blob: fa1117db4610f2ba44a7319c4bf674b523f77821 [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>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040016#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/spinlock.h>
18#include <linux/fs.h>
19#include <linux/mm.h>
20#include <linux/swap.h>
21#include <linux/slab.h>
22#include <linux/pagemap.h>
23#include <linux/writeback.h>
24#include <linux/init.h>
25#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080026#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070027#include <linux/blkdev.h>
28#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070029#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#include <linux/percpu.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
David Howells811d7362006-08-29 19:06:09 +010035#include <linux/pagevec.h>
Jan Karaeb608e32012-05-24 18:59:11 +020036#include <linux/timer.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060037#include <linux/sched/rt.h>
Ingo Molnarf361bf42017-02-03 23:47:37 +010038#include <linux/sched/signal.h>
Lisa Du6e543d52013-09-11 14:22:36 -070039#include <linux/mm_inline.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100040#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
Lisa Du6e543d52013-09-11 14:22:36 -070042#include "internal.h"
43
Linus Torvalds1da177e2005-04-16 15:20:36 -070044/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060045 * Sleep at most 200ms at a time in balance_dirty_pages().
46 */
47#define MAX_PAUSE max(HZ/5, 1)
48
49/*
Wu Fengguang5b9b3572011-12-06 13:17:17 -060050 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
52 */
53#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
54
55/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060056 * Estimate write bandwidth at 200ms intervals.
57 */
58#define BANDWIDTH_INTERVAL max(HZ/5, 1)
59
Wu Fengguang6c14ae12011-03-02 16:04:18 -060060#define RATELIMIT_CALC_SHIFT 10
61
Wu Fengguange98be2d2010-08-29 11:22:30 -060062/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
65 */
66static long ratelimit_pages = 32;
67
Linus Torvalds1da177e2005-04-16 15:20:36 -070068/* The following parameters are exported via /proc/sys/vm */
69
70/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020071 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080073int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
75/*
David Rientjes2da02992009-01-06 14:39:31 -080076 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
78 */
79unsigned long dirty_background_bytes;
80
81/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080082 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
84 */
85int vm_highmem_is_dirtyable;
86
87/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 * The generator of dirty data starts writeback at this percentage
89 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080090int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92/*
David Rientjes2da02992009-01-06 14:39:31 -080093 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
95 */
96unsigned long vm_dirty_bytes;
97
98/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070099 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Artem Bityutskiy91913a22012-03-21 22:33:00 -0400103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700106 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800111 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
112 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 */
114int laptop_mode;
115
116EXPORT_SYMBOL(laptop_mode);
117
118/* End of sysctl-exported parameters */
119
Tejun Heodcc25ae2015-05-22 18:23:22 -0400120struct wb_domain global_wb_domain;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400122/* consolidated parameters for balance_dirty_pages() and its subroutines */
123struct dirty_throttle_control {
Tejun Heoe9f07df2015-05-22 18:23:28 -0400124#ifdef CONFIG_CGROUP_WRITEBACK
125 struct wb_domain *dom;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400126 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
Tejun Heoe9f07df2015-05-22 18:23:28 -0400127#endif
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400128 struct bdi_writeback *wb;
Tejun Heoe9770b32015-05-22 18:23:27 -0400129 struct fprop_local_percpu *wb_completions;
Jan Karaeb608e32012-05-24 18:59:11 +0200130
Tejun Heo9fc3a432015-05-22 18:23:30 -0400131 unsigned long avail; /* dirtyable */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400132 unsigned long dirty; /* file_dirty + write + nfs */
133 unsigned long thresh; /* dirty threshold */
134 unsigned long bg_thresh; /* dirty background threshold */
135
136 unsigned long wb_dirty; /* per-wb counterparts */
137 unsigned long wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -0400138 unsigned long wb_bg_thresh;
Tejun Heodaddfa32015-05-22 18:23:26 -0400139
140 unsigned long pos_ratio;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400141};
142
Jan Karaeb608e32012-05-24 18:59:11 +0200143/*
144 * Length of period for aging writeout fractions of bdis. This is an
145 * arbitrarily chosen number. The longer the period, the slower fractions will
146 * reflect changes in current writeout rate.
147 */
148#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700149
Tejun Heo693108a2015-05-22 17:13:49 -0400150#ifdef CONFIG_CGROUP_WRITEBACK
151
Tejun Heod60d1bd2015-09-29 12:47:53 -0400152#define GDTC_INIT(__wb) .wb = (__wb), \
153 .dom = &global_wb_domain, \
154 .wb_completions = &(__wb)->completions
155
Tejun Heo9fc3a432015-05-22 18:23:30 -0400156#define GDTC_INIT_NO_WB .dom = &global_wb_domain
Tejun Heod60d1bd2015-09-29 12:47:53 -0400157
158#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
159 .dom = mem_cgroup_wb_domain(__wb), \
160 .wb_completions = &(__wb)->memcg_completions, \
161 .gdtc = __gdtc
Tejun Heoc2aa7232015-05-22 18:23:35 -0400162
163static bool mdtc_valid(struct dirty_throttle_control *dtc)
164{
165 return dtc->dom;
166}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400167
168static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
169{
170 return dtc->dom;
171}
172
Tejun Heo9fc3a432015-05-22 18:23:30 -0400173static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
174{
175 return mdtc->gdtc;
176}
177
Tejun Heo841710a2015-05-22 18:23:33 -0400178static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
179{
180 return &wb->memcg_completions;
181}
182
Tejun Heo693108a2015-05-22 17:13:49 -0400183static void wb_min_max_ratio(struct bdi_writeback *wb,
184 unsigned long *minp, unsigned long *maxp)
185{
Jan Kara20792eb2021-09-02 14:53:15 -0700186 unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
Tejun Heo693108a2015-05-22 17:13:49 -0400187 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
188 unsigned long long min = wb->bdi->min_ratio;
189 unsigned long long max = wb->bdi->max_ratio;
190
191 /*
192 * @wb may already be clean by the time control reaches here and
193 * the total may not include its bw.
194 */
195 if (this_bw < tot_bw) {
196 if (min) {
197 min *= this_bw;
Wen Yang6d9e8c62020-01-13 16:29:23 -0800198 min = div64_ul(min, tot_bw);
Tejun Heo693108a2015-05-22 17:13:49 -0400199 }
200 if (max < 100) {
201 max *= this_bw;
Wen Yang6d9e8c62020-01-13 16:29:23 -0800202 max = div64_ul(max, tot_bw);
Tejun Heo693108a2015-05-22 17:13:49 -0400203 }
204 }
205
206 *minp = min;
207 *maxp = max;
208}
209
210#else /* CONFIG_CGROUP_WRITEBACK */
211
Tejun Heod60d1bd2015-09-29 12:47:53 -0400212#define GDTC_INIT(__wb) .wb = (__wb), \
213 .wb_completions = &(__wb)->completions
Tejun Heo9fc3a432015-05-22 18:23:30 -0400214#define GDTC_INIT_NO_WB
Tejun Heoc2aa7232015-05-22 18:23:35 -0400215#define MDTC_INIT(__wb, __gdtc)
216
217static bool mdtc_valid(struct dirty_throttle_control *dtc)
218{
219 return false;
220}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400221
222static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
223{
224 return &global_wb_domain;
225}
226
Tejun Heo9fc3a432015-05-22 18:23:30 -0400227static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
228{
229 return NULL;
230}
231
Tejun Heo841710a2015-05-22 18:23:33 -0400232static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
233{
234 return NULL;
235}
236
Tejun Heo693108a2015-05-22 17:13:49 -0400237static void wb_min_max_ratio(struct bdi_writeback *wb,
238 unsigned long *minp, unsigned long *maxp)
239{
240 *minp = wb->bdi->min_ratio;
241 *maxp = wb->bdi->max_ratio;
242}
243
244#endif /* CONFIG_CGROUP_WRITEBACK */
245
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700246/*
Johannes Weinera756cf52012-01-10 15:07:49 -0800247 * In a memory zone, there is a certain amount of pages we consider
248 * available for the page cache, which is essentially the number of
249 * free and reclaimable pages, minus some zone reserves to protect
250 * lowmem and the ability to uphold the zone's watermarks without
251 * requiring writeback.
252 *
253 * This number of dirtyable pages is the base value of which the
Ethon Paule0857cf2020-06-04 16:49:40 -0700254 * user-configurable dirty ratio is the effective number of pages that
Johannes Weinera756cf52012-01-10 15:07:49 -0800255 * are allowed to be actually dirtied. Per individual zone, or
256 * globally by using the sum of dirtyable pages over all zones.
257 *
258 * Because the user is allowed to specify the dirty limit globally as
259 * absolute number of bytes, calculating the per-zone dirty limit can
260 * require translating the configured limit into a percentage of
261 * global dirtyable memory first.
262 */
263
Johannes Weinera804552b2014-01-29 14:05:39 -0800264/**
Mel Gorman281e3722016-07-28 15:46:11 -0700265 * node_dirtyable_memory - number of dirtyable pages in a node
266 * @pgdat: the node
Johannes Weinera804552b2014-01-29 14:05:39 -0800267 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800268 * Return: the node's number of pages potentially available for dirty
Mel Gorman281e3722016-07-28 15:46:11 -0700269 * page cache. This is the base value for the per-node dirty limits.
Johannes Weinera804552b2014-01-29 14:05:39 -0800270 */
Mel Gorman281e3722016-07-28 15:46:11 -0700271static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
Johannes Weinera804552b2014-01-29 14:05:39 -0800272{
Mel Gorman281e3722016-07-28 15:46:11 -0700273 unsigned long nr_pages = 0;
274 int z;
Johannes Weinera804552b2014-01-29 14:05:39 -0800275
Mel Gorman281e3722016-07-28 15:46:11 -0700276 for (z = 0; z < MAX_NR_ZONES; z++) {
277 struct zone *zone = pgdat->node_zones + z;
278
279 if (!populated_zone(zone))
280 continue;
281
282 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
283 }
284
Johannes Weinera8d01432016-01-14 15:20:15 -0800285 /*
286 * Pages reserved for the kernel should not be considered
287 * dirtyable, to prevent a situation where reclaim has to
288 * clean pages in order to balance the zones.
289 */
Mel Gorman281e3722016-07-28 15:46:11 -0700290 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
Johannes Weinera804552b2014-01-29 14:05:39 -0800291
Mel Gorman281e3722016-07-28 15:46:11 -0700292 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
293 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
Johannes Weinera804552b2014-01-29 14:05:39 -0800294
295 return nr_pages;
296}
297
Johannes Weiner1edf2232012-01-10 15:06:57 -0800298static unsigned long highmem_dirtyable_memory(unsigned long total)
299{
300#ifdef CONFIG_HIGHMEM
301 int node;
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700302 unsigned long x = 0;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700303 int i;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800304
305 for_each_node_state(node, N_HIGH_MEMORY) {
Mel Gorman281e3722016-07-28 15:46:11 -0700306 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
307 struct zone *z;
Minchan Kim9cb937e2016-07-28 15:47:08 -0700308 unsigned long nr_pages;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800309
Mel Gorman281e3722016-07-28 15:46:11 -0700310 if (!is_highmem_idx(i))
311 continue;
312
313 z = &NODE_DATA(node)->node_zones[i];
Minchan Kim9cb937e2016-07-28 15:47:08 -0700314 if (!populated_zone(z))
315 continue;
Mel Gorman281e3722016-07-28 15:46:11 -0700316
Minchan Kim9cb937e2016-07-28 15:47:08 -0700317 nr_pages = zone_page_state(z, NR_FREE_PAGES);
Mel Gorman281e3722016-07-28 15:46:11 -0700318 /* watch for underflows */
Minchan Kim9cb937e2016-07-28 15:47:08 -0700319 nr_pages -= min(nr_pages, high_wmark_pages(z));
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700320 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
321 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
322 x += nr_pages;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700323 }
Johannes Weiner1edf2232012-01-10 15:06:57 -0800324 }
Mel Gorman281e3722016-07-28 15:46:11 -0700325
Johannes Weiner1edf2232012-01-10 15:06:57 -0800326 /*
327 * Make sure that the number of highmem pages is never larger
328 * than the number of the total dirtyable memory. This can only
329 * occur in very strange VM situations but we want to make sure
330 * that this does not occur.
331 */
332 return min(x, total);
333#else
334 return 0;
335#endif
336}
337
338/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800339 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800340 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800341 * Return: the global number of pages potentially available for dirty
Johannes Weinerccafa282012-01-10 15:07:44 -0800342 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800343 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700344static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800345{
346 unsigned long x;
347
Michal Hockoc41f0122017-09-06 16:23:36 -0700348 x = global_zone_page_state(NR_FREE_PAGES);
Johannes Weinera8d01432016-01-14 15:20:15 -0800349 /*
350 * Pages reserved for the kernel should not be considered
351 * dirtyable, to prevent a situation where reclaim has to
352 * clean pages in order to balance the zones.
353 */
354 x -= min(x, totalreserve_pages);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800355
Mel Gorman599d0c92016-07-28 15:45:31 -0700356 x += global_node_page_state(NR_INACTIVE_FILE);
357 x += global_node_page_state(NR_ACTIVE_FILE);
Johannes Weinera804552b2014-01-29 14:05:39 -0800358
Johannes Weiner1edf2232012-01-10 15:06:57 -0800359 if (!vm_highmem_is_dirtyable)
360 x -= highmem_dirtyable_memory(x);
361
362 return x + 1; /* Ensure that we never return 0 */
363}
364
Tejun Heo9fc3a432015-05-22 18:23:30 -0400365/**
366 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
367 * @dtc: dirty_throttle_control of interest
Johannes Weinerccafa282012-01-10 15:07:44 -0800368 *
Tejun Heo9fc3a432015-05-22 18:23:30 -0400369 * Calculate @dtc->thresh and ->bg_thresh considering
370 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
371 * must ensure that @dtc->avail is set before calling this function. The
NeilBrowna37b0712020-06-01 21:48:18 -0700372 * dirty limits will be lifted by 1/4 for real-time tasks.
Johannes Weinerccafa282012-01-10 15:07:44 -0800373 */
Tejun Heo9fc3a432015-05-22 18:23:30 -0400374static void domain_dirty_limits(struct dirty_throttle_control *dtc)
375{
376 const unsigned long available_memory = dtc->avail;
377 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
378 unsigned long bytes = vm_dirty_bytes;
379 unsigned long bg_bytes = dirty_background_bytes;
Tejun Heo62a584f2016-05-27 14:34:46 -0400380 /* convert ratios to per-PAGE_SIZE for higher precision */
381 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
382 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400383 unsigned long thresh;
384 unsigned long bg_thresh;
385 struct task_struct *tsk;
386
387 /* gdtc is !NULL iff @dtc is for memcg domain */
388 if (gdtc) {
389 unsigned long global_avail = gdtc->avail;
390
391 /*
392 * The byte settings can't be applied directly to memcg
393 * domains. Convert them to ratios by scaling against
Tejun Heo62a584f2016-05-27 14:34:46 -0400394 * globally available memory. As the ratios are in
395 * per-PAGE_SIZE, they can be obtained by dividing bytes by
396 * number of pages.
