18:10:12 -06001544static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001545
Wu Fengguang54848d72011-04-05 13:21:19 -06001546/*
1547 * Normal tasks are throttled by
1548 * loop {
1549 * dirty tsk->nr_dirtied_pause pages;
1550 * take a snap in balance_dirty_pages();
1551 * }
1552 * However there is a worst case. If every task exit immediately when dirtied
1553 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1554 * called to throttle the page dirties. The solution is to save the not yet
1555 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1556 * randomly into the running tasks. This works well for the above worst case,
1557 * as the new task will pick up and accumulate the old task's leaked dirty
1558 * count and eventually get throttled.
1559 */
1560DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1561
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001563 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001564 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 *
1566 * Processes which are dirtying memory should call in here once for each page
1567 * which was newly dirtied. The function will periodically check the system's
1568 * dirty state and will initiate writeback if needed.
1569 *
1570 * On really big machines, get_writeback_state is expensive, so try to avoid
1571 * calling it too often (ratelimiting). But once we're over the dirty memory
1572 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1573 * from overshooting the limit by (ratelimit_pages) each.
1574 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001575void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576{
Wu Fengguang36715ce2011-06-11 17:53:57 -06001577 struct backing_dev_info *bdi = mapping->backing_dev_info;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001578 int ratelimit;
1579 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580
Wu Fengguang36715ce2011-06-11 17:53:57 -06001581 if (!bdi_cap_account_dirty(bdi))
1582 return;
1583
Wu Fengguang9d823e82011-06-11 18:10:12 -06001584 ratelimit = current->nr_dirtied_pause;
1585 if (bdi->dirty_exceeded)
1586 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001588 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001589 /*
1590 * This prevents one CPU to accumulate too many dirtied pages without
1591 * calling into balance_dirty_pages(), which can happen when there are
1592 * 1000+ tasks, all of them start dirtying pages at exactly the same
1593 * time, hence all honoured too large initial task->nr_dirtied_pause.
1594 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001595 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001596 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001597 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001598 else if (unlikely(*p >= ratelimit_pages)) {
1599 *p = 0;
1600 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001602 /*
1603 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1604 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1605 * the dirty throttling and livelock other long-run dirtiers.
1606 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001607 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06001608 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001609 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001610 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1611 *p -= nr_pages_dirtied;
1612 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001614 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001615
1616 if (unlikely(current->nr_dirtied >= ratelimit))
1617 balance_dirty_pages(mapping, current->nr_dirtied);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001619EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620
Andrew Morton232ea4d2007-02-28 20:13:21 -08001621void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622{
David Rientjes364aeb22009-01-06 14:39:29 -08001623 unsigned long background_thresh;
1624 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625
1626 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001627 global_dirty_limits(&background_thresh, &dirty_thresh);
Fengguang Wu47a13332012-03-21 16:34:09 -07001628 dirty_thresh = hard_dirty_limit(dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629
1630 /*
1631 * Boost the allowable dirty threshold a bit for page
1632 * allocators so they don't get DoS'ed by heavy writers
1633 */
1634 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1635
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001636 if (global_page_state(NR_UNSTABLE_NFS) +
1637 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1638 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001639 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001640
1641 /*
1642 * The caller might hold locks which can prevent IO completion
1643 * or progress in the filesystem. So we cannot just sit here
1644 * waiting for IO to complete.
1645 */
1646 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1647 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 }
1649}
1650
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1653 */
Joe Perchescccad5b2014-06-06 14:38:09 -07001654int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001655 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001657 proc_dointvec(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 return 0;
1659}
1660
Jens Axboec2c49862010-05-20 09:18:47 +02001661#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001662void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663{
Matthew Garrett31373d02010-04-06 14:25:14 +02001664 struct request_queue *q = (struct request_queue *)data;
1665 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1666 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667
Matthew Garrett31373d02010-04-06 14:25:14 +02001668 /*
1669 * We want to write everything out, not just down to the dirty
1670 * threshold
1671 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001672 if (bdi_has_dirty_io(&q->backing_dev_info))
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06001673 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1674 WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675}
1676
1677/*
1678 * We've spun up the disk and we're in laptop mode: schedule writeback
1679 * of all dirty data a few seconds from now. If the flush is already scheduled
1680 * then push it back - the user is still using the disk.
