Linux Audio

Check our new training course

Loading...
v4.17
 
   1/*
   2 * mm/page-writeback.c
   3 *
   4 * Copyright (C) 2002, Linus Torvalds.
   5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
   6 *
   7 * Contains functions related to writing back dirty pages at the
   8 * address_space level.
   9 *
  10 * 10Apr2002	Andrew Morton
  11 *		Initial version
  12 */
  13
  14#include <linux/kernel.h>
 
  15#include <linux/export.h>
  16#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>
  25#include <linux/task_io_accounting_ops.h>
  26#include <linux/blkdev.h>
  27#include <linux/mpage.h>
  28#include <linux/rmap.h>
  29#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>
  35#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
  36#include <linux/pagevec.h>
  37#include <linux/timer.h>
  38#include <linux/sched/rt.h>
  39#include <linux/sched/signal.h>
  40#include <linux/mm_inline.h>
  41#include <trace/events/writeback.h>
  42
  43#include "internal.h"
  44
  45/*
  46 * Sleep at most 200ms at a time in balance_dirty_pages().
  47 */
  48#define MAX_PAUSE		max(HZ/5, 1)
  49
  50/*
  51 * Try to keep balance_dirty_pages() call intervals higher than this many pages
  52 * by raising pause time to max_pause when falls below it.
  53 */
  54#define DIRTY_POLL_THRESH	(128 >> (PAGE_SHIFT - 10))
  55
  56/*
  57 * Estimate write bandwidth at 200ms intervals.
  58 */
  59#define BANDWIDTH_INTERVAL	max(HZ/5, 1)
  60
  61#define RATELIMIT_CALC_SHIFT	10
  62
  63/*
  64 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  65 * will look to see if it needs to force writeback or throttling.
  66 */
  67static long ratelimit_pages = 32;
  68
  69/* The following parameters are exported via /proc/sys/vm */
  70
  71/*
  72 * Start background writeback (via writeback threads) at this percentage
  73 */
  74int dirty_background_ratio = 10;
  75
  76/*
  77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
  78 * dirty_background_ratio * the amount of dirtyable memory
  79 */
  80unsigned long dirty_background_bytes;
  81
  82/*
  83 * free highmem will not be subtracted from the total free memory
  84 * for calculating free ratios if vm_highmem_is_dirtyable is true
  85 */
  86int vm_highmem_is_dirtyable;
  87
  88/*
  89 * The generator of dirty data starts writeback at this percentage
  90 */
  91int vm_dirty_ratio = 20;
  92
  93/*
  94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
  95 * vm_dirty_ratio * the amount of dirtyable memory
  96 */
  97unsigned long vm_dirty_bytes;
  98
  99/*
 100 * The interval between `kupdate'-style writebacks
 101 */
 102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
 103
 104EXPORT_SYMBOL_GPL(dirty_writeback_interval);
 105
 106/*
 107 * The longest time for which data is allowed to remain dirty
 108 */
 109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
 110
 111/*
 112 * Flag that makes the machine dump writes/reads and block dirtyings.
 113 */
 114int block_dump;
 115
 116/*
 117 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
 118 * a full sync is triggered after this time elapses without any disk activity.
 119 */
 120int laptop_mode;
 121
 122EXPORT_SYMBOL(laptop_mode);
 123
 124/* End of sysctl-exported parameters */
 125
 126struct wb_domain global_wb_domain;
 127
 128/* consolidated parameters for balance_dirty_pages() and its subroutines */
 129struct dirty_throttle_control {
 130#ifdef CONFIG_CGROUP_WRITEBACK
 131	struct wb_domain	*dom;
 132	struct dirty_throttle_control *gdtc;	/* only set in memcg dtc's */
 133#endif
 134	struct bdi_writeback	*wb;
 135	struct fprop_local_percpu *wb_completions;
 136
 137	unsigned long		avail;		/* dirtyable */
 138	unsigned long		dirty;		/* file_dirty + write + nfs */
 139	unsigned long		thresh;		/* dirty threshold */
 140	unsigned long		bg_thresh;	/* dirty background threshold */
 141
 142	unsigned long		wb_dirty;	/* per-wb counterparts */
 143	unsigned long		wb_thresh;
 144	unsigned long		wb_bg_thresh;
 145
 146	unsigned long		pos_ratio;
 
