blob: 498c924f2fcd609cb0e402c919de18fa301cf479 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra90eec102015-11-16 11:08:45 +01005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
Francois Camie1f8e872008-10-15 22:01:59 -070010 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Initial version
12 */
13
14#include <linux/kernel.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040015#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
Al Viroff01bb482011-09-16 02:31:11 -040035#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Jan Karaeb608e32012-05-24 18:59:11 +020037#include <linux/timer.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060038#include <linux/sched/rt.h>
Lisa Du6e543d52013-09-11 14:22:36 -070039#include <linux/mm_inline.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100040#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
Lisa Du6e543d52013-09-11 14:22:36 -070042#include "internal.h"
43
Linus Torvalds1da177e2005-04-16 15:20:36 -070044/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060045 * Sleep at most 200ms at a time in balance_dirty_pages().
46 */
47#define MAX_PAUSE max(HZ/5, 1)
48
49/*
Wu Fengguang5b9b3572011-12-06 13:17:17 -060050 * Try to keep balance_dirty_pages() call intervals higher than this many pages
51 * by raising pause time to max_pause when falls below it.
52 */
53#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
54
55/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060056 * Estimate write bandwidth at 200ms intervals.
57 */
58#define BANDWIDTH_INTERVAL max(HZ/5, 1)
59
Wu Fengguang6c14ae12011-03-02 16:04:18 -060060#define RATELIMIT_CALC_SHIFT 10
61
Wu Fengguange98be2d2010-08-29 11:22:30 -060062/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
64 * will look to see if it needs to force writeback or throttling.
65 */
66static long ratelimit_pages = 32;
67
Linus Torvalds1da177e2005-04-16 15:20:36 -070068/* The following parameters are exported via /proc/sys/vm */
69
70/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020071 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080073int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
75/*
David Rientjes2da02992009-01-06 14:39:31 -080076 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
77 * dirty_background_ratio * the amount of dirtyable memory
78 */
79unsigned long dirty_background_bytes;
80
81/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080082 * free highmem will not be subtracted from the total free memory
83 * for calculating free ratios if vm_highmem_is_dirtyable is true
84 */
85int vm_highmem_is_dirtyable;
86
87/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 * The generator of dirty data starts writeback at this percentage
89 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080090int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92/*
David Rientjes2da02992009-01-06 14:39:31 -080093 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
94 * vm_dirty_ratio * the amount of dirtyable memory
95 */
96unsigned long vm_dirty_bytes;
97
98/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070099 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Artem Bityutskiy91913a22012-03-21 22:33:00 -0400103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
104
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700106 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110/*
111 * Flag that makes the machine dump writes/reads and block dirtyings.
112 */
113int block_dump;
114
115/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
117 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118 */
119int laptop_mode;
120
121EXPORT_SYMBOL(laptop_mode);
122
123/* End of sysctl-exported parameters */
124
Tejun Heodcc25ae2015-05-22 18:23:22 -0400125struct wb_domain global_wb_domain;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400127/* consolidated parameters for balance_dirty_pages() and its subroutines */
128struct dirty_throttle_control {
Tejun Heoe9f07df2015-05-22 18:23:28 -0400129#ifdef CONFIG_CGROUP_WRITEBACK
130 struct wb_domain *dom;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400131 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
Tejun Heoe9f07df2015-05-22 18:23:28 -0400132#endif
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400133 struct bdi_writeback *wb;
Tejun Heoe9770b32015-05-22 18:23:27 -0400134 struct fprop_local_percpu *wb_completions;
Jan Karaeb608e32012-05-24 18:59:11 +0200135
Tejun Heo9fc3a432015-05-22 18:23:30 -0400136 unsigned long avail; /* dirtyable */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400137 unsigned long dirty; /* file_dirty + write + nfs */
138 unsigned long thresh; /* dirty threshold */
139 unsigned long bg_thresh; /* dirty background threshold */
140
141 unsigned long wb_dirty; /* per-wb counterparts */
142 unsigned long wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -0400143 unsigned long wb_bg_thresh;
Tejun Heodaddfa32015-05-22 18:23:26 -0400144
145 unsigned long pos_ratio;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400146};
147
Jan Karaeb608e32012-05-24 18:59:11 +0200148/*
149 * Length of period for aging writeout fractions of bdis. This is an
150 * arbitrarily chosen number. The longer the period, the slower fractions will
151 * reflect changes in current writeout rate.
152 */
153#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700154
Tejun Heo693108a2015-05-22 17:13:49 -0400155#ifdef CONFIG_CGROUP_WRITEBACK
156
Tejun Heod60d1bd2015-09-29 12:47:53 -0400157#define GDTC_INIT(__wb) .wb = (__wb), \
158 .dom = &global_wb_domain, \
159 .wb_completions = &(__wb)->completions
160
Tejun Heo9fc3a432015-05-22 18:23:30 -0400161#define GDTC_INIT_NO_WB .dom = &global_wb_domain
Tejun Heod60d1bd2015-09-29 12:47:53 -0400162
163#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
164 .dom = mem_cgroup_wb_domain(__wb), \
165 .wb_completions = &(__wb)->memcg_completions, \
166 .gdtc = __gdtc
Tejun Heoc2aa7232015-05-22 18:23:35 -0400167
168static bool mdtc_valid(struct dirty_throttle_control *dtc)
169{
170 return dtc->dom;
171}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400172
173static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
174{
175 return dtc->dom;
176}
177
Tejun Heo9fc3a432015-05-22 18:23:30 -0400178static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
179{
180 return mdtc->gdtc;
181}
182
Tejun Heo841710a2015-05-22 18:23:33 -0400183static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
184{
185 return &wb->memcg_completions;
186}
187
Tejun Heo693108a2015-05-22 17:13:49 -0400188static void wb_min_max_ratio(struct bdi_writeback *wb,
189 unsigned long *minp, unsigned long *maxp)
190{
191 unsigned long this_bw = wb->avg_write_bandwidth;
192 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
193 unsigned long long min = wb->bdi->min_ratio;
194 unsigned long long max = wb->bdi->max_ratio;
195
196 /*
197 * @wb may already be clean by the time control reaches here and
198 * the total may not include its bw.
199 */
200 if (this_bw < tot_bw) {
201 if (min) {
202 min *= this_bw;
Wen Yang28f04302020-01-13 16:29:23 -0800203 min = div64_ul(min, tot_bw);
Tejun Heo693108a2015-05-22 17:13:49 -0400204 }
205 if (max < 100) {
206 max *= this_bw;
Wen Yang28f04302020-01-13 16:29:23 -0800207 max = div64_ul(max, tot_bw);
Tejun Heo693108a2015-05-22 17:13:49 -0400208 }
209 }
210
211 *minp = min;
212 *maxp = max;
213}
214
215#else /* CONFIG_CGROUP_WRITEBACK */
216
Tejun Heod60d1bd2015-09-29 12:47:53 -0400217#define GDTC_INIT(__wb) .wb = (__wb), \
218 .wb_completions = &(__wb)->completions
Tejun Heo9fc3a432015-05-22 18:23:30 -0400219#define GDTC_INIT_NO_WB
Tejun Heoc2aa7232015-05-22 18:23:35 -0400220#define MDTC_INIT(__wb, __gdtc)
221
222static bool mdtc_valid(struct dirty_throttle_control *dtc)
223{
224 return false;
225}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400226
227static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
228{
229 return &global_wb_domain;
230}
231
Tejun Heo9fc3a432015-05-22 18:23:30 -0400232static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
233{
234 return NULL;
235}
236
Tejun Heo841710a2015-05-22 18:23:33 -0400237static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
238{
239 return NULL;
240}
241
Tejun Heo693108a2015-05-22 17:13:49 -0400242static void wb_min_max_ratio(struct bdi_writeback *wb,
243 unsigned long *minp, unsigned long *maxp)
244{
245 *minp = wb->bdi->min_ratio;
246 *maxp = wb->bdi->max_ratio;
247}
248
249#endif /* CONFIG_CGROUP_WRITEBACK */
250
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700251/*
Johannes Weinera756cf52012-01-10 15:07:49 -0800252 * In a memory zone, there is a certain amount of pages we consider
253 * available for the page cache, which is essentially the number of
254 * free and reclaimable pages, minus some zone reserves to protect
255 * lowmem and the ability to uphold the zone's watermarks without
256 * requiring writeback.
257 *
258 * This number of dirtyable pages is the base value of which the
259 * user-configurable dirty ratio is the effictive number of pages that
260 * are allowed to be actually dirtied. Per individual zone, or
261 * globally by using the sum of dirtyable pages over all zones.
262 *
263 * Because the user is allowed to specify the dirty limit globally as
264 * absolute number of bytes, calculating the per-zone dirty limit can
265 * require translating the configured limit into a percentage of
266 * global dirtyable memory first.
267 */
268
Johannes Weinera8045522014-01-29 14:05:39 -0800269/**
Mel Gorman281e3722016-07-28 15:46:11 -0700270 * node_dirtyable_memory - number of dirtyable pages in a node
271 * @pgdat: the node
Johannes Weinera8045522014-01-29 14:05:39 -0800272 *
Mel Gorman281e3722016-07-28 15:46:11 -0700273 * Returns the node's number of pages potentially available for dirty
274 * page cache. This is the base value for the per-node dirty limits.
Johannes Weinera8045522014-01-29 14:05:39 -0800275 */
Mel Gorman281e3722016-07-28 15:46:11 -0700276static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
Johannes Weinera8045522014-01-29 14:05:39 -0800277{
Mel Gorman281e3722016-07-28 15:46:11 -0700278 unsigned long nr_pages = 0;
279 int z;
Johannes Weinera8045522014-01-29 14:05:39 -0800280
Mel Gorman281e3722016-07-28 15:46:11 -0700281 for (z = 0; z < MAX_NR_ZONES; z++) {
282 struct zone *zone = pgdat->node_zones + z;
283
284 if (!populated_zone(zone))
285 continue;
286
287 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
288 }
289
Johannes Weinera8d01432016-01-14 15:20:15 -0800290 /*
291 * Pages reserved for the kernel should not be considered
292 * dirtyable, to prevent a situation where reclaim has to
293 * clean pages in order to balance the zones.
294 */
Mel Gorman281e3722016-07-28 15:46:11 -0700295 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
Johannes Weinera8045522014-01-29 14:05:39 -0800296
Mel Gorman281e3722016-07-28 15:46:11 -0700297 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
298 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800299
300 return nr_pages;
301}
302
Johannes Weiner1edf2232012-01-10 15:06:57 -0800303static unsigned long highmem_dirtyable_memory(unsigned long total)
304{
305#ifdef CONFIG_HIGHMEM
306 int node;
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700307 unsigned long x = 0;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700308 int i;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800309
310 for_each_node_state(node, N_HIGH_MEMORY) {
Mel Gorman281e3722016-07-28 15:46:11 -0700311 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
312 struct zone *z;
Minchan Kim9cb937e2016-07-28 15:47:08 -0700313 unsigned long nr_pages;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800314
Mel Gorman281e3722016-07-28 15:46:11 -0700315 if (!is_highmem_idx(i))
316 continue;
317
318 z = &NODE_DATA(node)->node_zones[i];
Minchan Kim9cb937e2016-07-28 15:47:08 -0700319 if (!populated_zone(z))
320 continue;
Mel Gorman281e3722016-07-28 15:46:11 -0700321
Minchan Kim9cb937e2016-07-28 15:47:08 -0700322 nr_pages = zone_page_state(z, NR_FREE_PAGES);
Mel Gorman281e3722016-07-28 15:46:11 -0700323 /* watch for underflows */
Minchan Kim9cb937e2016-07-28 15:47:08 -0700324 nr_pages -= min(nr_pages, high_wmark_pages(z));
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700325 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
326 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
327 x += nr_pages;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700328 }
Johannes Weiner1edf2232012-01-10 15:06:57 -0800329 }
Mel Gorman281e3722016-07-28 15:46:11 -0700330
Johannes Weiner1edf2232012-01-10 15:06:57 -0800331 /*
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800332 * Unreclaimable memory (kernel memory or anonymous memory
333 * without swap) can bring down the dirtyable pages below
334 * the zone's dirty balance reserve and the above calculation
335 * will underflow. However we still want to add in nodes
336 * which are below threshold (negative values) to get a more
337 * accurate calculation but make sure that the total never
338 * underflows.
339 */
340 if ((long)x < 0)
341 x = 0;
342
343 /*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800344 * Make sure that the number of highmem pages is never larger
345 * than the number of the total dirtyable memory. This can only
346 * occur in very strange VM situations but we want to make sure
347 * that this does not occur.
348 */
349 return min(x, total);
350#else
351 return 0;
352#endif
353}
354
355/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800356 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800357 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800358 * Returns the global number of pages potentially available for dirty
359 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800360 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700361static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800362{
363 unsigned long x;
364
Johannes Weinera8045522014-01-29 14:05:39 -0800365 x = global_page_state(NR_FREE_PAGES);
Johannes Weinera8d01432016-01-14 15:20:15 -0800366 /*
367 * Pages reserved for the kernel should not be considered
368 * dirtyable, to prevent a situation where reclaim has to
369 * clean pages in order to balance the zones.
370 */
371 x -= min(x, totalreserve_pages);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800372
Mel Gorman599d0c92016-07-28 15:45:31 -0700373 x += global_node_page_state(NR_INACTIVE_FILE);
374 x += global_node_page_state(NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800375
Johannes Weiner1edf2232012-01-10 15:06:57 -0800376 if (!vm_highmem_is_dirtyable)
377 x -= highmem_dirtyable_memory(x);
378
379 return x + 1; /* Ensure that we never return 0 */
380}
381
Tejun Heo9fc3a432015-05-22 18:23:30 -0400382/**
383 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
384 * @dtc: dirty_throttle_control of interest
Johannes Weinerccafa282012-01-10 15:07:44 -0800385 *
Tejun Heo9fc3a432015-05-22 18:23:30 -0400386 * Calculate @dtc->thresh and ->bg_thresh considering
387 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
388 * must ensure that @dtc->avail is set before calling this function. The
389 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Johannes Weinerccafa282012-01-10 15:07:44 -0800390 * real-time tasks.