Tejun Heo9fc3a432015-05-22 18:23:30 -0400397 */
398 if (bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400399 ratio = min(DIV_ROUND_UP(bytes, global_avail),
400 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400401 if (bg_bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400402 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
403 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400404 bytes = bg_bytes = 0;
405 }
406
407 if (bytes)
408 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
409 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400410 thresh = (ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400411
412 if (bg_bytes)
413 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
414 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400415 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400416
Michal Hocko90daf302017-11-29 16:10:58 -0800417 if (bg_thresh >= thresh)
Tejun Heo9fc3a432015-05-22 18:23:30 -0400418 bg_thresh = thresh / 2;
419 tsk = current;
NeilBrowna37b0712020-06-01 21:48:18 -0700420 if (rt_task(tsk)) {
NeilBrowna53eaff2016-05-20 16:58:53 -0700421 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
422 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400423 }
424 dtc->thresh = thresh;
425 dtc->bg_thresh = bg_thresh;
426
427 /* we should eventually report the domain in the TP */
428 if (!gdtc)
429 trace_global_dirty_state(bg_thresh, thresh);
430}
431
432/**
433 * global_dirty_limits - background-writeback and dirty-throttling thresholds
434 * @pbackground: out parameter for bg_thresh
435 * @pdirty: out parameter for thresh
436 *
437 * Calculate bg_thresh and thresh for global_wb_domain. See
438 * domain_dirty_limits() for details.
439 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800440void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
441{
Tejun Heo9fc3a432015-05-22 18:23:30 -0400442 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
Johannes Weinerccafa282012-01-10 15:07:44 -0800443
Tejun Heo9fc3a432015-05-22 18:23:30 -0400444 gdtc.avail = global_dirtyable_memory();
445 domain_dirty_limits(&gdtc);
Johannes Weinerccafa282012-01-10 15:07:44 -0800446
Tejun Heo9fc3a432015-05-22 18:23:30 -0400447 *pbackground = gdtc.bg_thresh;
448 *pdirty = gdtc.thresh;
Johannes Weinerccafa282012-01-10 15:07:44 -0800449}
450
Johannes Weinera756cf52012-01-10 15:07:49 -0800451/**
Mel Gorman281e3722016-07-28 15:46:11 -0700452 * node_dirty_limit - maximum number of dirty pages allowed in a node
453 * @pgdat: the node
Johannes Weinera756cf52012-01-10 15:07:49 -0800454 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800455 * Return: the maximum number of dirty pages allowed in a node, based
Mel Gorman281e3722016-07-28 15:46:11 -0700456 * on the node's dirtyable memory.
Johannes Weinera756cf52012-01-10 15:07:49 -0800457 */
Mel Gorman281e3722016-07-28 15:46:11 -0700458static unsigned long node_dirty_limit(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800459{
Mel Gorman281e3722016-07-28 15:46:11 -0700460 unsigned long node_memory = node_dirtyable_memory(pgdat);
Johannes Weinera756cf52012-01-10 15:07:49 -0800461 struct task_struct *tsk = current;
462 unsigned long dirty;
463
464 if (vm_dirty_bytes)
465 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
Mel Gorman281e3722016-07-28 15:46:11 -0700466 node_memory / global_dirtyable_memory();
Johannes Weinera756cf52012-01-10 15:07:49 -0800467 else
Mel Gorman281e3722016-07-28 15:46:11 -0700468 dirty = vm_dirty_ratio * node_memory / 100;
Johannes Weinera756cf52012-01-10 15:07:49 -0800469
NeilBrowna37b0712020-06-01 21:48:18 -0700470 if (rt_task(tsk))
Johannes Weinera756cf52012-01-10 15:07:49 -0800471 dirty += dirty / 4;
472
473 return dirty;
474}
475
476/**
Mel Gorman281e3722016-07-28 15:46:11 -0700477 * node_dirty_ok - tells whether a node is within its dirty limits
478 * @pgdat: the node to check
Johannes Weinera756cf52012-01-10 15:07:49 -0800479 *
Mike Rapoporta862f682019-03-05 15:48:42 -0800480 * Return: %true when the dirty pages in @pgdat are within the node's
Johannes Weinera756cf52012-01-10 15:07:49 -0800481 * dirty limit, %false if the limit is exceeded.
482 */
Mel Gorman281e3722016-07-28 15:46:11 -0700483bool node_dirty_ok(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800484{
Mel Gorman281e3722016-07-28 15:46:11 -0700485 unsigned long limit = node_dirty_limit(pgdat);
486 unsigned long nr_pages = 0;
Johannes Weinera756cf52012-01-10 15:07:49 -0800487
Mel Gorman11fb9982016-07-28 15:46:20 -0700488 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
Mel Gorman11fb9982016-07-28 15:46:20 -0700489 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
Mel Gorman281e3722016-07-28 15:46:11 -0700490
491 return nr_pages <= limit;
Johannes Weinera756cf52012-01-10 15:07:49 -0800492}
493
David Rientjes2da02992009-01-06 14:39:31 -0800494int dirty_background_ratio_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +0200495 void *buffer, size_t *lenp, loff_t *ppos)
David Rientjes2da02992009-01-06 14:39:31 -0800496{
497 int ret;
498
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700499 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800500 if (ret == 0 && write)
501 dirty_background_bytes = 0;
502 return ret;
503}
504
505int dirty_background_bytes_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +0200506 void *buffer, size_t *lenp, loff_t *ppos)
David Rientjes2da02992009-01-06 14:39:31 -0800507{
508 int ret;
509
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700510 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800511 if (ret == 0 && write)
512 dirty_background_ratio = 0;
513 return ret;
514}
515
Christoph Hellwig32927392020-04-24 08:43:38 +0200516int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
517 size_t *lenp, loff_t *ppos)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700518{
519 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800520 int ret;
521
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700522 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700523 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200524 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800525 vm_dirty_bytes = 0;
526 }
527 return ret;
528}
529
David Rientjes2da02992009-01-06 14:39:31 -0800530int dirty_bytes_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +0200531 void *buffer, size_t *lenp, loff_t *ppos)
David Rientjes2da02992009-01-06 14:39:31 -0800532{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800533 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800534 int ret;
535
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700536 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800537 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200538 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800539 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700540 }
541 return ret;
542}
543
Jan Karaeb608e32012-05-24 18:59:11 +0200544static unsigned long wp_next_time(unsigned long cur_time)
545{
546 cur_time += VM_COMPLETIONS_PERIOD_LEN;
547 /* 0 has a special meaning... */
548 if (!cur_time)
549 return 1;
550 return cur_time;
551}
552
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400553static void wb_domain_writeout_add(struct wb_domain *dom,
Tejun Heoc7981432015-05-22 18:23:29 -0400554 struct fprop_local_percpu *completions,
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400555 unsigned int max_prop_frac, long nr)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700556{
Matthew Wilcox (Oracle)be5f1792021-03-20 16:34:54 -0400557 __fprop_add_percpu_max(&dom->completions, completions,
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400558 max_prop_frac, nr);
Jan Karaeb608e32012-05-24 18:59:11 +0200559 /* First event after period switching was turned off? */
Steven Rostedt (VMware)517663e2017-02-24 14:59:24 -0800560 if (unlikely(!dom->period_time)) {
Jan Karaeb608e32012-05-24 18:59:11 +0200561 /*
562 * We can race with other __bdi_writeout_inc calls here but
563 * it does not cause any harm since the resulting time when
564 * timer will fire and what is in writeout_period_time will be
565 * roughly the same.
566 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400567 dom->period_time = wp_next_time(jiffies);
568 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200569 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700570}
571
Tejun Heoc7981432015-05-22 18:23:29 -0400572/*
573 * Increment @wb's writeout completion count and the global writeout
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -0500574 * completion count. Called from __folio_end_writeback().
Tejun Heoc7981432015-05-22 18:23:29 -0400575 */
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400576static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
Tejun Heoc7981432015-05-22 18:23:29 -0400577{
Tejun Heo841710a2015-05-22 18:23:33 -0400578 struct wb_domain *cgdom;
579
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400580 wb_stat_mod(wb, WB_WRITTEN, nr);
581 wb_domain_writeout_add(&global_wb_domain, &wb->completions,
582 wb->bdi->max_prop_frac, nr);
Tejun Heo841710a2015-05-22 18:23:33 -0400583
584 cgdom = mem_cgroup_wb_domain(wb);
585 if (cgdom)
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400586 wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
587 wb->bdi->max_prop_frac, nr);
Tejun Heoc7981432015-05-22 18:23:29 -0400588}
589
Tejun Heo93f78d82015-05-22 17:13:27 -0400590void wb_writeout_inc(struct bdi_writeback *wb)
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700591{
592 unsigned long flags;
593
594 local_irq_save(flags);
Matthew Wilcox (Oracle)cc24df42021-03-20 16:45:15 -0400595 __wb_writeout_add(wb, 1);
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700596 local_irq_restore(flags);
597}
Tejun Heo93f78d82015-05-22 17:13:27 -0400598EXPORT_SYMBOL_GPL(wb_writeout_inc);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700599
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700600/*
Jan Karaeb608e32012-05-24 18:59:11 +0200601 * On idle system, we can be called long after we scheduled because we use
602 * deferred timers so count with missed periods.
603 */
Kees Cook9823e512017-11-15 17:37:04 -0800604static void writeout_period(struct timer_list *t)
Jan Karaeb608e32012-05-24 18:59:11 +0200605{
Kees Cook9823e512017-11-15 17:37:04 -0800606 struct wb_domain *dom = from_timer(dom, t, period_timer);
Tejun Heo380c27c2015-05-22 18:23:21 -0400607 int miss_periods = (jiffies - dom->period_time) /
Jan Karaeb608e32012-05-24 18:59:11 +0200608 VM_COMPLETIONS_PERIOD_LEN;
609
Tejun Heo380c27c2015-05-22 18:23:21 -0400610 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
611 dom->period_time = wp_next_time(dom->period_time +
Jan Karaeb608e32012-05-24 18:59:11 +0200612 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
Tejun Heo380c27c2015-05-22 18:23:21 -0400613 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200614 } else {
615 /*
616 * Aging has zeroed all fractions. Stop wasting CPU on period
617 * updates.
618 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400619 dom->period_time = 0;
Jan Karaeb608e32012-05-24 18:59:11 +0200620 }
621}
622
Tejun Heo380c27c2015-05-22 18:23:21 -0400623int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
624{
625 memset(dom, 0, sizeof(*dom));
Tejun Heodcc25ae2015-05-22 18:23:22 -0400626
627 spin_lock_init(&dom->lock);
628
Kees Cook9823e512017-11-15 17:37:04 -0800629 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
Tejun Heodcc25ae2015-05-22 18:23:22 -0400630
631 dom->dirty_limit_tstamp = jiffies;
632
Tejun Heo380c27c2015-05-22 18:23:21 -0400633 return fprop_global_init(&dom->completions, gfp);
634}
635
Tejun Heo841710a2015-05-22 18:23:33 -0400636#ifdef CONFIG_CGROUP_WRITEBACK
637void wb_domain_exit(struct wb_domain *dom)
638{
639 del_timer_sync(&dom->period_timer);
640 fprop_global_destroy(&dom->completions);
641}
642#endif
643
Jan Karaeb608e32012-05-24 18:59:11 +0200644/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700645 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
646 * registered backing devices, which, for obvious reasons, can not
647 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700648 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700649static unsigned int bdi_min_ratio;
650
651int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
652{
653 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700654
Jens Axboecfc4ba52009-09-14 13:12:40 +0200655 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700656 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700657 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700658 } else {
659 min_ratio -= bdi->min_ratio;
660 if (bdi_min_ratio + min_ratio < 100) {
661 bdi_min_ratio += min_ratio;
662 bdi->min_ratio += min_ratio;
663 } else {
664 ret = -EINVAL;
665 }
666 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200667 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700668
669 return ret;
670}
671
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700672int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
673{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700674 int ret = 0;
675
676 if (max_ratio > 100)
677 return -EINVAL;
678
Jens Axboecfc4ba52009-09-14 13:12:40 +0200679 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700680 if (bdi->min_ratio > max_ratio) {
681 ret = -EINVAL;
682 } else {
683 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200684 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700685 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200686 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700687
688 return ret;
689}
690EXPORT_SYMBOL(bdi_set_max_ratio);
691
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600692static unsigned long dirty_freerun_ceiling(unsigned long thresh,
693 unsigned long bg_thresh)
694{
695 return (thresh + bg_thresh) / 2;
696}
697
Tejun Heoc7981432015-05-22 18:23:29 -0400698static unsigned long hard_dirty_limit(struct wb_domain *dom,
699 unsigned long thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -0600700{
Tejun Heodcc25ae2015-05-22 18:23:22 -0400701 return max(thresh, dom->dirty_limit);
Wu Fengguangffd1f602011-06-19 22:18:42 -0600702}
703
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400704/*
705 * Memory which can be further allocated to a memcg domain is capped by
706 * system-wide clean memory excluding the amount being used in the domain.
707 */
708static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
709 unsigned long filepages, unsigned long headroom)
Tejun Heoc2aa7232015-05-22 18:23:35 -0400710{
711 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400712 unsigned long clean = filepages - min(filepages, mdtc->dirty);
713 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
714 unsigned long other_clean = global_clean - min(global_clean, clean);
Tejun Heoc2aa7232015-05-22 18:23:35 -0400715
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400716 mdtc->avail = filepages + min(headroom, other_clean);
Christoph Lameter1b424462007-05-06 14:48:59 -0700717}
718
Wu Fengguang6f718652011-03-02 17:14:34 -0600719/**
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400720 * __wb_calc_thresh - @wb's share of dirty throttling threshold
721 * @dtc: dirty_throttle_context of interest
Wu Fengguang1babe182010-08-11 14:17:40 -0700722 *
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600723 * Note that balance_dirty_pages() will only seriously take it as a hard limit
724 * when sleeping max_pause per page is not enough to keep the dirty pages under
725 * control. For example, when the device is completely stalled due to some error
726 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
727 * In the other normal situations, it acts more gently by throttling the tasks
Tejun Heoa88a3412015-05-22 17:13:28 -0400728 * more (rather than completely block them) when the wb dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600729 *
730 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700731 * - starving fast devices
732 * - piling up dirty pages (that will take long time to sync) on slow devices
733 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400734 * The wb's share of dirty limit will be adapting to its throughput and
Wu Fengguang1babe182010-08-11 14:17:40 -0700735 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
Mike Rapoporta862f682019-03-05 15:48:42 -0800736 *
737 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
NeilBrown8d928902020-06-01 21:48:21 -0700738 * dirty balancing includes all PG_dirty and PG_writeback pages.
Wu Fengguang1babe182010-08-11 14:17:40 -0700739 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400740static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
Wu Fengguang16c40422010-08-11 14:17:39 -0700741{
Tejun Heoe9f07df2015-05-22 18:23:28 -0400742 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400743 unsigned long thresh = dtc->thresh;
Tejun Heo0d960a32015-05-22 18:23:19 -0400744 u64 wb_thresh;
Wen Yangd3ac9462020-01-13 16:29:26 -0800745 unsigned long numerator, denominator;
Tejun Heo693108a2015-05-22 17:13:49 -0400746 unsigned long wb_min_ratio, wb_max_ratio;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700747
Wu Fengguang16c40422010-08-11 14:17:39 -0700748 /*
Tejun Heo0d960a32015-05-22 18:23:19 -0400749 * Calculate this BDI's share of the thresh ratio.
