blob: 5cccc127ef81f1d64ca46f9ce9ad50f519d4ea9f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
Francois Camie1f8e872008-10-15 22:01:59 -070010 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Initial version
12 */
13
14#include <linux/kernel.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040015#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
Al Viroff01bb42011-09-16 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
Tejun Heoe9f07df2015-05-22 18:23:28 -0400148#define DTC_INIT_COMMON(__wb) .wb = (__wb), \
Tejun Heoe9770b32015-05-22 18:23:27 -0400149 .wb_completions = &(__wb)->completions
Jan Karaeb608e32012-05-24 18:59:11 +0200150
151/*
152 * Length of period for aging writeout fractions of bdis. This is an
153 * arbitrarily chosen number. The longer the period, the slower fractions will
154 * reflect changes in current writeout rate.
155 */
156#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700157
Tejun Heo693108a2015-05-22 17:13:49 -0400158#ifdef CONFIG_CGROUP_WRITEBACK
159
Tejun Heoe9f07df2015-05-22 18:23:28 -0400160#define GDTC_INIT(__wb) .dom = &global_wb_domain, \
161 DTC_INIT_COMMON(__wb)
Tejun Heo9fc3a432015-05-22 18:23:30 -0400162#define GDTC_INIT_NO_WB .dom = &global_wb_domain
Tejun Heoc2aa7232015-05-22 18:23:35 -0400163#define MDTC_INIT(__wb, __gdtc) .dom = mem_cgroup_wb_domain(__wb), \
164 .gdtc = __gdtc, \
165 DTC_INIT_COMMON(__wb)
166
167static bool mdtc_valid(struct dirty_throttle_control *dtc)
168{
169 return dtc->dom;
170}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400171
172static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
173{
174 return dtc->dom;
175}
176
Tejun Heo9fc3a432015-05-22 18:23:30 -0400177static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
178{
179 return mdtc->gdtc;
180}
181
Tejun Heo841710a2015-05-22 18:23:33 -0400182static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
183{
184 return &wb->memcg_completions;
185}
186
Tejun Heo693108a2015-05-22 17:13:49 -0400187static void wb_min_max_ratio(struct bdi_writeback *wb,
188 unsigned long *minp, unsigned long *maxp)
189{
190 unsigned long this_bw = wb->avg_write_bandwidth;
191 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
192 unsigned long long min = wb->bdi->min_ratio;
193 unsigned long long max = wb->bdi->max_ratio;
194
195 /*
196 * @wb may already be clean by the time control reaches here and
197 * the total may not include its bw.
198 */
199 if (this_bw < tot_bw) {
200 if (min) {
201 min *= this_bw;
202 do_div(min, tot_bw);
203 }
204 if (max < 100) {
205 max *= this_bw;
206 do_div(max, tot_bw);
207 }
208 }
209
210 *minp = min;
211 *maxp = max;
212}
213
214#else /* CONFIG_CGROUP_WRITEBACK */
215
Tejun Heoe9f07df2015-05-22 18:23:28 -0400216#define GDTC_INIT(__wb) DTC_INIT_COMMON(__wb)
Tejun Heo9fc3a432015-05-22 18:23:30 -0400217#define GDTC_INIT_NO_WB
Tejun Heoc2aa7232015-05-22 18:23:35 -0400218#define MDTC_INIT(__wb, __gdtc)
219
220static bool mdtc_valid(struct dirty_throttle_control *dtc)
221{
222 return false;
223}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400224
225static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
226{
227 return &global_wb_domain;
228}
229
Tejun Heo9fc3a432015-05-22 18:23:30 -0400230static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
231{
232 return NULL;
233}
234
Tejun Heo841710a2015-05-22 18:23:33 -0400235static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
236{
237 return NULL;
238}
239
Tejun Heo693108a2015-05-22 17:13:49 -0400240static void wb_min_max_ratio(struct bdi_writeback *wb,
241 unsigned long *minp, unsigned long *maxp)
242{
243 *minp = wb->bdi->min_ratio;
244 *maxp = wb->bdi->max_ratio;
245}
246
247#endif /* CONFIG_CGROUP_WRITEBACK */
248
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700249/*
Johannes Weinera756cf52012-01-10 15:07:49 -0800250 * In a memory zone, there is a certain amount of pages we consider
251 * available for the page cache, which is essentially the number of
252 * free and reclaimable pages, minus some zone reserves to protect
253 * lowmem and the ability to uphold the zone's watermarks without
254 * requiring writeback.
255 *
256 * This number of dirtyable pages is the base value of which the
257 * user-configurable dirty ratio is the effictive number of pages that
258 * are allowed to be actually dirtied. Per individual zone, or
259 * globally by using the sum of dirtyable pages over all zones.
260 *
261 * Because the user is allowed to specify the dirty limit globally as
262 * absolute number of bytes, calculating the per-zone dirty limit can
263 * require translating the configured limit into a percentage of
264 * global dirtyable memory first.
265 */
266
Johannes Weinera8045522014-01-29 14:05:39 -0800267/**
268 * zone_dirtyable_memory - number of dirtyable pages in a zone
269 * @zone: the zone
270 *
271 * Returns the zone's number of pages potentially available for dirty
272 * page cache. This is the base value for the per-zone dirty limits.
273 */
274static unsigned long zone_dirtyable_memory(struct zone *zone)
275{
276 unsigned long nr_pages;
277
278 nr_pages = zone_page_state(zone, NR_FREE_PAGES);
279 nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
280
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800281 nr_pages += zone_page_state(zone, NR_INACTIVE_FILE);
282 nr_pages += zone_page_state(zone, NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800283
284 return nr_pages;
285}
286
Johannes Weiner1edf2232012-01-10 15:06:57 -0800287static unsigned long highmem_dirtyable_memory(unsigned long total)
288{
289#ifdef CONFIG_HIGHMEM
290 int node;
291 unsigned long x = 0;
292
293 for_each_node_state(node, N_HIGH_MEMORY) {
Johannes Weinera8045522014-01-29 14:05:39 -0800294 struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
Johannes Weiner1edf2232012-01-10 15:06:57 -0800295
Johannes Weinera8045522014-01-29 14:05:39 -0800296 x += zone_dirtyable_memory(z);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800297 }
298 /*
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800299 * Unreclaimable memory (kernel memory or anonymous memory
300 * without swap) can bring down the dirtyable pages below
301 * the zone's dirty balance reserve and the above calculation
302 * will underflow. However we still want to add in nodes
303 * which are below threshold (negative values) to get a more
304 * accurate calculation but make sure that the total never
305 * underflows.
306 */
307 if ((long)x < 0)
308 x = 0;
309
310 /*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800311 * Make sure that the number of highmem pages is never larger
312 * than the number of the total dirtyable memory. This can only
313 * occur in very strange VM situations but we want to make sure
314 * that this does not occur.
315 */
316 return min(x, total);
317#else
318 return 0;
319#endif
320}
321
322/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800323 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800324 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800325 * Returns the global number of pages potentially available for dirty
326 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800327 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700328static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800329{
330 unsigned long x;
331
Johannes Weinera8045522014-01-29 14:05:39 -0800332 x = global_page_state(NR_FREE_PAGES);
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800333 x -= min(x, dirty_balance_reserve);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800334
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800335 x += global_page_state(NR_INACTIVE_FILE);
336 x += global_page_state(NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800337
Johannes Weiner1edf2232012-01-10 15:06:57 -0800338 if (!vm_highmem_is_dirtyable)
339 x -= highmem_dirtyable_memory(x);
340
341 return x + 1; /* Ensure that we never return 0 */
342}
343
Tejun Heo9fc3a432015-05-22 18:23:30 -0400344/**
345 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
346 * @dtc: dirty_throttle_control of interest
Johannes Weinerccafa282012-01-10 15:07:44 -0800347 *
Tejun Heo9fc3a432015-05-22 18:23:30 -0400348 * Calculate @dtc->thresh and ->bg_thresh considering
349 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
350 * must ensure that @dtc->avail is set before calling this function. The
351 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Johannes Weinerccafa282012-01-10 15:07:44 -0800352 * real-time tasks.
353 */
Tejun Heo9fc3a432015-05-22 18:23:30 -0400354static void domain_dirty_limits(struct dirty_throttle_control *dtc)
355{
356 const unsigned long available_memory = dtc->avail;
357 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
358 unsigned long bytes = vm_dirty_bytes;
359 unsigned long bg_bytes = dirty_background_bytes;
360 unsigned long ratio = vm_dirty_ratio;
361 unsigned long bg_ratio = dirty_background_ratio;
362 unsigned long thresh;
363 unsigned long bg_thresh;
364 struct task_struct *tsk;
365
366 /* gdtc is !NULL iff @dtc is for memcg domain */
367 if (gdtc) {
368 unsigned long global_avail = gdtc->avail;
369
370 /*
371 * The byte settings can't be applied directly to memcg
372 * domains. Convert them to ratios by scaling against
373 * globally available memory.
374 */
375 if (bytes)
376 ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 /
377 global_avail, 100UL);
378 if (bg_bytes)
379 bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 /
380 global_avail, 100UL);
381 bytes = bg_bytes = 0;
382 }
383
384 if (bytes)
385 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
386 else
387 thresh = (ratio * available_memory) / 100;
388
389 if (bg_bytes)
390 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
391 else
392 bg_thresh = (bg_ratio * available_memory) / 100;
393
394 if (bg_thresh >= thresh)
395 bg_thresh = thresh / 2;
396 tsk = current;
397 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
398 bg_thresh += bg_thresh / 4;
399 thresh += thresh / 4;
400 }
401 dtc->thresh = thresh;
402 dtc->bg_thresh = bg_thresh;
403
404 /* we should eventually report the domain in the TP */
405 if (!gdtc)
406 trace_global_dirty_state(bg_thresh, thresh);
407}
408
409/**
410 * global_dirty_limits - background-writeback and dirty-throttling thresholds
411 * @pbackground: out parameter for bg_thresh
412 * @pdirty: out parameter for thresh
413 *
414 * Calculate bg_thresh and thresh for global_wb_domain. See
415 * domain_dirty_limits() for details.
416 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800417void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
418{
Tejun Heo9fc3a432015-05-22 18:23:30 -0400419 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
Johannes Weinerccafa282012-01-10 15:07:44 -0800420
Tejun Heo9fc3a432015-05-22 18:23:30 -0400421 gdtc.avail = global_dirtyable_memory();
422 domain_dirty_limits(&gdtc);
Johannes Weinerccafa282012-01-10 15:07:44 -0800423
Tejun Heo9fc3a432015-05-22 18:23:30 -0400424 *pbackground = gdtc.bg_thresh;
425 *pdirty = gdtc.thresh;
Johannes Weinerccafa282012-01-10 15:07:44 -0800426}
427
Johannes Weinera756cf52012-01-10 15:07:49 -0800428/**
Johannes Weinera756cf52012-01-10 15:07:49 -0800429 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
430 * @zone: the zone
431 *
432 * Returns the maximum number of dirty pages allowed in a zone, based
433 * on the zone's dirtyable memory.
434 */
435static unsigned long zone_dirty_limit(struct zone *zone)
436{
437 unsigned long zone_memory = zone_dirtyable_memory(zone);
438 struct task_struct *tsk = current;
439 unsigned long dirty;
440
441 if (vm_dirty_bytes)
442 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
443 zone_memory / global_dirtyable_memory();
444 else
445 dirty = vm_dirty_ratio * zone_memory / 100;
446
447 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
448 dirty += dirty / 4;
449
450 return dirty;
451}
452
453/**
454 * zone_dirty_ok - tells whether a zone is within its dirty limits
455 * @zone: the zone to check
456 *
457 * Returns %true when the dirty pages in @zone are within the zone's
458 * dirty limit, %false if the limit is exceeded.
