blob: 11ff8f75863105b773ee279fee5d6853e4680abb [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 Zijlstra90eec102015-11-16 11:08:45 +01005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
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);
Johannes Weinera8d01432016-01-14 15:20:15 -0800281 /*
282 * Pages reserved for the kernel should not be considered
283 * dirtyable, to prevent a situation where reclaim has to
284 * clean pages in order to balance the zones.
285 */
286 nr_pages -= min(nr_pages, zone->totalreserve_pages);
Johannes Weinera8045522014-01-29 14:05:39 -0800287
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800288 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
289 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800290
291 return nr_pages;
292}
293
Johannes Weiner1edf2232012-01-10 15:06:57 -0800294static unsigned long highmem_dirtyable_memory(unsigned long total)
295{
296#ifdef CONFIG_HIGHMEM
297 int node;
298 unsigned long x = 0;
299
300 for_each_node_state(node, N_HIGH_MEMORY) {
Johannes Weinera8045522014-01-29 14:05:39 -0800301 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
Johannes Weiner1edf2232012-01-10 15:06:57 -0800302
Johannes Weinera8045522014-01-29 14:05:39 -0800303 x += zone_dirtyable_memory(z);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800304 }
305 /*
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800306 * Unreclaimable memory (kernel memory or anonymous memory
307 * without swap) can bring down the dirtyable pages below
308 * the zone's dirty balance reserve and the above calculation
309 * will underflow. However we still want to add in nodes
310 * which are below threshold (negative values) to get a more
311 * accurate calculation but make sure that the total never
312 * underflows.
313 */
314 if ((long)x < 0)
315 x = 0;
316
317 /*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800318 * Make sure that the number of highmem pages is never larger
319 * than the number of the total dirtyable memory. This can only
320 * occur in very strange VM situations but we want to make sure
321 * that this does not occur.
322 */
323 return min(x, total);
324#else
325 return 0;
326#endif
327}
328
329/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800330 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800331 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800332 * Returns the global number of pages potentially available for dirty
333 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800334 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700335static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800336{
337 unsigned long x;
338
Johannes Weinera8045522014-01-29 14:05:39 -0800339 x = global_page_state(NR_FREE_PAGES);
Johannes Weinera8d01432016-01-14 15:20:15 -0800340 /*
341 * Pages reserved for the kernel should not be considered
342 * dirtyable, to prevent a situation where reclaim has to
343 * clean pages in order to balance the zones.
344 */
345 x -= min(x, totalreserve_pages);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800346
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800347 x += global_page_state(NR_INACTIVE_FILE);
348 x += global_page_state(NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800349
Johannes Weiner1edf2232012-01-10 15:06:57 -0800350 if (!vm_highmem_is_dirtyable)
351 x -= highmem_dirtyable_memory(x);
352
353 return x + 1; /* Ensure that we never return 0 */
354}
355
Tejun Heo9fc3a432015-05-22 18:23:30 -0400356/**
357 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
358 * @dtc: dirty_throttle_control of interest
Johannes Weinerccafa282012-01-10 15:07:44 -0800359 *
Tejun Heo9fc3a432015-05-22 18:23:30 -0400360 * Calculate @dtc->thresh and ->bg_thresh considering
361 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
362 * must ensure that @dtc->avail is set before calling this function. The
363 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Johannes Weinerccafa282012-01-10 15:07:44 -0800364 * real-time tasks.
365 */
Tejun Heo9fc3a432015-05-22 18:23:30 -0400366static void domain_dirty_limits(struct dirty_throttle_control *dtc)
367{
368 const unsigned long available_memory = dtc->avail;
369 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
370 unsigned long bytes = vm_dirty_bytes;
371 unsigned long bg_bytes = dirty_background_bytes;
372 unsigned long ratio = vm_dirty_ratio;
373 unsigned long bg_ratio = dirty_background_ratio;
374 unsigned long thresh;
375 unsigned long bg_thresh;
376 struct task_struct *tsk;
377
378 /* gdtc is !NULL iff @dtc is for memcg domain */
379 if (gdtc) {
380 unsigned long global_avail = gdtc->avail;
381
382 /*
383 * The byte settings can't be applied directly to memcg
384 * domains. Convert them to ratios by scaling against
385 * globally available memory.
386 */
387 if (bytes)
388 ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
389 global_avail, 100UL);
390 if (bg_bytes)
391 bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
392 global_avail, 100UL);
393 bytes = bg_bytes = 0;
394 }
395
396 if (bytes)
397 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
398 else
399 thresh = (ratio * available_memory) / 100;
400
401 if (bg_bytes)
402 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
403 else
404 bg_thresh = (bg_ratio * available_memory) / 100;
405
406 if (bg_thresh >= thresh)
407 bg_thresh = thresh / 2;
408 tsk = current;
409 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
410 bg_thresh += bg_thresh / 4;
411 thresh += thresh / 4;
412 }
413 dtc->thresh = thresh;
414 dtc->bg_thresh = bg_thresh;
415
416 /* we should eventually report the domain in the TP */
417 if (!gdtc)
418 trace_global_dirty_state(bg_thresh, thresh);
419}
420
421/**
422 * global_dirty_limits - background-writeback and dirty-throttling thresholds
423 * @pbackground: out parameter for bg_thresh
424 * @pdirty: out parameter for thresh
425 *
426 * Calculate bg_thresh and thresh for global_wb_domain. See
427 * domain_dirty_limits() for details.
428 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800429void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
430{
Tejun Heo9fc3a432015-05-22 18:23:30 -0400431 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
Johannes Weinerccafa282012-01-10 15:07:44 -0800432
Tejun Heo9fc3a432015-05-22 18:23:30 -0400433 gdtc.avail = global_dirtyable_memory();
434 domain_dirty_limits(&gdtc);
Johannes Weinerccafa282012-01-10 15:07:44 -0800435
Tejun Heo9fc3a432015-05-22 18:23:30 -0400436 *pbackground = gdtc.bg_thresh;
437 *pdirty = gdtc.thresh;
Johannes Weinerccafa282012-01-10 15:07:44 -0800438}
439
Johannes Weinera756cf52012-01-10 15:07:49 -0800440/**
Johannes Weinera756cf52012-01-10 15:07:49 -0800441 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
442 * @zone: the zone
443 *
444 * Returns the maximum number of dirty pages allowed in a zone, based
445 * on the zone's dirtyable memory.
446 */
447static unsigned long zone_dirty_limit(struct zone *zone)
448{
449 unsigned long zone_memory = zone_dirtyable_memory(zone);
450 struct task_struct *tsk = current;
451 unsigned long dirty;
452
453 if (vm_dirty_bytes)
454 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
455 zone_memory / global_dirtyable_memory();
456 else
457 dirty = vm_dirty_ratio * zone_memory / 100;
458
459 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
460 dirty += dirty / 4;
461
462 return dirty;
463}
464
465/**
466 * zone_dirty_ok - tells whether a zone is within its dirty limits
467 * @zone: the zone to check
468 *
469 * Returns %true when the dirty pages in @zone are within the zone's
470 * dirty limit, %false if the limit is exceeded.
471 */
472bool zone_dirty_ok(struct zone *zone)
473{
474 unsigned long limit = zone_dirty_limit(zone);
475
476 return zone_page_state(zone, NR_FILE_DIRTY) +
477 zone_page_state(zone, NR_UNSTABLE_NFS) +
478 zone_page_state(zone, NR_WRITEBACK) <= limit;
479}
480
David Rientjes2da02992009-01-06 14:39:31 -0800481int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700482 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800483 loff_t *ppos)
484{
485 int ret;
486
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700487 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800488 if (ret == 0 && write)
489 dirty_background_bytes = 0;
490 return ret;
491}
492
493int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700494 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800495 loff_t *ppos)
496{
497 int ret;
498
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700499 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800500 if (ret == 0 && write)
501 dirty_background_ratio = 0;
502 return ret;
503}
504
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700505int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700506 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700507 loff_t *ppos)
508{
509 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800510 int ret;
511
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700512 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700513 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200514 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800515 vm_dirty_bytes = 0;
516 }
517 return ret;
518}
519
David Rientjes2da02992009-01-06 14:39:31 -0800520int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700521 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800522 loff_t *ppos)
523{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800524 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800525 int ret;
526
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700527 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800528 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200529 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800530 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700531 }
532 return ret;
533}
534
Jan Karaeb608e32012-05-24 18:59:11 +0200535static unsigned long wp_next_time(unsigned long cur_time)
536{
537 cur_time += VM_COMPLETIONS_PERIOD_LEN;
538 /* 0 has a special meaning... */
539 if (!cur_time)
540 return 1;
541 return cur_time;
542}
543
Tejun Heoc7981432015-05-22 18:23:29 -0400544static void wb_domain_writeout_inc(struct wb_domain *dom,
545 struct fprop_local_percpu *completions,
546 unsigned int max_prop_frac)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700547{
Tejun Heoc7981432015-05-22 18:23:29 -0400548 __fprop_inc_percpu_max(&dom->completions, completions,
549 max_prop_frac);
Jan Karaeb608e32012-05-24 18:59:11 +0200550 /* First event after period switching was turned off? */
Tejun Heo380c27c2015-05-22 18:23:21 -0400551 if (!unlikely(dom->period_time)) {
Jan Karaeb608e32012-05-24 18:59:11 +0200552 /*
553 * We can race with other __bdi_writeout_inc calls here but
554 * it does not cause any harm since the resulting time when
555 * timer will fire and what is in writeout_period_time will be
556 * roughly the same.
557 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400558 dom->period_time = wp_next_time(jiffies);
559 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200560 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700561}
562
Tejun Heoc7981432015-05-22 18:23:29 -0400563/*
564 * Increment @wb's writeout completion count and the global writeout
565 * completion count. Called from test_clear_page_writeback().
566 */
567static inline void __wb_writeout_inc(struct bdi_writeback *wb)
568{
Tejun Heo841710a2015-05-22 18:23:33 -0400569 struct wb_domain *cgdom;
570
Tejun Heoc7981432015-05-22 18:23:29 -0400571 __inc_wb_stat(wb, WB_WRITTEN);
572 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
573 wb->bdi->max_prop_frac);
Tejun Heo841710a2015-05-22 18:23:33 -0400574
575 cgdom = mem_cgroup_wb_domain(wb);
576 if (cgdom)
577 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
578 wb->bdi->max_prop_frac);
Tejun Heoc7981432015-05-22 18:23:29 -0400579}
580
Tejun Heo93f78d82015-05-22 17:13:27 -0400581void wb_writeout_inc(struct bdi_writeback *wb)
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700582{
583 unsigned long flags;
584
585 local_irq_save(flags);
Tejun Heo93f78d82015-05-22 17:13:27 -0400586 __wb_writeout_inc(wb);
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700587 local_irq_restore(flags);
588}
Tejun Heo93f78d82015-05-22 17:13:27 -0400589EXPORT_SYMBOL_GPL(wb_writeout_inc);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700590
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700591/*
Jan Karaeb608e32012-05-24 18:59:11 +0200592 * On idle system, we can be called long after we scheduled because we use
593 * deferred timers so count with missed periods.
594 */
595static void writeout_period(unsigned long t)
596{
Tejun Heo380c27c2015-05-22 18:23:21 -0400597 struct wb_domain *dom = (void *)t;
598 int miss_periods = (jiffies - dom->period_time) /
Jan Karaeb608e32012-05-24 18:59:11 +0200599 VM_COMPLETIONS_PERIOD_LEN;
600
Tejun Heo380c27c2015-05-22 18:23:21 -0400601 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
602 dom->period_time = wp_next_time(dom->period_time +
Jan Karaeb608e32012-05-24 18:59:11 +0200603 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
Tejun Heo380c27c2015-05-22 18:23:21 -0400604 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200605 } else {
606 /*
607 * Aging has zeroed all fractions. Stop wasting CPU on period
608 * updates.