Wu Fengguang16c40422010-08-11 14:17:39 -0700750 */
Tejun Heoe9770b32015-05-22 18:23:27 -0400751 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
Tejun Heo380c27c2015-05-22 18:23:21 -0400752 &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700753
Tejun Heo0d960a32015-05-22 18:23:19 -0400754 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
755 wb_thresh *= numerator;
Wen Yangd3ac9462020-01-13 16:29:26 -0800756 wb_thresh = div64_ul(wb_thresh, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700757
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400758 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
Wu Fengguang16c40422010-08-11 14:17:39 -0700759
Tejun Heo0d960a32015-05-22 18:23:19 -0400760 wb_thresh += (thresh * wb_min_ratio) / 100;
761 if (wb_thresh > (thresh * wb_max_ratio) / 100)
762 wb_thresh = thresh * wb_max_ratio / 100;
Wu Fengguang16c40422010-08-11 14:17:39 -0700763
Tejun Heo0d960a32015-05-22 18:23:19 -0400764 return wb_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765}
766
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400767unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
768{
769 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
770 .thresh = thresh };
771 return __wb_calc_thresh(&gdtc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772}
773
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600774/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700775 * setpoint - dirty 3
776 * f(dirty) := 1.0 + (----------------)
777 * limit - setpoint
778 *
779 * it's a 3rd order polynomial that subjects to
780 *
781 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
782 * (2) f(setpoint) = 1.0 => the balance point
783 * (3) f(limit) = 0 => the hard limit
784 * (4) df/dx <= 0 => negative feedback control
785 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
786 * => fast response on large errors; small oscillation near setpoint
787 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700788static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700789 unsigned long dirty,
790 unsigned long limit)
791{
792 long long pos_ratio;
793 long x;
794
Rik van Rield5c9fde2014-05-06 12:50:01 -0700795 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Tejun Heo464d1382015-04-21 16:49:13 -0400796 (limit - setpoint) | 1);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700797 pos_ratio = x;
798 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
799 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
800 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
801
802 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
803}
804
805/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600806 * Dirty position control.
807 *
808 * (o) global/bdi setpoints
809 *
Tejun Heode1fff32015-05-22 17:13:29 -0400810 * We want the dirty pages be balanced around the global/wb setpoints.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600811 * When the number of dirty pages is higher/lower than the setpoint, the
812 * dirty position control ratio (and hence task dirty ratelimit) will be
813 * decreased/increased to bring the dirty pages back to the setpoint.
814 *
815 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
816 *
817 * if (dirty < setpoint) scale up pos_ratio
818 * if (dirty > setpoint) scale down pos_ratio
819 *
Tejun Heode1fff32015-05-22 17:13:29 -0400820 * if (wb_dirty < wb_setpoint) scale up pos_ratio
821 * if (wb_dirty > wb_setpoint) scale down pos_ratio
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600822 *
823 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
824 *
825 * (o) global control line
826 *
827 * ^ pos_ratio
828 * |
829 * | |<===== global dirty control scope ======>|
Chi Wu03231552021-06-28 19:35:34 -0700830 * 2.0 * * * * * * *
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600831 * | .*
832 * | . *
833 * | . *
834 * | . *
835 * | . *
836 * | . *
837 * 1.0 ................................*
838 * | . . *
839 * | . . *
840 * | . . *
841 * | . . *
842 * | . . *
843 * 0 +------------.------------------.----------------------*------------->
844 * freerun^ setpoint^ limit^ dirty pages
845 *
Tejun Heode1fff32015-05-22 17:13:29 -0400846 * (o) wb control line
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600847 *
848 * ^ pos_ratio
849 * |
850 * | *
851 * | *
852 * | *
853 * | *
854 * | * |<=========== span ============>|
855 * 1.0 .......................*
856 * | . *
857 * | . *
858 * | . *
859 * | . *
860 * | . *
861 * | . *
862 * | . *
863 * | . *
864 * | . *
865 * | . *
866 * | . *
867 * 1/4 ...............................................* * * * * * * * * * * *
868 * | . .
869 * | . .
870 * | . .
871 * 0 +----------------------.-------------------------------.------------->
Tejun Heode1fff32015-05-22 17:13:29 -0400872 * wb_setpoint^ x_intercept^
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600873 *
Tejun Heode1fff32015-05-22 17:13:29 -0400874 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600875 * be smoothly throttled down to normal if it starts high in situations like
876 * - start writing to a slow SD card and a fast disk at the same time. The SD
Tejun Heode1fff32015-05-22 17:13:29 -0400877 * card's wb_dirty may rush to many times higher than wb_setpoint.
878 * - the wb dirty thresh drops quickly due to change of JBOD workload
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600879 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400880static void wb_position_ratio(struct dirty_throttle_control *dtc)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600881{
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400882 struct bdi_writeback *wb = dtc->wb;
Jan Kara20792eb2021-09-02 14:53:15 -0700883 unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400884 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -0400885 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400886 unsigned long wb_thresh = dtc->wb_thresh;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600887 unsigned long x_intercept;
888 unsigned long setpoint; /* dirty pages' target balance point */
Tejun Heode1fff32015-05-22 17:13:29 -0400889 unsigned long wb_setpoint;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600890 unsigned long span;
891 long long pos_ratio; /* for scaling up/down the rate limit */
892 long x;
893
Tejun Heodaddfa32015-05-22 18:23:26 -0400894 dtc->pos_ratio = 0;
895
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400896 if (unlikely(dtc->dirty >= limit))
Tejun Heodaddfa32015-05-22 18:23:26 -0400897 return;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600898
899 /*
900 * global setpoint
901 *
Maxim Patlasov5a537482013-09-11 14:22:46 -0700902 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600903 */
904 setpoint = (freerun + limit) / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400905 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700906
907 /*
908 * The strictlimit feature is a tool preventing mistrusted filesystems
909 * from growing a large number of dirty pages before throttling. For
Tejun Heode1fff32015-05-22 17:13:29 -0400910 * such filesystems balance_dirty_pages always checks wb counters
911 * against wb limits. Even if global "nr_dirty" is under "freerun".
Maxim Patlasov5a537482013-09-11 14:22:46 -0700912 * This is especially important for fuse which sets bdi->max_ratio to
913 * 1% by default. Without strictlimit feature, fuse writeback may
914 * consume arbitrary amount of RAM because it is accounted in
915 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
916 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400917 * Here, in wb_position_ratio(), we calculate pos_ratio based on
Tejun Heode1fff32015-05-22 17:13:29 -0400918 * two values: wb_dirty and wb_thresh. Let's consider an example:
Maxim Patlasov5a537482013-09-11 14:22:46 -0700919 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
920 * limits are set by default to 10% and 20% (background and throttle).
Tejun Heode1fff32015-05-22 17:13:29 -0400921 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
Tejun Heo0d960a32015-05-22 18:23:19 -0400922 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
Tejun Heode1fff32015-05-22 17:13:29 -0400923 * about ~6K pages (as the average of background and throttle wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700924 * limits). The 3rd order polynomial will provide positive feedback if
Tejun Heode1fff32015-05-22 17:13:29 -0400925 * wb_dirty is under wb_setpoint and vice versa.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700926 *
927 * Note, that we cannot use global counters in these calculations
Tejun Heode1fff32015-05-22 17:13:29 -0400928 * because we want to throttle process writing to a strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700929 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
930 * in the example above).
931 */
Tejun Heoa88a3412015-05-22 17:13:28 -0400932 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heode1fff32015-05-22 17:13:29 -0400933 long long wb_pos_ratio;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700934
Tejun Heodaddfa32015-05-22 18:23:26 -0400935 if (dtc->wb_dirty < 8) {
936 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
937 2 << RATELIMIT_CALC_SHIFT);
938 return;
939 }
Maxim Patlasov5a537482013-09-11 14:22:46 -0700940
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400941 if (dtc->wb_dirty >= wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400942 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700943
Tejun Heo970fb012015-05-22 18:23:24 -0400944 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
945 dtc->wb_bg_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700946
Tejun Heode1fff32015-05-22 17:13:29 -0400947 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400948 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700949
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400950 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
Tejun Heode1fff32015-05-22 17:13:29 -0400951 wb_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700952
953 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400954 * Typically, for strictlimit case, wb_setpoint << setpoint
955 * and pos_ratio >> wb_pos_ratio. In the other words global
Maxim Patlasov5a537482013-09-11 14:22:46 -0700956 * state ("dirty") is not limiting factor and we have to
Tejun Heode1fff32015-05-22 17:13:29 -0400957 * make decision based on wb counters. But there is an
Maxim Patlasov5a537482013-09-11 14:22:46 -0700958 * important case when global pos_ratio should get precedence:
959 * global limits are exceeded (e.g. due to activities on other
Tejun Heode1fff32015-05-22 17:13:29 -0400960 * wb's) while given strictlimit wb is below limit.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700961 *
Tejun Heode1fff32015-05-22 17:13:29 -0400962 * "pos_ratio * wb_pos_ratio" would work for the case above,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700963 * but it would look too non-natural for the case of all
Tejun Heode1fff32015-05-22 17:13:29 -0400964 * activity in the system coming from a single strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700965 * with bdi->max_ratio == 100%.
966 *
967 * Note that min() below somewhat changes the dynamics of the
968 * control system. Normally, pos_ratio value can be well over 3
Tejun Heode1fff32015-05-22 17:13:29 -0400969 * (when globally we are at freerun and wb is well below wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700970 * setpoint). Now the maximum pos_ratio in the same situation
971 * is 2. We might want to tweak this if we observe the control
972 * system is too slow to adapt.
973 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400974 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
975 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700976 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600977
978 /*
979 * We have computed basic pos_ratio above based on global situation. If
Tejun Heode1fff32015-05-22 17:13:29 -0400980 * the wb is over/under its share of dirty pages, we want to scale
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600981 * pos_ratio further down/up. That is done by the following mechanism.
982 */
983
984 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400985 * wb setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600986 *
Tejun Heode1fff32015-05-22 17:13:29 -0400987 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600988 *
Tejun Heode1fff32015-05-22 17:13:29 -0400989 * x_intercept - wb_dirty
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600990 * := --------------------------
Tejun Heode1fff32015-05-22 17:13:29 -0400991 * x_intercept - wb_setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600992 *
Tejun Heode1fff32015-05-22 17:13:29 -0400993 * The main wb control line is a linear function that subjects to
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600994 *
Tejun Heode1fff32015-05-22 17:13:29 -0400995 * (1) f(wb_setpoint) = 1.0
996 * (2) k = - 1 / (8 * write_bw) (in single wb case)
997 * or equally: x_intercept = wb_setpoint + 8 * write_bw
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600998 *
Tejun Heode1fff32015-05-22 17:13:29 -0400999 * For single wb case, the dirty pages are observed to fluctuate
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001000 * regularly within range
Tejun Heode1fff32015-05-22 17:13:29 -04001001 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001002 * for various filesystems, where (2) can yield in a reasonable 12.5%
1003 * fluctuation range for pos_ratio.
1004 *
Tejun Heode1fff32015-05-22 17:13:29 -04001005 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001006 * own size, so move the slope over accordingly and choose a slope that
Tejun Heode1fff32015-05-22 17:13:29 -04001007 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001008 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001009 if (unlikely(wb_thresh > dtc->thresh))
1010 wb_thresh = dtc->thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001011 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001012 * It's very possible that wb_thresh is close to 0 not because the
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001013 * device is slow, but that it has remained inactive for long time.
1014 * Honour such devices a reasonable good (hopefully IO efficient)
1015 * threshold, so that the occasional writes won't be blocked and active
1016 * writes can rampup the threshold quickly.
1017 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001018 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001019 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001020 * scale global setpoint to wb's:
1021 * wb_setpoint = setpoint * wb_thresh / thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001022 */
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001023 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
Tejun Heode1fff32015-05-22 17:13:29 -04001024 wb_setpoint = setpoint * (u64)x >> 16;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001025 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001026 * Use span=(8*write_bw) in single wb case as indicated by
1027 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001028 *
Tejun Heode1fff32015-05-22 17:13:29 -04001029 * wb_thresh thresh - wb_thresh
1030 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1031 * thresh thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001032 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001033 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
Tejun Heode1fff32015-05-22 17:13:29 -04001034 x_intercept = wb_setpoint + span;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001035
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001036 if (dtc->wb_dirty < x_intercept - span / 4) {
1037 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001038 (x_intercept - wb_setpoint) | 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001039 } else
1040 pos_ratio /= 4;
1041
Wu Fengguang8927f662011-08-04 22:16:46 -06001042 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001043 * wb reserve area, safeguard against dirty pool underrun and disk idle
Wu Fengguang8927f662011-08-04 22:16:46 -06001044 * It may push the desired control point of global dirty pages higher
1045 * than setpoint.
1046 */
Tejun Heode1fff32015-05-22 17:13:29 -04001047 x_intercept = wb_thresh / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001048 if (dtc->wb_dirty < x_intercept) {
1049 if (dtc->wb_dirty > x_intercept / 8)
1050 pos_ratio = div_u64(pos_ratio * x_intercept,
1051 dtc->wb_dirty);
Wu Fengguang50657fc2011-10-11 17:06:33 -06001052 else
Wu Fengguang8927f662011-08-04 22:16:46 -06001053 pos_ratio *= 8;
1054 }
1055
Tejun Heodaddfa32015-05-22 18:23:26 -04001056 dtc->pos_ratio = pos_ratio;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001057}
1058
Tejun Heoa88a3412015-05-22 17:13:28 -04001059static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1060 unsigned long elapsed,
1061 unsigned long written)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001062{
1063 const unsigned long period = roundup_pow_of_two(3 * HZ);
Tejun Heoa88a3412015-05-22 17:13:28 -04001064 unsigned long avg = wb->avg_write_bandwidth;
1065 unsigned long old = wb->write_bandwidth;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001066 u64 bw;
1067
1068 /*
1069 * bw = written * HZ / elapsed
1070 *
1071 * bw * elapsed + write_bandwidth * (period - elapsed)
1072 * write_bandwidth = ---------------------------------------------------
1073 * period
Tejun Heoc72efb62015-03-23 00:18:48 -04001074 *
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04001075 * @written may have decreased due to folio_account_redirty().
Tejun Heoc72efb62015-03-23 00:18:48 -04001076 * Avoid underflowing @bw calculation.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001077 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001078 bw = written - min(written, wb->written_stamp);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001079 bw *= HZ;
1080 if (unlikely(elapsed > period)) {
Wen Yang0a5d1a72020-01-13 16:29:29 -08001081 bw = div64_ul(bw, elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001082 avg = bw;
1083 goto out;
1084 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001085 bw += (u64)wb->write_bandwidth * (period - elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001086 bw >>= ilog2(period);
1087
1088 /*
1089 * one more level of smoothing, for filtering out sudden spikes
1090 */
1091 if (avg > old && old >= (unsigned long)bw)
1092 avg -= (avg - old) >> 3;
1093
1094 if (avg < old && old <= (unsigned long)bw)
1095 avg += (old - avg) >> 3;
1096
1097out:
Tejun Heo95a46c62015-05-22 17:13:47 -04001098 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1099 avg = max(avg, 1LU);
1100 if (wb_has_dirty_io(wb)) {
1101 long delta = avg - wb->avg_write_bandwidth;
1102 WARN_ON_ONCE(atomic_long_add_return(delta,
1103 &wb->bdi->tot_write_bandwidth) <= 0);
1104 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001105 wb->write_bandwidth = bw;
Jan Kara20792eb2021-09-02 14:53:15 -07001106 WRITE_ONCE(wb->avg_write_bandwidth, avg);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001107}
1108
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001109static void update_dirty_limit(struct dirty_throttle_control *dtc)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001110{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001111 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001112 unsigned long thresh = dtc->thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04001113 unsigned long limit = dom->dirty_limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001114
1115 /*
1116 * Follow up in one step.