391 */
Tejun Heo9fc3a432015-05-22 18:23:30 -0400392static void domain_dirty_limits(struct dirty_throttle_control *dtc)
393{
394 const unsigned long available_memory = dtc->avail;
395 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
396 unsigned long bytes = vm_dirty_bytes;
397 unsigned long bg_bytes = dirty_background_bytes;
Tejun Heo62a584f2016-05-27 14:34:46 -0400398 /* convert ratios to per-PAGE_SIZE for higher precision */
399 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
400 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400401 unsigned long thresh;
402 unsigned long bg_thresh;
403 struct task_struct *tsk;
404
405 /* gdtc is !NULL iff @dtc is for memcg domain */
406 if (gdtc) {
407 unsigned long global_avail = gdtc->avail;
408
409 /*
410 * The byte settings can't be applied directly to memcg
411 * domains. Convert them to ratios by scaling against
Tejun Heo62a584f2016-05-27 14:34:46 -0400412 * globally available memory. As the ratios are in
413 * per-PAGE_SIZE, they can be obtained by dividing bytes by
414 * number of pages.
Tejun Heo9fc3a432015-05-22 18:23:30 -0400415 */
416 if (bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400417 ratio = min(DIV_ROUND_UP(bytes, global_avail),
418 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400419 if (bg_bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400420 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
421 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400422 bytes = bg_bytes = 0;
423 }
424
425 if (bytes)
426 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
427 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400428 thresh = (ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400429
430 if (bg_bytes)
431 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
432 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400433 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400434
435 if (bg_thresh >= thresh)
436 bg_thresh = thresh / 2;
437 tsk = current;
438 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
NeilBrowna53eaff2016-05-20 16:58:53 -0700439 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
440 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400441 }
442 dtc->thresh = thresh;
443 dtc->bg_thresh = bg_thresh;
444
445 /* we should eventually report the domain in the TP */
446 if (!gdtc)
447 trace_global_dirty_state(bg_thresh, thresh);
448}
449
450/**
451 * global_dirty_limits - background-writeback and dirty-throttling thresholds
452 * @pbackground: out parameter for bg_thresh
453 * @pdirty: out parameter for thresh
454 *
455 * Calculate bg_thresh and thresh for global_wb_domain. See
456 * domain_dirty_limits() for details.
457 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800458void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
459{
Tejun Heo9fc3a432015-05-22 18:23:30 -0400460 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
Johannes Weinerccafa282012-01-10 15:07:44 -0800461
Tejun Heo9fc3a432015-05-22 18:23:30 -0400462 gdtc.avail = global_dirtyable_memory();
463 domain_dirty_limits(&gdtc);
Johannes Weinerccafa282012-01-10 15:07:44 -0800464
Tejun Heo9fc3a432015-05-22 18:23:30 -0400465 *pbackground = gdtc.bg_thresh;
466 *pdirty = gdtc.thresh;
Johannes Weinerccafa282012-01-10 15:07:44 -0800467}
468
Johannes Weinera756cf52012-01-10 15:07:49 -0800469/**
Mel Gorman281e3722016-07-28 15:46:11 -0700470 * node_dirty_limit - maximum number of dirty pages allowed in a node
471 * @pgdat: the node
Johannes Weinera756cf52012-01-10 15:07:49 -0800472 *
Mel Gorman281e3722016-07-28 15:46:11 -0700473 * Returns the maximum number of dirty pages allowed in a node, based
474 * on the node's dirtyable memory.
Johannes Weinera756cf52012-01-10 15:07:49 -0800475 */
Mel Gorman281e3722016-07-28 15:46:11 -0700476static unsigned long node_dirty_limit(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800477{
Mel Gorman281e3722016-07-28 15:46:11 -0700478 unsigned long node_memory = node_dirtyable_memory(pgdat);
Johannes Weinera756cf52012-01-10 15:07:49 -0800479 struct task_struct *tsk = current;
480 unsigned long dirty;
481
482 if (vm_dirty_bytes)
483 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
Mel Gorman281e3722016-07-28 15:46:11 -0700484 node_memory / global_dirtyable_memory();
Johannes Weinera756cf52012-01-10 15:07:49 -0800485 else
Mel Gorman281e3722016-07-28 15:46:11 -0700486 dirty = vm_dirty_ratio * node_memory / 100;
Johannes Weinera756cf52012-01-10 15:07:49 -0800487
488 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
489 dirty += dirty / 4;
490
491 return dirty;
492}
493
494/**
Mel Gorman281e3722016-07-28 15:46:11 -0700495 * node_dirty_ok - tells whether a node is within its dirty limits
496 * @pgdat: the node to check
Johannes Weinera756cf52012-01-10 15:07:49 -0800497 *
Mel Gorman281e3722016-07-28 15:46:11 -0700498 * Returns %true when the dirty pages in @pgdat are within the node's
Johannes Weinera756cf52012-01-10 15:07:49 -0800499 * dirty limit, %false if the limit is exceeded.
500 */
Mel Gorman281e3722016-07-28 15:46:11 -0700501bool node_dirty_ok(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800502{
Mel Gorman281e3722016-07-28 15:46:11 -0700503 unsigned long limit = node_dirty_limit(pgdat);
504 unsigned long nr_pages = 0;
Johannes Weinera756cf52012-01-10 15:07:49 -0800505
Mel Gorman11fb9982016-07-28 15:46:20 -0700506 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
507 nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
508 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
Mel Gorman281e3722016-07-28 15:46:11 -0700509
510 return nr_pages <= limit;
Johannes Weinera756cf52012-01-10 15:07:49 -0800511}
512
David Rientjes2da02992009-01-06 14:39:31 -0800513int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700514 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800515 loff_t *ppos)
516{
517 int ret;
518
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700519 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800520 if (ret == 0 && write)
521 dirty_background_bytes = 0;
522 return ret;
523}
524
525int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700526 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800527 loff_t *ppos)
528{
529 int ret;
530
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700531 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800532 if (ret == 0 && write)
533 dirty_background_ratio = 0;
534 return ret;
535}
536
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700537int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700538 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700539 loff_t *ppos)
540{
541 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800542 int ret;
543
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700544 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700545 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200546 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800547 vm_dirty_bytes = 0;
548 }
549 return ret;
550}
551
David Rientjes2da02992009-01-06 14:39:31 -0800552int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700553 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800554 loff_t *ppos)
555{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800556 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800557 int ret;
558
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700559 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800560 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200561 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800562 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700563 }
564 return ret;
565}
566
Jan Karaeb608e32012-05-24 18:59:11 +0200567static unsigned long wp_next_time(unsigned long cur_time)
568{
569 cur_time += VM_COMPLETIONS_PERIOD_LEN;
570 /* 0 has a special meaning... */
571 if (!cur_time)
572 return 1;
573 return cur_time;
574}
575
Tejun Heoc7981432015-05-22 18:23:29 -0400576static void wb_domain_writeout_inc(struct wb_domain *dom,
577 struct fprop_local_percpu *completions,
578 unsigned int max_prop_frac)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700579{
Tejun Heoc7981432015-05-22 18:23:29 -0400580 __fprop_inc_percpu_max(&dom->completions, completions,
581 max_prop_frac);
Jan Karaeb608e32012-05-24 18:59:11 +0200582 /* First event after period switching was turned off? */
Tejun Heo380c27c2015-05-22 18:23:21 -0400583 if (!unlikely(dom->period_time)) {
Jan Karaeb608e32012-05-24 18:59:11 +0200584 /*
585 * We can race with other __bdi_writeout_inc calls here but
586 * it does not cause any harm since the resulting time when
587 * timer will fire and what is in writeout_period_time will be
588 * roughly the same.
589 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400590 dom->period_time = wp_next_time(jiffies);
591 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200592 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700593}
594
Tejun Heoc7981432015-05-22 18:23:29 -0400595/*
596 * Increment @wb's writeout completion count and the global writeout
597 * completion count. Called from test_clear_page_writeback().
598 */
599static inline void __wb_writeout_inc(struct bdi_writeback *wb)
600{
Tejun Heo841710a2015-05-22 18:23:33 -0400601 struct wb_domain *cgdom;
602
Tejun Heoc7981432015-05-22 18:23:29 -0400603 __inc_wb_stat(wb, WB_WRITTEN);
604 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
605 wb->bdi->max_prop_frac);
Tejun Heo841710a2015-05-22 18:23:33 -0400606
607 cgdom = mem_cgroup_wb_domain(wb);
608 if (cgdom)
609 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
610 wb->bdi->max_prop_frac);
Tejun Heoc7981432015-05-22 18:23:29 -0400611}
612
Tejun Heo93f78d82015-05-22 17:13:27 -0400613void wb_writeout_inc(struct bdi_writeback *wb)
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700614{
615 unsigned long flags;
616
617 local_irq_save(flags);
Tejun Heo93f78d82015-05-22 17:13:27 -0400618 __wb_writeout_inc(wb);
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700619 local_irq_restore(flags);
620}
Tejun Heo93f78d82015-05-22 17:13:27 -0400621EXPORT_SYMBOL_GPL(wb_writeout_inc);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700622
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700623/*
Jan Karaeb608e32012-05-24 18:59:11 +0200624 * On idle system, we can be called long after we scheduled because we use
625 * deferred timers so count with missed periods.
626 */
627static void writeout_period(unsigned long t)
628{
Tejun Heo380c27c2015-05-22 18:23:21 -0400629 struct wb_domain *dom = (void *)t;
630 int miss_periods = (jiffies - dom->period_time) /
Jan Karaeb608e32012-05-24 18:59:11 +0200631 VM_COMPLETIONS_PERIOD_LEN;
632
Tejun Heo380c27c2015-05-22 18:23:21 -0400633 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
634 dom->period_time = wp_next_time(dom->period_time +
Jan Karaeb608e32012-05-24 18:59:11 +0200635 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
Tejun Heo380c27c2015-05-22 18:23:21 -0400636 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200637 } else {
638 /*
639 * Aging has zeroed all fractions. Stop wasting CPU on period
640 * updates.
641 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400642 dom->period_time = 0;
Jan Karaeb608e32012-05-24 18:59:11 +0200643 }
644}
645
Tejun Heo380c27c2015-05-22 18:23:21 -0400646int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
647{
648 memset(dom, 0, sizeof(*dom));
Tejun Heodcc25ae2015-05-22 18:23:22 -0400649
650 spin_lock_init(&dom->lock);
651
Tejun Heo380c27c2015-05-22 18:23:21 -0400652 init_timer_deferrable(&dom->period_timer);
653 dom->period_timer.function = writeout_period;
654 dom->period_timer.data = (unsigned long)dom;
Tejun Heodcc25ae2015-05-22 18:23:22 -0400655
656 dom->dirty_limit_tstamp = jiffies;
657
Tejun Heo380c27c2015-05-22 18:23:21 -0400658 return fprop_global_init(&dom->completions, gfp);
659}
660
Tejun Heo841710a2015-05-22 18:23:33 -0400661#ifdef CONFIG_CGROUP_WRITEBACK
662void wb_domain_exit(struct wb_domain *dom)
663{
664 del_timer_sync(&dom->period_timer);
665 fprop_global_destroy(&dom->completions);
666}
667#endif
668
Jan Karaeb608e32012-05-24 18:59:11 +0200669/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700670 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
671 * registered backing devices, which, for obvious reasons, can not
672 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700673 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700674static unsigned int bdi_min_ratio;
675
676int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
677{
678 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700679
Jens Axboecfc4ba52009-09-14 13:12:40 +0200680 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700681 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700682 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700683 } else {
684 min_ratio -= bdi->min_ratio;
685 if (bdi_min_ratio + min_ratio < 100) {
686 bdi_min_ratio += min_ratio;
687 bdi->min_ratio += min_ratio;
688 } else {
689 ret = -EINVAL;
690 }
691 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200692 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700693
694 return ret;
695}
696
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700697int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
698{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700699 int ret = 0;
700
701 if (max_ratio > 100)
702 return -EINVAL;
703
Jens Axboecfc4ba52009-09-14 13:12:40 +0200704 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700705 if (bdi->min_ratio > max_ratio) {
706 ret = -EINVAL;
707 } else {
708 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200709 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700710 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200711 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700712
713 return ret;
714}
715EXPORT_SYMBOL(bdi_set_max_ratio);
716
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600717static unsigned long dirty_freerun_ceiling(unsigned long thresh,
718 unsigned long bg_thresh)
719{
720 return (thresh + bg_thresh) / 2;
721}
722
Tejun Heoc7981432015-05-22 18:23:29 -0400723static unsigned long hard_dirty_limit(struct wb_domain *dom,
724 unsigned long thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -0600725{
Tejun Heodcc25ae2015-05-22 18:23:22 -0400726 return max(thresh, dom->dirty_limit);
Wu Fengguangffd1f602011-06-19 22:18:42 -0600727}
728
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400729/*
730 * Memory which can be further allocated to a memcg domain is capped by
731 * system-wide clean memory excluding the amount being used in the domain.