459 */
460bool zone_dirty_ok(struct zone *zone)
461{
462 unsigned long limit = zone_dirty_limit(zone);
463
464 return zone_page_state(zone, NR_FILE_DIRTY) +
465 zone_page_state(zone, NR_UNSTABLE_NFS) +
466 zone_page_state(zone, NR_WRITEBACK) <= limit;
467}
468
David Rientjes2da02992009-01-06 14:39:31 -0800469int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700470 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800471 loff_t *ppos)
472{
473 int ret;
474
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700475 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800476 if (ret == 0 && write)
477 dirty_background_bytes = 0;
478 return ret;
479}
480
481int dirty_background_bytes_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_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800488 if (ret == 0 && write)
489 dirty_background_ratio = 0;
490 return ret;
491}
492
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700493int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700494 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700495 loff_t *ppos)
496{
497 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800498 int ret;
499
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700500 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700501 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200502 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800503 vm_dirty_bytes = 0;
504 }
505 return ret;
506}
507
David Rientjes2da02992009-01-06 14:39:31 -0800508int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700509 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800510 loff_t *ppos)
511{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800512 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800513 int ret;
514
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700515 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800516 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200517 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800518 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700519 }
520 return ret;
521}
522
Jan Karaeb608e32012-05-24 18:59:11 +0200523static unsigned long wp_next_time(unsigned long cur_time)
524{
525 cur_time += VM_COMPLETIONS_PERIOD_LEN;
526 /* 0 has a special meaning... */
527 if (!cur_time)
528 return 1;
529 return cur_time;
530}
531
Tejun Heoc7981432015-05-22 18:23:29 -0400532static void wb_domain_writeout_inc(struct wb_domain *dom,
533 struct fprop_local_percpu *completions,
534 unsigned int max_prop_frac)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700535{
Tejun Heoc7981432015-05-22 18:23:29 -0400536 __fprop_inc_percpu_max(&dom->completions, completions,
537 max_prop_frac);
Jan Karaeb608e32012-05-24 18:59:11 +0200538 /* First event after period switching was turned off? */
Tejun Heo380c27c2015-05-22 18:23:21 -0400539 if (!unlikely(dom->period_time)) {
Jan Karaeb608e32012-05-24 18:59:11 +0200540 /*
541 * We can race with other __bdi_writeout_inc calls here but
542 * it does not cause any harm since the resulting time when
543 * timer will fire and what is in writeout_period_time will be
544 * roughly the same.
545 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400546 dom->period_time = wp_next_time(jiffies);
547 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200548 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700549}
550
Tejun Heoc7981432015-05-22 18:23:29 -0400551/*
552 * Increment @wb's writeout completion count and the global writeout
553 * completion count. Called from test_clear_page_writeback().
554 */
555static inline void __wb_writeout_inc(struct bdi_writeback *wb)
556{
Tejun Heo841710a2015-05-22 18:23:33 -0400557 struct wb_domain *cgdom;
558
Tejun Heoc7981432015-05-22 18:23:29 -0400559 __inc_wb_stat(wb, WB_WRITTEN);
560 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
561 wb->bdi->max_prop_frac);
Tejun Heo841710a2015-05-22 18:23:33 -0400562
563 cgdom = mem_cgroup_wb_domain(wb);
564 if (cgdom)
565 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
566 wb->bdi->max_prop_frac);
Tejun Heoc7981432015-05-22 18:23:29 -0400567}
568
Tejun Heo93f78d82015-05-22 17:13:27 -0400569void wb_writeout_inc(struct bdi_writeback *wb)
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700570{
571 unsigned long flags;
572
573 local_irq_save(flags);
Tejun Heo93f78d82015-05-22 17:13:27 -0400574 __wb_writeout_inc(wb);
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700575 local_irq_restore(flags);
576}
Tejun Heo93f78d82015-05-22 17:13:27 -0400577EXPORT_SYMBOL_GPL(wb_writeout_inc);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700578
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700579/*
Jan Karaeb608e32012-05-24 18:59:11 +0200580 * On idle system, we can be called long after we scheduled because we use
581 * deferred timers so count with missed periods.
582 */
583static void writeout_period(unsigned long t)
584{
Tejun Heo380c27c2015-05-22 18:23:21 -0400585 struct wb_domain *dom = (void *)t;
586 int miss_periods = (jiffies - dom->period_time) /
Jan Karaeb608e32012-05-24 18:59:11 +0200587 VM_COMPLETIONS_PERIOD_LEN;
588
Tejun Heo380c27c2015-05-22 18:23:21 -0400589 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
590 dom->period_time = wp_next_time(dom->period_time +
Jan Karaeb608e32012-05-24 18:59:11 +0200591 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
Tejun Heo380c27c2015-05-22 18:23:21 -0400592 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200593 } else {
594 /*
595 * Aging has zeroed all fractions. Stop wasting CPU on period
596 * updates.
597 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400598 dom->period_time = 0;
Jan Karaeb608e32012-05-24 18:59:11 +0200599 }
600}
601
Tejun Heo380c27c2015-05-22 18:23:21 -0400602int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
603{
604 memset(dom, 0, sizeof(*dom));
Tejun Heodcc25ae2015-05-22 18:23:22 -0400605
606 spin_lock_init(&dom->lock);
607
Tejun Heo380c27c2015-05-22 18:23:21 -0400608 init_timer_deferrable(&dom->period_timer);
609 dom->period_timer.function = writeout_period;
610 dom->period_timer.data = (unsigned long)dom;
Tejun Heodcc25ae2015-05-22 18:23:22 -0400611
612 dom->dirty_limit_tstamp = jiffies;
613
Tejun Heo380c27c2015-05-22 18:23:21 -0400614 return fprop_global_init(&dom->completions, gfp);
615}
616
Tejun Heo841710a2015-05-22 18:23:33 -0400617#ifdef CONFIG_CGROUP_WRITEBACK
618void wb_domain_exit(struct wb_domain *dom)
619{
620 del_timer_sync(&dom->period_timer);
621 fprop_global_destroy(&dom->completions);
622}
623#endif
624
Jan Karaeb608e32012-05-24 18:59:11 +0200625/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700626 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
627 * registered backing devices, which, for obvious reasons, can not
628 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700629 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700630static unsigned int bdi_min_ratio;
631
632int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
633{
634 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700635
Jens Axboecfc4ba52009-09-14 13:12:40 +0200636 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700637 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700638 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700639 } else {
640 min_ratio -= bdi->min_ratio;
641 if (bdi_min_ratio + min_ratio < 100) {
642 bdi_min_ratio += min_ratio;
643 bdi->min_ratio += min_ratio;
644 } else {
645 ret = -EINVAL;
646 }
647 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200648 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700649
650 return ret;
651}
652
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700653int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
654{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700655 int ret = 0;
656
657 if (max_ratio > 100)
658 return -EINVAL;
659
Jens Axboecfc4ba52009-09-14 13:12:40 +0200660 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700661 if (bdi->min_ratio > max_ratio) {
662 ret = -EINVAL;
663 } else {
664 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200665 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700666 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200667 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700668
669 return ret;
670}
671EXPORT_SYMBOL(bdi_set_max_ratio);
672
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600673static unsigned long dirty_freerun_ceiling(unsigned long thresh,
674 unsigned long bg_thresh)
675{
676 return (thresh + bg_thresh) / 2;
677}
678
Tejun Heoc7981432015-05-22 18:23:29 -0400679static unsigned long hard_dirty_limit(struct wb_domain *dom,
680 unsigned long thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -0600681{
Tejun Heodcc25ae2015-05-22 18:23:22 -0400682 return max(thresh, dom->dirty_limit);
Wu Fengguangffd1f602011-06-19 22:18:42 -0600683}
684
Tejun Heoc2aa7232015-05-22 18:23:35 -0400685/* memory available to a memcg domain is capped by system-wide clean memory */
686static void mdtc_cap_avail(struct dirty_throttle_control *mdtc)
687{
688 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
689 unsigned long clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
690
691 mdtc->avail = min(mdtc->avail, clean);
Christoph Lameter1b424462007-05-06 14:48:59 -0700692}
693
Wu Fengguang6f718652011-03-02 17:14:34 -0600694/**
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400695 * __wb_calc_thresh - @wb's share of dirty throttling threshold
696 * @dtc: dirty_throttle_context of interest
Wu Fengguang1babe182010-08-11 14:17:40 -0700697 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400698 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
Wu Fengguang6f718652011-03-02 17:14:34 -0600699 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600700 *
701 * Note that balance_dirty_pages() will only seriously take it as a hard limit
702 * when sleeping max_pause per page is not enough to keep the dirty pages under
703 * control. For example, when the device is completely stalled due to some error
704 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
705 * In the other normal situations, it acts more gently by throttling the tasks
Tejun Heoa88a3412015-05-22 17:13:28 -0400706 * more (rather than completely block them) when the wb dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600707 *
708 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700709 * - starving fast devices
710 * - piling up dirty pages (that will take long time to sync) on slow devices
711 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400712 * The wb's share of dirty limit will be adapting to its throughput and
Wu Fengguang1babe182010-08-11 14:17:40 -0700713 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
714 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400715static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
Wu Fengguang16c40422010-08-11 14:17:39 -0700716{
Tejun Heoe9f07df2015-05-22 18:23:28 -0400717 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400718 unsigned long thresh = dtc->thresh;
Tejun Heo0d960a32015-05-22 18:23:19 -0400719 u64 wb_thresh;
Wu Fengguang16c40422010-08-11 14:17:39 -0700720 long numerator, denominator;
Tejun Heo693108a2015-05-22 17:13:49 -0400721 unsigned long wb_min_ratio, wb_max_ratio;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700722
Wu Fengguang16c40422010-08-11 14:17:39 -0700723 /*
Tejun Heo0d960a32015-05-22 18:23:19 -0400724 * Calculate this BDI's share of the thresh ratio.
Wu Fengguang16c40422010-08-11 14:17:39 -0700725 */
Tejun Heoe9770b32015-05-22 18:23:27 -0400726 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
Tejun Heo380c27c2015-05-22 18:23:21 -0400727 &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700728
Tejun Heo0d960a32015-05-22 18:23:19 -0400729 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
730 wb_thresh *= numerator;
731 do_div(wb_thresh, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700732
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400733 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
Wu Fengguang16c40422010-08-11 14:17:39 -0700734
Tejun Heo0d960a32015-05-22 18:23:19 -0400735 wb_thresh += (thresh * wb_min_ratio) / 100;
736 if (wb_thresh > (thresh * wb_max_ratio) / 100)
737 wb_thresh = thresh * wb_max_ratio / 100;
Wu Fengguang16c40422010-08-11 14:17:39 -0700738
Tejun Heo0d960a32015-05-22 18:23:19 -0400739 return wb_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740}
741
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400742unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
743{
744 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
745 .thresh = thresh };
746 return __wb_calc_thresh(&gdtc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747}
748
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600749/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700750 * setpoint - dirty 3
751 * f(dirty) := 1.0 + (----------------)
752 * limit - setpoint
753 *
754 * it's a 3rd order polynomial that subjects to
755 *
756 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
757 * (2) f(setpoint) = 1.0 => the balance point
758 * (3) f(limit) = 0 => the hard limit
759 * (4) df/dx <= 0 => negative feedback control
760 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
761 * => fast response on large errors; small oscillation near setpoint
762 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700763static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700764 unsigned long dirty,
765 unsigned long limit)
766{
767 long long pos_ratio;
768 long x;
769
Rik van Rield5c9fde2014-05-06 12:50:01 -0700770 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Tejun Heo464d1382015-04-21 16:49:13 -0400771 (limit - setpoint) | 1);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700772 pos_ratio = x;
773 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
774 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
775 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
776
777 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
778}
779
780/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600781 * Dirty position control.
782 *
783 * (o) global/bdi setpoints
784 *
Tejun Heode1fff32015-05-22 17:13:29 -0400785 * We want the dirty pages be balanced around the global/wb setpoints.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600786 * When the number of dirty pages is higher/lower than the setpoint, the
787 * dirty position control ratio (and hence task dirty ratelimit) will be
788 * decreased/increased to bring the dirty pages back to the setpoint.
789 *
790 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
791 *
792 * if (dirty < setpoint) scale up pos_ratio
793 * if (dirty > setpoint) scale down pos_ratio
794 *
Tejun Heode1fff32015-05-22 17:13:29 -0400795 * if (wb_dirty < wb_setpoint) scale up pos_ratio
796 * if (wb_dirty > wb_setpoint) scale down pos_ratio
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600797 *
798 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
799 *
800 * (o) global control line
801 *
802 * ^ pos_ratio
803 * |
804 * | |<===== global dirty control scope ======>|
805 * 2.0 .............*
806 * | .*
807 * | . *
808 * | . *
809 * | . *
810 * | . *
811 * | . *
812 * 1.0 ................................*
813 * | . . *
814 * | . . *
815 * | . . *
816 * | . . *
817 * | . . *
818 * 0 +------------.------------------.----------------------*------------->
819 * freerun^ setpoint^ limit^ dirty pages
820 *
Tejun Heode1fff32015-05-22 17:13:29 -0400821 * (o) wb control line
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600822 *
823 * ^ pos_ratio
824 * |
825 * | *
826 * | *
827 * | *
828 * | *
829 * | * |<=========== span ============>|
830 * 1.0 .......................*
831 * | . *
832 * | . *
833 * | . *
834 * | . *
835 * | . *
836 * | . *
837 * | . *
838 * | . *
839 * | . *
840 * | . *
841 * | . *
842 * 1/4 ...............................................* * * * * * * * * * * *
843 * | . .
844 * | . .
845 * | . .
