blob: 3e4d65445fa71d7d629868c1db70af2bfb1fab28 [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
Tejun Heodcc25ae2015-05-22 18:23:22 -0400125struct wb_domain global_wb_domain;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400127/* consolidated parameters for balance_dirty_pages() and its subroutines */
128struct dirty_throttle_control {
Tejun Heoe9f07df2015-05-22 18:23:28 -0400129#ifdef CONFIG_CGROUP_WRITEBACK
130 struct wb_domain *dom;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400131 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
Tejun Heoe9f07df2015-05-22 18:23:28 -0400132#endif
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400133 struct bdi_writeback *wb;
Tejun Heoe9770b32015-05-22 18:23:27 -0400134 struct fprop_local_percpu *wb_completions;
Jan Karaeb608e32012-05-24 18:59:11 +0200135
Tejun Heo9fc3a432015-05-22 18:23:30 -0400136 unsigned long avail; /* dirtyable */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400137 unsigned long dirty; /* file_dirty + write + nfs */
138 unsigned long thresh; /* dirty threshold */
139 unsigned long bg_thresh; /* dirty background threshold */
140
141 unsigned long wb_dirty; /* per-wb counterparts */
142 unsigned long wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -0400143 unsigned long wb_bg_thresh;
Tejun Heodaddfa32015-05-22 18:23:26 -0400144
145 unsigned long pos_ratio;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400146};
147
Jan Karaeb608e32012-05-24 18:59:11 +0200148/*
149 * Length of period for aging writeout fractions of bdis. This is an
150 * arbitrarily chosen number. The longer the period, the slower fractions will
151 * reflect changes in current writeout rate.
152 */
153#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700154
Tejun Heo693108a2015-05-22 17:13:49 -0400155#ifdef CONFIG_CGROUP_WRITEBACK
156
Tejun Heod60d1bd2015-09-29 12:47:53 -0400157#define GDTC_INIT(__wb) .wb = (__wb), \
158 .dom = &global_wb_domain, \
159 .wb_completions = &(__wb)->completions
160
Tejun Heo9fc3a432015-05-22 18:23:30 -0400161#define GDTC_INIT_NO_WB .dom = &global_wb_domain
Tejun Heod60d1bd2015-09-29 12:47:53 -0400162
163#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
164 .dom = mem_cgroup_wb_domain(__wb), \
165 .wb_completions = &(__wb)->memcg_completions, \
166 .gdtc = __gdtc
Tejun Heoc2aa7232015-05-22 18:23:35 -0400167
168static bool mdtc_valid(struct dirty_throttle_control *dtc)
169{
170 return dtc->dom;
171}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400172
173static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
174{
175 return dtc->dom;
176}
177
Tejun Heo9fc3a432015-05-22 18:23:30 -0400178static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
179{
180 return mdtc->gdtc;
181}
182
Tejun Heo841710a2015-05-22 18:23:33 -0400183static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
184{
185 return &wb->memcg_completions;
186}
187
Tejun Heo693108a2015-05-22 17:13:49 -0400188static void wb_min_max_ratio(struct bdi_writeback *wb,
189 unsigned long *minp, unsigned long *maxp)
190{
191 unsigned long this_bw = wb->avg_write_bandwidth;
192 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
193 unsigned long long min = wb->bdi->min_ratio;
194 unsigned long long max = wb->bdi->max_ratio;
195
196 /*
197 * @wb may already be clean by the time control reaches here and
198 * the total may not include its bw.
199 */
200 if (this_bw < tot_bw) {
201 if (min) {
202 min *= this_bw;
203 do_div(min, tot_bw);
204 }
205 if (max < 100) {
206 max *= this_bw;
207 do_div(max, tot_bw);
208 }
209 }
210
211 *minp = min;
212 *maxp = max;
213}
214
215#else /* CONFIG_CGROUP_WRITEBACK */
216
Tejun Heod60d1bd2015-09-29 12:47:53 -0400217#define GDTC_INIT(__wb) .wb = (__wb), \
218 .wb_completions = &(__wb)->completions
Tejun Heo9fc3a432015-05-22 18:23:30 -0400219#define GDTC_INIT_NO_WB
Tejun Heoc2aa7232015-05-22 18:23:35 -0400220#define MDTC_INIT(__wb, __gdtc)
221
222static bool mdtc_valid(struct dirty_throttle_control *dtc)
223{
224 return false;
225}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400226
227static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
228{
229 return &global_wb_domain;
230}
231
Tejun Heo9fc3a432015-05-22 18:23:30 -0400232static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
233{
234 return NULL;
235}
236
Tejun Heo841710a2015-05-22 18:23:33 -0400237static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
238{
239 return NULL;
240}
241
Tejun Heo693108a2015-05-22 17:13:49 -0400242static void wb_min_max_ratio(struct bdi_writeback *wb,
243 unsigned long *minp, unsigned long *maxp)
244{
245 *minp = wb->bdi->min_ratio;
246 *maxp = wb->bdi->max_ratio;
247}
248
249#endif /* CONFIG_CGROUP_WRITEBACK */
250
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700251/*
Johannes Weinera756cf52012-01-10 15:07:49 -0800252 * In a memory zone, there is a certain amount of pages we consider
253 * available for the page cache, which is essentially the number of
254 * free and reclaimable pages, minus some zone reserves to protect
255 * lowmem and the ability to uphold the zone's watermarks without
256 * requiring writeback.
257 *
258 * This number of dirtyable pages is the base value of which the
259 * user-configurable dirty ratio is the effictive number of pages that
260 * are allowed to be actually dirtied. Per individual zone, or
261 * globally by using the sum of dirtyable pages over all zones.
262 *
263 * Because the user is allowed to specify the dirty limit globally as
264 * absolute number of bytes, calculating the per-zone dirty limit can
265 * require translating the configured limit into a percentage of
266 * global dirtyable memory first.
267 */
268
Johannes Weinera8045522014-01-29 14:05:39 -0800269/**
270 * zone_dirtyable_memory - number of dirtyable pages in a zone
271 * @zone: the zone
272 *
273 * Returns the zone's number of pages potentially available for dirty
274 * page cache. This is the base value for the per-zone dirty limits.
275 */
276static unsigned long zone_dirtyable_memory(struct zone *zone)
277{
278 unsigned long nr_pages;
279
280 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
281 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
282
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800283 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
284 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800285
286 return nr_pages;
287}
288
Johannes Weiner1edf2232012-01-10 15:06:57 -0800289static unsigned long highmem_dirtyable_memory(unsigned long total)
290{
291#ifdef CONFIG_HIGHMEM
292 int node;
293 unsigned long x = 0;
294
295 for_each_node_state(node, N_HIGH_MEMORY) {
Johannes Weinera8045522014-01-29 14:05:39 -0800296 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
Johannes Weiner1edf2232012-01-10 15:06:57 -0800297
Johannes Weinera8045522014-01-29 14:05:39 -0800298 x += zone_dirtyable_memory(z);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800299 }
300 /*
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800301 * Unreclaimable memory (kernel memory or anonymous memory
302 * without swap) can bring down the dirtyable pages below
303 * the zone's dirty balance reserve and the above calculation
304 * will underflow. However we still want to add in nodes
305 * which are below threshold (negative values) to get a more
306 * accurate calculation but make sure that the total never
307 * underflows.
308 */
309 if ((long)x < 0)
310 x = 0;
311
312 /*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800313 * Make sure that the number of highmem pages is never larger
314 * than the number of the total dirtyable memory. This can only
315 * occur in very strange VM situations but we want to make sure
316 * that this does not occur.
317 */
318 return min(x, total);
319#else
320 return 0;
321#endif
322}
323
324/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800325 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800326 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800327 * Returns the global number of pages potentially available for dirty
328 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800329 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700330static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800331{
332 unsigned long x;
333
Johannes Weinera8045522014-01-29 14:05:39 -0800334 x = global_page_state(NR_FREE_PAGES);
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800335 x -= min(x, dirty_balance_reserve);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800336
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800337 x += global_page_state(NR_INACTIVE_FILE);
338 x += global_page_state(NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800339
Johannes Weiner1edf2232012-01-10 15:06:57 -0800340 if (!vm_highmem_is_dirtyable)
341 x -= highmem_dirtyable_memory(x);
342
343 return x + 1; /* Ensure that we never return 0 */
344}
345
Tejun Heo9fc3a432015-05-22 18:23:30 -0400346/**
347 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
348 * @dtc: dirty_throttle_control of interest
Johannes Weinerccafa282012-01-10 15:07:44 -0800349 *
Tejun Heo9fc3a432015-05-22 18:23:30 -0400350 * Calculate @dtc->thresh and ->bg_thresh considering
351 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
352 * must ensure that @dtc->avail is set before calling this function. The
353 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Johannes Weinerccafa282012-01-10 15:07:44 -0800354 * real-time tasks.
355 */
Tejun Heo9fc3a432015-05-22 18:23:30 -0400356static void domain_dirty_limits(struct dirty_throttle_control *dtc)
357{
358 const unsigned long available_memory = dtc->avail;
359 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
360 unsigned long bytes = vm_dirty_bytes;
361 unsigned long bg_bytes = dirty_background_bytes;
362 unsigned long ratio = vm_dirty_ratio;
363 unsigned long bg_ratio = dirty_background_ratio;
364 unsigned long thresh;
365 unsigned long bg_thresh;
366 struct task_struct *tsk;
367
368 /* gdtc is !NULL iff @dtc is for memcg domain */
369 if (gdtc) {
370 unsigned long global_avail = gdtc->avail;
371
372 /*
373 * The byte settings can't be applied directly to memcg
374 * domains. Convert them to ratios by scaling against
375 * globally available memory.
376 */
377 if (bytes)
378 ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
379 global_avail, 100UL);
380 if (bg_bytes)
381 bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
382 global_avail, 100UL);
383 bytes = bg_bytes = 0;
384 }
385
386 if (bytes)
387 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
388 else
389 thresh = (ratio * available_memory) / 100;
390
391 if (bg_bytes)
392 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
393 else
394 bg_thresh = (bg_ratio * available_memory) / 100;
395
396 if (bg_thresh >= thresh)
397 bg_thresh = thresh / 2;
398 tsk = current;
399 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
400 bg_thresh += bg_thresh / 4;
401 thresh += thresh / 4;
402 }
403 dtc->thresh = thresh;
404 dtc->bg_thresh = bg_thresh;
405
406 /* we should eventually report the domain in the TP */
407 if (!gdtc)
408 trace_global_dirty_state(bg_thresh, thresh);
409}
410
411/**
412 * global_dirty_limits - background-writeback and dirty-throttling thresholds
413 * @pbackground: out parameter for bg_thresh
414 * @pdirty: out parameter for thresh
415 *
416 * Calculate bg_thresh and thresh for global_wb_domain. See
417 * domain_dirty_limits() for details.
418 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800419void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
420{
Tejun Heo9fc3a432015-05-22 18:23:30 -0400421 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
Johannes Weinerccafa282012-01-10 15:07:44 -0800422
Tejun Heo9fc3a432015-05-22 18:23:30 -0400423 gdtc.avail = global_dirtyable_memory();
424 domain_dirty_limits(&gdtc);
Johannes Weinerccafa282012-01-10 15:07:44 -0800425
Tejun Heo9fc3a432015-05-22 18:23:30 -0400426 *pbackground = gdtc.bg_thresh;
427 *pdirty = gdtc.thresh;
Johannes Weinerccafa282012-01-10 15:07:44 -0800428}
429
Johannes Weinera756cf52012-01-10 15:07:49 -0800430/**
Johannes Weinera756cf52012-01-10 15:07:49 -0800431 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
432 * @zone: the zone
433 *
434 * Returns the maximum number of dirty pages allowed in a zone, based
435 * on the zone's dirtyable memory.
436 */
437static unsigned long zone_dirty_limit(struct zone *zone)
438{
439 unsigned long zone_memory = zone_dirtyable_memory(zone);
440 struct task_struct *tsk = current;
441 unsigned long dirty;
442
443 if (vm_dirty_bytes)
444 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
445 zone_memory / global_dirtyable_memory();
446 else
447 dirty = vm_dirty_ratio * zone_memory / 100;
448
449 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
450 dirty += dirty / 4;
451
452 return dirty;
453}
454
455/**
456 * zone_dirty_ok - tells whether a zone is within its dirty limits
457 * @zone: the zone to check
458 *
459 * Returns %true when the dirty pages in @zone are within the zone's
460 * dirty limit, %false if the limit is exceeded.