732 */
733static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
734 unsigned long filepages, unsigned long headroom)
Tejun Heoc2aa7232015-05-22 18:23:35 -0400735{
736 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400737 unsigned long clean = filepages - min(filepages, mdtc->dirty);
738 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
739 unsigned long other_clean = global_clean - min(global_clean, clean);
Tejun Heoc2aa7232015-05-22 18:23:35 -0400740
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400741 mdtc->avail = filepages + min(headroom, other_clean);
Christoph Lameter1b424462007-05-06 14:48:59 -0700742}
743
Wu Fengguang6f718652011-03-02 17:14:34 -0600744/**
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400745 * __wb_calc_thresh - @wb's share of dirty throttling threshold
746 * @dtc: dirty_throttle_context of interest
Wu Fengguang1babe182010-08-11 14:17:40 -0700747 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400748 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
Wu Fengguang6f718652011-03-02 17:14:34 -0600749 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600750 *
751 * Note that balance_dirty_pages() will only seriously take it as a hard limit
752 * when sleeping max_pause per page is not enough to keep the dirty pages under
753 * control. For example, when the device is completely stalled due to some error
754 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
755 * In the other normal situations, it acts more gently by throttling the tasks
Tejun Heoa88a3412015-05-22 17:13:28 -0400756 * more (rather than completely block them) when the wb dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600757 *
758 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700759 * - starving fast devices
760 * - piling up dirty pages (that will take long time to sync) on slow devices
761 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400762 * The wb's share of dirty limit will be adapting to its throughput and
Wu Fengguang1babe182010-08-11 14:17:40 -0700763 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
764 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400765static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
Wu Fengguang16c40422010-08-11 14:17:39 -0700766{
Tejun Heoe9f07df2015-05-22 18:23:28 -0400767 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400768 unsigned long thresh = dtc->thresh;
Tejun Heo0d960a32015-05-22 18:23:19 -0400769 u64 wb_thresh;
Wu Fengguang16c40422010-08-11 14:17:39 -0700770 long numerator, denominator;
Tejun Heo693108a2015-05-22 17:13:49 -0400771 unsigned long wb_min_ratio, wb_max_ratio;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700772
Wu Fengguang16c40422010-08-11 14:17:39 -0700773 /*
Tejun Heo0d960a32015-05-22 18:23:19 -0400774 * Calculate this BDI's share of the thresh ratio.
Wu Fengguang16c40422010-08-11 14:17:39 -0700775 */
Tejun Heoe9770b32015-05-22 18:23:27 -0400776 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
Tejun Heo380c27c2015-05-22 18:23:21 -0400777 &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700778
Tejun Heo0d960a32015-05-22 18:23:19 -0400779 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
780 wb_thresh *= numerator;
781 do_div(wb_thresh, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700782
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400783 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
Wu Fengguang16c40422010-08-11 14:17:39 -0700784
Tejun Heo0d960a32015-05-22 18:23:19 -0400785 wb_thresh += (thresh * wb_min_ratio) / 100;
786 if (wb_thresh > (thresh * wb_max_ratio) / 100)
787 wb_thresh = thresh * wb_max_ratio / 100;
Wu Fengguang16c40422010-08-11 14:17:39 -0700788
Tejun Heo0d960a32015-05-22 18:23:19 -0400789 return wb_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790}
791
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400792unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
793{
794 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
795 .thresh = thresh };
796 return __wb_calc_thresh(&gdtc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797}
798
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600799/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700800 * setpoint - dirty 3
801 * f(dirty) := 1.0 + (----------------)
802 * limit - setpoint
803 *
804 * it's a 3rd order polynomial that subjects to
805 *
806 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
807 * (2) f(setpoint) = 1.0 => the balance point
808 * (3) f(limit) = 0 => the hard limit
809 * (4) df/dx <= 0 => negative feedback control
810 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
811 * => fast response on large errors; small oscillation near setpoint
812 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700813static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700814 unsigned long dirty,
815 unsigned long limit)
816{
817 long long pos_ratio;
818 long x;
819
Rik van Rield5c9fde2014-05-06 12:50:01 -0700820 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Tejun Heo464d1382015-04-21 16:49:13 -0400821 (limit - setpoint) | 1);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700822 pos_ratio = x;
823 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
824 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
825 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
826
827 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
828}
829
830/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600831 * Dirty position control.
832 *
833 * (o) global/bdi setpoints
834 *
Tejun Heode1fff32015-05-22 17:13:29 -0400835 * We want the dirty pages be balanced around the global/wb setpoints.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600836 * When the number of dirty pages is higher/lower than the setpoint, the
837 * dirty position control ratio (and hence task dirty ratelimit) will be
838 * decreased/increased to bring the dirty pages back to the setpoint.
839 *
840 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
841 *
842 * if (dirty < setpoint) scale up pos_ratio
843 * if (dirty > setpoint) scale down pos_ratio
844 *
Tejun Heode1fff32015-05-22 17:13:29 -0400845 * if (wb_dirty < wb_setpoint) scale up pos_ratio
846 * if (wb_dirty > wb_setpoint) scale down pos_ratio
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600847 *
848 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
849 *
850 * (o) global control line
851 *
852 * ^ pos_ratio
853 * |
854 * | |<===== global dirty control scope ======>|
855 * 2.0 .............*
856 * | .*
857 * | . *
858 * | . *
859 * | . *
860 * | . *
861 * | . *
862 * 1.0 ................................*
863 * | . . *
864 * | . . *
865 * | . . *
866 * | . . *
867 * | . . *
868 * 0 +------------.------------------.----------------------*------------->
869 * freerun^ setpoint^ limit^ dirty pages
870 *
Tejun Heode1fff32015-05-22 17:13:29 -0400871 * (o) wb control line
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600872 *
873 * ^ pos_ratio
874 * |
875 * | *
876 * | *
877 * | *
878 * | *
879 * | * |<=========== span ============>|
880 * 1.0 .......................*
881 * | . *
882 * | . *
883 * | . *
884 * | . *
885 * | . *
886 * | . *
887 * | . *
888 * | . *
889 * | . *
890 * | . *
891 * | . *
892 * 1/4 ...............................................* * * * * * * * * * * *
893 * | . .
894 * | . .
895 * | . .
896 * 0 +----------------------.-------------------------------.------------->
Tejun Heode1fff32015-05-22 17:13:29 -0400897 * wb_setpoint^ x_intercept^
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600898 *
Tejun Heode1fff32015-05-22 17:13:29 -0400899 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600900 * be smoothly throttled down to normal if it starts high in situations like
901 * - start writing to a slow SD card and a fast disk at the same time. The SD
Tejun Heode1fff32015-05-22 17:13:29 -0400902 * card's wb_dirty may rush to many times higher than wb_setpoint.
903 * - the wb dirty thresh drops quickly due to change of JBOD workload
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600904 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400905static void wb_position_ratio(struct dirty_throttle_control *dtc)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600906{
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400907 struct bdi_writeback *wb = dtc->wb;
Tejun Heoa88a3412015-05-22 17:13:28 -0400908 unsigned long write_bw = wb->avg_write_bandwidth;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400909 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -0400910 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400911 unsigned long wb_thresh = dtc->wb_thresh;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600912 unsigned long x_intercept;
913 unsigned long setpoint; /* dirty pages' target balance point */
Tejun Heode1fff32015-05-22 17:13:29 -0400914 unsigned long wb_setpoint;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600915 unsigned long span;
916 long long pos_ratio; /* for scaling up/down the rate limit */
917 long x;
918
Tejun Heodaddfa32015-05-22 18:23:26 -0400919 dtc->pos_ratio = 0;
920
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400921 if (unlikely(dtc->dirty >= limit))
Tejun Heodaddfa32015-05-22 18:23:26 -0400922 return;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600923
924 /*
925 * global setpoint
926 *
Maxim Patlasov5a537482013-09-11 14:22:46 -0700927 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600928 */
929 setpoint = (freerun + limit) / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400930 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700931
932 /*
933 * The strictlimit feature is a tool preventing mistrusted filesystems
934 * from growing a large number of dirty pages before throttling. For
Tejun Heode1fff32015-05-22 17:13:29 -0400935 * such filesystems balance_dirty_pages always checks wb counters
936 * against wb limits. Even if global "nr_dirty" is under "freerun".
Maxim Patlasov5a537482013-09-11 14:22:46 -0700937 * This is especially important for fuse which sets bdi->max_ratio to
938 * 1% by default. Without strictlimit feature, fuse writeback may
939 * consume arbitrary amount of RAM because it is accounted in
940 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
941 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400942 * Here, in wb_position_ratio(), we calculate pos_ratio based on
Tejun Heode1fff32015-05-22 17:13:29 -0400943 * two values: wb_dirty and wb_thresh. Let's consider an example:
Maxim Patlasov5a537482013-09-11 14:22:46 -0700944 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
945 * limits are set by default to 10% and 20% (background and throttle).
Tejun Heode1fff32015-05-22 17:13:29 -0400946 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
Tejun Heo0d960a32015-05-22 18:23:19 -0400947 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
Tejun Heode1fff32015-05-22 17:13:29 -0400948 * about ~6K pages (as the average of background and throttle wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700949 * limits). The 3rd order polynomial will provide positive feedback if
Tejun Heode1fff32015-05-22 17:13:29 -0400950 * wb_dirty is under wb_setpoint and vice versa.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700951 *
952 * Note, that we cannot use global counters in these calculations
Tejun Heode1fff32015-05-22 17:13:29 -0400953 * because we want to throttle process writing to a strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700954 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
955 * in the example above).
956 */
Tejun Heoa88a3412015-05-22 17:13:28 -0400957 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heode1fff32015-05-22 17:13:29 -0400958 long long wb_pos_ratio;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700959
Tejun Heodaddfa32015-05-22 18:23:26 -0400960 if (dtc->wb_dirty < 8) {
961 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
962 2 << RATELIMIT_CALC_SHIFT);
963 return;
964 }
Maxim Patlasov5a537482013-09-11 14:22:46 -0700965
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400966 if (dtc->wb_dirty >= wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400967 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700968
Tejun Heo970fb012015-05-22 18:23:24 -0400969 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
970 dtc->wb_bg_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700971
Tejun Heode1fff32015-05-22 17:13:29 -0400972 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400973 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700974
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400975 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
Tejun Heode1fff32015-05-22 17:13:29 -0400976 wb_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700977
978 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400979 * Typically, for strictlimit case, wb_setpoint << setpoint
980 * and pos_ratio >> wb_pos_ratio. In the other words global
Maxim Patlasov5a537482013-09-11 14:22:46 -0700981 * state ("dirty") is not limiting factor and we have to
Tejun Heode1fff32015-05-22 17:13:29 -0400982 * make decision based on wb counters. But there is an
Maxim Patlasov5a537482013-09-11 14:22:46 -0700983 * important case when global pos_ratio should get precedence:
984 * global limits are exceeded (e.g. due to activities on other
Tejun Heode1fff32015-05-22 17:13:29 -0400985 * wb's) while given strictlimit wb is below limit.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700986 *
Tejun Heode1fff32015-05-22 17:13:29 -0400987 * "pos_ratio * wb_pos_ratio" would work for the case above,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700988 * but it would look too non-natural for the case of all
Tejun Heode1fff32015-05-22 17:13:29 -0400989 * activity in the system coming from a single strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700990 * with bdi->max_ratio == 100%.
991 *
992 * Note that min() below somewhat changes the dynamics of the
993 * control system. Normally, pos_ratio value can be well over 3
Tejun Heode1fff32015-05-22 17:13:29 -0400994 * (when globally we are at freerun and wb is well below wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700995 * setpoint). Now the maximum pos_ratio in the same situation
996 * is 2. We might want to tweak this if we observe the control
997 * system is too slow to adapt.
998 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400999 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1000 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001001 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001002
1003 /*
1004 * We have computed basic pos_ratio above based on global situation. If
Tejun Heode1fff32015-05-22 17:13:29 -04001005 * the wb is over/under its share of dirty pages, we want to scale
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001006 * pos_ratio further down/up. That is done by the following mechanism.
1007 */
1008
1009 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001010 * wb setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001011 *
Tejun Heode1fff32015-05-22 17:13:29 -04001012 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001013 *
Tejun Heode1fff32015-05-22 17:13:29 -04001014 * x_intercept - wb_dirty
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001015 * := --------------------------
Tejun Heode1fff32015-05-22 17:13:29 -04001016 * x_intercept - wb_setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001017 *
Tejun Heode1fff32015-05-22 17:13:29 -04001018 * The main wb control line is a linear function that subjects to
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001019 *
Tejun Heode1fff32015-05-22 17:13:29 -04001020 * (1) f(wb_setpoint) = 1.0
1021 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1022 * or equally: x_intercept = wb_setpoint + 8 * write_bw
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001023 *
Tejun Heode1fff32015-05-22 17:13:29 -04001024 * For single wb case, the dirty pages are observed to fluctuate
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001025 * regularly within range
Tejun Heode1fff32015-05-22 17:13:29 -04001026 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001027 * for various filesystems, where (2) can yield in a reasonable 12.5%
1028 * fluctuation range for pos_ratio.
1029 *
Tejun Heode1fff32015-05-22 17:13:29 -04001030 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001031 * own size, so move the slope over accordingly and choose a slope that
Tejun Heode1fff32015-05-22 17:13:29 -04001032 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001033 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001034 if (unlikely(wb_thresh > dtc->thresh))
1035 wb_thresh = dtc->thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001036 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001037 * It's very possible that wb_thresh is close to 0 not because the
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001038 * device is slow, but that it has remained inactive for long time.
1039 * Honour such devices a reasonable good (hopefully IO efficient)
1040 * threshold, so that the occasional writes won't be blocked and active
1041 * writes can rampup the threshold quickly.
1042 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001043 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001044 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001045 * scale global setpoint to wb's:
1046 * wb_setpoint = setpoint * wb_thresh / thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001047 */
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001048 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
Tejun Heode1fff32015-05-22 17:13:29 -04001049 wb_setpoint = setpoint * (u64)x >> 16;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001050 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001051 * Use span=(8*write_bw) in single wb case as indicated by
1052 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001053 *
Tejun Heode1fff32015-05-22 17:13:29 -04001054 * wb_thresh thresh - wb_thresh
1055 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1056 * thresh thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001057 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001058 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
Tejun Heode1fff32015-05-22 17:13:29 -04001059 x_intercept = wb_setpoint + span;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001060
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001061 if (dtc->wb_dirty < x_intercept - span / 4) {
1062 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001063 (x_intercept - wb_setpoint) | 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001064 } else
1065 pos_ratio /= 4;
1066
Wu Fengguang8927f662011-08-04 22:16:46 -06001067 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001068 * wb reserve area, safeguard against dirty pool underrun and disk idle
Wu Fengguang8927f662011-08-04 22:16:46 -06001069 * It may push the desired control point of global dirty pages higher
1070 * than setpoint.