846 * 0 +----------------------.-------------------------------.------------->
Tejun Heode1fff32015-05-22 17:13:29 -0400847 * wb_setpoint^ x_intercept^
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600848 *
Tejun Heode1fff32015-05-22 17:13:29 -0400849 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600850 * be smoothly throttled down to normal if it starts high in situations like
851 * - start writing to a slow SD card and a fast disk at the same time. The SD
Tejun Heode1fff32015-05-22 17:13:29 -0400852 * card's wb_dirty may rush to many times higher than wb_setpoint.
853 * - the wb dirty thresh drops quickly due to change of JBOD workload
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600854 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400855static void wb_position_ratio(struct dirty_throttle_control *dtc)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600856{
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400857 struct bdi_writeback *wb = dtc->wb;
Tejun Heoa88a3412015-05-22 17:13:28 -0400858 unsigned long write_bw = wb->avg_write_bandwidth;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400859 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -0400860 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400861 unsigned long wb_thresh = dtc->wb_thresh;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600862 unsigned long x_intercept;
863 unsigned long setpoint; /* dirty pages' target balance point */
Tejun Heode1fff32015-05-22 17:13:29 -0400864 unsigned long wb_setpoint;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600865 unsigned long span;
866 long long pos_ratio; /* for scaling up/down the rate limit */
867 long x;
868
Tejun Heodaddfa32015-05-22 18:23:26 -0400869 dtc->pos_ratio = 0;
870
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400871 if (unlikely(dtc->dirty >= limit))
Tejun Heodaddfa32015-05-22 18:23:26 -0400872 return;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600873
874 /*
875 * global setpoint
876 *
Maxim Patlasov5a537482013-09-11 14:22:46 -0700877 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600878 */
879 setpoint = (freerun + limit) / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400880 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700881
882 /*
883 * The strictlimit feature is a tool preventing mistrusted filesystems
884 * from growing a large number of dirty pages before throttling. For
Tejun Heode1fff32015-05-22 17:13:29 -0400885 * such filesystems balance_dirty_pages always checks wb counters
886 * against wb limits. Even if global "nr_dirty" is under "freerun".
Maxim Patlasov5a537482013-09-11 14:22:46 -0700887 * This is especially important for fuse which sets bdi->max_ratio to
888 * 1% by default. Without strictlimit feature, fuse writeback may
889 * consume arbitrary amount of RAM because it is accounted in
890 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
891 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400892 * Here, in wb_position_ratio(), we calculate pos_ratio based on
Tejun Heode1fff32015-05-22 17:13:29 -0400893 * two values: wb_dirty and wb_thresh. Let's consider an example:
Maxim Patlasov5a537482013-09-11 14:22:46 -0700894 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
895 * limits are set by default to 10% and 20% (background and throttle).
Tejun Heode1fff32015-05-22 17:13:29 -0400896 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
Tejun Heo0d960a32015-05-22 18:23:19 -0400897 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
Tejun Heode1fff32015-05-22 17:13:29 -0400898 * about ~6K pages (as the average of background and throttle wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700899 * limits). The 3rd order polynomial will provide positive feedback if
Tejun Heode1fff32015-05-22 17:13:29 -0400900 * wb_dirty is under wb_setpoint and vice versa.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700901 *
902 * Note, that we cannot use global counters in these calculations
Tejun Heode1fff32015-05-22 17:13:29 -0400903 * because we want to throttle process writing to a strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700904 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
905 * in the example above).
906 */
Tejun Heoa88a3412015-05-22 17:13:28 -0400907 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heode1fff32015-05-22 17:13:29 -0400908 long long wb_pos_ratio;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700909
Tejun Heodaddfa32015-05-22 18:23:26 -0400910 if (dtc->wb_dirty < 8) {
911 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
912 2 << RATELIMIT_CALC_SHIFT);
913 return;
914 }
Maxim Patlasov5a537482013-09-11 14:22:46 -0700915
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400916 if (dtc->wb_dirty >= wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400917 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700918
Tejun Heo970fb012015-05-22 18:23:24 -0400919 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
920 dtc->wb_bg_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700921
Tejun Heode1fff32015-05-22 17:13:29 -0400922 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400923 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700924
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400925 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
Tejun Heode1fff32015-05-22 17:13:29 -0400926 wb_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700927
928 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400929 * Typically, for strictlimit case, wb_setpoint << setpoint
930 * and pos_ratio >> wb_pos_ratio. In the other words global
Maxim Patlasov5a537482013-09-11 14:22:46 -0700931 * state ("dirty") is not limiting factor and we have to
Tejun Heode1fff32015-05-22 17:13:29 -0400932 * make decision based on wb counters. But there is an
Maxim Patlasov5a537482013-09-11 14:22:46 -0700933 * important case when global pos_ratio should get precedence:
934 * global limits are exceeded (e.g. due to activities on other
Tejun Heode1fff32015-05-22 17:13:29 -0400935 * wb's) while given strictlimit wb is below limit.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700936 *
Tejun Heode1fff32015-05-22 17:13:29 -0400937 * "pos_ratio * wb_pos_ratio" would work for the case above,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700938 * but it would look too non-natural for the case of all
Tejun Heode1fff32015-05-22 17:13:29 -0400939 * activity in the system coming from a single strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700940 * with bdi->max_ratio == 100%.
941 *
942 * Note that min() below somewhat changes the dynamics of the
943 * control system. Normally, pos_ratio value can be well over 3
Tejun Heode1fff32015-05-22 17:13:29 -0400944 * (when globally we are at freerun and wb is well below wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700945 * setpoint). Now the maximum pos_ratio in the same situation
946 * is 2. We might want to tweak this if we observe the control
947 * system is too slow to adapt.
948 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400949 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
950 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700951 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600952
953 /*
954 * We have computed basic pos_ratio above based on global situation. If
Tejun Heode1fff32015-05-22 17:13:29 -0400955 * the wb is over/under its share of dirty pages, we want to scale
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600956 * pos_ratio further down/up. That is done by the following mechanism.
957 */
958
959 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400960 * wb setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600961 *
Tejun Heode1fff32015-05-22 17:13:29 -0400962 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600963 *
Tejun Heode1fff32015-05-22 17:13:29 -0400964 * x_intercept - wb_dirty
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600965 * := --------------------------
Tejun Heode1fff32015-05-22 17:13:29 -0400966 * x_intercept - wb_setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600967 *
Tejun Heode1fff32015-05-22 17:13:29 -0400968 * The main wb control line is a linear function that subjects to
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600969 *
Tejun Heode1fff32015-05-22 17:13:29 -0400970 * (1) f(wb_setpoint) = 1.0
971 * (2) k = - 1 / (8 * write_bw) (in single wb case)
972 * or equally: x_intercept = wb_setpoint + 8 * write_bw
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600973 *
Tejun Heode1fff32015-05-22 17:13:29 -0400974 * For single wb case, the dirty pages are observed to fluctuate
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600975 * regularly within range
Tejun Heode1fff32015-05-22 17:13:29 -0400976 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600977 * for various filesystems, where (2) can yield in a reasonable 12.5%
978 * fluctuation range for pos_ratio.
979 *
Tejun Heode1fff32015-05-22 17:13:29 -0400980 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600981 * own size, so move the slope over accordingly and choose a slope that
Tejun Heode1fff32015-05-22 17:13:29 -0400982 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600983 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400984 if (unlikely(wb_thresh > dtc->thresh))
985 wb_thresh = dtc->thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600986 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400987 * It's very possible that wb_thresh is close to 0 not because the
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600988 * device is slow, but that it has remained inactive for long time.
989 * Honour such devices a reasonable good (hopefully IO efficient)
990 * threshold, so that the occasional writes won't be blocked and active
991 * writes can rampup the threshold quickly.
992 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400993 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600994 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400995 * scale global setpoint to wb's:
996 * wb_setpoint = setpoint * wb_thresh / thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600997 */
Linus Torvaldse4bc13a2015-06-25 16:00:17 -0700998 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
Tejun Heode1fff32015-05-22 17:13:29 -0400999 wb_setpoint = setpoint * (u64)x >> 16;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001000 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001001 * Use span=(8*write_bw) in single wb case as indicated by
1002 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001003 *
Tejun Heode1fff32015-05-22 17:13:29 -04001004 * wb_thresh thresh - wb_thresh
1005 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1006 * thresh thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001007 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001008 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
Tejun Heode1fff32015-05-22 17:13:29 -04001009 x_intercept = wb_setpoint + span;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001010
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001011 if (dtc->wb_dirty < x_intercept - span / 4) {
1012 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001013 (x_intercept - wb_setpoint) | 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001014 } else
1015 pos_ratio /= 4;
1016
Wu Fengguang8927f662011-08-04 22:16:46 -06001017 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001018 * wb reserve area, safeguard against dirty pool underrun and disk idle
Wu Fengguang8927f662011-08-04 22:16:46 -06001019 * It may push the desired control point of global dirty pages higher
1020 * than setpoint.
1021 */
Tejun Heode1fff32015-05-22 17:13:29 -04001022 x_intercept = wb_thresh / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001023 if (dtc->wb_dirty < x_intercept) {
1024 if (dtc->wb_dirty > x_intercept / 8)
1025 pos_ratio = div_u64(pos_ratio * x_intercept,
1026 dtc->wb_dirty);
Wu Fengguang50657fc2011-10-11 17:06:33 -06001027 else
Wu Fengguang8927f662011-08-04 22:16:46 -06001028 pos_ratio *= 8;
1029 }
1030
Tejun Heodaddfa32015-05-22 18:23:26 -04001031 dtc->pos_ratio = pos_ratio;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001032}
1033
Tejun Heoa88a3412015-05-22 17:13:28 -04001034static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1035 unsigned long elapsed,
1036 unsigned long written)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001037{
1038 const unsigned long period = roundup_pow_of_two(3 * HZ);
Tejun Heoa88a3412015-05-22 17:13:28 -04001039 unsigned long avg = wb->avg_write_bandwidth;
1040 unsigned long old = wb->write_bandwidth;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001041 u64 bw;
1042
1043 /*
1044 * bw = written * HZ / elapsed
1045 *
1046 * bw * elapsed + write_bandwidth * (period - elapsed)
1047 * write_bandwidth = ---------------------------------------------------
1048 * period
Tejun Heoc72efb62015-03-23 00:18:48 -04001049 *
1050 * @written may have decreased due to account_page_redirty().
1051 * Avoid underflowing @bw calculation.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001052 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001053 bw = written - min(written, wb->written_stamp);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001054 bw *= HZ;
1055 if (unlikely(elapsed > period)) {
1056 do_div(bw, elapsed);
1057 avg = bw;
1058 goto out;
1059 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001060 bw += (u64)wb->write_bandwidth * (period - elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001061 bw >>= ilog2(period);
1062
1063 /*
1064 * one more level of smoothing, for filtering out sudden spikes
1065 */
1066 if (avg > old && old >= (unsigned long)bw)
1067 avg -= (avg - old) >> 3;
1068
1069 if (avg < old && old <= (unsigned long)bw)
1070 avg += (old - avg) >> 3;
1071
1072out:
Tejun Heo95a46c62015-05-22 17:13:47 -04001073 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1074 avg = max(avg, 1LU);
1075 if (wb_has_dirty_io(wb)) {
1076 long delta = avg - wb->avg_write_bandwidth;
1077 WARN_ON_ONCE(atomic_long_add_return(delta,
1078 &wb->bdi->tot_write_bandwidth) <= 0);
1079 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001080 wb->write_bandwidth = bw;
1081 wb->avg_write_bandwidth = avg;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001082}
1083
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001084static void update_dirty_limit(struct dirty_throttle_control *dtc)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001085{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001086 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001087 unsigned long thresh = dtc->thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04001088 unsigned long limit = dom->dirty_limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001089
1090 /*
1091 * Follow up in one step.