 
 147};
 148
 149/*
 150 * Length of period for aging writeout fractions of bdis. This is an
 151 * arbitrarily chosen number. The longer the period, the slower fractions will
 152 * reflect changes in current writeout rate.
 153 */
 154#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
 155
 156#ifdef CONFIG_CGROUP_WRITEBACK
 157
 158#define GDTC_INIT(__wb)		.wb = (__wb),				\
 159				.dom = &global_wb_domain,		\
 160				.wb_completions = &(__wb)->completions
 161
 162#define GDTC_INIT_NO_WB		.dom = &global_wb_domain
 163
 164#define MDTC_INIT(__wb, __gdtc)	.wb = (__wb),				\
 165				.dom = mem_cgroup_wb_domain(__wb),	\
 166				.wb_completions = &(__wb)->memcg_completions, \
 167				.gdtc = __gdtc
 168
 169static bool mdtc_valid(struct dirty_throttle_control *dtc)
 170{
 171	return dtc->dom;
 172}
 173
 174static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
 175{
 176	return dtc->dom;
 177}
 178
 179static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
 180{
 181	return mdtc->gdtc;
 182}
 183
 184static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
 185{
 186	return &wb->memcg_completions;
 187}
 188
 189static void wb_min_max_ratio(struct bdi_writeback *wb,
 190			     unsigned long *minp, unsigned long *maxp)
 191{
 192	unsigned long this_bw = wb->avg_write_bandwidth;
 193	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
 194	unsigned long long min = wb->bdi->min_ratio;
 195	unsigned long long max = wb->bdi->max_ratio;
 196
 197	/*
 198	 * @wb may already be clean by the time control reaches here and
 199	 * the total may not include its bw.
 200	 */
 201	if (this_bw < tot_bw) {
 202		if (min) {
 203			min *= this_bw;
 204			do_div(min, tot_bw);
 205		}
 206		if (max < 100) {
 207			max *= this_bw;
 208			do_div(max, tot_bw);
 209		}
 210	}
 211
 212	*minp = min;
 213	*maxp = max;
 214}
 215
 216#else	/* CONFIG_CGROUP_WRITEBACK */
 217
 218#define GDTC_INIT(__wb)		.wb = (__wb),                           \
 219				.wb_completions = &(__wb)->completions
 220#define GDTC_INIT_NO_WB
 221#define MDTC_INIT(__wb, __gdtc)
 222
 223static bool mdtc_valid(struct dirty_throttle_control *dtc)
 224{
 225	return false;
 226}
 227
 228static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
 229{
 230	return &global_wb_domain;
 231}
 232
 233static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
 234{
 235	return NULL;
 236}
 237
 238static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
 239{
 240	return NULL;
 241}
 242
 243static void wb_min_max_ratio(struct bdi_writeback *wb,
 244			     unsigned long *minp, unsigned long *maxp)
 245{
 246	*minp = wb->bdi->min_ratio;
 247	*maxp = wb->bdi->max_ratio;
 248}
 249
 250#endif	/* CONFIG_CGROUP_WRITEBACK */
 251
 252/*
 253 * In a memory zone, there is a certain amount of pages we consider
 254 * available for the page cache, which is essentially the number of
 255 * free and reclaimable pages, minus some zone reserves to protect
 256 * lowmem and the ability to uphold the zone's watermarks without
 257 * requiring writeback.
 258 *
 259 * This number of dirtyable pages is the base value of which the
 260 * user-configurable dirty ratio is the effictive number of pages that
 261 * are allowed to be actually dirtied.  Per individual zone, or
 262 * globally by using the sum of dirtyable pages over all zones.
 263 *
 264 * Because the user is allowed to specify the dirty limit globally as
 265 * absolute number of bytes, calculating the per-zone dirty limit can
 266 * require translating the configured limit into a percentage of
 267 * global dirtyable memory first.
 268 */
 269
 270/**
 271 * node_dirtyable_memory - number of dirtyable pages in a node
 272 * @pgdat: the node
 273 *
 274 * Returns the node's number of pages potentially available for dirty
 275 * page cache.  This is the base value for the per-node dirty limits.
 276 */
 277static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
 278{
 279	unsigned long nr_pages = 0;
 280	int z;
 281
 282	for (z = 0; z < MAX_NR_ZONES; z++) {
 283		struct zone *zone = pgdat->node_zones + z;
 284
 285		if (!populated_zone(zone))
 286			continue;
 287
 288		nr_pages += zone_page_state(zone, NR_FREE_PAGES);
 289	}
 290
 291	/*
 292	 * Pages reserved for the kernel should not be considered
 293	 * dirtyable, to prevent a situation where reclaim has to
 294	 * clean pages in order to balance the zones.
 295	 */
 296	nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
 297
 298	nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
 299	nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
 300
 301	return nr_pages;
 302}
 303