1071 */
Tejun Heode1fff32015-05-22 17:13:29 -04001072 x_intercept = wb_thresh / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001073 if (dtc->wb_dirty < x_intercept) {
1074 if (dtc->wb_dirty > x_intercept / 8)
1075 pos_ratio = div_u64(pos_ratio * x_intercept,
1076 dtc->wb_dirty);
Wu Fengguang50657fc2011-10-11 17:06:33 -06001077 else
Wu Fengguang8927f662011-08-04 22:16:46 -06001078 pos_ratio *= 8;
1079 }
1080
Tejun Heodaddfa32015-05-22 18:23:26 -04001081 dtc->pos_ratio = pos_ratio;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001082}
1083
Tejun Heoa88a3412015-05-22 17:13:28 -04001084static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1085 unsigned long elapsed,
1086 unsigned long written)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001087{
1088 const unsigned long period = roundup_pow_of_two(3 * HZ);
Tejun Heoa88a3412015-05-22 17:13:28 -04001089 unsigned long avg = wb->avg_write_bandwidth;
1090 unsigned long old = wb->write_bandwidth;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001091 u64 bw;
1092
1093 /*
1094 * bw = written * HZ / elapsed
1095 *
1096 * bw * elapsed + write_bandwidth * (period - elapsed)
1097 * write_bandwidth = ---------------------------------------------------
1098 * period
Tejun Heoc72efb62015-03-23 00:18:48 -04001099 *
1100 * @written may have decreased due to account_page_redirty().
1101 * Avoid underflowing @bw calculation.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001102 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001103 bw = written - min(written, wb->written_stamp);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001104 bw *= HZ;
1105 if (unlikely(elapsed > period)) {
1106 do_div(bw, elapsed);
1107 avg = bw;
1108 goto out;
1109 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001110 bw += (u64)wb->write_bandwidth * (period - elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001111 bw >>= ilog2(period);
1112
1113 /*
1114 * one more level of smoothing, for filtering out sudden spikes
1115 */
1116 if (avg > old && old >= (unsigned long)bw)
1117 avg -= (avg - old) >> 3;
1118
1119 if (avg < old && old <= (unsigned long)bw)
1120 avg += (old - avg) >> 3;
1121
1122out:
Tejun Heo95a46c62015-05-22 17:13:47 -04001123 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1124 avg = max(avg, 1LU);
1125 if (wb_has_dirty_io(wb)) {
1126 long delta = avg - wb->avg_write_bandwidth;
1127 WARN_ON_ONCE(atomic_long_add_return(delta,
1128 &wb->bdi->tot_write_bandwidth) <= 0);
1129 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001130 wb->write_bandwidth = bw;
1131 wb->avg_write_bandwidth = avg;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001132}
1133
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001134static void update_dirty_limit(struct dirty_throttle_control *dtc)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001135{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001136 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001137 unsigned long thresh = dtc->thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04001138 unsigned long limit = dom->dirty_limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001139
1140 /*
1141 * Follow up in one step.
1142 */
1143 if (limit < thresh) {
1144 limit = thresh;
1145 goto update;
1146 }
1147
1148 /*
1149 * Follow down slowly. Use the higher one as the target, because thresh
1150 * may drop below dirty. This is exactly the reason to introduce
Tejun Heodcc25ae2015-05-22 18:23:22 -04001151 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
Wu Fengguangc42843f2011-03-02 15:54:09 -06001152 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001153 thresh = max(thresh, dtc->dirty);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001154 if (limit > thresh) {
1155 limit -= (limit - thresh) >> 5;
1156 goto update;
1157 }
1158 return;
1159update:
Tejun Heodcc25ae2015-05-22 18:23:22 -04001160 dom->dirty_limit = limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001161}
1162
Tejun Heoe9f07df2015-05-22 18:23:28 -04001163static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001164 unsigned long now)
1165{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001166 struct wb_domain *dom = dtc_dom(dtc);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001167
1168 /*
1169 * check locklessly first to optimize away locking for the most time
1170 */
Tejun Heodcc25ae2015-05-22 18:23:22 -04001171 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
Wu Fengguangc42843f2011-03-02 15:54:09 -06001172 return;
1173
Tejun Heodcc25ae2015-05-22 18:23:22 -04001174 spin_lock(&dom->lock);
1175 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001176 update_dirty_limit(dtc);
Tejun Heodcc25ae2015-05-22 18:23:22 -04001177 dom->dirty_limit_tstamp = now;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001178 }
Tejun Heodcc25ae2015-05-22 18:23:22 -04001179 spin_unlock(&dom->lock);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001180}
1181
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001182/*
Tejun Heode1fff32015-05-22 17:13:29 -04001183 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001184 *
Tejun Heode1fff32015-05-22 17:13:29 -04001185 * Normal wb tasks will be curbed at or below it in long term.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001186 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1187 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001188static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
Tejun Heoa88a3412015-05-22 17:13:28 -04001189 unsigned long dirtied,
1190 unsigned long elapsed)
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001191{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001192 struct bdi_writeback *wb = dtc->wb;
1193 unsigned long dirty = dtc->dirty;
1194 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001195 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Wu Fengguang73811312011-08-26 15:53:24 -06001196 unsigned long setpoint = (freerun + limit) / 2;
Tejun Heoa88a3412015-05-22 17:13:28 -04001197 unsigned long write_bw = wb->avg_write_bandwidth;
1198 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001199 unsigned long dirty_rate;
1200 unsigned long task_ratelimit;
1201 unsigned long balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001202 unsigned long step;
1203 unsigned long x;
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001204 unsigned long shift;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001205
1206 /*
1207 * The dirty rate will match the writeout rate in long term, except
1208 * when dirty pages are truncated by userspace or re-dirtied by FS.
1209 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001210 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001211
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001212 /*
1213 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1214 */
1215 task_ratelimit = (u64)dirty_ratelimit *
Tejun Heodaddfa32015-05-22 18:23:26 -04001216 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001217 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1218
1219 /*
1220 * A linear estimation of the "balanced" throttle rate. The theory is,
Tejun Heode1fff32015-05-22 17:13:29 -04001221 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001222 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1223 * formula will yield the balanced rate limit (write_bw / N).
1224 *
1225 * Note that the expanded form is not a pure rate feedback:
1226 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1227 * but also takes pos_ratio into account:
1228 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1229 *
1230 * (1) is not realistic because pos_ratio also takes part in balancing
1231 * the dirty rate. Consider the state
1232 * pos_ratio = 0.5 (3)
1233 * rate = 2 * (write_bw / N) (4)
1234 * If (1) is used, it will stuck in that state! Because each dd will
1235 * be throttled at
1236 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1237 * yielding
1238 * dirty_rate = N * task_ratelimit = write_bw (6)
1239 * put (6) into (1) we get
1240 * rate_(i+1) = rate_(i) (7)
1241 *
1242 * So we end up using (2) to always keep
1243 * rate_(i+1) ~= (write_bw / N) (8)
1244 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1245 * pos_ratio is able to drive itself to 1.0, which is not only where
1246 * the dirty count meet the setpoint, but also where the slope of
1247 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1248 */
1249 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1250 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001251 /*
1252 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1253 */
1254 if (unlikely(balanced_dirty_ratelimit > write_bw))
1255 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001256
Wu Fengguang73811312011-08-26 15:53:24 -06001257 /*
1258 * We could safely do this and return immediately:
1259 *
Tejun Heode1fff32015-05-22 17:13:29 -04001260 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001261 *
1262 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001263 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001264 * limit the step size.
1265 *
1266 * The below code essentially only uses the relative value of
1267 *
1268 * task_ratelimit - dirty_ratelimit
1269 * = (pos_ratio - 1) * dirty_ratelimit
1270 *
1271 * which reflects the direction and size of dirty position error.
1272 */
1273
1274 /*
1275 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1276 * task_ratelimit is on the same side of dirty_ratelimit, too.
1277 * For example, when
1278 * - dirty_ratelimit > balanced_dirty_ratelimit
1279 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1280 * lowering dirty_ratelimit will help meet both the position and rate
1281 * control targets. Otherwise, don't update dirty_ratelimit if it will
1282 * only help meet the rate target. After all, what the users ultimately
1283 * feel and care are stable dirty rate and small position error.
1284 *
1285 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001286 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001287 * keeps jumping around randomly and can even leap far away at times
1288 * due to the small 200ms estimation period of dirty_rate (we want to
1289 * keep that period small to reduce time lags).
1290 */
1291 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001292
1293 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001294 * For strictlimit case, calculations above were based on wb counters
Tejun Heoa88a3412015-05-22 17:13:28 -04001295 * and limits (starting from pos_ratio = wb_position_ratio() and up to
Maxim Patlasov5a537482013-09-11 14:22:46 -07001296 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
Tejun Heode1fff32015-05-22 17:13:29 -04001297 * Hence, to calculate "step" properly, we have to use wb_dirty as
1298 * "dirty" and wb_setpoint as "setpoint".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001299 *
Tejun Heode1fff32015-05-22 17:13:29 -04001300 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1301 * it's possible that wb_thresh is close to zero due to inactivity
Tejun Heo970fb012015-05-22 18:23:24 -04001302 * of backing device.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001303 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001304 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001305 dirty = dtc->wb_dirty;
1306 if (dtc->wb_dirty < 8)
1307 setpoint = dtc->wb_dirty + 1;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001308 else
Tejun Heo970fb012015-05-22 18:23:24 -04001309 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001310 }
1311
Wu Fengguang73811312011-08-26 15:53:24 -06001312 if (dirty < setpoint) {
Tejun Heoa88a3412015-05-22 17:13:28 -04001313 x = min3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001314 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001315 if (dirty_ratelimit < x)
1316 step = x - dirty_ratelimit;
1317 } else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001318 x = max3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001319 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001320 if (dirty_ratelimit > x)
1321 step = dirty_ratelimit - x;
1322 }
1323
1324 /*
1325 * Don't pursue 100% rate matching. It's impossible since the balanced
1326 * rate itself is constantly fluctuating. So decrease the track speed
1327 * when it gets close to the target. Helps eliminate pointless tremors.
1328 */
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001329 shift = dirty_ratelimit / (2 * step + 1);
1330 if (shift < BITS_PER_LONG)
1331 step = DIV_ROUND_UP(step >> shift, 8);
1332 else
1333 step = 0;
Wu Fengguang73811312011-08-26 15:53:24 -06001334
1335 if (dirty_ratelimit < balanced_dirty_ratelimit)
1336 dirty_ratelimit += step;
1337 else
1338 dirty_ratelimit -= step;
1339
Tejun Heoa88a3412015-05-22 17:13:28 -04001340 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1341 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001342
Tejun Heo5634cc22015-08-18 14:54:56 -07001343 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001344}
1345
Tejun Heoc2aa7232015-05-22 18:23:35 -04001346static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1347 struct dirty_throttle_control *mdtc,
Tejun Heo8a731792015-05-22 18:23:20 -04001348 unsigned long start_time,
1349 bool update_ratelimit)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001350{
Tejun Heoc2aa7232015-05-22 18:23:35 -04001351 struct bdi_writeback *wb = gdtc->wb;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001352 unsigned long now = jiffies;
Tejun Heoa88a3412015-05-22 17:13:28 -04001353 unsigned long elapsed = now - wb->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001354 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001355 unsigned long written;
1356
Tejun Heo8a731792015-05-22 18:23:20 -04001357 lockdep_assert_held(&wb->list_lock);
1358
Wu Fengguange98be2d2010-08-29 11:22:30 -06001359 /*
1360 * rate-limit, only update once every 200ms.
1361 */
1362 if (elapsed < BANDWIDTH_INTERVAL)
1363 return;
1364
Tejun Heoa88a3412015-05-22 17:13:28 -04001365 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1366 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001367
1368 /*
1369 * Skip quiet periods when disk bandwidth is under-utilized.
1370 * (at least 1s idle time between two flusher runs)
1371 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001372 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
Wu Fengguange98be2d2010-08-29 11:22:30 -06001373 goto snapshot;
1374
Tejun Heo8a731792015-05-22 18:23:20 -04001375 if (update_ratelimit) {
Tejun Heoc2aa7232015-05-22 18:23:35 -04001376 domain_update_bandwidth(gdtc, now);
1377 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1378
1379 /*
1380 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1381 * compiler has no way to figure that out. Help it.
1382 */
1383 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1384 domain_update_bandwidth(mdtc, now);
1385 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1386 }
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001387 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001388 wb_update_write_bandwidth(wb, elapsed, written);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001389
1390snapshot:
Tejun Heoa88a3412015-05-22 17:13:28 -04001391 wb->dirtied_stamp = dirtied;
1392 wb->written_stamp = written;
1393 wb->bw_time_stamp = now;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001394}
1395
Tejun Heo8a731792015-05-22 18:23:20 -04001396void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001397{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001398 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1399
Tejun Heoc2aa7232015-05-22 18:23:35 -04001400 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001401}
1402
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001404 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001405 * will look to see if it needs to start dirty throttling.
1406 *
1407 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1408 * global_page_state() too often. So scale it near-sqrt to the safety margin
1409 * (the number of pages we may dirty without exceeding the dirty limits).
1410 */
1411static unsigned long dirty_poll_interval(unsigned long dirty,
1412 unsigned long thresh)
1413{
1414 if (thresh > dirty)
1415 return 1UL << (ilog2(thresh - dirty) >> 1);
1416
1417 return 1;
1418}
1419
Tejun Heoa88a3412015-05-22 17:13:28 -04001420static unsigned long wb_max_pause(struct bdi_writeback *wb,
Tejun Heode1fff32015-05-22 17:13:29 -04001421 unsigned long wb_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001422{
Tejun Heoa88a3412015-05-22 17:13:28 -04001423 unsigned long bw = wb->avg_write_bandwidth;
Fengguang Wue3b6c652013-10-16 13:47:03 -07001424 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001425
1426 /*
1427 * Limit pause time for small memory systems. If sleeping for too long
1428 * time, a small pool of dirty/writeback pages may go empty and disk go
1429 * idle.