1681 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001682void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
Matthew Garrett31373d02010-04-06 14:25:14 +02001684 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685}
1686
1687/*
1688 * We're in laptop mode and we've just synced. The sync's writes will have
1689 * caused another writeback to be scheduled by laptop_io_completion.
1690 * Nothing needs to be written back anymore, so we unschedule the writeback.
1691 */
1692void laptop_sync_completion(void)
1693{
Matthew Garrett31373d02010-04-06 14:25:14 +02001694 struct backing_dev_info *bdi;
1695
1696 rcu_read_lock();
1697
1698 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1699 del_timer(&bdi->laptop_mode_wb_timer);
1700
1701 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702}
Jens Axboec2c49862010-05-20 09:18:47 +02001703#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704
1705/*
1706 * If ratelimit_pages is too high then we can get into dirty-data overload
1707 * if a large number of processes all perform writes at the same time.
1708 * If it is too low then SMP machines will call the (expensive)
1709 * get_writeback_state too often.
1710 *
1711 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1712 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06001713 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 */
1715
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001716void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717{
Wu Fengguang9d823e82011-06-11 18:10:12 -06001718 unsigned long background_thresh;
1719 unsigned long dirty_thresh;
1720 global_dirty_limits(&background_thresh, &dirty_thresh);
Fengguang Wu68809c72012-05-06 13:21:42 +08001721 global_dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001722 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 if (ratelimit_pages < 16)
1724 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725}
1726
Paul Gortmaker0db06282013-06-19 14:53:51 -04001727static int
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08001728ratelimit_handler(struct notifier_block *self, unsigned long action,
1729 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730{
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08001731
1732 switch (action & ~CPU_TASKS_FROZEN) {
1733 case CPU_ONLINE:
1734 case CPU_DEAD:
1735 writeback_set_ratelimit();
1736 return NOTIFY_OK;
1737 default:
1738 return NOTIFY_DONE;
1739 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740}
1741
Paul Gortmaker0db06282013-06-19 14:53:51 -04001742static struct notifier_block ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 .notifier_call = ratelimit_handler,
1744 .next = NULL,
1745};
1746
1747/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08001748 * Called early on to tune the page writeback dirty limits.
1749 *
1750 * We used to scale dirty pages according to how total memory
1751 * related to pages that could be allocated for buffers (by
1752 * comparing nr_free_buffer_pages() to vm_total_pages.
1753 *
1754 * However, that was when we used "dirty_ratio" to scale with
1755 * all memory, and we don't do that any more. "dirty_ratio"
1756 * is now applied to total non-HIGHPAGE memory (by subtracting
1757 * totalhigh_pages from vm_total_pages), and as such we can't
1758 * get into the old insane situation any more where we had
1759 * large amounts of dirty pages compared to a small amount of
1760 * non-HIGHMEM memory.
1761 *
1762 * But we might still want to scale the dirty_ratio by how
1763 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 */
1765void __init page_writeback_init(void)
1766{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07001767 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001769
Tejun Heo20ae0072014-09-08 09:51:30 +09001770 fprop_global_init(&writeout_completions, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771}
1772
David Howells811d7362006-08-29 19:06:09 +01001773/**
Jan Karaf446daae2010-08-09 17:19:12 -07001774 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1775 * @mapping: address space structure to write
1776 * @start: starting page index
1777 * @end: ending page index (inclusive)
1778 *
1779 * This function scans the page range from @start to @end (inclusive) and tags
1780 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1781 * that write_cache_pages (or whoever calls this function) will then use
1782 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
1783 * used to avoid livelocking of writeback by a process steadily creating new
1784 * dirty pages in the file (thus it is important for this function to be quick
1785 * so that it can tag pages faster than a dirtying process can create them).