 304static unsigned long highmem_dirtyable_memory(unsigned long total)
 305{
 306#ifdef CONFIG_HIGHMEM
 307	int node;
 308	unsigned long x = 0;
 309	int i;
 310
 311	for_each_node_state(node, N_HIGH_MEMORY) {
 312		for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
 313			struct zone *z;
 314			unsigned long nr_pages;
 315
 316			if (!is_highmem_idx(i))
 317				continue;
 318
 319			z = &NODE_DATA(node)->node_zones[i];
 320			if (!populated_zone(z))
 321				continue;
 322
 323			nr_pages = zone_page_state(z, NR_FREE_PAGES);
 324			/* watch for underflows */
 325			nr_pages -= min(nr_pages, high_wmark_pages(z));
 326			nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
 327			nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
 328			x += nr_pages;
 329		}
 330	}
 331
 332	/*
 333	 * Unreclaimable memory (kernel memory or anonymous memory
 334	 * without swap) can bring down the dirtyable pages below
 335	 * the zone's dirty balance reserve and the above calculation
 336	 * will underflow.  However we still want to add in nodes
 337	 * which are below threshold (negative values) to get a more
 338	 * accurate calculation but make sure that the total never
 339	 * underflows.
 340	 */
 341	if ((long)x < 0)
 342		x = 0;
 343
 344	/*
 345	 * Make sure that the number of highmem pages is never larger
 346	 * than the number of the total dirtyable memory. This can only
 347	 * occur in very strange VM situations but we want to make sure
 348	 * that this does not occur.
 349	 */
 350	return min(x, total);
 351#else
 352	return 0;
 353#endif
 354}
 355
 356/**
 357 * global_dirtyable_memory - number of globally dirtyable pages
 358 *
 359 * Returns the global number of pages potentially available for dirty
 360 * page cache.  This is the base value for the global dirty limits.
 361 */
 362static unsigned long global_dirtyable_memory(void)
 363{
 364	unsigned long x;
 365
 366	x = global_zone_page_state(NR_FREE_PAGES);
 367	/*
 368	 * Pages reserved for the kernel should not be considered
 369	 * dirtyable, to prevent a situation where reclaim has to
 370	 * clean pages in order to balance the zones.
 371	 */
 372	x -= min(x, totalreserve_pages);
 373
 374	x += global_node_page_state(NR_INACTIVE_FILE);
 375	x += global_node_page_state(NR_ACTIVE_FILE);
 376
 377	if (!vm_highmem_is_dirtyable)
 378		x -= highmem_dirtyable_memory(x);
 379
 380	return x + 1;	/* Ensure that we never return 0 */
 381}
 382
 383/**
 384 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
 385 * @dtc: dirty_throttle_control of interest
 386 *
 387 * Calculate @dtc->thresh and ->bg_thresh considering
 388 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
 389 * must ensure that @dtc->avail is set before calling this function.  The
 390 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
 391 * real-time tasks.
 392 */
 393static void domain_dirty_limits(struct dirty_throttle_control *dtc)
 394{
 395	const unsigned long available_memory = dtc->avail;
 396	struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
 397	unsigned long bytes = vm_dirty_bytes;
 398	unsigned long bg_bytes = dirty_background_bytes;
 399	/* convert ratios to per-PAGE_SIZE for higher precision */
 400	unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
 401	unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
 402	unsigned long thresh;
 403	unsigned long bg_thresh;
 404	struct task_struct *tsk;
 405
 406	/* gdtc is !NULL iff @dtc is for memcg domain */
 407	if (gdtc) {
 408		unsigned long global_avail = gdtc->avail;
 409
 410		/*
 411		 * The byte settings can't be applied directly to memcg
 412		 * domains.  Convert them to ratios by scaling against
 413		 * globally available memory.  As the ratios are in
 414		 * per-PAGE_SIZE, they can be obtained by dividing bytes by
 415		 * number of pages.
 416		 */
 417		if (bytes)
 418			ratio = min(DIV_ROUND_UP(bytes, global_avail),
 419				    PAGE_SIZE);
 420		if (bg_bytes)
 421			bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
 422				       PAGE_SIZE);
 423		bytes = bg_bytes = 0;
 424	}
 425
 426	if (bytes)
 427		thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
 428	else
 429		thresh = (ratio * available_memory) / PAGE_SIZE;
 430
 431	if (bg_bytes)
 432		bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
 433	else
 434		bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
 435
 436	if (bg_thresh >= thresh)
 437		bg_thresh = thresh / 2;
 438	tsk = current;
 439	if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
 440		bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
 441		thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
 442	}
 