1430 *
1431 * 8 serves as the safety ratio.
1432 */
Tejun Heode1fff32015-05-22 17:13:29 -04001433 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001434 t++;
1435
Fengguang Wue3b6c652013-10-16 13:47:03 -07001436 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001437}
1438
Tejun Heoa88a3412015-05-22 17:13:28 -04001439static long wb_min_pause(struct bdi_writeback *wb,
1440 long max_pause,
1441 unsigned long task_ratelimit,
1442 unsigned long dirty_ratelimit,
1443 int *nr_dirtied_pause)
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001444{
Tejun Heoa88a3412015-05-22 17:13:28 -04001445 long hi = ilog2(wb->avg_write_bandwidth);
1446 long lo = ilog2(wb->dirty_ratelimit);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001447 long t; /* target pause */
1448 long pause; /* estimated next pause */
1449 int pages; /* target nr_dirtied_pause */
1450
1451 /* target for 10ms pause on 1-dd case */
1452 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001453
1454 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001455 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1456 * overheads.
1457 *
1458 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001459 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001460 if (hi > lo)
1461 t += (hi - lo) * (10 * HZ) / 1024;
1462
1463 /*
1464 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1465 * on the much more stable dirty_ratelimit. However the next pause time
1466 * will be computed based on task_ratelimit and the two rate limits may
1467 * depart considerably at some time. Especially if task_ratelimit goes
1468 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1469 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1470 * result task_ratelimit won't be executed faithfully, which could
1471 * eventually bring down dirty_ratelimit.
1472 *
1473 * We apply two rules to fix it up:
1474 * 1) try to estimate the next pause time and if necessary, use a lower
1475 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1476 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1477 * 2) limit the target pause time to max_pause/2, so that the normal
1478 * small fluctuations of task_ratelimit won't trigger rule (1) and
1479 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1480 */
1481 t = min(t, 1 + max_pause / 2);
1482 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1483
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001484 /*
1485 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1486 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1487 * When the 16 consecutive reads are often interrupted by some dirty
1488 * throttling pause during the async writes, cfq will go into idles
1489 * (deadline is fine). So push nr_dirtied_pause as high as possible
1490 * until reaches DIRTY_POLL_THRESH=32 pages.
1491 */
1492 if (pages < DIRTY_POLL_THRESH) {
1493 t = max_pause;
1494 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1495 if (pages > DIRTY_POLL_THRESH) {
1496 pages = DIRTY_POLL_THRESH;
1497 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1498 }
1499 }
1500
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001501 pause = HZ * pages / (task_ratelimit + 1);
1502 if (pause > max_pause) {
1503 t = max_pause;
1504 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1505 }
1506
1507 *nr_dirtied_pause = pages;
1508 /*
1509 * The minimal pause time will normally be half the target pause time.
1510 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001511 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001512}
1513
Tejun Heo970fb012015-05-22 18:23:24 -04001514static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
Maxim Patlasov5a537482013-09-11 14:22:46 -07001515{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001516 struct bdi_writeback *wb = dtc->wb;
Tejun Heo93f78d82015-05-22 17:13:27 -04001517 unsigned long wb_reclaimable;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001518
1519 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001520 * wb_thresh is not treated as some limiting factor as
Maxim Patlasov5a537482013-09-11 14:22:46 -07001521 * dirty_thresh, due to reasons
Tejun Heode1fff32015-05-22 17:13:29 -04001522 * - in JBOD setup, wb_thresh can fluctuate a lot
Maxim Patlasov5a537482013-09-11 14:22:46 -07001523 * - in a system with HDD and USB key, the USB key may somehow
Tejun Heode1fff32015-05-22 17:13:29 -04001524 * go into state (wb_dirty >> wb_thresh) either because
1525 * wb_dirty starts high, or because wb_thresh drops low.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001526 * In this case we don't want to hard throttle the USB key
Tejun Heode1fff32015-05-22 17:13:29 -04001527 * dirtiers for 100 seconds until wb_dirty drops under
1528 * wb_thresh. Instead the auxiliary wb control line in
Tejun Heoa88a3412015-05-22 17:13:28 -04001529 * wb_position_ratio() will let the dirtier task progress
Tejun Heode1fff32015-05-22 17:13:29 -04001530 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001531 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -04001532 dtc->wb_thresh = __wb_calc_thresh(dtc);
Tejun Heo970fb012015-05-22 18:23:24 -04001533 dtc->wb_bg_thresh = dtc->thresh ?
1534 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001535
1536 /*
1537 * In order to avoid the stacked BDI deadlock we need
1538 * to ensure we accurately count the 'dirty' pages when
1539 * the threshold is low.
1540 *
1541 * Otherwise it would be possible to get thresh+n pages
1542 * reported dirty, even though there are thresh-m pages
1543 * actually dirty; with m+n sitting in the percpu
1544 * deltas.
1545 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001546 if (dtc->wb_thresh < 2 * wb_stat_error(wb)) {
Tejun Heo93f78d82015-05-22 17:13:27 -04001547 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001548 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001549 } else {
Tejun Heo93f78d82015-05-22 17:13:27 -04001550 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001551 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001552 }
1553}
1554
Wu Fengguang9d823e82011-06-11 18:10:12 -06001555/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 * balance_dirty_pages() must be called by processes which are generating dirty
1557 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001558 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001559 * If we're over `background_thresh' then the writeback threads are woken to
1560 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +08001562static void balance_dirty_pages(struct address_space *mapping,
Tejun Heodfb8ae52015-05-22 17:13:40 -04001563 struct bdi_writeback *wb,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001564 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001566 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001567 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001568 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001569 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1570 &mdtc_stor : NULL;
1571 struct dirty_throttle_control *sdtc;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001572 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
Wu Fengguang83712352011-06-11 19:25:42 -06001573 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001574 long pause;
1575 long max_pause;
1576 long min_pause;
1577 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001578 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001579 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001580 unsigned long dirty_ratelimit;
Tejun Heodfb8ae52015-05-22 17:13:40 -04001581 struct backing_dev_info *bdi = wb->bdi;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001582 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001583 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584
1585 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001586 unsigned long now = jiffies;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001587 unsigned long dirty, thresh, bg_thresh;
Yang Shi50e55bf2015-11-20 15:57:10 -08001588 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1589 unsigned long m_thresh = 0;
1590 unsigned long m_bg_thresh = 0;
Wu Fengguang83712352011-06-11 19:25:42 -06001591
Wu Fengguang143dfe82010-08-27 18:45:12 -06001592 /*
1593 * Unstable writes are a feature of certain networked
1594 * filesystems (i.e. NFS) in which data may have been
1595 * written to the server's write cache, but has not yet
1596 * been flushed to permanent storage.
1597 */
Mel Gorman11fb9982016-07-28 15:46:20 -07001598 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1599 global_node_page_state(NR_UNSTABLE_NFS);
Tejun Heo9fc3a432015-05-22 18:23:30 -04001600 gdtc->avail = global_dirtyable_memory();
Mel Gorman11fb9982016-07-28 15:46:20 -07001601 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001602
Tejun Heo9fc3a432015-05-22 18:23:30 -04001603 domain_dirty_limits(gdtc);
Wu Fengguang16c40422010-08-11 14:17:39 -07001604
Maxim Patlasov5a537482013-09-11 14:22:46 -07001605 if (unlikely(strictlimit)) {
Tejun Heo970fb012015-05-22 18:23:24 -04001606 wb_dirty_limits(gdtc);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001607
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001608 dirty = gdtc->wb_dirty;
1609 thresh = gdtc->wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -04001610 bg_thresh = gdtc->wb_bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001611 } else {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001612 dirty = gdtc->dirty;
1613 thresh = gdtc->thresh;
1614 bg_thresh = gdtc->bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001615 }
1616
Tejun Heoc2aa7232015-05-22 18:23:35 -04001617 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001618 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001619
1620 /*
1621 * If @wb belongs to !root memcg, repeat the same
1622 * basic calculations for the memcg domain.
1623 */
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001624 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1625 &mdtc->dirty, &writeback);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001626 mdtc->dirty += writeback;
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001627 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001628
1629 domain_dirty_limits(mdtc);
1630
1631 if (unlikely(strictlimit)) {
1632 wb_dirty_limits(mdtc);
1633 m_dirty = mdtc->wb_dirty;
1634 m_thresh = mdtc->wb_thresh;
1635 m_bg_thresh = mdtc->wb_bg_thresh;
1636 } else {
1637 m_dirty = mdtc->dirty;
1638 m_thresh = mdtc->thresh;
1639 m_bg_thresh = mdtc->bg_thresh;
1640 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001641 }
1642
1643 /*
1644 * Throttle it only when the background writeback cannot
1645 * catch-up. This avoids (excessively) small writeouts
Tejun Heode1fff32015-05-22 17:13:29 -04001646 * when the wb limits are ramping up in case of !strictlimit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001647 *
Tejun Heode1fff32015-05-22 17:13:29 -04001648 * In strictlimit case make decision based on the wb counters
1649 * and limits. Small writeouts when the wb limits are ramping
Maxim Patlasov5a537482013-09-11 14:22:46 -07001650 * up are the price we consciously pay for strictlimit-ing.
Tejun Heoc2aa7232015-05-22 18:23:35 -04001651 *
1652 * If memcg domain is in effect, @dirty should be under
1653 * both global and memcg freerun ceilings.
Wu Fengguang16c40422010-08-11 14:17:39 -07001654 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001655 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1656 (!mdtc ||
1657 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1658 unsigned long intv = dirty_poll_interval(dirty, thresh);
1659 unsigned long m_intv = ULONG_MAX;
1660
Wu Fengguang83712352011-06-11 19:25:42 -06001661 current->dirty_paused_when = now;
1662 current->nr_dirtied = 0;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001663 if (mdtc)
1664 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1665 current->nr_dirtied_pause = min(intv, m_intv);
Wu Fengguang16c40422010-08-11 14:17:39 -07001666 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001667 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001668
Tejun Heobc058732015-05-22 17:13:53 -04001669 if (unlikely(!writeback_in_progress(wb)))
Tejun Heo9ecf48662015-05-22 17:13:54 -04001670 wb_start_background_writeback(wb);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001671
Tejun Heoc2aa7232015-05-22 18:23:35 -04001672 /*
1673 * Calculate global domain's pos_ratio and select the
1674 * global dtc by default.
1675 */
Maxim Patlasov5a537482013-09-11 14:22:46 -07001676 if (!strictlimit)
Tejun Heo970fb012015-05-22 18:23:24 -04001677 wb_dirty_limits(gdtc);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001678
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001679 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1680 ((gdtc->dirty > gdtc->thresh) || strictlimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
Tejun Heodaddfa32015-05-22 18:23:26 -04001682 wb_position_ratio(gdtc);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001683 sdtc = gdtc;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001684
Tejun Heoc2aa7232015-05-22 18:23:35 -04001685 if (mdtc) {
1686 /*
1687 * If memcg domain is in effect, calculate its
1688 * pos_ratio. @wb should satisfy constraints from
1689 * both global and memcg domains. Choose the one
1690 * w/ lower pos_ratio.
1691 */
1692 if (!strictlimit)
1693 wb_dirty_limits(mdtc);
1694
1695 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1696 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1697
1698 wb_position_ratio(mdtc);
1699 if (mdtc->pos_ratio < gdtc->pos_ratio)
1700 sdtc = mdtc;
1701 }
Tejun Heodaddfa32015-05-22 18:23:26 -04001702
Tejun Heoa88a3412015-05-22 17:13:28 -04001703 if (dirty_exceeded && !wb->dirty_exceeded)
1704 wb->dirty_exceeded = 1;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001705
Tejun Heo8a731792015-05-22 18:23:20 -04001706 if (time_is_before_jiffies(wb->bw_time_stamp +
1707 BANDWIDTH_INTERVAL)) {
1708 spin_lock(&wb->list_lock);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001709 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
Tejun Heo8a731792015-05-22 18:23:20 -04001710 spin_unlock(&wb->list_lock);
1711 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712
Tejun Heoc2aa7232015-05-22 18:23:35 -04001713 /* throttle according to the chosen dtc */
Tejun Heoa88a3412015-05-22 17:13:28 -04001714 dirty_ratelimit = wb->dirty_ratelimit;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001715 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001716 RATELIMIT_CALC_SHIFT;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001717 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
Tejun Heoa88a3412015-05-22 17:13:28 -04001718 min_pause = wb_min_pause(wb, max_pause,
1719 task_ratelimit, dirty_ratelimit,
1720 &nr_dirtied_pause);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001721
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001722 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001723 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001724 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001725 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 }
Wu Fengguang83712352011-06-11 19:25:42 -06001727 period = HZ * pages_dirtied / task_ratelimit;
1728 pause = period;
1729 if (current->dirty_paused_when)
1730 pause -= now - current->dirty_paused_when;
1731 /*
1732 * For less than 1s think time (ext3/4 may block the dirtier
1733 * for up to 800ms from time to time on 1-HDD; so does xfs,
1734 * however at much less frequency), try to compensate it in
1735 * future periods by updating the virtual time; otherwise just
1736 * do a reset, as it may be a light dirtier.