609 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400610 dom->period_time = 0;
Jan Karaeb608e32012-05-24 18:59:11 +0200611 }
612}
613
Tejun Heo380c27c2015-05-22 18:23:21 -0400614int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
615{
616 memset(dom, 0, sizeof(*dom));
Tejun Heodcc25ae2015-05-22 18:23:22 -0400617
618 spin_lock_init(&dom->lock);
619
Tejun Heo380c27c2015-05-22 18:23:21 -0400620 init_timer_deferrable(&dom->period_timer);
621 dom->period_timer.function = writeout_period;
622 dom->period_timer.data = (unsigned long)dom;
Tejun Heodcc25ae2015-05-22 18:23:22 -0400623
624 dom->dirty_limit_tstamp = jiffies;
625
Tejun Heo380c27c2015-05-22 18:23:21 -0400626 return fprop_global_init(&dom->completions, gfp);
627}
628
Tejun Heo841710a2015-05-22 18:23:33 -0400629#ifdef CONFIG_CGROUP_WRITEBACK
630void wb_domain_exit(struct wb_domain *dom)
631{
632 del_timer_sync(&dom->period_timer);
633 fprop_global_destroy(&dom->completions);
634}
635#endif
636
Jan Karaeb608e32012-05-24 18:59:11 +0200637/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700638 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
639 * registered backing devices, which, for obvious reasons, can not
640 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700641 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700642static unsigned int bdi_min_ratio;
643
644int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
645{
646 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700647
Jens Axboecfc4ba52009-09-14 13:12:40 +0200648 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700649 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700650 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700651 } else {
652 min_ratio -= bdi->min_ratio;
653 if (bdi_min_ratio + min_ratio < 100) {
654 bdi_min_ratio += min_ratio;
655 bdi->min_ratio += min_ratio;
656 } else {
657 ret = -EINVAL;
658 }
659 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200660 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700661
662 return ret;
663}
664
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700665int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
666{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700667 int ret = 0;
668
669 if (max_ratio > 100)
670 return -EINVAL;
671
Jens Axboecfc4ba52009-09-14 13:12:40 +0200672 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700673 if (bdi->min_ratio > max_ratio) {
674 ret = -EINVAL;
675 } else {
676 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200677 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700678 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200679 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700680
681 return ret;
682}
683EXPORT_SYMBOL(bdi_set_max_ratio);
684
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600685static unsigned long dirty_freerun_ceiling(unsigned long thresh,
686 unsigned long bg_thresh)
687{
688 return (thresh + bg_thresh) / 2;
689}
690
Tejun Heoc7981432015-05-22 18:23:29 -0400691static unsigned long hard_dirty_limit(struct wb_domain *dom,
692 unsigned long thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -0600693{
Tejun Heodcc25ae2015-05-22 18:23:22 -0400694 return max(thresh, dom->dirty_limit);
Wu Fengguangffd1f602011-06-19 22:18:42 -0600695}
696
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400697/*
698 * Memory which can be further allocated to a memcg domain is capped by
699 * system-wide clean memory excluding the amount being used in the domain.
700 */
701static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
702 unsigned long filepages, unsigned long headroom)
Tejun Heoc2aa7232015-05-22 18:23:35 -0400703{
704 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400705 unsigned long clean = filepages - min(filepages, mdtc->dirty);
706 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
707 unsigned long other_clean = global_clean - min(global_clean, clean);
Tejun Heoc2aa7232015-05-22 18:23:35 -0400708
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400709 mdtc->avail = filepages + min(headroom, other_clean);
Christoph Lameter1b424462007-05-06 14:48:59 -0700710}
711
Wu Fengguang6f718652011-03-02 17:14:34 -0600712/**
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400713 * __wb_calc_thresh - @wb's share of dirty throttling threshold
714 * @dtc: dirty_throttle_context of interest
Wu Fengguang1babe182010-08-11 14:17:40 -0700715 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400716 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
Wu Fengguang6f718652011-03-02 17:14:34 -0600717 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600718 *
719 * Note that balance_dirty_pages() will only seriously take it as a hard limit
720 * when sleeping max_pause per page is not enough to keep the dirty pages under
721 * control. For example, when the device is completely stalled due to some error
722 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
723 * In the other normal situations, it acts more gently by throttling the tasks
Tejun Heoa88a3412015-05-22 17:13:28 -0400724 * more (rather than completely block them) when the wb dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600725 *
726 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700727 * - starving fast devices
728 * - piling up dirty pages (that will take long time to sync) on slow devices
729 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400730 * The wb's share of dirty limit will be adapting to its throughput and
Wu Fengguang1babe182010-08-11 14:17:40 -0700731 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
732 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400733static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
Wu Fengguang16c40422010-08-11 14:17:39 -0700734{
Tejun Heoe9f07df2015-05-22 18:23:28 -0400735 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400736 unsigned long thresh = dtc->thresh;
Tejun Heo0d960a32015-05-22 18:23:19 -0400737 u64 wb_thresh;
Wu Fengguang16c40422010-08-11 14:17:39 -0700738 long numerator, denominator;
Tejun Heo693108a2015-05-22 17:13:49 -0400739 unsigned long wb_min_ratio, wb_max_ratio;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700740
Wu Fengguang16c40422010-08-11 14:17:39 -0700741 /*
Tejun Heo0d960a32015-05-22 18:23:19 -0400742 * Calculate this BDI's share of the thresh ratio.
Wu Fengguang16c40422010-08-11 14:17:39 -0700743 */
Tejun Heoe9770b32015-05-22 18:23:27 -0400744 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
Tejun Heo380c27c2015-05-22 18:23:21 -0400745 &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700746
Tejun Heo0d960a32015-05-22 18:23:19 -0400747 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
748 wb_thresh *= numerator;
749 do_div(wb_thresh, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700750
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400751 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
Wu Fengguang16c40422010-08-11 14:17:39 -0700752
Tejun Heo0d960a32015-05-22 18:23:19 -0400753 wb_thresh += (thresh * wb_min_ratio) / 100;
754 if (wb_thresh > (thresh * wb_max_ratio) / 100)
755 wb_thresh = thresh * wb_max_ratio / 100;
Wu Fengguang16c40422010-08-11 14:17:39 -0700756
Tejun Heo0d960a32015-05-22 18:23:19 -0400757 return wb_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758}
759
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400760unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
761{
762 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
763 .thresh = thresh };
764 return __wb_calc_thresh(&gdtc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765}
766
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600767/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700768 * setpoint - dirty 3
769 * f(dirty) := 1.0 + (----------------)
770 * limit - setpoint
771 *
772 * it's a 3rd order polynomial that subjects to
773 *
774 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
775 * (2) f(setpoint) = 1.0 => the balance point
776 * (3) f(limit) = 0 => the hard limit
777 * (4) df/dx <= 0 => negative feedback control
778 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
779 * => fast response on large errors; small oscillation near setpoint
780 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700781static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700782 unsigned long dirty,
783 unsigned long limit)
784{
785 long long pos_ratio;
786 long x;
787
Rik van Rield5c9fde2014-05-06 12:50:01 -0700788 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Tejun Heo464d1382015-04-21 16:49:13 -0400789 (limit - setpoint) | 1);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700790 pos_ratio = x;
791 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
792 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
793 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
794
795 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
796}
797
798/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600799 * Dirty position control.
800 *
801 * (o) global/bdi setpoints
802 *
Tejun Heode1fff32015-05-22 17:13:29 -0400803 * We want the dirty pages be balanced around the global/wb setpoints.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600804 * When the number of dirty pages is higher/lower than the setpoint, the
805 * dirty position control ratio (and hence task dirty ratelimit) will be
806 * decreased/increased to bring the dirty pages back to the setpoint.
807 *
808 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
809 *
810 * if (dirty < setpoint) scale up pos_ratio
811 * if (dirty > setpoint) scale down pos_ratio
812 *
Tejun Heode1fff32015-05-22 17:13:29 -0400813 * if (wb_dirty < wb_setpoint) scale up pos_ratio
814 * if (wb_dirty > wb_setpoint) scale down pos_ratio
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600815 *
816 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
817 *
818 * (o) global control line
819 *
820 * ^ pos_ratio
821 * |
822 * | |<===== global dirty control scope ======>|
823 * 2.0 .............*
824 * | .*
825 * | . *
826 * | . *
827 * | . *
828 * | . *
829 * | . *
830 * 1.0 ................................*
831 * | . . *
832 * | . . *
833 * | . . *
834 * | . . *
835 * | . . *
836 * 0 +------------.------------------.----------------------*------------->
837 * freerun^ setpoint^ limit^ dirty pages
838 *
Tejun Heode1fff32015-05-22 17:13:29 -0400839 * (o) wb control line
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600840 *
841 * ^ pos_ratio
842 * |
843 * | *
844 * | *
845 * | *
846 * | *
847 * | * |<=========== span ============>|
848 * 1.0 .......................*
849 * | . *
850 * | . *
851 * | . *
852 * | . *
853 * | . *
854 * | . *
855 * | . *
856 * | . *
857 * | . *
858 * | . *
859 * | . *
860 * 1/4 ...............................................* * * * * * * * * * * *
861 * | . .
862 * | . .
863 * | . .
864 * 0 +----------------------.-------------------------------.------------->
Tejun Heode1fff32015-05-22 17:13:29 -0400865 * wb_setpoint^ x_intercept^
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600866 *
Tejun Heode1fff32015-05-22 17:13:29 -0400867 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600868 * be smoothly throttled down to normal if it starts high in situations like
869 * - start writing to a slow SD card and a fast disk at the same time. The SD
Tejun Heode1fff32015-05-22 17:13:29 -0400870 * card's wb_dirty may rush to many times higher than wb_setpoint.
871 * - the wb dirty thresh drops quickly due to change of JBOD workload
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600872 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400873static void wb_position_ratio(struct dirty_throttle_control *dtc)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600874{
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400875 struct bdi_writeback *wb = dtc->wb;
Tejun Heoa88a3412015-05-22 17:13:28 -0400876 unsigned long write_bw = wb->avg_write_bandwidth;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400877 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -0400878 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400879 unsigned long wb_thresh = dtc->wb_thresh;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600880 unsigned long x_intercept;
881 unsigned long setpoint; /* dirty pages' target balance point */
Tejun Heode1fff32015-05-22 17:13:29 -0400882 unsigned long wb_setpoint;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600883 unsigned long span;
884 long long pos_ratio; /* for scaling up/down the rate limit */
885 long x;
886
Tejun Heodaddfa32015-05-22 18:23:26 -0400887 dtc->pos_ratio = 0;
888
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400889 if (unlikely(dtc->dirty >= limit))
Tejun Heodaddfa32015-05-22 18:23:26 -0400890 return;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600891
892 /*
893 * global setpoint
894 *
Maxim Patlasov5a537482013-09-11 14:22:46 -0700895 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600896 */
897 setpoint = (freerun + limit) / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400898 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700899
900 /*
901 * The strictlimit feature is a tool preventing mistrusted filesystems
902 * from growing a large number of dirty pages before throttling. For
Tejun Heode1fff32015-05-22 17:13:29 -0400903 * such filesystems balance_dirty_pages always checks wb counters
904 * against wb limits. Even if global "nr_dirty" is under "freerun".
Maxim Patlasov5a537482013-09-11 14:22:46 -0700905 * This is especially important for fuse which sets bdi->max_ratio to
906 * 1% by default. Without strictlimit feature, fuse writeback may
907 * consume arbitrary amount of RAM because it is accounted in
908 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
909 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400910 * Here, in wb_position_ratio(), we calculate pos_ratio based on
Tejun Heode1fff32015-05-22 17:13:29 -0400911 * two values: wb_dirty and wb_thresh. Let's consider an example:
Maxim Patlasov5a537482013-09-11 14:22:46 -0700912 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
913 * limits are set by default to 10% and 20% (background and throttle).