1117 */
1118 if (limit < thresh) {
1119 limit = thresh;
1120 goto update;
1121 }
1122
1123 /*
1124 * Follow down slowly. Use the higher one as the target, because thresh
1125 * may drop below dirty. This is exactly the reason to introduce
Tejun Heodcc25ae2015-05-22 18:23:22 -04001126 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
Wu Fengguangc42843f2011-03-02 15:54:09 -06001127 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001128 thresh = max(thresh, dtc->dirty);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001129 if (limit > thresh) {
1130 limit -= (limit - thresh) >> 5;
1131 goto update;
1132 }
1133 return;
1134update:
Tejun Heodcc25ae2015-05-22 18:23:22 -04001135 dom->dirty_limit = limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001136}
1137
Jan Kara42dd2352021-09-02 14:53:12 -07001138static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1139 unsigned long now)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001140{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001141 struct wb_domain *dom = dtc_dom(dtc);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001142
1143 /*
1144 * check locklessly first to optimize away locking for the most time
1145 */
Tejun Heodcc25ae2015-05-22 18:23:22 -04001146 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
Wu Fengguangc42843f2011-03-02 15:54:09 -06001147 return;
1148
Tejun Heodcc25ae2015-05-22 18:23:22 -04001149 spin_lock(&dom->lock);
1150 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001151 update_dirty_limit(dtc);
Tejun Heodcc25ae2015-05-22 18:23:22 -04001152 dom->dirty_limit_tstamp = now;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001153 }
Tejun Heodcc25ae2015-05-22 18:23:22 -04001154 spin_unlock(&dom->lock);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001155}
1156
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001157/*
Tejun Heode1fff32015-05-22 17:13:29 -04001158 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001159 *
Tejun Heode1fff32015-05-22 17:13:29 -04001160 * Normal wb tasks will be curbed at or below it in long term.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001161 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1162 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001163static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
Tejun Heoa88a3412015-05-22 17:13:28 -04001164 unsigned long dirtied,
1165 unsigned long elapsed)
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001166{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001167 struct bdi_writeback *wb = dtc->wb;
1168 unsigned long dirty = dtc->dirty;
1169 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001170 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Wu Fengguang73811312011-08-26 15:53:24 -06001171 unsigned long setpoint = (freerun + limit) / 2;
Tejun Heoa88a3412015-05-22 17:13:28 -04001172 unsigned long write_bw = wb->avg_write_bandwidth;
1173 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001174 unsigned long dirty_rate;
1175 unsigned long task_ratelimit;
1176 unsigned long balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001177 unsigned long step;
1178 unsigned long x;
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001179 unsigned long shift;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001180
1181 /*
1182 * The dirty rate will match the writeout rate in long term, except
1183 * when dirty pages are truncated by userspace or re-dirtied by FS.
1184 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001185 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001186
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001187 /*
1188 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1189 */
1190 task_ratelimit = (u64)dirty_ratelimit *
Tejun Heodaddfa32015-05-22 18:23:26 -04001191 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001192 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1193
1194 /*
1195 * A linear estimation of the "balanced" throttle rate. The theory is,
Tejun Heode1fff32015-05-22 17:13:29 -04001196 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001197 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1198 * formula will yield the balanced rate limit (write_bw / N).
1199 *
1200 * Note that the expanded form is not a pure rate feedback:
1201 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1202 * but also takes pos_ratio into account:
1203 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1204 *
1205 * (1) is not realistic because pos_ratio also takes part in balancing
1206 * the dirty rate. Consider the state
1207 * pos_ratio = 0.5 (3)
1208 * rate = 2 * (write_bw / N) (4)
1209 * If (1) is used, it will stuck in that state! Because each dd will
1210 * be throttled at
1211 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1212 * yielding
1213 * dirty_rate = N * task_ratelimit = write_bw (6)
1214 * put (6) into (1) we get
1215 * rate_(i+1) = rate_(i) (7)
1216 *
1217 * So we end up using (2) to always keep
1218 * rate_(i+1) ~= (write_bw / N) (8)
1219 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1220 * pos_ratio is able to drive itself to 1.0, which is not only where
1221 * the dirty count meet the setpoint, but also where the slope of
1222 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1223 */
1224 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1225 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001226 /*
1227 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1228 */
1229 if (unlikely(balanced_dirty_ratelimit > write_bw))
1230 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001231
Wu Fengguang73811312011-08-26 15:53:24 -06001232 /*
1233 * We could safely do this and return immediately:
1234 *
Tejun Heode1fff32015-05-22 17:13:29 -04001235 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001236 *
1237 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001238 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001239 * limit the step size.
1240 *
1241 * The below code essentially only uses the relative value of
1242 *
1243 * task_ratelimit - dirty_ratelimit
1244 * = (pos_ratio - 1) * dirty_ratelimit
1245 *
1246 * which reflects the direction and size of dirty position error.
1247 */
1248
1249 /*
1250 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1251 * task_ratelimit is on the same side of dirty_ratelimit, too.
1252 * For example, when
1253 * - dirty_ratelimit > balanced_dirty_ratelimit
1254 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1255 * lowering dirty_ratelimit will help meet both the position and rate
1256 * control targets. Otherwise, don't update dirty_ratelimit if it will
1257 * only help meet the rate target. After all, what the users ultimately
1258 * feel and care are stable dirty rate and small position error.
1259 *
1260 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001261 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001262 * keeps jumping around randomly and can even leap far away at times
1263 * due to the small 200ms estimation period of dirty_rate (we want to
1264 * keep that period small to reduce time lags).
1265 */
1266 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001267
1268 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001269 * For strictlimit case, calculations above were based on wb counters
Tejun Heoa88a3412015-05-22 17:13:28 -04001270 * and limits (starting from pos_ratio = wb_position_ratio() and up to
Maxim Patlasov5a537482013-09-11 14:22:46 -07001271 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
Tejun Heode1fff32015-05-22 17:13:29 -04001272 * Hence, to calculate "step" properly, we have to use wb_dirty as
1273 * "dirty" and wb_setpoint as "setpoint".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001274 *
Tejun Heode1fff32015-05-22 17:13:29 -04001275 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1276 * it's possible that wb_thresh is close to zero due to inactivity
Tejun Heo970fb012015-05-22 18:23:24 -04001277 * of backing device.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001278 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001279 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001280 dirty = dtc->wb_dirty;
1281 if (dtc->wb_dirty < 8)
1282 setpoint = dtc->wb_dirty + 1;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001283 else
Tejun Heo970fb012015-05-22 18:23:24 -04001284 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001285 }
1286
Wu Fengguang73811312011-08-26 15:53:24 -06001287 if (dirty < setpoint) {
Tejun Heoa88a3412015-05-22 17:13:28 -04001288 x = min3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001289 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001290 if (dirty_ratelimit < x)
1291 step = x - dirty_ratelimit;
1292 } else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001293 x = max3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001294 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001295 if (dirty_ratelimit > x)
1296 step = dirty_ratelimit - x;
1297 }
1298
1299 /*
1300 * Don't pursue 100% rate matching. It's impossible since the balanced
1301 * rate itself is constantly fluctuating. So decrease the track speed
1302 * when it gets close to the target. Helps eliminate pointless tremors.
1303 */
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001304 shift = dirty_ratelimit / (2 * step + 1);
1305 if (shift < BITS_PER_LONG)
1306 step = DIV_ROUND_UP(step >> shift, 8);
1307 else
1308 step = 0;
Wu Fengguang73811312011-08-26 15:53:24 -06001309
1310 if (dirty_ratelimit < balanced_dirty_ratelimit)
1311 dirty_ratelimit += step;
1312 else
1313 dirty_ratelimit -= step;
1314
Jan Kara20792eb2021-09-02 14:53:15 -07001315 WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
Tejun Heoa88a3412015-05-22 17:13:28 -04001316 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001317
Tejun Heo5634cc22015-08-18 14:54:56 -07001318 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001319}
1320
Tejun Heoc2aa7232015-05-22 18:23:35 -04001321static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1322 struct dirty_throttle_control *mdtc,
Tejun Heo8a731792015-05-22 18:23:20 -04001323 bool update_ratelimit)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001324{
Tejun Heoc2aa7232015-05-22 18:23:35 -04001325 struct bdi_writeback *wb = gdtc->wb;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001326 unsigned long now = jiffies;
Jan Kara45a29662021-09-02 14:53:09 -07001327 unsigned long elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001328 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001329 unsigned long written;
1330
Jan Kara45a29662021-09-02 14:53:09 -07001331 spin_lock(&wb->list_lock);
Tejun Heo8a731792015-05-22 18:23:20 -04001332
Wu Fengguange98be2d2010-08-29 11:22:30 -06001333 /*
Jan Kara45a29662021-09-02 14:53:09 -07001334 * Lockless checks for elapsed time are racy and delayed update after
1335 * IO completion doesn't do it at all (to make sure written pages are
1336 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1337 * division errors.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001338 */
Jan Kara45a29662021-09-02 14:53:09 -07001339 elapsed = max(now - wb->bw_time_stamp, 1UL);
Tejun Heoa88a3412015-05-22 17:13:28 -04001340 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1341 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001342
Tejun Heo8a731792015-05-22 18:23:20 -04001343 if (update_ratelimit) {
Jan Kara42dd2352021-09-02 14:53:12 -07001344 domain_update_dirty_limit(gdtc, now);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001345 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1346
1347 /*
1348 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1349 * compiler has no way to figure that out. Help it.
1350 */
1351 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
Jan Kara42dd2352021-09-02 14:53:12 -07001352 domain_update_dirty_limit(mdtc, now);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001353 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1354 }
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001355 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001356 wb_update_write_bandwidth(wb, elapsed, written);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001357
Tejun Heoa88a3412015-05-22 17:13:28 -04001358 wb->dirtied_stamp = dirtied;
1359 wb->written_stamp = written;
Jan Kara20792eb2021-09-02 14:53:15 -07001360 WRITE_ONCE(wb->bw_time_stamp, now);
Jan Kara45a29662021-09-02 14:53:09 -07001361 spin_unlock(&wb->list_lock);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001362}
1363
Jan Kara45a29662021-09-02 14:53:09 -07001364void wb_update_bandwidth(struct bdi_writeback *wb)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001365{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001366 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1367
Jan Karafee468f2021-09-02 14:53:06 -07001368 __wb_update_bandwidth(&gdtc, NULL, false);
Jan Karafee468f2021-09-02 14:53:06 -07001369}
1370
1371/* Interval after which we consider wb idle and don't estimate bandwidth */
1372#define WB_BANDWIDTH_IDLE_JIF (HZ)
1373
1374static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1375{
1376 unsigned long now = jiffies;
1377 unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1378
1379 if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1380 !atomic_read(&wb->writeback_inodes)) {
1381 spin_lock(&wb->list_lock);
1382 wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1383 wb->written_stamp = wb_stat(wb, WB_WRITTEN);
Jan Kara20792eb2021-09-02 14:53:15 -07001384 WRITE_ONCE(wb->bw_time_stamp, now);
Jan Karafee468f2021-09-02 14:53:06 -07001385 spin_unlock(&wb->list_lock);
1386 }
Wu Fengguange98be2d2010-08-29 11:22:30 -06001387}
1388
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001390 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001391 * will look to see if it needs to start dirty throttling.
1392 *
1393 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
Michal Hockoc41f0122017-09-06 16:23:36 -07001394 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
Wu Fengguang9d823e82011-06-11 18:10:12 -06001395 * (the number of pages we may dirty without exceeding the dirty limits).
1396 */
1397static unsigned long dirty_poll_interval(unsigned long dirty,
1398 unsigned long thresh)
1399{
1400 if (thresh > dirty)
1401 return 1UL << (ilog2(thresh - dirty) >> 1);
1402
1403 return 1;
1404}
1405
Tejun Heoa88a3412015-05-22 17:13:28 -04001406static unsigned long wb_max_pause(struct bdi_writeback *wb,
Tejun Heode1fff32015-05-22 17:13:29 -04001407 unsigned long wb_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001408{
Jan Kara20792eb2021-09-02 14:53:15 -07001409 unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
Fengguang Wue3b6c652013-10-16 13:47:03 -07001410 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001411
1412 /*
1413 * Limit pause time for small memory systems. If sleeping for too long
1414 * time, a small pool of dirty/writeback pages may go empty and disk go
1415 * idle.
1416 *
1417 * 8 serves as the safety ratio.
1418 */
Tejun Heode1fff32015-05-22 17:13:29 -04001419 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001420 t++;
1421
Fengguang Wue3b6c652013-10-16 13:47:03 -07001422 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001423}
1424
Tejun Heoa88a3412015-05-22 17:13:28 -04001425static long wb_min_pause(struct bdi_writeback *wb,
1426 long max_pause,
1427 unsigned long task_ratelimit,
1428 unsigned long dirty_ratelimit,
1429 int *nr_dirtied_pause)
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001430{
Jan Kara20792eb2021-09-02 14:53:15 -07001431 long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1432 long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001433 long t; /* target pause */
1434 long pause; /* estimated next pause */
1435 int pages; /* target nr_dirtied_pause */
1436
1437 /* target for 10ms pause on 1-dd case */
1438 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001439
1440 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001441 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1442 * overheads.
1443 *
1444 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001445 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001446 if (hi > lo)
1447 t += (hi - lo) * (10 * HZ) / 1024;
1448
1449 /*
1450 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1451 * on the much more stable dirty_ratelimit. However the next pause time
1452 * will be computed based on task_ratelimit and the two rate limits may
1453 * depart considerably at some time. Especially if task_ratelimit goes
1454 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1455 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1456 * result task_ratelimit won't be executed faithfully, which could
1457 * eventually bring down dirty_ratelimit.
1458 *
1459 * We apply two rules to fix it up:
1460 * 1) try to estimate the next pause time and if necessary, use a lower
1461 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1462 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1463 * 2) limit the target pause time to max_pause/2, so that the normal
1464 * small fluctuations of task_ratelimit won't trigger rule (1) and
1465 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1466 */
1467 t = min(t, 1 + max_pause / 2);
1468 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1469
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001470 /*
1471 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1472 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1473 * When the 16 consecutive reads are often interrupted by some dirty
1474 * throttling pause during the async writes, cfq will go into idles
1475 * (deadline is fine). So push nr_dirtied_pause as high as possible
1476 * until reaches DIRTY_POLL_THRESH=32 pages.
1477 */
1478 if (pages < DIRTY_POLL_THRESH) {
1479 t = max_pause;
1480 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1481 if (pages > DIRTY_POLL_THRESH) {
1482 pages = DIRTY_POLL_THRESH;
1483 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1484 }
1485 }
1486
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001487 pause = HZ * pages / (task_ratelimit + 1);
1488 if (pause > max_pause) {
1489 t = max_pause;
1490 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1491 }
1492
1493 *nr_dirtied_pause = pages;
1494 /*
1495 * The minimal pause time will normally be half the target pause time.
1496 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001497 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001498}
1499
Tejun Heo970fb012015-05-22 18:23:24 -04001500static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
Maxim Patlasov5a537482013-09-11 14:22:46 -07001501{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001502 struct bdi_writeback *wb = dtc->wb;
Tejun Heo93f78d82015-05-22 17:13:27 -04001503 unsigned long wb_reclaimable;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001504
1505 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001506 * wb_thresh is not treated as some limiting factor as
Maxim Patlasov5a537482013-09-11 14:22:46 -07001507 * dirty_thresh, due to reasons
Tejun Heode1fff32015-05-22 17:13:29 -04001508 * - in JBOD setup, wb_thresh can fluctuate a lot
Maxim Patlasov5a537482013-09-11 14:22:46 -07001509 * - in a system with HDD and USB key, the USB key may somehow
Tejun Heode1fff32015-05-22 17:13:29 -04001510 * go into state (wb_dirty >> wb_thresh) either because
1511 * wb_dirty starts high, or because wb_thresh drops low.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001512 * In this case we don't want to hard throttle the USB key
Tejun Heode1fff32015-05-22 17:13:29 -04001513 * dirtiers for 100 seconds until wb_dirty drops under
1514 * wb_thresh. Instead the auxiliary wb control line in
Tejun Heoa88a3412015-05-22 17:13:28 -04001515 * wb_position_ratio() will let the dirtier task progress
Tejun Heode1fff32015-05-22 17:13:29 -04001516 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001517 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -04001518 dtc->wb_thresh = __wb_calc_thresh(dtc);
Tejun Heo970fb012015-05-22 18:23:24 -04001519 dtc->wb_bg_thresh = dtc->thresh ?