1737 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001738 if (pause < min_pause) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001739 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001740 sdtc->thresh,
1741 sdtc->bg_thresh,
1742 sdtc->dirty,
1743 sdtc->wb_thresh,
1744 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001745 dirty_ratelimit,
1746 task_ratelimit,
1747 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001748 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001749 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001750 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001751 if (pause < -HZ) {
1752 current->dirty_paused_when = now;
1753 current->nr_dirtied = 0;
1754 } else if (period) {
1755 current->dirty_paused_when += period;
1756 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001757 } else if (current->nr_dirtied_pause <= pages_dirtied)
1758 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001759 break;
1760 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001761 if (unlikely(pause > max_pause)) {
1762 /* for occasional dropped task_ratelimit */
1763 now += min(pause - max_pause, max_pause);
1764 pause = max_pause;
1765 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001766
1767pause:
Tejun Heo5634cc22015-08-18 14:54:56 -07001768 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001769 sdtc->thresh,
1770 sdtc->bg_thresh,
1771 sdtc->dirty,
1772 sdtc->wb_thresh,
1773 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001774 dirty_ratelimit,
1775 task_ratelimit,
1776 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001777 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001778 pause,
1779 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001780 __set_current_state(TASK_KILLABLE);
Wu Fengguangd25105e2009-10-09 12:40:42 +02001781 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001782
Wu Fengguang83712352011-06-11 19:25:42 -06001783 current->dirty_paused_when = now + pause;
1784 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001785 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001786
Wu Fengguangffd1f602011-06-19 22:18:42 -06001787 /*
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001788 * This is typically equal to (dirty < thresh) and can also
1789 * keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001790 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001791 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001792 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001793
Wu Fengguangc5c63432011-12-02 10:21:33 -06001794 /*
1795 * In the case of an unresponding NFS server and the NFS dirty
Tejun Heode1fff32015-05-22 17:13:29 -04001796 * pages exceeds dirty_thresh, give the other good wb's a pipe
Wu Fengguangc5c63432011-12-02 10:21:33 -06001797 * to go through, so that tasks on them still remain responsive.
1798 *
1799 * In theory 1 page is enough to keep the comsumer-producer
1800 * pipe going: the flusher cleans 1 page => the task dirties 1
Tejun Heode1fff32015-05-22 17:13:29 -04001801 * more page. However wb_dirty has accounting errors. So use
Tejun Heo93f78d82015-05-22 17:13:27 -04001802 * the larger and more IO friendly wb_stat_error.
Wu Fengguangc5c63432011-12-02 10:21:33 -06001803 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001804 if (sdtc->wb_dirty <= wb_stat_error(wb))
Wu Fengguangc5c63432011-12-02 10:21:33 -06001805 break;
1806
Jan Kara499d05e2011-11-16 19:34:48 +08001807 if (fatal_signal_pending(current))
1808 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809 }
1810
Tejun Heoa88a3412015-05-22 17:13:28 -04001811 if (!dirty_exceeded && wb->dirty_exceeded)
1812 wb->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813
Tejun Heobc058732015-05-22 17:13:53 -04001814 if (writeback_in_progress(wb))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001815 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816
1817 /*
1818 * In laptop mode, we wait until hitting the higher threshold before
1819 * starting background writeout, and then write out all the way down
1820 * to the lower threshold. So slow writers cause minimal disk activity.
1821 *
1822 * In normal mode, we start background writeout at the lower
1823 * background_thresh, to keep the amount of dirty memory low.
1824 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001825 if (laptop_mode)
1826 return;
1827
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001828 if (nr_reclaimable > gdtc->bg_thresh)
Tejun Heo9ecf48662015-05-22 17:13:54 -04001829 wb_start_background_writeback(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830}
1831
Wu Fengguang9d823e82011-06-11 18:10:12 -06001832static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001833
Wu Fengguang54848d72011-04-05 13:21:19 -06001834/*
1835 * Normal tasks are throttled by
1836 * loop {
1837 * dirty tsk->nr_dirtied_pause pages;
1838 * take a snap in balance_dirty_pages();
1839 * }
1840 * However there is a worst case. If every task exit immediately when dirtied
1841 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1842 * called to throttle the page dirties. The solution is to save the not yet
1843 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1844 * randomly into the running tasks. This works well for the above worst case,
1845 * as the new task will pick up and accumulate the old task's leaked dirty
1846 * count and eventually get throttled.
1847 */
1848DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1849
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001851 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001852 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 *
1854 * Processes which are dirtying memory should call in here once for each page
1855 * which was newly dirtied. The function will periodically check the system's
1856 * dirty state and will initiate writeback if needed.
1857 *
1858 * On really big machines, get_writeback_state is expensive, so try to avoid
1859 * calling it too often (ratelimiting). But once we're over the dirty memory
1860 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1861 * from overshooting the limit by (ratelimit_pages) each.
1862 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001863void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864{
Tejun Heodfb8ae52015-05-22 17:13:40 -04001865 struct inode *inode = mapping->host;
1866 struct backing_dev_info *bdi = inode_to_bdi(inode);
1867 struct bdi_writeback *wb = NULL;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001868 int ratelimit;
1869 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870
Wu Fengguang36715ce2011-06-11 17:53:57 -06001871 if (!bdi_cap_account_dirty(bdi))
1872 return;
1873
Tejun Heodfb8ae52015-05-22 17:13:40 -04001874 if (inode_cgwb_enabled(inode))
1875 wb = wb_get_create_current(bdi, GFP_KERNEL);
1876 if (!wb)
1877 wb = &bdi->wb;
1878
Wu Fengguang9d823e82011-06-11 18:10:12 -06001879 ratelimit = current->nr_dirtied_pause;
Tejun Heoa88a3412015-05-22 17:13:28 -04001880 if (wb->dirty_exceeded)
Wu Fengguang9d823e82011-06-11 18:10:12 -06001881 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001883 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001884 /*
1885 * This prevents one CPU to accumulate too many dirtied pages without
1886 * calling into balance_dirty_pages(), which can happen when there are
1887 * 1000+ tasks, all of them start dirtying pages at exactly the same
1888 * time, hence all honoured too large initial task->nr_dirtied_pause.
1889 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001890 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001891 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001892 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001893 else if (unlikely(*p >= ratelimit_pages)) {
1894 *p = 0;
1895 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001897 /*
1898 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1899 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1900 * the dirty throttling and livelock other long-run dirtiers.
1901 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001902 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06001903 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001904 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001905 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1906 *p -= nr_pages_dirtied;
1907 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001909 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001910
1911 if (unlikely(current->nr_dirtied >= ratelimit))
Tejun Heodfb8ae52015-05-22 17:13:40 -04001912 balance_dirty_pages(mapping, wb, current->nr_dirtied);
1913
1914 wb_put(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001916EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
Tejun Heoaa661bb2015-05-22 18:23:31 -04001918/**
1919 * wb_over_bg_thresh - does @wb need to be written back?
1920 * @wb: bdi_writeback of interest
1921 *
1922 * Determines whether background writeback should keep writing @wb or it's
1923 * clean enough. Returns %true if writeback should continue.
1924 */
1925bool wb_over_bg_thresh(struct bdi_writeback *wb)
1926{
Tejun Heo947e9762015-05-22 18:23:32 -04001927 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001928 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo947e9762015-05-22 18:23:32 -04001929 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001930 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1931 &mdtc_stor : NULL;
Tejun Heoaa661bb2015-05-22 18:23:31 -04001932
Tejun Heo947e9762015-05-22 18:23:32 -04001933 /*
1934 * Similar to balance_dirty_pages() but ignores pages being written
1935 * as we're trying to decide whether to put more under writeback.
1936 */
1937 gdtc->avail = global_dirtyable_memory();
Mel Gorman11fb9982016-07-28 15:46:20 -07001938 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1939 global_node_page_state(NR_UNSTABLE_NFS);
Tejun Heo947e9762015-05-22 18:23:32 -04001940 domain_dirty_limits(gdtc);
Tejun Heoaa661bb2015-05-22 18:23:31 -04001941
Tejun Heo947e9762015-05-22 18:23:32 -04001942 if (gdtc->dirty > gdtc->bg_thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04001943 return true;
1944
Howard Cochran74d36942016-03-10 01:12:39 -05001945 if (wb_stat(wb, WB_RECLAIMABLE) >
1946 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
Tejun Heoaa661bb2015-05-22 18:23:31 -04001947 return true;
1948
Tejun Heoc2aa7232015-05-22 18:23:35 -04001949 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001950 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001951
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001952 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1953 &writeback);
1954 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001955 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1956
1957 if (mdtc->dirty > mdtc->bg_thresh)
1958 return true;
1959
Howard Cochran74d36942016-03-10 01:12:39 -05001960 if (wb_stat(wb, WB_RECLAIMABLE) >
1961 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
Tejun Heoc2aa7232015-05-22 18:23:35 -04001962 return true;
1963 }
1964
Tejun Heoaa661bb2015-05-22 18:23:31 -04001965 return false;
1966}
1967
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1970 */
Joe Perchescccad5b2014-06-06 14:38:09 -07001971int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001972 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001974 proc_dointvec(table, write, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 return 0;
1976}
1977
Jens Axboec2c49862010-05-20 09:18:47 +02001978#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +02001979void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980{
Matthew Garrett31373d02010-04-06 14:25:14 +02001981 struct request_queue *q = (struct request_queue *)data;
Mel Gorman11fb9982016-07-28 15:46:20 -07001982 int nr_pages = global_node_page_state(NR_FILE_DIRTY) +
1983 global_node_page_state(NR_UNSTABLE_NFS);
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001984 struct bdi_writeback *wb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985
Matthew Garrett31373d02010-04-06 14:25:14 +02001986 /*
1987 * We want to write everything out, not just down to the dirty
1988 * threshold
1989 */
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001990 if (!bdi_has_dirty_io(&q->backing_dev_info))
1991 return;
1992
Tejun Heo9ad18ab2015-09-29 12:47:50 -04001993 rcu_read_lock();
Tejun Heob8175252015-10-02 14:47:05 -04001994 list_for_each_entry_rcu(wb, &q->backing_dev_info.wb_list, bdi_node)
Tejun Heoa06fd6b2015-05-22 17:13:52 -04001995 if (wb_has_dirty_io(wb))
1996 wb_start_writeback(wb, nr_pages, true,
1997 WB_REASON_LAPTOP_TIMER);
Tejun Heo9ad18ab2015-09-29 12:47:50 -04001998 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999}
2000
2001/*
2002 * We've spun up the disk and we're in laptop mode: schedule writeback
2003 * of all dirty data a few seconds from now. If the flush is already scheduled
2004 * then push it back - the user is still using the disk.
2005 */
Matthew Garrett31373d02010-04-06 14:25:14 +02002006void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007{
Matthew Garrett31373d02010-04-06 14:25:14 +02002008 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009}
2010
2011/*
2012 * We're in laptop mode and we've just synced. The sync's writes will have
2013 * caused another writeback to be scheduled by laptop_io_completion.
2014 * Nothing needs to be written back anymore, so we unschedule the writeback.
2015 */
2016void laptop_sync_completion(void)
2017{
Matthew Garrett31373d02010-04-06 14:25:14 +02002018 struct backing_dev_info *bdi;
2019
2020 rcu_read_lock();
2021
2022 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2023 del_timer(&bdi->laptop_mode_wb_timer);
2024
2025 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026}
Jens Axboec2c49862010-05-20 09:18:47 +02002027#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
2029/*
2030 * If ratelimit_pages is too high then we can get into dirty-data overload
2031 * if a large number of processes all perform writes at the same time.
2032 * If it is too low then SMP machines will call the (expensive)
2033 * get_writeback_state too often.
2034 *
2035 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2036 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06002037 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 */
2039
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002040void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041{
Tejun Heodcc25ae2015-05-22 18:23:22 -04002042 struct wb_domain *dom = &global_wb_domain;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002043 unsigned long background_thresh;
2044 unsigned long dirty_thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04002045
Wu Fengguang9d823e82011-06-11 18:10:12 -06002046 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heodcc25ae2015-05-22 18:23:22 -04002047 dom->dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002048 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 if (ratelimit_pages < 16)
2050 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051}
2052
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002053static int page_writeback_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054{
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002055 writeback_set_ratelimit();
2056 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057}
2058
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002060 * Called early on to tune the page writeback dirty limits.
2061 *
2062 * We used to scale dirty pages according to how total memory
2063 * related to pages that could be allocated for buffers (by
2064 * comparing nr_free_buffer_pages() to vm_total_pages.
2065 *
2066 * However, that was when we used "dirty_ratio" to scale with
2067 * all memory, and we don't do that any more. "dirty_ratio"
2068 * is now applied to total non-HIGHPAGE memory (by subtracting
2069 * totalhigh_pages from vm_total_pages), and as such we can't
2070 * get into the old insane situation any more where we had
2071 * large amounts of dirty pages compared to a small amount of
2072 * non-HIGHMEM memory.
2073 *
2074 * But we might still want to scale the dirty_ratio by how
2075 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 */
2077void __init page_writeback_init(void)
2078{
Rabin Vincenta50fcb52015-08-06 15:47:14 -07002079 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2080
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002081 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2082 page_writeback_cpu_online, NULL);
2083 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2084 page_writeback_cpu_online);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085}
2086
David Howells811d7362006-08-29 19:06:09 +01002087/**
Jan Karaf446daae2010-08-09 17:19:12 -07002088 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2089 * @mapping: address space structure to write
2090 * @start: starting page index
2091 * @end: ending page index (inclusive)
2092 *
2093 * This function scans the page range from @start to @end (inclusive) and tags
2094 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2095 * that write_cache_pages (or whoever calls this function) will then use
2096 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2097 * used to avoid livelocking of writeback by a process steadily creating new
2098 * dirty pages in the file (thus it is important for this function to be quick
2099 * so that it can tag pages faster than a dirtying process can create them).
2100 */
2101/*
2102 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
2103 */
Jan Karaf446daae2010-08-09 17:19:12 -07002104void tag_pages_for_writeback(struct address_space *mapping,
2105 pgoff_t start, pgoff_t end)
2106{
Randy Dunlap3c111a02010-08-11 14:17:30 -07002107#define WRITEBACK_TAG_BATCH 4096
Jan Karaf446daae2010-08-09 17:19:12 -07002108 unsigned long tagged;
2109
2110 do {
2111 spin_lock_irq(&mapping->tree_lock);
2112 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
2113 &start, end, WRITEBACK_TAG_BATCH,
2114 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
2115 spin_unlock_irq(&mapping->tree_lock);
2116 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
2117 cond_resched();
Jan Karad5ed3a42010-08-19 14:13:33 -07002118 /* We check 'start' to handle wrapping when end == ~0UL */
2119 } while (tagged >= WRITEBACK_TAG_BATCH && start);
Jan Karaf446daae2010-08-09 17:19:12 -07002120}
2121EXPORT_SYMBOL(tag_pages_for_writeback);
2122
2123/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002124 * 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 +01002125 * @mapping: address space structure to write
2126 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002127 * @writepage: function called for each page
2128 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01002129 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002130 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01002131 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2132 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2133 * and msync() need to guarantee that all the data which was dirty at the time
2134 * the call was made get new I/O started against them. If wbc->sync_mode is
2135 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2136 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07002137 *
2138 * To avoid livelocks (when other process dirties new pages), we first tag
2139 * pages which should be written back with TOWRITE tag and only then start
2140 * writing them. For data-integrity sync we have to be careful so that we do
2141 * not miss some pages (e.g., because some other process has cleared TOWRITE
2142 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2143 * by the process clearing the DIRTY tag (and submitting the page for IO).