Tejun Heode1fff32015-05-22 17:13:29 -0400914 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
Tejun Heo0d960a32015-05-22 18:23:19 -0400915 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
Tejun Heode1fff32015-05-22 17:13:29 -0400916 * about ~6K pages (as the average of background and throttle wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700917 * limits). The 3rd order polynomial will provide positive feedback if
Tejun Heode1fff32015-05-22 17:13:29 -0400918 * wb_dirty is under wb_setpoint and vice versa.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700919 *
920 * Note, that we cannot use global counters in these calculations
Tejun Heode1fff32015-05-22 17:13:29 -0400921 * because we want to throttle process writing to a strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700922 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
923 * in the example above).
924 */
Tejun Heoa88a3412015-05-22 17:13:28 -0400925 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heode1fff32015-05-22 17:13:29 -0400926 long long wb_pos_ratio;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700927
Tejun Heodaddfa32015-05-22 18:23:26 -0400928 if (dtc->wb_dirty < 8) {
929 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
930 2 << RATELIMIT_CALC_SHIFT);
931 return;
932 }
Maxim Patlasov5a537482013-09-11 14:22:46 -0700933
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400934 if (dtc->wb_dirty >= wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400935 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700936
Tejun Heo970fb012015-05-22 18:23:24 -0400937 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
938 dtc->wb_bg_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700939
Tejun Heode1fff32015-05-22 17:13:29 -0400940 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400941 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700942
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400943 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
Tejun Heode1fff32015-05-22 17:13:29 -0400944 wb_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700945
946 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400947 * Typically, for strictlimit case, wb_setpoint << setpoint
948 * and pos_ratio >> wb_pos_ratio. In the other words global
Maxim Patlasov5a537482013-09-11 14:22:46 -0700949 * state ("dirty") is not limiting factor and we have to
Tejun Heode1fff32015-05-22 17:13:29 -0400950 * make decision based on wb counters. But there is an
Maxim Patlasov5a537482013-09-11 14:22:46 -0700951 * important case when global pos_ratio should get precedence:
952 * global limits are exceeded (e.g. due to activities on other
Tejun Heode1fff32015-05-22 17:13:29 -0400953 * wb's) while given strictlimit wb is below limit.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700954 *
Tejun Heode1fff32015-05-22 17:13:29 -0400955 * "pos_ratio * wb_pos_ratio" would work for the case above,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700956 * but it would look too non-natural for the case of all
Tejun Heode1fff32015-05-22 17:13:29 -0400957 * activity in the system coming from a single strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700958 * with bdi->max_ratio == 100%.
959 *
960 * Note that min() below somewhat changes the dynamics of the
961 * control system. Normally, pos_ratio value can be well over 3
Tejun Heode1fff32015-05-22 17:13:29 -0400962 * (when globally we are at freerun and wb is well below wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700963 * setpoint). Now the maximum pos_ratio in the same situation
964 * is 2. We might want to tweak this if we observe the control
965 * system is too slow to adapt.
966 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400967 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
968 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700969 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600970
971 /*
972 * We have computed basic pos_ratio above based on global situation. If
Tejun Heode1fff32015-05-22 17:13:29 -0400973 * the wb is over/under its share of dirty pages, we want to scale
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600974 * pos_ratio further down/up. That is done by the following mechanism.
975 */
976
977 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400978 * wb setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600979 *
Tejun Heode1fff32015-05-22 17:13:29 -0400980 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600981 *
Tejun Heode1fff32015-05-22 17:13:29 -0400982 * x_intercept - wb_dirty
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600983 * := --------------------------
Tejun Heode1fff32015-05-22 17:13:29 -0400984 * x_intercept - wb_setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600985 *
Tejun Heode1fff32015-05-22 17:13:29 -0400986 * The main wb control line is a linear function that subjects to
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600987 *
Tejun Heode1fff32015-05-22 17:13:29 -0400988 * (1) f(wb_setpoint) = 1.0
989 * (2) k = - 1 / (8 * write_bw) (in single wb case)
990 * or equally: x_intercept = wb_setpoint + 8 * write_bw
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600991 *
Tejun Heode1fff32015-05-22 17:13:29 -0400992 * For single wb case, the dirty pages are observed to fluctuate
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600993 * regularly within range
Tejun Heode1fff32015-05-22 17:13:29 -0400994 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600995 * for various filesystems, where (2) can yield in a reasonable 12.5%
996 * fluctuation range for pos_ratio.
997 *
Tejun Heode1fff32015-05-22 17:13:29 -0400998 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600999 * own size, so move the slope over accordingly and choose a slope that
Tejun Heode1fff32015-05-22 17:13:29 -04001000 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001001 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001002 if (unlikely(wb_thresh > dtc->thresh))
1003 wb_thresh = dtc->thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001004 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001005 * It's very possible that wb_thresh is close to 0 not because the
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001006 * device is slow, but that it has remained inactive for long time.
1007 * Honour such devices a reasonable good (hopefully IO efficient)
1008 * threshold, so that the occasional writes won't be blocked and active
1009 * writes can rampup the threshold quickly.
1010 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001011 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001012 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001013 * scale global setpoint to wb's:
1014 * wb_setpoint = setpoint * wb_thresh / thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001015 */
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001016 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
Tejun Heode1fff32015-05-22 17:13:29 -04001017 wb_setpoint = setpoint * (u64)x >> 16;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001018 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001019 * Use span=(8*write_bw) in single wb case as indicated by
1020 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001021 *
Tejun Heode1fff32015-05-22 17:13:29 -04001022 * wb_thresh thresh - wb_thresh
1023 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1024 * thresh thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001025 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001026 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
Tejun Heode1fff32015-05-22 17:13:29 -04001027 x_intercept = wb_setpoint + span;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001028
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001029 if (dtc->wb_dirty < x_intercept - span / 4) {
1030 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001031 (x_intercept - wb_setpoint) | 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001032 } else
1033 pos_ratio /= 4;
1034
Wu Fengguang8927f662011-08-04 22:16:46 -06001035 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001036 * wb reserve area, safeguard against dirty pool underrun and disk idle
Wu Fengguang8927f662011-08-04 22:16:46 -06001037 * It may push the desired control point of global dirty pages higher
1038 * than setpoint.
1039 */
Tejun Heode1fff32015-05-22 17:13:29 -04001040 x_intercept = wb_thresh / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001041 if (dtc->wb_dirty < x_intercept) {
1042 if (dtc->wb_dirty > x_intercept / 8)
1043 pos_ratio = div_u64(pos_ratio * x_intercept,
1044 dtc->wb_dirty);
Wu Fengguang50657fc2011-10-11 17:06:33 -06001045 else
Wu Fengguang8927f662011-08-04 22:16:46 -06001046 pos_ratio *= 8;
1047 }
1048
Tejun Heodaddfa32015-05-22 18:23:26 -04001049 dtc->pos_ratio = pos_ratio;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001050}
1051
Tejun Heoa88a3412015-05-22 17:13:28 -04001052static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1053 unsigned long elapsed,
1054 unsigned long written)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001055{
1056 const unsigned long period = roundup_pow_of_two(3 * HZ);
Tejun Heoa88a3412015-05-22 17:13:28 -04001057 unsigned long avg = wb->avg_write_bandwidth;
1058 unsigned long old = wb->write_bandwidth;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001059 u64 bw;
1060
1061 /*
1062 * bw = written * HZ / elapsed
1063 *
1064 * bw * elapsed + write_bandwidth * (period - elapsed)
1065 * write_bandwidth = ---------------------------------------------------
1066 * period
Tejun Heoc72efb62015-03-23 00:18:48 -04001067 *
1068 * @written may have decreased due to account_page_redirty().
1069 * Avoid underflowing @bw calculation.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001070 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001071 bw = written - min(written, wb->written_stamp);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001072 bw *= HZ;
1073 if (unlikely(elapsed > period)) {
1074 do_div(bw, elapsed);
1075 avg = bw;
1076 goto out;
1077 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001078 bw += (u64)wb->write_bandwidth * (period - elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001079 bw >>= ilog2(period);
1080
1081 /*
1082 * one more level of smoothing, for filtering out sudden spikes
1083 */
1084 if (avg > old && old >= (unsigned long)bw)
1085 avg -= (avg - old) >> 3;
1086
1087 if (avg < old && old <= (unsigned long)bw)
1088 avg += (old - avg) >> 3;
1089
1090out:
Tejun Heo95a46c62015-05-22 17:13:47 -04001091 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1092 avg = max(avg, 1LU);
1093 if (wb_has_dirty_io(wb)) {
1094 long delta = avg - wb->avg_write_bandwidth;
1095 WARN_ON_ONCE(atomic_long_add_return(delta,
1096 &wb->bdi->tot_write_bandwidth) <= 0);
1097 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001098 wb->write_bandwidth = bw;
1099 wb->avg_write_bandwidth = avg;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001100}
1101
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001102static void update_dirty_limit(struct dirty_throttle_control *dtc)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001103{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001104 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001105 unsigned long thresh = dtc->thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04001106 unsigned long limit = dom->dirty_limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001107
1108 /*
1109 * Follow up in one step.
1110 */
1111 if (limit < thresh) {
1112 limit = thresh;
1113 goto update;
1114 }
1115
1116 /*
1117 * Follow down slowly. Use the higher one as the target, because thresh
1118 * may drop below dirty. This is exactly the reason to introduce
Tejun Heodcc25ae2015-05-22 18:23:22 -04001119 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
Wu Fengguangc42843f2011-03-02 15:54:09 -06001120 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001121 thresh = max(thresh, dtc->dirty);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001122 if (limit > thresh) {
1123 limit -= (limit - thresh) >> 5;
1124 goto update;
1125 }
1126 return;
1127update:
Tejun Heodcc25ae2015-05-22 18:23:22 -04001128 dom->dirty_limit = limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001129}
1130
Tejun Heoe9f07df2015-05-22 18:23:28 -04001131static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001132 unsigned long now)
1133{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001134 struct wb_domain *dom = dtc_dom(dtc);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001135
1136 /*
1137 * check locklessly first to optimize away locking for the most time
1138 */
Tejun Heodcc25ae2015-05-22 18:23:22 -04001139 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
Wu Fengguangc42843f2011-03-02 15:54:09 -06001140 return;
1141
Tejun Heodcc25ae2015-05-22 18:23:22 -04001142 spin_lock(&dom->lock);
1143 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001144 update_dirty_limit(dtc);
Tejun Heodcc25ae2015-05-22 18:23:22 -04001145 dom->dirty_limit_tstamp = now;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001146 }
Tejun Heodcc25ae2015-05-22 18:23:22 -04001147 spin_unlock(&dom->lock);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001148}
1149
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001150/*
Tejun Heode1fff32015-05-22 17:13:29 -04001151 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001152 *
Tejun Heode1fff32015-05-22 17:13:29 -04001153 * Normal wb tasks will be curbed at or below it in long term.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001154 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1155 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001156static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
Tejun Heoa88a3412015-05-22 17:13:28 -04001157 unsigned long dirtied,
1158 unsigned long elapsed)
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001159{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001160 struct bdi_writeback *wb = dtc->wb;
1161 unsigned long dirty = dtc->dirty;
1162 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001163 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Wu Fengguang73811312011-08-26 15:53:24 -06001164 unsigned long setpoint = (freerun + limit) / 2;
Tejun Heoa88a3412015-05-22 17:13:28 -04001165 unsigned long write_bw = wb->avg_write_bandwidth;
1166 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001167 unsigned long dirty_rate;
1168 unsigned long task_ratelimit;
1169 unsigned long balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001170 unsigned long step;
1171 unsigned long x;
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001172 unsigned long shift;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001173
1174 /*
1175 * The dirty rate will match the writeout rate in long term, except
1176 * when dirty pages are truncated by userspace or re-dirtied by FS.
1177 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001178 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001179
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001180 /*
1181 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1182 */
1183 task_ratelimit = (u64)dirty_ratelimit *
Tejun Heodaddfa32015-05-22 18:23:26 -04001184 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001185 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1186
1187 /*
1188 * A linear estimation of the "balanced" throttle rate. The theory is,
Tejun Heode1fff32015-05-22 17:13:29 -04001189 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001190 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1191 * formula will yield the balanced rate limit (write_bw / N).
