blob: 692c0da04cbd030952c6037a33cbd06e3362c39d [file] [log] [blame]
Thomas Gleixner457c8992019-05-19 13:08:55 +01001// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02003 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
5 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra90eec102015-11-16 11:08:45 +01006 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
Linus Torvalds1da177e2005-04-16 15:20:36 -07007 *
8 * Contains functions related to writing back dirty pages at the
9 * address_space level.
10 *
Francois Camie1f8e872008-10-15 22:01:59 -070011 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 * Initial version
13 */
14
15#include <linux/kernel.h>
Stefan Roesch1bf27e92022-11-18 16:52:02 -080016#include <linux/math64.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040017#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018#include <linux/spinlock.h>
19#include <linux/fs.h>
20#include <linux/mm.h>
21#include <linux/swap.h>
22#include <linux/slab.h>
23#include <linux/pagemap.h>
24#include <linux/writeback.h>
25#include <linux/init.h>
26#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080027#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/blkdev.h>
29#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070030#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#include <linux/percpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/smp.h>
33#include <linux/sysctl.h>
34#include <linux/cpu.h>
35#include <linux/syscalls.h>
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>
Ingo Molnarf361bf42017-02-03 23:47:37 +010039#include <linux/sched/signal.h>
Lisa Du6e543d52013-09-11 14:22:36 -070040#include <linux/mm_inline.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100041#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042
Lisa Du6e543d52013-09-11 14:22:36 -070043#include "internal.h"
44
Linus Torvalds1da177e2005-04-16 15:20:36 -070045/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060046 * Sleep at most 200ms at a time in balance_dirty_pages().
47 */
48#define MAX_PAUSE max(HZ/5, 1)
49
50/*
Wu Fengguang5b9b3572011-12-06 13:17:17 -060051 * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 * by raising pause time to max_pause when falls below it.
53 */
54#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
55
56/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060057 * Estimate write bandwidth at 200ms intervals.
58 */
59#define BANDWIDTH_INTERVAL max(HZ/5, 1)
60
Wu Fengguang6c14ae12011-03-02 16:04:18 -060061#define RATELIMIT_CALC_SHIFT 10
62
Wu Fengguange98be2d2010-08-29 11:22:30 -060063/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 * will look to see if it needs to force writeback or throttling.
66 */
67static long ratelimit_pages = 32;
68
Linus Torvalds1da177e2005-04-16 15:20:36 -070069/* The following parameters are exported via /proc/sys/vm */
70
71/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020072 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070073 */
zhanglianjieaa779e52022-02-17 18:51:51 -080074static int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070075
76/*
David Rientjes2da02992009-01-06 14:39:31 -080077 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
79 */
zhanglianjieaa779e52022-02-17 18:51:51 -080080static unsigned long dirty_background_bytes;
David Rientjes2da02992009-01-06 14:39:31 -080081
82/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080083 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 */
zhanglianjieaa779e52022-02-17 18:51:51 -080086static int vm_highmem_is_dirtyable;
Bron Gondwana195cf4532008-02-04 22:29:20 -080087
88/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 * The generator of dirty data starts writeback at this percentage
90 */
zhanglianjieaa779e52022-02-17 18:51:51 -080091static int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
93/*
David Rientjes2da02992009-01-06 14:39:31 -080094 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
96 */
zhanglianjieaa779e52022-02-17 18:51:51 -080097static unsigned long vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -080098
99/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700100 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Artem Bityutskiy91913a22012-03-21 22:33:00 -0400104EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700107 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800112 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
113 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 */
115int laptop_mode;
116
117EXPORT_SYMBOL(laptop_mode);
118
119/* End of sysctl-exported parameters */
120
Tejun Heodcc25ae2015-05-22 18:23:22 -0400121struct wb_domain global_wb_domain;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400123/* consolidated parameters for balance_dirty_pages() and its subroutines */
124struct dirty_throttle_control {
Tejun Heoe9f07df2015-05-22 18:23:28 -0400125#ifdef CONFIG_CGROUP_WRITEBACK
126 struct wb_domain *dom;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400127 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
Tejun Heoe9f07df2015-05-22 18:23:28 -0400128#endif
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400129 struct bdi_writeback *wb;
Tejun Heoe9770b32015-05-22 18:23:27 -0400130 struct fprop_local_percpu *wb_completions;
Jan Karaeb608e32012-05-24 18:59:11 +0200131
Tejun Heo9fc3a432015-05-22 18:23:30 -0400132 unsigned long avail; /* dirtyable */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400133 unsigned long dirty; /* file_dirty + write + nfs */
134 unsigned long thresh; /* dirty threshold */
135 unsigned long bg_thresh; /* dirty background threshold */
136
137 unsigned long wb_dirty; /* per-wb counterparts */
138 unsigned long wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -0400139 unsigned long wb_bg_thresh;
Tejun Heodaddfa32015-05-22 18:23:26 -0400140
141 unsigned long pos_ratio;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400142};
143
Jan Karaeb608e32012-05-24 18:59:11 +0200144/*
145 * Length of period for aging writeout fractions of bdis. This is an
146 * arbitrarily chosen number. The longer the period, the slower fractions will
147 * reflect changes in current writeout rate.
148 */
149#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700150
Tejun Heo693108a2015-05-22 17:13:49 -0400151#ifdef CONFIG_CGROUP_WRITEBACK
152
Tejun Heod60d1bd2015-09-29 12:47:53 -0400153#define GDTC_INIT(__wb) .wb = (__wb), \
154 .dom = &global_wb_domain, \
155 .wb_completions = &(__wb)->completions
156
Tejun Heo9fc3a432015-05-22 18:23:30 -0400157#define GDTC_INIT_NO_WB .dom = &global_wb_domain
Tejun Heod60d1bd2015-09-29 12:47:53 -0400158
159#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
160 .dom = mem_cgroup_wb_domain(__wb), \
161 .wb_completions = &(__wb)->memcg_completions, \
162 .gdtc = __gdtc
Tejun Heoc2aa7232015-05-22 18:23:35 -0400163
164static bool mdtc_valid(struct dirty_throttle_control *dtc)
165{
166 return dtc->dom;
167}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400168
169static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
170{
171 return dtc->dom;
172}
173
Tejun Heo9fc3a432015-05-22 18:23:30 -0400174static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
175{
176 return mdtc->gdtc;
177}
178
Tejun Heo841710a2015-05-22 18:23:33 -0400179static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
180{
181 return &wb->memcg_completions;
182}
183
Tejun Heo693108a2015-05-22 17:13:49 -0400184static void wb_min_max_ratio(struct bdi_writeback *wb,
185 unsigned long *minp, unsigned long *maxp)
186{
Jan Kara20792eb2021-09-02 14:53:15 -0700187 unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
Tejun Heo693108a2015-05-22 17:13:49 -0400188 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
189 unsigned long long min = wb->bdi->min_ratio;
190 unsigned long long max = wb->bdi->max_ratio;
191
192 /*
193 * @wb may already be clean by the time control reaches here and
194 * the total may not include its bw.
195 */
196 if (this_bw < tot_bw) {
197 if (min) {
198 min *= this_bw;
Wen Yang6d9e8c62020-01-13 16:29:23 -0800199 min = div64_ul(min, tot_bw);
Tejun Heo693108a2015-05-22 17:13:49 -0400200 }
Stefan Roeschae822912022-11-18 16:51:59 -0800201 if (max < 100 * BDI_RATIO_SCALE) {
Tejun Heo693108a2015-05-22 17:13:49 -0400202 max *= this_bw;
Wen Yang6d9e8c62020-01-13 16:29:23 -0800203 max = div64_ul(max, tot_bw);
Tejun Heo693108a2015-05-22 17:13:49 -0400204 }
205 }
206
207 *minp = min;
208 *maxp = max;
209}
210
211#else /* CONFIG_CGROUP_WRITEBACK */
212
Tejun Heod60d1bd2015-09-29 12:47:53 -0400213#define GDTC_INIT(__wb) .wb = (__wb), \
214 .wb_completions = &(__wb)->completions
Tejun Heo9fc3a432015-05-22 18:23:30 -0400215#define GDTC_INIT_NO_WB
Tejun Heoc2aa7232015-05-22 18:23:35 -0400216#define MDTC_INIT(__wb, __gdtc)
217
218static bool mdtc_valid(struct dirty_throttle_control *dtc)
219{
220 return false;
221}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400222
223static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
224{
225 return &global_wb_domain;
226}
227
Tejun Heo9fc3a432015-05-22 18:23:30 -0400228static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
229{
230 return NULL;
231}
232
Tejun Heo841710a2015-05-22 18:23:33 -0400233static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
234{
235 return NULL;
236}
237
Tejun Heo693108a2015-05-22 17:13:49 -0400238static void wb_min_max_ratio(struct bdi_writeback *wb,
239 unsigned long *minp, unsigned long *maxp)
240{
241 *minp = wb->bdi->min_ratio;
242 *maxp = wb->bdi->max_ratio;
243}
244
245#endif /* CONFIG_CGROUP_WRITEBACK */
246
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700247/*
Johannes Weinera756cf52012-01-10 15:07:49 -0800248 * In a memory zone, there is a certain amount of pages we consider
249 * available for the page cache, which is essentially the number of
250 * free and reclaimable pages, minus some zone reserves to protect
251 * lowmem and the ability to uphold the zone's watermarks without
252 * requiring writeback.
253 *
254 * This number of dirtyable pages is the base value of which the
Ethon Paule0857cf2020-06-04 16:49:40 -0700255 * user-configurable dirty ratio is the effective number of pages that
Johannes Weinera756cf52012-01-10 15:07:49 -0800256 * are allowed to be actually dirtied. Per individual zone, or
257 * globally by using the sum of dirtyable pages over all zones.
258 *
259 * Because the user is allowed to specify the dirty limit globally as
260 * absolute number of bytes, calculating the per-zone dirty limit can
261 * require translating the configured limit into a percentage of
262 * global dirtyable memory first.
263 */
264
Johannes Weinera804552b2014-01-29 14:05:39 -0800265/**
Mel Gorman281e3722016-07-28 15:46:11 -0700266 * node_dirtyable_memory - number of dirtyable pages in a node
267 * @pgdat: the node
Johannes Weinera804552b2014-01-29 14:05:39 -0800268 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800269 * Return: the node's number of pages potentially available for dirty
Mel Gorman281e3722016-07-28 15:46:11 -0700270 * page cache. This is the base value for the per-node dirty limits.
Johannes Weinera804552b2014-01-29 14:05:39 -0800271 */
Mel Gorman281e3722016-07-28 15:46:11 -0700272static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
Johannes Weinera804552b2014-01-29 14:05:39 -0800273{
Mel Gorman281e3722016-07-28 15:46:11 -0700274 unsigned long nr_pages = 0;
275 int z;
Johannes Weinera804552b2014-01-29 14:05:39 -0800276
Mel Gorman281e3722016-07-28 15:46:11 -0700277 for (z = 0; z < MAX_NR_ZONES; z++) {
278 struct zone *zone = pgdat->node_zones + z;
279
280 if (!populated_zone(zone))
281 continue;
282
283 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
284 }
285
Johannes Weinera8d01432016-01-14 15:20:15 -0800286 /*
287 * Pages reserved for the kernel should not be considered
288 * dirtyable, to prevent a situation where reclaim has to
289 * clean pages in order to balance the zones.
290 */
Mel Gorman281e3722016-07-28 15:46:11 -0700291 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
Johannes Weinera804552b2014-01-29 14:05:39 -0800292
Mel Gorman281e3722016-07-28 15:46:11 -0700293 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
294 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
Johannes Weinera804552b2014-01-29 14:05:39 -0800295
296 return nr_pages;
297}
298
Johannes Weiner1edf2232012-01-10 15:06:57 -0800299static unsigned long highmem_dirtyable_memory(unsigned long total)
300{
301#ifdef CONFIG_HIGHMEM
302 int node;
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700303 unsigned long x = 0;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700304 int i;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800305
306 for_each_node_state(node, N_HIGH_MEMORY) {
Mel Gorman281e3722016-07-28 15:46:11 -0700307 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
308 struct zone *z;
Minchan Kim9cb937e2016-07-28 15:47:08 -0700309 unsigned long nr_pages;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800310
Mel Gorman281e3722016-07-28 15:46:11 -0700311 if (!is_highmem_idx(i))
312 continue;
313
314 z = &NODE_DATA(node)->node_zones[i];
Minchan Kim9cb937e2016-07-28 15:47:08 -0700315 if (!populated_zone(z))
316 continue;
Mel Gorman281e3722016-07-28 15:46:11 -0700317
Minchan Kim9cb937e2016-07-28 15:47:08 -0700318 nr_pages = zone_page_state(z, NR_FREE_PAGES);
Mel Gorman281e3722016-07-28 15:46:11 -0700319 /* watch for underflows */
Minchan Kim9cb937e2016-07-28 15:47:08 -0700320 nr_pages -= min(nr_pages, high_wmark_pages(z));
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700321 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
322 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
323 x += nr_pages;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700324 }
Johannes Weiner1edf2232012-01-10 15:06:57 -0800325 }
Mel Gorman281e3722016-07-28 15:46:11 -0700326
Johannes Weiner1edf2232012-01-10 15:06:57 -0800327 /*
328 * Make sure that the number of highmem pages is never larger
329 * than the number of the total dirtyable memory. This can only
330 * occur in very strange VM situations but we want to make sure
331 * that this does not occur.
332 */
333 return min(x, total);
334#else
335 return 0;
336#endif
337}
338
339/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800340 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800341 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800342 * Return: the global number of pages potentially available for dirty
Johannes Weinerccafa282012-01-10 15:07:44 -0800343 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800344 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700345static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800346{
347 unsigned long x;
348
Michal Hockoc41f0122017-09-06 16:23:36 -0700349 x = global_zone_page_state(NR_FREE_PAGES);
Johannes Weinera8d01432016-01-14 15:20:15 -0800350 /*
351 * Pages reserved for the kernel should not be considered
352 * dirtyable, to prevent a situation where reclaim has to
353 * clean pages in order to balance the zones.
354 */
355 x -= min(x, totalreserve_pages);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800356
Mel Gorman599d0c92016-07-28 15:45:31 -0700357 x += global_node_page_state(NR_INACTIVE_FILE);
358 x += global_node_page_state(NR_ACTIVE_FILE);
Johannes Weinera804552b2014-01-29 14:05:39 -0800359
Johannes Weiner1edf2232012-01-10 15:06:57 -0800360 if (!vm_highmem_is_dirtyable)
361 x -= highmem_dirtyable_memory(x);
362
363 return x + 1; /* Ensure that we never return 0 */
364}
365
Tejun Heo9fc3a432015-05-22 18:23:30 -0400366/**
367 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
368 * @dtc: dirty_throttle_control of interest
Johannes Weinerccafa282012-01-10 15:07:44 -0800369 *
Tejun Heo9fc3a432015-05-22 18:23:30 -0400370 * Calculate @dtc->thresh and ->bg_thresh considering
371 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
372 * must ensure that @dtc->avail is set before calling this function. The
NeilBrowna37b0712020-06-01 21:48:18 -0700373 * dirty limits will be lifted by 1/4 for real-time tasks.
Johannes Weinerccafa282012-01-10 15:07:44 -0800374 */
Tejun Heo9fc3a432015-05-22 18:23:30 -0400375static void domain_dirty_limits(struct dirty_throttle_control *dtc)
376{
377 const unsigned long available_memory = dtc->avail;
378 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
379 unsigned long bytes = vm_dirty_bytes;
380 unsigned long bg_bytes = dirty_background_bytes;
Tejun Heo62a584f2016-05-27 14:34:46 -0400381 /* convert ratios to per-PAGE_SIZE for higher precision */
382 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
383 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400384 unsigned long thresh;
385 unsigned long bg_thresh;
386 struct task_struct *tsk;
387
388 /* gdtc is !NULL iff @dtc is for memcg domain */
389 if (gdtc) {
390 unsigned long global_avail = gdtc->avail;
391
392 /*
393 * The byte settings can't be applied directly to memcg
394 * domains. Convert them to ratios by scaling against
Tejun Heo62a584f2016-05-27 14:34:46 -0400395 * globally available memory. As the ratios are in
396 * per-PAGE_SIZE, they can be obtained by dividing bytes by
397 * number of pages.
Tejun Heo9fc3a432015-05-22 18:23:30 -0400398 */
399 if (bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400400 ratio = min(DIV_ROUND_UP(bytes, global_avail),
401 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400402 if (bg_bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400403 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
404 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400405 bytes = bg_bytes = 0;
406 }
407
408 if (bytes)
409 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
410 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400411 thresh = (ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400412
413 if (bg_bytes)
414 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
415 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400416 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400417
Michal Hocko90daf302017-11-29 16:10:58 -0800418 if (bg_thresh >= thresh)
Tejun Heo9fc3a432015-05-22 18:23:30 -0400419 bg_thresh = thresh / 2;
420 tsk = current;
NeilBrowna37b0712020-06-01 21:48:18 -0700421 if (rt_task(tsk)) {
NeilBrowna53eaff2016-05-20 16:58:53 -0700422 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
423 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400424 }
425 dtc->thresh = thresh;
426 dtc->bg_thresh = bg_thresh;
427
428 /* we should eventually report the domain in the TP */
429 if (!gdtc)
430 trace_global_dirty_state(bg_thresh, thresh);
431}
432
433/**
434 * global_dirty_limits - background-writeback and dirty-throttling thresholds
435 * @pbackground: out parameter for bg_thresh
436 * @pdirty: out parameter for thresh
437 *
438 * Calculate bg_thresh and thresh for global_wb_domain. See
439 * domain_dirty_limits() for details.