1520 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001521
1522 /*
1523 * In order to avoid the stacked BDI deadlock we need
1524 * to ensure we accurately count the 'dirty' pages when
1525 * the threshold is low.
1526 *
1527 * Otherwise it would be possible to get thresh+n pages
1528 * reported dirty, even though there are thresh-m pages
1529 * actually dirty; with m+n sitting in the percpu
1530 * deltas.
1531 */
Wang Long2bce7742017-11-15 17:39:03 -08001532 if (dtc->wb_thresh < 2 * wb_stat_error()) {
Tejun Heo93f78d82015-05-22 17:13:27 -04001533 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001534 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001535 } else {
Tejun Heo93f78d82015-05-22 17:13:27 -04001536 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001537 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001538 }
1539}
1540
Wu Fengguang9d823e82011-06-11 18:10:12 -06001541/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 * balance_dirty_pages() must be called by processes which are generating dirty
1543 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001544 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001545 * If we're over `background_thresh' then the writeback threads are woken to
1546 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 */
Tahsin Erdogan4c578dc2017-11-15 17:33:19 -08001548static void balance_dirty_pages(struct bdi_writeback *wb,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001549 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001551 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001552 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001553 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001554 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1555 &mdtc_stor : NULL;
1556 struct dirty_throttle_control *sdtc;
NeilBrown8d928902020-06-01 21:48:21 -07001557 unsigned long nr_reclaimable; /* = file_dirty */
Wu Fengguang83712352011-06-11 19:25:42 -06001558 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001559 long pause;
1560 long max_pause;
1561 long min_pause;
1562 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001563 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001564 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001565 unsigned long dirty_ratelimit;
Tejun Heodfb8ae52015-05-22 17:13:40 -04001566 struct backing_dev_info *bdi = wb->bdi;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001567 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001568 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569
1570 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001571 unsigned long now = jiffies;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001572 unsigned long dirty, thresh, bg_thresh;
Yang Shi50e55bf2015-11-20 15:57:10 -08001573 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1574 unsigned long m_thresh = 0;
1575 unsigned long m_bg_thresh = 0;
Wu Fengguang83712352011-06-11 19:25:42 -06001576
NeilBrown8d928902020-06-01 21:48:21 -07001577 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
Tejun Heo9fc3a432015-05-22 18:23:30 -04001578 gdtc->avail = global_dirtyable_memory();
Mel Gorman11fb9982016-07-28 15:46:20 -07001579 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001580
Tejun Heo9fc3a432015-05-22 18:23:30 -04001581 domain_dirty_limits(gdtc);
Wu Fengguang16c40422010-08-11 14:17:39 -07001582
Maxim Patlasov5a537482013-09-11 14:22:46 -07001583 if (unlikely(strictlimit)) {
Tejun Heo970fb012015-05-22 18:23:24 -04001584 wb_dirty_limits(gdtc);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001585
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001586 dirty = gdtc->wb_dirty;
1587 thresh = gdtc->wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -04001588 bg_thresh = gdtc->wb_bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001589 } else {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001590 dirty = gdtc->dirty;
1591 thresh = gdtc->thresh;
1592 bg_thresh = gdtc->bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001593 }
1594
Tejun Heoc2aa7232015-05-22 18:23:35 -04001595 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001596 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001597
1598 /*
1599 * If @wb belongs to !root memcg, repeat the same
1600 * basic calculations for the memcg domain.
1601 */
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001602 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1603 &mdtc->dirty, &writeback);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001604 mdtc->dirty += writeback;
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001605 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001606
1607 domain_dirty_limits(mdtc);
1608
1609 if (unlikely(strictlimit)) {
1610 wb_dirty_limits(mdtc);
1611 m_dirty = mdtc->wb_dirty;
1612 m_thresh = mdtc->wb_thresh;
1613 m_bg_thresh = mdtc->wb_bg_thresh;
1614 } else {
1615 m_dirty = mdtc->dirty;
1616 m_thresh = mdtc->thresh;
1617 m_bg_thresh = mdtc->bg_thresh;
1618 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001619 }
1620
1621 /*
1622 * Throttle it only when the background writeback cannot
1623 * catch-up. This avoids (excessively) small writeouts
Tejun Heode1fff32015-05-22 17:13:29 -04001624 * when the wb limits are ramping up in case of !strictlimit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001625 *
Tejun Heode1fff32015-05-22 17:13:29 -04001626 * In strictlimit case make decision based on the wb counters
1627 * and limits. Small writeouts when the wb limits are ramping
Maxim Patlasov5a537482013-09-11 14:22:46 -07001628 * up are the price we consciously pay for strictlimit-ing.
Tejun Heoc2aa7232015-05-22 18:23:35 -04001629 *
1630 * If memcg domain is in effect, @dirty should be under
1631 * both global and memcg freerun ceilings.
Wu Fengguang16c40422010-08-11 14:17:39 -07001632 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001633 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1634 (!mdtc ||
1635 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
NeilBrowna37b0712020-06-01 21:48:18 -07001636 unsigned long intv;
1637 unsigned long m_intv;
1638
1639free_running:
1640 intv = dirty_poll_interval(dirty, thresh);
1641 m_intv = ULONG_MAX;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001642
Wu Fengguang83712352011-06-11 19:25:42 -06001643 current->dirty_paused_when = now;
1644 current->nr_dirtied = 0;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001645 if (mdtc)
1646 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1647 current->nr_dirtied_pause = min(intv, m_intv);
Wu Fengguang16c40422010-08-11 14:17:39 -07001648 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001649 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001650
Tejun Heobc058732015-05-22 17:13:53 -04001651 if (unlikely(!writeback_in_progress(wb)))
Tejun Heo9ecf48662015-05-22 17:13:54 -04001652 wb_start_background_writeback(wb);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001653
Tejun Heo97b27822019-08-26 09:06:56 -07001654 mem_cgroup_flush_foreign(wb);
1655
Tejun Heoc2aa7232015-05-22 18:23:35 -04001656 /*
1657 * Calculate global domain's pos_ratio and select the
1658 * global dtc by default.
1659 */
NeilBrowna37b0712020-06-01 21:48:18 -07001660 if (!strictlimit) {
Tejun Heo970fb012015-05-22 18:23:24 -04001661 wb_dirty_limits(gdtc);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001662
NeilBrowna37b0712020-06-01 21:48:18 -07001663 if ((current->flags & PF_LOCAL_THROTTLE) &&
1664 gdtc->wb_dirty <
1665 dirty_freerun_ceiling(gdtc->wb_thresh,
1666 gdtc->wb_bg_thresh))
1667 /*
1668 * LOCAL_THROTTLE tasks must not be throttled
1669 * when below the per-wb freerun ceiling.
1670 */
1671 goto free_running;
1672 }
1673
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001674 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1675 ((gdtc->dirty > gdtc->thresh) || strictlimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676
Tejun Heodaddfa32015-05-22 18:23:26 -04001677 wb_position_ratio(gdtc);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001678 sdtc = gdtc;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001679
Tejun Heoc2aa7232015-05-22 18:23:35 -04001680 if (mdtc) {
1681 /*
1682 * If memcg domain is in effect, calculate its
1683 * pos_ratio. @wb should satisfy constraints from
1684 * both global and memcg domains. Choose the one
1685 * w/ lower pos_ratio.
1686 */
NeilBrowna37b0712020-06-01 21:48:18 -07001687 if (!strictlimit) {
Tejun Heoc2aa7232015-05-22 18:23:35 -04001688 wb_dirty_limits(mdtc);
1689
NeilBrowna37b0712020-06-01 21:48:18 -07001690 if ((current->flags & PF_LOCAL_THROTTLE) &&
1691 mdtc->wb_dirty <
1692 dirty_freerun_ceiling(mdtc->wb_thresh,
1693 mdtc->wb_bg_thresh))
1694 /*
1695 * LOCAL_THROTTLE tasks must not be
1696 * throttled when below the per-wb
1697 * freerun ceiling.
1698 */
1699 goto free_running;
1700 }
Tejun Heoc2aa7232015-05-22 18:23:35 -04001701 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1702 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1703
1704 wb_position_ratio(mdtc);
1705 if (mdtc->pos_ratio < gdtc->pos_ratio)
1706 sdtc = mdtc;
1707 }
Tejun Heodaddfa32015-05-22 18:23:26 -04001708
Tejun Heoa88a3412015-05-22 17:13:28 -04001709 if (dirty_exceeded && !wb->dirty_exceeded)
1710 wb->dirty_exceeded = 1;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001711
Jan Kara20792eb2021-09-02 14:53:15 -07001712 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
Jan Kara45a29662021-09-02 14:53:09 -07001713 BANDWIDTH_INTERVAL))
Jan Karafee468f2021-09-02 14:53:06 -07001714 __wb_update_bandwidth(gdtc, mdtc, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715
Tejun Heoc2aa7232015-05-22 18:23:35 -04001716 /* throttle according to the chosen dtc */
Jan Kara20792eb2021-09-02 14:53:15 -07001717 dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001718 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001719 RATELIMIT_CALC_SHIFT;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001720 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
Tejun Heoa88a3412015-05-22 17:13:28 -04001721 min_pause = wb_min_pause(wb, max_pause,
1722 task_ratelimit, dirty_ratelimit,
1723 &nr_dirtied_pause);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001724
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001725 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001726 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001727 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001728 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 }
Wu Fengguang83712352011-06-11 19:25:42 -06001730 period = HZ * pages_dirtied / task_ratelimit;
1731 pause = period;
1732 if (current->dirty_paused_when)
1733 pause -= now - current->dirty_paused_when;
1734 /*
1735 * For less than 1s think time (ext3/4 may block the dirtier
1736 * for up to 800ms from time to time on 1-HDD; so does xfs,
1737 * however at much less frequency), try to compensate it in
1738 * future periods by updating the virtual time; otherwise just
1739 * do a reset, as it may be a light dirtier.
1740 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001741 if (pause < min_pause) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001742 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001743 sdtc->thresh,
1744 sdtc->bg_thresh,
1745 sdtc->dirty,
1746 sdtc->wb_thresh,
1747 sdtc->wb_dirty,
Wu Fengguangece13ac32010-08-29 23:33:20 -06001748 dirty_ratelimit,
1749 task_ratelimit,
1750 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001751 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001752 min(pause, 0L),
Wu Fengguangece13ac32010-08-29 23:33:20 -06001753 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001754 if (pause < -HZ) {
1755 current->dirty_paused_when = now;
1756 current->nr_dirtied = 0;
1757 } else if (period) {
1758 current->dirty_paused_when += period;
1759 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001760 } else if (current->nr_dirtied_pause <= pages_dirtied)
1761 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001762 break;
1763 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001764 if (unlikely(pause > max_pause)) {
1765 /* for occasional dropped task_ratelimit */
1766 now += min(pause - max_pause, max_pause);
1767 pause = max_pause;
1768 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001769
1770pause:
Tejun Heo5634cc22015-08-18 14:54:56 -07001771 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001772 sdtc->thresh,
1773 sdtc->bg_thresh,
1774 sdtc->dirty,
1775 sdtc->wb_thresh,
1776 sdtc->wb_dirty,
Wu Fengguangece13ac32010-08-29 23:33:20 -06001777 dirty_ratelimit,
1778 task_ratelimit,
1779 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001780 period,
Wu Fengguangece13ac32010-08-29 23:33:20 -06001781 pause,
1782 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001783 __set_current_state(TASK_KILLABLE);
Jens Axboeb57d74a2016-09-01 10:20:33 -06001784 wb->dirty_sleep = now;
Wu Fengguangd25105e2009-10-09 12:40:42 +02001785 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001786
Wu Fengguang83712352011-06-11 19:25:42 -06001787 current->dirty_paused_when = now + pause;
1788 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001789 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001790
Wu Fengguangffd1f602011-06-19 22:18:42 -06001791 /*
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001792 * This is typically equal to (dirty < thresh) and can also
1793 * keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001794 */
Wu Fengguang1df647192011-11-13 19:47:32 -06001795 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001796 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001797
Wu Fengguangc5c63432011-12-02 10:21:33 -06001798 /*
Ingo Molnarf0953a12021-05-06 18:06:47 -07001799 * In the case of an unresponsive NFS server and the NFS dirty
Tejun Heode1fff32015-05-22 17:13:29 -04001800 * pages exceeds dirty_thresh, give the other good wb's a pipe
Wu Fengguangc5c63432011-12-02 10:21:33 -06001801 * to go through, so that tasks on them still remain responsive.
1802 *
Masahiro Yamada3f8b6fb2017-02-27 14:29:25 -08001803 * In theory 1 page is enough to keep the consumer-producer
Wu Fengguangc5c63432011-12-02 10:21:33 -06001804 * pipe going: the flusher cleans 1 page => the task dirties 1
Tejun Heode1fff32015-05-22 17:13:29 -04001805 * more page. However wb_dirty has accounting errors. So use
Tejun Heo93f78d82015-05-22 17:13:27 -04001806 * the larger and more IO friendly wb_stat_error.
Wu Fengguangc5c63432011-12-02 10:21:33 -06001807 */
Wang Long2bce7742017-11-15 17:39:03 -08001808 if (sdtc->wb_dirty <= wb_stat_error())
Wu Fengguangc5c63432011-12-02 10:21:33 -06001809 break;
1810
Jan Kara499d05e2011-11-16 19:34:48 +08001811 if (fatal_signal_pending(current))
1812 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 }
1814
Tejun Heoa88a3412015-05-22 17:13:28 -04001815 if (!dirty_exceeded && wb->dirty_exceeded)
1816 wb->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817
Tejun Heobc058732015-05-22 17:13:53 -04001818 if (writeback_in_progress(wb))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001819 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820
1821 /*
1822 * In laptop mode, we wait until hitting the higher threshold before
1823 * starting background writeout, and then write out all the way down
1824 * to the lower threshold. So slow writers cause minimal disk activity.
1825 *
1826 * In normal mode, we start background writeout at the lower
1827 * background_thresh, to keep the amount of dirty memory low.
1828 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001829 if (laptop_mode)
1830 return;
1831
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001832 if (nr_reclaimable > gdtc->bg_thresh)
Tejun Heo9ecf48662015-05-22 17:13:54 -04001833 wb_start_background_writeback(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834}
1835
Wu Fengguang9d823e82011-06-11 18:10:12 -06001836static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001837
Wu Fengguang54848d72011-04-05 13:21:19 -06001838/*
1839 * Normal tasks are throttled by
1840 * loop {
1841 * dirty tsk->nr_dirtied_pause pages;
1842 * take a snap in balance_dirty_pages();
1843 * }
1844 * However there is a worst case. If every task exit immediately when dirtied
1845 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1846 * called to throttle the page dirties. The solution is to save the not yet
1847 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1848 * randomly into the running tasks. This works well for the above worst case,
1849 * as the new task will pick up and accumulate the old task's leaked dirty
1850 * count and eventually get throttled.
1851 */
1852DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1853
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001855 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001856 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 *
1858 * Processes which are dirtying memory should call in here once for each page
1859 * which was newly dirtied. The function will periodically check the system's
1860 * dirty state and will initiate writeback if needed.