1192 *
1193 * Note that the expanded form is not a pure rate feedback:
1194 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1195 * but also takes pos_ratio into account:
1196 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1197 *
1198 * (1) is not realistic because pos_ratio also takes part in balancing
1199 * the dirty rate. Consider the state
1200 * pos_ratio = 0.5 (3)
1201 * rate = 2 * (write_bw / N) (4)
1202 * If (1) is used, it will stuck in that state! Because each dd will
1203 * be throttled at
1204 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1205 * yielding
1206 * dirty_rate = N * task_ratelimit = write_bw (6)
1207 * put (6) into (1) we get
1208 * rate_(i+1) = rate_(i) (7)
1209 *
1210 * So we end up using (2) to always keep
1211 * rate_(i+1) ~= (write_bw / N) (8)
1212 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1213 * pos_ratio is able to drive itself to 1.0, which is not only where
1214 * the dirty count meet the setpoint, but also where the slope of
1215 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1216 */
1217 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1218 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001219 /*
1220 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1221 */
1222 if (unlikely(balanced_dirty_ratelimit > write_bw))
1223 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001224
Wu Fengguang73811312011-08-26 15:53:24 -06001225 /*
1226 * We could safely do this and return immediately:
1227 *
Tejun Heode1fff32015-05-22 17:13:29 -04001228 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001229 *
1230 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001231 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001232 * limit the step size.
1233 *
1234 * The below code essentially only uses the relative value of
1235 *
1236 * task_ratelimit - dirty_ratelimit
1237 * = (pos_ratio - 1) * dirty_ratelimit
1238 *
1239 * which reflects the direction and size of dirty position error.
1240 */
1241
1242 /*
1243 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1244 * task_ratelimit is on the same side of dirty_ratelimit, too.
1245 * For example, when
1246 * - dirty_ratelimit > balanced_dirty_ratelimit
1247 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1248 * lowering dirty_ratelimit will help meet both the position and rate
1249 * control targets. Otherwise, don't update dirty_ratelimit if it will
1250 * only help meet the rate target. After all, what the users ultimately
1251 * feel and care are stable dirty rate and small position error.
1252 *
1253 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001254 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001255 * keeps jumping around randomly and can even leap far away at times
1256 * due to the small 200ms estimation period of dirty_rate (we want to
1257 * keep that period small to reduce time lags).
1258 */
1259 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001260
1261 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001262 * For strictlimit case, calculations above were based on wb counters
Tejun Heoa88a3412015-05-22 17:13:28 -04001263 * and limits (starting from pos_ratio = wb_position_ratio() and up to
Maxim Patlasov5a537482013-09-11 14:22:46 -07001264 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
Tejun Heode1fff32015-05-22 17:13:29 -04001265 * Hence, to calculate "step" properly, we have to use wb_dirty as
1266 * "dirty" and wb_setpoint as "setpoint".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001267 *
Tejun Heode1fff32015-05-22 17:13:29 -04001268 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1269 * it's possible that wb_thresh is close to zero due to inactivity
Tejun Heo970fb012015-05-22 18:23:24 -04001270 * of backing device.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001271 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001272 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001273 dirty = dtc->wb_dirty;
1274 if (dtc->wb_dirty < 8)
1275 setpoint = dtc->wb_dirty + 1;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001276 else
Tejun Heo970fb012015-05-22 18:23:24 -04001277 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001278 }
1279
Wu Fengguang73811312011-08-26 15:53:24 -06001280 if (dirty < setpoint) {
Tejun Heoa88a3412015-05-22 17:13:28 -04001281 x = min3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001282 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001283 if (dirty_ratelimit < x)
1284 step = x - dirty_ratelimit;
1285 } else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001286 x = max3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001287 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001288 if (dirty_ratelimit > x)
1289 step = dirty_ratelimit - x;
1290 }
1291
1292 /*
1293 * Don't pursue 100% rate matching. It's impossible since the balanced
1294 * rate itself is constantly fluctuating. So decrease the track speed
1295 * when it gets close to the target. Helps eliminate pointless tremors.
1296 */
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001297 shift = dirty_ratelimit / (2 * step + 1);
1298 if (shift < BITS_PER_LONG)
1299 step = DIV_ROUND_UP(step >> shift, 8);
1300 else
1301 step = 0;
Wu Fengguang73811312011-08-26 15:53:24 -06001302
1303 if (dirty_ratelimit < balanced_dirty_ratelimit)
1304 dirty_ratelimit += step;
1305 else
1306 dirty_ratelimit -= step;
1307
Tejun Heoa88a3412015-05-22 17:13:28 -04001308 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1309 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001310
Tejun Heo5634cc22015-08-18 14:54:56 -07001311 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001312}
1313
Tejun Heoc2aa7232015-05-22 18:23:35 -04001314static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1315 struct dirty_throttle_control *mdtc,
Tejun Heo8a731792015-05-22 18:23:20 -04001316 unsigned long start_time,
1317 bool update_ratelimit)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001318{
Tejun Heoc2aa7232015-05-22 18:23:35 -04001319 struct bdi_writeback *wb = gdtc->wb;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001320 unsigned long now = jiffies;
Tejun Heoa88a3412015-05-22 17:13:28 -04001321 unsigned long elapsed = now - wb->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001322 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001323 unsigned long written;
1324
Tejun Heo8a731792015-05-22 18:23:20 -04001325 lockdep_assert_held(&wb->list_lock);
1326
Wu Fengguange98be2d2010-08-29 11:22:30 -06001327 /*
1328 * rate-limit, only update once every 200ms.
1329 */
1330 if (elapsed < BANDWIDTH_INTERVAL)
1331 return;
1332
Tejun Heoa88a3412015-05-22 17:13:28 -04001333 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1334 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001335
1336 /*
1337 * Skip quiet periods when disk bandwidth is under-utilized.
1338 * (at least 1s idle time between two flusher runs)
1339 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001340 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
Wu Fengguange98be2d2010-08-29 11:22:30 -06001341 goto snapshot;
1342
Tejun Heo8a731792015-05-22 18:23:20 -04001343 if (update_ratelimit) {
Tejun Heoc2aa7232015-05-22 18:23:35 -04001344 domain_update_bandwidth(gdtc, now);
1345 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1346
1347 /*
1348 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1349 * compiler has no way to figure that out. Help it.
1350 */
1351 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1352 domain_update_bandwidth(mdtc, now);
1353 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1354 }
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001355 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001356 wb_update_write_bandwidth(wb, elapsed, written);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001357
1358snapshot:
Tejun Heoa88a3412015-05-22 17:13:28 -04001359 wb->dirtied_stamp = dirtied;
1360 wb->written_stamp = written;
1361 wb->bw_time_stamp = now;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001362}
1363
Tejun Heo8a731792015-05-22 18:23:20 -04001364void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001365{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001366 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1367
Tejun Heoc2aa7232015-05-22 18:23:35 -04001368 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001369}
1370
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001372 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001373 * will look to see if it needs to start dirty throttling.
1374 *
1375 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1376 * global_page_state() too often. So scale it near-sqrt to the safety margin
1377 * (the number of pages we may dirty without exceeding the dirty limits).
1378 */
1379static unsigned long dirty_poll_interval(unsigned long dirty,
1380 unsigned long thresh)
1381{
1382 if (thresh > dirty)
1383 return 1UL << (ilog2(thresh - dirty) >> 1);
1384
1385 return 1;
1386}
1387
Tejun Heoa88a3412015-05-22 17:13:28 -04001388static unsigned long wb_max_pause(struct bdi_writeback *wb,
Tejun Heode1fff32015-05-22 17:13:29 -04001389 unsigned long wb_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001390{
Tejun Heoa88a3412015-05-22 17:13:28 -04001391 unsigned long bw = wb->avg_write_bandwidth;
Fengguang Wue3b6c652013-10-16 13:47:03 -07001392 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001393
1394 /*
1395 * Limit pause time for small memory systems. If sleeping for too long
1396 * time, a small pool of dirty/writeback pages may go empty and disk go
1397 * idle.
1398 *
1399 * 8 serves as the safety ratio.
1400 */
Tejun Heode1fff32015-05-22 17:13:29 -04001401 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001402 t++;
1403
Fengguang Wue3b6c652013-10-16 13:47:03 -07001404 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001405}
1406
Tejun Heoa88a3412015-05-22 17:13:28 -04001407static long wb_min_pause(struct bdi_writeback *wb,
1408 long max_pause,
1409 unsigned long task_ratelimit,
1410 unsigned long dirty_ratelimit,
1411 int *nr_dirtied_pause)
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001412{
Tejun Heoa88a3412015-05-22 17:13:28 -04001413 long hi = ilog2(wb->avg_write_bandwidth);
1414 long lo = ilog2(wb->dirty_ratelimit);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001415 long t; /* target pause */
1416 long pause; /* estimated next pause */
1417 int pages; /* target nr_dirtied_pause */
1418
1419 /* target for 10ms pause on 1-dd case */
1420 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001421
1422 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001423 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1424 * overheads.
1425 *
1426 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001427 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001428 if (hi > lo)
1429 t += (hi - lo) * (10 * HZ) / 1024;
1430
1431 /*
1432 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1433 * on the much more stable dirty_ratelimit. However the next pause time
1434 * will be computed based on task_ratelimit and the two rate limits may
1435 * depart considerably at some time. Especially if task_ratelimit goes
1436 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1437 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1438 * result task_ratelimit won't be executed faithfully, which could
1439 * eventually bring down dirty_ratelimit.
1440 *
1441 * We apply two rules to fix it up:
1442 * 1) try to estimate the next pause time and if necessary, use a lower
1443 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1444 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1445 * 2) limit the target pause time to max_pause/2, so that the normal
1446 * small fluctuations of task_ratelimit won't trigger rule (1) and
1447 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1448 */
1449 t = min(t, 1 + max_pause / 2);
1450 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1451
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001452 /*
1453 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1454 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1455 * When the 16 consecutive reads are often interrupted by some dirty
1456 * throttling pause during the async writes, cfq will go into idles
1457 * (deadline is fine). So push nr_dirtied_pause as high as possible
1458 * until reaches DIRTY_POLL_THRESH=32 pages.
1459 */
1460 if (pages < DIRTY_POLL_THRESH) {
1461 t = max_pause;
1462 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1463 if (pages > DIRTY_POLL_THRESH) {
1464 pages = DIRTY_POLL_THRESH;
1465 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1466 }
1467 }
1468
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001469 pause = HZ * pages / (task_ratelimit + 1);
1470 if (pause > max_pause) {
1471 t = max_pause;
1472 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1473 }
1474
1475 *nr_dirtied_pause = pages;
1476 /*
1477 * The minimal pause time will normally be half the target pause time.
1478 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001479 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001480}
1481
Tejun Heo970fb012015-05-22 18:23:24 -04001482static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
Maxim Patlasov5a537482013-09-11 14:22:46 -07001483{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001484 struct bdi_writeback *wb = dtc->wb;
Tejun Heo93f78d82015-05-22 17:13:27 -04001485 unsigned long wb_reclaimable;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001486
1487 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001488 * wb_thresh is not treated as some limiting factor as
Maxim Patlasov5a537482013-09-11 14:22:46 -07001489 * dirty_thresh, due to reasons
Tejun Heode1fff32015-05-22 17:13:29 -04001490 * - in JBOD setup, wb_thresh can fluctuate a lot
Maxim Patlasov5a537482013-09-11 14:22:46 -07001491 * - in a system with HDD and USB key, the USB key may somehow
Tejun Heode1fff32015-05-22 17:13:29 -04001492 * go into state (wb_dirty >> wb_thresh) either because
1493 * wb_dirty starts high, or because wb_thresh drops low.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001494 * In this case we don't want to hard throttle the USB key
Tejun Heode1fff32015-05-22 17:13:29 -04001495 * dirtiers for 100 seconds until wb_dirty drops under
1496 * wb_thresh. Instead the auxiliary wb control line in
Tejun Heoa88a3412015-05-22 17:13:28 -04001497 * wb_position_ratio() will let the dirtier task progress
Tejun Heode1fff32015-05-22 17:13:29 -04001498 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001499 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -04001500 dtc->wb_thresh = __wb_calc_thresh(dtc);
Tejun Heo970fb012015-05-22 18:23:24 -04001501 dtc->wb_bg_thresh = dtc->thresh ?