440 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800441void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
442{
Tejun Heo9fc3a432015-05-22 18:23:30 -0400443 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
Johannes Weinerccafa282012-01-10 15:07:44 -0800444
Tejun Heo9fc3a432015-05-22 18:23:30 -0400445 gdtc.avail = global_dirtyable_memory();
446 domain_dirty_limits(&gdtc);
Johannes Weinerccafa282012-01-10 15:07:44 -0800447
Tejun Heo9fc3a432015-05-22 18:23:30 -0400448 *pbackground = gdtc.bg_thresh;
449 *pdirty = gdtc.thresh;
Johannes Weinerccafa282012-01-10 15:07:44 -0800450}
451
Johannes Weinera756cf52012-01-10 15:07:49 -0800452/**
Mel Gorman281e3722016-07-28 15:46:11 -0700453 * node_dirty_limit - maximum number of dirty pages allowed in a node
454 * @pgdat: the node
Johannes Weinera756cf52012-01-10 15:07:49 -0800455 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800456 * Return: the maximum number of dirty pages allowed in a node, based
Mel Gorman281e3722016-07-28 15:46:11 -0700457 * on the node's dirtyable memory.
Johannes Weinera756cf52012-01-10 15:07:49 -0800458 */
Mel Gorman281e3722016-07-28 15:46:11 -0700459static unsigned long node_dirty_limit(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800460{
Mel Gorman281e3722016-07-28 15:46:11 -0700461 unsigned long node_memory = node_dirtyable_memory(pgdat);
Johannes Weinera756cf52012-01-10 15:07:49 -0800462 struct task_struct *tsk = current;
463 unsigned long dirty;
464
465 if (vm_dirty_bytes)
466 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
Mel Gorman281e3722016-07-28 15:46:11 -0700467 node_memory / global_dirtyable_memory();
Johannes Weinera756cf52012-01-10 15:07:49 -0800468 else
Mel Gorman281e3722016-07-28 15:46:11 -0700469 dirty = vm_dirty_ratio * node_memory / 100;
Johannes Weinera756cf52012-01-10 15:07:49 -0800470
NeilBrowna37b0712020-06-01 21:48:18 -0700471 if (rt_task(tsk))
Johannes Weinera756cf52012-01-10 15:07:49 -0800472 dirty += dirty / 4;
473
474 return dirty;
475}
476
477/**
Mel Gorman281e3722016-07-28 15:46:11 -0700478 * node_dirty_ok - tells whether a node is within its dirty limits
479 * @pgdat: the node to check
Johannes Weinera756cf52012-01-10 15:07:49 -0800480 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800481 * Return: %true when the dirty pages in @pgdat are within the node's
Johannes Weinera756cf52012-01-10 15:07:49 -0800482 * dirty limit, %false if the limit is exceeded.
483 */
Mel Gorman281e3722016-07-28 15:46:11 -0700484bool node_dirty_ok(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800485{
Mel Gorman281e3722016-07-28 15:46:11 -0700486 unsigned long limit = node_dirty_limit(pgdat);
487 unsigned long nr_pages = 0;
Johannes Weinera756cf52012-01-10 15:07:49 -0800488
Mel Gorman11fb9982016-07-28 15:46:20 -0700489 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman11fb9982016-07-28 15:46:20 -0700490 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
Mel Gorman281e3722016-07-28 15:46:11 -0700491
492 return nr_pages <= limit;
Johannes Weinera756cf52012-01-10 15:07:49 -0800493}
494
zhanglianjieaa779e52022-02-17 18:51:51 -0800495#ifdef CONFIG_SYSCTL
496static int dirty_background_ratio_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +0200497 void *buffer, size_t *lenp, loff_t *ppos)
David Rientjes2da02992009-01-06 14:39:31 -0800498{
499 int ret;
500
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700501 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800502 if (ret == 0 && write)
503 dirty_background_bytes = 0;
504 return ret;
505}
506
zhanglianjieaa779e52022-02-17 18:51:51 -0800507static int dirty_background_bytes_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +0200508 void *buffer, size_t *lenp, loff_t *ppos)
David Rientjes2da02992009-01-06 14:39:31 -0800509{
510 int ret;
511
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700512 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800513 if (ret == 0 && write)
514 dirty_background_ratio = 0;
515 return ret;
516}
517
zhanglianjieaa779e52022-02-17 18:51:51 -0800518static int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
Christoph Hellwig32927392020-04-24 08:43:38 +0200519 size_t *lenp, loff_t *ppos)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700520{
521 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800522 int ret;
523
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700524 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700525 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200526 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800527 vm_dirty_bytes = 0;
528 }
529 return ret;
530}
531
zhanglianjieaa779e52022-02-17 18:51:51 -0800532static int dirty_bytes_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +0200533 void *buffer, size_t *lenp, loff_t *ppos)
David Rientjes2da02992009-01-06 14:39:31 -0800534{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800535 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800536 int ret;
537
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700538 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800539 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200540 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800541 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700542 }
543 return ret;
544}
zhanglianjieaa779e52022-02-17 18:51:51 -0800545#endif
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700546
Jan Karaeb608e32012-05-24 18:59:11 +0200547static unsigned long wp_next_time(unsigned long cur_time)
548{
549 cur_time += VM_COMPLETIONS_PERIOD_LEN;
550 /* 0 has a special meaning... */
551 if (!cur_time)
552 return 1;
553 return cur_time;
554}
555
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400556static void wb_domain_writeout_add(struct wb_domain *dom,
Tejun Heoc7981432015-05-22 18:23:29 -0400557 struct fprop_local_percpu *completions,
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400558 unsigned int max_prop_frac, long nr)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700559{
Matthew Wilcox (Oracle)be5f1792021-03-20 16:34:54 -0400560 __fprop_add_percpu_max(&dom->completions, completions,
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400561 max_prop_frac, nr);
Jan Karaeb608e32012-05-24 18:59:11 +0200562 /* First event after period switching was turned off? */
Steven Rostedt (VMware)517663e2017-02-24 14:59:24 -0800563 if (unlikely(!dom->period_time)) {
Jan Karaeb608e32012-05-24 18:59:11 +0200564 /*
565 * We can race with other __bdi_writeout_inc calls here but
566 * it does not cause any harm since the resulting time when
567 * timer will fire and what is in writeout_period_time will be
568 * roughly the same.
569 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400570 dom->period_time = wp_next_time(jiffies);
571 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200572 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700573}
574
Tejun Heoc7981432015-05-22 18:23:29 -0400575/*
576 * Increment @wb's writeout completion count and the global writeout
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -0500577 * completion count. Called from __folio_end_writeback().
Tejun Heoc7981432015-05-22 18:23:29 -0400578 */
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400579static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
Tejun Heoc7981432015-05-22 18:23:29 -0400580{
Tejun Heo841710a2015-05-22 18:23:33 -0400581 struct wb_domain *cgdom;
582
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400583 wb_stat_mod(wb, WB_WRITTEN, nr);
584 wb_domain_writeout_add(&global_wb_domain, &wb->completions,
585 wb->bdi->max_prop_frac, nr);
Tejun Heo841710a2015-05-22 18:23:33 -0400586
587 cgdom = mem_cgroup_wb_domain(wb);
588 if (cgdom)
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400589 wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
590 wb->bdi->max_prop_frac, nr);
Tejun Heoc7981432015-05-22 18:23:29 -0400591}
592
Tejun Heo93f78d82015-05-22 17:13:27 -0400593void wb_writeout_inc(struct bdi_writeback *wb)
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700594{
595 unsigned long flags;
596
597 local_irq_save(flags);
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400598 __wb_writeout_add(wb, 1);
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700599 local_irq_restore(flags);
600}
Tejun Heo93f78d82015-05-22 17:13:27 -0400601EXPORT_SYMBOL_GPL(wb_writeout_inc);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700602
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700603/*
Jan Karaeb608e32012-05-24 18:59:11 +0200604 * On idle system, we can be called long after we scheduled because we use
605 * deferred timers so count with missed periods.
606 */
Kees Cook9823e512017-11-15 17:37:04 -0800607static void writeout_period(struct timer_list *t)
Jan Karaeb608e32012-05-24 18:59:11 +0200608{
Kees Cook9823e512017-11-15 17:37:04 -0800609 struct wb_domain *dom = from_timer(dom, t, period_timer);
Tejun Heo380c27c2015-05-22 18:23:21 -0400610 int miss_periods = (jiffies - dom->period_time) /
Jan Karaeb608e32012-05-24 18:59:11 +0200611 VM_COMPLETIONS_PERIOD_LEN;
612
Tejun Heo380c27c2015-05-22 18:23:21 -0400613 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
614 dom->period_time = wp_next_time(dom->period_time +
Jan Karaeb608e32012-05-24 18:59:11 +0200615 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
Tejun Heo380c27c2015-05-22 18:23:21 -0400616 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200617 } else {
618 /*
619 * Aging has zeroed all fractions. Stop wasting CPU on period
620 * updates.
621 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400622 dom->period_time = 0;
Jan Karaeb608e32012-05-24 18:59:11 +0200623 }
624}
625
Tejun Heo380c27c2015-05-22 18:23:21 -0400626int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
627{
628 memset(dom, 0, sizeof(*dom));
Tejun Heodcc25ae2015-05-22 18:23:22 -0400629
630 spin_lock_init(&dom->lock);
631
Kees Cook9823e512017-11-15 17:37:04 -0800632 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
Tejun Heodcc25ae2015-05-22 18:23:22 -0400633
634 dom->dirty_limit_tstamp = jiffies;
635
Tejun Heo380c27c2015-05-22 18:23:21 -0400636 return fprop_global_init(&dom->completions, gfp);
637}
638
Tejun Heo841710a2015-05-22 18:23:33 -0400639#ifdef CONFIG_CGROUP_WRITEBACK
640void wb_domain_exit(struct wb_domain *dom)
641{
642 del_timer_sync(&dom->period_timer);
643 fprop_global_destroy(&dom->completions);
644}
645#endif
646
Jan Karaeb608e32012-05-24 18:59:11 +0200647/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700648 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
649 * registered backing devices, which, for obvious reasons, can not
650 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700651 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700652static unsigned int bdi_min_ratio;
653
Stefan Roesch1bf27e92022-11-18 16:52:02 -0800654static int bdi_check_pages_limit(unsigned long pages)
655{
656 unsigned long max_dirty_pages = global_dirtyable_memory();
657
658 if (pages > max_dirty_pages)
659 return -EINVAL;
660
661 return 0;
662}
663
664static unsigned long bdi_ratio_from_pages(unsigned long pages)
665{
666 unsigned long background_thresh;
667 unsigned long dirty_thresh;
668 unsigned long ratio;
669
670 global_dirty_limits(&background_thresh, &dirty_thresh);
671 ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh);
672
673 return ratio;
674}
675
Stefan Roesch00df7d52022-11-18 16:52:00 -0800676static u64 bdi_get_bytes(unsigned int ratio)
677{
678 unsigned long background_thresh;
679 unsigned long dirty_thresh;
680 u64 bytes;
681
682 global_dirty_limits(&background_thresh, &dirty_thresh);
683 bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
684
685 return bytes;
686}
687
Stefan Roesch8021fb32022-11-18 16:52:06 -0800688static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700689{
Chen Wandun21f0dd82022-04-28 23:15:57 -0700690 unsigned int delta;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700691 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700692
Stefan Roesch2c44af4f2022-11-18 16:52:13 -0800693 if (min_ratio > 100 * BDI_RATIO_SCALE)
694 return -EINVAL;
Stefan Roeschae822912022-11-18 16:51:59 -0800695
Jens Axboecfc4ba52009-09-14 13:12:40 +0200696 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700697 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700698 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700699 } else {
Chen Wandun21f0dd82022-04-28 23:15:57 -0700700 if (min_ratio < bdi->min_ratio) {
701 delta = bdi->min_ratio - min_ratio;
702 bdi_min_ratio -= delta;
703 bdi->min_ratio = min_ratio;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700704 } else {
Chen Wandun21f0dd82022-04-28 23:15:57 -0700705 delta = min_ratio - bdi->min_ratio;
Stefan Roeschae822912022-11-18 16:51:59 -0800706 if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
Chen Wandun21f0dd82022-04-28 23:15:57 -0700707 bdi_min_ratio += delta;
708 bdi->min_ratio = min_ratio;
709 } else {
710 ret = -EINVAL;
711 }
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700712 }
713 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200714 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700715
716 return ret;
717}
718
Stefan Roeschefc3e6a2022-11-18 16:52:01 -0800719static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700720{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700721 int ret = 0;
722
Stefan Roesch4e230b42022-11-18 16:52:10 -0800723 if (max_ratio > 100 * BDI_RATIO_SCALE)
724 return -EINVAL;
725
Jens Axboecfc4ba52009-09-14 13:12:40 +0200726 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700727 if (bdi->min_ratio > max_ratio) {
728 ret = -EINVAL;
729 } else {
730 bdi->max_ratio = max_ratio;
Jingbo Xufa151a32023-12-19 22:25:08 +0800731 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) /
732 (100 * BDI_RATIO_SCALE);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700733 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200734 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700735
736 return ret;
737}
Stefan Roeschefc3e6a2022-11-18 16:52:01 -0800738
Stefan Roesch2c44af4f2022-11-18 16:52:13 -0800739int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio)
740{
741 return __bdi_set_min_ratio(bdi, min_ratio);
742}
743
Stefan Roesch4e230b42022-11-18 16:52:10 -0800744int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio)
745{
746 return __bdi_set_max_ratio(bdi, max_ratio);
747}
748
Stefan Roesch8021fb32022-11-18 16:52:06 -0800749int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
750{
751 return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE);
752}
753
Stefan Roeschefc3e6a2022-11-18 16:52:01 -0800754int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
755{
Stefan Roeschefc3e6a2022-11-18 16:52:01 -0800756 return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
757}
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700758EXPORT_SYMBOL(bdi_set_max_ratio);
759
Stefan Roesch712c00d2022-11-18 16:52:05 -0800760u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
761{
762 return bdi_get_bytes(bdi->min_ratio);
763}
764
Stefan Roesch803c9802022-11-18 16:52:07 -0800765int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes)
766{
767 int ret;
768 unsigned long pages = min_bytes >> PAGE_SHIFT;
769 unsigned long min_ratio;
770
771 ret = bdi_check_pages_limit(pages);
772 if (ret)
773 return ret;
774
775 min_ratio = bdi_ratio_from_pages(pages);
776 return __bdi_set_min_ratio(bdi, min_ratio);
777}
778
Stefan Roesch00df7d52022-11-18 16:52:00 -0800779u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
780{
781 return bdi_get_bytes(bdi->max_ratio);
782}
783
Stefan Roesch1bf27e92022-11-18 16:52:02 -0800784int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
785{
786 int ret;
787 unsigned long pages = max_bytes >> PAGE_SHIFT;
788 unsigned long max_ratio;
789
790 ret = bdi_check_pages_limit(pages);
791 if (ret)
792 return ret;
793
794 max_ratio = bdi_ratio_from_pages(pages);
795 return __bdi_set_max_ratio(bdi, max_ratio);
796}
797
Stefan Roesch8e9d5ea2022-11-18 16:51:56 -0800798int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
799{
800 if (strict_limit > 1)
801 return -EINVAL;
802
803 spin_lock_bh(&bdi_lock);
804 if (strict_limit)
805 bdi->capabilities |= BDI_CAP_STRICTLIMIT;
806 else
807 bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
808 spin_unlock_bh(&bdi_lock);
809
810 return 0;
811}
812
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600813static unsigned long dirty_freerun_ceiling(unsigned long thresh,
814 unsigned long bg_thresh)
815{
816 return (thresh + bg_thresh) / 2;
817}
818
Tejun Heoc7981432015-05-22 18:23:29 -0400819static unsigned long hard_dirty_limit(struct wb_domain *dom,
820 unsigned long thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -0600821{
Tejun Heodcc25ae2015-05-22 18:23:22 -0400822 return max(thresh, dom->dirty_limit);
Wu Fengguangffd1f602011-06-19 22:18:42 -0600823}
824
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400825/*
826 * Memory which can be further allocated to a memcg domain is capped by
827 * system-wide clean memory excluding the amount being used in the domain.