 
 
 
 
 
 
 
 
 443	dtc->thresh = thresh;
 444	dtc->bg_thresh = bg_thresh;
 445
 446	/* we should eventually report the domain in the TP */
 447	if (!gdtc)
 448		trace_global_dirty_state(bg_thresh, thresh);
 449}
 450
 451/**
 452 * global_dirty_limits - background-writeback and dirty-throttling thresholds
 453 * @pbackground: out parameter for bg_thresh
 454 * @pdirty: out parameter for thresh
 455 *
 456 * Calculate bg_thresh and thresh for global_wb_domain.  See
 457 * domain_dirty_limits() for details.
 458 */
 459void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
 460{
 461	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
 462
 463	gdtc.avail = global_dirtyable_memory();
 464	domain_dirty_limits(&gdtc);
 465
 466	*pbackground = gdtc.bg_thresh;
 467	*pdirty = gdtc.thresh;
 468}
 469
 470/**
 471 * node_dirty_limit - maximum number of dirty pages allowed in a node
 472 * @pgdat: the node
 473 *
 474 * Returns the maximum number of dirty pages allowed in a node, based
 475 * on the node's dirtyable memory.
 476 */
 477static unsigned long node_dirty_limit(struct pglist_data *pgdat)
 478{
 479	unsigned long node_memory = node_dirtyable_memory(pgdat);
 480	struct task_struct *tsk = current;
 481	unsigned long dirty;
 482
 483	if (vm_dirty_bytes)
 484		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
 485			node_memory / global_dirtyable_memory();
 486	else
 487		dirty = vm_dirty_ratio * node_memory / 100;
 488
 489	if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
 490		dirty += dirty / 4;
 491
 492	return dirty;
 
 
 
 
 493}
 494
 495/**
 496 * node_dirty_ok - tells whether a node is within its dirty limits
 497 * @pgdat: the node to check
 498 *
 499 * Returns %true when the dirty pages in @pgdat are within the node's
 500 * dirty limit, %false if the limit is exceeded.
 501 */
 502bool node_dirty_ok(struct pglist_data *pgdat)
 503{
 504	unsigned long limit = node_dirty_limit(pgdat);
 505	unsigned long nr_pages = 0;
 506
 507	nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
 508	nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
 509	nr_pages += node_page_state(pgdat, NR_WRITEBACK);
 510
 511	return nr_pages <= limit;
 512}
 513
 514int dirty_background_ratio_handler(struct ctl_table *table, int write,
 515		void __user *buffer, size_t *lenp,
 516		loff_t *ppos)
 517{
 518	int ret;
 519
 520	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 521	if (ret == 0 && write)
 522		dirty_background_bytes = 0;
 523	return ret;
 524}
 525
 526int dirty_background_bytes_handler(struct ctl_table *table, int write,
 527		void __user *buffer, size_t *lenp,
 528		loff_t *ppos)
 529{
 530	int ret;
 