1092 */
1093 if (limit < thresh) {
1094 limit = thresh;
1095 goto update;
1096 }
1097
1098 /*
1099 * Follow down slowly. Use the higher one as the target, because thresh
1100 * may drop below dirty. This is exactly the reason to introduce
Tejun Heodcc25ae2015-05-22 18:23:22 -04001101 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
Wu Fengguangc42843f2011-03-02 15:54:09 -06001102 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001103 thresh = max(thresh, dtc->dirty);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001104 if (limit > thresh) {
1105 limit -= (limit - thresh) >> 5;
1106 goto update;
1107 }
1108 return;
1109update:
Tejun Heodcc25ae2015-05-22 18:23:22 -04001110 dom->dirty_limit = limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001111}
1112
Tejun Heoe9f07df2015-05-22 18:23:28 -04001113static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001114 unsigned long now)
1115{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001116 struct wb_domain *dom = dtc_dom(dtc);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001117
1118 /*
1119 * check locklessly first to optimize away locking for the most time
1120 */
Tejun Heodcc25ae2015-05-22 18:23:22 -04001121 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
Wu Fengguangc42843f2011-03-02 15:54:09 -06001122 return;
1123
Tejun Heodcc25ae2015-05-22 18:23:22 -04001124 spin_lock(&dom->lock);
1125 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001126 update_dirty_limit(dtc);
Tejun Heodcc25ae2015-05-22 18:23:22 -04001127 dom->dirty_limit_tstamp = now;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001128 }
Tejun Heodcc25ae2015-05-22 18:23:22 -04001129 spin_unlock(&dom->lock);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001130}
1131
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001132/*
Tejun Heode1fff32015-05-22 17:13:29 -04001133 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001134 *
Tejun Heode1fff32015-05-22 17:13:29 -04001135 * Normal wb tasks will be curbed at or below it in long term.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001136 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1137 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001138static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
Tejun Heoa88a3412015-05-22 17:13:28 -04001139 unsigned long dirtied,
1140 unsigned long elapsed)
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001141{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001142 struct bdi_writeback *wb = dtc->wb;
1143 unsigned long dirty = dtc->dirty;
1144 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001145 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Wu Fengguang73811312011-08-26 15:53:24 -06001146 unsigned long setpoint = (freerun + limit) / 2;
Tejun Heoa88a3412015-05-22 17:13:28 -04001147 unsigned long write_bw = wb->avg_write_bandwidth;
1148 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001149 unsigned long dirty_rate;
1150 unsigned long task_ratelimit;
1151 unsigned long balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001152 unsigned long step;
1153 unsigned long x;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001154
1155 /*
1156 * The dirty rate will match the writeout rate in long term, except
1157 * when dirty pages are truncated by userspace or re-dirtied by FS.
1158 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001159 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001160
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001161 /*
1162 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1163 */
1164 task_ratelimit = (u64)dirty_ratelimit *
Tejun Heodaddfa32015-05-22 18:23:26 -04001165 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001166 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1167
1168 /*
1169 * A linear estimation of the "balanced" throttle rate. The theory is,
Tejun Heode1fff32015-05-22 17:13:29 -04001170 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001171 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1172 * formula will yield the balanced rate limit (write_bw / N).
1173 *
1174 * Note that the expanded form is not a pure rate feedback:
1175 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1176 * but also takes pos_ratio into account:
1177 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1178 *
1179 * (1) is not realistic because pos_ratio also takes part in balancing
1180 * the dirty rate. Consider the state
1181 * pos_ratio = 0.5 (3)
1182 * rate = 2 * (write_bw / N) (4)
1183 * If (1) is used, it will stuck in that state! Because each dd will
1184 * be throttled at
1185 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1186 * yielding
1187 * dirty_rate = N * task_ratelimit = write_bw (6)
1188 * put (6) into (1) we get
1189 * rate_(i+1) = rate_(i) (7)
1190 *
1191 * So we end up using (2) to always keep
1192 * rate_(i+1) ~= (write_bw / N) (8)
1193 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1194 * pos_ratio is able to drive itself to 1.0, which is not only where
1195 * the dirty count meet the setpoint, but also where the slope of
1196 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1197 */
1198 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1199 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001200 /*
1201 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1202 */
1203 if (unlikely(balanced_dirty_ratelimit > write_bw))
1204 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001205
Wu Fengguang73811312011-08-26 15:53:24 -06001206 /*
1207 * We could safely do this and return immediately:
1208 *
Tejun Heode1fff32015-05-22 17:13:29 -04001209 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001210 *
1211 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001212 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001213 * limit the step size.
1214 *
1215 * The below code essentially only uses the relative value of
1216 *
1217 * task_ratelimit - dirty_ratelimit
1218 * = (pos_ratio - 1) * dirty_ratelimit
1219 *
1220 * which reflects the direction and size of dirty position error.
1221 */
1222
1223 /*
1224 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1225 * task_ratelimit is on the same side of dirty_ratelimit, too.
1226 * For example, when
1227 * - dirty_ratelimit > balanced_dirty_ratelimit
1228 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1229 * lowering dirty_ratelimit will help meet both the position and rate
1230 * control targets. Otherwise, don't update dirty_ratelimit if it will
1231 * only help meet the rate target. After all, what the users ultimately
1232 * feel and care are stable dirty rate and small position error.
1233 *
1234 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001235 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001236 * keeps jumping around randomly and can even leap far away at times
1237 * due to the small 200ms estimation period of dirty_rate (we want to
1238 * keep that period small to reduce time lags).
1239 */
1240 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001241
1242 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001243 * For strictlimit case, calculations above were based on wb counters
Tejun Heoa88a3412015-05-22 17:13:28 -04001244 * and limits (starting from pos_ratio = wb_position_ratio() and up to
Maxim Patlasov5a537482013-09-11 14:22:46 -07001245 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
Tejun Heode1fff32015-05-22 17:13:29 -04001246 * Hence, to calculate "step" properly, we have to use wb_dirty as
1247 * "dirty" and wb_setpoint as "setpoint".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001248 *
Tejun Heode1fff32015-05-22 17:13:29 -04001249 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1250 * it's possible that wb_thresh is close to zero due to inactivity
Tejun Heo970fb012015-05-22 18:23:24 -04001251 * of backing device.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001252 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001253 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001254 dirty = dtc->wb_dirty;
1255 if (dtc->wb_dirty < 8)
1256 setpoint = dtc->wb_dirty + 1;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001257 else
Tejun Heo970fb012015-05-22 18:23:24 -04001258 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001259 }
1260
Wu Fengguang73811312011-08-26 15:53:24 -06001261 if (dirty < setpoint) {
Tejun Heoa88a3412015-05-22 17:13:28 -04001262 x = min3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001263 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001264 if (dirty_ratelimit < x)
1265 step = x - dirty_ratelimit;
1266 } else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001267 x = max3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001268 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001269 if (dirty_ratelimit > x)
1270 step = dirty_ratelimit - x;
1271 }
1272
1273 /*
1274 * Don't pursue 100% rate matching. It's impossible since the balanced
1275 * rate itself is constantly fluctuating. So decrease the track speed
1276 * when it gets close to the target. Helps eliminate pointless tremors.
1277 */
1278 step >>= dirty_ratelimit / (2 * step + 1);
1279 /*
1280 * Limit the tracking speed to avoid overshooting.
1281 */
1282 step = (step + 7) / 8;
1283
1284 if (dirty_ratelimit < balanced_dirty_ratelimit)
1285 dirty_ratelimit += step;
1286 else
1287 dirty_ratelimit -= step;
1288
Tejun Heoa88a3412015-05-22 17:13:28 -04001289 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1290 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001291
Tejun Heoa88a3412015-05-22 17:13:28 -04001292 trace_bdi_dirty_ratelimit(wb->bdi, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001293}
1294
Tejun Heoc2aa7232015-05-22 18:23:35 -04001295static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1296 struct dirty_throttle_control *mdtc,
Tejun Heo8a731792015-05-22 18:23:20 -04001297 unsigned long start_time,
1298 bool update_ratelimit)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001299{
Tejun Heoc2aa7232015-05-22 18:23:35 -04001300 struct bdi_writeback *wb = gdtc->wb;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001301 unsigned long now = jiffies;
Tejun Heoa88a3412015-05-22 17:13:28 -04001302 unsigned long elapsed = now - wb->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001303 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001304 unsigned long written;
1305
Tejun Heo8a731792015-05-22 18:23:20 -04001306 lockdep_assert_held(&wb->list_lock);
1307
Wu Fengguange98be2d2010-08-29 11:22:30 -06001308 /*
1309 * rate-limit, only update once every 200ms.
1310 */
1311 if (elapsed < BANDWIDTH_INTERVAL)
1312 return;
1313
Tejun Heoa88a3412015-05-22 17:13:28 -04001314 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1315 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001316
1317 /*
1318 * Skip quiet periods when disk bandwidth is under-utilized.
1319 * (at least 1s idle time between two flusher runs)
1320 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001321 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
Wu Fengguange98be2d2010-08-29 11:22:30 -06001322 goto snapshot;
1323
Tejun Heo8a731792015-05-22 18:23:20 -04001324 if (update_ratelimit) {
Tejun Heoc2aa7232015-05-22 18:23:35 -04001325 domain_update_bandwidth(gdtc, now);
1326 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1327
1328 /*
1329 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1330 * compiler has no way to figure that out. Help it.
1331 */
1332 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1333 domain_update_bandwidth(mdtc, now);
1334 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1335 }
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001336 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001337 wb_update_write_bandwidth(wb, elapsed, written);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001338
1339snapshot:
Tejun Heoa88a3412015-05-22 17:13:28 -04001340 wb->dirtied_stamp = dirtied;
1341 wb->written_stamp = written;
1342 wb->bw_time_stamp = now;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001343}
1344
Tejun Heo8a731792015-05-22 18:23:20 -04001345void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001346{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001347 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1348
Tejun Heoc2aa7232015-05-22 18:23:35 -04001349 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001350}
1351
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001353 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001354 * will look to see if it needs to start dirty throttling.
1355 *
1356 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1357 * global_page_state() too often. So scale it near-sqrt to the safety margin
1358 * (the number of pages we may dirty without exceeding the dirty limits).
1359 */
1360static unsigned long dirty_poll_interval(unsigned long dirty,
1361 unsigned long thresh)
1362{
1363 if (thresh > dirty)
1364 return 1UL << (ilog2(thresh - dirty) >> 1);
1365
1366 return 1;
1367}
1368
Tejun Heoa88a3412015-05-22 17:13:28 -04001369static unsigned long wb_max_pause(struct bdi_writeback *wb,
Tejun Heode1fff32015-05-22 17:13:29 -04001370 unsigned long wb_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001371{
Tejun Heoa88a3412015-05-22 17:13:28 -04001372 unsigned long bw = wb->avg_write_bandwidth;
Fengguang Wue3b6c652013-10-16 13:47:03 -07001373 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001374
1375 /*
1376 * Limit pause time for small memory systems. If sleeping for too long
1377 * time, a small pool of dirty/writeback pages may go empty and disk go
1378 * idle.
1379 *
1380 * 8 serves as the safety ratio.
1381 */
Tejun Heode1fff32015-05-22 17:13:29 -04001382 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001383 t++;
1384
Fengguang Wue3b6c652013-10-16 13:47:03 -07001385 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001386}
1387
Tejun Heoa88a3412015-05-22 17:13:28 -04001388static long wb_min_pause(struct bdi_writeback *wb,
1389 long max_pause,
1390 unsigned long task_ratelimit,
1391 unsigned long dirty_ratelimit,
1392 int *nr_dirtied_pause)
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001393{
Tejun Heoa88a3412015-05-22 17:13:28 -04001394 long hi = ilog2(wb->avg_write_bandwidth);
1395 long lo = ilog2(wb->dirty_ratelimit);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001396 long t; /* target pause */
1397 long pause; /* estimated next pause */
1398 int pages; /* target nr_dirtied_pause */
1399
1400 /* target for 10ms pause on 1-dd case */
1401 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001402
1403 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001404 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1405 * overheads.
1406 *
1407 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001408 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001409 if (hi > lo)
1410 t += (hi - lo) * (10 * HZ) / 1024;
1411
1412 /*
1413 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1414 * on the much more stable dirty_ratelimit. However the next pause time
1415 * will be computed based on task_ratelimit and the two rate limits may
1416 * depart considerably at some time. Especially if task_ratelimit goes
1417 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1418 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1419 * result task_ratelimit won't be executed faithfully, which could
1420 * eventually bring down dirty_ratelimit.
1421 *
1422 * We apply two rules to fix it up:
1423 * 1) try to estimate the next pause time and if necessary, use a lower
1424 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1425 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1426 * 2) limit the target pause time to max_pause/2, so that the normal
1427 * small fluctuations of task_ratelimit won't trigger rule (1) and
1428 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1429 */
1430 t = min(t, 1 + max_pause / 2);
1431 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1432
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001433 /*
1434 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1435 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1436 * When the 16 consecutive reads are often interrupted by some dirty
1437 * throttling pause during the async writes, cfq will go into idles
1438 * (deadline is fine). So push nr_dirtied_pause as high as possible
1439 * until reaches DIRTY_POLL_THRESH=32 pages.
1440 */
1441 if (pages < DIRTY_POLL_THRESH) {
1442 t = max_pause;
1443 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1444 if (pages > DIRTY_POLL_THRESH) {
1445 pages = DIRTY_POLL_THRESH;
1446 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1447 }
1448 }
1449
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001450 pause = HZ * pages / (task_ratelimit + 1);
1451 if (pause > max_pause) {
1452 t = max_pause;
1453 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1454 }
1455
1456 *nr_dirtied_pause = pages;
1457 /*
1458 * The minimal pause time will normally be half the target pause time.