1861 *
Kefeng Wang5defd492021-06-28 19:35:28 -07001862 * Once we're over the dirty memory limit we decrease the ratelimiting
1863 * by a lot, to prevent individual processes from overshooting the limit
1864 * by (ratelimit_pages) each.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001866void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867{
Tejun Heodfb8ae52015-05-22 17:13:40 -04001868 struct inode *inode = mapping->host;
1869 struct backing_dev_info *bdi = inode_to_bdi(inode);
1870 struct bdi_writeback *wb = NULL;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001871 int ratelimit;
1872 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873
Christoph Hellwigf56753a2020-09-24 08:51:40 +02001874 if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
Wu Fengguang36715ce2011-06-11 17:53:57 -06001875 return;
1876
Tejun Heodfb8ae52015-05-22 17:13:40 -04001877 if (inode_cgwb_enabled(inode))
1878 wb = wb_get_create_current(bdi, GFP_KERNEL);
1879 if (!wb)
1880 wb = &bdi->wb;
1881
Wu Fengguang9d823e82011-06-11 18:10:12 -06001882 ratelimit = current->nr_dirtied_pause;
Tejun Heoa88a3412015-05-22 17:13:28 -04001883 if (wb->dirty_exceeded)
Wu Fengguang9d823e82011-06-11 18:10:12 -06001884 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001886 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001887 /*
1888 * This prevents one CPU to accumulate too many dirtied pages without
1889 * calling into balance_dirty_pages(), which can happen when there are
1890 * 1000+ tasks, all of them start dirtying pages at exactly the same
1891 * time, hence all honoured too large initial task->nr_dirtied_pause.
1892 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001893 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001894 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001895 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001896 else if (unlikely(*p >= ratelimit_pages)) {
1897 *p = 0;
1898 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001900 /*
1901 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1902 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1903 * the dirty throttling and livelock other long-run dirtiers.
1904 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001905 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06001906 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001907 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001908 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1909 *p -= nr_pages_dirtied;
1910 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001912 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001913
1914 if (unlikely(current->nr_dirtied >= ratelimit))
Tahsin Erdogan4c578dc2017-11-15 17:33:19 -08001915 balance_dirty_pages(wb, current->nr_dirtied);
Tejun Heodfb8ae52015-05-22 17:13:40 -04001916
1917 wb_put(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001919EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920
Tejun Heoaa661bb2015-05-22 18:23:31 -04001921/**
1922 * wb_over_bg_thresh - does @wb need to be written back?
1923 * @wb: bdi_writeback of interest
1924 *
1925 * Determines whether background writeback should keep writing @wb or it's
Mike Rapoporta862f682019-03-05 15:48:42 -08001926 * clean enough.
1927 *
1928 * Return: %true if writeback should continue.
Tejun Heoaa661bb2015-05-22 18:23:31 -04001929 */
1930bool wb_over_bg_thresh(struct bdi_writeback *wb)
1931{
Tejun Heo947e9762015-05-22 18:23:32 -04001932 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001933 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo947e9762015-05-22 18:23:32 -04001934 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001935 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1936 &mdtc_stor : NULL;
Chi Wuab199392021-06-28 19:35:31 -07001937 unsigned long reclaimable;
1938 unsigned long thresh;
Tejun Heoaa661bb2015-05-22 18:23:31 -04001939
Tejun Heo947e9762015-05-22 18:23:32 -04001940 /*
1941 * Similar to balance_dirty_pages() but ignores pages being written
1942 * as we're trying to decide whether to put more under writeback.
1943 */
1944 gdtc->avail = global_dirtyable_memory();
NeilBrown8d928902020-06-01 21:48:21 -07001945 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
Tejun Heo947e9762015-05-22 18:23:32 -04001946 domain_dirty_limits(gdtc);
Tejun Heoaa661bb2015-05-22 18:23:31 -04001947
Tejun Heo947e9762015-05-22 18:23:32 -04001948 if (gdtc->dirty > gdtc->bg_thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04001949 return true;
1950
Chi Wuab199392021-06-28 19:35:31 -07001951 thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
1952 if (thresh < 2 * wb_stat_error())
1953 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1954 else
1955 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1956
1957 if (reclaimable > thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04001958 return true;
1959
Tejun Heoc2aa7232015-05-22 18:23:35 -04001960 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001961 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001962
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001963 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1964 &writeback);
1965 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001966 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1967
1968 if (mdtc->dirty > mdtc->bg_thresh)
1969 return true;
1970
Chi Wuab199392021-06-28 19:35:31 -07001971 thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
1972 if (thresh < 2 * wb_stat_error())
1973 reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1974 else
1975 reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1976
1977 if (reclaimable > thresh)
Tejun Heoc2aa7232015-05-22 18:23:35 -04001978 return true;
1979 }
1980
Tejun Heoaa661bb2015-05-22 18:23:31 -04001981 return false;
1982}
1983
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1986 */
Joe Perchescccad5b2014-06-06 14:38:09 -07001987int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02001988 void *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989{
Yafang Shao94af5842017-10-10 05:44:13 +08001990 unsigned int old_interval = dirty_writeback_interval;
1991 int ret;
1992
1993 ret = proc_dointvec(table, write, buffer, length, ppos);
Yafang Shao515c24c2017-10-14 16:38:27 +08001994
1995 /*
1996 * Writing 0 to dirty_writeback_interval will disable periodic writeback
1997 * and a different non-zero value will wakeup the writeback threads.
1998 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1999 * iterate over all bdis and wbs.
2000 * The reason we do this is to make the change take effect immediately.
2001 */
2002 if (!ret && write && dirty_writeback_interval &&
2003 dirty_writeback_interval != old_interval)
Yafang Shao94af5842017-10-10 05:44:13 +08002004 wakeup_flusher_threads(WB_REASON_PERIODIC);
2005
2006 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007}
2008
Kees Cookbca237a2017-08-28 15:03:41 -07002009void laptop_mode_timer_fn(struct timer_list *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010{
Kees Cookbca237a2017-08-28 15:03:41 -07002011 struct backing_dev_info *backing_dev_info =
2012 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013
Kees Cookbca237a2017-08-28 15:03:41 -07002014 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015}
2016
2017/*
2018 * We've spun up the disk and we're in laptop mode: schedule writeback
2019 * of all dirty data a few seconds from now. If the flush is already scheduled
2020 * then push it back - the user is still using the disk.
2021 */
Matthew Garrett31373d02010-04-06 14:25:14 +02002022void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023{
Matthew Garrett31373d02010-04-06 14:25:14 +02002024 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025}
2026
2027/*
2028 * We're in laptop mode and we've just synced. The sync's writes will have
2029 * caused another writeback to be scheduled by laptop_io_completion.
2030 * Nothing needs to be written back anymore, so we unschedule the writeback.
2031 */
2032void laptop_sync_completion(void)
2033{
Matthew Garrett31373d02010-04-06 14:25:14 +02002034 struct backing_dev_info *bdi;
2035
2036 rcu_read_lock();
2037
2038 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2039 del_timer(&bdi->laptop_mode_wb_timer);
2040
2041 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042}
2043
2044/*
2045 * If ratelimit_pages is too high then we can get into dirty-data overload
2046 * if a large number of processes all perform writes at the same time.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 *
2048 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2049 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06002050 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 */
2052
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002053void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054{
Tejun Heodcc25ae2015-05-22 18:23:22 -04002055 struct wb_domain *dom = &global_wb_domain;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002056 unsigned long background_thresh;
2057 unsigned long dirty_thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04002058
Wu Fengguang9d823e82011-06-11 18:10:12 -06002059 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heodcc25ae2015-05-22 18:23:22 -04002060 dom->dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002061 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 if (ratelimit_pages < 16)
2063 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064}
2065
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002066static int page_writeback_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067{
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002068 writeback_set_ratelimit();
2069 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070}
2071
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002073 * Called early on to tune the page writeback dirty limits.
2074 *
2075 * We used to scale dirty pages according to how total memory
David Hildenbrand0a18e602020-08-06 23:25:27 -07002076 * related to pages that could be allocated for buffers.
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002077 *
2078 * However, that was when we used "dirty_ratio" to scale with
2079 * all memory, and we don't do that any more. "dirty_ratio"
David Hildenbrand0a18e602020-08-06 23:25:27 -07002080 * is now applied to total non-HIGHPAGE memory, and as such we can't
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002081 * get into the old insane situation any more where we had
2082 * large amounts of dirty pages compared to a small amount of
2083 * non-HIGHMEM memory.
2084 *
2085 * But we might still want to scale the dirty_ratio by how
2086 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 */
2088void __init page_writeback_init(void)
2089{
Rabin Vincenta50fcb52015-08-06 15:47:14 -07002090 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2091
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002092 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2093 page_writeback_cpu_online, NULL);
2094 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2095 page_writeback_cpu_online);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096}
2097
David Howells811d7362006-08-29 19:06:09 +01002098/**
Jan Karaf446daae2010-08-09 17:19:12 -07002099 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2100 * @mapping: address space structure to write
2101 * @start: starting page index
2102 * @end: ending page index (inclusive)
2103 *
2104 * This function scans the page range from @start to @end (inclusive) and tags
2105 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2106 * that write_cache_pages (or whoever calls this function) will then use
2107 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2108 * used to avoid livelocking of writeback by a process steadily creating new
2109 * dirty pages in the file (thus it is important for this function to be quick
2110 * so that it can tag pages faster than a dirtying process can create them).
2111 */
Jan Karaf446daae2010-08-09 17:19:12 -07002112void tag_pages_for_writeback(struct address_space *mapping,
2113 pgoff_t start, pgoff_t end)
2114{
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002115 XA_STATE(xas, &mapping->i_pages, start);
2116 unsigned int tagged = 0;
2117 void *page;
Jan Karaf446daae2010-08-09 17:19:12 -07002118
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002119 xas_lock_irq(&xas);
2120 xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2121 xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2122 if (++tagged % XA_CHECK_SCHED)
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002123 continue;
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002124
2125 xas_pause(&xas);
2126 xas_unlock_irq(&xas);
Jan Karaf446daae2010-08-09 17:19:12 -07002127 cond_resched();
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002128 xas_lock_irq(&xas);
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002129 }
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002130 xas_unlock_irq(&xas);
Jan Karaf446daae2010-08-09 17:19:12 -07002131}
2132EXPORT_SYMBOL(tag_pages_for_writeback);
2133
2134/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002135 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
David Howells811d7362006-08-29 19:06:09 +01002136 * @mapping: address space structure to write
2137 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002138 * @writepage: function called for each page
2139 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01002140 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002141 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01002142 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2143 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2144 * and msync() need to guarantee that all the data which was dirty at the time
2145 * the call was made get new I/O started against them. If wbc->sync_mode is
2146 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2147 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07002148 *
2149 * To avoid livelocks (when other process dirties new pages), we first tag
2150 * pages which should be written back with TOWRITE tag and only then start
2151 * writing them. For data-integrity sync we have to be careful so that we do
2152 * not miss some pages (e.g., because some other process has cleared TOWRITE
2153 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2154 * by the process clearing the DIRTY tag (and submitting the page for IO).
Dave Chinner64081362018-10-26 15:09:45 -07002155 *
2156 * To avoid deadlocks between range_cyclic writeback and callers that hold
2157 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2158 * we do not loop back to the start of the file. Doing so causes a page
2159 * lock/page writeback access order inversion - we should only ever lock
2160 * multiple pages in ascending page->index order, and looping back to the start
2161 * of the file violates that rule and causes deadlocks.
Mike Rapoporta862f682019-03-05 15:48:42 -08002162 *
2163 * Return: %0 on success, negative error code otherwise
David Howells811d7362006-08-29 19:06:09 +01002164 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002165int write_cache_pages(struct address_space *mapping,
2166 struct writeback_control *wbc, writepage_t writepage,
2167 void *data)
David Howells811d7362006-08-29 19:06:09 +01002168{
David Howells811d7362006-08-29 19:06:09 +01002169 int ret = 0;
2170 int done = 0;
Brian Foster3fa750d2018-12-28 00:37:20 -08002171 int error;
David Howells811d7362006-08-29 19:06:09 +01002172 struct pagevec pvec;
2173 int nr_pages;
2174 pgoff_t index;
2175 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08002176 pgoff_t done_index;
David Howells811d7362006-08-29 19:06:09 +01002177 int range_whole = 0;
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002178 xa_mark_t tag;
David Howells811d7362006-08-29 19:06:09 +01002179
Mel Gorman86679822017-11-15 17:37:52 -08002180 pagevec_init(&pvec);
David Howells811d7362006-08-29 19:06:09 +01002181 if (wbc->range_cyclic) {
Chao Yu28659cc2020-06-01 21:48:15 -07002182 index = mapping->writeback_index; /* prev offset */
David Howells811d7362006-08-29 19:06:09 +01002183 end = -1;
2184 } else {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002185 index = wbc->range_start >> PAGE_SHIFT;
2186 end = wbc->range_end >> PAGE_SHIFT;
David Howells811d7362006-08-29 19:06:09 +01002187 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2188 range_whole = 1;
David Howells811d7362006-08-29 19:06:09 +01002189 }
Mauricio Faria de Oliveiracc7b8f62020-04-01 21:04:43 -07002190 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
Jan Karaf446daae2010-08-09 17:19:12 -07002191 tag_pages_for_writeback(mapping, index, end);
Mauricio Faria de Oliveiracc7b8f62020-04-01 21:04:43 -07002192 tag = PAGECACHE_TAG_TOWRITE;
2193 } else {
2194 tag = PAGECACHE_TAG_DIRTY;
2195 }
Nick Pigginbd19e012009-01-06 14:39:06 -08002196 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002197 while (!done && (index <= end)) {
2198 int i;
2199
Jan Kara2b9775a2017-11-15 17:35:09 -08002200 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
Jan Kara67fd7072017-11-15 17:35:19 -08002201 tag);
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002202 if (nr_pages == 0)
2203 break;
David Howells811d7362006-08-29 19:06:09 +01002204
David Howells811d7362006-08-29 19:06:09 +01002205 for (i = 0; i < nr_pages; i++) {
2206 struct page *page = pvec.pages[i];
2207
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002208 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08002209
David Howells811d7362006-08-29 19:06:09 +01002210 lock_page(page);
2211
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002212 /*
2213 * Page truncated or invalidated. We can freely skip it
2214 * then, even for data integrity operations: the page
2215 * has disappeared concurrently, so there could be no
Ingo Molnarf0953a12021-05-06 18:06:47 -07002216 * real expectation of this data integrity operation
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002217 * even if there is now a new, dirty page at the same
2218 * pagecache address.
2219 */
David Howells811d7362006-08-29 19:06:09 +01002220 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002221continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01002222 unlock_page(page);
2223 continue;
2224 }
2225
Nick Piggin515f4a02009-01-06 14:39:10 -08002226 if (!PageDirty(page)) {
2227 /* someone wrote it for us */
2228 goto continue_unlock;
2229 }
David Howells811d7362006-08-29 19:06:09 +01002230
Nick Piggin515f4a02009-01-06 14:39:10 -08002231 if (PageWriteback(page)) {
2232 if (wbc->sync_mode != WB_SYNC_NONE)
2233 wait_on_page_writeback(page);
2234 else
2235 goto continue_unlock;
2236 }
2237
2238 BUG_ON(PageWriteback(page));
2239 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002240 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01002241
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002242 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
Brian Foster3fa750d2018-12-28 00:37:20 -08002243 error = (*writepage)(page, wbc, data);
2244 if (unlikely(error)) {
2245 /*
2246 * Handle errors according to the type of
2247 * writeback. There's no need to continue for
2248 * background writeback. Just push done_index
2249 * past this page so media errors won't choke
2250 * writeout for the entire file. For integrity
2251 * writeback, we must process the entire dirty
2252 * set regardless of errors because the fs may
2253 * still have state to clear for each page. In
2254 * that case we continue processing and return
2255 * the first error.