1786 */
1787/*
1788 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1789 */
Jan Karaf446daae2010-08-09 17:19:12 -07001790void tag_pages_for_writeback(struct address_space *mapping,
1791 pgoff_t start, pgoff_t end)
1792{
Randy Dunlap3c111a02010-08-11 14:17:30 -07001793#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07001794 unsigned long tagged;
1795
1796 do {
1797 spin_lock_irq(&mapping->tree_lock);
1798 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1799 &start, end, WRITEBACK_TAG_BATCH,
1800 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1801 spin_unlock_irq(&mapping->tree_lock);
1802 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1803 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07001804 /* We check 'start' to handle wrapping when end == ~0UL */
1805 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07001806}
1807EXPORT_SYMBOL(tag_pages_for_writeback);
1808
1809/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001810 * 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 +01001811 * @mapping: address space structure to write
1812 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001813 * @writepage: function called for each page
1814 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01001815 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001816 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01001817 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
1818 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
1819 * and msync() need to guarantee that all the data which was dirty at the time
1820 * the call was made get new I/O started against them. If wbc->sync_mode is
1821 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1822 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07001823 *
1824 * To avoid livelocks (when other process dirties new pages), we first tag
1825 * pages which should be written back with TOWRITE tag and only then start
1826 * writing them. For data-integrity sync we have to be careful so that we do
1827 * not miss some pages (e.g., because some other process has cleared TOWRITE
1828 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1829 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01001830 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001831int write_cache_pages(struct address_space *mapping,
1832 struct writeback_control *wbc, writepage_t writepage,
1833 void *data)
David Howells811d7362006-08-29 19:06:09 +01001834{
David Howells811d7362006-08-29 19:06:09 +01001835 int ret = 0;
1836 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01001837 struct pagevec pvec;
1838 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08001839 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01001840 pgoff_t index;
1841 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08001842 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08001843 int cycled;
David Howells811d7362006-08-29 19:06:09 +01001844 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07001845 int tag;
David Howells811d7362006-08-29 19:06:09 +01001846
David Howells811d7362006-08-29 19:06:09 +01001847 pagevec_init(&pvec, 0);
1848 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08001849 writeback_index = mapping->writeback_index; /* prev offset */
1850 index = writeback_index;
1851 if (index == 0)
1852 cycled = 1;
1853 else
1854 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01001855 end = -1;
1856 } else {
1857 index = wbc->range_start >> PAGE_CACHE_SHIFT;
1858 end = wbc->range_end >> PAGE_CACHE_SHIFT;
1859 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1860 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08001861 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01001862 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001863 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001864 tag = PAGECACHE_TAG_TOWRITE;
1865 else
1866 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01001867retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06001868 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07001869 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08001870 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001871 while (!done && (index <= end)) {
1872 int i;
1873
Jan Karaf446daae2010-08-09 17:19:12 -07001874 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001875 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1876 if (nr_pages == 0)
1877 break;
David Howells811d7362006-08-29 19:06:09 +01001878
David Howells811d7362006-08-29 19:06:09 +01001879 for (i = 0; i < nr_pages; i++) {
1880 struct page *page = pvec.pages[i];
1881
Nick Piggind5482cd2009-01-06 14:39:11 -08001882 /*
1883 * At this point, the page may be truncated or
1884 * invalidated (changing page->mapping to NULL), or
1885 * even swizzled back from swapper_space to tmpfs file
1886 * mapping. However, page->index will not change
1887 * because we have a reference on the page.
1888 */
1889 if (page->index > end) {
1890 /*
1891 * can't be range_cyclic (1st pass) because
1892 * end == -1 in that case.
1893 */
1894 done = 1;
1895 break;
1896 }
1897
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001898 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08001899
David Howells811d7362006-08-29 19:06:09 +01001900 lock_page(page);
1901
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001902 /*
1903 * Page truncated or invalidated. We can freely skip it
1904 * then, even for data integrity operations: the page
1905 * has disappeared concurrently, so there could be no
1906 * real expectation of this data interity operation
1907 * even if there is now a new, dirty page at the same
1908 * pagecache address.