461 */
462bool zone_dirty_ok(struct zone *zone)
463{
464 unsigned long limit = zone_dirty_limit(zone);
465
466 return zone_page_state(zone, NR_FILE_DIRTY) +
467 zone_page_state(zone, NR_UNSTABLE_NFS) +
468 zone_page_state(zone, NR_WRITEBACK) <= limit;
469}
470
David Rientjes2da02992009-01-06 14:39:31 -0800471int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700472 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800473 loff_t *ppos)
474{
475 int ret;
476
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700477 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800478 if (ret == 0 && write)
479 dirty_background_bytes = 0;
480 return ret;
481}
482
483int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700484 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800485 loff_t *ppos)
486{
487 int ret;
488
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700489 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800490 if (ret == 0 && write)
491 dirty_background_ratio = 0;
492 return ret;
493}
494
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700495int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700496 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700497 loff_t *ppos)
498{
499 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800500 int ret;
501
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700502 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700503 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200504 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800505 vm_dirty_bytes = 0;
506 }
507 return ret;
508}
509
David Rientjes2da02992009-01-06 14:39:31 -0800510int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700511 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800512 loff_t *ppos)
513{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800514 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800515 int ret;
516
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700517 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800518 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200519 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800520 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700521 }
522 return ret;
523}
524
Jan Karaeb608e32012-05-24 18:59:11 +0200525static unsigned long wp_next_time(unsigned long cur_time)
526{
527 cur_time += VM_COMPLETIONS_PERIOD_LEN;
528 /* 0 has a special meaning... */
529 if (!cur_time)
530 return 1;
531 return cur_time;
532}
533
Tejun Heoc7981432015-05-22 18:23:29 -0400534static void wb_domain_writeout_inc(struct wb_domain *dom,
535 struct fprop_local_percpu *completions,
536 unsigned int max_prop_frac)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700537{
Tejun Heoc7981432015-05-22 18:23:29 -0400538 __fprop_inc_percpu_max(&dom->completions, completions,
539 max_prop_frac);
Jan Karaeb608e32012-05-24 18:59:11 +0200540 /* First event after period switching was turned off? */
Tejun Heo380c27c2015-05-22 18:23:21 -0400541 if (!unlikely(dom->period_time)) {
Jan Karaeb608e32012-05-24 18:59:11 +0200542 /*
543 * We can race with other __bdi_writeout_inc calls here but
544 * it does not cause any harm since the resulting time when
545 * timer will fire and what is in writeout_period_time will be
546 * roughly the same.
547 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400548 dom->period_time = wp_next_time(jiffies);
549 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200550 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700551}
552
Tejun Heoc7981432015-05-22 18:23:29 -0400553/*
554 * Increment @wb's writeout completion count and the global writeout
555 * completion count. Called from test_clear_page_writeback().
556 */
557static inline void __wb_writeout_inc(struct bdi_writeback *wb)
558{
Tejun Heo841710a2015-05-22 18:23:33 -0400559 struct wb_domain *cgdom;
560
Tejun Heoc7981432015-05-22 18:23:29 -0400561 __inc_wb_stat(wb, WB_WRITTEN);
562 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
563 wb->bdi->max_prop_frac);
Tejun Heo841710a2015-05-22 18:23:33 -0400564
565 cgdom = mem_cgroup_wb_domain(wb);
566 if (cgdom)
567 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
568 wb->bdi->max_prop_frac);
Tejun Heoc7981432015-05-22 18:23:29 -0400569}
570
Tejun Heo93f78d82015-05-22 17:13:27 -0400571void wb_writeout_inc(struct bdi_writeback *wb)
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700572{
573 unsigned long flags;
574
575 local_irq_save(flags);
Tejun Heo93f78d82015-05-22 17:13:27 -0400576 __wb_writeout_inc(wb);
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700577 local_irq_restore(flags);
578}
Tejun Heo93f78d82015-05-22 17:13:27 -0400579EXPORT_SYMBOL_GPL(wb_writeout_inc);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700580
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700581/*
Jan Karaeb608e32012-05-24 18:59:11 +0200582 * On idle system, we can be called long after we scheduled because we use
583 * deferred timers so count with missed periods.
584 */
585static void writeout_period(unsigned long t)
586{
Tejun Heo380c27c2015-05-22 18:23:21 -0400587 struct wb_domain *dom = (void *)t;
588 int miss_periods = (jiffies - dom->period_time) /
Jan Karaeb608e32012-05-24 18:59:11 +0200589 VM_COMPLETIONS_PERIOD_LEN;
590
Tejun Heo380c27c2015-05-22 18:23:21 -0400591 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
592 dom->period_time = wp_next_time(dom->period_time +
Jan Karaeb608e32012-05-24 18:59:11 +0200593 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
Tejun Heo380c27c2015-05-22 18:23:21 -0400594 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200595 } else {
596 /*
597 * Aging has zeroed all fractions. Stop wasting CPU on period
598 * updates.
599 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400600 dom->period_time = 0;
Jan Karaeb608e32012-05-24 18:59:11 +0200601 }
602}
603
Tejun Heo380c27c2015-05-22 18:23:21 -0400604int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
605{
606 memset(dom, 0, sizeof(*dom));
Tejun Heodcc25ae2015-05-22 18:23:22 -0400607
608 spin_lock_init(&dom->lock);
609
Tejun Heo380c27c2015-05-22 18:23:21 -0400610 init_timer_deferrable(&dom->period_timer);
611 dom->period_timer.function = writeout_period;
612 dom->period_timer.data = (unsigned long)dom;
Tejun Heodcc25ae2015-05-22 18:23:22 -0400613
614 dom->dirty_limit_tstamp = jiffies;
615
Tejun Heo380c27c2015-05-22 18:23:21 -0400616 return fprop_global_init(&dom->completions, gfp);
617}
618
Tejun Heo841710a2015-05-22 18:23:33 -0400619#ifdef CONFIG_CGROUP_WRITEBACK
620void wb_domain_exit(struct wb_domain *dom)
621{
622 del_timer_sync(&dom->period_timer);
623 fprop_global_destroy(&dom->completions);
624}
625#endif
626
Jan Karaeb608e32012-05-24 18:59:11 +0200627/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700628 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
629 * registered backing devices, which, for obvious reasons, can not
630 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700631 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700632static unsigned int bdi_min_ratio;
633
634int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
635{
636 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700637
Jens Axboecfc4ba52009-09-14 13:12:40 +0200638 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700639 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700640 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700641 } else {
642 min_ratio -= bdi->min_ratio;
643 if (bdi_min_ratio + min_ratio < 100) {
644 bdi_min_ratio += min_ratio;
645 bdi->min_ratio += min_ratio;
646 } else {
647 ret = -EINVAL;
648 }
649 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200650 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700651
652 return ret;
653}
654
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700655int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
656{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700657 int ret = 0;
658
659 if (max_ratio > 100)
660 return -EINVAL;
661
Jens Axboecfc4ba52009-09-14 13:12:40 +0200662 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700663 if (bdi->min_ratio > max_ratio) {
664 ret = -EINVAL;
665 } else {
666 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200667 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700668 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200669 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700670
671 return ret;
672}
673EXPORT_SYMBOL(bdi_set_max_ratio);
674
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600675static unsigned long dirty_freerun_ceiling(unsigned long thresh,
676 unsigned long bg_thresh)
677{
678 return (thresh + bg_thresh) / 2;
679}
680
Tejun Heoc7981432015-05-22 18:23:29 -0400681static unsigned long hard_dirty_limit(struct wb_domain *dom,
682 unsigned long thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -0600683{
Tejun Heodcc25ae2015-05-22 18:23:22 -0400684 return max(thresh, dom->dirty_limit);
Wu Fengguangffd1f602011-06-19 22:18:42 -0600685}
686
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400687/*
688 * Memory which can be further allocated to a memcg domain is capped by
689 * system-wide clean memory excluding the amount being used in the domain.
690 */
691static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
692 unsigned long filepages, unsigned long headroom)
Tejun Heoc2aa7232015-05-22 18:23:35 -0400693{
694 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400695 unsigned long clean = filepages - min(filepages, mdtc->dirty);
696 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
697 unsigned long other_clean = global_clean - min(global_clean, clean);
Tejun Heoc2aa7232015-05-22 18:23:35 -0400698
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400699 mdtc->avail = filepages + min(headroom, other_clean);
Christoph Lameter1b424462007-05-06 14:48:59 -0700700}
701
Wu Fengguang6f718652011-03-02 17:14:34 -0600702/**
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400703 * __wb_calc_thresh - @wb's share of dirty throttling threshold
704 * @dtc: dirty_throttle_context of interest
Wu Fengguang1babe182010-08-11 14:17:40 -0700705 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400706 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
Wu Fengguang6f718652011-03-02 17:14:34 -0600707 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600708 *
709 * Note that balance_dirty_pages() will only seriously take it as a hard limit
710 * when sleeping max_pause per page is not enough to keep the dirty pages under
711 * control. For example, when the device is completely stalled due to some error
712 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
713 * In the other normal situations, it acts more gently by throttling the tasks
Tejun Heoa88a3412015-05-22 17:13:28 -0400714 * more (rather than completely block them) when the wb dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600715 *
716 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700717 * - starving fast devices
718 * - piling up dirty pages (that will take long time to sync) on slow devices
719 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400720 * The wb's share of dirty limit will be adapting to its throughput and
Wu Fengguang1babe182010-08-11 14:17:40 -0700721 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
722 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400723static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
Wu Fengguang16c40422010-08-11 14:17:39 -0700724{
Tejun Heoe9f07df2015-05-22 18:23:28 -0400725 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400726 unsigned long thresh = dtc->thresh;
Tejun Heo0d960a32015-05-22 18:23:19 -0400727 u64 wb_thresh;
Wu Fengguang16c40422010-08-11 14:17:39 -0700728 long numerator, denominator;
Tejun Heo693108a2015-05-22 17:13:49 -0400729 unsigned long wb_min_ratio, wb_max_ratio;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700730
Wu Fengguang16c40422010-08-11 14:17:39 -0700731 /*
Tejun Heo0d960a32015-05-22 18:23:19 -0400732 * Calculate this BDI's share of the thresh ratio.
Wu Fengguang16c40422010-08-11 14:17:39 -0700733 */
Tejun Heoe9770b32015-05-22 18:23:27 -0400734 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
Tejun Heo380c27c2015-05-22 18:23:21 -0400735 &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700736
Tejun Heo0d960a32015-05-22 18:23:19 -0400737 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
738 wb_thresh *= numerator;
739 do_div(wb_thresh, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700740
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400741 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
Wu Fengguang16c40422010-08-11 14:17:39 -0700742
Tejun Heo0d960a32015-05-22 18:23:19 -0400743 wb_thresh += (thresh * wb_min_ratio) / 100;
744 if (wb_thresh > (thresh * wb_max_ratio) / 100)
745 wb_thresh = thresh * wb_max_ratio / 100;
Wu Fengguang16c40422010-08-11 14:17:39 -0700746
Tejun Heo0d960a32015-05-22 18:23:19 -0400747 return wb_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748}
749
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400750unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
751{
752 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
753 .thresh = thresh };
754 return __wb_calc_thresh(&gdtc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755}
756
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600757/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700758 * setpoint - dirty 3
759 * f(dirty) := 1.0 + (----------------)
760 * limit - setpoint
761 *
762 * it's a 3rd order polynomial that subjects to
763 *
764 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
765 * (2) f(setpoint) = 1.0 => the balance point
766 * (3) f(limit) = 0 => the hard limit
767 * (4) df/dx <= 0 => negative feedback control
768 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
769 * => fast response on large errors; small oscillation near setpoint
770 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700771static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700772 unsigned long dirty,
773 unsigned long limit)
774{
775 long long pos_ratio;
776 long x;
777
Rik van Rield5c9fde2014-05-06 12:50:01 -0700778 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Tejun Heo464d1382015-04-21 16:49:13 -0400779 (limit - setpoint) | 1);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700780 pos_ratio = x;
781 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
782 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
783 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
784
785 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
786}
787
788/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600789 * Dirty position control.
790 *
791 * (o) global/bdi setpoints
792 *
Tejun Heode1fff32015-05-22 17:13:29 -0400793 * We want the dirty pages be balanced around the global/wb setpoints.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600794 * When the number of dirty pages is higher/lower than the setpoint, the
795 * dirty position control ratio (and hence task dirty ratelimit) will be
796 * decreased/increased to bring the dirty pages back to the setpoint.