828 */
829static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
830 unsigned long filepages, unsigned long headroom)
Tejun Heoc2aa7232015-05-22 18:23:35 -0400831{
832 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400833 unsigned long clean = filepages - min(filepages, mdtc->dirty);
834 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
835 unsigned long other_clean = global_clean - min(global_clean, clean);
Tejun Heoc2aa7232015-05-22 18:23:35 -0400836
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400837 mdtc->avail = filepages + min(headroom, other_clean);
Christoph Lameter1b424462007-05-06 14:48:59 -0700838}
839
Wu Fengguang6f718652011-03-02 17:14:34 -0600840/**
Kemeng Shi13fc4412024-04-25 21:17:21 +0800841 * __wb_calc_thresh - @wb's share of dirty threshold
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400842 * @dtc: dirty_throttle_context of interest
Kemeng Shi13fc4412024-04-25 21:17:21 +0800843 * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc
Wu Fengguang1babe182010-08-11 14:17:40 -0700844 *
Kemeng Shi13fc4412024-04-25 21:17:21 +0800845 * Note that balance_dirty_pages() will only seriously take dirty throttling
846 * threshold as a hard limit when sleeping max_pause per page is not enough
847 * to keep the dirty pages under control. For example, when the device is
848 * completely stalled due to some error conditions, or when there are 1000
849 * dd tasks writing to a slow 10MB/s USB key.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600850 * In the other normal situations, it acts more gently by throttling the tasks
Tejun Heoa88a3412015-05-22 17:13:28 -0400851 * more (rather than completely block them) when the wb dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600852 *
853 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700854 * - starving fast devices
855 * - piling up dirty pages (that will take long time to sync) on slow devices
856 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400857 * The wb's share of dirty limit will be adapting to its throughput and
Wu Fengguang1babe182010-08-11 14:17:40 -0700858 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
Mike Rapoporta862f682019-03-05 15:48:42 -0800859 *
Kemeng Shi13fc4412024-04-25 21:17:21 +0800860 * Return: @wb's dirty limit in pages. For dirty throttling limit, the term
861 * "dirty" in the context of dirty balancing includes all PG_dirty and
862 * PG_writeback pages.
Wu Fengguang1babe182010-08-11 14:17:40 -0700863 */
Kemeng Shi13fc4412024-04-25 21:17:21 +0800864static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc,
865 unsigned long thresh)
Wu Fengguang16c40422010-08-11 14:17:39 -0700866{
Tejun Heoe9f07df2015-05-22 18:23:28 -0400867 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo0d960a32015-05-22 18:23:19 -0400868 u64 wb_thresh;
Wen Yangd3ac9462020-01-13 16:29:26 -0800869 unsigned long numerator, denominator;
Tejun Heo693108a2015-05-22 17:13:49 -0400870 unsigned long wb_min_ratio, wb_max_ratio;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700871
Wu Fengguang16c40422010-08-11 14:17:39 -0700872 /*
Kemeng Shi13fc4412024-04-25 21:17:21 +0800873 * Calculate this wb's share of the thresh ratio.
Wu Fengguang16c40422010-08-11 14:17:39 -0700874 */
Tejun Heoe9770b32015-05-22 18:23:27 -0400875 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
Tejun Heo380c27c2015-05-22 18:23:21 -0400876 &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700877
Stefan Roeschae822912022-11-18 16:51:59 -0800878 wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
Tejun Heo0d960a32015-05-22 18:23:19 -0400879 wb_thresh *= numerator;
Wen Yangd3ac9462020-01-13 16:29:26 -0800880 wb_thresh = div64_ul(wb_thresh, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700881
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400882 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
Wu Fengguang16c40422010-08-11 14:17:39 -0700883
Stefan Roeschae822912022-11-18 16:51:59 -0800884 wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
885 if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE))
886 wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
Wu Fengguang16c40422010-08-11 14:17:39 -0700887
Tejun Heo0d960a32015-05-22 18:23:19 -0400888 return wb_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889}
890
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400891unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
892{
Kemeng Shi13fc4412024-04-25 21:17:21 +0800893 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
894
895 return __wb_calc_thresh(&gdtc, thresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896}
897
Kemeng Shi4b5bbc32024-04-23 11:46:40 +0800898unsigned long cgwb_calc_thresh(struct bdi_writeback *wb)
899{
900 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
901 struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
902 unsigned long filepages = 0, headroom = 0, writeback = 0;
903
904 gdtc.avail = global_dirtyable_memory();
905 gdtc.dirty = global_node_page_state(NR_FILE_DIRTY) +
906 global_node_page_state(NR_WRITEBACK);
907
908 mem_cgroup_wb_stats(wb, &filepages, &headroom,
909 &mdtc.dirty, &writeback);
910 mdtc.dirty += writeback;
911 mdtc_calc_avail(&mdtc, filepages, headroom);
912 domain_dirty_limits(&mdtc);
913
Kemeng Shi13fc4412024-04-25 21:17:21 +0800914 return __wb_calc_thresh(&mdtc, mdtc.thresh);
Kemeng Shi4b5bbc32024-04-23 11:46:40 +0800915}
916
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600917/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700918 * setpoint - dirty 3
919 * f(dirty) := 1.0 + (----------------)
920 * limit - setpoint
921 *
922 * it's a 3rd order polynomial that subjects to
923 *
924 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
925 * (2) f(setpoint) = 1.0 => the balance point
926 * (3) f(limit) = 0 => the hard limit
927 * (4) df/dx <= 0 => negative feedback control
928 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
929 * => fast response on large errors; small oscillation near setpoint
930 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700931static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700932 unsigned long dirty,
933 unsigned long limit)
934{
935 long long pos_ratio;
936 long x;
937
Rik van Rield5c9fde2014-05-06 12:50:01 -0700938 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Tejun Heo464d1382015-04-21 16:49:13 -0400939 (limit - setpoint) | 1);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700940 pos_ratio = x;
941 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
942 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
943 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
944
945 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
946}
947
948/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600949 * Dirty position control.
950 *
951 * (o) global/bdi setpoints
952 *
Tejun Heode1fff32015-05-22 17:13:29 -0400953 * We want the dirty pages be balanced around the global/wb setpoints.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600954 * When the number of dirty pages is higher/lower than the setpoint, the
955 * dirty position control ratio (and hence task dirty ratelimit) will be
956 * decreased/increased to bring the dirty pages back to the setpoint.
957 *
958 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
959 *
960 * if (dirty < setpoint) scale up pos_ratio
961 * if (dirty > setpoint) scale down pos_ratio
962 *
Tejun Heode1fff32015-05-22 17:13:29 -0400963 * if (wb_dirty < wb_setpoint) scale up pos_ratio
964 * if (wb_dirty > wb_setpoint) scale down pos_ratio
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600965 *
966 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
967 *
968 * (o) global control line
969 *
970 * ^ pos_ratio
971 * |
972 * | |<===== global dirty control scope ======>|
Chi Wu03231552021-06-28 19:35:34 -0700973 * 2.0 * * * * * * *
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600974 * | .*
975 * | . *
976 * | . *
977 * | . *
978 * | . *
979 * | . *
980 * 1.0 ................................*
981 * | . . *
982 * | . . *
983 * | . . *
984 * | . . *
985 * | . . *
986 * 0 +------------.------------------.----------------------*------------->
987 * freerun^ setpoint^ limit^ dirty pages
988 *
Tejun Heode1fff32015-05-22 17:13:29 -0400989 * (o) wb control line
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600990 *
991 * ^ pos_ratio
992 * |
993 * | *
994 * | *
995 * | *
996 * | *
997 * | * |<=========== span ============>|
998 * 1.0 .......................*
999 * | . *
1000 * | . *
1001 * | . *
1002 * | . *
1003 * | . *
1004 * | . *
1005 * | . *
1006 * | . *
1007 * | . *
1008 * | . *
1009 * | . *
1010 * 1/4 ...............................................* * * * * * * * * * * *
1011 * | . .
1012 * | . .
1013 * | . .
1014 * 0 +----------------------.-------------------------------.------------->
Tejun Heode1fff32015-05-22 17:13:29 -04001015 * wb_setpoint^ x_intercept^
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001016 *
Tejun Heode1fff32015-05-22 17:13:29 -04001017 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001018 * be smoothly throttled down to normal if it starts high in situations like
1019 * - start writing to a slow SD card and a fast disk at the same time. The SD
Tejun Heode1fff32015-05-22 17:13:29 -04001020 * card's wb_dirty may rush to many times higher than wb_setpoint.
1021 * - the wb dirty thresh drops quickly due to change of JBOD workload
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001022 */
Tejun Heodaddfa32015-05-22 18:23:26 -04001023static void wb_position_ratio(struct dirty_throttle_control *dtc)
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001024{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001025 struct bdi_writeback *wb = dtc->wb;
Jan Kara20792eb2021-09-02 14:53:15 -07001026 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001027 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001028 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001029 unsigned long wb_thresh = dtc->wb_thresh;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001030 unsigned long x_intercept;
1031 unsigned long setpoint; /* dirty pages' target balance point */
Tejun Heode1fff32015-05-22 17:13:29 -04001032 unsigned long wb_setpoint;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001033 unsigned long span;
1034 long long pos_ratio; /* for scaling up/down the rate limit */
1035 long x;
1036
Tejun Heodaddfa32015-05-22 18:23:26 -04001037 dtc->pos_ratio = 0;
1038
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001039 if (unlikely(dtc->dirty >= limit))
Tejun Heodaddfa32015-05-22 18:23:26 -04001040 return;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001041
1042 /*
1043 * global setpoint
1044 *
Maxim Patlasov5a537482013-09-11 14:22:46 -07001045 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001046 */
1047 setpoint = (freerun + limit) / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001048 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001049
1050 /*
1051 * The strictlimit feature is a tool preventing mistrusted filesystems
1052 * from growing a large number of dirty pages before throttling. For
Tejun Heode1fff32015-05-22 17:13:29 -04001053 * such filesystems balance_dirty_pages always checks wb counters
1054 * against wb limits. Even if global "nr_dirty" is under "freerun".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001055 * This is especially important for fuse which sets bdi->max_ratio to
1056 * 1% by default. Without strictlimit feature, fuse writeback may
1057 * consume arbitrary amount of RAM because it is accounted in
1058 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
1059 *
Tejun Heoa88a3412015-05-22 17:13:28 -04001060 * Here, in wb_position_ratio(), we calculate pos_ratio based on
Tejun Heode1fff32015-05-22 17:13:29 -04001061 * two values: wb_dirty and wb_thresh. Let's consider an example:
Maxim Patlasov5a537482013-09-11 14:22:46 -07001062 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
1063 * limits are set by default to 10% and 20% (background and throttle).
Tejun Heode1fff32015-05-22 17:13:29 -04001064 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
Tejun Heo0d960a32015-05-22 18:23:19 -04001065 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
Tejun Heode1fff32015-05-22 17:13:29 -04001066 * about ~6K pages (as the average of background and throttle wb
Maxim Patlasov5a537482013-09-11 14:22:46 -07001067 * limits). The 3rd order polynomial will provide positive feedback if
Tejun Heode1fff32015-05-22 17:13:29 -04001068 * wb_dirty is under wb_setpoint and vice versa.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001069 *
1070 * Note, that we cannot use global counters in these calculations
Tejun Heode1fff32015-05-22 17:13:29 -04001071 * because we want to throttle process writing to a strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -07001072 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
1073 * in the example above).
1074 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001075 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heode1fff32015-05-22 17:13:29 -04001076 long long wb_pos_ratio;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001077
Tejun Heodaddfa32015-05-22 18:23:26 -04001078 if (dtc->wb_dirty < 8) {
1079 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
1080 2 << RATELIMIT_CALC_SHIFT);
1081 return;
1082 }
Maxim Patlasov5a537482013-09-11 14:22:46 -07001083
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001084 if (dtc->wb_dirty >= wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -04001085 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001086
Tejun Heo970fb012015-05-22 18:23:24 -04001087 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
1088 dtc->wb_bg_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001089
Tejun Heode1fff32015-05-22 17:13:29 -04001090 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -04001091 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001092
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001093 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
Tejun Heode1fff32015-05-22 17:13:29 -04001094 wb_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001095
1096 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001097 * Typically, for strictlimit case, wb_setpoint << setpoint
1098 * and pos_ratio >> wb_pos_ratio. In the other words global
Maxim Patlasov5a537482013-09-11 14:22:46 -07001099 * state ("dirty") is not limiting factor and we have to
Tejun Heode1fff32015-05-22 17:13:29 -04001100 * make decision based on wb counters. But there is an
Maxim Patlasov5a537482013-09-11 14:22:46 -07001101 * important case when global pos_ratio should get precedence:
1102 * global limits are exceeded (e.g. due to activities on other
Tejun Heode1fff32015-05-22 17:13:29 -04001103 * wb's) while given strictlimit wb is below limit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001104 *
Tejun Heode1fff32015-05-22 17:13:29 -04001105 * "pos_ratio * wb_pos_ratio" would work for the case above,
Maxim Patlasov5a537482013-09-11 14:22:46 -07001106 * but it would look too non-natural for the case of all
Tejun Heode1fff32015-05-22 17:13:29 -04001107 * activity in the system coming from a single strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -07001108 * with bdi->max_ratio == 100%.
1109 *
1110 * Note that min() below somewhat changes the dynamics of the
1111 * control system. Normally, pos_ratio value can be well over 3
Tejun Heode1fff32015-05-22 17:13:29 -04001112 * (when globally we are at freerun and wb is well below wb
Maxim Patlasov5a537482013-09-11 14:22:46 -07001113 * setpoint). Now the maximum pos_ratio in the same situation
1114 * is 2. We might want to tweak this if we observe the control
1115 * system is too slow to adapt.
1116 */
Tejun Heodaddfa32015-05-22 18:23:26 -04001117 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1118 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001119 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001120
1121 /*
1122 * We have computed basic pos_ratio above based on global situation. If
Tejun Heode1fff32015-05-22 17:13:29 -04001123 * the wb is over/under its share of dirty pages, we want to scale
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001124 * pos_ratio further down/up. That is done by the following mechanism.
1125 */
1126
1127 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001128 * wb setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001129 *
Tejun Heode1fff32015-05-22 17:13:29 -04001130 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001131 *
Tejun Heode1fff32015-05-22 17:13:29 -04001132 * x_intercept - wb_dirty
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001133 * := --------------------------
Tejun Heode1fff32015-05-22 17:13:29 -04001134 * x_intercept - wb_setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001135 *
Tejun Heode1fff32015-05-22 17:13:29 -04001136 * The main wb control line is a linear function that subjects to
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001137 *
Tejun Heode1fff32015-05-22 17:13:29 -04001138 * (1) f(wb_setpoint) = 1.0
1139 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1140 * or equally: x_intercept = wb_setpoint + 8 * write_bw
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001141 *
Tejun Heode1fff32015-05-22 17:13:29 -04001142 * For single wb case, the dirty pages are observed to fluctuate
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001143 * regularly within range
Tejun Heode1fff32015-05-22 17:13:29 -04001144 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001145 * for various filesystems, where (2) can yield in a reasonable 12.5%
1146 * fluctuation range for pos_ratio.
1147 *
Tejun Heode1fff32015-05-22 17:13:29 -04001148 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001149 * own size, so move the slope over accordingly and choose a slope that
Tejun Heode1fff32015-05-22 17:13:29 -04001150 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001151 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001152 if (unlikely(wb_thresh > dtc->thresh))
1153 wb_thresh = dtc->thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001154 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001155 * It's very possible that wb_thresh is close to 0 not because the
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001156 * device is slow, but that it has remained inactive for long time.
1157 * Honour such devices a reasonable good (hopefully IO efficient)
1158 * threshold, so that the occasional writes won't be blocked and active
1159 * writes can rampup the threshold quickly.
1160 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001161 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001162 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001163 * scale global setpoint to wb's:
1164 * wb_setpoint = setpoint * wb_thresh / thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001165 */
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001166 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
Tejun Heode1fff32015-05-22 17:13:29 -04001167 wb_setpoint = setpoint * (u64)x >> 16;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001168 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001169 * Use span=(8*write_bw) in single wb case as indicated by
1170 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001171 *
Tejun Heode1fff32015-05-22 17:13:29 -04001172 * wb_thresh thresh - wb_thresh
1173 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1174 * thresh thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001175 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001176 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
Tejun Heode1fff32015-05-22 17:13:29 -04001177 x_intercept = wb_setpoint + span;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001178
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001179 if (dtc->wb_dirty < x_intercept - span / 4) {
1180 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001181 (x_intercept - wb_setpoint) | 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001182 } else
1183 pos_ratio /= 4;
1184
Wu Fengguang8927f662011-08-04 22:16:46 -06001185 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001186 * wb reserve area, safeguard against dirty pool underrun and disk idle
Wu Fengguang8927f662011-08-04 22:16:46 -06001187 * It may push the desired control point of global dirty pages higher
1188 * than setpoint.
1189 */
Tejun Heode1fff32015-05-22 17:13:29 -04001190 x_intercept = wb_thresh / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001191 if (dtc->wb_dirty < x_intercept) {
1192 if (dtc->wb_dirty > x_intercept / 8)
1193 pos_ratio = div_u64(pos_ratio * x_intercept,
1194 dtc->wb_dirty);
Wu Fengguang50657fc2011-10-11 17:06:33 -06001195 else
Wu Fengguang8927f662011-08-04 22:16:46 -06001196 pos_ratio *= 8;
1197 }
1198
Tejun Heodaddfa32015-05-22 18:23:26 -04001199 dtc->pos_ratio = pos_ratio;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001200}
1201
Tejun Heoa88a3412015-05-22 17:13:28 -04001202static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1203 unsigned long elapsed,
1204 unsigned long written)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001205{
1206 const unsigned long period = roundup_pow_of_two(3 * HZ);
Tejun Heoa88a3412015-05-22 17:13:28 -04001207 unsigned long avg = wb->avg_write_bandwidth;
1208 unsigned long old = wb->write_bandwidth;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001209 u64 bw;
1210
1211 /*
1212 * bw = written * HZ / elapsed
1213 *
1214 * bw * elapsed + write_bandwidth * (period - elapsed)
1215 * write_bandwidth = ---------------------------------------------------
1216 * period
Tejun Heoc72efb62015-03-23 00:18:48 -04001217 *
Christoph Hellwiged2da922023-06-28 17:31:44 +02001218 * @written may have decreased due to folio_redirty_for_writepage().