1459 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001460 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001461}
1462
Tejun Heo970fb012015-05-22 18:23:24 -04001463static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
Maxim Patlasov5a537482013-09-11 14:22:46 -07001464{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001465 struct bdi_writeback *wb = dtc->wb;
Tejun Heo93f78d82015-05-22 17:13:27 -04001466 unsigned long wb_reclaimable;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001467
1468 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001469 * wb_thresh is not treated as some limiting factor as
Maxim Patlasov5a537482013-09-11 14:22:46 -07001470 * dirty_thresh, due to reasons
Tejun Heode1fff32015-05-22 17:13:29 -04001471 * - in JBOD setup, wb_thresh can fluctuate a lot
Maxim Patlasov5a537482013-09-11 14:22:46 -07001472 * - in a system with HDD and USB key, the USB key may somehow
Tejun Heode1fff32015-05-22 17:13:29 -04001473 * go into state (wb_dirty >> wb_thresh) either because
1474 * wb_dirty starts high, or because wb_thresh drops low.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001475 * In this case we don't want to hard throttle the USB key
Tejun Heode1fff32015-05-22 17:13:29 -04001476 * dirtiers for 100 seconds until wb_dirty drops under
1477 * wb_thresh. Instead the auxiliary wb control line in
Tejun Heoa88a3412015-05-22 17:13:28 -04001478 * wb_position_ratio() will let the dirtier task progress
Tejun Heode1fff32015-05-22 17:13:29 -04001479 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001480 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -04001481 dtc->wb_thresh = __wb_calc_thresh(dtc);
Tejun Heo970fb012015-05-22 18:23:24 -04001482 dtc->wb_bg_thresh = dtc->thresh ?
1483 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001484
1485 /*
1486 * In order to avoid the stacked BDI deadlock we need
1487 * to ensure we accurately count the 'dirty' pages when
1488 * the threshold is low.
1489 *
1490 * Otherwise it would be possible to get thresh+n pages
1491 * reported dirty, even though there are thresh-m pages
1492 * actually dirty; with m+n sitting in the percpu
1493 * deltas.
1494 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001495 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
Tejun Heo93f78d82015-05-22 17:13:27 -04001496 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001497 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001498 } else {
Tejun Heo93f78d82015-05-22 17:13:27 -04001499 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001500 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001501 }
1502}
1503
Wu Fengguang9d823e82011-06-11 18:10:12 -06001504/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 * balance_dirty_pages() must be called by processes which are generating dirty
1506 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001507 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001508 * If we're over `background_thresh' then the writeback threads are woken to
1509 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001511static void balance_dirty_pages(struct address_space *mapping,
Tejun Heodfb8ae52015-05-22 17:13:40 -04001512 struct bdi_writeback *wb,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001513 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001515 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001516 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001517 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001518 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1519 &mdtc_stor : NULL;
1520 struct dirty_throttle_control *sdtc;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001521 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
Wu Fengguang83712352011-06-11 19:25:42 -06001522 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001523 long pause;
1524 long max_pause;
1525 long min_pause;
1526 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001527 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001528 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001529 unsigned long dirty_ratelimit;
Tejun Heodfb8ae52015-05-22 17:13:40 -04001530 struct backing_dev_info *bdi = wb->bdi;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001531 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001532 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533
1534 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001535 unsigned long now = jiffies;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001536 unsigned long dirty, thresh, bg_thresh;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001537 unsigned long m_dirty, m_thresh, m_bg_thresh;
Wu Fengguang83712352011-06-11 19:25:42 -06001538
Wu Fengguang143dfe82010-08-27 18:45:12 -06001539 /*
1540 * Unstable writes are a feature of certain networked
1541 * filesystems (i.e. NFS) in which data may have been
1542 * written to the server's write cache, but has not yet
1543 * been flushed to permanent storage.
1544 */
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001545 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1546 global_page_state(NR_UNSTABLE_NFS);
Tejun Heo9fc3a432015-05-22 18:23:30 -04001547 gdtc->avail = global_dirtyable_memory();
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001548 gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001549
Tejun Heo9fc3a432015-05-22 18:23:30 -04001550 domain_dirty_limits(gdtc);
Wu Fengguang16c40422010-08-11 14:17:39 -07001551
Maxim Patlasov5a537482013-09-11 14:22:46 -07001552 if (unlikely(strictlimit)) {
Tejun Heo970fb012015-05-22 18:23:24 -04001553 wb_dirty_limits(gdtc);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001554
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001555 dirty = gdtc->wb_dirty;
1556 thresh = gdtc->wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -04001557 bg_thresh = gdtc->wb_bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001558 } else {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001559 dirty = gdtc->dirty;
1560 thresh = gdtc->thresh;
1561 bg_thresh = gdtc->bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001562 }
1563
Tejun Heoc2aa7232015-05-22 18:23:35 -04001564 if (mdtc) {
1565 unsigned long writeback;
1566
1567 /*
1568 * If @wb belongs to !root memcg, repeat the same
1569 * basic calculations for the memcg domain.
1570 */
1571 mem_cgroup_wb_stats(wb, &mdtc->avail, &mdtc->dirty,
1572 &writeback);
1573 mdtc_cap_avail(mdtc);
1574 mdtc->dirty += writeback;
1575
1576 domain_dirty_limits(mdtc);
1577
1578 if (unlikely(strictlimit)) {
1579 wb_dirty_limits(mdtc);
1580 m_dirty = mdtc->wb_dirty;
1581 m_thresh = mdtc->wb_thresh;
1582 m_bg_thresh = mdtc->wb_bg_thresh;
1583 } else {
1584 m_dirty = mdtc->dirty;
1585 m_thresh = mdtc->thresh;
1586 m_bg_thresh = mdtc->bg_thresh;
1587 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001588 }
1589
1590 /*
1591 * Throttle it only when the background writeback cannot
1592 * catch-up. This avoids (excessively) small writeouts
Tejun Heode1fff32015-05-22 17:13:29 -04001593 * when the wb limits are ramping up in case of !strictlimit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001594 *
Tejun Heode1fff32015-05-22 17:13:29 -04001595 * In strictlimit case make decision based on the wb counters
1596 * and limits. Small writeouts when the wb limits are ramping
Maxim Patlasov5a537482013-09-11 14:22:46 -07001597 * up are the price we consciously pay for strictlimit-ing.
Tejun Heoc2aa7232015-05-22 18:23:35 -04001598 *
1599 * If memcg domain is in effect, @dirty should be under
1600 * both global and memcg freerun ceilings.
Wu Fengguang16c40422010-08-11 14:17:39 -07001601 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001602 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1603 (!mdtc ||
1604 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1605 unsigned long intv = dirty_poll_interval(dirty, thresh);
1606 unsigned long m_intv = ULONG_MAX;
1607
Wu Fengguang83712352011-06-11 19:25:42 -06001608 current->dirty_paused_when = now;
1609 current->nr_dirtied = 0;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001610 if (mdtc)
1611 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1612 current->nr_dirtied_pause = min(intv, m_intv);
Wu Fengguang16c40422010-08-11 14:17:39 -07001613 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001614 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001615
Tejun Heobc058732015-05-22 17:13:53 -04001616 if (unlikely(!writeback_in_progress(wb)))
Tejun Heo9ecf48662015-05-22 17:13:54 -04001617 wb_start_background_writeback(wb);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001618
Tejun Heoc2aa7232015-05-22 18:23:35 -04001619 /*
1620 * Calculate global domain's pos_ratio and select the
1621 * global dtc by default.
1622 */
Maxim Patlasov5a537482013-09-11 14:22:46 -07001623 if (!strictlimit)
Tejun Heo970fb012015-05-22 18:23:24 -04001624 wb_dirty_limits(gdtc);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001625
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001626 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1627 ((gdtc->dirty > gdtc->thresh) || strictlimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628
Tejun Heodaddfa32015-05-22 18:23:26 -04001629 wb_position_ratio(gdtc);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001630 sdtc = gdtc;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001631
Tejun Heoc2aa7232015-05-22 18:23:35 -04001632 if (mdtc) {
1633 /*
1634 * If memcg domain is in effect, calculate its
1635 * pos_ratio. @wb should satisfy constraints from
1636 * both global and memcg domains. Choose the one
1637 * w/ lower pos_ratio.
1638 */
1639 if (!strictlimit)
1640 wb_dirty_limits(mdtc);
1641
1642 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1643 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1644
1645 wb_position_ratio(mdtc);
1646 if (mdtc->pos_ratio < gdtc->pos_ratio)
1647 sdtc = mdtc;
1648 }
Tejun Heodaddfa32015-05-22 18:23:26 -04001649
Tejun Heoa88a3412015-05-22 17:13:28 -04001650 if (dirty_exceeded && !wb->dirty_exceeded)
1651 wb->dirty_exceeded = 1;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001652
Tejun Heo8a731792015-05-22 18:23:20 -04001653 if (time_is_before_jiffies(wb->bw_time_stamp +
1654 BANDWIDTH_INTERVAL)) {
1655 spin_lock(&wb->list_lock);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001656 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
Tejun Heo8a731792015-05-22 18:23:20 -04001657 spin_unlock(&wb->list_lock);
1658 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659
Tejun Heoc2aa7232015-05-22 18:23:35 -04001660 /* throttle according to the chosen dtc */
Tejun Heoa88a3412015-05-22 17:13:28 -04001661 dirty_ratelimit = wb->dirty_ratelimit;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001662 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001663 RATELIMIT_CALC_SHIFT;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001664 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
Tejun Heoa88a3412015-05-22 17:13:28 -04001665 min_pause = wb_min_pause(wb, max_pause,
1666 task_ratelimit, dirty_ratelimit,
1667 &nr_dirtied_pause);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001668
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001669 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001670 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001671 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001672 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 }
Wu Fengguang83712352011-06-11 19:25:42 -06001674 period = HZ * pages_dirtied / task_ratelimit;
1675 pause = period;
1676 if (current->dirty_paused_when)
1677 pause -= now - current->dirty_paused_when;
1678 /*
1679 * For less than 1s think time (ext3/4 may block the dirtier
1680 * for up to 800ms from time to time on 1-HDD; so does xfs,
1681 * however at much less frequency), try to compensate it in
1682 * future periods by updating the virtual time; otherwise just
1683 * do a reset, as it may be a light dirtier.
1684 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001685 if (pause < min_pause) {
Wu Fengguangece13ac2010-08-29 23:33:20 -06001686 trace_balance_dirty_pages(bdi,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001687 sdtc->thresh,
1688 sdtc->bg_thresh,
1689 sdtc->dirty,
1690 sdtc->wb_thresh,
1691 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001692 dirty_ratelimit,
1693 task_ratelimit,
1694 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001695 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001696 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001697 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001698 if (pause < -HZ) {
1699 current->dirty_paused_when = now;
1700 current->nr_dirtied = 0;
1701 } else if (period) {
1702 current->dirty_paused_when += period;
1703 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001704 } else if (current->nr_dirtied_pause <= pages_dirtied)
1705 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001706 break;
1707 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001708 if (unlikely(pause > max_pause)) {
1709 /* for occasional dropped task_ratelimit */
1710 now += min(pause - max_pause, max_pause);
1711 pause = max_pause;
1712 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001713
1714pause:
Wu Fengguangece13ac2010-08-29 23:33:20 -06001715 trace_balance_dirty_pages(bdi,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001716 sdtc->thresh,
1717 sdtc->bg_thresh,
1718 sdtc->dirty,
1719 sdtc->wb_thresh,
1720 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001721 dirty_ratelimit,
1722 task_ratelimit,
1723 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001724 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001725 pause,
1726 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001727 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001728 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001729
Wu Fengguang83712352011-06-11 19:25:42 -06001730 current->dirty_paused_when = now + pause;
1731 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001732 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001733
Wu Fengguangffd1f602011-06-19 22:18:42 -06001734 /*
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001735 * This is typically equal to (dirty < thresh) and can also
1736 * keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001737 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001738 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001739 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001740
Wu Fengguangc5c63432011-12-02 10:21:33 -06001741 /*
1742 * In the case of an unresponding NFS server and the NFS dirty
Tejun Heode1fff32015-05-22 17:13:29 -04001743 * pages exceeds dirty_thresh, give the other good wb's a pipe
Wu Fengguangc5c63432011-12-02 10:21:33 -06001744 * to go through, so that tasks on them still remain responsive.
1745 *
1746 * In theory 1 page is enough to keep the comsumer-producer
1747 * pipe going: the flusher cleans 1 page => the task dirties 1
Tejun Heode1fff32015-05-22 17:13:29 -04001748 * more page. However wb_dirty has accounting errors. So use
Tejun Heo93f78d82015-05-22 17:13:27 -04001749 * the larger and more IO friendly wb_stat_error.