 531
 532	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 533	if (ret == 0 && write)
 
 
 
 
 
 534		dirty_background_ratio = 0;
 
 535	return ret;
 536}
 537
 538int dirty_ratio_handler(struct ctl_table *table, int write,
 539		void __user *buffer, size_t *lenp,
 540		loff_t *ppos)
 541{
 542	int old_ratio = vm_dirty_ratio;
 543	int ret;
 544
 545	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 546	if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
 547		writeback_set_ratelimit();
 548		vm_dirty_bytes = 0;
 549	}
 550	return ret;
 551}
 552
 553int dirty_bytes_handler(struct ctl_table *table, int write,
 554		void __user *buffer, size_t *lenp,
 555		loff_t *ppos)
 556{
 557	unsigned long old_bytes = vm_dirty_bytes;
 558	int ret;
 559
 560	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 561	if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
 
 
 
 
 562		writeback_set_ratelimit();
 563		vm_dirty_ratio = 0;
 564	}
 565	return ret;
 566}
 
 567
 568static unsigned long wp_next_time(unsigned long cur_time)
 569{
 570	cur_time += VM_COMPLETIONS_PERIOD_LEN;
 571	/* 0 has a special meaning... */
 572	if (!cur_time)
 573		return 1;
 574	return cur_time;
 575}
 576
 577static void wb_domain_writeout_inc(struct wb_domain *dom,
 578				   struct fprop_local_percpu *completions,
 579				   unsigned int max_prop_frac)
 580{
 581	__fprop_inc_percpu_max(&dom->completions, completions,
 582			       max_prop_frac);
 583	/* First event after period switching was turned off? */
 584	if (unlikely(!dom->period_time)) {
 585		/*
 586		 * We can race with other __bdi_writeout_inc calls here but
 587		 * it does not cause any harm since the resulting time when
 588		 * timer will fire and what is in writeout_period_time will be
 589		 * roughly the same.
 590		 */
 591		dom->period_time = wp_next_time(jiffies);
 592		mod_timer(&dom->period_timer, dom->period_time);
 593	}
 594}
 595
 596/*
 597 * Increment @wb's writeout completion count and the global writeout
 598 * completion count. Called from test_clear_page_writeback().
 599 */
 600static inline void __wb_writeout_inc(struct bdi_writeback *wb)
 601{
 602	struct wb_domain *cgdom;
 603
 604	inc_wb_stat(wb, WB_WRITTEN);
 605	wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
 606			       wb->bdi->max_prop_frac);
 607
 608	cgdom = mem_cgroup_wb_domain(wb);
 609	if (cgdom)
 610		wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
 611				       wb->bdi->max_prop_frac);
 612}
 613
 614void wb_writeout_inc(struct bdi_writeback *wb)
 615{
 616	unsigned long flags;
 617
 618	local_irq_save(flags);
 619	__wb_writeout_inc(wb);
 620	local_irq_restore(flags);
 621}
 622EXPORT_SYMBOL_GPL(wb_writeout_inc);
 623
 624/*
 625 * On idle system, we can be called long after we scheduled because we use
 626 * deferred timers so count with missed periods.
 627 */
 628static void writeout_period(struct timer_list *t)
 629{
 630	struct wb_domain *dom = from_timer(dom, t, period_timer);
 631	int miss_periods = (jiffies - dom->period_time) /
 632						 VM_COMPLETIONS_PERIOD_LEN;
 633
 634	if (fprop_new_period(&dom->completions, miss_periods + 1)) {
 635		dom->period_time = wp_next_time(dom->period_time +
 636				miss_periods * VM_COMPLETIONS_PERIOD_LEN);
 637		mod_timer(&dom->period_timer, dom->period_time);
 638	} else {
 639		/*
 640		 * Aging has zeroed all fractions. Stop wasting CPU on period
 641		 * updates.
 642		 */
 643		dom->period_time = 0;
 644	}
 645}
 646
 647int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
 648{
 649	memset(dom, 0, sizeof(*dom));
 650
 651	spin_lock_init(&dom->lock);
 652
 653	timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
 654
 655	dom->dirty_limit_tstamp = jiffies;
 656
 657	return fprop_global_init(&dom->completions, gfp);
 658}
 659
 660#ifdef CONFIG_CGROUP_WRITEBACK
 661void wb_domain_exit(struct wb_domain *dom)
 662{
 663	del_timer_sync(&dom->period_timer);
 664	fprop_global_destroy(&dom->completions);
 665}
 666#endif
 667
 668/*
 669 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
 670 * registered backing devices, which, for obvious reasons, can not
 671 * exceed 100%.
 672 */
 673static unsigned int bdi_min_ratio;
 674
 675int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 676{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 677	int ret = 0;
 678
 