Tejun Heoc72efb62015-03-23 00:18:48 -04001219 * Avoid underflowing @bw calculation.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001220 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001221 bw = written - min(written, wb->written_stamp);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001222 bw *= HZ;
1223 if (unlikely(elapsed > period)) {
Wen Yang0a5d1a72020-01-13 16:29:29 -08001224 bw = div64_ul(bw, elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001225 avg = bw;
1226 goto out;
1227 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001228 bw += (u64)wb->write_bandwidth * (period - elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001229 bw >>= ilog2(period);
1230
1231 /*
1232 * one more level of smoothing, for filtering out sudden spikes
1233 */
1234 if (avg > old && old >= (unsigned long)bw)
1235 avg -= (avg - old) >> 3;
1236
1237 if (avg < old && old <= (unsigned long)bw)
1238 avg += (old - avg) >> 3;
1239
1240out:
Tejun Heo95a46c62015-05-22 17:13:47 -04001241 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1242 avg = max(avg, 1LU);
1243 if (wb_has_dirty_io(wb)) {
1244 long delta = avg - wb->avg_write_bandwidth;
1245 WARN_ON_ONCE(atomic_long_add_return(delta,
1246 &wb->bdi->tot_write_bandwidth) <= 0);
1247 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001248 wb->write_bandwidth = bw;
Jan Kara20792eb2021-09-02 14:53:15 -07001249 WRITE_ONCE(wb->avg_write_bandwidth, avg);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001250}
1251
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001252static void update_dirty_limit(struct dirty_throttle_control *dtc)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001253{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001254 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001255 unsigned long thresh = dtc->thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04001256 unsigned long limit = dom->dirty_limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001257
1258 /*
1259 * Follow up in one step.
1260 */
1261 if (limit < thresh) {
1262 limit = thresh;
1263 goto update;
1264 }
1265
1266 /*
1267 * Follow down slowly. Use the higher one as the target, because thresh
1268 * may drop below dirty. This is exactly the reason to introduce
Tejun Heodcc25ae2015-05-22 18:23:22 -04001269 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
Wu Fengguangc42843f2011-03-02 15:54:09 -06001270 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001271 thresh = max(thresh, dtc->dirty);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001272 if (limit > thresh) {
1273 limit -= (limit - thresh) >> 5;
1274 goto update;
1275 }
1276 return;
1277update:
Tejun Heodcc25ae2015-05-22 18:23:22 -04001278 dom->dirty_limit = limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001279}
1280
Jan Kara42dd2352021-09-02 14:53:12 -07001281static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1282 unsigned long now)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001283{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001284 struct wb_domain *dom = dtc_dom(dtc);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001285
1286 /*
1287 * check locklessly first to optimize away locking for the most time
1288 */
Tejun Heodcc25ae2015-05-22 18:23:22 -04001289 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
Wu Fengguangc42843f2011-03-02 15:54:09 -06001290 return;
1291
Tejun Heodcc25ae2015-05-22 18:23:22 -04001292 spin_lock(&dom->lock);
1293 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001294 update_dirty_limit(dtc);
Tejun Heodcc25ae2015-05-22 18:23:22 -04001295 dom->dirty_limit_tstamp = now;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001296 }
Tejun Heodcc25ae2015-05-22 18:23:22 -04001297 spin_unlock(&dom->lock);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001298}
1299
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001300/*
Tejun Heode1fff32015-05-22 17:13:29 -04001301 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001302 *
Tejun Heode1fff32015-05-22 17:13:29 -04001303 * Normal wb tasks will be curbed at or below it in long term.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001304 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1305 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001306static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
Tejun Heoa88a3412015-05-22 17:13:28 -04001307 unsigned long dirtied,
1308 unsigned long elapsed)
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001309{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001310 struct bdi_writeback *wb = dtc->wb;
1311 unsigned long dirty = dtc->dirty;
1312 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001313 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Wu Fengguang73811312011-08-26 15:53:24 -06001314 unsigned long setpoint = (freerun + limit) / 2;
Tejun Heoa88a3412015-05-22 17:13:28 -04001315 unsigned long write_bw = wb->avg_write_bandwidth;
1316 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001317 unsigned long dirty_rate;
1318 unsigned long task_ratelimit;
1319 unsigned long balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001320 unsigned long step;
1321 unsigned long x;
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001322 unsigned long shift;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001323
1324 /*
1325 * The dirty rate will match the writeout rate in long term, except
1326 * when dirty pages are truncated by userspace or re-dirtied by FS.
1327 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001328 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001329
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001330 /*
1331 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1332 */
1333 task_ratelimit = (u64)dirty_ratelimit *
Tejun Heodaddfa32015-05-22 18:23:26 -04001334 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001335 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1336
1337 /*
1338 * A linear estimation of the "balanced" throttle rate. The theory is,
Tejun Heode1fff32015-05-22 17:13:29 -04001339 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001340 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1341 * formula will yield the balanced rate limit (write_bw / N).
1342 *
1343 * Note that the expanded form is not a pure rate feedback:
1344 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1345 * but also takes pos_ratio into account:
1346 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1347 *
1348 * (1) is not realistic because pos_ratio also takes part in balancing
1349 * the dirty rate. Consider the state
1350 * pos_ratio = 0.5 (3)
1351 * rate = 2 * (write_bw / N) (4)
1352 * If (1) is used, it will stuck in that state! Because each dd will
1353 * be throttled at
1354 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1355 * yielding
1356 * dirty_rate = N * task_ratelimit = write_bw (6)
1357 * put (6) into (1) we get
1358 * rate_(i+1) = rate_(i) (7)
1359 *
1360 * So we end up using (2) to always keep
1361 * rate_(i+1) ~= (write_bw / N) (8)
1362 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1363 * pos_ratio is able to drive itself to 1.0, which is not only where
1364 * the dirty count meet the setpoint, but also where the slope of
1365 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1366 */
1367 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1368 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001369 /*
1370 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1371 */
1372 if (unlikely(balanced_dirty_ratelimit > write_bw))
1373 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001374
Wu Fengguang73811312011-08-26 15:53:24 -06001375 /*
1376 * We could safely do this and return immediately:
1377 *
Tejun Heode1fff32015-05-22 17:13:29 -04001378 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001379 *
1380 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001381 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001382 * limit the step size.
1383 *
1384 * The below code essentially only uses the relative value of
1385 *
1386 * task_ratelimit - dirty_ratelimit
1387 * = (pos_ratio - 1) * dirty_ratelimit
1388 *
1389 * which reflects the direction and size of dirty position error.
1390 */
1391
1392 /*
1393 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1394 * task_ratelimit is on the same side of dirty_ratelimit, too.
1395 * For example, when
1396 * - dirty_ratelimit > balanced_dirty_ratelimit
1397 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1398 * lowering dirty_ratelimit will help meet both the position and rate
1399 * control targets. Otherwise, don't update dirty_ratelimit if it will
1400 * only help meet the rate target. After all, what the users ultimately
1401 * feel and care are stable dirty rate and small position error.
1402 *
1403 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001404 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001405 * keeps jumping around randomly and can even leap far away at times
1406 * due to the small 200ms estimation period of dirty_rate (we want to
1407 * keep that period small to reduce time lags).
1408 */
1409 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001410
1411 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001412 * For strictlimit case, calculations above were based on wb counters
Tejun Heoa88a3412015-05-22 17:13:28 -04001413 * and limits (starting from pos_ratio = wb_position_ratio() and up to
Maxim Patlasov5a537482013-09-11 14:22:46 -07001414 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
Tejun Heode1fff32015-05-22 17:13:29 -04001415 * Hence, to calculate "step" properly, we have to use wb_dirty as
1416 * "dirty" and wb_setpoint as "setpoint".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001417 *
Tejun Heode1fff32015-05-22 17:13:29 -04001418 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1419 * it's possible that wb_thresh is close to zero due to inactivity
Tejun Heo970fb012015-05-22 18:23:24 -04001420 * of backing device.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001421 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001422 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001423 dirty = dtc->wb_dirty;
1424 if (dtc->wb_dirty < 8)
1425 setpoint = dtc->wb_dirty + 1;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001426 else
Tejun Heo970fb012015-05-22 18:23:24 -04001427 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001428 }
1429
Wu Fengguang73811312011-08-26 15:53:24 -06001430 if (dirty < setpoint) {
Tejun Heoa88a3412015-05-22 17:13:28 -04001431 x = min3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001432 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001433 if (dirty_ratelimit < x)
1434 step = x - dirty_ratelimit;
1435 } else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001436 x = max3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001437 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001438 if (dirty_ratelimit > x)
1439 step = dirty_ratelimit - x;
1440 }
1441
1442 /*
1443 * Don't pursue 100% rate matching. It's impossible since the balanced
1444 * rate itself is constantly fluctuating. So decrease the track speed
1445 * when it gets close to the target. Helps eliminate pointless tremors.
1446 */
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001447 shift = dirty_ratelimit / (2 * step + 1);
1448 if (shift < BITS_PER_LONG)
1449 step = DIV_ROUND_UP(step >> shift, 8);
1450 else
1451 step = 0;
Wu Fengguang73811312011-08-26 15:53:24 -06001452
1453 if (dirty_ratelimit < balanced_dirty_ratelimit)
1454 dirty_ratelimit += step;
1455 else
1456 dirty_ratelimit -= step;
1457
Jan Kara20792eb2021-09-02 14:53:15 -07001458 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
Tejun Heoa88a3412015-05-22 17:13:28 -04001459 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001460
Tejun Heo5634cc22015-08-18 14:54:56 -07001461 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001462}
1463
Tejun Heoc2aa7232015-05-22 18:23:35 -04001464static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1465 struct dirty_throttle_control *mdtc,
Tejun Heo8a731792015-05-22 18:23:20 -04001466 bool update_ratelimit)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001467{
Tejun Heoc2aa7232015-05-22 18:23:35 -04001468 struct bdi_writeback *wb = gdtc->wb;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001469 unsigned long now = jiffies;
Jan Kara45a29662021-09-02 14:53:09 -07001470 unsigned long elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001471 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001472 unsigned long written;
1473
Jan Kara45a29662021-09-02 14:53:09 -07001474 spin_lock(&wb->list_lock);
Tejun Heo8a731792015-05-22 18:23:20 -04001475
Wu Fengguange98be2d2010-08-29 11:22:30 -06001476 /*
Jan Kara45a29662021-09-02 14:53:09 -07001477 * Lockless checks for elapsed time are racy and delayed update after
1478 * IO completion doesn't do it at all (to make sure written pages are
1479 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1480 * division errors.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001481 */
Jan Kara45a29662021-09-02 14:53:09 -07001482 elapsed = max(now - wb->bw_time_stamp, 1UL);
Tejun Heoa88a3412015-05-22 17:13:28 -04001483 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1484 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001485
Tejun Heo8a731792015-05-22 18:23:20 -04001486 if (update_ratelimit) {
Jan Kara42dd2352021-09-02 14:53:12 -07001487 domain_update_dirty_limit(gdtc, now);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001488 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1489
1490 /*
1491 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1492 * compiler has no way to figure that out. Help it.
1493 */
1494 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
Jan Kara42dd2352021-09-02 14:53:12 -07001495 domain_update_dirty_limit(mdtc, now);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001496 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1497 }
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001498 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001499 wb_update_write_bandwidth(wb, elapsed, written);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001500
Tejun Heoa88a3412015-05-22 17:13:28 -04001501 wb->dirtied_stamp = dirtied;
1502 wb->written_stamp = written;
Jan Kara20792eb2021-09-02 14:53:15 -07001503 WRITE_ONCE(wb->bw_time_stamp, now);
Jan Kara45a29662021-09-02 14:53:09 -07001504 spin_unlock(&wb->list_lock);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001505}
1506
Jan Kara45a29662021-09-02 14:53:09 -07001507void wb_update_bandwidth(struct bdi_writeback *wb)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001508{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001509 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1510
Jan Karafee468f2021-09-02 14:53:06 -07001511 __wb_update_bandwidth(&gdtc, NULL, false);
Jan Karafee468f2021-09-02 14:53:06 -07001512}
1513
1514/* Interval after which we consider wb idle and don't estimate bandwidth */
1515#define WB_BANDWIDTH_IDLE_JIF (HZ)
1516
1517static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1518{
1519 unsigned long now = jiffies;
1520 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1521
1522 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1523 !atomic_read(&wb->writeback_inodes)) {
1524 spin_lock(&wb->list_lock);
1525 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1526 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
Jan Kara20792eb2021-09-02 14:53:15 -07001527 WRITE_ONCE(wb->bw_time_stamp, now);
Jan Karafee468f2021-09-02 14:53:06 -07001528 spin_unlock(&wb->list_lock);
1529 }
Wu Fengguange98be2d2010-08-29 11:22:30 -06001530}
1531
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001533 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001534 * will look to see if it needs to start dirty throttling.
1535 *
1536 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
Michal Hockoc41f0122017-09-06 16:23:36 -07001537 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
Wu Fengguang9d823e82011-06-11 18:10:12 -06001538 * (the number of pages we may dirty without exceeding the dirty limits).
1539 */
1540static unsigned long dirty_poll_interval(unsigned long dirty,
1541 unsigned long thresh)
1542{
1543 if (thresh > dirty)
1544 return 1UL << (ilog2(thresh - dirty) >> 1);
1545
1546 return 1;
1547}
1548
Tejun Heoa88a3412015-05-22 17:13:28 -04001549static unsigned long wb_max_pause(struct bdi_writeback *wb,
Tejun Heode1fff32015-05-22 17:13:29 -04001550 unsigned long wb_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001551{
Jan Kara20792eb2021-09-02 14:53:15 -07001552 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
Fengguang Wue3b6c652013-10-16 13:47:03 -07001553 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001554
1555 /*
1556 * Limit pause time for small memory systems. If sleeping for too long
1557 * time, a small pool of dirty/writeback pages may go empty and disk go
1558 * idle.
1559 *
1560 * 8 serves as the safety ratio.
1561 */
Tejun Heode1fff32015-05-22 17:13:29 -04001562 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001563 t++;
1564
Fengguang Wue3b6c652013-10-16 13:47:03 -07001565 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001566}
1567
Tejun Heoa88a3412015-05-22 17:13:28 -04001568static long wb_min_pause(struct bdi_writeback *wb,
1569 long max_pause,
1570 unsigned long task_ratelimit,
1571 unsigned long dirty_ratelimit,
1572 int *nr_dirtied_pause)
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001573{
Jan Kara20792eb2021-09-02 14:53:15 -07001574 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1575 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001576 long t; /* target pause */
1577 long pause; /* estimated next pause */
1578 int pages; /* target nr_dirtied_pause */
1579
1580 /* target for 10ms pause on 1-dd case */
1581 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001582
1583 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001584 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1585 * overheads.
1586 *
1587 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001588 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001589 if (hi > lo)
1590 t += (hi - lo) * (10 * HZ) / 1024;
1591
1592 /*
1593 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1594 * on the much more stable dirty_ratelimit. However the next pause time
1595 * will be computed based on task_ratelimit and the two rate limits may
1596 * depart considerably at some time. Especially if task_ratelimit goes
1597 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1598 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1599 * result task_ratelimit won't be executed faithfully, which could
1600 * eventually bring down dirty_ratelimit.
1601 *
1602 * We apply two rules to fix it up:
1603 * 1) try to estimate the next pause time and if necessary, use a lower
1604 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1605 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1606 * 2) limit the target pause time to max_pause/2, so that the normal
1607 * small fluctuations of task_ratelimit won't trigger rule (1) and
1608 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1609 */
1610 t = min(t, 1 + max_pause / 2);
1611 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1612
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001613 /*
1614 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1615 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1616 * When the 16 consecutive reads are often interrupted by some dirty
1617 * throttling pause during the async writes, cfq will go into idles
1618 * (deadline is fine). So push nr_dirtied_pause as high as possible
1619 * until reaches DIRTY_POLL_THRESH=32 pages.
1620 */
1621 if (pages < DIRTY_POLL_THRESH) {
1622 t = max_pause;
1623 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1624 if (pages > DIRTY_POLL_THRESH) {
1625 pages = DIRTY_POLL_THRESH;
1626 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1627 }
1628 }
1629
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001630 pause = HZ * pages / (task_ratelimit + 1);
1631 if (pause > max_pause) {
1632 t = max_pause;
1633 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1634 }
1635
1636 *nr_dirtied_pause = pages;
1637 /*
1638 * The minimal pause time will normally be half the target pause time.
1639 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001640 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001641}
1642
Tejun Heo970fb012015-05-22 18:23:24 -04001643static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
Maxim Patlasov5a537482013-09-11 14:22:46 -07001644{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001645 struct bdi_writeback *wb = dtc->wb;
Tejun Heo93f78d82015-05-22 17:13:27 -04001646 unsigned long wb_reclaimable;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001647
1648 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001649 * wb_thresh is not treated as some limiting factor as
Maxim Patlasov5a537482013-09-11 14:22:46 -07001650 * dirty_thresh, due to reasons
Tejun Heode1fff32015-05-22 17:13:29 -04001651 * - in JBOD setup, wb_thresh can fluctuate a lot
Maxim Patlasov5a537482013-09-11 14:22:46 -07001652 * - in a system with HDD and USB key, the USB key may somehow
Tejun Heode1fff32015-05-22 17:13:29 -04001653 * go into state (wb_dirty >> wb_thresh) either because
1654 * wb_dirty starts high, or because wb_thresh drops low.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001655 * In this case we don't want to hard throttle the USB key
Tejun Heode1fff32015-05-22 17:13:29 -04001656 * dirtiers for 100 seconds until wb_dirty drops under
1657 * wb_thresh. Instead the auxiliary wb control line in
Tejun Heoa88a3412015-05-22 17:13:28 -04001658 * wb_position_ratio() will let the dirtier task progress
Tejun Heode1fff32015-05-22 17:13:29 -04001659 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001660 */
Kemeng Shi13fc4412024-04-25 21:17:21 +08001661 dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh);
Tejun Heo970fb012015-05-22 18:23:24 -04001662 dtc->wb_bg_thresh = dtc->thresh ?