Wu Fengguangc5c63432011-12-02 10:21:33 -06001750 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001751 if (sdtc->wb_dirty <= wb_stat_error(wb))
Wu Fengguangc5c63432011-12-02 10:21:33 -06001752 break;
1753
Jan Kara499d05e2011-11-16 19:34:48 +08001754 if (fatal_signal_pending(current))
1755 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 }
1757
Tejun Heoa88a3412015-05-22 17:13:28 -04001758 if (!dirty_exceeded && wb->dirty_exceeded)
1759 wb->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760
Tejun Heobc058732015-05-22 17:13:53 -04001761 if (writeback_in_progress(wb))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001762 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763
1764 /*
1765 * In laptop mode, we wait until hitting the higher threshold before
1766 * starting background writeout, and then write out all the way down
1767 * to the lower threshold. So slow writers cause minimal disk activity.
1768 *
1769 * In normal mode, we start background writeout at the lower
1770 * background_thresh, to keep the amount of dirty memory low.
1771 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001772 if (laptop_mode)
1773 return;
1774
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001775 if (nr_reclaimable > gdtc->bg_thresh)
Tejun Heo9ecf48662015-05-22 17:13:54 -04001776 wb_start_background_writeback(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777}
1778
Wu Fengguang9d823e82011-06-11 18:10:12 -06001779static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001780
Wu Fengguang54848d72011-04-05 13:21:19 -06001781/*
1782 * Normal tasks are throttled by
1783 * loop {
1784 * dirty tsk->nr_dirtied_pause pages;
1785 * take a snap in balance_dirty_pages();
1786 * }
1787 * However there is a worst case. If every task exit immediately when dirtied
1788 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1789 * called to throttle the page dirties. The solution is to save the not yet
1790 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1791 * randomly into the running tasks. This works well for the above worst case,
1792 * as the new task will pick up and accumulate the old task's leaked dirty
1793 * count and eventually get throttled.
1794 */
1795DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1796
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001798 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001799 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 *
1801 * Processes which are dirtying memory should call in here once for each page
1802 * which was newly dirtied. The function will periodically check the system's
1803 * dirty state and will initiate writeback if needed.
1804 *
1805 * On really big machines, get_writeback_state is expensive, so try to avoid
1806 * calling it too often (ratelimiting). But once we're over the dirty memory
1807 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1808 * from overshooting the limit by (ratelimit_pages) each.
1809 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001810void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811{
Tejun Heodfb8ae52015-05-22 17:13:40 -04001812 struct inode *inode = mapping->host;
1813 struct backing_dev_info *bdi = inode_to_bdi(inode);
1814 struct bdi_writeback *wb = NULL;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001815 int ratelimit;
1816 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817
Wu Fengguang36715ce2011-06-11 17:53:57 -06001818 if (!bdi_cap_account_dirty(bdi))
1819 return;
1820
Tejun Heodfb8ae52015-05-22 17:13:40 -04001821 if (inode_cgwb_enabled(inode))
1822 wb = wb_get_create_current(bdi, GFP_KERNEL);
1823 if (!wb)
1824 wb = &bdi->wb;
1825
Wu Fengguang9d823e82011-06-11 18:10:12 -06001826 ratelimit = current->nr_dirtied_pause;
Tejun Heoa88a3412015-05-22 17:13:28 -04001827 if (wb->dirty_exceeded)
Wu Fengguang9d823e82011-06-11 18:10:12 -06001828 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001830 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001831 /*
1832 * This prevents one CPU to accumulate too many dirtied pages without
1833 * calling into balance_dirty_pages(), which can happen when there are
1834 * 1000+ tasks, all of them start dirtying pages at exactly the same
1835 * time, hence all honoured too large initial task->nr_dirtied_pause.
1836 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001837 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001838 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001839 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001840 else if (unlikely(*p >= ratelimit_pages)) {
1841 *p = 0;
1842 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001844 /*
1845 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1846 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1847 * the dirty throttling and livelock other long-run dirtiers.
1848 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001849 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06001850 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001851 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001852 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1853 *p -= nr_pages_dirtied;
1854 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001856 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001857
1858 if (unlikely(current->nr_dirtied >= ratelimit))
Tejun Heodfb8ae52015-05-22 17:13:40 -04001859 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1860
1861 wb_put(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001863EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864
Tejun Heoaa661bb2015-05-22 18:23:31 -04001865/**
1866 * wb_over_bg_thresh - does @wb need to be written back?
1867 * @wb: bdi_writeback of interest
1868 *
1869 * Determines whether background writeback should keep writing @wb or it's
1870 * clean enough. Returns %true if writeback should continue.
1871 */
1872bool wb_over_bg_thresh(struct bdi_writeback *wb)
1873{
Tejun Heo947e9762015-05-22 18:23:32 -04001874 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001875 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo947e9762015-05-22 18:23:32 -04001876 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001877 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1878 &mdtc_stor : NULL;
Tejun Heoaa661bb2015-05-22 18:23:31 -04001879
Tejun Heo947e9762015-05-22 18:23:32 -04001880 /*
1881 * Similar to balance_dirty_pages() but ignores pages being written
1882 * as we're trying to decide whether to put more under writeback.
1883 */
1884 gdtc->avail = global_dirtyable_memory();
1885 gdtc->dirty = global_page_state(NR_FILE_DIRTY) +
1886 global_page_state(NR_UNSTABLE_NFS);
1887 domain_dirty_limits(gdtc);
Tejun Heoaa661bb2015-05-22 18:23:31 -04001888
Tejun Heo947e9762015-05-22 18:23:32 -04001889 if (gdtc->dirty > gdtc->bg_thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04001890 return true;
1891
Tejun Heo947e9762015-05-22 18:23:32 -04001892 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(gdtc))
Tejun Heoaa661bb2015-05-22 18:23:31 -04001893 return true;
1894
Tejun Heoc2aa7232015-05-22 18:23:35 -04001895 if (mdtc) {
1896 unsigned long writeback;
1897
1898 mem_cgroup_wb_stats(wb, &mdtc->avail, &mdtc->dirty, &writeback);
1899 mdtc_cap_avail(mdtc);
1900 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1901
1902 if (mdtc->dirty > mdtc->bg_thresh)
1903 return true;
1904
1905 if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(mdtc))
1906 return true;
1907 }
1908
Tejun Heoaa661bb2015-05-22 18:23:31 -04001909 return false;
1910}
1911
Andrew Morton232ea4d2007-02-28 20:13:21 -08001912void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913{
David Rientjes364aeb22009-01-06 14:39:29 -08001914 unsigned long background_thresh;
1915 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
1917 for ( ; ; ) {
Wu Fengguang16c40422010-08-11 14:17:39 -07001918 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001919 dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920
1921 /*
1922 * Boost the allowable dirty threshold a bit for page
1923 * allocators so they don't get DoS'ed by heavy writers
1924 */
1925 dirty_thresh += dirty_thresh / 10; /* wheeee... */
1926
Christoph Lameterc24f21b2006-06-30 01:55:42 -07001927 if (global_page_state(NR_UNSTABLE_NFS) +
1928 global_page_state(NR_WRITEBACK) <= dirty_thresh)
1929 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +02001930 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -07001931
1932 /*
1933 * The caller might hold locks which can prevent IO completion
1934 * or progress in the filesystem. So we cannot just sit here
1935 * waiting for IO to complete.
1936 */
1937 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1938 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 }
1940}
1941
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1944 */
Joe Perchescccad5b2014-06-06 14:38:09 -07001945int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001946 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001948 proc_dointvec(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 return 0;
1950}
1951
Jens Axboec2c49862010-05-20 09:18:47 +02001952#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001953void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954{
Matthew Garrett31373d02010-04-06 14:25:14 +02001955 struct request_queue *q = (struct request_queue *)data;
1956 int nr_pages = global_page_state(NR_FILE_DIRTY) +
1957 global_page_state(NR_UNSTABLE_NFS);
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001958 struct bdi_writeback *wb;
1959 struct wb_iter iter;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960
Matthew Garrett31373d02010-04-06 14:25:14 +02001961 /*
1962 * We want to write everything out, not just down to the dirty
1963 * threshold
1964 */
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001965 if (!bdi_has_dirty_io(&q->backing_dev_info))
1966 return;
1967
1968 bdi_for_each_wb(wb, &q->backing_dev_info, &iter, 0)
1969 if (wb_has_dirty_io(wb))
1970 wb_start_writeback(wb, nr_pages, true,
1971 WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972}
1973
1974/*
1975 * We've spun up the disk and we're in laptop mode: schedule writeback
1976 * of all dirty data a few seconds from now. If the flush is already scheduled
1977 * then push it back - the user is still using the disk.
1978 */
Matthew Garrett31373d02010-04-06 14:25:14 +02001979void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980{
Matthew Garrett31373d02010-04-06 14:25:14 +02001981 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982}
1983
1984/*
1985 * We're in laptop mode and we've just synced. The sync's writes will have
1986 * caused another writeback to be scheduled by laptop_io_completion.
1987 * Nothing needs to be written back anymore, so we unschedule the writeback.
1988 */
1989void laptop_sync_completion(void)
1990{
Matthew Garrett31373d02010-04-06 14:25:14 +02001991 struct backing_dev_info *bdi;
1992
1993 rcu_read_lock();
1994
1995 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1996 del_timer(&bdi->laptop_mode_wb_timer);
1997
1998 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999}
Jens Axboec2c49862010-05-20 09:18:47 +02002000#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001
2002/*
2003 * If ratelimit_pages is too high then we can get into dirty-data overload
2004 * if a large number of processes all perform writes at the same time.
2005 * If it is too low then SMP machines will call the (expensive)
2006 * get_writeback_state too often.
2007 *
2008 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2009 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06002010 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 */
2012
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002013void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014{
Tejun Heodcc25ae2015-05-22 18:23:22 -04002015 struct wb_domain *dom = &global_wb_domain;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002016 unsigned long background_thresh;
2017 unsigned long dirty_thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04002018
Wu Fengguang9d823e82011-06-11 18:10:12 -06002019 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heodcc25ae2015-05-22 18:23:22 -04002020 dom->dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002021 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 if (ratelimit_pages < 16)
2023 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024}
2025
Paul Gortmaker0db06282013-06-19 14:53:51 -04002026static int
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08002027ratelimit_handler(struct notifier_block *self, unsigned long action,
2028 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029{
Srivatsa S. Bhat2f60d622012-09-28 20:27:49 +08002030
2031 switch (action & ~CPU_TASKS_FROZEN) {
2032 case CPU_ONLINE:
2033 case CPU_DEAD:
2034 writeback_set_ratelimit();
2035 return NOTIFY_OK;
2036 default:
2037 return NOTIFY_DONE;
2038 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039}
2040
Paul Gortmaker0db06282013-06-19 14:53:51 -04002041static struct notifier_block ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 .notifier_call = ratelimit_handler,
2043 .next = NULL,
2044};
2045
2046/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002047 * Called early on to tune the page writeback dirty limits.
2048 *
2049 * We used to scale dirty pages according to how total memory
2050 * related to pages that could be allocated for buffers (by
2051 * comparing nr_free_buffer_pages() to vm_total_pages.
2052 *
2053 * However, that was when we used "dirty_ratio" to scale with
2054 * all memory, and we don't do that any more. "dirty_ratio"
2055 * is now applied to total non-HIGHPAGE memory (by subtracting
2056 * totalhigh_pages from vm_total_pages), and as such we can't
2057 * get into the old insane situation any more where we had
2058 * large amounts of dirty pages compared to a small amount of
2059 * non-HIGHMEM memory.
2060 *
2061 * But we might still want to scale the dirty_ratio by how
2062 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 */
2064void __init page_writeback_init(void)
2065{
Rabin Vincenta50fcb52015-08-06 15:47:14 -07002066 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2067
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002068 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 register_cpu_notifier(&ratelimit_nb);
2070}
2071
David Howells811d7362006-08-29 19:06:09 +01002072/**
Jan Karaf446daae2010-08-09 17:19:12 -07002073 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2074 * @mapping: address space structure to write
2075 * @start: starting page index
2076 * @end: ending page index (inclusive)
2077 *
2078 * This function scans the page range from @start to @end (inclusive) and tags
2079 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2080 * that write_cache_pages (or whoever calls this function) will then use
2081 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2082 * used to avoid livelocking of writeback by a process steadily creating new
2083 * dirty pages in the file (thus it is important for this function to be quick
2084 * so that it can tag pages faster than a dirtying process can create them).