1909 */
David Howells811d7362006-08-29 19:06:09 +01001910 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001911continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01001912 unlock_page(page);
1913 continue;
1914 }
1915
Nick Piggin515f4a02009-01-06 14:39:10 -08001916 if (!PageDirty(page)) {
1917 /* someone wrote it for us */
1918 goto continue_unlock;
1919 }
David Howells811d7362006-08-29 19:06:09 +01001920
Nick Piggin515f4a02009-01-06 14:39:10 -08001921 if (PageWriteback(page)) {
1922 if (wbc->sync_mode != WB_SYNC_NONE)
1923 wait_on_page_writeback(page);
1924 else
1925 goto continue_unlock;
1926 }
1927
1928 BUG_ON(PageWriteback(page));
1929 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08001930 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01001931
Dave Chinner9e094382010-07-07 13:24:08 +10001932 trace_wbc_writepage(wbc, mapping->backing_dev_info);
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001933 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08001934 if (unlikely(ret)) {
1935 if (ret == AOP_WRITEPAGE_ACTIVATE) {
1936 unlock_page(page);
1937 ret = 0;
1938 } else {
1939 /*
1940 * done_index is set past this page,
1941 * so media errors will not choke
1942 * background writeout for the entire
1943 * file. This has consequences for
1944 * range_cyclic semantics (ie. it may
1945 * not be suitable for data integrity
1946 * writeout).
1947 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07001948 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08001949 done = 1;
1950 break;
1951 }
Dave Chinner0b564922010-06-09 10:37:18 +10001952 }
David Howells811d7362006-08-29 19:06:09 +01001953
Dave Chinner546a1922010-08-24 11:44:34 +10001954 /*
1955 * We stop writing back only if we are not doing
1956 * integrity sync. In case of integrity sync we have to
1957 * keep going until we have written all the pages
1958 * we tagged for writeback prior to entering this loop.
1959 */
1960 if (--wbc->nr_to_write <= 0 &&
1961 wbc->sync_mode == WB_SYNC_NONE) {
1962 done = 1;
1963 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08001964 }
David Howells811d7362006-08-29 19:06:09 +01001965 }
1966 pagevec_release(&pvec);
1967 cond_resched();
1968 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01001969 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01001970 /*
Nick Piggin31a12662009-01-06 14:39:04 -08001971 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01001972 * We hit the last page and there is more work to be done: wrap
1973 * back to the start of the file
1974 */
Nick Piggin31a12662009-01-06 14:39:04 -08001975 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01001976 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08001977 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01001978 goto retry;
1979 }
Dave Chinner0b564922010-06-09 10:37:18 +10001980 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1981 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04001982
David Howells811d7362006-08-29 19:06:09 +01001983 return ret;
1984}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07001985EXPORT_SYMBOL(write_cache_pages);
1986
1987/*
1988 * Function used by generic_writepages to call the real writepage
1989 * function and set the mapping flags on error
1990 */
1991static int __writepage(struct page *page, struct writeback_control *wbc,
1992 void *data)
1993{
1994 struct address_space *mapping = data;
1995 int ret = mapping->a_ops->writepage(page, wbc);
1996 mapping_set_error(mapping, ret);
1997 return ret;
1998}
1999
2000/**
2001 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2002 * @mapping: address space structure to write
2003 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2004 *
2005 * This is a library function, which implements the writepages()
2006 * address_space_operation.
2007 */
2008int generic_writepages(struct address_space *mapping,
2009 struct writeback_control *wbc)
2010{
Shaohua Li9b6096a2011-03-17 10:47:06 +01002011 struct blk_plug plug;
2012 int ret;
2013
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002014 /* deal with chardevs and other special file */
2015 if (!mapping->a_ops->writepage)
2016 return 0;
2017
Shaohua Li9b6096a2011-03-17 10:47:06 +01002018 blk_start_plug(&plug);
2019 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2020 blk_finish_plug(&plug);
2021 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002022}
David Howells811d7362006-08-29 19:06:09 +01002023
2024EXPORT_SYMBOL(generic_writepages);
2025
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2027{
Andrew Morton22905f72005-11-16 15:07:01 -08002028 int ret;
2029
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 if (wbc->nr_to_write <= 0)
2031 return 0;
2032 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002033 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002034 else
2035 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002036 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037}
2038
2039/**
2040 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07002041 * @page: the page to write
2042 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 *
2044 * The page must be locked by the caller and will be unlocked upon return.