Zach O'Keefe9319b642024-01-18 10:19:53 -08001663 div64_u64(dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001664
1665 /*
1666 * In order to avoid the stacked BDI deadlock we need
1667 * to ensure we accurately count the 'dirty' pages when
1668 * the threshold is low.
1669 *
1670 * Otherwise it would be possible to get thresh+n pages
1671 * reported dirty, even though there are thresh-m pages
1672 * actually dirty; with m+n sitting in the percpu
1673 * deltas.
1674 */
Wang Long2bce7742017-11-15 17:39:03 -08001675 if (dtc->wb_thresh < 2 * wb_stat_error()) {
Tejun Heo93f78d82015-05-22 17:13:27 -04001676 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001677 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001678 } else {
Tejun Heo93f78d82015-05-22 17:13:27 -04001679 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001680 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001681 }
1682}
1683
Wu Fengguang9d823e82011-06-11 18:10:12 -06001684/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 * balance_dirty_pages() must be called by processes which are generating dirty
1686 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001687 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001688 * If we're over `background_thresh' then the writeback threads are woken to
1689 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 */
Jan Karafe6c9c62022-06-23 10:51:46 -07001691static int balance_dirty_pages(struct bdi_writeback *wb,
1692 unsigned long pages_dirtied, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001694 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001695 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001696 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001697 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1698 &mdtc_stor : NULL;
1699 struct dirty_throttle_control *sdtc;
Kemeng Shi826881a2024-04-23 11:46:43 +08001700 unsigned long nr_dirty;
Wu Fengguang83712352011-06-11 19:25:42 -06001701 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001702 long pause;
1703 long max_pause;
1704 long min_pause;
1705 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001706 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001707 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001708 unsigned long dirty_ratelimit;
Tejun Heodfb8ae52015-05-22 17:13:40 -04001709 struct backing_dev_info *bdi = wb->bdi;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001710 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001711 unsigned long start_time = jiffies;
Jan Karafe6c9c62022-06-23 10:51:46 -07001712 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
1714 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001715 unsigned long now = jiffies;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001716 unsigned long dirty, thresh, bg_thresh;
Yang Shi50e55bf2015-11-20 15:57:10 -08001717 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1718 unsigned long m_thresh = 0;
1719 unsigned long m_bg_thresh = 0;
Wu Fengguang83712352011-06-11 19:25:42 -06001720
Kemeng Shi826881a2024-04-23 11:46:43 +08001721 nr_dirty = global_node_page_state(NR_FILE_DIRTY);
Tejun Heo9fc3a432015-05-22 18:23:30 -04001722 gdtc->avail = global_dirtyable_memory();
Kemeng Shi826881a2024-04-23 11:46:43 +08001723 gdtc->dirty = nr_dirty + global_node_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001724
Tejun Heo9fc3a432015-05-22 18:23:30 -04001725 domain_dirty_limits(gdtc);
Wu Fengguang16c40422010-08-11 14:17:39 -07001726
Maxim Patlasov5a537482013-09-11 14:22:46 -07001727 if (unlikely(strictlimit)) {
Tejun Heo970fb012015-05-22 18:23:24 -04001728 wb_dirty_limits(gdtc);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001729
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001730 dirty = gdtc->wb_dirty;
1731 thresh = gdtc->wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -04001732 bg_thresh = gdtc->wb_bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001733 } else {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001734 dirty = gdtc->dirty;
1735 thresh = gdtc->thresh;
1736 bg_thresh = gdtc->bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001737 }
1738
Tejun Heoc2aa7232015-05-22 18:23:35 -04001739 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001740 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001741
1742 /*
1743 * If @wb belongs to !root memcg, repeat the same
1744 * basic calculations for the memcg domain.
1745 */
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001746 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1747 &mdtc->dirty, &writeback);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001748 mdtc->dirty += writeback;
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001749 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001750
1751 domain_dirty_limits(mdtc);
1752
1753 if (unlikely(strictlimit)) {
1754 wb_dirty_limits(mdtc);
1755 m_dirty = mdtc->wb_dirty;
1756 m_thresh = mdtc->wb_thresh;
1757 m_bg_thresh = mdtc->wb_bg_thresh;
1758 } else {
1759 m_dirty = mdtc->dirty;
1760 m_thresh = mdtc->thresh;
1761 m_bg_thresh = mdtc->bg_thresh;
1762 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001763 }
1764
1765 /*
Jan Karaea6813b2022-06-23 10:51:44 -07001766 * In laptop mode, we wait until hitting the higher threshold
1767 * before starting background writeout, and then write out all
1768 * the way down to the lower threshold. So slow writers cause
1769 * minimal disk activity.
1770 *
1771 * In normal mode, we start background writeout at the lower
1772 * background_thresh, to keep the amount of dirty memory low.
1773 */
Kemeng Shi826881a2024-04-23 11:46:43 +08001774 if (!laptop_mode && nr_dirty > gdtc->bg_thresh &&
Jan Karaea6813b2022-06-23 10:51:44 -07001775 !writeback_in_progress(wb))
1776 wb_start_background_writeback(wb);
1777
1778 /*
Wu Fengguang16c40422010-08-11 14:17:39 -07001779 * Throttle it only when the background writeback cannot
1780 * catch-up. This avoids (excessively) small writeouts
Tejun Heode1fff32015-05-22 17:13:29 -04001781 * when the wb limits are ramping up in case of !strictlimit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001782 *
Tejun Heode1fff32015-05-22 17:13:29 -04001783 * In strictlimit case make decision based on the wb counters
1784 * and limits. Small writeouts when the wb limits are ramping
Maxim Patlasov5a537482013-09-11 14:22:46 -07001785 * up are the price we consciously pay for strictlimit-ing.
Tejun Heoc2aa7232015-05-22 18:23:35 -04001786 *
1787 * If memcg domain is in effect, @dirty should be under
1788 * both global and memcg freerun ceilings.
Wu Fengguang16c40422010-08-11 14:17:39 -07001789 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001790 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1791 (!mdtc ||
1792 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
NeilBrowna37b0712020-06-01 21:48:18 -07001793 unsigned long intv;
1794 unsigned long m_intv;
1795
1796free_running:
1797 intv = dirty_poll_interval(dirty, thresh);
1798 m_intv = ULONG_MAX;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001799
Wu Fengguang83712352011-06-11 19:25:42 -06001800 current->dirty_paused_when = now;
1801 current->nr_dirtied = 0;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001802 if (mdtc)
1803 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1804 current->nr_dirtied_pause = min(intv, m_intv);
Wu Fengguang16c40422010-08-11 14:17:39 -07001805 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001806 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001807
Jan Karaea6813b2022-06-23 10:51:44 -07001808 /* Start writeback even when in laptop mode */
Tejun Heobc058732015-05-22 17:13:53 -04001809 if (unlikely(!writeback_in_progress(wb)))
Tejun Heo9ecf48662015-05-22 17:13:54 -04001810 wb_start_background_writeback(wb);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001811
Tejun Heo97b27822019-08-26 09:06:56 -07001812 mem_cgroup_flush_foreign(wb);
1813
Tejun Heoc2aa7232015-05-22 18:23:35 -04001814 /*
1815 * Calculate global domain's pos_ratio and select the
1816 * global dtc by default.
1817 */
NeilBrowna37b0712020-06-01 21:48:18 -07001818 if (!strictlimit) {
Tejun Heo970fb012015-05-22 18:23:24 -04001819 wb_dirty_limits(gdtc);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001820
NeilBrowna37b0712020-06-01 21:48:18 -07001821 if ((current->flags & PF_LOCAL_THROTTLE) &&
1822 gdtc->wb_dirty <
1823 dirty_freerun_ceiling(gdtc->wb_thresh,
1824 gdtc->wb_bg_thresh))
1825 /*
1826 * LOCAL_THROTTLE tasks must not be throttled
1827 * when below the per-wb freerun ceiling.
1828 */
1829 goto free_running;
1830 }
1831
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001832 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1833 ((gdtc->dirty > gdtc->thresh) || strictlimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834
Tejun Heodaddfa32015-05-22 18:23:26 -04001835 wb_position_ratio(gdtc);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001836 sdtc = gdtc;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001837
Tejun Heoc2aa7232015-05-22 18:23:35 -04001838 if (mdtc) {
1839 /*
1840 * If memcg domain is in effect, calculate its
1841 * pos_ratio. @wb should satisfy constraints from
1842 * both global and memcg domains. Choose the one
1843 * w/ lower pos_ratio.
1844 */
NeilBrowna37b0712020-06-01 21:48:18 -07001845 if (!strictlimit) {
Tejun Heoc2aa7232015-05-22 18:23:35 -04001846 wb_dirty_limits(mdtc);
1847
NeilBrowna37b0712020-06-01 21:48:18 -07001848 if ((current->flags & PF_LOCAL_THROTTLE) &&
1849 mdtc->wb_dirty <
1850 dirty_freerun_ceiling(mdtc->wb_thresh,
1851 mdtc->wb_bg_thresh))
1852 /*
1853 * LOCAL_THROTTLE tasks must not be
1854 * throttled when below the per-wb
1855 * freerun ceiling.
1856 */
1857 goto free_running;
1858 }
Tejun Heoc2aa7232015-05-22 18:23:35 -04001859 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1860 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1861
1862 wb_position_ratio(mdtc);
1863 if (mdtc->pos_ratio < gdtc->pos_ratio)
1864 sdtc = mdtc;
1865 }
Tejun Heodaddfa32015-05-22 18:23:26 -04001866
Jan Karae92eebbb2022-06-23 10:51:45 -07001867 if (dirty_exceeded != wb->dirty_exceeded)
1868 wb->dirty_exceeded = dirty_exceeded;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001869
Jan Kara20792eb2021-09-02 14:53:15 -07001870 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
Jan Kara45a29662021-09-02 14:53:09 -07001871 BANDWIDTH_INTERVAL))
Jan Karafee468f2021-09-02 14:53:06 -07001872 __wb_update_bandwidth(gdtc, mdtc, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873
Tejun Heoc2aa7232015-05-22 18:23:35 -04001874 /* throttle according to the chosen dtc */
Jan Kara20792eb2021-09-02 14:53:15 -07001875 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001876 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001877 RATELIMIT_CALC_SHIFT;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001878 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
Tejun Heoa88a3412015-05-22 17:13:28 -04001879 min_pause = wb_min_pause(wb, max_pause,
1880 task_ratelimit, dirty_ratelimit,
1881 &nr_dirtied_pause);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001882
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001883 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001884 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001885 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001886 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887 }
Wu Fengguang83712352011-06-11 19:25:42 -06001888 period = HZ * pages_dirtied / task_ratelimit;
1889 pause = period;
1890 if (current->dirty_paused_when)
1891 pause -= now - current->dirty_paused_when;
1892 /*
1893 * For less than 1s think time (ext3/4 may block the dirtier
1894 * for up to 800ms from time to time on 1-HDD; so does xfs,
1895 * however at much less frequency), try to compensate it in
1896 * future periods by updating the virtual time; otherwise just
1897 * do a reset, as it may be a light dirtier.
1898 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001899 if (pause < min_pause) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001900 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001901 sdtc->thresh,
1902 sdtc->bg_thresh,
1903 sdtc->dirty,
1904 sdtc->wb_thresh,
1905 sdtc->wb_dirty,
Wu Fengguangece13ac32010-08-29 23:33:20 -06001906 dirty_ratelimit,
1907 task_ratelimit,
1908 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001909 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001910 min(pause, 0L),
Wu Fengguangece13ac32010-08-29 23:33:20 -06001911 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001912 if (pause < -HZ) {
1913 current->dirty_paused_when = now;
1914 current->nr_dirtied = 0;
1915 } else if (period) {
1916 current->dirty_paused_when += period;
1917 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001918 } else if (current->nr_dirtied_pause <= pages_dirtied)
1919 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001920 break;
1921 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001922 if (unlikely(pause > max_pause)) {
1923 /* for occasional dropped task_ratelimit */
1924 now += min(pause - max_pause, max_pause);
1925 pause = max_pause;
1926 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001927
1928pause:
Tejun Heo5634cc22015-08-18 14:54:56 -07001929 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001930 sdtc->thresh,
1931 sdtc->bg_thresh,
1932 sdtc->dirty,
1933 sdtc->wb_thresh,
1934 sdtc->wb_dirty,
Wu Fengguangece13ac32010-08-29 23:33:20 -06001935 dirty_ratelimit,
1936 task_ratelimit,
1937 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001938 period,
Wu Fengguangece13ac32010-08-29 23:33:20 -06001939 pause,
1940 start_time);
Jan Karafe6c9c62022-06-23 10:51:46 -07001941 if (flags & BDP_ASYNC) {
1942 ret = -EAGAIN;
1943 break;
1944 }
Jan Kara499d05e2011-11-16 19:34:48 +08001945 __set_current_state(TASK_KILLABLE);
Jan Karaf814bdd2024-01-23 18:58:26 +01001946 bdi->last_bdp_sleep = jiffies;
Wu Fengguangd25105e2009-10-09 12:40:42 +02001947 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001948
Wu Fengguang83712352011-06-11 19:25:42 -06001949 current->dirty_paused_when = now + pause;
1950 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001951 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001952
Wu Fengguangffd1f602011-06-19 22:18:42 -06001953 /*
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001954 * This is typically equal to (dirty < thresh) and can also
1955 * keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001956 */
Wu Fengguang1df647192011-11-13 19:47:32 -06001957 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001958 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001959
Wu Fengguangc5c63432011-12-02 10:21:33 -06001960 /*
Ingo Molnarf0953a12021-05-06 18:06:47 -07001961 * In the case of an unresponsive NFS server and the NFS dirty
Tejun Heode1fff32015-05-22 17:13:29 -04001962 * pages exceeds dirty_thresh, give the other good wb's a pipe
Wu Fengguangc5c63432011-12-02 10:21:33 -06001963 * to go through, so that tasks on them still remain responsive.
1964 *
Masahiro Yamada3f8b6fb2017-02-27 14:29:25 -08001965 * In theory 1 page is enough to keep the consumer-producer
Wu Fengguangc5c63432011-12-02 10:21:33 -06001966 * pipe going: the flusher cleans 1 page => the task dirties 1
Tejun Heode1fff32015-05-22 17:13:29 -04001967 * more page. However wb_dirty has accounting errors. So use
Tejun Heo93f78d82015-05-22 17:13:27 -04001968 * the larger and more IO friendly wb_stat_error.
Wu Fengguangc5c63432011-12-02 10:21:33 -06001969 */
Wang Long2bce7742017-11-15 17:39:03 -08001970 if (sdtc->wb_dirty <= wb_stat_error())
Wu Fengguangc5c63432011-12-02 10:21:33 -06001971 break;
1972
Jan Kara499d05e2011-11-16 19:34:48 +08001973 if (fatal_signal_pending(current))
1974 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 }
Jan Karafe6c9c62022-06-23 10:51:46 -07001976 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977}
1978
Wu Fengguang9d823e82011-06-11 18:10:12 -06001979static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001980
Wu Fengguang54848d72011-04-05 13:21:19 -06001981/*
1982 * Normal tasks are throttled by
1983 * loop {
1984 * dirty tsk->nr_dirtied_pause pages;
1985 * take a snap in balance_dirty_pages();
1986 * }
1987 * However there is a worst case. If every task exit immediately when dirtied
1988 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1989 * called to throttle the page dirties. The solution is to save the not yet
1990 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1991 * randomly into the running tasks. This works well for the above worst case,
1992 * as the new task will pick up and accumulate the old task's leaked dirty
1993 * count and eventually get throttled.
1994 */
1995DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1996
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997/**
Jan Karafe6c9c62022-06-23 10:51:46 -07001998 * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
1999 * @mapping: address_space which was dirtied.
2000 * @flags: BDP flags.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 *
2002 * Processes which are dirtying memory should call in here once for each page
2003 * which was newly dirtied. The function will periodically check the system's
2004 * dirty state and will initiate writeback if needed.
2005 *
Jan Karafe6c9c62022-06-23 10:51:46 -07002006 * See balance_dirty_pages_ratelimited() for details.