2085 */
2086/*
2087 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2088 */
Jan Karaf446daae2010-08-09 17:19:12 -07002089void tag_pages_for_writeback(struct address_space *mapping,
2090 pgoff_t start, pgoff_t end)
2091{
Randy Dunlap3c111a02010-08-11 14:17:30 -07002092#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07002093 unsigned long tagged;
2094
2095 do {
2096 spin_lock_irq(&mapping->tree_lock);
2097 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2098 &start, end, WRITEBACK_TAG_BATCH,
2099 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2100 spin_unlock_irq(&mapping->tree_lock);
2101 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2102 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07002103 /* We check 'start' to handle wrapping when end == ~0UL */
2104 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07002105}
2106EXPORT_SYMBOL(tag_pages_for_writeback);
2107
2108/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002109 * 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 +01002110 * @mapping: address space structure to write
2111 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002112 * @writepage: function called for each page
2113 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01002114 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002115 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01002116 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2117 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2118 * and msync() need to guarantee that all the data which was dirty at the time
2119 * the call was made get new I/O started against them. If wbc->sync_mode is
2120 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2121 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07002122 *
2123 * To avoid livelocks (when other process dirties new pages), we first tag
2124 * pages which should be written back with TOWRITE tag and only then start
2125 * writing them. For data-integrity sync we have to be careful so that we do
2126 * not miss some pages (e.g., because some other process has cleared TOWRITE
2127 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2128 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01002129 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002130int write_cache_pages(struct address_space *mapping,
2131 struct writeback_control *wbc, writepage_t writepage,
2132 void *data)
David Howells811d7362006-08-29 19:06:09 +01002133{
David Howells811d7362006-08-29 19:06:09 +01002134 int ret = 0;
2135 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01002136 struct pagevec pvec;
2137 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08002138 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01002139 pgoff_t index;
2140 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08002141 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08002142 int cycled;
David Howells811d7362006-08-29 19:06:09 +01002143 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07002144 int tag;
David Howells811d7362006-08-29 19:06:09 +01002145
David Howells811d7362006-08-29 19:06:09 +01002146 pagevec_init(&pvec, 0);
2147 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08002148 writeback_index = mapping->writeback_index; /* prev offset */
2149 index = writeback_index;
2150 if (index == 0)
2151 cycled = 1;
2152 else
2153 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01002154 end = -1;
2155 } else {
2156 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2157 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2158 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2159 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08002160 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01002161 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002162 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002163 tag = PAGECACHE_TAG_TOWRITE;
2164 else
2165 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01002166retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002167 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002168 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08002169 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002170 while (!done && (index <= end)) {
2171 int i;
2172
Jan Karaf446daae2010-08-09 17:19:12 -07002173 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002174 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2175 if (nr_pages == 0)
2176 break;
David Howells811d7362006-08-29 19:06:09 +01002177
David Howells811d7362006-08-29 19:06:09 +01002178 for (i = 0; i < nr_pages; i++) {
2179 struct page *page = pvec.pages[i];
2180
Nick Piggind5482cd2009-01-06 14:39:11 -08002181 /*
2182 * At this point, the page may be truncated or
2183 * invalidated (changing page->mapping to NULL), or
2184 * even swizzled back from swapper_space to tmpfs file
2185 * mapping. However, page->index will not change
2186 * because we have a reference on the page.
2187 */
2188 if (page->index > end) {
2189 /*
2190 * can't be range_cyclic (1st pass) because
2191 * end == -1 in that case.
2192 */
2193 done = 1;
2194 break;
2195 }
2196
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002197 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08002198
David Howells811d7362006-08-29 19:06:09 +01002199 lock_page(page);
2200
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002201 /*
2202 * Page truncated or invalidated. We can freely skip it
2203 * then, even for data integrity operations: the page
2204 * has disappeared concurrently, so there could be no
2205 * real expectation of this data interity operation
2206 * even if there is now a new, dirty page at the same
2207 * pagecache address.
2208 */
David Howells811d7362006-08-29 19:06:09 +01002209 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002210continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01002211 unlock_page(page);
2212 continue;
2213 }
2214
Nick Piggin515f4a02009-01-06 14:39:10 -08002215 if (!PageDirty(page)) {
2216 /* someone wrote it for us */
2217 goto continue_unlock;
2218 }
David Howells811d7362006-08-29 19:06:09 +01002219
Nick Piggin515f4a02009-01-06 14:39:10 -08002220 if (PageWriteback(page)) {
2221 if (wbc->sync_mode != WB_SYNC_NONE)
2222 wait_on_page_writeback(page);
2223 else
2224 goto continue_unlock;
2225 }
2226
2227 BUG_ON(PageWriteback(page));
2228 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002229 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01002230
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002231 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002232 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08002233 if (unlikely(ret)) {
2234 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2235 unlock_page(page);
2236 ret = 0;
2237 } else {
2238 /*
2239 * done_index is set past this page,
2240 * so media errors will not choke
2241 * background writeout for the entire
2242 * file. This has consequences for
2243 * range_cyclic semantics (ie. it may
2244 * not be suitable for data integrity
2245 * writeout).
2246 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002247 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08002248 done = 1;
2249 break;
2250 }
Dave Chinner0b564922010-06-09 10:37:18 +10002251 }
David Howells811d7362006-08-29 19:06:09 +01002252
Dave Chinner546a1922010-08-24 11:44:34 +10002253 /*
2254 * We stop writing back only if we are not doing
2255 * integrity sync. In case of integrity sync we have to
2256 * keep going until we have written all the pages
2257 * we tagged for writeback prior to entering this loop.
2258 */
2259 if (--wbc->nr_to_write <= 0 &&
2260 wbc->sync_mode == WB_SYNC_NONE) {
2261 done = 1;
2262 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08002263 }
David Howells811d7362006-08-29 19:06:09 +01002264 }
2265 pagevec_release(&pvec);
2266 cond_resched();
2267 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01002268 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01002269 /*
Nick Piggin31a12662009-01-06 14:39:04 -08002270 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01002271 * We hit the last page and there is more work to be done: wrap
2272 * back to the start of the file
2273 */
Nick Piggin31a12662009-01-06 14:39:04 -08002274 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01002275 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08002276 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01002277 goto retry;
2278 }
Dave Chinner0b564922010-06-09 10:37:18 +10002279 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2280 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04002281
David Howells811d7362006-08-29 19:06:09 +01002282 return ret;
2283}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002284EXPORT_SYMBOL(write_cache_pages);
2285
2286/*
2287 * Function used by generic_writepages to call the real writepage
2288 * function and set the mapping flags on error
2289 */
2290static int __writepage(struct page *page, struct writeback_control *wbc,
2291 void *data)
2292{
2293 struct address_space *mapping = data;
2294 int ret = mapping->a_ops->writepage(page, wbc);
2295 mapping_set_error(mapping, ret);
2296 return ret;
2297}
2298
2299/**
2300 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2301 * @mapping: address space structure to write
2302 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2303 *
2304 * This is a library function, which implements the writepages()
2305 * address_space_operation.
2306 */
2307int generic_writepages(struct address_space *mapping,
2308 struct writeback_control *wbc)
2309{
Shaohua Li9b6096a2011-03-17 10:47:06 +01002310 struct blk_plug plug;
2311 int ret;
2312
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002313 /* deal with chardevs and other special file */
2314 if (!mapping->a_ops->writepage)
2315 return 0;
2316
Shaohua Li9b6096a2011-03-17 10:47:06 +01002317 blk_start_plug(&plug);
2318 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2319 blk_finish_plug(&plug);
2320 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002321}
David Howells811d7362006-08-29 19:06:09 +01002322
2323EXPORT_SYMBOL(generic_writepages);
2324
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2326{
Andrew Morton22905f72005-11-16 15:07:01 -08002327 int ret;
2328
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 if (wbc->nr_to_write <= 0)
2330 return 0;
2331 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002332 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002333 else
2334 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002335 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336}
2337
2338/**
2339 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07002340 * @page: the page to write
2341 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 *
2343 * The page must be locked by the caller and will be unlocked upon return.
2344 *
2345 * write_one_page() returns a negative error code if I/O failed.
2346 */
2347int write_one_page(struct page *page, int wait)
2348{
2349 struct address_space *mapping = page->mapping;
2350 int ret = 0;
2351 struct writeback_control wbc = {
2352 .sync_mode = WB_SYNC_ALL,
2353 .nr_to_write = 1,
2354 };
2355
2356 BUG_ON(!PageLocked(page));
2357
2358 if (wait)
2359 wait_on_page_writeback(page);
2360
2361 if (clear_page_dirty_for_io(page)) {
2362 page_cache_get(page);
2363 ret = mapping->a_ops->writepage(page, &wbc);
2364 if (ret == 0 && wait) {
2365 wait_on_page_writeback(page);
2366 if (PageError(page))
2367 ret = -EIO;
2368 }
2369 page_cache_release(page);
2370 } else {
2371 unlock_page(page);
2372 }
2373 return ret;
2374}
2375EXPORT_SYMBOL(write_one_page);
2376
2377/*
Ken Chen76719322007-02-10 01:43:15 -08002378 * For address_spaces which do not use buffers nor write back.
2379 */
2380int __set_page_dirty_no_writeback(struct page *page)
2381{
2382 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08002383 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08002384 return 0;
2385}
2386
2387/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002388 * Helper function for set_page_dirty family.
Greg Thelenc4843a72015-05-22 17:13:16 -04002389 *
2390 * Caller must hold mem_cgroup_begin_page_stat().
2391 *
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002392 * NOTE: This relies on being atomic wrt interrupts.
2393 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002394void account_page_dirtied(struct page *page, struct address_space *mapping,
2395 struct mem_cgroup *memcg)
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002396{
Tejun Heo52ebea72015-05-22 17:13:37 -04002397 struct inode *inode = mapping->host;
2398
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002399 trace_writeback_dirty_page(page, mapping);
2400
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002401 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo52ebea72015-05-22 17:13:37 -04002402 struct bdi_writeback *wb;
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002403
Tejun Heo52ebea72015-05-22 17:13:37 -04002404 inode_attach_wb(inode, page);
2405 wb = inode_to_wb(inode);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002406
Greg Thelenc4843a72015-05-22 17:13:16 -04002407 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002408 __inc_zone_page_state(page, NR_FILE_DIRTY);
Michael Rubinea941f02010-10-26 14:21:35 -07002409 __inc_zone_page_state(page, NR_DIRTIED);
Tejun Heo52ebea72015-05-22 17:13:37 -04002410 __inc_wb_stat(wb, WB_RECLAIMABLE);
2411 __inc_wb_stat(wb, WB_DIRTIED);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002412 task_io_account_write(PAGE_CACHE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06002413 current->nr_dirtied++;
2414 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002415 }
2416}
Michael Rubin679ceac2010-08-20 02:31:26 -07002417EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002418
2419/*
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002420 * Helper function for deaccounting dirty page without writeback.
2421 *
Greg Thelenc4843a72015-05-22 17:13:16 -04002422 * Caller must hold mem_cgroup_begin_page_stat().
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002423 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002424void account_page_cleaned(struct page *page, struct address_space *mapping,
Tejun Heo682aa8e2015-05-28 14:50:53 -04002425 struct mem_cgroup *memcg, struct bdi_writeback *wb)
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002426{
2427 if (mapping_cap_account_dirty(mapping)) {
Greg Thelenc4843a72015-05-22 17:13:16 -04002428 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002429 dec_zone_page_state(page, NR_FILE_DIRTY);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002430 dec_wb_stat(wb, WB_RECLAIMABLE);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002431 task_io_account_cancelled_write(PAGE_CACHE_SIZE);
2432 }
2433}
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002434
2435/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 * For address_spaces which do not use buffers. Just tag the page as dirty in
2437 * its radix tree.
2438 *
2439 * This is also used when a single buffer is being dirtied: we want to set the
2440 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2441 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2442 *
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002443 * The caller must ensure this doesn't race with truncation. Most will simply
2444 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2445 * the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 */
2447int __set_page_dirty_nobuffers(struct page *page)
2448{
Greg Thelenc4843a72015-05-22 17:13:16 -04002449 struct mem_cgroup *memcg;
2450
2451 memcg = mem_cgroup_begin_page_stat(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 if (!TestSetPageDirty(page)) {
2453 struct address_space *mapping = page_mapping(page);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002454 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
Greg Thelenc4843a72015-05-22 17:13:16 -04002456 if (!mapping) {
2457 mem_cgroup_end_page_stat(memcg);
Andrew Morton8c085402006-12-10 02:19:24 -08002458 return 1;
Greg Thelenc4843a72015-05-22 17:13:16 -04002459 }
Andrew Morton8c085402006-12-10 02:19:24 -08002460
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002461 spin_lock_irqsave(&mapping->tree_lock, flags);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002462 BUG_ON(page_mapping(page) != mapping);
2463 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Greg Thelenc4843a72015-05-22 17:13:16 -04002464 account_page_dirtied(page, mapping, memcg);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002465 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2466 PAGECACHE_TAG_DIRTY);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002467 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Greg Thelenc4843a72015-05-22 17:13:16 -04002468 mem_cgroup_end_page_stat(memcg);
2469
Andrew Morton8c085402006-12-10 02:19:24 -08002470 if (mapping->host) {
2471 /* !PageAnon && !swapper_space */
2472 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002474 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475 }
Greg Thelenc4843a72015-05-22 17:13:16 -04002476 mem_cgroup_end_page_stat(memcg);
Andrew Morton4741c9f2006-03-24 03:18:11 -08002477 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478}
2479EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2480
2481/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002482 * Call this whenever redirtying a page, to de-account the dirty counters
2483 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2484 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2485 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2486 * control.