797 *
798 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
799 *
800 * if (dirty < setpoint) scale up pos_ratio
801 * if (dirty > setpoint) scale down pos_ratio
802 *
Tejun Heode1fff32015-05-22 17:13:29 -0400803 * if (wb_dirty < wb_setpoint) scale up pos_ratio
804 * if (wb_dirty > wb_setpoint) scale down pos_ratio
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600805 *
806 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
807 *
808 * (o) global control line
809 *
810 * ^ pos_ratio
811 * |
812 * | |<===== global dirty control scope ======>|
813 * 2.0 .............*
814 * | .*
815 * | . *
816 * | . *
817 * | . *
818 * | . *
819 * | . *
820 * 1.0 ................................*
821 * | . . *
822 * | . . *
823 * | . . *
824 * | . . *
825 * | . . *
826 * 0 +------------.------------------.----------------------*------------->
827 * freerun^ setpoint^ limit^ dirty pages
828 *
Tejun Heode1fff32015-05-22 17:13:29 -0400829 * (o) wb control line
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600830 *
831 * ^ pos_ratio
832 * |
833 * | *
834 * | *
835 * | *
836 * | *
837 * | * |<=========== span ============>|
838 * 1.0 .......................*
839 * | . *
840 * | . *
841 * | . *
842 * | . *
843 * | . *
844 * | . *
845 * | . *
846 * | . *
847 * | . *
848 * | . *
849 * | . *
850 * 1/4 ...............................................* * * * * * * * * * * *
851 * | . .
852 * | . .
853 * | . .
854 * 0 +----------------------.-------------------------------.------------->
Tejun Heode1fff32015-05-22 17:13:29 -0400855 * wb_setpoint^ x_intercept^
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600856 *
Tejun Heode1fff32015-05-22 17:13:29 -0400857 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600858 * be smoothly throttled down to normal if it starts high in situations like
859 * - start writing to a slow SD card and a fast disk at the same time. The SD
Tejun Heode1fff32015-05-22 17:13:29 -0400860 * card's wb_dirty may rush to many times higher than wb_setpoint.
861 * - the wb dirty thresh drops quickly due to change of JBOD workload
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600862 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400863static void wb_position_ratio(struct dirty_throttle_control *dtc)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600864{
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400865 struct bdi_writeback *wb = dtc->wb;
Tejun Heoa88a3412015-05-22 17:13:28 -0400866 unsigned long write_bw = wb->avg_write_bandwidth;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400867 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -0400868 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400869 unsigned long wb_thresh = dtc->wb_thresh;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600870 unsigned long x_intercept;
871 unsigned long setpoint; /* dirty pages' target balance point */
Tejun Heode1fff32015-05-22 17:13:29 -0400872 unsigned long wb_setpoint;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600873 unsigned long span;
874 long long pos_ratio; /* for scaling up/down the rate limit */
875 long x;
876
Tejun Heodaddfa32015-05-22 18:23:26 -0400877 dtc->pos_ratio = 0;
878
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400879 if (unlikely(dtc->dirty >= limit))
Tejun Heodaddfa32015-05-22 18:23:26 -0400880 return;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600881
882 /*
883 * global setpoint
884 *
Maxim Patlasov5a537482013-09-11 14:22:46 -0700885 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600886 */
887 setpoint = (freerun + limit) / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400888 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700889
890 /*
891 * The strictlimit feature is a tool preventing mistrusted filesystems
892 * from growing a large number of dirty pages before throttling. For
Tejun Heode1fff32015-05-22 17:13:29 -0400893 * such filesystems balance_dirty_pages always checks wb counters
894 * against wb limits. Even if global "nr_dirty" is under "freerun".
Maxim Patlasov5a537482013-09-11 14:22:46 -0700895 * This is especially important for fuse which sets bdi->max_ratio to
896 * 1% by default. Without strictlimit feature, fuse writeback may
897 * consume arbitrary amount of RAM because it is accounted in
898 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
899 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400900 * Here, in wb_position_ratio(), we calculate pos_ratio based on
Tejun Heode1fff32015-05-22 17:13:29 -0400901 * two values: wb_dirty and wb_thresh. Let's consider an example:
Maxim Patlasov5a537482013-09-11 14:22:46 -0700902 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
903 * limits are set by default to 10% and 20% (background and throttle).
Tejun Heode1fff32015-05-22 17:13:29 -0400904 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
Tejun Heo0d960a32015-05-22 18:23:19 -0400905 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
Tejun Heode1fff32015-05-22 17:13:29 -0400906 * about ~6K pages (as the average of background and throttle wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700907 * limits). The 3rd order polynomial will provide positive feedback if
Tejun Heode1fff32015-05-22 17:13:29 -0400908 * wb_dirty is under wb_setpoint and vice versa.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700909 *
910 * Note, that we cannot use global counters in these calculations
Tejun Heode1fff32015-05-22 17:13:29 -0400911 * because we want to throttle process writing to a strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700912 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
913 * in the example above).
914 */
Tejun Heoa88a3412015-05-22 17:13:28 -0400915 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heode1fff32015-05-22 17:13:29 -0400916 long long wb_pos_ratio;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700917
Tejun Heodaddfa32015-05-22 18:23:26 -0400918 if (dtc->wb_dirty < 8) {
919 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
920 2 << RATELIMIT_CALC_SHIFT);
921 return;
922 }
Maxim Patlasov5a537482013-09-11 14:22:46 -0700923
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400924 if (dtc->wb_dirty >= wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400925 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700926
Tejun Heo970fb012015-05-22 18:23:24 -0400927 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
928 dtc->wb_bg_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700929
Tejun Heode1fff32015-05-22 17:13:29 -0400930 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400931 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700932
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400933 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
Tejun Heode1fff32015-05-22 17:13:29 -0400934 wb_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700935
936 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400937 * Typically, for strictlimit case, wb_setpoint << setpoint
938 * and pos_ratio >> wb_pos_ratio. In the other words global
Maxim Patlasov5a537482013-09-11 14:22:46 -0700939 * state ("dirty") is not limiting factor and we have to
Tejun Heode1fff32015-05-22 17:13:29 -0400940 * make decision based on wb counters. But there is an
Maxim Patlasov5a537482013-09-11 14:22:46 -0700941 * important case when global pos_ratio should get precedence:
942 * global limits are exceeded (e.g. due to activities on other
Tejun Heode1fff32015-05-22 17:13:29 -0400943 * wb's) while given strictlimit wb is below limit.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700944 *
Tejun Heode1fff32015-05-22 17:13:29 -0400945 * "pos_ratio * wb_pos_ratio" would work for the case above,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700946 * but it would look too non-natural for the case of all
Tejun Heode1fff32015-05-22 17:13:29 -0400947 * activity in the system coming from a single strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700948 * with bdi->max_ratio == 100%.
949 *
950 * Note that min() below somewhat changes the dynamics of the
951 * control system. Normally, pos_ratio value can be well over 3
Tejun Heode1fff32015-05-22 17:13:29 -0400952 * (when globally we are at freerun and wb is well below wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700953 * setpoint). Now the maximum pos_ratio in the same situation
954 * is 2. We might want to tweak this if we observe the control
955 * system is too slow to adapt.
956 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400957 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
958 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700959 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600960
961 /*
962 * We have computed basic pos_ratio above based on global situation. If
Tejun Heode1fff32015-05-22 17:13:29 -0400963 * the wb is over/under its share of dirty pages, we want to scale
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600964 * pos_ratio further down/up. That is done by the following mechanism.
965 */
966
967 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400968 * wb setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600969 *
Tejun Heode1fff32015-05-22 17:13:29 -0400970 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600971 *
Tejun Heode1fff32015-05-22 17:13:29 -0400972 * x_intercept - wb_dirty
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600973 * := --------------------------
Tejun Heode1fff32015-05-22 17:13:29 -0400974 * x_intercept - wb_setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600975 *
Tejun Heode1fff32015-05-22 17:13:29 -0400976 * The main wb control line is a linear function that subjects to
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600977 *
Tejun Heode1fff32015-05-22 17:13:29 -0400978 * (1) f(wb_setpoint) = 1.0
979 * (2) k = - 1 / (8 * write_bw) (in single wb case)
980 * or equally: x_intercept = wb_setpoint + 8 * write_bw
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600981 *
Tejun Heode1fff32015-05-22 17:13:29 -0400982 * For single wb case, the dirty pages are observed to fluctuate
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600983 * regularly within range
Tejun Heode1fff32015-05-22 17:13:29 -0400984 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600985 * for various filesystems, where (2) can yield in a reasonable 12.5%
986 * fluctuation range for pos_ratio.
987 *
Tejun Heode1fff32015-05-22 17:13:29 -0400988 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600989 * own size, so move the slope over accordingly and choose a slope that
Tejun Heode1fff32015-05-22 17:13:29 -0400990 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600991 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400992 if (unlikely(wb_thresh > dtc->thresh))
993 wb_thresh = dtc->thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600994 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400995 * It's very possible that wb_thresh is close to 0 not because the
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600996 * device is slow, but that it has remained inactive for long time.
997 * Honour such devices a reasonable good (hopefully IO efficient)
998 * threshold, so that the occasional writes won't be blocked and active
999 * writes can rampup the threshold quickly.
1000 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001001 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001002 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001003 * scale global setpoint to wb's:
1004 * wb_setpoint = setpoint * wb_thresh / thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001005 */
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001006 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
Tejun Heode1fff32015-05-22 17:13:29 -04001007 wb_setpoint = setpoint * (u64)x >> 16;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001008 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001009 * Use span=(8*write_bw) in single wb case as indicated by
1010 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001011 *
Tejun Heode1fff32015-05-22 17:13:29 -04001012 * wb_thresh thresh - wb_thresh
1013 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1014 * thresh thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001015 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001016 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
Tejun Heode1fff32015-05-22 17:13:29 -04001017 x_intercept = wb_setpoint + span;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001018
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001019 if (dtc->wb_dirty < x_intercept - span / 4) {
1020 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001021 (x_intercept - wb_setpoint) | 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001022 } else
1023 pos_ratio /= 4;
1024
Wu Fengguang8927f662011-08-04 22:16:46 -06001025 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001026 * wb reserve area, safeguard against dirty pool underrun and disk idle
Wu Fengguang8927f662011-08-04 22:16:46 -06001027 * It may push the desired control point of global dirty pages higher
1028 * than setpoint.
1029 */
Tejun Heode1fff32015-05-22 17:13:29 -04001030 x_intercept = wb_thresh / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001031 if (dtc->wb_dirty < x_intercept) {
1032 if (dtc->wb_dirty > x_intercept / 8)
1033 pos_ratio = div_u64(pos_ratio * x_intercept,
1034 dtc->wb_dirty);
Wu Fengguang50657fc2011-10-11 17:06:33 -06001035 else
Wu Fengguang8927f662011-08-04 22:16:46 -06001036 pos_ratio *= 8;
1037 }
1038
Tejun Heodaddfa32015-05-22 18:23:26 -04001039 dtc->pos_ratio = pos_ratio;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001040}
1041
Tejun Heoa88a3412015-05-22 17:13:28 -04001042static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1043 unsigned long elapsed,
1044 unsigned long written)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001045{
1046 const unsigned long period = roundup_pow_of_two(3 * HZ);
Tejun Heoa88a3412015-05-22 17:13:28 -04001047 unsigned long avg = wb->avg_write_bandwidth;
1048 unsigned long old = wb->write_bandwidth;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001049 u64 bw;
1050
1051 /*
1052 * bw = written * HZ / elapsed
1053 *
1054 * bw * elapsed + write_bandwidth * (period - elapsed)
1055 * write_bandwidth = ---------------------------------------------------
1056 * period
Tejun Heoc72efb62015-03-23 00:18:48 -04001057 *
1058 * @written may have decreased due to account_page_redirty().
1059 * Avoid underflowing @bw calculation.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001060 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001061 bw = written - min(written, wb->written_stamp);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001062 bw *= HZ;
1063 if (unlikely(elapsed > period)) {
1064 do_div(bw, elapsed);
1065 avg = bw;
1066 goto out;
1067 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001068 bw += (u64)wb->write_bandwidth * (period - elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001069 bw >>= ilog2(period);
1070
1071 /*
1072 * one more level of smoothing, for filtering out sudden spikes
1073 */
1074 if (avg > old && old >= (unsigned long)bw)
1075 avg -= (avg - old) >> 3;
1076
1077 if (avg < old && old <= (unsigned long)bw)
1078 avg += (old - avg) >> 3;
1079
1080out:
Tejun Heo95a46c62015-05-22 17:13:47 -04001081 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1082 avg = max(avg, 1LU);
1083 if (wb_has_dirty_io(wb)) {
1084 long delta = avg - wb->avg_write_bandwidth;
1085 WARN_ON_ONCE(atomic_long_add_return(delta,
1086 &wb->bdi->tot_write_bandwidth) <= 0);
1087 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001088 wb->write_bandwidth = bw;
1089 wb->avg_write_bandwidth = avg;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001090}
1091
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001092static void update_dirty_limit(struct dirty_throttle_control *dtc)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001093{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001094 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001095 unsigned long thresh = dtc->thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04001096 unsigned long limit = dom->dirty_limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001097
1098 /*
1099 * Follow up in one step.