Dave Chinner8a7b0a52018-10-26 15:09:45 -07002144 *
2145 * To avoid deadlocks between range_cyclic writeback and callers that hold
2146 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2147 * we do not loop back to the start of the file. Doing so causes a page
2148 * lock/page writeback access order inversion - we should only ever lock
2149 * multiple pages in ascending page->index order, and looping back to the start
2150 * of the file violates that rule and causes deadlocks.
David Howells811d7362006-08-29 19:06:09 +01002151 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002152int write_cache_pages(struct address_space *mapping,
2153 struct writeback_control *wbc, writepage_t writepage,
2154 void *data)
David Howells811d7362006-08-29 19:06:09 +01002155{
David Howells811d7362006-08-29 19:06:09 +01002156 int ret = 0;
2157 int done = 0;
Brian Fosterc6e4be62018-12-28 00:37:20 -08002158 int error;
David Howells811d7362006-08-29 19:06:09 +01002159 struct pagevec pvec;
2160 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08002161 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01002162 pgoff_t index;
2163 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08002164 pgoff_t done_index;
David Howells811d7362006-08-29 19:06:09 +01002165 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07002166 int tag;
David Howells811d7362006-08-29 19:06:09 +01002167
David Howells811d7362006-08-29 19:06:09 +01002168 pagevec_init(&pvec, 0);
2169 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08002170 writeback_index = mapping->writeback_index; /* prev offset */
2171 index = writeback_index;
David Howells811d7362006-08-29 19:06:09 +01002172 end = -1;
2173 } else {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002174 index = wbc->range_start >> PAGE_SHIFT;
2175 end = wbc->range_end >> PAGE_SHIFT;
David Howells811d7362006-08-29 19:06:09 +01002176 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2177 range_whole = 1;
David Howells811d7362006-08-29 19:06:09 +01002178 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002179 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002180 tag = PAGECACHE_TAG_TOWRITE;
2181 else
2182 tag = PAGECACHE_TAG_DIRTY;
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002183 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002184 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08002185 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002186 while (!done && (index <= end)) {
2187 int i;
2188
Jan Karaf446daae2010-08-09 17:19:12 -07002189 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002190 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2191 if (nr_pages == 0)
2192 break;
David Howells811d7362006-08-29 19:06:09 +01002193
David Howells811d7362006-08-29 19:06:09 +01002194 for (i = 0; i < nr_pages; i++) {
2195 struct page *page = pvec.pages[i];
2196
Nick Piggind5482cd2009-01-06 14:39:11 -08002197 /*
2198 * At this point, the page may be truncated or
2199 * invalidated (changing page->mapping to NULL), or
2200 * even swizzled back from swapper_space to tmpfs file
2201 * mapping. However, page->index will not change
2202 * because we have a reference on the page.
2203 */
2204 if (page->index > end) {
2205 /*
2206 * can't be range_cyclic (1st pass) because
2207 * end == -1 in that case.
2208 */
2209 done = 1;
2210 break;
2211 }
2212
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002213 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08002214
David Howells811d7362006-08-29 19:06:09 +01002215 lock_page(page);
2216
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002217 /*
2218 * Page truncated or invalidated. We can freely skip it
2219 * then, even for data integrity operations: the page
2220 * has disappeared concurrently, so there could be no
2221 * real expectation of this data interity operation
2222 * even if there is now a new, dirty page at the same
2223 * pagecache address.
2224 */
David Howells811d7362006-08-29 19:06:09 +01002225 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002226continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01002227 unlock_page(page);
2228 continue;
2229 }
2230
Nick Piggin515f4a02009-01-06 14:39:10 -08002231 if (!PageDirty(page)) {
2232 /* someone wrote it for us */
2233 goto continue_unlock;
2234 }
David Howells811d7362006-08-29 19:06:09 +01002235
Nick Piggin515f4a02009-01-06 14:39:10 -08002236 if (PageWriteback(page)) {
2237 if (wbc->sync_mode != WB_SYNC_NONE)
2238 wait_on_page_writeback(page);
2239 else
2240 goto continue_unlock;
2241 }
2242
2243 BUG_ON(PageWriteback(page));
2244 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002245 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01002246
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002247 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
Brian Fosterc6e4be62018-12-28 00:37:20 -08002248 error = (*writepage)(page, wbc, data);
2249 if (unlikely(error)) {
2250 /*
2251 * Handle errors according to the type of
2252 * writeback. There's no need to continue for
2253 * background writeback. Just push done_index
2254 * past this page so media errors won't choke
2255 * writeout for the entire file. For integrity
2256 * writeback, we must process the entire dirty
2257 * set regardless of errors because the fs may
2258 * still have state to clear for each page. In
2259 * that case we continue processing and return
2260 * the first error.
2261 */
2262 if (error == AOP_WRITEPAGE_ACTIVATE) {
Nick Piggin00266772009-01-06 14:39:06 -08002263 unlock_page(page);
Brian Fosterc6e4be62018-12-28 00:37:20 -08002264 error = 0;
2265 } else if (wbc->sync_mode != WB_SYNC_ALL) {
2266 ret = error;
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002267 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08002268 done = 1;
2269 break;
2270 }
Brian Fosterc6e4be62018-12-28 00:37:20 -08002271 if (!ret)
2272 ret = error;
Dave Chinner0b564922010-06-09 10:37:18 +10002273 }
David Howells811d7362006-08-29 19:06:09 +01002274
Dave Chinner546a1922010-08-24 11:44:34 +10002275 /*
2276 * We stop writing back only if we are not doing
2277 * integrity sync. In case of integrity sync we have to
2278 * keep going until we have written all the pages
2279 * we tagged for writeback prior to entering this loop.
2280 */
2281 if (--wbc->nr_to_write <= 0 &&
2282 wbc->sync_mode == WB_SYNC_NONE) {
2283 done = 1;
2284 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08002285 }
David Howells811d7362006-08-29 19:06:09 +01002286 }
2287 pagevec_release(&pvec);
2288 cond_resched();
2289 }
Dave Chinner8a7b0a52018-10-26 15:09:45 -07002290
2291 /*
2292 * If we hit the last page and there is more work to be done: wrap
2293 * back the index back to the start of the file for the next
2294 * time we are called.
2295 */
2296 if (wbc->range_cyclic && !done)
2297 done_index = 0;
Dave Chinner0b564922010-06-09 10:37:18 +10002298 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2299 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04002300
David Howells811d7362006-08-29 19:06:09 +01002301 return ret;
2302}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002303EXPORT_SYMBOL(write_cache_pages);
2304
2305/*
2306 * Function used by generic_writepages to call the real writepage
2307 * function and set the mapping flags on error
2308 */
2309static int __writepage(struct page *page, struct writeback_control *wbc,
2310 void *data)
2311{
2312 struct address_space *mapping = data;
2313 int ret = mapping->a_ops->writepage(page, wbc);
2314 mapping_set_error(mapping, ret);
2315 return ret;
2316}
2317
2318/**
2319 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2320 * @mapping: address space structure to write
2321 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2322 *
2323 * This is a library function, which implements the writepages()
2324 * address_space_operation.
2325 */
2326int generic_writepages(struct address_space *mapping,
2327 struct writeback_control *wbc)
2328{
Shaohua Li9b6096a2011-03-17 10:47:06 +01002329 struct blk_plug plug;
2330 int ret;
2331
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002332 /* deal with chardevs and other special file */
2333 if (!mapping->a_ops->writepage)
2334 return 0;
2335
Shaohua Li9b6096a2011-03-17 10:47:06 +01002336 blk_start_plug(&plug);
2337 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2338 blk_finish_plug(&plug);
2339 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002340}
David Howells811d7362006-08-29 19:06:09 +01002341
2342EXPORT_SYMBOL(generic_writepages);
2343
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2345{
Andrew Morton22905f72005-11-16 15:07:01 -08002346 int ret;
2347
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 if (wbc->nr_to_write <= 0)
2349 return 0;
2350 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002351 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002352 else
2353 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08002354 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355}
2356
2357/**
2358 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07002359 * @page: the page to write
2360 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 *
2362 * The page must be locked by the caller and will be unlocked upon return.
2363 *
2364 * write_one_page() returns a negative error code if I/O failed.
2365 */
2366int write_one_page(struct page *page, int wait)
2367{
2368 struct address_space *mapping = page->mapping;
2369 int ret = 0;
2370 struct writeback_control wbc = {
2371 .sync_mode = WB_SYNC_ALL,
2372 .nr_to_write = 1,
2373 };
2374
2375 BUG_ON(!PageLocked(page));
2376
2377 if (wait)
2378 wait_on_page_writeback(page);
2379
2380 if (clear_page_dirty_for_io(page)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002381 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 ret = mapping->a_ops->writepage(page, &wbc);
2383 if (ret == 0 && wait) {
2384 wait_on_page_writeback(page);
2385 if (PageError(page))
2386 ret = -EIO;
2387 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002388 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 } else {
2390 unlock_page(page);
2391 }
2392 return ret;
2393}
2394EXPORT_SYMBOL(write_one_page);
2395
2396/*
Ken Chen76719322007-02-10 01:43:15 -08002397 * For address_spaces which do not use buffers nor write back.
2398 */
2399int __set_page_dirty_no_writeback(struct page *page)
2400{
2401 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08002402 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08002403 return 0;
2404}
2405
2406/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002407 * Helper function for set_page_dirty family.
Greg Thelenc4843a72015-05-22 17:13:16 -04002408 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002409 * Caller must hold lock_page_memcg().
Greg Thelenc4843a72015-05-22 17:13:16 -04002410 *
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002411 * NOTE: This relies on being atomic wrt interrupts.
2412 */
Johannes Weiner62cccb82016-03-15 14:57:22 -07002413void account_page_dirtied(struct page *page, struct address_space *mapping)
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002414{
Tejun Heo52ebea72015-05-22 17:13:37 -04002415 struct inode *inode = mapping->host;
2416
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002417 trace_writeback_dirty_page(page, mapping);
2418
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002419 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo52ebea72015-05-22 17:13:37 -04002420 struct bdi_writeback *wb;
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002421
Tejun Heo52ebea72015-05-22 17:13:37 -04002422 inode_attach_wb(inode, page);
2423 wb = inode_to_wb(inode);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002424
Johannes Weiner62cccb82016-03-15 14:57:22 -07002425 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_DIRTY);
Mel Gorman11fb9982016-07-28 15:46:20 -07002426 __inc_node_page_state(page, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002427 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Mel Gormanc4a25632016-07-28 15:46:23 -07002428 __inc_node_page_state(page, NR_DIRTIED);
Tejun Heo52ebea72015-05-22 17:13:37 -04002429 __inc_wb_stat(wb, WB_RECLAIMABLE);
2430 __inc_wb_stat(wb, WB_DIRTIED);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002431 task_io_account_write(PAGE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06002432 current->nr_dirtied++;
2433 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002434 }
2435}
Michael Rubin679ceac2010-08-20 02:31:26 -07002436EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002437
2438/*
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002439 * Helper function for deaccounting dirty page without writeback.
2440 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002441 * Caller must hold lock_page_memcg().
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002442 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002443void account_page_cleaned(struct page *page, struct address_space *mapping,
Johannes Weiner62cccb82016-03-15 14:57:22 -07002444 struct bdi_writeback *wb)
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002445{
2446 if (mapping_cap_account_dirty(mapping)) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002447 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
Mel Gorman11fb9982016-07-28 15:46:20 -07002448 dec_node_page_state(page, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002449 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002450 dec_wb_stat(wb, WB_RECLAIMABLE);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002451 task_io_account_cancelled_write(PAGE_SIZE);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002452 }
2453}
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002454
2455/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 * For address_spaces which do not use buffers. Just tag the page as dirty in
2457 * its radix tree.
2458 *
2459 * This is also used when a single buffer is being dirtied: we want to set the
2460 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2461 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2462 *
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002463 * The caller must ensure this doesn't race with truncation. Most will simply
2464 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2465 * the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466 */
2467int __set_page_dirty_nobuffers(struct page *page)
2468{
Johannes Weiner62cccb82016-03-15 14:57:22 -07002469 lock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 if (!TestSetPageDirty(page)) {
2471 struct address_space *mapping = page_mapping(page);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002472 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473
Greg Thelenc4843a72015-05-22 17:13:16 -04002474 if (!mapping) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002475 unlock_page_memcg(page);
Andrew Morton8c085402006-12-10 02:19:24 -08002476 return 1;
Greg Thelenc4843a72015-05-22 17:13:16 -04002477 }
Andrew Morton8c085402006-12-10 02:19:24 -08002478
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002479 spin_lock_irqsave(&mapping->tree_lock, flags);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002480 BUG_ON(page_mapping(page) != mapping);
2481 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Johannes Weiner62cccb82016-03-15 14:57:22 -07002482 account_page_dirtied(page, mapping);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002483 radix_tree_tag_set(&mapping->page_tree, page_index(page),
2484 PAGECACHE_TAG_DIRTY);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002485 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner62cccb82016-03-15 14:57:22 -07002486 unlock_page_memcg(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002487
Andrew Morton8c085402006-12-10 02:19:24 -08002488 if (mapping->host) {
2489 /* !PageAnon && !swapper_space */
2490 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002492 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07002494 unlock_page_memcg(page);
Andrew Morton4741c9f2006-03-24 03:18:11 -08002495 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496}
2497EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2498
2499/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002500 * Call this whenever redirtying a page, to de-account the dirty counters
2501 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2502 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2503 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2504 * control.