 
 
 679	spin_lock_bh(&bdi_lock);
 680	if (min_ratio > bdi->max_ratio) {
 681		ret = -EINVAL;
 682	} else {
 683		min_ratio -= bdi->min_ratio;
 684		if (bdi_min_ratio + min_ratio < 100) {
 685			bdi_min_ratio += min_ratio;
 686			bdi->min_ratio += min_ratio;
 687		} else {
 688			ret = -EINVAL;
 
 
 
 
 
 
 689		}
 690	}
 691	spin_unlock_bh(&bdi_lock);
 692
 693	return ret;
 694}
 695
 696int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
 697{
 698	int ret = 0;
 699
 700	if (max_ratio > 100)
 701		return -EINVAL;
 702
 703	spin_lock_bh(&bdi_lock);
 704	if (bdi->min_ratio > max_ratio) {
 705		ret = -EINVAL;
 706	} else {
 707		bdi->max_ratio = max_ratio;
 708		bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
 
 709	}
 710	spin_unlock_bh(&bdi_lock);
 711
 712	return ret;
 713}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 714EXPORT_SYMBOL(bdi_set_max_ratio);
 715
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 716static unsigned long dirty_freerun_ceiling(unsigned long thresh,
 717					   unsigned long bg_thresh)
 718{
 719	return (thresh + bg_thresh) / 2;
 720}
 721
 722static unsigned long hard_dirty_limit(struct wb_domain *dom,
 723				      unsigned long thresh)
 724{
 725	return max(thresh, dom->dirty_limit);
 726}
 727
 728/*
 729 * Memory which can be further allocated to a memcg domain is capped by
 730 * system-wide clean memory excluding the amount being used in the domain.
 731 */
 732static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
 733			    unsigned long filepages, unsigned long headroom)
 734{
 735	struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
 736	unsigned long clean = filepages - min(filepages, mdtc->dirty);
 737	unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
 738	unsigned long other_clean = global_clean - min(global_clean, clean);
 739
 740	mdtc->avail = filepages + min(headroom, other_clean);
 741}
 742
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 743/**
 744 * __wb_calc_thresh - @wb's share of dirty throttling threshold
 745 * @dtc: dirty_throttle_context of interest
 
 746 *
 747 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
 748 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
 749 *
 750 * Note that balance_dirty_pages() will only seriously take it as a hard limit
 751 * when sleeping max_pause per page is not enough to keep the dirty pages under
 752 * control. For example, when the device is completely stalled due to some error
 753 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
 754 * In the other normal situations, it acts more gently by throttling the tasks
 755 * more (rather than completely block them) when the wb dirty pages go high.
 756 *
 757 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
 758 * - starving fast devices
 759 * - piling up dirty pages (that will take long time to sync) on slow devices
 760 *
 761 * The wb's share of dirty limit will be adapting to its throughput and
 762 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
 
 
 
 
 763 */
 764static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
 
 765{
 766	struct wb_domain *dom = dtc_dom(dtc);
 767	unsigned long thresh = dtc->thresh;
 768	u64 wb_thresh;
 769	long numerator, denominator;
 
 770	unsigned long wb_min_ratio, wb_max_ratio;
 771
 772	/*
 773	 * Calculate this BDI's share of the thresh ratio.
 774	 */
 775	fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
 776			      &numerator, &denominator);
 777
 778	wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
 779	wb_thresh *= numerator;
 780	do_div(wb_thresh, denominator);
 