1100 */
1101 if (limit < thresh) {
1102 limit = thresh;
1103 goto update;
1104 }
1105
1106 /*
1107 * Follow down slowly. Use the higher one as the target, because thresh
1108 * may drop below dirty. This is exactly the reason to introduce
Tejun Heodcc25ae2015-05-22 18:23:22 -04001109 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
Wu Fengguangc42843f2011-03-02 15:54:09 -06001110 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001111 thresh = max(thresh, dtc->dirty);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001112 if (limit > thresh) {
1113 limit -= (limit - thresh) >> 5;
1114 goto update;
1115 }
1116 return;
1117update:
Tejun Heodcc25ae2015-05-22 18:23:22 -04001118 dom->dirty_limit = limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001119}
1120
Tejun Heoe9f07df2015-05-22 18:23:28 -04001121static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001122 unsigned long now)
1123{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001124 struct wb_domain *dom = dtc_dom(dtc);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001125
1126 /*
1127 * check locklessly first to optimize away locking for the most time
1128 */
Tejun Heodcc25ae2015-05-22 18:23:22 -04001129 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
Wu Fengguangc42843f2011-03-02 15:54:09 -06001130 return;
1131
Tejun Heodcc25ae2015-05-22 18:23:22 -04001132 spin_lock(&dom->lock);
1133 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001134 update_dirty_limit(dtc);
Tejun Heodcc25ae2015-05-22 18:23:22 -04001135 dom->dirty_limit_tstamp = now;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001136 }
Tejun Heodcc25ae2015-05-22 18:23:22 -04001137 spin_unlock(&dom->lock);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001138}
1139
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001140/*
Tejun Heode1fff32015-05-22 17:13:29 -04001141 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001142 *
Tejun Heode1fff32015-05-22 17:13:29 -04001143 * Normal wb tasks will be curbed at or below it in long term.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001144 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1145 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001146static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
Tejun Heoa88a3412015-05-22 17:13:28 -04001147 unsigned long dirtied,
1148 unsigned long elapsed)
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001149{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001150 struct bdi_writeback *wb = dtc->wb;
1151 unsigned long dirty = dtc->dirty;
1152 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001153 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Wu Fengguang73811312011-08-26 15:53:24 -06001154 unsigned long setpoint = (freerun + limit) / 2;
Tejun Heoa88a3412015-05-22 17:13:28 -04001155 unsigned long write_bw = wb->avg_write_bandwidth;
1156 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001157 unsigned long dirty_rate;
1158 unsigned long task_ratelimit;
1159 unsigned long balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001160 unsigned long step;
1161 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001162
1163 /*
1164 * The dirty rate will match the writeout rate in long term, except
1165 * when dirty pages are truncated by userspace or re-dirtied by FS.
1166 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001167 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001168
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001169 /*
1170 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1171 */
1172 task_ratelimit = (u64)dirty_ratelimit *
Tejun Heodaddfa32015-05-22 18:23:26 -04001173 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001174 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1175
1176 /*
1177 * A linear estimation of the "balanced" throttle rate. The theory is,
Tejun Heode1fff32015-05-22 17:13:29 -04001178 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001179 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1180 * formula will yield the balanced rate limit (write_bw / N).
1181 *
1182 * Note that the expanded form is not a pure rate feedback:
1183 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1184 * but also takes pos_ratio into account:
1185 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1186 *
1187 * (1) is not realistic because pos_ratio also takes part in balancing
1188 * the dirty rate. Consider the state
1189 * pos_ratio = 0.5 (3)
1190 * rate = 2 * (write_bw / N) (4)
1191 * If (1) is used, it will stuck in that state! Because each dd will
1192 * be throttled at
1193 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1194 * yielding
1195 * dirty_rate = N * task_ratelimit = write_bw (6)
1196 * put (6) into (1) we get
1197 * rate_(i+1) = rate_(i) (7)
1198 *
1199 * So we end up using (2) to always keep
1200 * rate_(i+1) ~= (write_bw / N) (8)
1201 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1202 * pos_ratio is able to drive itself to 1.0, which is not only where
1203 * the dirty count meet the setpoint, but also where the slope of
1204 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1205 */
1206 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1207 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001208 /*
1209 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1210 */
1211 if (unlikely(balanced_dirty_ratelimit > write_bw))
1212 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001213
Wu Fengguang73811312011-08-26 15:53:24 -06001214 /*
1215 * We could safely do this and return immediately:
1216 *
Tejun Heode1fff32015-05-22 17:13:29 -04001217 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001218 *
1219 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001220 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001221 * limit the step size.
1222 *
1223 * The below code essentially only uses the relative value of
1224 *
1225 * task_ratelimit - dirty_ratelimit
1226 * = (pos_ratio - 1) * dirty_ratelimit
1227 *
1228 * which reflects the direction and size of dirty position error.
1229 */
1230
1231 /*
1232 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1233 * task_ratelimit is on the same side of dirty_ratelimit, too.
1234 * For example, when
1235 * - dirty_ratelimit > balanced_dirty_ratelimit
1236 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1237 * lowering dirty_ratelimit will help meet both the position and rate
1238 * control targets. Otherwise, don't update dirty_ratelimit if it will
1239 * only help meet the rate target. After all, what the users ultimately
1240 * feel and care are stable dirty rate and small position error.
1241 *
1242 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001243 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001244 * keeps jumping around randomly and can even leap far away at times
1245 * due to the small 200ms estimation period of dirty_rate (we want to
1246 * keep that period small to reduce time lags).
1247 */
1248 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001249
1250 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001251 * For strictlimit case, calculations above were based on wb counters
Tejun Heoa88a3412015-05-22 17:13:28 -04001252 * and limits (starting from pos_ratio = wb_position_ratio() and up to
Maxim Patlasov5a537482013-09-11 14:22:46 -07001253 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
Tejun Heode1fff32015-05-22 17:13:29 -04001254 * Hence, to calculate "step" properly, we have to use wb_dirty as
1255 * "dirty" and wb_setpoint as "setpoint".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001256 *
Tejun Heode1fff32015-05-22 17:13:29 -04001257 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1258 * it's possible that wb_thresh is close to zero due to inactivity
Tejun Heo970fb012015-05-22 18:23:24 -04001259 * of backing device.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001260 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001261 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001262 dirty = dtc->wb_dirty;
1263 if (dtc->wb_dirty < 8)
1264 setpoint = dtc->wb_dirty + 1;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001265 else
Tejun Heo970fb012015-05-22 18:23:24 -04001266 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001267 }
1268
Wu Fengguang73811312011-08-26 15:53:24 -06001269 if (dirty < setpoint) {
Tejun Heoa88a3412015-05-22 17:13:28 -04001270 x = min3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001271 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001272 if (dirty_ratelimit < x)
1273 step = x - dirty_ratelimit;
1274 } else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001275 x = max3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001276 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001277 if (dirty_ratelimit > x)
1278 step = dirty_ratelimit - x;
1279 }
1280
1281 /*
1282 * Don't pursue 100% rate matching. It's impossible since the balanced
1283 * rate itself is constantly fluctuating. So decrease the track speed
1284 * when it gets close to the target. Helps eliminate pointless tremors.
1285 */
1286 step >>= dirty_ratelimit / (2 * step + 1);
1287 /*
1288 * Limit the tracking speed to avoid overshooting.
1289 */
1290 step = (step + 7) / 8;
1291
1292 if (dirty_ratelimit < balanced_dirty_ratelimit)
1293 dirty_ratelimit += step;
1294 else
1295 dirty_ratelimit -= step;
1296
Tejun Heoa88a3412015-05-22 17:13:28 -04001297 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1298 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001299
Tejun Heo5634cc22015-08-18 14:54:56 -07001300 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001301}
1302
Tejun Heoc2aa7232015-05-22 18:23:35 -04001303static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1304 struct dirty_throttle_control *mdtc,
Tejun Heo8a731792015-05-22 18:23:20 -04001305 unsigned long start_time,
1306 bool update_ratelimit)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001307{
Tejun Heoc2aa7232015-05-22 18:23:35 -04001308 struct bdi_writeback *wb = gdtc->wb;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001309 unsigned long now = jiffies;
Tejun Heoa88a3412015-05-22 17:13:28 -04001310 unsigned long elapsed = now - wb->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001311 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001312 unsigned long written;
1313
Tejun Heo8a731792015-05-22 18:23:20 -04001314 lockdep_assert_held(&wb->list_lock);
1315
Wu Fengguange98be2d2010-08-29 11:22:30 -06001316 /*
1317 * rate-limit, only update once every 200ms.
1318 */
1319 if (elapsed < BANDWIDTH_INTERVAL)
1320 return;
1321
Tejun Heoa88a3412015-05-22 17:13:28 -04001322 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1323 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001324
1325 /*
1326 * Skip quiet periods when disk bandwidth is under-utilized.
1327 * (at least 1s idle time between two flusher runs)
1328 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001329 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
Wu Fengguange98be2d2010-08-29 11:22:30 -06001330 goto snapshot;
1331
Tejun Heo8a731792015-05-22 18:23:20 -04001332 if (update_ratelimit) {
Tejun Heoc2aa7232015-05-22 18:23:35 -04001333 domain_update_bandwidth(gdtc, now);
1334 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1335
1336 /*
1337 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1338 * compiler has no way to figure that out. Help it.
1339 */
1340 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1341 domain_update_bandwidth(mdtc, now);
1342 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1343 }
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001344 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001345 wb_update_write_bandwidth(wb, elapsed, written);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001346
1347snapshot:
Tejun Heoa88a3412015-05-22 17:13:28 -04001348 wb->dirtied_stamp = dirtied;
1349 wb->written_stamp = written;
1350 wb->bw_time_stamp = now;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001351}
1352
Tejun Heo8a731792015-05-22 18:23:20 -04001353void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001354{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001355 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1356
Tejun Heoc2aa7232015-05-22 18:23:35 -04001357 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001358}
1359
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001361 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001362 * will look to see if it needs to start dirty throttling.
1363 *
1364 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1365 * global_page_state() too often. So scale it near-sqrt to the safety margin
1366 * (the number of pages we may dirty without exceeding the dirty limits).
1367 */
1368static unsigned long dirty_poll_interval(unsigned long dirty,
1369 unsigned long thresh)
1370{
1371 if (thresh > dirty)
1372 return 1UL << (ilog2(thresh - dirty) >> 1);
1373
1374 return 1;
1375}
1376
Tejun Heoa88a3412015-05-22 17:13:28 -04001377static unsigned long wb_max_pause(struct bdi_writeback *wb,
Tejun Heode1fff32015-05-22 17:13:29 -04001378 unsigned long wb_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001379{
Tejun Heoa88a3412015-05-22 17:13:28 -04001380 unsigned long bw = wb->avg_write_bandwidth;
Fengguang Wue3b6c652013-10-16 13:47:03 -07001381 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001382
1383 /*
1384 * Limit pause time for small memory systems. If sleeping for too long
1385 * time, a small pool of dirty/writeback pages may go empty and disk go
1386 * idle.
1387 *
1388 * 8 serves as the safety ratio.
1389 */
Tejun Heode1fff32015-05-22 17:13:29 -04001390 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001391 t++;
1392
Fengguang Wue3b6c652013-10-16 13:47:03 -07001393 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001394}
1395
Tejun Heoa88a3412015-05-22 17:13:28 -04001396static long wb_min_pause(struct bdi_writeback *wb,
1397 long max_pause,
1398 unsigned long task_ratelimit,
1399 unsigned long dirty_ratelimit,
1400 int *nr_dirtied_pause)
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001401{
Tejun Heoa88a3412015-05-22 17:13:28 -04001402 long hi = ilog2(wb->avg_write_bandwidth);
1403 long lo = ilog2(wb->dirty_ratelimit);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001404 long t; /* target pause */
1405 long pause; /* estimated next pause */
1406 int pages; /* target nr_dirtied_pause */
1407
1408 /* target for 10ms pause on 1-dd case */
1409 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001410
1411 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001412 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1413 * overheads.
1414 *
1415 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001416 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001417 if (hi > lo)
1418 t += (hi - lo) * (10 * HZ) / 1024;
1419
1420 /*
1421 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1422 * on the much more stable dirty_ratelimit. However the next pause time
1423 * will be computed based on task_ratelimit and the two rate limits may
1424 * depart considerably at some time. Especially if task_ratelimit goes
1425 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1426 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1427 * result task_ratelimit won't be executed faithfully, which could
1428 * eventually bring down dirty_ratelimit.