2007 *
2008 * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
2009 * indicate that memory is out of balance and the caller must wait
2010 * for I/O to complete. Otherwise, it will return 0 to indicate
2011 * that either memory was already in balance, or it was able to sleep
2012 * until the amount of dirty memory returned to balance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 */
Jan Karafe6c9c62022-06-23 10:51:46 -07002014int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
2015 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016{
Tejun Heodfb8ae52015-05-22 17:13:40 -04002017 struct inode *inode = mapping->host;
2018 struct backing_dev_info *bdi = inode_to_bdi(inode);
2019 struct bdi_writeback *wb = NULL;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002020 int ratelimit;
Jan Karafe6c9c62022-06-23 10:51:46 -07002021 int ret = 0;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002022 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002024 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
Jan Karafe6c9c62022-06-23 10:51:46 -07002025 return ret;
Wu Fengguang36715ce2011-06-11 17:53:57 -06002026
Tejun Heodfb8ae52015-05-22 17:13:40 -04002027 if (inode_cgwb_enabled(inode))
2028 wb = wb_get_create_current(bdi, GFP_KERNEL);
2029 if (!wb)
2030 wb = &bdi->wb;
2031
Wu Fengguang9d823e82011-06-11 18:10:12 -06002032 ratelimit = current->nr_dirtied_pause;
Tejun Heoa88a3412015-05-22 17:13:28 -04002033 if (wb->dirty_exceeded)
Wu Fengguang9d823e82011-06-11 18:10:12 -06002034 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035
Andrew Mortonfa5a7342006-03-24 03:18:10 -08002036 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06002037 /*
2038 * This prevents one CPU to accumulate too many dirtied pages without
2039 * calling into balance_dirty_pages(), which can happen when there are
2040 * 1000+ tasks, all of them start dirtying pages at exactly the same
2041 * time, hence all honoured too large initial task->nr_dirtied_pause.
2042 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07002043 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06002044 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08002045 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06002046 else if (unlikely(*p >= ratelimit_pages)) {
2047 *p = 0;
2048 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 }
Wu Fengguang54848d72011-04-05 13:21:19 -06002050 /*
2051 * Pick up the dirtied pages by the exited tasks. This avoids lots of
2052 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
2053 * the dirty throttling and livelock other long-run dirtiers.
2054 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07002055 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06002056 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08002057 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06002058 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
2059 *p -= nr_pages_dirtied;
2060 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08002062 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06002063
2064 if (unlikely(current->nr_dirtied >= ratelimit))
Jan Karafe6c9c62022-06-23 10:51:46 -07002065 ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
Tejun Heodfb8ae52015-05-22 17:13:40 -04002066
2067 wb_put(wb);
Jan Karafe6c9c62022-06-23 10:51:46 -07002068 return ret;
2069}
Stefan Roesch611df5d2022-09-12 12:27:41 -07002070EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
Jan Karafe6c9c62022-06-23 10:51:46 -07002071
2072/**
2073 * balance_dirty_pages_ratelimited - balance dirty memory state.
2074 * @mapping: address_space which was dirtied.
2075 *
2076 * Processes which are dirtying memory should call in here once for each page
2077 * which was newly dirtied. The function will periodically check the system's
2078 * dirty state and will initiate writeback if needed.
2079 *
2080 * Once we're over the dirty memory limit we decrease the ratelimiting
2081 * by a lot, to prevent individual processes from overshooting the limit
2082 * by (ratelimit_pages) each.
2083 */
2084void balance_dirty_pages_ratelimited(struct address_space *mapping)
2085{
2086 balance_dirty_pages_ratelimited_flags(mapping, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087}
Namjae Jeond0e1d662012-12-11 16:00:21 -08002088EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089
Tejun Heoaa661bb2015-05-22 18:23:31 -04002090/**
2091 * wb_over_bg_thresh - does @wb need to be written back?
2092 * @wb: bdi_writeback of interest
2093 *
2094 * Determines whether background writeback should keep writing @wb or it's
Mike Rapoporta862f682019-03-05 15:48:42 -08002095 * clean enough.
2096 *
2097 * Return: %true if writeback should continue.
Tejun Heoaa661bb2015-05-22 18:23:31 -04002098 */
2099bool wb_over_bg_thresh(struct bdi_writeback *wb)
2100{
Tejun Heo947e9762015-05-22 18:23:32 -04002101 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04002102 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo947e9762015-05-22 18:23:32 -04002103 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04002104 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
2105 &mdtc_stor : NULL;
Chi Wuab199392021-06-28 19:35:31 -07002106 unsigned long reclaimable;
2107 unsigned long thresh;
Tejun Heoaa661bb2015-05-22 18:23:31 -04002108
Tejun Heo947e9762015-05-22 18:23:32 -04002109 /*
2110 * Similar to balance_dirty_pages() but ignores pages being written
2111 * as we're trying to decide whether to put more under writeback.
2112 */
2113 gdtc->avail = global_dirtyable_memory();
NeilBrown8d928902020-06-01 21:48:21 -07002114 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
Tejun Heo947e9762015-05-22 18:23:32 -04002115 domain_dirty_limits(gdtc);
Tejun Heoaa661bb2015-05-22 18:23:31 -04002116
Tejun Heo947e9762015-05-22 18:23:32 -04002117 if (gdtc->dirty > gdtc->bg_thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04002118 return true;
2119
Kemeng Shi3b3e4122024-04-25 21:17:23 +08002120 thresh = __wb_calc_thresh(gdtc, gdtc->bg_thresh);
Chi Wuab199392021-06-28 19:35:31 -07002121 if (thresh < 2 * wb_stat_error())
2122 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2123 else
2124 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2125
2126 if (reclaimable > thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04002127 return true;
2128
Tejun Heoc2aa7232015-05-22 18:23:35 -04002129 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04002130 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04002131
Tejun Heoc5edf9c2015-09-29 13:04:26 -04002132 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
2133 &writeback);
2134 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04002135 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
2136
2137 if (mdtc->dirty > mdtc->bg_thresh)
2138 return true;
2139
Kemeng Shifabd2e42024-04-25 21:17:22 +08002140 thresh = __wb_calc_thresh(mdtc, mdtc->bg_thresh);
Chi Wuab199392021-06-28 19:35:31 -07002141 if (thresh < 2 * wb_stat_error())
2142 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
2143 else
2144 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
2145
2146 if (reclaimable > thresh)
Tejun Heoc2aa7232015-05-22 18:23:35 -04002147 return true;
2148 }
2149
Tejun Heoaa661bb2015-05-22 18:23:31 -04002150 return false;
2151}
2152
zhanglianjieaa779e52022-02-17 18:51:51 -08002153#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2156 */
zhanglianjieaa779e52022-02-17 18:51:51 -08002157static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02002158 void *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159{
Yafang Shao94af5842017-10-10 05:44:13 +08002160 unsigned int old_interval = dirty_writeback_interval;
2161 int ret;
2162
2163 ret = proc_dointvec(table, write, buffer, length, ppos);
Yafang Shao515c24c2017-10-14 16:38:27 +08002164
2165 /*
2166 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2167 * and a different non-zero value will wakeup the writeback threads.
2168 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2169 * iterate over all bdis and wbs.
2170 * The reason we do this is to make the change take effect immediately.
2171 */
2172 if (!ret && write && dirty_writeback_interval &&
2173 dirty_writeback_interval != old_interval)
Yafang Shao94af5842017-10-10 05:44:13 +08002174 wakeup_flusher_threads(WB_REASON_PERIODIC);
2175
2176 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177}
zhanglianjieaa779e52022-02-17 18:51:51 -08002178#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179
Kees Cookbca237a2017-08-28 15:03:41 -07002180void laptop_mode_timer_fn(struct timer_list *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181{
Kees Cookbca237a2017-08-28 15:03:41 -07002182 struct backing_dev_info *backing_dev_info =
2183 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184
Kees Cookbca237a2017-08-28 15:03:41 -07002185 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186}
2187
2188/*
2189 * We've spun up the disk and we're in laptop mode: schedule writeback
2190 * of all dirty data a few seconds from now. If the flush is already scheduled
2191 * then push it back - the user is still using the disk.
2192 */
Matthew Garrett31373d02010-04-06 14:25:14 +02002193void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194{
Matthew Garrett31373d02010-04-06 14:25:14 +02002195 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196}
2197
2198/*
2199 * We're in laptop mode and we've just synced. The sync's writes will have
2200 * caused another writeback to be scheduled by laptop_io_completion.
2201 * Nothing needs to be written back anymore, so we unschedule the writeback.
2202 */
2203void laptop_sync_completion(void)
2204{
Matthew Garrett31373d02010-04-06 14:25:14 +02002205 struct backing_dev_info *bdi;
2206
2207 rcu_read_lock();
2208
2209 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2210 del_timer(&bdi->laptop_mode_wb_timer);
2211
2212 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213}
2214
2215/*
2216 * If ratelimit_pages is too high then we can get into dirty-data overload
2217 * if a large number of processes all perform writes at the same time.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 *
2219 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2220 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06002221 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 */
2223
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002224void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225{
Tejun Heodcc25ae2015-05-22 18:23:22 -04002226 struct wb_domain *dom = &global_wb_domain;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002227 unsigned long background_thresh;
2228 unsigned long dirty_thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04002229
Wu Fengguang9d823e82011-06-11 18:10:12 -06002230 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heodcc25ae2015-05-22 18:23:22 -04002231 dom->dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002232 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 if (ratelimit_pages < 16)
2234 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235}
2236
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002237static int page_writeback_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238{
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002239 writeback_set_ratelimit();
2240 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241}
2242
zhanglianjieaa779e52022-02-17 18:51:51 -08002243#ifdef CONFIG_SYSCTL
Luis Chamberlain3c6a4cb2022-04-15 15:08:02 -07002244
2245/* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2246static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2247
zhanglianjieaa779e52022-02-17 18:51:51 -08002248static struct ctl_table vm_page_writeback_sysctls[] = {
2249 {
2250 .procname = "dirty_background_ratio",
2251 .data = &dirty_background_ratio,
2252 .maxlen = sizeof(dirty_background_ratio),
2253 .mode = 0644,
2254 .proc_handler = dirty_background_ratio_handler,
2255 .extra1 = SYSCTL_ZERO,
2256 .extra2 = SYSCTL_ONE_HUNDRED,
2257 },
2258 {
2259 .procname = "dirty_background_bytes",
2260 .data = &dirty_background_bytes,
2261 .maxlen = sizeof(dirty_background_bytes),
2262 .mode = 0644,
2263 .proc_handler = dirty_background_bytes_handler,
2264 .extra1 = SYSCTL_LONG_ONE,
2265 },
2266 {
2267 .procname = "dirty_ratio",
2268 .data = &vm_dirty_ratio,
2269 .maxlen = sizeof(vm_dirty_ratio),
2270 .mode = 0644,
2271 .proc_handler = dirty_ratio_handler,
2272 .extra1 = SYSCTL_ZERO,
2273 .extra2 = SYSCTL_ONE_HUNDRED,
2274 },
2275 {
2276 .procname = "dirty_bytes",
2277 .data = &vm_dirty_bytes,
2278 .maxlen = sizeof(vm_dirty_bytes),
2279 .mode = 0644,
2280 .proc_handler = dirty_bytes_handler,
2281 .extra1 = (void *)&dirty_bytes_min,
2282 },
2283 {
2284 .procname = "dirty_writeback_centisecs",
2285 .data = &dirty_writeback_interval,
2286 .maxlen = sizeof(dirty_writeback_interval),
2287 .mode = 0644,
2288 .proc_handler = dirty_writeback_centisecs_handler,
2289 },
2290 {
2291 .procname = "dirty_expire_centisecs",
2292 .data = &dirty_expire_interval,
2293 .maxlen = sizeof(dirty_expire_interval),
2294 .mode = 0644,
2295 .proc_handler = proc_dointvec_minmax,
2296 .extra1 = SYSCTL_ZERO,
2297 },
2298#ifdef CONFIG_HIGHMEM
2299 {
2300 .procname = "highmem_is_dirtyable",
2301 .data = &vm_highmem_is_dirtyable,
2302 .maxlen = sizeof(vm_highmem_is_dirtyable),
2303 .mode = 0644,
2304 .proc_handler = proc_dointvec_minmax,
2305 .extra1 = SYSCTL_ZERO,
2306 .extra2 = SYSCTL_ONE,
2307 },
2308#endif
2309 {
2310 .procname = "laptop_mode",
2311 .data = &laptop_mode,
2312 .maxlen = sizeof(laptop_mode),
2313 .mode = 0644,
2314 .proc_handler = proc_dointvec_jiffies,
2315 },
zhanglianjieaa779e52022-02-17 18:51:51 -08002316};
2317#endif
2318
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002320 * Called early on to tune the page writeback dirty limits.
2321 *
2322 * We used to scale dirty pages according to how total memory
David Hildenbrand0a18e602020-08-06 23:25:27 -07002323 * related to pages that could be allocated for buffers.
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002324 *
2325 * However, that was when we used "dirty_ratio" to scale with
2326 * all memory, and we don't do that any more. "dirty_ratio"
David Hildenbrand0a18e602020-08-06 23:25:27 -07002327 * is now applied to total non-HIGHPAGE memory, and as such we can't
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002328 * get into the old insane situation any more where we had
2329 * large amounts of dirty pages compared to a small amount of
2330 * non-HIGHMEM memory.
2331 *
2332 * But we might still want to scale the dirty_ratio by how
2333 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 */
2335void __init page_writeback_init(void)
2336{
Rabin Vincenta50fcb52015-08-06 15:47:14 -07002337 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2338
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002339 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2340 page_writeback_cpu_online, NULL);
2341 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2342 page_writeback_cpu_online);
zhanglianjieaa779e52022-02-17 18:51:51 -08002343#ifdef CONFIG_SYSCTL
2344 register_sysctl_init("vm", vm_page_writeback_sysctls);
2345#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346}
2347
David Howells811d7362006-08-29 19:06:09 +01002348/**
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002349 * tag_pages_for_writeback - tag pages to be written by writeback
Jan Karaf446daae2010-08-09 17:19:12 -07002350 * @mapping: address space structure to write
2351 * @start: starting page index
2352 * @end: ending page index (inclusive)
2353 *
2354 * This function scans the page range from @start to @end (inclusive) and tags
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002355 * all pages that have DIRTY tag set with a special TOWRITE tag. The caller
2356 * can then use the TOWRITE tag to identify pages eligible for writeback.
2357 * This mechanism is used to avoid livelocking of writeback by a process
2358 * steadily creating new dirty pages in the file (thus it is important for this
2359 * function to be quick so that it can tag pages faster than a dirtying process
2360 * can create them).
Jan Karaf446daae2010-08-09 17:19:12 -07002361 */
Jan Karaf446daae2010-08-09 17:19:12 -07002362void tag_pages_for_writeback(struct address_space *mapping,
2363 pgoff_t start, pgoff_t end)
2364{
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002365 XA_STATE(xas, &mapping->i_pages, start);
2366 unsigned int tagged = 0;
2367 void *page;
Jan Karaf446daae2010-08-09 17:19:12 -07002368
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002369 xas_lock_irq(&xas);
2370 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2371 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2372 if (++tagged % XA_CHECK_SCHED)
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002373 continue;
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002374
2375 xas_pause(&xas);
2376 xas_unlock_irq(&xas);
Jan Karaf446daae2010-08-09 17:19:12 -07002377 cond_resched();
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002378 xas_lock_irq(&xas);
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002379 }
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002380 xas_unlock_irq(&xas);
Jan Karaf446daae2010-08-09 17:19:12 -07002381}
2382EXPORT_SYMBOL(tag_pages_for_writeback);
2383
Matthew Wilcox (Oracle)b1793922024-02-15 07:36:42 +01002384static bool folio_prepare_writeback(struct address_space *mapping,
2385 struct writeback_control *wbc, struct folio *folio)
2386{
2387 /*
2388 * Folio truncated or invalidated. We can freely skip it then,
2389 * even for data integrity operations: the folio has disappeared
2390 * concurrently, so there could be no real expectation of this
2391 * data integrity operation even if there is now a new, dirty
2392 * folio at the same pagecache index.
2393 */
2394 if (unlikely(folio->mapping != mapping))
2395 return false;
2396
2397 /*
2398 * Did somebody else write it for us?
2399 */
2400 if (!folio_test_dirty(folio))
2401 return false;
2402
2403 if (folio_test_writeback(folio)) {
2404 if (wbc->sync_mode == WB_SYNC_NONE)
2405 return false;
2406 folio_wait_writeback(folio);
2407 }
2408 BUG_ON(folio_test_writeback(folio));
2409
2410 if (!folio_clear_dirty_for_io(folio))
2411 return false;
2412
2413 return true;
2414}
2415
Matthew Wilcox (Oracle)751e0d52024-02-15 07:36:43 +01002416static xa_mark_t wbc_to_tag(struct writeback_control *wbc)
2417{
2418 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2419 return PAGECACHE_TAG_TOWRITE;
2420 return PAGECACHE_TAG_DIRTY;
2421}
2422
2423static pgoff_t wbc_end(struct writeback_control *wbc)
2424{
2425 if (wbc->range_cyclic)
2426 return -1;
2427 return wbc->range_end >> PAGE_SHIFT;
2428}
2429
Matthew Wilcox (Oracle)e6d0ab82024-02-15 07:36:46 +01002430static struct folio *writeback_get_folio(struct address_space *mapping,
Matthew Wilcox (Oracle)751e0d52024-02-15 07:36:43 +01002431 struct writeback_control *wbc)
2432{
Matthew Wilcox (Oracle)e6d0ab82024-02-15 07:36:46 +01002433 struct folio *folio;
2434
Matthew Wilcox (Oracle)a2cbc132024-02-15 07:36:47 +01002435retry:
Matthew Wilcox (Oracle)e6d0ab82024-02-15 07:36:46 +01002436 folio = folio_batch_next(&wbc->fbatch);
2437 if (!folio) {
2438 folio_batch_release(&wbc->fbatch);
2439 cond_resched();
2440 filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc),
2441 wbc_to_tag(wbc), &wbc->fbatch);
2442 folio = folio_batch_next(&wbc->fbatch);
Matthew Wilcox (Oracle)a2cbc132024-02-15 07:36:47 +01002443 if (!folio)
2444 return NULL;
Matthew Wilcox (Oracle)e6d0ab82024-02-15 07:36:46 +01002445 }
2446
Matthew Wilcox (Oracle)a2cbc132024-02-15 07:36:47 +01002447 folio_lock(folio);
2448 if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) {
2449 folio_unlock(folio);
2450 goto retry;
2451 }
2452
2453 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
Matthew Wilcox (Oracle)e6d0ab82024-02-15 07:36:46 +01002454 return folio;
Matthew Wilcox (Oracle)751e0d52024-02-15 07:36:43 +01002455}
2456
Jan Karaf446daae2010-08-09 17:19:12 -07002457/**
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002458 * writeback_iter - iterate folio of a mapping for writeback
David Howells811d7362006-08-29 19:06:09 +01002459 * @mapping: address space structure to write
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002460 * @wbc: writeback context
2461 * @folio: previously iterated folio (%NULL to start)
2462 * @error: in-out pointer for writeback errors (see below)
David Howells811d7362006-08-29 19:06:09 +01002463 *
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002464 * This function returns the next folio for the writeback operation described by
2465 * @wbc on @mapping and should be called in a while loop in the ->writepages
2466 * implementation.