2256 */
2257 if (error == AOP_WRITEPAGE_ACTIVATE) {
Nick Piggin00266772009-01-06 14:39:06 -08002258 unlock_page(page);
Brian Foster3fa750d2018-12-28 00:37:20 -08002259 error = 0;
2260 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2261 ret = error;
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002262 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08002263 done = 1;
2264 break;
2265 }
Brian Foster3fa750d2018-12-28 00:37:20 -08002266 if (!ret)
2267 ret = error;
Dave Chinner0b564922010-06-09 10:37:18 +10002268 }
David Howells811d7362006-08-29 19:06:09 +01002269
Dave Chinner546a1922010-08-24 11:44:34 +10002270 /*
2271 * We stop writing back only if we are not doing
2272 * integrity sync. In case of integrity sync we have to
2273 * keep going until we have written all the pages
2274 * we tagged for writeback prior to entering this loop.
2275 */
2276 if (--wbc->nr_to_write <= 0 &&
2277 wbc->sync_mode == WB_SYNC_NONE) {
2278 done = 1;
2279 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08002280 }
David Howells811d7362006-08-29 19:06:09 +01002281 }
2282 pagevec_release(&pvec);
2283 cond_resched();
2284 }
Dave Chinner64081362018-10-26 15:09:45 -07002285
2286 /*
2287 * If we hit the last page and there is more work to be done: wrap
2288 * back the index back to the start of the file for the next
2289 * time we are called.
2290 */
2291 if (wbc->range_cyclic && !done)
2292 done_index = 0;
Dave Chinner0b564922010-06-09 10:37:18 +10002293 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2294 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04002295
David Howells811d7362006-08-29 19:06:09 +01002296 return ret;
2297}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002298EXPORT_SYMBOL(write_cache_pages);
2299
2300/*
2301 * Function used by generic_writepages to call the real writepage
2302 * function and set the mapping flags on error
2303 */
2304static int __writepage(struct page *page, struct writeback_control *wbc,
2305 void *data)
2306{
2307 struct address_space *mapping = data;
2308 int ret = mapping->a_ops->writepage(page, wbc);
2309 mapping_set_error(mapping, ret);
2310 return ret;
2311}
2312
2313/**
2314 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2315 * @mapping: address space structure to write
2316 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2317 *
2318 * This is a library function, which implements the writepages()
2319 * address_space_operation.
Mike Rapoporta862f682019-03-05 15:48:42 -08002320 *
2321 * Return: %0 on success, negative error code otherwise
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002322 */
2323int generic_writepages(struct address_space *mapping,
2324 struct writeback_control *wbc)
2325{
Shaohua Li9b6096a2011-03-17 10:47:06 +01002326 struct blk_plug plug;
2327 int ret;
2328
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002329 /* deal with chardevs and other special file */
2330 if (!mapping->a_ops->writepage)
2331 return 0;
2332
Shaohua Li9b6096a2011-03-17 10:47:06 +01002333 blk_start_plug(&plug);
2334 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2335 blk_finish_plug(&plug);
2336 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002337}
David Howells811d7362006-08-29 19:06:09 +01002338
2339EXPORT_SYMBOL(generic_writepages);
2340
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2342{
Andrew Morton22905f72005-11-16 15:07:01 -08002343 int ret;
Jan Karafee468f2021-09-02 14:53:06 -07002344 struct bdi_writeback *wb;
Andrew Morton22905f72005-11-16 15:07:01 -08002345
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 if (wbc->nr_to_write <= 0)
2347 return 0;
Jan Karafee468f2021-09-02 14:53:06 -07002348 wb = inode_to_wb_wbc(mapping->host, wbc);
2349 wb_bandwidth_estimate_start(wb);
Theodore Ts'o80a2ea92017-04-28 09:51:54 -04002350 while (1) {
2351 if (mapping->a_ops->writepages)
2352 ret = mapping->a_ops->writepages(mapping, wbc);
2353 else
2354 ret = generic_writepages(mapping, wbc);
2355 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2356 break;
Mel Gorman8d588022021-11-05 13:42:35 -07002357
2358 /*
2359 * Lacking an allocation context or the locality or writeback
2360 * state of any of the inode's pages, throttle based on
2361 * writeback activity on the local node. It's as good a
2362 * guess as any.
2363 */
2364 reclaim_throttle(NODE_DATA(numa_node_id()),
Mel Gormanc3f4a9a2021-11-05 13:42:42 -07002365 VMSCAN_THROTTLE_WRITEBACK);
Theodore Ts'o80a2ea92017-04-28 09:51:54 -04002366 }
Jan Kara45a29662021-09-02 14:53:09 -07002367 /*
2368 * Usually few pages are written by now from those we've just submitted
2369 * but if there's constant writeback being submitted, this makes sure
2370 * writeback bandwidth is updated once in a while.
2371 */
Jan Kara20792eb2021-09-02 14:53:15 -07002372 if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2373 BANDWIDTH_INTERVAL))
Jan Kara45a29662021-09-02 14:53:09 -07002374 wb_update_bandwidth(wb);
Andrew Morton22905f72005-11-16 15:07:01 -08002375 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376}
2377
2378/**
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002379 * folio_write_one - write out a single folio and wait on I/O.
2380 * @folio: The folio to write.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 *
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002382 * The folio must be locked by the caller and will be unlocked upon return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 *
Jeff Layton37e51a72017-07-05 15:17:19 -04002384 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2385 * function returns.
Mike Rapoporta862f682019-03-05 15:48:42 -08002386 *
2387 * Return: %0 on success, negative error code otherwise
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 */
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002389int folio_write_one(struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390{
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002391 struct address_space *mapping = folio->mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 int ret = 0;
2393 struct writeback_control wbc = {
2394 .sync_mode = WB_SYNC_ALL,
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002395 .nr_to_write = folio_nr_pages(folio),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 };
2397
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002398 BUG_ON(!folio_test_locked(folio));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002400 folio_wait_writeback(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002402 if (folio_clear_dirty_for_io(folio)) {
2403 folio_get(folio);
2404 ret = mapping->a_ops->writepage(&folio->page, &wbc);
Jeff Layton37e51a72017-07-05 15:17:19 -04002405 if (ret == 0)
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002406 folio_wait_writeback(folio);
2407 folio_put(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 } else {
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002409 folio_unlock(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 }
Jeff Layton37e51a72017-07-05 15:17:19 -04002411
2412 if (!ret)
2413 ret = filemap_check_errors(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 return ret;
2415}
Matthew Wilcox (Oracle)121703c2021-03-09 13:48:03 -05002416EXPORT_SYMBOL(folio_write_one);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417
2418/*
Ken Chen76719322007-02-10 01:43:15 -08002419 * For address_spaces which do not use buffers nor write back.
2420 */
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +00002421bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
Ken Chen76719322007-02-10 01:43:15 -08002422{
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +00002423 if (!folio_test_dirty(folio))
2424 return !folio_test_set_dirty(folio);
2425 return false;
Ken Chen76719322007-02-10 01:43:15 -08002426}
Matthew Wilcox (Oracle)46de8b972022-02-09 20:22:13 +00002427EXPORT_SYMBOL(noop_dirty_folio);
Ken Chen76719322007-02-10 01:43:15 -08002428
2429/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002430 * Helper function for set_page_dirty family.
Greg Thelenc4843a72015-05-22 17:13:16 -04002431 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002432 * Caller must hold lock_page_memcg().
Greg Thelenc4843a72015-05-22 17:13:16 -04002433 *
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002434 * NOTE: This relies on being atomic wrt interrupts.
2435 */
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002436static void folio_account_dirtied(struct folio *folio,
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002437 struct address_space *mapping)
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002438{
Tejun Heo52ebea72015-05-22 17:13:37 -04002439 struct inode *inode = mapping->host;
2440
Matthew Wilcox (Oracle)b9b0ff62021-07-13 22:45:50 -04002441 trace_writeback_dirty_folio(folio, mapping);
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002442
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002443 if (mapping_can_writeback(mapping)) {
Tejun Heo52ebea72015-05-22 17:13:37 -04002444 struct bdi_writeback *wb;
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002445 long nr = folio_nr_pages(folio);
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002446
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002447 inode_attach_wb(inode, &folio->page);
Tejun Heo52ebea72015-05-22 17:13:37 -04002448 wb = inode_to_wb(inode);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002449
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002450 __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2451 __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2452 __node_stat_mod_folio(folio, NR_DIRTIED, nr);
2453 wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2454 wb_stat_mod(wb, WB_DIRTIED, nr);
2455 task_io_account_write(nr * PAGE_SIZE);
2456 current->nr_dirtied += nr;
2457 __this_cpu_add(bdp_ratelimits, nr);
Tejun Heo97b27822019-08-26 09:06:56 -07002458
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002459 mem_cgroup_track_foreign_dirty(folio, wb);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002460 }
2461}
2462
2463/*
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002464 * Helper function for deaccounting dirty page without writeback.
2465 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002466 * Caller must hold lock_page_memcg().
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002467 */
Hugh Dickins566d3362022-03-24 18:13:59 -07002468void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002469{
Hugh Dickins566d3362022-03-24 18:13:59 -07002470 long nr = folio_nr_pages(folio);
2471
2472 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2473 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2474 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2475 task_io_account_cancelled_write(nr * PAGE_SIZE);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002476}
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002477
2478/*
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002479 * Mark the folio dirty, and set it dirty in the page cache, and mark
2480 * the inode dirty.
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002481 *
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002482 * 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 -07002483 * not been truncated.
2484 *
Matthew Wilcox (Oracle)a229a4f2020-12-20 06:44:51 -05002485 * The caller must hold lock_page_memcg(). Most callers have the folio
2486 * locked. A few have the folio blocked from truncation through other
2487 * means (eg zap_page_range() has it mapped and is holding the page table
2488 * lock). This can also be called from mark_buffer_dirty(), which I
2489 * cannot prove is always protected against truncate.
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002490 */
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002491void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002492 int warn)
2493{
2494 unsigned long flags;
2495
2496 xa_lock_irqsave(&mapping->i_pages, flags);
Matthew Wilcox (Oracle)203a3152021-05-04 11:01:10 -04002497 if (folio->mapping) { /* Race with truncate? */
2498 WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2499 folio_account_dirtied(folio, mapping);
2500 __xa_set_mark(&mapping->i_pages, folio_index(folio),
Matthew Wilcox (Oracle)6e1cae882021-06-28 19:36:15 -07002501 PAGECACHE_TAG_DIRTY);
2502 }
2503 xa_unlock_irqrestore(&mapping->i_pages, flags);
2504}
2505
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002506/**
2507 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2508 * @mapping: Address space this folio belongs to.
2509 * @folio: Folio to be marked as dirty.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510 *
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002511 * Filesystems which do not use buffer heads should call this function
2512 * from their set_page_dirty address space operation. It ignores the
2513 * contents of folio_get_private(), so if the filesystem marks individual
2514 * blocks as dirty, the filesystem should handle that itself.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 *
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002516 * This is also sometimes used by filesystems which use buffer_heads when
2517 * a single buffer is being dirtied: we want to set the folio dirty in
2518 * that case, but not all the buffers. This is a "bottom-up" dirtying,
Matthew Wilcox (Oracle)e6219002022-02-09 20:22:12 +00002519 * whereas block_dirty_folio() is a "top-down" dirtying.
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002520 *
2521 * The caller must ensure this doesn't race with truncation. Most will
2522 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2523 * folio mapped and the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 */
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002525bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526{
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002527 folio_memcg_lock(folio);
2528 if (folio_test_set_dirty(folio)) {
2529 folio_memcg_unlock(folio);
2530 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 }
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002532
2533 __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2534 folio_memcg_unlock(folio);
2535
2536 if (mapping->host) {
2537 /* !PageAnon && !swapper_space */
2538 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2539 }
2540 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541}
Matthew Wilcox (Oracle)85d4d2e2021-05-03 23:30:44 -04002542EXPORT_SYMBOL(filemap_dirty_folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04002544/**
2545 * folio_account_redirty - Manually account for redirtying a page.
2546 * @folio: The folio which is being redirtied.
2547 *
2548 * Most filesystems should call folio_redirty_for_writepage() instead
2549 * of this fuction. If your filesystem is doing writeback outside the
2550 * context of a writeback_control(), it can call this when redirtying
2551 * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED,
2552 * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN,
2553 * WB_WRITTEN) in long term. The mismatches will lead to systematic errors
2554 * in balanced_dirty_ratelimit and the dirty pages position control.
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002555 */
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04002556void folio_account_redirty(struct folio *folio)
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002557{
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04002558 struct address_space *mapping = folio->mapping;
Tejun Heo91018132015-05-22 17:13:39 -04002559
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002560 if (mapping && mapping_can_writeback(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002561 struct inode *inode = mapping->host;
2562 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002563 struct wb_lock_cookie cookie = {};
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04002564 long nr = folio_nr_pages(folio);
Tejun Heo91018132015-05-22 17:13:39 -04002565
Greg Thelen2e898e42018-04-20 14:55:42 -07002566 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04002567 current->nr_dirtied -= nr;
2568 node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2569 wb_stat_mod(wb, WB_DIRTIED, -nr);
Greg Thelen2e898e42018-04-20 14:55:42 -07002570 unlocked_inode_to_wb_end(inode, &cookie);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002571 }
2572}
Matthew Wilcox (Oracle)25ff8b12021-05-03 10:06:55 -04002573EXPORT_SYMBOL(folio_account_redirty);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002574
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002575/**
2576 * folio_redirty_for_writepage - Decline to write a dirty folio.
2577 * @wbc: The writeback control.
2578 * @folio: The folio.
2579 *
2580 * When a writepage implementation decides that it doesn't want to write
2581 * @folio for some reason, it should call this function, unlock @folio and
2582 * return 0.
2583 *
2584 * Return: True if we redirtied the folio. False if someone else dirtied
2585 * it first.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586 */
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002587bool folio_redirty_for_writepage(struct writeback_control *wbc,
2588 struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589{
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002590 bool ret;
2591 long nr = folio_nr_pages(folio);
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002592
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002593 wbc->pages_skipped += nr;
2594 ret = filemap_dirty_folio(folio->mapping, folio);
2595 folio_account_redirty(folio);
2596
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002597 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598}
Matthew Wilcox (Oracle)cd78ab12021-05-02 23:22:52 -04002599EXPORT_SYMBOL(folio_redirty_for_writepage);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002601/**
2602 * folio_mark_dirty - Mark a folio as being modified.
2603 * @folio: The folio.
Wu Fengguang6746aff2009-09-16 11:50:14 +02002604 *
Matthew Wilcox (Oracle)2ca456c2022-04-28 14:21:02 -04002605 * The folio may not be truncated while this function is running.
2606 * Holding the folio lock is sufficient to prevent truncation, but some
2607 * callers cannot acquire a sleeping lock. These callers instead hold
2608 * the page table lock for a page table which contains at least one page
2609 * in this folio. Truncation will block on the page table lock as it
2610 * unmaps pages before removing the folio from its mapping.
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002611 *
2612 * Return: True if the folio was newly dirtied, false if it was already dirty.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613 */
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002614bool folio_mark_dirty(struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615{
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002616 struct address_space *mapping = folio_mapping(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617
2618 if (likely(mapping)) {
Minchan Kim278df9f2011-03-22 16:32:54 -07002619 /*
2620 * readahead/lru_deactivate_page could remain
Matthew Wilcox (Oracle)6f31a5a2022-02-09 20:22:00 +00002621 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2622 * About readahead, if the folio is written, the flags would be
Minchan Kim278df9f2011-03-22 16:32:54 -07002623 * reset. So no problem.