2487 */
2488void account_page_redirty(struct page *page)
2489{
2490 struct address_space *mapping = page->mapping;
Tejun Heo91018132015-05-22 17:13:39 -04002491
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002492 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002493 struct inode *inode = mapping->host;
2494 struct bdi_writeback *wb;
2495 bool locked;
Tejun Heo91018132015-05-22 17:13:39 -04002496
Tejun Heo682aa8e2015-05-28 14:50:53 -04002497 wb = unlocked_inode_to_wb_begin(inode, &locked);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002498 current->nr_dirtied--;
2499 dec_zone_page_state(page, NR_DIRTIED);
Tejun Heo91018132015-05-22 17:13:39 -04002500 dec_wb_stat(wb, WB_DIRTIED);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002501 unlocked_inode_to_wb_end(inode, locked);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002502 }
2503}
2504EXPORT_SYMBOL(account_page_redirty);
2505
2506/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 * When a writepage implementation decides that it doesn't want to write this
2508 * page for some reason, it should redirty the locked page via
2509 * redirty_page_for_writepage() and it should then unlock the page and return 0
2510 */
2511int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2512{
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002513 int ret;
2514
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 wbc->pages_skipped++;
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002516 ret = __set_page_dirty_nobuffers(page);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002517 account_page_redirty(page);
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002518 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519}
2520EXPORT_SYMBOL(redirty_page_for_writepage);
2521
2522/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002523 * Dirty a page.
2524 *
2525 * For pages with a mapping this should be done under the page lock
2526 * for the benefit of asynchronous memory errors who prefer a consistent
2527 * dirty state. This rule can be broken in some special cases,
2528 * but should be better not to.
2529 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 * If the mapping doesn't provide a set_page_dirty a_op, then
2531 * just fall through and assume that it wants buffer_heads.
2532 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002533int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534{
2535 struct address_space *mapping = page_mapping(page);
2536
2537 if (likely(mapping)) {
2538 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002539 /*
2540 * readahead/lru_deactivate_page could remain
2541 * PG_readahead/PG_reclaim due to race with end_page_writeback
2542 * About readahead, if the page is written, the flags would be
2543 * reset. So no problem.
2544 * About lru_deactivate_page, if the page is redirty, the flag
2545 * will be reset. So no problem. but if the page is used by readahead
2546 * it will confuse readahead and make it restart the size rampup
2547 * process. But it's a trivial problem.
2548 */
Naoya Horiguchia4bb3ec2015-04-15 16:13:17 -07002549 if (PageReclaim(page))
2550 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002551#ifdef CONFIG_BLOCK
2552 if (!spd)
2553 spd = __set_page_dirty_buffers;
2554#endif
2555 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002557 if (!PageDirty(page)) {
2558 if (!TestSetPageDirty(page))
2559 return 1;
2560 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 return 0;
2562}
2563EXPORT_SYMBOL(set_page_dirty);
2564
2565/*
2566 * set_page_dirty() is racy if the caller has no reference against
2567 * page->mapping->host, and if the page is unlocked. This is because another
2568 * CPU could truncate the page off the mapping and then free the mapping.
2569 *
2570 * Usually, the page _is_ locked, or the caller is a user-space process which
2571 * holds a reference on the inode by having an open file.
2572 *
2573 * In other cases, the page should be locked before running set_page_dirty().
2574 */
2575int set_page_dirty_lock(struct page *page)
2576{
2577 int ret;
2578
Jens Axboe7eaceac2011-03-10 08:52:07 +01002579 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 ret = set_page_dirty(page);
2581 unlock_page(page);
2582 return ret;
2583}
2584EXPORT_SYMBOL(set_page_dirty_lock);
2585
2586/*
Tejun Heo11f81be2015-05-22 17:13:15 -04002587 * This cancels just the dirty bit on the kernel page itself, it does NOT
2588 * actually remove dirty bits on any mmap's that may be around. It also
2589 * leaves the page tagged dirty, so any sync activity will still find it on
2590 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2591 * look at the dirty bits in the VM.
2592 *
2593 * Doing this should *normally* only ever be done when a page is truncated,
2594 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2595 * this when it notices that somebody has cleaned out all the buffers on a
2596 * page without actually doing it through the VM. Can you say "ext3 is
2597 * horribly ugly"? Thought you could.
2598 */
2599void cancel_dirty_page(struct page *page)
2600{
Greg Thelenc4843a72015-05-22 17:13:16 -04002601 struct address_space *mapping = page_mapping(page);
2602
2603 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002604 struct inode *inode = mapping->host;
2605 struct bdi_writeback *wb;
Greg Thelenc4843a72015-05-22 17:13:16 -04002606 struct mem_cgroup *memcg;
Tejun Heo682aa8e2015-05-28 14:50:53 -04002607 bool locked;
Greg Thelenc4843a72015-05-22 17:13:16 -04002608
2609 memcg = mem_cgroup_begin_page_stat(page);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002610 wb = unlocked_inode_to_wb_begin(inode, &locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002611
2612 if (TestClearPageDirty(page))
Tejun Heo682aa8e2015-05-28 14:50:53 -04002613 account_page_cleaned(page, mapping, memcg, wb);
Greg Thelenc4843a72015-05-22 17:13:16 -04002614
Tejun Heo682aa8e2015-05-28 14:50:53 -04002615 unlocked_inode_to_wb_end(inode, locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002616 mem_cgroup_end_page_stat(memcg);
2617 } else {
2618 ClearPageDirty(page);
2619 }
Tejun Heo11f81be2015-05-22 17:13:15 -04002620}
2621EXPORT_SYMBOL(cancel_dirty_page);
2622
2623/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 * Clear a page's dirty flag, while caring for dirty memory accounting.
2625 * Returns true if the page was previously dirty.
2626 *
2627 * This is for preparing to put the page under writeout. We leave the page
2628 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2629 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2630 * implementation will run either set_page_writeback() or set_page_dirty(),
2631 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2632 * back into sync.
2633 *
2634 * This incoherency between the page's dirty flag and radix-tree tag is
2635 * unfortunate, but it only exists while the page is locked.
2636 */
2637int clear_page_dirty_for_io(struct page *page)
2638{
2639 struct address_space *mapping = page_mapping(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002640 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641
Nick Piggin79352892007-07-19 01:47:22 -07002642 BUG_ON(!PageLocked(page));
2643
Linus Torvalds7658cc22006-12-29 10:00:58 -08002644 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002645 struct inode *inode = mapping->host;
2646 struct bdi_writeback *wb;
2647 struct mem_cgroup *memcg;
2648 bool locked;
2649
Linus Torvalds7658cc22006-12-29 10:00:58 -08002650 /*
2651 * Yes, Virginia, this is indeed insane.
2652 *
2653 * We use this sequence to make sure that
2654 * (a) we account for dirty stats properly
2655 * (b) we tell the low-level filesystem to
2656 * mark the whole page dirty if it was
2657 * dirty in a pagetable. Only to then
2658 * (c) clean the page again and return 1 to
2659 * cause the writeback.
2660 *
2661 * This way we avoid all nasty races with the
2662 * dirty bit in multiple places and clearing
2663 * them concurrently from different threads.
2664 *
2665 * Note! Normally the "set_page_dirty(page)"
2666 * has no effect on the actual dirty bit - since
2667 * that will already usually be set. But we
2668 * need the side effects, and it can help us
2669 * avoid races.
2670 *
2671 * We basically use the page "master dirty bit"
2672 * as a serialization point for all the different
2673 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002674 */
2675 if (page_mkclean(page))
2676 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002677 /*
2678 * We carefully synchronise fault handlers against
2679 * installing a dirty pte and marking the page dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002680 * at this point. We do this by having them hold the
2681 * page lock while dirtying the page, and pages are
2682 * always locked coming in here, so we get the desired
2683 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002684 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002685 memcg = mem_cgroup_begin_page_stat(page);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002686 wb = unlocked_inode_to_wb_begin(inode, &locked);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002687 if (TestClearPageDirty(page)) {
Greg Thelenc4843a72015-05-22 17:13:16 -04002688 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY);
Andrew Morton8c085402006-12-10 02:19:24 -08002689 dec_zone_page_state(page, NR_FILE_DIRTY);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002690 dec_wb_stat(wb, WB_RECLAIMABLE);
Greg Thelenc4843a72015-05-22 17:13:16 -04002691 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 }
Tejun Heo682aa8e2015-05-28 14:50:53 -04002693 unlocked_inode_to_wb_end(inode, locked);
Greg Thelenc4843a72015-05-22 17:13:16 -04002694 mem_cgroup_end_page_stat(memcg);
2695 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002697 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002699EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700
2701int test_clear_page_writeback(struct page *page)
2702{
2703 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002704 struct mem_cgroup *memcg;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002705 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706
Johannes Weiner6de22612015-02-11 15:25:01 -08002707 memcg = mem_cgroup_begin_page_stat(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 if (mapping) {
Tejun Heo91018132015-05-22 17:13:39 -04002709 struct inode *inode = mapping->host;
2710 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 unsigned long flags;
2712
Nick Piggin19fd6232008-07-25 19:45:32 -07002713 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002715 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 radix_tree_tag_clear(&mapping->page_tree,
2717 page_index(page),
2718 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002719 if (bdi_cap_account_writeback(bdi)) {
Tejun Heo91018132015-05-22 17:13:39 -04002720 struct bdi_writeback *wb = inode_to_wb(inode);
2721
2722 __dec_wb_stat(wb, WB_WRITEBACK);
2723 __wb_writeout_inc(wb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002724 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002725 }
Nick Piggin19fd6232008-07-25 19:45:32 -07002726 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 } else {
2728 ret = TestClearPageWriteback(page);
2729 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002730 if (ret) {
Johannes Weinerd7365e72014-10-29 14:50:48 -07002731 mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
Andrew Mortond688abf2007-07-19 01:49:17 -07002732 dec_zone_page_state(page, NR_WRITEBACK);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002733 inc_zone_page_state(page, NR_WRITTEN);
2734 }
Johannes Weiner6de22612015-02-11 15:25:01 -08002735 mem_cgroup_end_page_stat(memcg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 return ret;
2737}
2738
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002739int __test_set_page_writeback(struct page *page, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740{
2741 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002742 struct mem_cgroup *memcg;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002743 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744
Johannes Weiner6de22612015-02-11 15:25:01 -08002745 memcg = mem_cgroup_begin_page_stat(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 if (mapping) {
Tejun Heo91018132015-05-22 17:13:39 -04002747 struct inode *inode = mapping->host;
2748 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 unsigned long flags;
2750
Nick Piggin19fd6232008-07-25 19:45:32 -07002751 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002753 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 radix_tree_tag_set(&mapping->page_tree,
2755 page_index(page),
2756 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002757 if (bdi_cap_account_writeback(bdi))
Tejun Heo91018132015-05-22 17:13:39 -04002758 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002759 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 if (!PageDirty(page))
2761 radix_tree_tag_clear(&mapping->page_tree,
2762 page_index(page),
2763 PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002764 if (!keep_write)
2765 radix_tree_tag_clear(&mapping->page_tree,
2766 page_index(page),
2767 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002768 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769 } else {
2770 ret = TestSetPageWriteback(page);
2771 }
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002772 if (!ret) {
Johannes Weinerd7365e72014-10-29 14:50:48 -07002773 mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002774 inc_zone_page_state(page, NR_WRITEBACK);
2775 }
Johannes Weiner6de22612015-02-11 15:25:01 -08002776 mem_cgroup_end_page_stat(memcg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 return ret;
2778
2779}
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002780EXPORT_SYMBOL(__test_set_page_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781
2782/*
Nick Piggin00128182007-10-16 01:24:40 -07002783 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 * passed tag.
2785 */
2786int mapping_tagged(struct address_space *mapping, int tag)
2787{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002788 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789}
2790EXPORT_SYMBOL(mapping_tagged);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002791
2792/**
2793 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2794 * @page: The page to wait on.
2795 *
2796 * This function determines if the given page is related to a backing device
2797 * that requires page contents to be held stable during writeback. If so, then
2798 * it will wait for any pending writeback to complete.
2799 */
2800void wait_for_stable_page(struct page *page)
2801{
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002802 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2803 wait_on_page_writeback(page);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002804}
2805EXPORT_SYMBOL_GPL(wait_for_stable_page);