2505 */
2506void account_page_redirty(struct page *page)
2507{
2508 struct address_space *mapping = page->mapping;
Tejun Heo91018132015-05-22 17:13:39 -04002509
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002510 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002511 struct inode *inode = mapping->host;
2512 struct bdi_writeback *wb;
Greg Thelen18484eb2018-04-20 14:55:42 -07002513 struct wb_lock_cookie cookie = {};
Tejun Heo91018132015-05-22 17:13:39 -04002514
Greg Thelen18484eb2018-04-20 14:55:42 -07002515 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002516 current->nr_dirtied--;
Mel Gormanc4a25632016-07-28 15:46:23 -07002517 dec_node_page_state(page, NR_DIRTIED);
Tejun Heo91018132015-05-22 17:13:39 -04002518 dec_wb_stat(wb, WB_DIRTIED);
Greg Thelen18484eb2018-04-20 14:55:42 -07002519 unlocked_inode_to_wb_end(inode, &cookie);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002520 }
2521}
2522EXPORT_SYMBOL(account_page_redirty);
2523
2524/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 * When a writepage implementation decides that it doesn't want to write this
2526 * page for some reason, it should redirty the locked page via
2527 * redirty_page_for_writepage() and it should then unlock the page and return 0
2528 */
2529int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2530{
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002531 int ret;
2532
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 wbc->pages_skipped++;
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002534 ret = __set_page_dirty_nobuffers(page);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002535 account_page_redirty(page);
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002536 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537}
2538EXPORT_SYMBOL(redirty_page_for_writepage);
2539
2540/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002541 * Dirty a page.
2542 *
2543 * For pages with a mapping this should be done under the page lock
2544 * for the benefit of asynchronous memory errors who prefer a consistent
2545 * dirty state. This rule can be broken in some special cases,
2546 * but should be better not to.
2547 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 * If the mapping doesn't provide a set_page_dirty a_op, then
2549 * just fall through and assume that it wants buffer_heads.
2550 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002551int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552{
2553 struct address_space *mapping = page_mapping(page);
2554
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002555 page = compound_head(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 if (likely(mapping)) {
2557 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002558 /*
2559 * readahead/lru_deactivate_page could remain
2560 * PG_readahead/PG_reclaim due to race with end_page_writeback
2561 * About readahead, if the page is written, the flags would be
2562 * reset. So no problem.
2563 * About lru_deactivate_page, if the page is redirty, the flag
2564 * will be reset. So no problem. but if the page is used by readahead
2565 * it will confuse readahead and make it restart the size rampup
2566 * process. But it's a trivial problem.
2567 */
Naoya Horiguchia4bb3ec2015-04-15 16:13:17 -07002568 if (PageReclaim(page))
2569 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002570#ifdef CONFIG_BLOCK
2571 if (!spd)
2572 spd = __set_page_dirty_buffers;
2573#endif
2574 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002576 if (!PageDirty(page)) {
2577 if (!TestSetPageDirty(page))
2578 return 1;
2579 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 return 0;
2581}
2582EXPORT_SYMBOL(set_page_dirty);
2583
2584/*
2585 * set_page_dirty() is racy if the caller has no reference against
2586 * page->mapping->host, and if the page is unlocked. This is because another
2587 * CPU could truncate the page off the mapping and then free the mapping.
2588 *
2589 * Usually, the page _is_ locked, or the caller is a user-space process which
2590 * holds a reference on the inode by having an open file.
2591 *
2592 * In other cases, the page should be locked before running set_page_dirty().
2593 */
2594int set_page_dirty_lock(struct page *page)
2595{
2596 int ret;
2597
Jens Axboe7eaceac2011-03-10 08:52:07 +01002598 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 ret = set_page_dirty(page);
2600 unlock_page(page);
2601 return ret;
2602}
2603EXPORT_SYMBOL(set_page_dirty_lock);
2604
2605/*
Tejun Heo11f81be2015-05-22 17:13:15 -04002606 * This cancels just the dirty bit on the kernel page itself, it does NOT
2607 * actually remove dirty bits on any mmap's that may be around. It also
2608 * leaves the page tagged dirty, so any sync activity will still find it on
2609 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2610 * look at the dirty bits in the VM.
2611 *
2612 * Doing this should *normally* only ever be done when a page is truncated,
2613 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2614 * this when it notices that somebody has cleaned out all the buffers on a
2615 * page without actually doing it through the VM. Can you say "ext3 is
2616 * horribly ugly"? Thought you could.
2617 */
2618void cancel_dirty_page(struct page *page)
2619{
Greg Thelenc4843a72015-05-22 17:13:16 -04002620 struct address_space *mapping = page_mapping(page);
2621
2622 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002623 struct inode *inode = mapping->host;
2624 struct bdi_writeback *wb;
Greg Thelen18484eb2018-04-20 14:55:42 -07002625 struct wb_lock_cookie cookie = {};
Greg Thelenc4843a72015-05-22 17:13:16 -04002626
Johannes Weiner62cccb82016-03-15 14:57:22 -07002627 lock_page_memcg(page);
Greg Thelen18484eb2018-04-20 14:55:42 -07002628 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002629
2630 if (TestClearPageDirty(page))
Johannes Weiner62cccb82016-03-15 14:57:22 -07002631 account_page_cleaned(page, mapping, wb);
Greg Thelenc4843a72015-05-22 17:13:16 -04002632
Greg Thelen18484eb2018-04-20 14:55:42 -07002633 unlocked_inode_to_wb_end(inode, &cookie);
Johannes Weiner62cccb82016-03-15 14:57:22 -07002634 unlock_page_memcg(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002635 } else {
2636 ClearPageDirty(page);
2637 }
Tejun Heo11f81be2015-05-22 17:13:15 -04002638}
2639EXPORT_SYMBOL(cancel_dirty_page);
2640
2641/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 * Clear a page's dirty flag, while caring for dirty memory accounting.
2643 * Returns true if the page was previously dirty.
2644 *
2645 * This is for preparing to put the page under writeout. We leave the page
2646 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2647 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2648 * implementation will run either set_page_writeback() or set_page_dirty(),
2649 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2650 * back into sync.
2651 *
2652 * This incoherency between the page's dirty flag and radix-tree tag is
2653 * unfortunate, but it only exists while the page is locked.
2654 */
2655int clear_page_dirty_for_io(struct page *page)
2656{
2657 struct address_space *mapping = page_mapping(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002658 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659
Nick Piggin79352892007-07-19 01:47:22 -07002660 BUG_ON(!PageLocked(page));
2661
Linus Torvalds7658cc22006-12-29 10:00:58 -08002662 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002663 struct inode *inode = mapping->host;
2664 struct bdi_writeback *wb;
Greg Thelen18484eb2018-04-20 14:55:42 -07002665 struct wb_lock_cookie cookie = {};
Tejun Heo682aa8e2015-05-28 14:50:53 -04002666
Linus Torvalds7658cc22006-12-29 10:00:58 -08002667 /*
2668 * Yes, Virginia, this is indeed insane.
2669 *
2670 * We use this sequence to make sure that
2671 * (a) we account for dirty stats properly
2672 * (b) we tell the low-level filesystem to
2673 * mark the whole page dirty if it was
2674 * dirty in a pagetable. Only to then
2675 * (c) clean the page again and return 1 to
2676 * cause the writeback.
2677 *
2678 * This way we avoid all nasty races with the
2679 * dirty bit in multiple places and clearing
2680 * them concurrently from different threads.
2681 *
2682 * Note! Normally the "set_page_dirty(page)"
2683 * has no effect on the actual dirty bit - since
2684 * that will already usually be set. But we
2685 * need the side effects, and it can help us
2686 * avoid races.
2687 *
2688 * We basically use the page "master dirty bit"
2689 * as a serialization point for all the different
2690 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002691 */
2692 if (page_mkclean(page))
2693 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002694 /*
2695 * We carefully synchronise fault handlers against
2696 * installing a dirty pte and marking the page dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002697 * at this point. We do this by having them hold the
2698 * page lock while dirtying the page, and pages are
2699 * always locked coming in here, so we get the desired
2700 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002701 */
Greg Thelen18484eb2018-04-20 14:55:42 -07002702 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002703 if (TestClearPageDirty(page)) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002704 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_DIRTY);
Mel Gorman11fb9982016-07-28 15:46:20 -07002705 dec_node_page_state(page, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002706 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002707 dec_wb_stat(wb, WB_RECLAIMABLE);
Greg Thelenc4843a72015-05-22 17:13:16 -04002708 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 }
Greg Thelen18484eb2018-04-20 14:55:42 -07002710 unlocked_inode_to_wb_end(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002711 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002713 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002715EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716
2717int test_clear_page_writeback(struct page *page)
2718{
2719 struct address_space *mapping = page_mapping(page);
Johannes Weinerdee92932017-08-18 15:15:48 -07002720 struct mem_cgroup *memcg;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002721 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722
Johannes Weinerdee92932017-08-18 15:15:48 -07002723 memcg = lock_page_memcg(page);
Huang Ying371a0962016-10-07 16:59:30 -07002724 if (mapping && mapping_use_writeback_tags(mapping)) {
Tejun Heo91018132015-05-22 17:13:39 -04002725 struct inode *inode = mapping->host;
2726 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 unsigned long flags;
2728
Nick Piggin19fd6232008-07-25 19:45:32 -07002729 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002731 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 radix_tree_tag_clear(&mapping->page_tree,
2733 page_index(page),
2734 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002735 if (bdi_cap_account_writeback(bdi)) {
Tejun Heo91018132015-05-22 17:13:39 -04002736 struct bdi_writeback *wb = inode_to_wb(inode);
2737
2738 __dec_wb_stat(wb, WB_WRITEBACK);
2739 __wb_writeout_inc(wb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002740 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002741 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002742
2743 if (mapping->host && !mapping_tagged(mapping,
2744 PAGECACHE_TAG_WRITEBACK))
2745 sb_clear_inode_writeback(mapping->host);
2746
Nick Piggin19fd6232008-07-25 19:45:32 -07002747 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 } else {
2749 ret = TestClearPageWriteback(page);
2750 }
Johannes Weinerdee92932017-08-18 15:15:48 -07002751 /*
2752 * NOTE: Page might be free now! Writeback doesn't hold a page
2753 * reference on its own, it relies on truncation to wait for
2754 * the clearing of PG_writeback. The below can only access
2755 * page state that is static across allocation cycles.
2756 */
Wu Fengguang99b12e32011-07-25 17:12:37 -07002757 if (ret) {
Johannes Weinerdee92932017-08-18 15:15:48 -07002758 __mem_cgroup_update_page_stat(page, memcg,
2759 MEM_CGROUP_STAT_WRITEBACK, -1);
Mel Gorman11fb9982016-07-28 15:46:20 -07002760 dec_node_page_state(page, NR_WRITEBACK);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002761 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Mel Gormanc4a25632016-07-28 15:46:23 -07002762 inc_node_page_state(page, NR_WRITTEN);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002763 }
Johannes Weinerdee92932017-08-18 15:15:48 -07002764 __unlock_page_memcg(memcg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 return ret;
2766}
2767
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002768int __test_set_page_writeback(struct page *page, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769{
2770 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002771 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772
Johannes Weiner62cccb82016-03-15 14:57:22 -07002773 lock_page_memcg(page);
Huang Ying371a0962016-10-07 16:59:30 -07002774 if (mapping && mapping_use_writeback_tags(mapping)) {
Tejun Heo91018132015-05-22 17:13:39 -04002775 struct inode *inode = mapping->host;
2776 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 unsigned long flags;
2778
Nick Piggin19fd6232008-07-25 19:45:32 -07002779 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002781 if (!ret) {
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002782 bool on_wblist;
2783
2784 on_wblist = mapping_tagged(mapping,
2785 PAGECACHE_TAG_WRITEBACK);
2786
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 radix_tree_tag_set(&mapping->page_tree,
2788 page_index(page),
2789 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002790 if (bdi_cap_account_writeback(bdi))
Tejun Heo91018132015-05-22 17:13:39 -04002791 __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002792
2793 /*
2794 * We can come through here when swapping anonymous
2795 * pages, so we don't necessarily have an inode to track
2796 * for sync.
2797 */
2798 if (mapping->host && !on_wblist)
2799 sb_mark_inode_writeback(mapping->host);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002800 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 if (!PageDirty(page))
2802 radix_tree_tag_clear(&mapping->page_tree,
2803 page_index(page),
2804 PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002805 if (!keep_write)
2806 radix_tree_tag_clear(&mapping->page_tree,
2807 page_index(page),
2808 PAGECACHE_TAG_TOWRITE);
Nick Piggin19fd6232008-07-25 19:45:32 -07002809 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 } else {
2811 ret = TestSetPageWriteback(page);
2812 }
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002813 if (!ret) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002814 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
Mel Gorman11fb9982016-07-28 15:46:20 -07002815 inc_node_page_state(page, NR_WRITEBACK);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002816 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002817 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07002818 unlock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 return ret;
2820
2821}
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002822EXPORT_SYMBOL(__test_set_page_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
2824/*
Nick Piggin00128182007-10-16 01:24:40 -07002825 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 * passed tag.
2827 */
2828int mapping_tagged(struct address_space *mapping, int tag)
2829{
Konstantin Khlebnikov72c47832011-07-25 17:12:31 -07002830 return radix_tree_tagged(&mapping->page_tree, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831}
2832EXPORT_SYMBOL(mapping_tagged);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002833
2834/**
2835 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2836 * @page: The page to wait on.
2837 *
2838 * This function determines if the given page is related to a backing device
2839 * that requires page contents to be held stable during writeback. If so, then
2840 * it will wait for any pending writeback to complete.
2841 */
2842void wait_for_stable_page(struct page *page)
2843{
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002844 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2845 wait_on_page_writeback(page);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002846}
2847EXPORT_SYMBOL_GPL(wait_for_stable_page);