Jan Karaf446daae2010-08-09 17:19:12 -07002467 *
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002468 * To start the writeback operation, %NULL is passed in the @folio argument, and
2469 * for every subsequent iteration the folio returned previously should be passed
2470 * back in.
Dave Chinner64081362018-10-26 15:09:45 -07002471 *
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002472 * If there was an error in the per-folio writeback inside the writeback_iter()
2473 * loop, @error should be set to the error value.
Mike Rapoporta862f682019-03-05 15:48:42 -08002474 *
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002475 * Once the writeback described in @wbc has finished, this function will return
2476 * %NULL and if there was an error in any iteration restore it to @error.
2477 *
2478 * Note: callers should not manually break out of the loop using break or goto
2479 * but must keep calling writeback_iter() until it returns %NULL.
2480 *
2481 * Return: the folio to write or %NULL if the loop is done.
David Howells811d7362006-08-29 19:06:09 +01002482 */
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002483struct folio *writeback_iter(struct address_space *mapping,
2484 struct writeback_control *wbc, struct folio *folio, int *error)
David Howells811d7362006-08-29 19:06:09 +01002485{
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002486 if (!folio) {
2487 folio_batch_init(&wbc->fbatch);
2488 wbc->saved_err = *error = 0;
David Howells811d7362006-08-29 19:06:09 +01002489
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002490 /*
2491 * For range cyclic writeback we remember where we stopped so
2492 * that we can continue where we stopped.
2493 *
2494 * For non-cyclic writeback we always start at the beginning of
2495 * the passed in range.
2496 */
2497 if (wbc->range_cyclic)
2498 wbc->index = mapping->writeback_index;
2499 else
2500 wbc->index = wbc->range_start >> PAGE_SHIFT;
2501
2502 /*
2503 * To avoid livelocks when other processes dirty new pages, we
2504 * first tag pages which should be written back and only then
2505 * start writing them.
2506 *
2507 * For data-integrity writeback we have to be careful so that we
2508 * do not miss some pages (e.g., because some other process has
2509 * cleared the TOWRITE tag we set). The rule we follow is that
2510 * TOWRITE tag can be cleared only by the process clearing the
2511 * DIRTY tag (and submitting the page for I/O).
2512 */
2513 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2514 tag_pages_for_writeback(mapping, wbc->index,
2515 wbc_end(wbc));
David Howells811d7362006-08-29 19:06:09 +01002516 } else {
Matthew Wilcox (Oracle)807d1fe2024-02-15 07:36:44 +01002517 wbc->nr_to_write -= folio_nr_pages(folio);
Christoph Hellwigf946e0d2024-02-15 07:36:41 +01002518
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002519 WARN_ON_ONCE(*error > 0);
David Howells811d7362006-08-29 19:06:09 +01002520
Matthew Wilcox (Oracle)807d1fe2024-02-15 07:36:44 +01002521 /*
2522 * For integrity writeback we have to keep going until we have
2523 * written all the folios we tagged for writeback above, even if
2524 * we run past wbc->nr_to_write or encounter errors.
2525 * We stash away the first error we encounter in wbc->saved_err
2526 * so that it can be retrieved when we're done. This is because
2527 * the file system may still have state to clear for each folio.
2528 *
2529 * For background writeback we exit as soon as we run past
2530 * wbc->nr_to_write or encounter the first error.
2531 */
2532 if (wbc->sync_mode == WB_SYNC_ALL) {
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002533 if (*error && !wbc->saved_err)
2534 wbc->saved_err = *error;
Matthew Wilcox (Oracle)807d1fe2024-02-15 07:36:44 +01002535 } else {
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002536 if (*error || wbc->nr_to_write <= 0)
Matthew Wilcox (Oracle)807d1fe2024-02-15 07:36:44 +01002537 goto done;
David Howells811d7362006-08-29 19:06:09 +01002538 }
David Howells811d7362006-08-29 19:06:09 +01002539 }
Dave Chinner64081362018-10-26 15:09:45 -07002540
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002541 folio = writeback_get_folio(mapping, wbc);
2542 if (!folio) {
2543 /*
2544 * To avoid deadlocks between range_cyclic writeback and callers
2545 * that hold pages in PageWriteback to aggregate I/O until
2546 * the writeback iteration finishes, we do not loop back to the
2547 * start of the file. Doing so causes a page lock/page
2548 * writeback access order inversion - we should only ever lock
2549 * multiple pages in ascending page->index order, and looping
2550 * back to the start of the file violates that rule and causes
2551 * deadlocks.
2552 */
2553 if (wbc->range_cyclic)
2554 mapping->writeback_index = 0;
2555
2556 /*
2557 * Return the first error we encountered (if there was any) to
2558 * the caller.
2559 */
2560 *error = wbc->saved_err;
2561 }
2562 return folio;
Christoph Hellwigf946e0d2024-02-15 07:36:41 +01002563
2564done:
2565 if (wbc->range_cyclic)
2566 mapping->writeback_index = folio->index + folio_nr_pages(folio);
Matthew Wilcox (Oracle)751e0d52024-02-15 07:36:43 +01002567 folio_batch_release(&wbc->fbatch);
Christoph Hellwigcdc150b2024-02-15 07:36:48 +01002568 return NULL;
2569}
2570
2571/**
2572 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2573 * @mapping: address space structure to write
2574 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2575 * @writepage: function called for each page
2576 * @data: data passed to writepage function
2577 *
2578 * Return: %0 on success, negative error code otherwise
2579 *
2580 * Note: please use writeback_iter() instead.
2581 */
2582int write_cache_pages(struct address_space *mapping,
2583 struct writeback_control *wbc, writepage_t writepage,
2584 void *data)
2585{
2586 struct folio *folio = NULL;
2587 int error;
2588
2589 while ((folio = writeback_iter(mapping, wbc, folio, &error))) {
2590 error = writepage(folio, wbc, data);
2591 if (error == AOP_WRITEPAGE_ACTIVATE) {
2592 folio_unlock(folio);
2593 error = 0;
2594 }
2595 }
2596
Christoph Hellwigf946e0d2024-02-15 07:36:41 +01002597 return error;
David Howells811d7362006-08-29 19:06:09 +01002598}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002599EXPORT_SYMBOL(write_cache_pages);
2600
Matthew Wilcox (Oracle)c44ed5b2024-02-15 07:36:49 +01002601static int writeback_use_writepage(struct address_space *mapping,
2602 struct writeback_control *wbc)
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002603{
Matthew Wilcox (Oracle)c44ed5b2024-02-15 07:36:49 +01002604 struct folio *folio = NULL;
2605 struct blk_plug plug;
2606 int err;
Christoph Hellwig67689072024-02-15 07:36:36 +01002607
Matthew Wilcox (Oracle)c44ed5b2024-02-15 07:36:49 +01002608 blk_start_plug(&plug);
2609 while ((folio = writeback_iter(mapping, wbc, folio, &err))) {
2610 err = mapping->a_ops->writepage(&folio->page, wbc);
2611 if (err == AOP_WRITEPAGE_ACTIVATE) {
2612 folio_unlock(folio);
2613 err = 0;
2614 }
2615 mapping_set_error(mapping, err);
2616 }
2617 blk_finish_plug(&plug);
2618
2619 return err;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002620}
2621
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2623{
Andrew Morton22905f72005-11-16 15:07:01 -08002624 int ret;
Jan Karafee468f2021-09-02 14:53:06 -07002625 struct bdi_writeback *wb;
Andrew Morton22905f72005-11-16 15:07:01 -08002626
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 if (wbc->nr_to_write <= 0)
2628 return 0;
Jan Karafee468f2021-09-02 14:53:06 -07002629 wb = inode_to_wb_wbc(mapping->host, wbc);
2630 wb_bandwidth_estimate_start(wb);
Theodore Ts'o80a2ea92017-04-28 09:51:54 -04002631 while (1) {
Christoph Hellwigc2ca7a52022-12-29 06:10:31 -10002632 if (mapping->a_ops->writepages) {
Theodore Ts'o80a2ea92017-04-28 09:51:54 -04002633 ret = mapping->a_ops->writepages(mapping, wbc);
Christoph Hellwigc2ca7a52022-12-29 06:10:31 -10002634 } else if (mapping->a_ops->writepage) {
Matthew Wilcox (Oracle)c44ed5b2024-02-15 07:36:49 +01002635 ret = writeback_use_writepage(mapping, wbc);
Christoph Hellwigc2ca7a52022-12-29 06:10:31 -10002636 } else {
2637 /* deal with chardevs and other special files */
2638 ret = 0;
2639 }
2640 if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL)
Theodore Ts'o80a2ea92017-04-28 09:51:54 -04002641 break;
Mel Gorman8d588022021-11-05 13:42:35 -07002642
2643 /*
2644 * Lacking an allocation context or the locality or writeback
2645 * state of any of the inode's pages, throttle based on
2646 * writeback activity on the local node. It's as good a
2647 * guess as any.
2648 */
2649 reclaim_throttle(NODE_DATA(numa_node_id()),
Mel Gormanc3f4a9a2021-11-05 13:42:42 -07002650 VMSCAN_THROTTLE_WRITEBACK);
Theodore Ts'o80a2ea92017-04-28 09:51:54 -04002651 }
Jan Kara45a29662021-09-02 14:53:09 -07002652 /*
2653 * Usually few pages are written by now from those we've just submitted
2654 * but if there's constant writeback being submitted, this makes sure
2655 * writeback bandwidth is updated once in a while.
2656 */
Jan Kara20792eb2021-09-02 14:53:15 -07002657 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2658 BANDWIDTH_INTERVAL))
Jan Kara45a29662021-09-02 14:53:09 -07002659 wb_update_bandwidth(wb);
Andrew Morton22905f72005-11-16 15:07:01 -08002660 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661}
2662
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663/*
Ken Chen76719322007-02-10 01:43:15 -08002664 * For address_spaces which do not use buffers nor write back.
2665 */
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +00002666bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
Ken Chen76719322007-02-10 01:43:15 -08002667{
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +00002668 if (!folio_test_dirty(folio))
2669 return !folio_test_set_dirty(folio);
2670 return false;
Ken Chen76719322007-02-10 01:43:15 -08002671}
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +00002672EXPORT_SYMBOL(noop_dirty_folio);
Ken Chen76719322007-02-10 01:43:15 -08002673
2674/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002675 * Helper function for set_page_dirty family.
Greg Thelenc4843a72015-05-22 17:13:16 -04002676 *
Kefeng Wang6c77b602023-06-14 22:36:12 +08002677 * Caller must hold folio_memcg_lock().
Greg Thelenc4843a72015-05-22 17:13:16 -04002678 *
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002679 * NOTE: This relies on being atomic wrt interrupts.
2680 */
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002681static void folio_account_dirtied(struct folio *folio,
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002682 struct address_space *mapping)
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002683{
Tejun Heo52ebea72015-05-22 17:13:37 -04002684 struct inode *inode = mapping->host;
2685
Matthew Wilcox (Oracle)b9b0ff62021-07-13 22:45:50 -04002686 trace_writeback_dirty_folio(folio, mapping);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002687
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002688 if (mapping_can_writeback(mapping)) {
Tejun Heo52ebea72015-05-22 17:13:37 -04002689 struct bdi_writeback *wb;
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002690 long nr = folio_nr_pages(folio);
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002691
Matthew Wilcox (Oracle)9cfb8162023-01-16 19:25:06 +00002692 inode_attach_wb(inode, folio);
Tejun Heo52ebea72015-05-22 17:13:37 -04002693 wb = inode_to_wb(inode);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002694
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002695 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2696 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2697 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2698 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2699 wb_stat_mod(wb, WB_DIRTIED, nr);
2700 task_io_account_write(nr * PAGE_SIZE);
2701 current->nr_dirtied += nr;
2702 __this_cpu_add(bdp_ratelimits, nr);
Tejun Heo97b27822019-08-26 09:06:56 -07002703
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002704 mem_cgroup_track_foreign_dirty(folio, wb);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002705 }
2706}
2707
2708/*
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002709 * Helper function for deaccounting dirty page without writeback.
2710 *
Kefeng Wang6c77b602023-06-14 22:36:12 +08002711 * Caller must hold folio_memcg_lock().
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002712 */
Hugh Dickins566d3362022-03-24 18:13:59 -07002713void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002714{
Hugh Dickins566d3362022-03-24 18:13:59 -07002715 long nr = folio_nr_pages(folio);
2716
2717 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2718 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2719 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2720 task_io_account_cancelled_write(nr * PAGE_SIZE);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002721}
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002722
2723/*
Kemeng Shidc6e0ae2024-04-25 21:17:24 +08002724 * Mark the folio dirty, and set it dirty in the page cache.
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002725 *
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002726 * If warn is true, then emit a warning if the folio is not uptodate and has
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002727 * not been truncated.
2728 *
Matthew Wilcox (Oracle)3d84d892024-04-16 04:17:45 +01002729 * The caller must hold folio_memcg_lock(). It is the caller's
2730 * responsibility to prevent the folio from being truncated while
2731 * this function is in progress, although it may have been truncated
2732 * before this function is called. Most callers have the folio locked.
2733 * A few have the folio blocked from truncation through other means (e.g.
2734 * zap_vma_pages() has it mapped and is holding the page table lock).
2735 * When called from mark_buffer_dirty(), the filesystem should hold a
2736 * reference to the buffer_head that is being marked dirty, which causes
2737 * try_to_free_buffers() to fail.
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002738 */
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002739void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002740 int warn)
2741{
2742 unsigned long flags;
2743
2744 xa_lock_irqsave(&mapping->i_pages, flags);
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002745 if (folio->mapping) { /* Race with truncate? */
2746 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2747 folio_account_dirtied(folio, mapping);
2748 __xa_set_mark(&mapping->i_pages, folio_index(folio),
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002749 PAGECACHE_TAG_DIRTY);
2750 }
2751 xa_unlock_irqrestore(&mapping->i_pages, flags);
2752}
2753
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002754/**
2755 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2756 * @mapping: Address space this folio belongs to.
2757 * @folio: Folio to be marked as dirty.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 *
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002759 * Filesystems which do not use buffer heads should call this function
Jianguo Baoab428b42023-09-17 23:04:01 +08002760 * from their dirty_folio address space operation. It ignores the
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002761 * contents of folio_get_private(), so if the filesystem marks individual
2762 * blocks as dirty, the filesystem should handle that itself.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 *
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002764 * This is also sometimes used by filesystems which use buffer_heads when
2765 * a single buffer is being dirtied: we want to set the folio dirty in
2766 * that case, but not all the buffers. This is a "bottom-up" dirtying,
Matthew Wilcox (Oracle)e6219002022-02-09 20:22:12 +00002767 * whereas block_dirty_folio() is a "top-down" dirtying.
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002768 *
2769 * The caller must ensure this doesn't race with truncation. Most will
2770 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2771 * folio mapped and the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 */
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002773bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774{
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002775 folio_memcg_lock(folio);
2776 if (folio_test_set_dirty(folio)) {
2777 folio_memcg_unlock(folio);
2778 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 }
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002780
2781 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2782 folio_memcg_unlock(folio);
2783
2784 if (mapping->host) {
2785 /* !PageAnon && !swapper_space */
2786 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2787 }
2788 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789}
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002790EXPORT_SYMBOL(filemap_dirty_folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04002792/**
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002793 * folio_redirty_for_writepage - Decline to write a dirty folio.
2794 * @wbc: The writeback control.
2795 * @folio: The folio.
2796 *
2797 * When a writepage implementation decides that it doesn't want to write
2798 * @folio for some reason, it should call this function, unlock @folio and
2799 * return 0.