2045 *
2046 * write_one_page() returns a negative error code if I/O failed.
2047 */
2048int write_one_page(struct page *page, int wait)
2049{
2050 struct address_space *mapping = page->mapping;
2051 int ret = 0;
2052 struct writeback_control wbc = {
2053 .sync_mode = WB_SYNC_ALL,
2054 .nr_to_write = 1,
2055 };
2056
2057 BUG_ON(!PageLocked(page));
2058
2059 if (wait)
2060 wait_on_page_writeback(page);
2061
2062 if (clear_page_dirty_for_io(page)) {
2063 page_cache_get(page);
2064 ret = mapping->a_ops->writepage(page, &wbc);
2065 if (ret == 0 && wait) {
2066 wait_on_page_writeback(page);
2067 if (PageError(page))
2068 ret = -EIO;
2069 }
2070 page_cache_release(page);
2071 } else {
2072 unlock_page(page);
2073 }
2074 return ret;
2075}
2076EXPORT_SYMBOL(write_one_page);
2077
2078/*
Ken Chen76719322007-02-10 01:43:15 -08002079 * For address_spaces which do not use buffers nor write back.
2080 */
2081int __set_page_dirty_no_writeback(struct page *page)
2082{
2083 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08002084 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08002085 return 0;
2086}
2087
2088/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002089 * Helper function for set_page_dirty family.
2090 * NOTE: This relies on being atomic wrt interrupts.
2091 */
2092void account_page_dirtied(struct page *page, struct address_space *mapping)
2093{
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002094 trace_writeback_dirty_page(page, mapping);
2095
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002096 if (mapping_cap_account_dirty(mapping)) {
2097 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07002098 __inc_zone_page_state(page, NR_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002099 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
Wu Fengguangc8e28ce2011-01-23 10:07:47 -06002100 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002101 task_io_account_write(PAGE_CACHE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06002102 current->nr_dirtied++;
2103 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002104 }
2105}
Michael Rubin679ceac2010-08-20 02:31:26 -07002106EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002107
2108/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 * For address_spaces which do not use buffers. Just tag the page as dirty in
2110 * its radix tree.
2111 *
2112 * This is also used when a single buffer is being dirtied: we want to set the
2113 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2114 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2115 *
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002116 * The caller must ensure this doesn't race with truncation. Most will simply
2117 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2118 * the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 */
2120int __set_page_dirty_nobuffers(struct page *page)
2121{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 if (!TestSetPageDirty(page)) {
2123 struct address_space *mapping = page_mapping(page);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002124 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125
Andrew Morton8c085402006-12-10 02:19:24 -08002126 if (!mapping)
2127 return 1;
2128
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002129 spin_lock_irqsave(&mapping->tree_lock, flags);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002130 BUG_ON(page_mapping(page) != mapping);
2131 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2132 account_page_dirtied(page, mapping);
2133 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2134 PAGECACHE_TAG_DIRTY);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002135 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Andrew Morton8c085402006-12-10 02:19:24 -08002136 if (mapping->host) {
2137 /* !PageAnon && !swapper_space */
2138 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002140 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002142 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143}
2144EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2145
2146/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002147 * Call this whenever redirtying a page, to de-account the dirty counters
2148 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2149 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2150 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2151 * control.
2152 */
2153void account_page_redirty(struct page *page)
2154{
2155 struct address_space *mapping = page->mapping;
2156 if (mapping && mapping_cap_account_dirty(mapping)) {
2157 current->nr_dirtied--;
2158 dec_zone_page_state(page, NR_DIRTIED);
2159 dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2160 }
2161}
2162EXPORT_SYMBOL(account_page_redirty);
2163
2164/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 * When a writepage implementation decides that it doesn't want to write this
2166 * page for some reason, it should redirty the locked page via
2167 * redirty_page_for_writepage() and it should then unlock the page and return 0
2168 */
2169int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2170{
2171 wbc->pages_skipped++;
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002172 account_page_redirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 return __set_page_dirty_nobuffers(page);
2174}
2175EXPORT_SYMBOL(redirty_page_for_writepage);
2176
2177/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002178 * Dirty a page.