1502 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001503
1504 /*
1505 * In order to avoid the stacked BDI deadlock we need
1506 * to ensure we accurately count the 'dirty' pages when
1507 * the threshold is low.
1508 *
1509 * Otherwise it would be possible to get thresh+n pages
1510 * reported dirty, even though there are thresh-m pages
1511 * actually dirty; with m+n sitting in the percpu
1512 * deltas.
1513 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001514 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
Tejun Heo93f78d82015-05-22 17:13:27 -04001515 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001516 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001517 } else {
Tejun Heo93f78d82015-05-22 17:13:27 -04001518 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001519 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001520 }
1521}
1522
Wu Fengguang9d823e82011-06-11 18:10:12 -06001523/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 * balance_dirty_pages() must be called by processes which are generating dirty
1525 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001526 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001527 * If we're over `background_thresh' then the writeback threads are woken to
1528 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001530static void balance_dirty_pages(struct address_space *mapping,
Tejun Heodfb8ae52015-05-22 17:13:40 -04001531 struct bdi_writeback *wb,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001532 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001534 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001535 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001536 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001537 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1538 &mdtc_stor : NULL;
1539 struct dirty_throttle_control *sdtc;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001540 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
Wu Fengguang83712352011-06-11 19:25:42 -06001541 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001542 long pause;
1543 long max_pause;
1544 long min_pause;
1545 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001546 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001547 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001548 unsigned long dirty_ratelimit;
Tejun Heodfb8ae52015-05-22 17:13:40 -04001549 struct backing_dev_info *bdi = wb->bdi;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001550 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001551 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552
1553 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001554 unsigned long now = jiffies;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001555 unsigned long dirty, thresh, bg_thresh;
Yang Shi50e55bf2015-11-20 15:57:10 -08001556 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1557 unsigned long m_thresh = 0;
1558 unsigned long m_bg_thresh = 0;
Wu Fengguang83712352011-06-11 19:25:42 -06001559
Wu Fengguang143dfe82010-08-27 18:45:12 -06001560 /*
1561 * Unstable writes are a feature of certain networked
1562 * filesystems (i.e. NFS) in which data may have been
1563 * written to the server's write cache, but has not yet
1564 * been flushed to permanent storage.
1565 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001566 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1567 global_page_state(NR_UNSTABLE_NFS);
Tejun Heo9fc3a432015-05-22 18:23:30 -04001568 gdtc->avail = global_dirtyable_memory();
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001569 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001570
Tejun Heo9fc3a432015-05-22 18:23:30 -04001571 domain_dirty_limits(gdtc);
Wu Fengguang16c40422010-08-11 14:17:39 -07001572
Maxim Patlasov5a537482013-09-11 14:22:46 -07001573 if (unlikely(strictlimit)) {
Tejun Heo970fb012015-05-22 18:23:24 -04001574 wb_dirty_limits(gdtc);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001575
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001576 dirty = gdtc->wb_dirty;
1577 thresh = gdtc->wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -04001578 bg_thresh = gdtc->wb_bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001579 } else {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001580 dirty = gdtc->dirty;
1581 thresh = gdtc->thresh;
1582 bg_thresh = gdtc->bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001583 }
1584
Tejun Heoc2aa7232015-05-22 18:23:35 -04001585 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001586 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001587
1588 /*
1589 * If @wb belongs to !root memcg, repeat the same
1590 * basic calculations for the memcg domain.
1591 */
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001592 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1593 &mdtc->dirty, &writeback);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001594 mdtc->dirty += writeback;
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001595 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001596
1597 domain_dirty_limits(mdtc);
1598
1599 if (unlikely(strictlimit)) {
1600 wb_dirty_limits(mdtc);
1601 m_dirty = mdtc->wb_dirty;
1602 m_thresh = mdtc->wb_thresh;
1603 m_bg_thresh = mdtc->wb_bg_thresh;
1604 } else {
1605 m_dirty = mdtc->dirty;
1606 m_thresh = mdtc->thresh;
1607 m_bg_thresh = mdtc->bg_thresh;
1608 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001609 }
1610
1611 /*
1612 * Throttle it only when the background writeback cannot
1613 * catch-up. This avoids (excessively) small writeouts
Tejun Heode1fff32015-05-22 17:13:29 -04001614 * when the wb limits are ramping up in case of !strictlimit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001615 *
Tejun Heode1fff32015-05-22 17:13:29 -04001616 * In strictlimit case make decision based on the wb counters
1617 * and limits. Small writeouts when the wb limits are ramping
Maxim Patlasov5a537482013-09-11 14:22:46 -07001618 * up are the price we consciously pay for strictlimit-ing.
Tejun Heoc2aa7232015-05-22 18:23:35 -04001619 *
1620 * If memcg domain is in effect, @dirty should be under
1621 * both global and memcg freerun ceilings.
Wu Fengguang16c40422010-08-11 14:17:39 -07001622 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001623 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1624 (!mdtc ||
1625 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1626 unsigned long intv = dirty_poll_interval(dirty, thresh);
1627 unsigned long m_intv = ULONG_MAX;
1628
Wu Fengguang83712352011-06-11 19:25:42 -06001629 current->dirty_paused_when = now;
1630 current->nr_dirtied = 0;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001631 if (mdtc)
1632 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1633 current->nr_dirtied_pause = min(intv, m_intv);
Wu Fengguang16c40422010-08-11 14:17:39 -07001634 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001635 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001636
Tejun Heobc058732015-05-22 17:13:53 -04001637 if (unlikely(!writeback_in_progress(wb)))
Tejun Heo9ecf48662015-05-22 17:13:54 -04001638 wb_start_background_writeback(wb);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001639
Tejun Heoc2aa7232015-05-22 18:23:35 -04001640 /*
1641 * Calculate global domain's pos_ratio and select the
1642 * global dtc by default.
1643 */
Maxim Patlasov5a537482013-09-11 14:22:46 -07001644 if (!strictlimit)
Tejun Heo970fb012015-05-22 18:23:24 -04001645 wb_dirty_limits(gdtc);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001646
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001647 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1648 ((gdtc->dirty > gdtc->thresh) || strictlimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Tejun Heodaddfa32015-05-22 18:23:26 -04001650 wb_position_ratio(gdtc);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001651 sdtc = gdtc;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001652
Tejun Heoc2aa7232015-05-22 18:23:35 -04001653 if (mdtc) {
1654 /*
1655 * If memcg domain is in effect, calculate its
1656 * pos_ratio. @wb should satisfy constraints from
1657 * both global and memcg domains. Choose the one
1658 * w/ lower pos_ratio.
1659 */
1660 if (!strictlimit)
1661 wb_dirty_limits(mdtc);
1662
1663 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1664 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1665
1666 wb_position_ratio(mdtc);
1667 if (mdtc->pos_ratio < gdtc->pos_ratio)
1668 sdtc = mdtc;
1669 }
Tejun Heodaddfa32015-05-22 18:23:26 -04001670
Tejun Heoa88a3412015-05-22 17:13:28 -04001671 if (dirty_exceeded && !wb->dirty_exceeded)
1672 wb->dirty_exceeded = 1;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001673
Tejun Heo8a731792015-05-22 18:23:20 -04001674 if (time_is_before_jiffies(wb->bw_time_stamp +
1675 BANDWIDTH_INTERVAL)) {
1676 spin_lock(&wb->list_lock);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001677 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
Tejun Heo8a731792015-05-22 18:23:20 -04001678 spin_unlock(&wb->list_lock);
1679 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680
Tejun Heoc2aa7232015-05-22 18:23:35 -04001681 /* throttle according to the chosen dtc */
Tejun Heoa88a3412015-05-22 17:13:28 -04001682 dirty_ratelimit = wb->dirty_ratelimit;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001683 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001684 RATELIMIT_CALC_SHIFT;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001685 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
Tejun Heoa88a3412015-05-22 17:13:28 -04001686 min_pause = wb_min_pause(wb, max_pause,
1687 task_ratelimit, dirty_ratelimit,
1688 &nr_dirtied_pause);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001689
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001690 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001691 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001692 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001693 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 }
Wu Fengguang83712352011-06-11 19:25:42 -06001695 period = HZ * pages_dirtied / task_ratelimit;
1696 pause = period;
1697 if (current->dirty_paused_when)
1698 pause -= now - current->dirty_paused_when;
1699 /*
1700 * For less than 1s think time (ext3/4 may block the dirtier
1701 * for up to 800ms from time to time on 1-HDD; so does xfs,
1702 * however at much less frequency), try to compensate it in
1703 * future periods by updating the virtual time; otherwise just
1704 * do a reset, as it may be a light dirtier.
1705 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001706 if (pause < min_pause) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001707 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001708 sdtc->thresh,
1709 sdtc->bg_thresh,
1710 sdtc->dirty,
1711 sdtc->wb_thresh,
1712 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001713 dirty_ratelimit,
1714 task_ratelimit,
1715 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001716 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001717 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001718 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001719 if (pause < -HZ) {
1720 current->dirty_paused_when = now;
1721 current->nr_dirtied = 0;
1722 } else if (period) {
1723 current->dirty_paused_when += period;
1724 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001725 } else if (current->nr_dirtied_pause <= pages_dirtied)
1726 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001727 break;
1728 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001729 if (unlikely(pause > max_pause)) {
1730 /* for occasional dropped task_ratelimit */
1731 now += min(pause - max_pause, max_pause);
1732 pause = max_pause;
1733 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001734
1735pause:
Tejun Heo5634cc22015-08-18 14:54:56 -07001736 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001737 sdtc->thresh,
1738 sdtc->bg_thresh,
1739 sdtc->dirty,
1740 sdtc->wb_thresh,
1741 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001742 dirty_ratelimit,
1743 task_ratelimit,
1744 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001745 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001746 pause,
1747 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001748 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001749 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001750
Wu Fengguang83712352011-06-11 19:25:42 -06001751 current->dirty_paused_when = now + pause;
1752 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001753 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001754
Wu Fengguangffd1f602011-06-19 22:18:42 -06001755 /*
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001756 * This is typically equal to (dirty < thresh) and can also
1757 * keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001758 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001759 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001760 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001761
Wu Fengguangc5c63432011-12-02 10:21:33 -06001762 /*
1763 * In the case of an unresponding NFS server and the NFS dirty
Tejun Heode1fff32015-05-22 17:13:29 -04001764 * pages exceeds dirty_thresh, give the other good wb's a pipe
Wu Fengguangc5c63432011-12-02 10:21:33 -06001765 * to go through, so that tasks on them still remain responsive.
1766 *
1767 * In theory 1 page is enough to keep the comsumer-producer
1768 * pipe going: the flusher cleans 1 page => the task dirties 1
Tejun Heode1fff32015-05-22 17:13:29 -04001769 * more page. However wb_dirty has accounting errors. So use
Tejun Heo93f78d82015-05-22 17:13:27 -04001770 * the larger and more IO friendly wb_stat_error.
Wu Fengguangc5c63432011-12-02 10:21:33 -06001771 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001772 if (sdtc->wb_dirty <= wb_stat_error(wb))
Wu Fengguangc5c63432011-12-02 10:21:33 -06001773 break;
1774
Jan Kara499d05e2011-11-16 19:34:48 +08001775 if (fatal_signal_pending(current))
1776 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 }
1778
Tejun Heoa88a3412015-05-22 17:13:28 -04001779 if (!dirty_exceeded && wb->dirty_exceeded)
1780 wb->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781
Tejun Heobc058732015-05-22 17:13:53 -04001782 if (writeback_in_progress(wb))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001783 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784
1785 /*
1786 * In laptop mode, we wait until hitting the higher threshold before
1787 * starting background writeout, and then write out all the way down
1788 * to the lower threshold. So slow writers cause minimal disk activity.