1429 *
1430 * We apply two rules to fix it up:
1431 * 1) try to estimate the next pause time and if necessary, use a lower
1432 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1433 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1434 * 2) limit the target pause time to max_pause/2, so that the normal
1435 * small fluctuations of task_ratelimit won't trigger rule (1) and
1436 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1437 */
1438 t = min(t, 1 + max_pause / 2);
1439 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1440
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001441 /*
1442 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1443 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1444 * When the 16 consecutive reads are often interrupted by some dirty
1445 * throttling pause during the async writes, cfq will go into idles
1446 * (deadline is fine). So push nr_dirtied_pause as high as possible
1447 * until reaches DIRTY_POLL_THRESH=32 pages.
1448 */
1449 if (pages < DIRTY_POLL_THRESH) {
1450 t = max_pause;
1451 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1452 if (pages > DIRTY_POLL_THRESH) {
1453 pages = DIRTY_POLL_THRESH;
1454 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1455 }
1456 }
1457
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001458 pause = HZ * pages / (task_ratelimit + 1);
1459 if (pause > max_pause) {
1460 t = max_pause;
1461 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1462 }
1463
1464 *nr_dirtied_pause = pages;
1465 /*
1466 * The minimal pause time will normally be half the target pause time.
1467 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001468 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001469}
1470
Tejun Heo970fb012015-05-22 18:23:24 -04001471static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
Maxim Patlasov5a537482013-09-11 14:22:46 -07001472{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001473 struct bdi_writeback *wb = dtc->wb;
Tejun Heo93f78d82015-05-22 17:13:27 -04001474 unsigned long wb_reclaimable;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001475
1476 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001477 * wb_thresh is not treated as some limiting factor as
Maxim Patlasov5a537482013-09-11 14:22:46 -07001478 * dirty_thresh, due to reasons
Tejun Heode1fff32015-05-22 17:13:29 -04001479 * - in JBOD setup, wb_thresh can fluctuate a lot
Maxim Patlasov5a537482013-09-11 14:22:46 -07001480 * - in a system with HDD and USB key, the USB key may somehow
Tejun Heode1fff32015-05-22 17:13:29 -04001481 * go into state (wb_dirty >> wb_thresh) either because
1482 * wb_dirty starts high, or because wb_thresh drops low.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001483 * In this case we don't want to hard throttle the USB key
Tejun Heode1fff32015-05-22 17:13:29 -04001484 * dirtiers for 100 seconds until wb_dirty drops under
1485 * wb_thresh. Instead the auxiliary wb control line in
Tejun Heoa88a3412015-05-22 17:13:28 -04001486 * wb_position_ratio() will let the dirtier task progress
Tejun Heode1fff32015-05-22 17:13:29 -04001487 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001488 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -04001489 dtc->wb_thresh = __wb_calc_thresh(dtc);
Tejun Heo970fb012015-05-22 18:23:24 -04001490 dtc->wb_bg_thresh = dtc->thresh ?
1491 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001492
1493 /*
1494 * In order to avoid the stacked BDI deadlock we need
1495 * to ensure we accurately count the 'dirty' pages when
1496 * the threshold is low.
1497 *
1498 * Otherwise it would be possible to get thresh+n pages
1499 * reported dirty, even though there are thresh-m pages
1500 * actually dirty; with m+n sitting in the percpu
1501 * deltas.
1502 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001503 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
Tejun Heo93f78d82015-05-22 17:13:27 -04001504 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001505 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001506 } else {
Tejun Heo93f78d82015-05-22 17:13:27 -04001507 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001508 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001509 }
1510}
1511
Wu Fengguang9d823e82011-06-11 18:10:12 -06001512/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 * balance_dirty_pages() must be called by processes which are generating dirty
1514 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001515 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001516 * If we're over `background_thresh' then the writeback threads are woken to
1517 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001519static void balance_dirty_pages(struct address_space *mapping,
Tejun Heodfb8ae52015-05-22 17:13:40 -04001520 struct bdi_writeback *wb,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001521 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001523 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001524 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001525 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001526 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1527 &mdtc_stor : NULL;
1528 struct dirty_throttle_control *sdtc;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001529 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
Wu Fengguang83712352011-06-11 19:25:42 -06001530 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001531 long pause;
1532 long max_pause;
1533 long min_pause;
1534 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001535 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001536 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001537 unsigned long dirty_ratelimit;
Tejun Heodfb8ae52015-05-22 17:13:40 -04001538 struct backing_dev_info *bdi = wb->bdi;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001539 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001540 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541
1542 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001543 unsigned long now = jiffies;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001544 unsigned long dirty, thresh, bg_thresh;
Yang Shi50e55bf2015-11-20 15:57:10 -08001545 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1546 unsigned long m_thresh = 0;
1547 unsigned long m_bg_thresh = 0;
Wu Fengguang83712352011-06-11 19:25:42 -06001548
Wu Fengguang143dfe82010-08-27 18:45:12 -06001549 /*
1550 * Unstable writes are a feature of certain networked
1551 * filesystems (i.e. NFS) in which data may have been
1552 * written to the server's write cache, but has not yet
1553 * been flushed to permanent storage.
1554 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001555 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1556 global_page_state(NR_UNSTABLE_NFS);
Tejun Heo9fc3a432015-05-22 18:23:30 -04001557 gdtc->avail = global_dirtyable_memory();
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001558 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001559
Tejun Heo9fc3a432015-05-22 18:23:30 -04001560 domain_dirty_limits(gdtc);
Wu Fengguang16c40422010-08-11 14:17:39 -07001561
Maxim Patlasov5a537482013-09-11 14:22:46 -07001562 if (unlikely(strictlimit)) {
Tejun Heo970fb012015-05-22 18:23:24 -04001563 wb_dirty_limits(gdtc);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001564
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001565 dirty = gdtc->wb_dirty;
1566 thresh = gdtc->wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -04001567 bg_thresh = gdtc->wb_bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001568 } else {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001569 dirty = gdtc->dirty;
1570 thresh = gdtc->thresh;
1571 bg_thresh = gdtc->bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001572 }
1573
Tejun Heoc2aa7232015-05-22 18:23:35 -04001574 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001575 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001576
1577 /*
1578 * If @wb belongs to !root memcg, repeat the same
1579 * basic calculations for the memcg domain.
1580 */
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001581 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1582 &mdtc->dirty, &writeback);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001583 mdtc->dirty += writeback;
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001584 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001585
1586 domain_dirty_limits(mdtc);
1587
1588 if (unlikely(strictlimit)) {
1589 wb_dirty_limits(mdtc);
1590 m_dirty = mdtc->wb_dirty;
1591 m_thresh = mdtc->wb_thresh;
1592 m_bg_thresh = mdtc->wb_bg_thresh;
1593 } else {
1594 m_dirty = mdtc->dirty;
1595 m_thresh = mdtc->thresh;
1596 m_bg_thresh = mdtc->bg_thresh;
1597 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001598 }
1599
1600 /*
1601 * Throttle it only when the background writeback cannot
1602 * catch-up. This avoids (excessively) small writeouts
Tejun Heode1fff32015-05-22 17:13:29 -04001603 * when the wb limits are ramping up in case of !strictlimit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001604 *
Tejun Heode1fff32015-05-22 17:13:29 -04001605 * In strictlimit case make decision based on the wb counters
1606 * and limits. Small writeouts when the wb limits are ramping
Maxim Patlasov5a537482013-09-11 14:22:46 -07001607 * up are the price we consciously pay for strictlimit-ing.
Tejun Heoc2aa7232015-05-22 18:23:35 -04001608 *
1609 * If memcg domain is in effect, @dirty should be under
1610 * both global and memcg freerun ceilings.
Wu Fengguang16c40422010-08-11 14:17:39 -07001611 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001612 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1613 (!mdtc ||
1614 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1615 unsigned long intv = dirty_poll_interval(dirty, thresh);
1616 unsigned long m_intv = ULONG_MAX;
1617
Wu Fengguang83712352011-06-11 19:25:42 -06001618 current->dirty_paused_when = now;
1619 current->nr_dirtied = 0;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001620 if (mdtc)
1621 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1622 current->nr_dirtied_pause = min(intv, m_intv);
Wu Fengguang16c40422010-08-11 14:17:39 -07001623 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001624 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001625
Tejun Heobc058732015-05-22 17:13:53 -04001626 if (unlikely(!writeback_in_progress(wb)))
Tejun Heo9ecf48662015-05-22 17:13:54 -04001627 wb_start_background_writeback(wb);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001628
Tejun Heoc2aa7232015-05-22 18:23:35 -04001629 /*
1630 * Calculate global domain's pos_ratio and select the
1631 * global dtc by default.
1632 */
Maxim Patlasov5a537482013-09-11 14:22:46 -07001633 if (!strictlimit)
Tejun Heo970fb012015-05-22 18:23:24 -04001634 wb_dirty_limits(gdtc);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001635
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001636 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1637 ((gdtc->dirty > gdtc->thresh) || strictlimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638
Tejun Heodaddfa32015-05-22 18:23:26 -04001639 wb_position_ratio(gdtc);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001640 sdtc = gdtc;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001641
Tejun Heoc2aa7232015-05-22 18:23:35 -04001642 if (mdtc) {
1643 /*
1644 * If memcg domain is in effect, calculate its
1645 * pos_ratio. @wb should satisfy constraints from
1646 * both global and memcg domains. Choose the one
1647 * w/ lower pos_ratio.
1648 */
1649 if (!strictlimit)
1650 wb_dirty_limits(mdtc);
1651
1652 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1653 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1654
1655 wb_position_ratio(mdtc);
1656 if (mdtc->pos_ratio < gdtc->pos_ratio)
1657 sdtc = mdtc;
1658 }
Tejun Heodaddfa32015-05-22 18:23:26 -04001659
Tejun Heoa88a3412015-05-22 17:13:28 -04001660 if (dirty_exceeded && !wb->dirty_exceeded)
1661 wb->dirty_exceeded = 1;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001662
Tejun Heo8a731792015-05-22 18:23:20 -04001663 if (time_is_before_jiffies(wb->bw_time_stamp +
1664 BANDWIDTH_INTERVAL)) {
1665 spin_lock(&wb->list_lock);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001666 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
Tejun Heo8a731792015-05-22 18:23:20 -04001667 spin_unlock(&wb->list_lock);
1668 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669
Tejun Heoc2aa7232015-05-22 18:23:35 -04001670 /* throttle according to the chosen dtc */
Tejun Heoa88a3412015-05-22 17:13:28 -04001671 dirty_ratelimit = wb->dirty_ratelimit;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001672 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001673 RATELIMIT_CALC_SHIFT;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001674 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
Tejun Heoa88a3412015-05-22 17:13:28 -04001675 min_pause = wb_min_pause(wb, max_pause,
1676 task_ratelimit, dirty_ratelimit,
1677 &nr_dirtied_pause);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001678
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001679 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001680 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001681 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001682 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 }
Wu Fengguang83712352011-06-11 19:25:42 -06001684 period = HZ * pages_dirtied / task_ratelimit;
1685 pause = period;
1686 if (current->dirty_paused_when)
1687 pause -= now - current->dirty_paused_when;
1688 /*
1689 * For less than 1s think time (ext3/4 may block the dirtier
1690 * for up to 800ms from time to time on 1-HDD; so does xfs,
1691 * however at much less frequency), try to compensate it in
1692 * future periods by updating the virtual time; otherwise just
1693 * do a reset, as it may be a light dirtier.
1694 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001695 if (pause < min_pause) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001696 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001697 sdtc->thresh,
1698 sdtc->bg_thresh,
1699 sdtc->dirty,
1700 sdtc->wb_thresh,
1701 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001702 dirty_ratelimit,
1703 task_ratelimit,
1704 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001705 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001706 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001707 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001708 if (pause < -HZ) {
1709 current->dirty_paused_when = now;
1710 current->nr_dirtied = 0;
1711 } else if (period) {
1712 current->dirty_paused_when += period;
1713 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001714 } else if (current->nr_dirtied_pause <= pages_dirtied)
1715 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001716 break;
1717 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001718 if (unlikely(pause > max_pause)) {
1719 /* for occasional dropped task_ratelimit */
1720 now += min(pause - max_pause, max_pause);
1721 pause = max_pause;
1722 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001723
1724pause:
Tejun Heo5634cc22015-08-18 14:54:56 -07001725 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001726 sdtc->thresh,
1727 sdtc->bg_thresh,
1728 sdtc->dirty,
1729 sdtc->wb_thresh,
1730 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001731 dirty_ratelimit,
1732 task_ratelimit,
1733 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001734 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001735 pause,
1736 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001737 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001738 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001739
Wu Fengguang83712352011-06-11 19:25:42 -06001740 current->dirty_paused_when = now + pause;
1741 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001742 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001743
Wu Fengguangffd1f602011-06-19 22:18:42 -06001744 /*
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001745 * This is typically equal to (dirty < thresh) and can also
1746 * keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001747 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001748 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001749 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001750
Wu Fengguangc5c63432011-12-02 10:21:33 -06001751 /*
1752 * In the case of an unresponding NFS server and the NFS dirty
Tejun Heode1fff32015-05-22 17:13:29 -04001753 * pages exceeds dirty_thresh, give the other good wb's a pipe
Wu Fengguangc5c63432011-12-02 10:21:33 -06001754 * to go through, so that tasks on them still remain responsive.