Matthew Wilcox (Oracle)6f31a5a2022-02-09 20:22:00 +00002624 * About lru_deactivate_page, if the folio is redirtied,
2625 * the flag will be reset. So no problem. but if the
2626 * folio is used by readahead it will confuse readahead
2627 * and make it restart the size rampup process. But it's
2628 * a trivial problem.
Minchan Kim278df9f2011-03-22 16:32:54 -07002629 */
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002630 if (folio_test_reclaim(folio))
2631 folio_clear_reclaim(folio);
Matthew Wilcox (Oracle)3a3bae52022-02-09 20:22:15 +00002632 return mapping->a_ops->dirty_folio(mapping, folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 }
Matthew Wilcox (Oracle)3a3bae52022-02-09 20:22:15 +00002634
2635 return noop_dirty_folio(mapping, folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636}
Matthew Wilcox (Oracle)b5e84592021-04-26 23:53:10 -04002637EXPORT_SYMBOL(folio_mark_dirty);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638
2639/*
2640 * set_page_dirty() is racy if the caller has no reference against
2641 * page->mapping->host, and if the page is unlocked. This is because another
2642 * CPU could truncate the page off the mapping and then free the mapping.
2643 *
2644 * Usually, the page _is_ locked, or the caller is a user-space process which
2645 * holds a reference on the inode by having an open file.
2646 *
2647 * In other cases, the page should be locked before running set_page_dirty().
2648 */
2649int set_page_dirty_lock(struct page *page)
2650{
2651 int ret;
2652
Jens Axboe7eaceac2011-03-10 08:52:07 +01002653 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 ret = set_page_dirty(page);
2655 unlock_page(page);
2656 return ret;
2657}
2658EXPORT_SYMBOL(set_page_dirty_lock);
2659
2660/*
Tejun Heo11f81be2015-05-22 17:13:15 -04002661 * This cancels just the dirty bit on the kernel page itself, it does NOT
2662 * actually remove dirty bits on any mmap's that may be around. It also
2663 * leaves the page tagged dirty, so any sync activity will still find it on
2664 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2665 * look at the dirty bits in the VM.
2666 *
2667 * Doing this should *normally* only ever be done when a page is truncated,
2668 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2669 * this when it notices that somebody has cleaned out all the buffers on a
2670 * page without actually doing it through the VM. Can you say "ext3 is
2671 * horribly ugly"? Thought you could.
2672 */
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002673void __folio_cancel_dirty(struct folio *folio)
Tejun Heo11f81be2015-05-22 17:13:15 -04002674{
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002675 struct address_space *mapping = folio_mapping(folio);
Greg Thelenc4843a72015-05-22 17:13:16 -04002676
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002677 if (mapping_can_writeback(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002678 struct inode *inode = mapping->host;
2679 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002680 struct wb_lock_cookie cookie = {};
Greg Thelenc4843a72015-05-22 17:13:16 -04002681
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002682 folio_memcg_lock(folio);
Greg Thelen2e898e42018-04-20 14:55:42 -07002683 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002684
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002685 if (folio_test_clear_dirty(folio))
Hugh Dickins566d3362022-03-24 18:13:59 -07002686 folio_account_cleaned(folio, wb);
Greg Thelenc4843a72015-05-22 17:13:16 -04002687
Greg Thelen2e898e42018-04-20 14:55:42 -07002688 unlocked_inode_to_wb_end(inode, &cookie);
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002689 folio_memcg_unlock(folio);
Greg Thelenc4843a72015-05-22 17:13:16 -04002690 } else {
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002691 folio_clear_dirty(folio);
Greg Thelenc4843a72015-05-22 17:13:16 -04002692 }
Tejun Heo11f81be2015-05-22 17:13:15 -04002693}
Matthew Wilcox (Oracle)fdaf5322021-03-08 16:43:04 -05002694EXPORT_SYMBOL(__folio_cancel_dirty);
Tejun Heo11f81be2015-05-22 17:13:15 -04002695
2696/*
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002697 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2698 * Returns true if the folio was previously dirty.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002700 * This is for preparing to put the folio under writeout. We leave
2701 * the folio tagged as dirty in the xarray so that a concurrent
2702 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2703 * The ->writepage implementation will run either folio_start_writeback()
2704 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2705 * and xarray dirty tag back into sync.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002707 * This incoherency between the folio's dirty flag and xarray tag is
2708 * unfortunate, but it only exists while the folio is locked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 */
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002710bool folio_clear_dirty_for_io(struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711{
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002712 struct address_space *mapping = folio_mapping(folio);
2713 bool ret = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002715 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
Nick Piggin79352892007-07-19 01:47:22 -07002716
Christoph Hellwigf56753a2020-09-24 08:51:40 +02002717 if (mapping && mapping_can_writeback(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002718 struct inode *inode = mapping->host;
2719 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002720 struct wb_lock_cookie cookie = {};
Tejun Heo682aa8e2015-05-28 14:50:53 -04002721
Linus Torvalds7658cc22006-12-29 10:00:58 -08002722 /*
2723 * Yes, Virginia, this is indeed insane.
2724 *
2725 * We use this sequence to make sure that
2726 * (a) we account for dirty stats properly
2727 * (b) we tell the low-level filesystem to
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002728 * mark the whole folio dirty if it was
Linus Torvalds7658cc22006-12-29 10:00:58 -08002729 * dirty in a pagetable. Only to then
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002730 * (c) clean the folio again and return 1 to
Linus Torvalds7658cc22006-12-29 10:00:58 -08002731 * cause the writeback.
2732 *
2733 * This way we avoid all nasty races with the
2734 * dirty bit in multiple places and clearing
2735 * them concurrently from different threads.
2736 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002737 * Note! Normally the "folio_mark_dirty(folio)"
Linus Torvalds7658cc22006-12-29 10:00:58 -08002738 * has no effect on the actual dirty bit - since
2739 * that will already usually be set. But we
2740 * need the side effects, and it can help us
2741 * avoid races.
2742 *
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002743 * We basically use the folio "master dirty bit"
Linus Torvalds7658cc22006-12-29 10:00:58 -08002744 * as a serialization point for all the different
2745 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002746 */
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002747 if (folio_mkclean(folio))
2748 folio_mark_dirty(folio);
Nick Piggin79352892007-07-19 01:47:22 -07002749 /*
2750 * We carefully synchronise fault handlers against
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002751 * installing a dirty pte and marking the folio dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002752 * at this point. We do this by having them hold the
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002753 * page lock while dirtying the folio, and folios are
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002754 * always locked coming in here, so we get the desired
2755 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002756 */
Greg Thelen2e898e42018-04-20 14:55:42 -07002757 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002758 if (folio_test_clear_dirty(folio)) {
2759 long nr = folio_nr_pages(folio);
2760 lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2761 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2762 wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2763 ret = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 }
Greg Thelen2e898e42018-04-20 14:55:42 -07002765 unlocked_inode_to_wb_end(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002766 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 }
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002768 return folio_test_clear_dirty(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769}
Matthew Wilcox (Oracle)9350f202021-02-28 16:21:20 -05002770EXPORT_SYMBOL(folio_clear_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
Jan Kara633a2ab2021-09-02 14:53:04 -07002772static void wb_inode_writeback_start(struct bdi_writeback *wb)
2773{
2774 atomic_inc(&wb->writeback_inodes);
2775}
2776
2777static void wb_inode_writeback_end(struct bdi_writeback *wb)
2778{
2779 atomic_dec(&wb->writeback_inodes);
Jan Kara45a29662021-09-02 14:53:09 -07002780 /*
2781 * Make sure estimate of writeback throughput gets updated after
2782 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
2783 * (which is the interval other bandwidth updates use for batching) so
2784 * that if multiple inodes end writeback at a similar time, they get
2785 * batched into one bandwidth update.
2786 */
2787 queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
Jan Kara633a2ab2021-09-02 14:53:04 -07002788}
2789
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002790bool __folio_end_writeback(struct folio *folio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791{
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002792 long nr = folio_nr_pages(folio);
2793 struct address_space *mapping = folio_mapping(folio);
2794 bool ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002796 folio_memcg_lock(folio);
Huang Ying371a0962016-10-07 16:59:30 -07002797 if (mapping && mapping_use_writeback_tags(mapping)) {
Tejun Heo91018132015-05-22 17:13:39 -04002798 struct inode *inode = mapping->host;
2799 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 unsigned long flags;
2801
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002802 xa_lock_irqsave(&mapping->i_pages, flags);
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002803 ret = folio_test_clear_writeback(folio);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002804 if (ret) {
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002805 __xa_clear_mark(&mapping->i_pages, folio_index(folio),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 PAGECACHE_TAG_WRITEBACK);
Christoph Hellwig823423e2020-09-24 08:51:39 +02002807 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
Tejun Heo91018132015-05-22 17:13:39 -04002808 struct bdi_writeback *wb = inode_to_wb(inode);
2809
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002810 wb_stat_mod(wb, WB_WRITEBACK, -nr);
2811 __wb_writeout_add(wb, nr);
Jan Kara633a2ab2021-09-02 14:53:04 -07002812 if (!mapping_tagged(mapping,
2813 PAGECACHE_TAG_WRITEBACK))
2814 wb_inode_writeback_end(wb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002815 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002816 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002817
2818 if (mapping->host && !mapping_tagged(mapping,
2819 PAGECACHE_TAG_WRITEBACK))
2820 sb_clear_inode_writeback(mapping->host);
2821
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002822 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 } else {
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002824 ret = folio_test_clear_writeback(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 }
Wu Fengguang99b12e32011-07-25 17:12:37 -07002826 if (ret) {
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002827 lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
2828 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2829 node_stat_mod_folio(folio, NR_WRITTEN, nr);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002830 }
Matthew Wilcox (Oracle)269ccca32021-01-15 23:34:16 -05002831 folio_memcg_unlock(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 return ret;
2833}
2834
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002835bool __folio_start_writeback(struct folio *folio, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836{
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002837 long nr = folio_nr_pages(folio);
2838 struct address_space *mapping = folio_mapping(folio);
2839 bool ret;
2840 int access_ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002842 folio_memcg_lock(folio);
Huang Ying371a0962016-10-07 16:59:30 -07002843 if (mapping && mapping_use_writeback_tags(mapping)) {
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002844 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
Tejun Heo91018132015-05-22 17:13:39 -04002845 struct inode *inode = mapping->host;
2846 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 unsigned long flags;
2848
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002849 xas_lock_irqsave(&xas, flags);
2850 xas_load(&xas);
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002851 ret = folio_test_set_writeback(folio);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002852 if (!ret) {
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002853 bool on_wblist;
2854
2855 on_wblist = mapping_tagged(mapping,
2856 PAGECACHE_TAG_WRITEBACK);
2857
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002858 xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
Jan Kara633a2ab2021-09-02 14:53:04 -07002859 if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2860 struct bdi_writeback *wb = inode_to_wb(inode);
2861
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002862 wb_stat_mod(wb, WB_WRITEBACK, nr);
Jan Kara633a2ab2021-09-02 14:53:04 -07002863 if (!on_wblist)
2864 wb_inode_writeback_start(wb);
2865 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002866
2867 /*
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002868 * We can come through here when swapping
2869 * anonymous folios, so we don't necessarily
2870 * have an inode to track for sync.
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002871 */
2872 if (mapping->host && !on_wblist)
2873 sb_mark_inode_writeback(mapping->host);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002874 }
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002875 if (!folio_test_dirty(folio))
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002876 xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002877 if (!keep_write)
Matthew Wilcoxff9c7452017-11-22 11:41:23 -05002878 xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2879 xas_unlock_irqrestore(&xas, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 } else {
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002881 ret = folio_test_set_writeback(folio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 }
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002883 if (!ret) {
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002884 lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
2885 zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002886 }
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002887 folio_memcg_unlock(folio);
2888 access_ret = arch_make_folio_accessible(folio);
Claudio Imbrendaf28d4362020-04-01 21:05:56 -07002889 /*
2890 * If writeback has been triggered on a page that cannot be made
2891 * accessible, it is too late to recover here.
2892 */
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002893 VM_BUG_ON_FOLIO(access_ret != 0, folio);
Claudio Imbrendaf28d4362020-04-01 21:05:56 -07002894
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896}
Matthew Wilcox (Oracle)f143f1e2021-04-24 12:00:48 -04002897EXPORT_SYMBOL(__folio_start_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002899/**
2900 * folio_wait_writeback - Wait for a folio to finish writeback.
2901 * @folio: The folio to wait for.
2902 *
2903 * If the folio is currently being written back to storage, wait for the
2904 * I/O to complete.
2905 *
2906 * Context: Sleeps. Must be called in process context and with
2907 * no spinlocks held. Caller should hold a reference on the folio.
2908 * If the folio is not locked, writeback may start again after writeback
2909 * has finished.
Yafang Shao19343b52019-05-13 17:23:11 -07002910 */
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002911void folio_wait_writeback(struct folio *folio)
Yafang Shao19343b52019-05-13 17:23:11 -07002912{
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002913 while (folio_test_writeback(folio)) {
Matthew Wilcox (Oracle)b9b0ff62021-07-13 22:45:50 -04002914 trace_folio_wait_writeback(folio, folio_mapping(folio));
Matthew Wilcox (Oracle)101c0bf2021-03-04 12:02:54 -05002915 folio_wait_bit(folio, PG_writeback);
Yafang Shao19343b52019-05-13 17:23:11 -07002916 }
2917}
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002918EXPORT_SYMBOL_GPL(folio_wait_writeback);
Yafang Shao19343b52019-05-13 17:23:11 -07002919
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002920/**
2921 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
2922 * @folio: The folio to wait for.
2923 *
2924 * If the folio is currently being written back to storage, wait for the
2925 * I/O to complete or a fatal signal to arrive.
2926 *
2927 * Context: Sleeps. Must be called in process context and with
2928 * no spinlocks held. Caller should hold a reference on the folio.
2929 * If the folio is not locked, writeback may start again after writeback
2930 * has finished.
2931 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00002932 */
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002933int folio_wait_writeback_killable(struct folio *folio)
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00002934{
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002935 while (folio_test_writeback(folio)) {
Matthew Wilcox (Oracle)b9b0ff62021-07-13 22:45:50 -04002936 trace_folio_wait_writeback(folio, folio_mapping(folio));
Matthew Wilcox (Oracle)101c0bf2021-03-04 12:02:54 -05002937 if (folio_wait_bit_killable(folio, PG_writeback))
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00002938 return -EINTR;
2939 }
2940
2941 return 0;
2942}
Matthew Wilcox (Oracle)490e0162021-03-04 11:09:17 -05002943EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
Matthew Wilcox (Oracle)e5dbd332021-03-20 05:40:39 +00002944
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002945/**
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05002946 * folio_wait_stable() - wait for writeback to finish, if necessary.
2947 * @folio: The folio to wait on.
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002948 *
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05002949 * This function determines if the given folio is related to a backing
2950 * device that requires folio contents to be held stable during writeback.
2951 * If so, then it will wait for any pending writeback to complete.
2952 *
2953 * Context: Sleeps. Must be called in process context and with
2954 * no spinlocks held. Caller should hold a reference on the folio.
2955 * If the folio is not locked, writeback may start again after writeback
2956 * has finished.
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002957 */
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05002958void folio_wait_stable(struct folio *folio)
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002959{
David Howells452c4722021-08-12 22:09:57 +01002960 if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES)
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05002961 folio_wait_writeback(folio);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002962}
Matthew Wilcox (Oracle)a49d0c52021-03-04 11:25:25 -05002963EXPORT_SYMBOL_GPL(folio_wait_stable);