1789 *
1790 * In normal mode, we start background writeout at the lower
1791 * background_thresh, to keep the amount of dirty memory low.
1792 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001793 if (laptop_mode)
1794 return;
1795
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001796 if (nr_reclaimable > gdtc->bg_thresh)
Tejun Heo9ecf48662015-05-22 17:13:54 -04001797 wb_start_background_writeback(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798}
1799
Wu Fengguang9d823e82011-06-11 18:10:12 -06001800static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001801
Wu Fengguang54848d72011-04-05 13:21:19 -06001802/*
1803 * Normal tasks are throttled by
1804 * loop {
1805 * dirty tsk->nr_dirtied_pause pages;
1806 * take a snap in balance_dirty_pages();
1807 * }
1808 * However there is a worst case. If every task exit immediately when dirtied
1809 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1810 * called to throttle the page dirties. The solution is to save the not yet
1811 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1812 * randomly into the running tasks. This works well for the above worst case,
1813 * as the new task will pick up and accumulate the old task's leaked dirty
1814 * count and eventually get throttled.
1815 */
1816DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1817
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001819 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001820 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 *
1822 * Processes which are dirtying memory should call in here once for each page
1823 * which was newly dirtied. The function will periodically check the system's
1824 * dirty state and will initiate writeback if needed.
1825 *
1826 * On really big machines, get_writeback_state is expensive, so try to avoid
1827 * calling it too often (ratelimiting). But once we're over the dirty memory
1828 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1829 * from overshooting the limit by (ratelimit_pages) each.
1830 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001831void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832{
Tejun Heodfb8ae52015-05-22 17:13:40 -04001833 struct inode *inode = mapping->host;
1834 struct backing_dev_info *bdi = inode_to_bdi(inode);
1835 struct bdi_writeback *wb = NULL;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001836 int ratelimit;
1837 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838
Wu Fengguang36715ce2011-06-11 17:53:57 -06001839 if (!bdi_cap_account_dirty(bdi))
1840 return;
1841
Tejun Heodfb8ae52015-05-22 17:13:40 -04001842 if (inode_cgwb_enabled(inode))
1843 wb = wb_get_create_current(bdi, GFP_KERNEL);
1844 if (!wb)
1845 wb = &bdi->wb;
1846
Wu Fengguang9d823e82011-06-11 18:10:12 -06001847 ratelimit = current->nr_dirtied_pause;
Tejun Heoa88a3412015-05-22 17:13:28 -04001848 if (wb->dirty_exceeded)
Wu Fengguang9d823e82011-06-11 18:10:12 -06001849 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001851 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001852 /*
1853 * This prevents one CPU to accumulate too many dirtied pages without
1854 * calling into balance_dirty_pages(), which can happen when there are
1855 * 1000+ tasks, all of them start dirtying pages at exactly the same
1856 * time, hence all honoured too large initial task->nr_dirtied_pause.
1857 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001858 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001859 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001860 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001861 else if (unlikely(*p >= ratelimit_pages)) {
1862 *p = 0;
1863 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001865 /*
1866 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1867 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1868 * the dirty throttling and livelock other long-run dirtiers.
1869 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001870 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06001871 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001872 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001873 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1874 *p -= nr_pages_dirtied;
1875 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001877 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001878
1879 if (unlikely(current->nr_dirtied >= ratelimit))
Tejun Heodfb8ae52015-05-22 17:13:40 -04001880 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1881
1882 wb_put(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001884EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885
Tejun Heoaa661bb2015-05-22 18:23:31 -04001886/**
1887 * wb_over_bg_thresh - does @wb need to be written back?
1888 * @wb: bdi_writeback of interest
1889 *
1890 * Determines whether background writeback should keep writing @wb or it's
1891 * clean enough. Returns %true if writeback should continue.
1892 */
1893bool wb_over_bg_thresh(struct bdi_writeback *wb)
1894{
Tejun Heo947e9762015-05-22 18:23:32 -04001895 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001896 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo947e9762015-05-22 18:23:32 -04001897 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001898 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1899 &mdtc_stor : NULL;
Tejun Heoaa661bb2015-05-22 18:23:31 -04001900
Tejun Heo947e9762015-05-22 18:23:32 -04001901 /*
1902 * Similar to balance_dirty_pages() but ignores pages being written
1903 * as we're trying to decide whether to put more under writeback.
1904 */
1905 gdtc->avail = global_dirtyable_memory();
1906 gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1907 global_page_state(NR_UNSTABLE_NFS);
1908 domain_dirty_limits(gdtc);
Tejun Heoaa661bb2015-05-22 18:23:31 -04001909
Tejun Heo947e9762015-05-22 18:23:32 -04001910 if (gdtc->dirty > gdtc->bg_thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04001911 return true;
1912
Tejun Heo947e9762015-05-22 18:23:32 -04001913 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(gdtc))
Tejun Heoaa661bb2015-05-22 18:23:31 -04001914 return true;
1915
Tejun Heoc2aa7232015-05-22 18:23:35 -04001916 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001917 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001918
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001919 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1920 &writeback);
1921 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001922 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1923
1924 if (mdtc->dirty > mdtc->bg_thresh)
1925 return true;
1926
1927 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(mdtc))
1928 return true;
1929 }
1930
Tejun Heoaa661bb2015-05-22 18:23:31 -04001931 return false;
1932}
1933
Andrew Morton232ea4d2007-02-28 20:13:21 -08001934void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935{
David Rientjes364aeb22009-01-06 14:39:29 -08001936 unsigned long background_thresh;
1937 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
1939 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001940 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001941 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942
1943 /*
1944 * Boost the allowable dirty threshold a bit for page
1945 * allocators so they don't get DoS'ed by heavy writers
1946 */
1947 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1948
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001949 if (global_page_state(NR_UNSTABLE_NFS) +
1950 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1951 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001952 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001953
1954 /*
1955 * The caller might hold locks which can prevent IO completion
1956 * or progress in the filesystem. So we cannot just sit here
1957 * waiting for IO to complete.
1958 */
1959 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1960 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 }
1962}
1963
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1966 */
Joe Perchescccad5b2014-06-06 14:38:09 -07001967int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001968 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001970 proc_dointvec(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 return 0;
1972}
1973
Jens Axboec2c49862010-05-20 09:18:47 +02001974#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001975void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976{
Matthew Garrett31373d02010-04-06 14:25:14 +02001977 struct request_queue *q = (struct request_queue *)data;
1978 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1979 global_page_state(NR_UNSTABLE_NFS);
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001980 struct bdi_writeback *wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981
Matthew Garrett31373d02010-04-06 14:25:14 +02001982 /*
1983 * We want to write everything out, not just down to the dirty
1984 * threshold
1985 */
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001986 if (!bdi_has_dirty_io(&q->backing_dev_info))
1987 return;
1988
Tejun Heo9ad18ab2015-09-29 12:47:50 -04001989 rcu_read_lock();
Tejun Heob8175252015-10-02 14:47:05 -04001990 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001991 if (wb_has_dirty_io(wb))
1992 wb_start_writeback(wb, nr_pages, true,
1993 WB_REASON_LAPTOP_TIMER);
Tejun Heo9ad18ab2015-09-29 12:47:50 -04001994 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995}
1996
1997/*
1998 * We've spun up the disk and we're in laptop mode: schedule writeback
1999 * of all dirty data a few seconds from now. If the flush is already scheduled
2000 * then push it back - the user is still using the disk.
2001 */
Matthew Garrett31373d02010-04-06 14:25:14 +02002002void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003{
Matthew Garrett31373d02010-04-06 14:25:14 +02002004 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005}
2006
2007/*
2008 * We're in laptop mode and we've just synced. The sync's writes will have
2009 * caused another writeback to be scheduled by laptop_io_completion.
2010 * Nothing needs to be written back anymore, so we unschedule the writeback.
2011 */
2012void laptop_sync_completion(void)
2013{
Matthew Garrett31373d02010-04-06 14:25:14 +02002014 struct backing_dev_info *bdi;
2015
2016 rcu_read_lock();
2017
2018 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2019 del_timer(&bdi->laptop_mode_wb_timer);
2020
2021 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022}
Jens Axboec2c49862010-05-20 09:18:47 +02002023#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024
2025/*
2026 * If ratelimit_pages is too high then we can get into dirty-data overload
2027 * if a large number of processes all perform writes at the same time.
2028 * If it is too low then SMP machines will call the (expensive)
2029 * get_writeback_state too often.
2030 *
2031 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2032 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06002033 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 */
2035
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002036void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037{
Tejun Heodcc25ae2015-05-22 18:23:22 -04002038 struct wb_domain *dom = &global_wb_domain;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002039 unsigned long background_thresh;
2040 unsigned long dirty_thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04002041
Wu Fengguang9d823e82011-06-11 18:10:12 -06002042 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heodcc25ae2015-05-22 18:23:22 -04002043 dom->dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002044 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 if (ratelimit_pages < 16)
2046 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047}
2048
Paul Gortmaker0db06282013-06-19 14:53:51 -04002049static int
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08002050ratelimit_handler(struct notifier_block *self, unsigned long action,
2051 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052{
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08002053
2054 switch (action & ~CPU_TASKS_FROZEN) {
2055 case CPU_ONLINE:
2056 case CPU_DEAD:
2057 writeback_set_ratelimit();
2058 return NOTIFY_OK;
2059 default:
2060 return NOTIFY_DONE;
2061 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062}
2063
Paul Gortmaker0db06282013-06-19 14:53:51 -04002064static struct notifier_block ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 .notifier_call = ratelimit_handler,
2066 .next = NULL,
2067};
2068
2069/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002070 * Called early on to tune the page writeback dirty limits.
2071 *
2072 * We used to scale dirty pages according to how total memory
2073 * related to pages that could be allocated for buffers (by
2074 * comparing nr_free_buffer_pages() to vm_total_pages.
2075 *
2076 * However, that was when we used "dirty_ratio" to scale with
2077 * all memory, and we don't do that any more. "dirty_ratio"
2078 * is now applied to total non-HIGHPAGE memory (by subtracting
2079 * totalhigh_pages from vm_total_pages), and as such we can't
2080 * get into the old insane situation any more where we had
2081 * large amounts of dirty pages compared to a small amount of
2082 * non-HIGHMEM memory.
2083 *
2084 * But we might still want to scale the dirty_ratio by how
2085 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 */
2087void __init page_writeback_init(void)
2088{
Rabin Vincenta50fcb52015-08-06 15:47:14 -07002089 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2090
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002091 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 register_cpu_notifier(&ratelimit_nb);
2093}
2094
David Howells811d7362006-08-29 19:06:09 +01002095/**
Jan Karaf446daae2010-08-09 17:19:12 -07002096 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2097 * @mapping: address space structure to write
2098 * @start: starting page index
2099 * @end: ending page index (inclusive)
2100 *
2101 * This function scans the page range from @start to @end (inclusive) and tags
2102 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2103 * that write_cache_pages (or whoever calls this function) will then use
2104 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2105 * used to avoid livelocking of writeback by a process steadily creating new
2106 * dirty pages in the file (thus it is important for this function to be quick
2107 * so that it can tag pages faster than a dirtying process can create them).