1755 *
1756 * In theory 1 page is enough to keep the comsumer-producer
1757 * pipe going: the flusher cleans 1 page => the task dirties 1
Tejun Heode1fff32015-05-22 17:13:29 -04001758 * more page. However wb_dirty has accounting errors. So use
Tejun Heo93f78d82015-05-22 17:13:27 -04001759 * the larger and more IO friendly wb_stat_error.
Wu Fengguangc5c63432011-12-02 10:21:33 -06001760 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001761 if (sdtc->wb_dirty <= wb_stat_error(wb))
Wu Fengguangc5c63432011-12-02 10:21:33 -06001762 break;
1763
Jan Kara499d05e2011-11-16 19:34:48 +08001764 if (fatal_signal_pending(current))
1765 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 }
1767
Tejun Heoa88a3412015-05-22 17:13:28 -04001768 if (!dirty_exceeded && wb->dirty_exceeded)
1769 wb->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770
Tejun Heobc058732015-05-22 17:13:53 -04001771 if (writeback_in_progress(wb))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001772 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773
1774 /*
1775 * In laptop mode, we wait until hitting the higher threshold before
1776 * starting background writeout, and then write out all the way down
1777 * to the lower threshold. So slow writers cause minimal disk activity.
1778 *
1779 * In normal mode, we start background writeout at the lower
1780 * background_thresh, to keep the amount of dirty memory low.
1781 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001782 if (laptop_mode)
1783 return;
1784
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001785 if (nr_reclaimable > gdtc->bg_thresh)
Tejun Heo9ecf48662015-05-22 17:13:54 -04001786 wb_start_background_writeback(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787}
1788
Wu Fengguang9d823e82011-06-11 18:10:12 -06001789static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001790
Wu Fengguang54848d72011-04-05 13:21:19 -06001791/*
1792 * Normal tasks are throttled by
1793 * loop {
1794 * dirty tsk->nr_dirtied_pause pages;
1795 * take a snap in balance_dirty_pages();
1796 * }
1797 * However there is a worst case. If every task exit immediately when dirtied
1798 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1799 * called to throttle the page dirties. The solution is to save the not yet
1800 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1801 * randomly into the running tasks. This works well for the above worst case,
1802 * as the new task will pick up and accumulate the old task's leaked dirty
1803 * count and eventually get throttled.
1804 */
1805DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1806
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001808 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001809 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 *
1811 * Processes which are dirtying memory should call in here once for each page
1812 * which was newly dirtied. The function will periodically check the system's
1813 * dirty state and will initiate writeback if needed.
1814 *
1815 * On really big machines, get_writeback_state is expensive, so try to avoid
1816 * calling it too often (ratelimiting). But once we're over the dirty memory
1817 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1818 * from overshooting the limit by (ratelimit_pages) each.
1819 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001820void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821{
Tejun Heodfb8ae52015-05-22 17:13:40 -04001822 struct inode *inode = mapping->host;
1823 struct backing_dev_info *bdi = inode_to_bdi(inode);
1824 struct bdi_writeback *wb = NULL;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001825 int ratelimit;
1826 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827
Wu Fengguang36715ce2011-06-11 17:53:57 -06001828 if (!bdi_cap_account_dirty(bdi))
1829 return;
1830
Tejun Heodfb8ae52015-05-22 17:13:40 -04001831 if (inode_cgwb_enabled(inode))
1832 wb = wb_get_create_current(bdi, GFP_KERNEL);
1833 if (!wb)
1834 wb = &bdi->wb;
1835
Wu Fengguang9d823e82011-06-11 18:10:12 -06001836 ratelimit = current->nr_dirtied_pause;
Tejun Heoa88a3412015-05-22 17:13:28 -04001837 if (wb->dirty_exceeded)
Wu Fengguang9d823e82011-06-11 18:10:12 -06001838 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001840 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001841 /*
1842 * This prevents one CPU to accumulate too many dirtied pages without
1843 * calling into balance_dirty_pages(), which can happen when there are
1844 * 1000+ tasks, all of them start dirtying pages at exactly the same
1845 * time, hence all honoured too large initial task->nr_dirtied_pause.
1846 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001847 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001848 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001849 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001850 else if (unlikely(*p >= ratelimit_pages)) {
1851 *p = 0;
1852 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001854 /*
1855 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1856 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1857 * the dirty throttling and livelock other long-run dirtiers.
1858 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001859 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06001860 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001861 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001862 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1863 *p -= nr_pages_dirtied;
1864 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001866 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001867
1868 if (unlikely(current->nr_dirtied >= ratelimit))
Tejun Heodfb8ae52015-05-22 17:13:40 -04001869 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1870
1871 wb_put(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001873EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874
Tejun Heoaa661bb2015-05-22 18:23:31 -04001875/**
1876 * wb_over_bg_thresh - does @wb need to be written back?
1877 * @wb: bdi_writeback of interest
1878 *
1879 * Determines whether background writeback should keep writing @wb or it's
1880 * clean enough. Returns %true if writeback should continue.
1881 */
1882bool wb_over_bg_thresh(struct bdi_writeback *wb)
1883{
Tejun Heo947e9762015-05-22 18:23:32 -04001884 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001885 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo947e9762015-05-22 18:23:32 -04001886 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001887 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1888 &mdtc_stor : NULL;
Tejun Heoaa661bb2015-05-22 18:23:31 -04001889
Tejun Heo947e9762015-05-22 18:23:32 -04001890 /*
1891 * Similar to balance_dirty_pages() but ignores pages being written
1892 * as we're trying to decide whether to put more under writeback.
1893 */
1894 gdtc->avail = global_dirtyable_memory();
1895 gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1896 global_page_state(NR_UNSTABLE_NFS);
1897 domain_dirty_limits(gdtc);
Tejun Heoaa661bb2015-05-22 18:23:31 -04001898
Tejun Heo947e9762015-05-22 18:23:32 -04001899 if (gdtc->dirty > gdtc->bg_thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04001900 return true;
1901
Tejun Heo947e9762015-05-22 18:23:32 -04001902 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(gdtc))
Tejun Heoaa661bb2015-05-22 18:23:31 -04001903 return true;
1904
Tejun Heoc2aa7232015-05-22 18:23:35 -04001905 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001906 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001907
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001908 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1909 &writeback);
1910 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001911 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1912
1913 if (mdtc->dirty > mdtc->bg_thresh)
1914 return true;
1915
1916 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(mdtc))
1917 return true;
1918 }
1919
Tejun Heoaa661bb2015-05-22 18:23:31 -04001920 return false;
1921}
1922
Andrew Morton232ea4d2007-02-28 20:13:21 -08001923void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924{
David Rientjes364aeb22009-01-06 14:39:29 -08001925 unsigned long background_thresh;
1926 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927
1928 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001929 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001930 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931
1932 /*
1933 * Boost the allowable dirty threshold a bit for page
1934 * allocators so they don't get DoS'ed by heavy writers
1935 */
1936 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1937
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001938 if (global_page_state(NR_UNSTABLE_NFS) +
1939 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1940 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001941 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001942
1943 /*
1944 * The caller might hold locks which can prevent IO completion
1945 * or progress in the filesystem. So we cannot just sit here
1946 * waiting for IO to complete.
1947 */
1948 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1949 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 }
1951}
1952
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1955 */
Joe Perchescccad5b2014-06-06 14:38:09 -07001956int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001957 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001959 proc_dointvec(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960 return 0;
1961}
1962
Jens Axboec2c49862010-05-20 09:18:47 +02001963#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001964void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965{
Matthew Garrett31373d02010-04-06 14:25:14 +02001966 struct request_queue *q = (struct request_queue *)data;
1967 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1968 global_page_state(NR_UNSTABLE_NFS);
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001969 struct bdi_writeback *wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970
Matthew Garrett31373d02010-04-06 14:25:14 +02001971 /*
1972 * We want to write everything out, not just down to the dirty
1973 * threshold
1974 */
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001975 if (!bdi_has_dirty_io(&q->backing_dev_info))
1976 return;
1977
Tejun Heo9ad18ab2015-09-29 12:47:50 -04001978 rcu_read_lock();
Tejun Heob8175252015-10-02 14:47:05 -04001979 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001980 if (wb_has_dirty_io(wb))
1981 wb_start_writeback(wb, nr_pages, true,
1982 WB_REASON_LAPTOP_TIMER);
Tejun Heo9ad18ab2015-09-29 12:47:50 -04001983 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984}
1985
1986/*
1987 * We've spun up the disk and we're in laptop mode: schedule writeback
1988 * of all dirty data a few seconds from now. If the flush is already scheduled
1989 * then push it back - the user is still using the disk.
1990 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001991void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992{
Matthew Garrett31373d02010-04-06 14:25:14 +02001993 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994}
1995
1996/*
1997 * We're in laptop mode and we've just synced. The sync's writes will have
1998 * caused another writeback to be scheduled by laptop_io_completion.
1999 * Nothing needs to be written back anymore, so we unschedule the writeback.
2000 */
2001void laptop_sync_completion(void)
2002{
Matthew Garrett31373d02010-04-06 14:25:14 +02002003 struct backing_dev_info *bdi;
2004
2005 rcu_read_lock();
2006
2007 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2008 del_timer(&bdi->laptop_mode_wb_timer);
2009
2010 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011}
Jens Axboec2c49862010-05-20 09:18:47 +02002012#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013
2014/*
2015 * If ratelimit_pages is too high then we can get into dirty-data overload
2016 * if a large number of processes all perform writes at the same time.
2017 * If it is too low then SMP machines will call the (expensive)
2018 * get_writeback_state too often.
2019 *
2020 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2021 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06002022 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 */
2024
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002025void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026{
Tejun Heodcc25ae2015-05-22 18:23:22 -04002027 struct wb_domain *dom = &global_wb_domain;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002028 unsigned long background_thresh;
2029 unsigned long dirty_thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04002030
Wu Fengguang9d823e82011-06-11 18:10:12 -06002031 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heodcc25ae2015-05-22 18:23:22 -04002032 dom->dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002033 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 if (ratelimit_pages < 16)
2035 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036}
2037
Paul Gortmaker0db06282013-06-19 14:53:51 -04002038static int
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08002039ratelimit_handler(struct notifier_block *self, unsigned long action,
2040 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041{
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08002042
2043 switch (action & ~CPU_TASKS_FROZEN) {
2044 case CPU_ONLINE:
2045 case CPU_DEAD:
2046 writeback_set_ratelimit();
2047 return NOTIFY_OK;
2048 default:
2049 return NOTIFY_DONE;
2050 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051}
2052
Paul Gortmaker0db06282013-06-19 14:53:51 -04002053static struct notifier_block ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054 .notifier_call = ratelimit_handler,
2055 .next = NULL,
2056};
2057
2058/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002059 * Called early on to tune the page writeback dirty limits.
2060 *
2061 * We used to scale dirty pages according to how total memory
2062 * related to pages that could be allocated for buffers (by
2063 * comparing nr_free_buffer_pages() to vm_total_pages.
2064 *
2065 * However, that was when we used "dirty_ratio" to scale with
2066 * all memory, and we don't do that any more. "dirty_ratio"
2067 * is now applied to total non-HIGHPAGE memory (by subtracting
2068 * totalhigh_pages from vm_total_pages), and as such we can't
2069 * get into the old insane situation any more where we had
2070 * large amounts of dirty pages compared to a small amount of
2071 * non-HIGHMEM memory.
2072 *
2073 * But we might still want to scale the dirty_ratio by how
2074 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 */
2076void __init page_writeback_init(void)
2077{
Rabin Vincenta50fcb52015-08-06 15:47:14 -07002078 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2079
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002080 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 register_cpu_notifier(&ratelimit_nb);
2082}
2083
David Howells811d7362006-08-29 19:06:09 +01002084/**
Jan Karaf446daae2010-08-09 17:19:12 -07002085 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2086 * @mapping: address space structure to write
2087 * @start: starting page index
2088 * @end: ending page index (inclusive)
2089 *
2090 * This function scans the page range from @start to @end (inclusive) and tags
2091 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2092 * that write_cache_pages (or whoever calls this function) will then use
2093 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2094 * used to avoid livelocking of writeback by a process steadily creating new
2095 * dirty pages in the file (thus it is important for this function to be quick
2096 * so that it can tag pages faster than a dirtying process can create them).