2179 *
2180 * For pages with a mapping this should be done under the page lock
2181 * for the benefit of asynchronous memory errors who prefer a consistent
2182 * dirty state. This rule can be broken in some special cases,
2183 * but should be better not to.
2184 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 * If the mapping doesn't provide a set_page_dirty a_op, then
2186 * just fall through and assume that it wants buffer_heads.
2187 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002188int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189{
2190 struct address_space *mapping = page_mapping(page);
2191
2192 if (likely(mapping)) {
2193 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002194 /*
2195 * readahead/lru_deactivate_page could remain
2196 * PG_readahead/PG_reclaim due to race with end_page_writeback
2197 * About readahead, if the page is written, the flags would be
2198 * reset. So no problem.
2199 * About lru_deactivate_page, if the page is redirty, the flag
2200 * will be reset. So no problem. but if the page is used by readahead
2201 * it will confuse readahead and make it restart the size rampup
2202 * process. But it's a trivial problem.
2203 */
2204 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002205#ifdef CONFIG_BLOCK
2206 if (!spd)
2207 spd = __set_page_dirty_buffers;
2208#endif
2209 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002211 if (!PageDirty(page)) {
2212 if (!TestSetPageDirty(page))
2213 return 1;
2214 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 return 0;
2216}
2217EXPORT_SYMBOL(set_page_dirty);
2218
2219/*
2220 * set_page_dirty() is racy if the caller has no reference against
2221 * page->mapping->host, and if the page is unlocked. This is because another
2222 * CPU could truncate the page off the mapping and then free the mapping.
2223 *
2224 * Usually, the page _is_ locked, or the caller is a user-space process which
2225 * holds a reference on the inode by having an open file.
2226 *
2227 * In other cases, the page should be locked before running set_page_dirty().
2228 */
2229int set_page_dirty_lock(struct page *page)
2230{
2231 int ret;
2232
Jens Axboe7eaceac2011-03-10 08:52:07 +01002233 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 ret = set_page_dirty(page);
2235 unlock_page(page);
2236 return ret;
2237}
2238EXPORT_SYMBOL(set_page_dirty_lock);
2239
2240/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 * Clear a page's dirty flag, while caring for dirty memory accounting.
2242 * Returns true if the page was previously dirty.
2243 *
2244 * This is for preparing to put the page under writeout. We leave the page
2245 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2246 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2247 * implementation will run either set_page_writeback() or set_page_dirty(),
2248 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2249 * back into sync.
2250 *
2251 * This incoherency between the page's dirty flag and radix-tree tag is
2252 * unfortunate, but it only exists while the page is locked.
2253 */
2254int clear_page_dirty_for_io(struct page *page)
2255{
2256 struct address_space *mapping = page_mapping(page);
2257
Nick Piggin79352892007-07-19 01:47:22 -07002258 BUG_ON(!PageLocked(page));
2259
Linus Torvalds7658cc22006-12-29 10:00:58 -08002260 if (mapping && mapping_cap_account_dirty(mapping)) {
2261 /*
2262 * Yes, Virginia, this is indeed insane.
2263 *
2264 * We use this sequence to make sure that
2265 * (a) we account for dirty stats properly
2266 * (b) we tell the low-level filesystem to
2267 * mark the whole page dirty if it was
2268 * dirty in a pagetable. Only to then
2269 * (c) clean the page again and return 1 to
2270 * cause the writeback.
2271 *
2272 * This way we avoid all nasty races with the
2273 * dirty bit in multiple places and clearing
2274 * them concurrently from different threads.
2275 *
2276 * Note! Normally the "set_page_dirty(page)"
2277 * has no effect on the actual dirty bit - since
2278 * that will already usually be set. But we
2279 * need the side effects, and it can help us
2280 * avoid races.