2108 */
2109/*
2110 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2111 */
Jan Karaf446daae2010-08-09 17:19:12 -07002112void tag_pages_for_writeback(struct address_space *mapping,
2113 pgoff_t start, pgoff_t end)
2114{
Randy Dunlap3c111a02010-08-11 14:17:30 -07002115#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07002116 unsigned long tagged;
2117
2118 do {
2119 spin_lock_irq(&mapping->tree_lock);
2120 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2121 &start, end, WRITEBACK_TAG_BATCH,
2122 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2123 spin_unlock_irq(&mapping->tree_lock);
2124 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2125 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07002126 /* We check 'start' to handle wrapping when end == ~0UL */
2127 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07002128}
2129EXPORT_SYMBOL(tag_pages_for_writeback);
2130
2131/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002132 * 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 +01002133 * @mapping: address space structure to write
2134 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002135 * @writepage: function called for each page
2136 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01002137 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002138 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01002139 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2140 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2141 * and msync() need to guarantee that all the data which was dirty at the time
2142 * the call was made get new I/O started against them. If wbc->sync_mode is
2143 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2144 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07002145 *
2146 * To avoid livelocks (when other process dirties new pages), we first tag
2147 * pages which should be written back with TOWRITE tag and only then start
2148 * writing them. For data-integrity sync we have to be careful so that we do
2149 * not miss some pages (e.g., because some other process has cleared TOWRITE
2150 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2151 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01002152 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002153int write_cache_pages(struct address_space *mapping,
2154 struct writeback_control *wbc, writepage_t writepage,
2155 void *data)
David Howells811d7362006-08-29 19:06:09 +01002156{
David Howells811d7362006-08-29 19:06:09 +01002157 int ret = 0;
2158 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01002159 struct pagevec pvec;
2160 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08002161 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01002162 pgoff_t index;
2163 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08002164 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08002165 int cycled;
David Howells811d7362006-08-29 19:06:09 +01002166 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07002167 int tag;
David Howells811d7362006-08-29 19:06:09 +01002168
David Howells811d7362006-08-29 19:06:09 +01002169 pagevec_init(&pvec, 0);
2170 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08002171 writeback_index = mapping->writeback_index; /* prev offset */
2172 index = writeback_index;
2173 if (index == 0)
2174 cycled = 1;
2175 else
2176 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01002177 end = -1;
2178 } else {
2179 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2180 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2181 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2182 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08002183 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01002184 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002185 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002186 tag = PAGECACHE_TAG_TOWRITE;
2187 else
2188 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01002189retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002190 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002191 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08002192 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002193 while (!done && (index <= end)) {
2194 int i;
2195
Jan Karaf446daae2010-08-09 17:19:12 -07002196 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002197 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2198 if (nr_pages == 0)
2199 break;
David Howells811d7362006-08-29 19:06:09 +01002200
David Howells811d7362006-08-29 19:06:09 +01002201 for (i = 0; i < nr_pages; i++) {
2202 struct page *page = pvec.pages[i];
2203
Nick Piggind5482cd2009-01-06 14:39:11 -08002204 /*
2205 * At this point, the page may be truncated or
2206 * invalidated (changing page->mapping to NULL), or
2207 * even swizzled back from swapper_space to tmpfs file
2208 * mapping. However, page->index will not change
2209 * because we have a reference on the page.
2210 */
2211 if (page->index > end) {
2212 /*
2213 * can't be range_cyclic (1st pass) because
2214 * end == -1 in that case.
2215 */
2216 done = 1;
2217 break;
2218 }
2219
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002220 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08002221
David Howells811d7362006-08-29 19:06:09 +01002222 lock_page(page);
2223
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002224 /*
2225 * Page truncated or invalidated. We can freely skip it
2226 * then, even for data integrity operations: the page
2227 * has disappeared concurrently, so there could be no
2228 * real expectation of this data interity operation
2229 * even if there is now a new, dirty page at the same
2230 * pagecache address.
2231 */
David Howells811d7362006-08-29 19:06:09 +01002232 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002233continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01002234 unlock_page(page);
2235 continue;
2236 }
2237
Nick Piggin515f4a02009-01-06 14:39:10 -08002238 if (!PageDirty(page)) {
2239 /* someone wrote it for us */
2240 goto continue_unlock;
2241 }
David Howells811d7362006-08-29 19:06:09 +01002242
Nick Piggin515f4a02009-01-06 14:39:10 -08002243 if (PageWriteback(page)) {
2244 if (wbc->sync_mode != WB_SYNC_NONE)
2245 wait_on_page_writeback(page);
2246 else
2247 goto continue_unlock;
2248 }
2249
2250 BUG_ON(PageWriteback(page));
2251 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002252 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01002253
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002254 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002255 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08002256 if (unlikely(ret)) {
2257 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2258 unlock_page(page);
2259 ret = 0;
2260 } else {
2261 /*
2262 * done_index is set past this page,
2263 * so media errors will not choke
2264 * background writeout for the entire
2265 * file. This has consequences for
2266 * range_cyclic semantics (ie. it may
2267 * not be suitable for data integrity
2268 * writeout).
2269 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002270 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08002271 done = 1;
2272 break;
2273 }
Dave Chinner0b564922010-06-09 10:37:18 +10002274 }
David Howells811d7362006-08-29 19:06:09 +01002275
Dave Chinner546a1922010-08-24 11:44:34 +10002276 /*
2277 * We stop writing back only if we are not doing
2278 * integrity sync. In case of integrity sync we have to
2279 * keep going until we have written all the pages
2280 * we tagged for writeback prior to entering this loop.
2281 */
2282 if (--wbc->nr_to_write <= 0 &&
2283 wbc->sync_mode == WB_SYNC_NONE) {
2284 done = 1;
2285 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08002286 }
David Howells811d7362006-08-29 19:06:09 +01002287 }
2288 pagevec_release(&pvec);
2289 cond_resched();
2290 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01002291 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01002292 /*
Nick Piggin31a12662009-01-06 14:39:04 -08002293 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01002294 * We hit the last page and there is more work to be done: wrap
2295 * back to the start of the file
2296 */
Nick Piggin31a12662009-01-06 14:39:04 -08002297 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01002298 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08002299 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01002300 goto retry;
2301 }
Dave Chinner0b564922010-06-09 10:37:18 +10002302 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2303 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04002304
David Howells811d7362006-08-29 19:06:09 +01002305 return ret;
2306}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002307EXPORT_SYMBOL(write_cache_pages);
2308
2309/*
2310 * Function used by generic_writepages to call the real writepage
2311 * function and set the mapping flags on error
2312 */
2313static int __writepage(struct page *page, struct writeback_control *wbc,
2314 void *data)
2315{
2316 struct address_space *mapping = data;
2317 int ret = mapping->a_ops->writepage(page, wbc);
2318 mapping_set_error(mapping, ret);
2319 return ret;
2320}
2321
2322/**
2323 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2324 * @mapping: address space structure to write
2325 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2326 *
2327 * This is a library function, which implements the writepages()
2328 * address_space_operation.
2329 */
2330int generic_writepages(struct address_space *mapping,
2331 struct writeback_control *wbc)
2332{
Shaohua Li9b6096a2011-03-17 10:47:06 +01002333 struct blk_plug plug;
2334 int ret;
2335
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002336 /* deal with chardevs and other special file */
2337 if (!mapping->a_ops->writepage)
2338 return 0;
2339
Shaohua Li9b6096a2011-03-17 10:47:06 +01002340 blk_start_plug(&plug);
2341 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2342 blk_finish_plug(&plug);
2343 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002344}
David Howells811d7362006-08-29 19:06:09 +01002345
2346EXPORT_SYMBOL(generic_writepages);
2347
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2349{
Andrew Morton22905f72005-11-16 15:07:01 -08002350 int ret;
2351
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 if (wbc->nr_to_write <= 0)
2353 return 0;
2354 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002355 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002356 else
2357 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002358 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359}
2360
2361/**
2362 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07002363 * @page: the page to write
2364 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 *
2366 * The page must be locked by the caller and will be unlocked upon return.
2367 *
2368 * write_one_page() returns a negative error code if I/O failed.
2369 */
2370int write_one_page(struct page *page, int wait)
2371{
2372 struct address_space *mapping = page->mapping;
2373 int ret = 0;
2374 struct writeback_control wbc = {
2375 .sync_mode = WB_SYNC_ALL,
2376 .nr_to_write = 1,
2377 };
2378
2379 BUG_ON(!PageLocked(page));
2380
2381 if (wait)
2382 wait_on_page_writeback(page);
2383
2384 if (clear_page_dirty_for_io(page)) {
2385 page_cache_get(page);
2386 ret = mapping->a_ops->writepage(page, &wbc);
2387 if (ret == 0 && wait) {
2388 wait_on_page_writeback(page);
2389 if (PageError(page))
2390 ret = -EIO;
2391 }
2392 page_cache_release(page);
2393 } else {
2394 unlock_page(page);
2395 }
2396 return ret;
2397}
2398EXPORT_SYMBOL(write_one_page);
2399
2400/*
Ken Chen76719322007-02-10 01:43:15 -08002401 * For address_spaces which do not use buffers nor write back.
2402 */
2403int __set_page_dirty_no_writeback(struct page *page)
2404{
2405 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08002406 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08002407 return 0;
2408}
2409
2410/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002411 * Helper function for set_page_dirty family.
Greg Thelenc4843a72015-05-22 17:13:16 -04002412 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002413 * Caller must hold lock_page_memcg().
Greg Thelenc4843a72015-05-22 17:13:16 -04002414 *
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002415 * NOTE: This relies on being atomic wrt interrupts.
2416 */
Johannes Weiner62cccb82016-03-15 14:57:22 -07002417void account_page_dirtied(struct page *page, struct address_space *mapping)
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002418{
Tejun Heo52ebea72015-05-22 17:13:37 -04002419 struct inode *inode = mapping->host;
2420
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002421 trace_writeback_dirty_page(page, mapping);
2422
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002423 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo52ebea72015-05-22 17:13:37 -04002424 struct bdi_writeback *wb;
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002425
Tejun Heo52ebea72015-05-22 17:13:37 -04002426 inode_attach_wb(inode, page);
2427 wb = inode_to_wb(inode);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002428
Johannes Weiner62cccb82016-03-15 14:57:22 -07002429 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_DIRTY);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002430 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07002431 __inc_zone_page_state(page, NR_DIRTIED);
Tejun Heo52ebea72015-05-22 17:13:37 -04002432 __inc_wb_stat(wb, WB_RECLAIMABLE);
2433 __inc_wb_stat(wb, WB_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002434 task_io_account_write(PAGE_CACHE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06002435 current->nr_dirtied++;
2436 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002437 }
2438}
Michael Rubin679ceac2010-08-20 02:31:26 -07002439EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002440
2441/*
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002442 * Helper function for deaccounting dirty page without writeback.
2443 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002444 * Caller must hold lock_page_memcg().
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002445 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002446void account_page_cleaned(struct page *page, struct address_space *mapping,
Johannes Weiner62cccb82016-03-15 14:57:22 -07002447 struct bdi_writeback *wb)
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002448{
2449 if (mapping_cap_account_dirty(mapping)) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002450 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002451 dec_zone_page_state(page, NR_FILE_DIRTY);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002452 dec_wb_stat(wb, WB_RECLAIMABLE);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002453 task_io_account_cancelled_write(PAGE_CACHE_SIZE);
2454 }
2455}
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002456
2457/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 * For address_spaces which do not use buffers. Just tag the page as dirty in
2459 * its radix tree.
2460 *
2461 * This is also used when a single buffer is being dirtied: we want to set the
2462 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2463 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2464 *
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002465 * The caller must ensure this doesn't race with truncation. Most will simply
2466 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2467 * the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 */
2469int __set_page_dirty_nobuffers(struct page *page)
2470{
Johannes Weiner62cccb82016-03-15 14:57:22 -07002471 lock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 if (!TestSetPageDirty(page)) {
2473 struct address_space *mapping = page_mapping(page);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002474 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Greg Thelenc4843a72015-05-22 17:13:16 -04002476 if (!mapping) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002477 unlock_page_memcg(page);
Andrew Morton8c085402006-12-10 02:19:24 -08002478 return 1;
Greg Thelenc4843a72015-05-22 17:13:16 -04002479 }
Andrew Morton8c085402006-12-10 02:19:24 -08002480
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002481 spin_lock_irqsave(&mapping->tree_lock, flags);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002482 BUG_ON(page_mapping(page) != mapping);
2483 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Johannes Weiner62cccb82016-03-15 14:57:22 -07002484 account_page_dirtied(page, mapping);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002485 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2486 PAGECACHE_TAG_DIRTY);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002487 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner62cccb82016-03-15 14:57:22 -07002488 unlock_page_memcg(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002489
Andrew Morton8c085402006-12-10 02:19:24 -08002490 if (mapping->host) {
2491 /* !PageAnon && !swapper_space */
2492 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002494 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07002496 unlock_page_memcg(page);
Andrew Morton4741c9f2006-03-24 03:18:11 -08002497 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498}
2499EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2500
2501/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002502 * Call this whenever redirtying a page, to de-account the dirty counters
2503 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2504 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2505 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2506 * control.