2097 */
2098/*
2099 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2100 */
Jan Karaf446daae2010-08-09 17:19:12 -07002101void tag_pages_for_writeback(struct address_space *mapping,
2102 pgoff_t start, pgoff_t end)
2103{
Randy Dunlap3c111a02010-08-11 14:17:30 -07002104#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07002105 unsigned long tagged;
2106
2107 do {
2108 spin_lock_irq(&mapping->tree_lock);
2109 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2110 &start, end, WRITEBACK_TAG_BATCH,
2111 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2112 spin_unlock_irq(&mapping->tree_lock);
2113 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2114 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07002115 /* We check 'start' to handle wrapping when end == ~0UL */
2116 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07002117}
2118EXPORT_SYMBOL(tag_pages_for_writeback);
2119
2120/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002121 * 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 +01002122 * @mapping: address space structure to write
2123 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002124 * @writepage: function called for each page
2125 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01002126 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002127 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01002128 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2129 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2130 * and msync() need to guarantee that all the data which was dirty at the time
2131 * the call was made get new I/O started against them. If wbc->sync_mode is
2132 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2133 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07002134 *
2135 * To avoid livelocks (when other process dirties new pages), we first tag
2136 * pages which should be written back with TOWRITE tag and only then start
2137 * writing them. For data-integrity sync we have to be careful so that we do
2138 * not miss some pages (e.g., because some other process has cleared TOWRITE
2139 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2140 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01002141 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002142int write_cache_pages(struct address_space *mapping,
2143 struct writeback_control *wbc, writepage_t writepage,
2144 void *data)
David Howells811d7362006-08-29 19:06:09 +01002145{
David Howells811d7362006-08-29 19:06:09 +01002146 int ret = 0;
2147 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01002148 struct pagevec pvec;
2149 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08002150 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01002151 pgoff_t index;
2152 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08002153 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08002154 int cycled;
David Howells811d7362006-08-29 19:06:09 +01002155 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07002156 int tag;
David Howells811d7362006-08-29 19:06:09 +01002157
David Howells811d7362006-08-29 19:06:09 +01002158 pagevec_init(&pvec, 0);
2159 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08002160 writeback_index = mapping->writeback_index; /* prev offset */
2161 index = writeback_index;
2162 if (index == 0)
2163 cycled = 1;
2164 else
2165 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01002166 end = -1;
2167 } else {
2168 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2169 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2170 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2171 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08002172 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01002173 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002174 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002175 tag = PAGECACHE_TAG_TOWRITE;
2176 else
2177 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01002178retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002179 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002180 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08002181 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002182 while (!done && (index <= end)) {
2183 int i;
2184
Jan Karaf446daae2010-08-09 17:19:12 -07002185 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002186 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2187 if (nr_pages == 0)
2188 break;
David Howells811d7362006-08-29 19:06:09 +01002189
David Howells811d7362006-08-29 19:06:09 +01002190 for (i = 0; i < nr_pages; i++) {
2191 struct page *page = pvec.pages[i];
2192
Nick Piggind5482cd2009-01-06 14:39:11 -08002193 /*
2194 * At this point, the page may be truncated or
2195 * invalidated (changing page->mapping to NULL), or
2196 * even swizzled back from swapper_space to tmpfs file
2197 * mapping. However, page->index will not change
2198 * because we have a reference on the page.
2199 */
2200 if (page->index > end) {
2201 /*
2202 * can't be range_cyclic (1st pass) because
2203 * end == -1 in that case.
2204 */
2205 done = 1;
2206 break;
2207 }
2208
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002209 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08002210
David Howells811d7362006-08-29 19:06:09 +01002211 lock_page(page);
2212
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002213 /*
2214 * Page truncated or invalidated. We can freely skip it
2215 * then, even for data integrity operations: the page
2216 * has disappeared concurrently, so there could be no
2217 * real expectation of this data interity operation
2218 * even if there is now a new, dirty page at the same
2219 * pagecache address.
2220 */
David Howells811d7362006-08-29 19:06:09 +01002221 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002222continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01002223 unlock_page(page);
2224 continue;
2225 }
2226
Nick Piggin515f4a02009-01-06 14:39:10 -08002227 if (!PageDirty(page)) {
2228 /* someone wrote it for us */
2229 goto continue_unlock;
2230 }
David Howells811d7362006-08-29 19:06:09 +01002231
Nick Piggin515f4a02009-01-06 14:39:10 -08002232 if (PageWriteback(page)) {
2233 if (wbc->sync_mode != WB_SYNC_NONE)
2234 wait_on_page_writeback(page);
2235 else
2236 goto continue_unlock;
2237 }
2238
2239 BUG_ON(PageWriteback(page));
2240 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002241 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01002242
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002243 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002244 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08002245 if (unlikely(ret)) {
2246 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2247 unlock_page(page);
2248 ret = 0;
2249 } else {
2250 /*
2251 * done_index is set past this page,
2252 * so media errors will not choke
2253 * background writeout for the entire
2254 * file. This has consequences for
2255 * range_cyclic semantics (ie. it may
2256 * not be suitable for data integrity
2257 * writeout).
2258 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002259 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08002260 done = 1;
2261 break;
2262 }
Dave Chinner0b564922010-06-09 10:37:18 +10002263 }
David Howells811d7362006-08-29 19:06:09 +01002264
Dave Chinner546a1922010-08-24 11:44:34 +10002265 /*
2266 * We stop writing back only if we are not doing
2267 * integrity sync. In case of integrity sync we have to
2268 * keep going until we have written all the pages
2269 * we tagged for writeback prior to entering this loop.
2270 */
2271 if (--wbc->nr_to_write <= 0 &&
2272 wbc->sync_mode == WB_SYNC_NONE) {
2273 done = 1;
2274 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08002275 }
David Howells811d7362006-08-29 19:06:09 +01002276 }
2277 pagevec_release(&pvec);
2278 cond_resched();
2279 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01002280 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01002281 /*
Nick Piggin31a12662009-01-06 14:39:04 -08002282 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01002283 * We hit the last page and there is more work to be done: wrap
2284 * back to the start of the file
2285 */
Nick Piggin31a12662009-01-06 14:39:04 -08002286 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01002287 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08002288 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01002289 goto retry;
2290 }
Dave Chinner0b564922010-06-09 10:37:18 +10002291 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2292 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04002293
David Howells811d7362006-08-29 19:06:09 +01002294 return ret;
2295}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002296EXPORT_SYMBOL(write_cache_pages);
2297
2298/*
2299 * Function used by generic_writepages to call the real writepage
2300 * function and set the mapping flags on error
2301 */
2302static int __writepage(struct page *page, struct writeback_control *wbc,
2303 void *data)
2304{
2305 struct address_space *mapping = data;
2306 int ret = mapping->a_ops->writepage(page, wbc);
2307 mapping_set_error(mapping, ret);
2308 return ret;
2309}
2310
2311/**
2312 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2313 * @mapping: address space structure to write
2314 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2315 *
2316 * This is a library function, which implements the writepages()
2317 * address_space_operation.
2318 */
2319int generic_writepages(struct address_space *mapping,
2320 struct writeback_control *wbc)
2321{
Shaohua Li9b6096a2011-03-17 10:47:06 +01002322 struct blk_plug plug;
2323 int ret;
2324
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002325 /* deal with chardevs and other special file */
2326 if (!mapping->a_ops->writepage)
2327 return 0;
2328
Shaohua Li9b6096a2011-03-17 10:47:06 +01002329 blk_start_plug(&plug);
2330 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2331 blk_finish_plug(&plug);
2332 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002333}
David Howells811d7362006-08-29 19:06:09 +01002334
2335EXPORT_SYMBOL(generic_writepages);
2336
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2338{
Andrew Morton22905f72005-11-16 15:07:01 -08002339 int ret;
2340
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 if (wbc->nr_to_write <= 0)
2342 return 0;
2343 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002344 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002345 else
2346 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002347 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348}
2349
2350/**
2351 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07002352 * @page: the page to write
2353 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 *
2355 * The page must be locked by the caller and will be unlocked upon return.
2356 *
2357 * write_one_page() returns a negative error code if I/O failed.
2358 */
2359int write_one_page(struct page *page, int wait)
2360{
2361 struct address_space *mapping = page->mapping;
2362 int ret = 0;
2363 struct writeback_control wbc = {
2364 .sync_mode = WB_SYNC_ALL,
2365 .nr_to_write = 1,
2366 };
2367
2368 BUG_ON(!PageLocked(page));
2369
2370 if (wait)
2371 wait_on_page_writeback(page);
2372
2373 if (clear_page_dirty_for_io(page)) {
2374 page_cache_get(page);
2375 ret = mapping->a_ops->writepage(page, &wbc);
2376 if (ret == 0 && wait) {
2377 wait_on_page_writeback(page);
2378 if (PageError(page))
2379 ret = -EIO;
2380 }
2381 page_cache_release(page);
2382 } else {
2383 unlock_page(page);
2384 }
2385 return ret;
2386}
2387EXPORT_SYMBOL(write_one_page);
2388
2389/*
Ken Chen76719322007-02-10 01:43:15 -08002390 * For address_spaces which do not use buffers nor write back.
2391 */
2392int __set_page_dirty_no_writeback(struct page *page)
2393{
2394 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08002395 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08002396 return 0;
2397}
2398
2399/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002400 * Helper function for set_page_dirty family.
Greg Thelenc4843a72015-05-22 17:13:16 -04002401 *
2402 * Caller must hold mem_cgroup_begin_page_stat().
2403 *
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002404 * NOTE: This relies on being atomic wrt interrupts.
2405 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002406void account_page_dirtied(struct page *page, struct address_space *mapping,
2407 struct mem_cgroup *memcg)
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002408{
Tejun Heo52ebea72015-05-22 17:13:37 -04002409 struct inode *inode = mapping->host;
2410
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002411 trace_writeback_dirty_page(page, mapping);
2412
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002413 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo52ebea72015-05-22 17:13:37 -04002414 struct bdi_writeback *wb;
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002415
Tejun Heo52ebea72015-05-22 17:13:37 -04002416 inode_attach_wb(inode, page);
2417 wb = inode_to_wb(inode);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002418
Greg Thelenc4843a72015-05-22 17:13:16 -04002419 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002420 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07002421 __inc_zone_page_state(page, NR_DIRTIED);
Tejun Heo52ebea72015-05-22 17:13:37 -04002422 __inc_wb_stat(wb, WB_RECLAIMABLE);
2423 __inc_wb_stat(wb, WB_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002424 task_io_account_write(PAGE_CACHE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06002425 current->nr_dirtied++;
2426 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002427 }
2428}
Michael Rubin679ceac2010-08-20 02:31:26 -07002429EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002430
2431/*
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002432 * Helper function for deaccounting dirty page without writeback.
2433 *
Greg Thelenc4843a72015-05-22 17:13:16 -04002434 * Caller must hold mem_cgroup_begin_page_stat().
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002435 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002436void account_page_cleaned(struct page *page, struct address_space *mapping,
Tejun Heo682aa8e2015-05-28 14:50:53 -04002437 struct mem_cgroup *memcg, struct bdi_writeback *wb)
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002438{
2439 if (mapping_cap_account_dirty(mapping)) {
Greg Thelenc4843a72015-05-22 17:13:16 -04002440 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002441 dec_zone_page_state(page, NR_FILE_DIRTY);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002442 dec_wb_stat(wb, WB_RECLAIMABLE);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002443 task_io_account_cancelled_write(PAGE_CACHE_SIZE);
2444 }
2445}
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002446
2447/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 * For address_spaces which do not use buffers. Just tag the page as dirty in
2449 * its radix tree.
2450 *
2451 * This is also used when a single buffer is being dirtied: we want to set the
2452 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2453 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2454 *
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002455 * The caller must ensure this doesn't race with truncation. Most will simply
2456 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2457 * the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 */
2459int __set_page_dirty_nobuffers(struct page *page)
2460{
Greg Thelenc4843a72015-05-22 17:13:16 -04002461 struct mem_cgroup *memcg;
2462
2463 memcg = mem_cgroup_begin_page_stat(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 if (!TestSetPageDirty(page)) {
2465 struct address_space *mapping = page_mapping(page);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002466 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467
Greg Thelenc4843a72015-05-22 17:13:16 -04002468 if (!mapping) {
2469 mem_cgroup_end_page_stat(memcg);
Andrew Morton8c085402006-12-10 02:19:24 -08002470 return 1;
Greg Thelenc4843a72015-05-22 17:13:16 -04002471 }
Andrew Morton8c085402006-12-10 02:19:24 -08002472
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002473 spin_lock_irqsave(&mapping->tree_lock, flags);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002474 BUG_ON(page_mapping(page) != mapping);
2475 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Greg Thelenc4843a72015-05-22 17:13:16 -04002476 account_page_dirtied(page, mapping, memcg);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002477 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2478 PAGECACHE_TAG_DIRTY);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002479 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Greg Thelenc4843a72015-05-22 17:13:16 -04002480 mem_cgroup_end_page_stat(memcg);
2481
Andrew Morton8c085402006-12-10 02:19:24 -08002482 if (mapping->host) {
2483 /* !PageAnon && !swapper_space */
2484 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002486 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 }
Greg Thelenc4843a72015-05-22 17:13:16 -04002488 mem_cgroup_end_page_stat(memcg);
Andrew Morton4741c9f2006-03-24 03:18:11 -08002489 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490}
2491EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2492
2493/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002494 * Call this whenever redirtying a page, to de-account the dirty counters
2495 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2496 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2497 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2498 * control.