2281 *
2282 * We basically use the page "master dirty bit"
2283 * as a serialization point for all the different
2284 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002285 */
2286 if (page_mkclean(page))
2287 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002288 /*
2289 * We carefully synchronise fault handlers against
2290 * installing a dirty pte and marking the page dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002291 * at this point. We do this by having them hold the
2292 * page lock while dirtying the page, and pages are
2293 * always locked coming in here, so we get the desired
2294 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002295 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08002296 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08002297 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07002298 dec_bdi_stat(mapping->backing_dev_info,
2299 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002300 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002302 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002304 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002306EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307
2308int test_clear_page_writeback(struct page *page)
2309{
2310 struct address_space *mapping = page_mapping(page);
Sha Zhengju3ea67d02013-09-12 15:13:53 -07002311 unsigned long memcg_flags;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002312 struct mem_cgroup *memcg;
2313 bool locked;
2314 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315
Johannes Weinerd7365e72014-10-29 14:50:48 -07002316 memcg = mem_cgroup_begin_page_stat(page, &locked, &memcg_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002318 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 unsigned long flags;
2320
Nick Piggin19fd6232008-07-25 19:45:32 -07002321 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002323 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 radix_tree_tag_clear(&mapping->page_tree,
2325 page_index(page),
2326 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002327 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002328 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002329 __bdi_writeout_inc(bdi);
2330 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002331 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002332 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 } else {
2334 ret = TestClearPageWriteback(page);
2335 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002336 if (ret) {
Johannes Weinerd7365e72014-10-29 14:50:48 -07002337 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
Andrew Mortond688abf2007-07-19 01:49:17 -07002338 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002339 inc_zone_page_state(page, NR_WRITTEN);
2340 }
Michal Hockoe4bd6a02014-12-10 15:44:39 -08002341 mem_cgroup_end_page_stat(memcg, &locked, &memcg_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 return ret;
2343}
2344
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002345int __test_set_page_writeback(struct page *page, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346{
2347 struct address_space *mapping = page_mapping(page);
Sha Zhengju3ea67d02013-09-12 15:13:53 -07002348 unsigned long memcg_flags;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002349 struct mem_cgroup *memcg;
2350 bool locked;
2351 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352
Johannes Weinerd7365e72014-10-29 14:50:48 -07002353 memcg = mem_cgroup_begin_page_stat(page, &locked, &memcg_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002355 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 unsigned long flags;
2357
Nick Piggin19fd6232008-07-25 19:45:32 -07002358 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002360 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 radix_tree_tag_set(&mapping->page_tree,
2362 page_index(page),
2363 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002364 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002365 __inc_bdi_stat(bdi, BDI_WRITEBACK);
2366 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 if (!PageDirty(page))
2368 radix_tree_tag_clear(&mapping->page_tree,
2369 page_index(page),
2370 PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002371 if (!keep_write)
2372 radix_tree_tag_clear(&mapping->page_tree,
2373 page_index(page),
2374 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002375 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 } else {
2377 ret = TestSetPageWriteback(page);
2378 }
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002379 if (!ret) {
Johannes Weinerd7365e72014-10-29 14:50:48 -07002380 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002381 inc_zone_page_state(page, NR_WRITEBACK);
2382 }
Michal Hockoe4bd6a02014-12-10 15:44:39 -08002383 mem_cgroup_end_page_stat(memcg, &locked, &memcg_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 return ret;
2385
2386}
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002387EXPORT_SYMBOL(__test_set_page_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
2389/*
Nick Piggin00128182007-10-16 01:24:40 -07002390 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 * passed tag.
2392 */
2393int mapping_tagged(struct address_space *mapping, int tag)
2394{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002395 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396}
2397EXPORT_SYMBOL(mapping_tagged);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002398
2399/**
2400 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2401 * @page: The page to wait on.
2402 *
2403 * This function determines if the given page is related to a backing device
2404 * that requires page contents to be held stable during writeback. If so, then
2405 * it will wait for any pending writeback to complete.
2406 */
2407void wait_for_stable_page(struct page *page)
2408{
2409 struct address_space *mapping = page_mapping(page);
2410 struct backing_dev_info *bdi = mapping->backing_dev_info;
2411
2412 if (!bdi_cap_stable_pages_required(bdi))
2413 return;
2414
2415 wait_on_page_writeback(page);
2416}
2417EXPORT_SYMBOL_GPL(wait_for_stable_page);