2507 */
2508void account_page_redirty(struct page *page)
2509{
2510 struct address_space *mapping = page->mapping;
Tejun Heo91018132015-05-22 17:13:39 -04002511
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002512 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002513 struct inode *inode = mapping->host;
2514 struct bdi_writeback *wb;
2515 bool locked;
Tejun Heo91018132015-05-22 17:13:39 -04002516
Tejun Heo682aa8e2015-05-28 14:50:53 -04002517 wb = unlocked_inode_to_wb_begin(inode, &locked);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002518 current->nr_dirtied--;
2519 dec_zone_page_state(page, NR_DIRTIED);
Tejun Heo91018132015-05-22 17:13:39 -04002520 dec_wb_stat(wb, WB_DIRTIED);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002521 unlocked_inode_to_wb_end(inode, locked);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002522 }
2523}
2524EXPORT_SYMBOL(account_page_redirty);
2525
2526/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 * When a writepage implementation decides that it doesn't want to write this
2528 * page for some reason, it should redirty the locked page via
2529 * redirty_page_for_writepage() and it should then unlock the page and return 0
2530 */
2531int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2532{
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002533 int ret;
2534
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 wbc->pages_skipped++;
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002536 ret = __set_page_dirty_nobuffers(page);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002537 account_page_redirty(page);
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002538 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539}
2540EXPORT_SYMBOL(redirty_page_for_writepage);
2541
2542/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002543 * Dirty a page.
2544 *
2545 * For pages with a mapping this should be done under the page lock
2546 * for the benefit of asynchronous memory errors who prefer a consistent
2547 * dirty state. This rule can be broken in some special cases,
2548 * but should be better not to.
2549 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 * If the mapping doesn't provide a set_page_dirty a_op, then
2551 * just fall through and assume that it wants buffer_heads.
2552 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002553int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554{
2555 struct address_space *mapping = page_mapping(page);
2556
2557 if (likely(mapping)) {
2558 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002559 /*
2560 * readahead/lru_deactivate_page could remain
2561 * PG_readahead/PG_reclaim due to race with end_page_writeback
2562 * About readahead, if the page is written, the flags would be
2563 * reset. So no problem.
2564 * About lru_deactivate_page, if the page is redirty, the flag
2565 * will be reset. So no problem. but if the page is used by readahead
2566 * it will confuse readahead and make it restart the size rampup
2567 * process. But it's a trivial problem.
2568 */
Naoya Horiguchia4bb3ec2015-04-15 16:13:17 -07002569 if (PageReclaim(page))
2570 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002571#ifdef CONFIG_BLOCK
2572 if (!spd)
2573 spd = __set_page_dirty_buffers;
2574#endif
2575 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002577 if (!PageDirty(page)) {
2578 if (!TestSetPageDirty(page))
2579 return 1;
2580 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581 return 0;
2582}
2583EXPORT_SYMBOL(set_page_dirty);
2584
2585/*
2586 * set_page_dirty() is racy if the caller has no reference against
2587 * page->mapping->host, and if the page is unlocked. This is because another
2588 * CPU could truncate the page off the mapping and then free the mapping.
2589 *
2590 * Usually, the page _is_ locked, or the caller is a user-space process which
2591 * holds a reference on the inode by having an open file.
2592 *
2593 * In other cases, the page should be locked before running set_page_dirty().
2594 */
2595int set_page_dirty_lock(struct page *page)
2596{
2597 int ret;
2598
Jens Axboe7eaceac2011-03-10 08:52:07 +01002599 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 ret = set_page_dirty(page);
2601 unlock_page(page);
2602 return ret;
2603}
2604EXPORT_SYMBOL(set_page_dirty_lock);
2605
2606/*
Tejun Heo11f81be2015-05-22 17:13:15 -04002607 * This cancels just the dirty bit on the kernel page itself, it does NOT
2608 * actually remove dirty bits on any mmap's that may be around. It also
2609 * leaves the page tagged dirty, so any sync activity will still find it on
2610 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2611 * look at the dirty bits in the VM.
2612 *
2613 * Doing this should *normally* only ever be done when a page is truncated,
2614 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2615 * this when it notices that somebody has cleaned out all the buffers on a
2616 * page without actually doing it through the VM. Can you say "ext3 is
2617 * horribly ugly"? Thought you could.
2618 */
2619void cancel_dirty_page(struct page *page)
2620{
Greg Thelenc4843a72015-05-22 17:13:16 -04002621 struct address_space *mapping = page_mapping(page);
2622
2623 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002624 struct inode *inode = mapping->host;
2625 struct bdi_writeback *wb;
Tejun Heo682aa8e2015-05-28 14:50:53 -04002626 bool locked;
Greg Thelenc4843a72015-05-22 17:13:16 -04002627
Johannes Weiner62cccb82016-03-15 14:57:22 -07002628 lock_page_memcg(page);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002629 wb = unlocked_inode_to_wb_begin(inode, &locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002630
2631 if (TestClearPageDirty(page))
Johannes Weiner62cccb82016-03-15 14:57:22 -07002632 account_page_cleaned(page, mapping, wb);
Greg Thelenc4843a72015-05-22 17:13:16 -04002633
Tejun Heo682aa8e2015-05-28 14:50:53 -04002634 unlocked_inode_to_wb_end(inode, locked);
Johannes Weiner62cccb82016-03-15 14:57:22 -07002635 unlock_page_memcg(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002636 } else {
2637 ClearPageDirty(page);
2638 }
Tejun Heo11f81be2015-05-22 17:13:15 -04002639}
2640EXPORT_SYMBOL(cancel_dirty_page);
2641
2642/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 * Clear a page's dirty flag, while caring for dirty memory accounting.
2644 * Returns true if the page was previously dirty.
2645 *
2646 * This is for preparing to put the page under writeout. We leave the page
2647 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2648 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2649 * implementation will run either set_page_writeback() or set_page_dirty(),
2650 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2651 * back into sync.
2652 *
2653 * This incoherency between the page's dirty flag and radix-tree tag is
2654 * unfortunate, but it only exists while the page is locked.
2655 */
2656int clear_page_dirty_for_io(struct page *page)
2657{
2658 struct address_space *mapping = page_mapping(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002659 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660
Nick Piggin79352892007-07-19 01:47:22 -07002661 BUG_ON(!PageLocked(page));
2662
Linus Torvalds7658cc22006-12-29 10:00:58 -08002663 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002664 struct inode *inode = mapping->host;
2665 struct bdi_writeback *wb;
Tejun Heo682aa8e2015-05-28 14:50:53 -04002666 bool locked;
2667
Linus Torvalds7658cc22006-12-29 10:00:58 -08002668 /*
2669 * Yes, Virginia, this is indeed insane.
2670 *
2671 * We use this sequence to make sure that
2672 * (a) we account for dirty stats properly
2673 * (b) we tell the low-level filesystem to
2674 * mark the whole page dirty if it was
2675 * dirty in a pagetable. Only to then
2676 * (c) clean the page again and return 1 to
2677 * cause the writeback.
2678 *
2679 * This way we avoid all nasty races with the
2680 * dirty bit in multiple places and clearing
2681 * them concurrently from different threads.
2682 *
2683 * Note! Normally the "set_page_dirty(page)"
2684 * has no effect on the actual dirty bit - since
2685 * that will already usually be set. But we
2686 * need the side effects, and it can help us
2687 * avoid races.
2688 *
2689 * We basically use the page "master dirty bit"
2690 * as a serialization point for all the different
2691 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002692 */
2693 if (page_mkclean(page))
2694 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002695 /*
2696 * We carefully synchronise fault handlers against
2697 * installing a dirty pte and marking the page dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002698 * at this point. We do this by having them hold the
2699 * page lock while dirtying the page, and pages are
2700 * always locked coming in here, so we get the desired
2701 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002702 */
Tejun Heo682aa8e2015-05-28 14:50:53 -04002703 wb = unlocked_inode_to_wb_begin(inode, &locked);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002704 if (TestClearPageDirty(page)) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002705 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
Andrew Morton8c085402006-12-10 02:19:24 -08002706 dec_zone_page_state(page, NR_FILE_DIRTY);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002707 dec_wb_stat(wb, WB_RECLAIMABLE);
Greg Thelenc4843a72015-05-22 17:13:16 -04002708 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 }
Tejun Heo682aa8e2015-05-28 14:50:53 -04002710 unlocked_inode_to_wb_end(inode, locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002711 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002713 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002715EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716
2717int test_clear_page_writeback(struct page *page)
2718{
2719 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002720 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
Johannes Weiner62cccb82016-03-15 14:57:22 -07002722 lock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 if (mapping) {
Tejun Heo91018132015-05-22 17:13:39 -04002724 struct inode *inode = mapping->host;
2725 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 unsigned long flags;
2727
Nick Piggin19fd6232008-07-25 19:45:32 -07002728 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002730 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 radix_tree_tag_clear(&mapping->page_tree,
2732 page_index(page),
2733 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002734 if (bdi_cap_account_writeback(bdi)) {
Tejun Heo91018132015-05-22 17:13:39 -04002735 struct bdi_writeback *wb = inode_to_wb(inode);
2736
2737 __dec_wb_stat(wb, WB_WRITEBACK);
2738 __wb_writeout_inc(wb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002739 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002740 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002741 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 } else {
2743 ret = TestClearPageWriteback(page);
2744 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002745 if (ret) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002746 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
Andrew Mortond688abf2007-07-19 01:49:17 -07002747 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002748 inc_zone_page_state(page, NR_WRITTEN);
2749 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07002750 unlock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 return ret;
2752}
2753
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002754int __test_set_page_writeback(struct page *page, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755{
2756 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002757 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758
Johannes Weiner62cccb82016-03-15 14:57:22 -07002759 lock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 if (mapping) {
Tejun Heo91018132015-05-22 17:13:39 -04002761 struct inode *inode = mapping->host;
2762 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 unsigned long flags;
2764
Nick Piggin19fd6232008-07-25 19:45:32 -07002765 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002767 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 radix_tree_tag_set(&mapping->page_tree,
2769 page_index(page),
2770 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002771 if (bdi_cap_account_writeback(bdi))
Tejun Heo91018132015-05-22 17:13:39 -04002772 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002773 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 if (!PageDirty(page))
2775 radix_tree_tag_clear(&mapping->page_tree,
2776 page_index(page),
2777 PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002778 if (!keep_write)
2779 radix_tree_tag_clear(&mapping->page_tree,
2780 page_index(page),
2781 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002782 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 } else {
2784 ret = TestSetPageWriteback(page);
2785 }
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002786 if (!ret) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002787 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002788 inc_zone_page_state(page, NR_WRITEBACK);
2789 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07002790 unlock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 return ret;
2792
2793}
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002794EXPORT_SYMBOL(__test_set_page_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795
2796/*
Nick Piggin00128182007-10-16 01:24:40 -07002797 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 * passed tag.
2799 */
2800int mapping_tagged(struct address_space *mapping, int tag)
2801{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002802 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803}
2804EXPORT_SYMBOL(mapping_tagged);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002805
2806/**
2807 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2808 * @page: The page to wait on.
2809 *
2810 * This function determines if the given page is related to a backing device
2811 * that requires page contents to be held stable during writeback. If so, then
2812 * it will wait for any pending writeback to complete.
2813 */
2814void wait_for_stable_page(struct page *page)
2815{
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002816 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2817 wait_on_page_writeback(page);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002818}
2819EXPORT_SYMBOL_GPL(wait_for_stable_page);