2499 */
2500void account_page_redirty(struct page *page)
2501{
2502 struct address_space *mapping = page->mapping;
Tejun Heo91018132015-05-22 17:13:39 -04002503
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002504 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002505 struct inode *inode = mapping->host;
2506 struct bdi_writeback *wb;
2507 bool locked;
Tejun Heo91018132015-05-22 17:13:39 -04002508
Tejun Heo682aa8e2015-05-28 14:50:53 -04002509 wb = unlocked_inode_to_wb_begin(inode, &locked);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002510 current->nr_dirtied--;
2511 dec_zone_page_state(page, NR_DIRTIED);
Tejun Heo91018132015-05-22 17:13:39 -04002512 dec_wb_stat(wb, WB_DIRTIED);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002513 unlocked_inode_to_wb_end(inode, locked);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002514 }
2515}
2516EXPORT_SYMBOL(account_page_redirty);
2517
2518/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 * When a writepage implementation decides that it doesn't want to write this
2520 * page for some reason, it should redirty the locked page via
2521 * redirty_page_for_writepage() and it should then unlock the page and return 0
2522 */
2523int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2524{
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002525 int ret;
2526
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 wbc->pages_skipped++;
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002528 ret = __set_page_dirty_nobuffers(page);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002529 account_page_redirty(page);
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002530 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531}
2532EXPORT_SYMBOL(redirty_page_for_writepage);
2533
2534/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002535 * Dirty a page.
2536 *
2537 * For pages with a mapping this should be done under the page lock
2538 * for the benefit of asynchronous memory errors who prefer a consistent
2539 * dirty state. This rule can be broken in some special cases,
2540 * but should be better not to.
2541 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 * If the mapping doesn't provide a set_page_dirty a_op, then
2543 * just fall through and assume that it wants buffer_heads.
2544 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002545int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546{
2547 struct address_space *mapping = page_mapping(page);
2548
2549 if (likely(mapping)) {
2550 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002551 /*
2552 * readahead/lru_deactivate_page could remain
2553 * PG_readahead/PG_reclaim due to race with end_page_writeback
2554 * About readahead, if the page is written, the flags would be
2555 * reset. So no problem.
2556 * About lru_deactivate_page, if the page is redirty, the flag
2557 * will be reset. So no problem. but if the page is used by readahead
2558 * it will confuse readahead and make it restart the size rampup
2559 * process. But it's a trivial problem.
2560 */
Naoya Horiguchia4bb3ec2015-04-15 16:13:17 -07002561 if (PageReclaim(page))
2562 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002563#ifdef CONFIG_BLOCK
2564 if (!spd)
2565 spd = __set_page_dirty_buffers;
2566#endif
2567 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002569 if (!PageDirty(page)) {
2570 if (!TestSetPageDirty(page))
2571 return 1;
2572 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 return 0;
2574}
2575EXPORT_SYMBOL(set_page_dirty);
2576
2577/*
2578 * set_page_dirty() is racy if the caller has no reference against
2579 * page->mapping->host, and if the page is unlocked. This is because another
2580 * CPU could truncate the page off the mapping and then free the mapping.
2581 *
2582 * Usually, the page _is_ locked, or the caller is a user-space process which
2583 * holds a reference on the inode by having an open file.
2584 *
2585 * In other cases, the page should be locked before running set_page_dirty().
2586 */
2587int set_page_dirty_lock(struct page *page)
2588{
2589 int ret;
2590
Jens Axboe7eaceac2011-03-10 08:52:07 +01002591 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 ret = set_page_dirty(page);
2593 unlock_page(page);
2594 return ret;
2595}
2596EXPORT_SYMBOL(set_page_dirty_lock);
2597
2598/*
Tejun Heo11f81be2015-05-22 17:13:15 -04002599 * This cancels just the dirty bit on the kernel page itself, it does NOT
2600 * actually remove dirty bits on any mmap's that may be around. It also
2601 * leaves the page tagged dirty, so any sync activity will still find it on
2602 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2603 * look at the dirty bits in the VM.
2604 *
2605 * Doing this should *normally* only ever be done when a page is truncated,
2606 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2607 * this when it notices that somebody has cleaned out all the buffers on a
2608 * page without actually doing it through the VM. Can you say "ext3 is
2609 * horribly ugly"? Thought you could.
2610 */
2611void cancel_dirty_page(struct page *page)
2612{
Greg Thelenc4843a72015-05-22 17:13:16 -04002613 struct address_space *mapping = page_mapping(page);
2614
2615 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002616 struct inode *inode = mapping->host;
2617 struct bdi_writeback *wb;
Greg Thelenc4843a72015-05-22 17:13:16 -04002618 struct mem_cgroup *memcg;
Tejun Heo682aa8e2015-05-28 14:50:53 -04002619 bool locked;
Greg Thelenc4843a72015-05-22 17:13:16 -04002620
2621 memcg = mem_cgroup_begin_page_stat(page);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002622 wb = unlocked_inode_to_wb_begin(inode, &locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002623
2624 if (TestClearPageDirty(page))
Tejun Heo682aa8e2015-05-28 14:50:53 -04002625 account_page_cleaned(page, mapping, memcg, wb);
Greg Thelenc4843a72015-05-22 17:13:16 -04002626
Tejun Heo682aa8e2015-05-28 14:50:53 -04002627 unlocked_inode_to_wb_end(inode, locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002628 mem_cgroup_end_page_stat(memcg);
2629 } else {
2630 ClearPageDirty(page);
2631 }
Tejun Heo11f81be2015-05-22 17:13:15 -04002632}
2633EXPORT_SYMBOL(cancel_dirty_page);
2634
2635/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 * Clear a page's dirty flag, while caring for dirty memory accounting.
2637 * Returns true if the page was previously dirty.
2638 *
2639 * This is for preparing to put the page under writeout. We leave the page
2640 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2641 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2642 * implementation will run either set_page_writeback() or set_page_dirty(),
2643 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2644 * back into sync.
2645 *
2646 * This incoherency between the page's dirty flag and radix-tree tag is
2647 * unfortunate, but it only exists while the page is locked.
2648 */
2649int clear_page_dirty_for_io(struct page *page)
2650{
2651 struct address_space *mapping = page_mapping(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002652 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653
Nick Piggin79352892007-07-19 01:47:22 -07002654 BUG_ON(!PageLocked(page));
2655
Linus Torvalds7658cc22006-12-29 10:00:58 -08002656 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002657 struct inode *inode = mapping->host;
2658 struct bdi_writeback *wb;
2659 struct mem_cgroup *memcg;
2660 bool locked;
2661
Linus Torvalds7658cc22006-12-29 10:00:58 -08002662 /*
2663 * Yes, Virginia, this is indeed insane.
2664 *
2665 * We use this sequence to make sure that
2666 * (a) we account for dirty stats properly
2667 * (b) we tell the low-level filesystem to
2668 * mark the whole page dirty if it was
2669 * dirty in a pagetable. Only to then
2670 * (c) clean the page again and return 1 to
2671 * cause the writeback.
2672 *
2673 * This way we avoid all nasty races with the
2674 * dirty bit in multiple places and clearing
2675 * them concurrently from different threads.
2676 *
2677 * Note! Normally the "set_page_dirty(page)"
2678 * has no effect on the actual dirty bit - since
2679 * that will already usually be set. But we
2680 * need the side effects, and it can help us
2681 * avoid races.
2682 *
2683 * We basically use the page "master dirty bit"
2684 * as a serialization point for all the different
2685 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002686 */
2687 if (page_mkclean(page))
2688 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002689 /*
2690 * We carefully synchronise fault handlers against
2691 * installing a dirty pte and marking the page dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002692 * at this point. We do this by having them hold the
2693 * page lock while dirtying the page, and pages are
2694 * always locked coming in here, so we get the desired
2695 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002696 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002697 memcg = mem_cgroup_begin_page_stat(page);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002698 wb = unlocked_inode_to_wb_begin(inode, &locked);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002699 if (TestClearPageDirty(page)) {
Greg Thelenc4843a72015-05-22 17:13:16 -04002700 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
Andrew Morton8c085402006-12-10 02:19:24 -08002701 dec_zone_page_state(page, NR_FILE_DIRTY);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002702 dec_wb_stat(wb, WB_RECLAIMABLE);
Greg Thelenc4843a72015-05-22 17:13:16 -04002703 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 }
Tejun Heo682aa8e2015-05-28 14:50:53 -04002705 unlocked_inode_to_wb_end(inode, locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002706 mem_cgroup_end_page_stat(memcg);
2707 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002709 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002711EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712
2713int test_clear_page_writeback(struct page *page)
2714{
2715 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002716 struct mem_cgroup *memcg;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002717 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718
Johannes Weiner6de22612015-02-11 15:25:01 -08002719 memcg = mem_cgroup_begin_page_stat(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 if (mapping) {
Tejun Heo91018132015-05-22 17:13:39 -04002721 struct inode *inode = mapping->host;
2722 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 unsigned long flags;
2724
Nick Piggin19fd6232008-07-25 19:45:32 -07002725 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002727 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 radix_tree_tag_clear(&mapping->page_tree,
2729 page_index(page),
2730 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002731 if (bdi_cap_account_writeback(bdi)) {
Tejun Heo91018132015-05-22 17:13:39 -04002732 struct bdi_writeback *wb = inode_to_wb(inode);
2733
2734 __dec_wb_stat(wb, WB_WRITEBACK);
2735 __wb_writeout_inc(wb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002736 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002737 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002738 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 } else {
2740 ret = TestClearPageWriteback(page);
2741 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002742 if (ret) {
Johannes Weinerd7365e72014-10-29 14:50:48 -07002743 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
Andrew Mortond688abf2007-07-19 01:49:17 -07002744 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002745 inc_zone_page_state(page, NR_WRITTEN);
2746 }
Johannes Weiner6de22612015-02-11 15:25:01 -08002747 mem_cgroup_end_page_stat(memcg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 return ret;
2749}
2750
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002751int __test_set_page_writeback(struct page *page, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752{
2753 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002754 struct mem_cgroup *memcg;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002755 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756
Johannes Weiner6de22612015-02-11 15:25:01 -08002757 memcg = mem_cgroup_begin_page_stat(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 if (mapping) {
Tejun Heo91018132015-05-22 17:13:39 -04002759 struct inode *inode = mapping->host;
2760 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 unsigned long flags;
2762
Nick Piggin19fd6232008-07-25 19:45:32 -07002763 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002765 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 radix_tree_tag_set(&mapping->page_tree,
2767 page_index(page),
2768 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002769 if (bdi_cap_account_writeback(bdi))
Tejun Heo91018132015-05-22 17:13:39 -04002770 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002771 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 if (!PageDirty(page))
2773 radix_tree_tag_clear(&mapping->page_tree,
2774 page_index(page),
2775 PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002776 if (!keep_write)
2777 radix_tree_tag_clear(&mapping->page_tree,
2778 page_index(page),
2779 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002780 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 } else {
2782 ret = TestSetPageWriteback(page);
2783 }
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002784 if (!ret) {
Johannes Weinerd7365e72014-10-29 14:50:48 -07002785 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002786 inc_zone_page_state(page, NR_WRITEBACK);
2787 }
Johannes Weiner6de22612015-02-11 15:25:01 -08002788 mem_cgroup_end_page_stat(memcg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 return ret;
2790
2791}
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002792EXPORT_SYMBOL(__test_set_page_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
2794/*
Nick Piggin00128182007-10-16 01:24:40 -07002795 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 * passed tag.
2797 */
2798int mapping_tagged(struct address_space *mapping, int tag)
2799{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002800 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801}
2802EXPORT_SYMBOL(mapping_tagged);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002803
2804/**
2805 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2806 * @page: The page to wait on.
2807 *
2808 * This function determines if the given page is related to a backing device
2809 * that requires page contents to be held stable during writeback. If so, then
2810 * it will wait for any pending writeback to complete.
2811 */
2812void wait_for_stable_page(struct page *page)
2813{
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002814 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2815 wait_on_page_writeback(page);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002816}
2817EXPORT_SYMBOL_GPL(wait_for_stable_page);