2800 *
2801 * Return: True if we redirtied the folio. False if someone else dirtied
2802 * it first.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 */
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002804bool folio_redirty_for_writepage(struct writeback_control *wbc,
2805 struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806{
Christoph Hellwiged2da922023-06-28 17:31:44 +02002807 struct address_space *mapping = folio->mapping;
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002808 long nr = folio_nr_pages(folio);
Christoph Hellwiged2da922023-06-28 17:31:44 +02002809 bool ret;
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002810
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002811 wbc->pages_skipped += nr;
Christoph Hellwiged2da922023-06-28 17:31:44 +02002812 ret = filemap_dirty_folio(mapping, folio);
2813 if (mapping && mapping_can_writeback(mapping)) {
2814 struct inode *inode = mapping->host;
2815 struct bdi_writeback *wb;
2816 struct wb_lock_cookie cookie = {};
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002817
Christoph Hellwiged2da922023-06-28 17:31:44 +02002818 wb = unlocked_inode_to_wb_begin(inode, &cookie);
2819 current->nr_dirtied -= nr;
2820 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2821 wb_stat_mod(wb, WB_DIRTIED, -nr);
2822 unlocked_inode_to_wb_end(inode, &cookie);
2823 }
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002824 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825}
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002826EXPORT_SYMBOL(folio_redirty_for_writepage);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002828/**
2829 * folio_mark_dirty - Mark a folio as being modified.
2830 * @folio: The folio.
Wu Fengguang6746aff2009-09-16 11:50:14 +02002831 *
Matthew Wilcox (Oracle)2ca456c2022-04-28 14:21:02 -04002832 * The folio may not be truncated while this function is running.
2833 * Holding the folio lock is sufficient to prevent truncation, but some
2834 * callers cannot acquire a sleeping lock. These callers instead hold
2835 * the page table lock for a page table which contains at least one page
2836 * in this folio. Truncation will block on the page table lock as it
2837 * unmaps pages before removing the folio from its mapping.
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002838 *
2839 * Return: True if the folio was newly dirtied, false if it was already dirty.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 */
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002841bool folio_mark_dirty(struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842{
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002843 struct address_space *mapping = folio_mapping(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844
2845 if (likely(mapping)) {
Minchan Kim278df9f2011-03-22 16:32:54 -07002846 /*
Vishal Moola (Oracle)5a9e3472022-12-21 10:08:48 -08002847 * readahead/folio_deactivate could remain
Matthew Wilcox (Oracle)6f31a5a2022-02-09 20:22:00 +00002848 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2849 * About readahead, if the folio is written, the flags would be
Minchan Kim278df9f2011-03-22 16:32:54 -07002850 * reset. So no problem.
Vishal Moola (Oracle)5a9e3472022-12-21 10:08:48 -08002851 * About folio_deactivate, if the folio is redirtied,
Matthew Wilcox (Oracle)6f31a5a2022-02-09 20:22:00 +00002852 * the flag will be reset. So no problem. but if the
2853 * folio is used by readahead it will confuse readahead
2854 * and make it restart the size rampup process. But it's
2855 * a trivial problem.
Minchan Kim278df9f2011-03-22 16:32:54 -07002856 */
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002857 if (folio_test_reclaim(folio))
2858 folio_clear_reclaim(folio);
Matthew Wilcox (Oracle)3a3bae52022-02-09 20:22:15 +00002859 return mapping->a_ops->dirty_folio(mapping, folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860 }
Matthew Wilcox (Oracle)3a3bae52022-02-09 20:22:15 +00002861
2862 return noop_dirty_folio(mapping, folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863}
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002864EXPORT_SYMBOL(folio_mark_dirty);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865
2866/*
2867 * set_page_dirty() is racy if the caller has no reference against
2868 * page->mapping->host, and if the page is unlocked. This is because another
2869 * CPU could truncate the page off the mapping and then free the mapping.
2870 *
2871 * Usually, the page _is_ locked, or the caller is a user-space process which
2872 * holds a reference on the inode by having an open file.
2873 *
2874 * In other cases, the page should be locked before running set_page_dirty().
2875 */
2876int set_page_dirty_lock(struct page *page)
2877{
2878 int ret;
2879
Jens Axboe7eaceac2011-03-10 08:52:07 +01002880 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 ret = set_page_dirty(page);
2882 unlock_page(page);
2883 return ret;
2884}
2885EXPORT_SYMBOL(set_page_dirty_lock);
2886
2887/*
Tejun Heo11f81be2015-05-22 17:13:15 -04002888 * This cancels just the dirty bit on the kernel page itself, it does NOT
2889 * actually remove dirty bits on any mmap's that may be around. It also
2890 * leaves the page tagged dirty, so any sync activity will still find it on
2891 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2892 * look at the dirty bits in the VM.
2893 *
2894 * Doing this should *normally* only ever be done when a page is truncated,
2895 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2896 * this when it notices that somebody has cleaned out all the buffers on a
2897 * page without actually doing it through the VM. Can you say "ext3 is
2898 * horribly ugly"? Thought you could.
2899 */
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002900void __folio_cancel_dirty(struct folio *folio)
Tejun Heo11f81be2015-05-22 17:13:15 -04002901{
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002902 struct address_space *mapping = folio_mapping(folio);
Greg Thelenc4843a72015-05-22 17:13:16 -04002903
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002904 if (mapping_can_writeback(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002905 struct inode *inode = mapping->host;
2906 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002907 struct wb_lock_cookie cookie = {};
Greg Thelenc4843a72015-05-22 17:13:16 -04002908
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002909 folio_memcg_lock(folio);
Greg Thelen2e898e42018-04-20 14:55:42 -07002910 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002911
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002912 if (folio_test_clear_dirty(folio))
Hugh Dickins566d3362022-03-24 18:13:59 -07002913 folio_account_cleaned(folio, wb);
Greg Thelenc4843a72015-05-22 17:13:16 -04002914
Greg Thelen2e898e42018-04-20 14:55:42 -07002915 unlocked_inode_to_wb_end(inode, &cookie);
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002916 folio_memcg_unlock(folio);
Greg Thelenc4843a72015-05-22 17:13:16 -04002917 } else {
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002918 folio_clear_dirty(folio);
Greg Thelenc4843a72015-05-22 17:13:16 -04002919 }
Tejun Heo11f81be2015-05-22 17:13:15 -04002920}
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002921EXPORT_SYMBOL(__folio_cancel_dirty);
Tejun Heo11f81be2015-05-22 17:13:15 -04002922
2923/*
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002924 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2925 * Returns true if the folio was previously dirty.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002927 * This is for preparing to put the folio under writeout. We leave
2928 * the folio tagged as dirty in the xarray so that a concurrent
2929 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2930 * The ->writepage implementation will run either folio_start_writeback()
2931 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2932 * and xarray dirty tag back into sync.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002934 * This incoherency between the folio's dirty flag and xarray tag is
2935 * unfortunate, but it only exists while the folio is locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 */
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002937bool folio_clear_dirty_for_io(struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938{
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002939 struct address_space *mapping = folio_mapping(folio);
2940 bool ret = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002942 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
Nick Piggin79352892007-07-19 01:47:22 -07002943
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002944 if (mapping && mapping_can_writeback(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002945 struct inode *inode = mapping->host;
2946 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002947 struct wb_lock_cookie cookie = {};
Tejun Heo682aa8e2015-05-28 14:50:53 -04002948
Linus Torvalds7658cc22006-12-29 10:00:58 -08002949 /*
2950 * Yes, Virginia, this is indeed insane.
2951 *
2952 * We use this sequence to make sure that
2953 * (a) we account for dirty stats properly
2954 * (b) we tell the low-level filesystem to
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002955 * mark the whole folio dirty if it was
Linus Torvalds7658cc22006-12-29 10:00:58 -08002956 * dirty in a pagetable. Only to then
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002957 * (c) clean the folio again and return 1 to
Linus Torvalds7658cc22006-12-29 10:00:58 -08002958 * cause the writeback.
2959 *
2960 * This way we avoid all nasty races with the
2961 * dirty bit in multiple places and clearing
2962 * them concurrently from different threads.
2963 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002964 * Note! Normally the "folio_mark_dirty(folio)"
Linus Torvalds7658cc22006-12-29 10:00:58 -08002965 * has no effect on the actual dirty bit - since
2966 * that will already usually be set. But we
2967 * need the side effects, and it can help us
2968 * avoid races.
2969 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002970 * We basically use the folio "master dirty bit"
Linus Torvalds7658cc22006-12-29 10:00:58 -08002971 * as a serialization point for all the different
2972 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002973 */
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002974 if (folio_mkclean(folio))
2975 folio_mark_dirty(folio);
Nick Piggin79352892007-07-19 01:47:22 -07002976 /*
2977 * We carefully synchronise fault handlers against
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002978 * installing a dirty pte and marking the folio dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002979 * at this point. We do this by having them hold the
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002980 * page lock while dirtying the folio, and folios are
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002981 * always locked coming in here, so we get the desired
2982 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002983 */
Greg Thelen2e898e42018-04-20 14:55:42 -07002984 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002985 if (folio_test_clear_dirty(folio)) {
2986 long nr = folio_nr_pages(folio);
2987 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2988 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2989 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2990 ret = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991 }
Greg Thelen2e898e42018-04-20 14:55:42 -07002992 unlocked_inode_to_wb_end(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002993 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 }
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002995 return folio_test_clear_dirty(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996}
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002997EXPORT_SYMBOL(folio_clear_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998
Jan Kara633a2ab2021-09-02 14:53:04 -07002999static void wb_inode_writeback_start(struct bdi_writeback *wb)
3000{
3001 atomic_inc(&wb->writeback_inodes);
3002}
3003
3004static void wb_inode_writeback_end(struct bdi_writeback *wb)
3005{
Khazhismel Kumykovf87904c2022-08-01 08:50:34 -07003006 unsigned long flags;
Jan Kara633a2ab2021-09-02 14:53:04 -07003007 atomic_dec(&wb->writeback_inodes);
Jan Kara45a29662021-09-02 14:53:09 -07003008 /*
3009 * Make sure estimate of writeback throughput gets updated after
3010 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
3011 * (which is the interval other bandwidth updates use for batching) so
3012 * that if multiple inodes end writeback at a similar time, they get
3013 * batched into one bandwidth update.
3014 */
Khazhismel Kumykovf87904c2022-08-01 08:50:34 -07003015 spin_lock_irqsave(&wb->work_lock, flags);
3016 if (test_bit(WB_registered, &wb->state))
3017 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
3018 spin_unlock_irqrestore(&wb->work_lock, flags);
Jan Kara633a2ab2021-09-02 14:53:04 -07003019}
3020
Matthew Wilcox (Oracle)2580d552023-10-04 17:53:17 +01003021bool __folio_end_writeback(struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022{
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05003023 long nr = folio_nr_pages(folio);
3024 struct address_space *mapping = folio_mapping(folio);
Matthew Wilcox (Oracle)2580d552023-10-04 17:53:17 +01003025 bool ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05003027 folio_memcg_lock(folio);
Huang Ying371a0962016-10-07 16:59:30 -07003028 if (mapping && mapping_use_writeback_tags(mapping)) {
Tejun Heo91018132015-05-22 17:13:39 -04003029 struct inode *inode = mapping->host;
3030 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 unsigned long flags;
3032
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07003033 xa_lock_irqsave(&mapping->i_pages, flags);
Matthew Wilcox (Oracle)2580d552023-10-04 17:53:17 +01003034 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
Matthew Wilcox (Oracle)7d0795d2023-10-04 17:53:16 +01003035 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
3036 PAGECACHE_TAG_WRITEBACK);
3037 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3038 struct bdi_writeback *wb = inode_to_wb(inode);
Tejun Heo91018132015-05-22 17:13:39 -04003039
Matthew Wilcox (Oracle)7d0795d2023-10-04 17:53:16 +01003040 wb_stat_mod(wb, WB_WRITEBACK, -nr);
3041 __wb_writeout_add(wb, nr);
3042 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
3043 wb_inode_writeback_end(wb);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07003044 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07003045
3046 if (mapping->host && !mapping_tagged(mapping,
3047 PAGECACHE_TAG_WRITEBACK))
3048 sb_clear_inode_writeback(mapping->host);
3049
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07003050 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 } else {
Matthew Wilcox (Oracle)2580d552023-10-04 17:53:17 +01003052 ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 }
Matthew Wilcox (Oracle)7d0795d2023-10-04 17:53:16 +01003054
3055 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
3056 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3057 node_stat_mod_folio(folio, NR_WRITTEN, nr);
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05003058 folio_memcg_unlock(folio);
Matthew Wilcox (Oracle)2580d552023-10-04 17:53:17 +01003059
3060 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061}
3062
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003063void __folio_start_writeback(struct folio *folio, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064{
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003065 long nr = folio_nr_pages(folio);
3066 struct address_space *mapping = folio_mapping(folio);
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003067 int access_ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003069 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
3070
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003071 folio_memcg_lock(folio);
Huang Ying371a0962016-10-07 16:59:30 -07003072 if (mapping && mapping_use_writeback_tags(mapping)) {
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003073 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
Tejun Heo91018132015-05-22 17:13:39 -04003074 struct inode *inode = mapping->host;
3075 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076 unsigned long flags;
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003077 bool on_wblist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05003079 xas_lock_irqsave(&xas, flags);
3080 xas_load(&xas);
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003081 folio_test_set_writeback(folio);
Dave Chinner6c60d2b2016-07-26 15:21:50 -07003082
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003083 on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
Dave Chinner6c60d2b2016-07-26 15:21:50 -07003084
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003085 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3086 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3087 struct bdi_writeback *wb = inode_to_wb(inode);
Jan Kara633a2ab2021-09-02 14:53:04 -07003088
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003089 wb_stat_mod(wb, WB_WRITEBACK, nr);
3090 if (!on_wblist)
3091 wb_inode_writeback_start(wb);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07003092 }
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003093
3094 /*
3095 * We can come through here when swapping anonymous
3096 * folios, so we don't necessarily have an inode to
3097 * track for sync.
3098 */
3099 if (mapping->host && !on_wblist)
3100 sb_mark_inode_writeback(mapping->host);
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003101 if (!folio_test_dirty(folio))
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05003102 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04003103 if (!keep_write)
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05003104 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3105 xas_unlock_irqrestore(&xas, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 } else {
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003107 folio_test_set_writeback(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 }
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003109
3110 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3111 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003112 folio_memcg_unlock(folio);
Matthew Wilcox (Oracle)b5612c32023-11-08 20:46:05 +00003113
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003114 access_ret = arch_make_folio_accessible(folio);
Claudio Imbrendaf28d4362020-04-01 21:05:56 -07003115 /*
3116 * If writeback has been triggered on a page that cannot be made
3117 * accessible, it is too late to recover here.
3118 */
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003119 VM_BUG_ON_FOLIO(access_ret != 0, folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120}
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04003121EXPORT_SYMBOL(__folio_start_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003123/**
3124 * folio_wait_writeback - Wait for a folio to finish writeback.
3125 * @folio: The folio to wait for.
3126 *
3127 * If the folio is currently being written back to storage, wait for the
3128 * I/O to complete.
3129 *
3130 * Context: Sleeps. Must be called in process context and with
3131 * no spinlocks held. Caller should hold a reference on the folio.
3132 * If the folio is not locked, writeback may start again after writeback
3133 * has finished.
Yafang Shao19343b52019-05-13 17:23:11 -07003134 */
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003135void folio_wait_writeback(struct folio *folio)
Yafang Shao19343b52019-05-13 17:23:11 -07003136{
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003137 while (folio_test_writeback(folio)) {
Matthew Wilcox (Oracle)b9b0ff62021-07-13 22:45:50 -04003138 trace_folio_wait_writeback(folio, folio_mapping(folio));
Matthew Wilcox (Oracle)101c0bf2021-03-04 12:02:54 -05003139 folio_wait_bit(folio, PG_writeback);
Yafang Shao19343b52019-05-13 17:23:11 -07003140 }
3141}
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003142EXPORT_SYMBOL_GPL(folio_wait_writeback);
Yafang Shao19343b52019-05-13 17:23:11 -07003143
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003144/**
3145 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3146 * @folio: The folio to wait for.
3147 *
3148 * If the folio is currently being written back to storage, wait for the
3149 * I/O to complete or a fatal signal to arrive.
3150 *
3151 * Context: Sleeps. Must be called in process context and with
3152 * no spinlocks held. Caller should hold a reference on the folio.
3153 * If the folio is not locked, writeback may start again after writeback
3154 * has finished.
3155 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00003156 */
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003157int folio_wait_writeback_killable(struct folio *folio)
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00003158{
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003159 while (folio_test_writeback(folio)) {
Matthew Wilcox (Oracle)b9b0ff62021-07-13 22:45:50 -04003160 trace_folio_wait_writeback(folio, folio_mapping(folio));
Matthew Wilcox (Oracle)101c0bf2021-03-04 12:02:54 -05003161 if (folio_wait_bit_killable(folio, PG_writeback))
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00003162 return -EINTR;
3163 }
3164
3165 return 0;
3166}
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05003167EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00003168
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08003169/**
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05003170 * folio_wait_stable() - wait for writeback to finish, if necessary.
3171 * @folio: The folio to wait on.
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08003172 *
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05003173 * This function determines if the given folio is related to a backing
3174 * device that requires folio contents to be held stable during writeback.
3175 * If so, then it will wait for any pending writeback to complete.
3176 *
3177 * Context: Sleeps. Must be called in process context and with
3178 * no spinlocks held. Caller should hold a reference on the folio.
3179 * If the folio is not locked, writeback may start again after writeback
3180 * has finished.
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08003181 */
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05003182void folio_wait_stable(struct folio *folio)
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08003183{
Christoph Hellwig762321d2023-10-25 16:10:17 +02003184 if (mapping_stable_writes(folio_mapping(folio)))
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05003185 folio_wait_writeback(folio);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08003186}
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05003187EXPORT_SYMBOL_GPL(folio_wait_stable);