 
 781
 782	wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 783
 784	wb_thresh += (thresh * wb_min_ratio) / 100;
 785	if (wb_thresh > (thresh * wb_max_ratio) / 100)
 786		wb_thresh = thresh * wb_max_ratio / 100;
 
 787
 788	return wb_thresh;
 789}
 790
 791unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
 792{
 793	struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
 794					       .thresh = thresh };
 795	return __wb_calc_thresh(&gdtc);
 
 
 
 
 
 
 
 
 
 
 
 
 796}
 797
 798/*
 799 *                           setpoint - dirty 3
 800 *        f(dirty) := 1.0 + (----------------)
 801 *                           limit - setpoint
 802 *
 803 * it's a 3rd order polynomial that subjects to
 804 *
 805 * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
 806 * (2) f(setpoint) = 1.0 => the balance point
 807 * (3) f(limit)    = 0   => the hard limit
 808 * (4) df/dx      <= 0	 => negative feedback control
 809 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
 810 *     => fast response on large errors; small oscillation near setpoint
 811 */
 812static long long pos_ratio_polynom(unsigned long setpoint,
 813					  unsigned long dirty,
 814					  unsigned long limit)
 815{
 816	long long pos_ratio;
 817	long x;
 818
 819	x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
 820		      (limit - setpoint) | 1);
 821	pos_ratio = x;
 822	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 823	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 824	pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
 825
 826	return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
 827}
 828
 829/*
 830 * Dirty position control.
 831 *
 832 * (o) global/bdi setpoints
 833 *
 834 * We want the dirty pages be balanced around the global/wb setpoints.
 835 * When the number of dirty pages is higher/lower than the setpoint, the
 836 * dirty position control ratio (and hence task dirty ratelimit) will be
 837 * decreased/increased to bring the dirty pages back to the setpoint.
 838 *
 839 *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
 840 *
 841 *     if (dirty < setpoint) scale up   pos_ratio
 842 *     if (dirty > setpoint) scale down pos_ratio
 843 *
 844 *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
 845 *     if (wb_dirty > wb_setpoint) scale down pos_ratio
 846 *
 847 *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
 848 *
 849 * (o) global control line
 850 *
 851 *     ^ pos_ratio
 852 *     |
 853 *     |            |<===== global dirty control scope ======>|
 854 * 2.0 .............*
 855 *     |            .*
 856 *     |            . *
 857 *     |            .   *
 858 *     |            .     *
 859 *     |            .        *
 860 *     |            .            *
 861 * 1.0 ................................*
 862 *     |            .                  .     *
 863 *     |            .                  .          *
 864 *     |            .                  .              *
 865 *     |            .                  .                 *
 866 *     |            .                  .                    *
 867 *   0 +------------.------------------.----------------------*------------->
 868 *           freerun^          setpoint^                 limit^   dirty pages
 869 *
 870 * (o) wb control line
 871 *
 872 *     ^ pos_ratio
 873 *     |
 874 *     |            *
 875 *     |              *
 876 *     |                *
 877 *     |                  *
 878 *     |                    * |<=========== span ============>|
 879 * 1.0 .......................*
 880 *     |                      . *
 881 *     |                      .   *
 882 *     |                      .     *
 883 *     |                      .       *
 884 *     |                      .         *
 885 *     |                      .           *
 886 *     |                      .             *
 887 *     |                      .               *
 888 *     |                      .                 *
 889 *     |                      .                   *
 890 *     |                      .                     *
 891 * 1/4 ...............................................* * * * * * * * * * * *
 892 *     |                      .                         .
 893 *     |                      .                           .
 894 *     |                      .                             .
 895 *   0 +----------------------.-------------------------------.------------->
 896 *                wb_setpoint^                    x_intercept^
 897 *
 898 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
 899 * be smoothly throttled down to normal if it starts high in situations like
 900 * - start writing to a slow SD card and a fast disk at the same time. The SD
 901 *   card's wb_dirty may rush to many times higher than wb_setpoint.
 902 * - the wb dirty thresh drops quickly due to change of JBOD workload
 903 */
 904static void wb_position_ratio(struct dirty_throttle_control *dtc)
 905{
 906	struct bdi_writeback *wb = dtc->wb;
 907	unsigned long write_bw = wb->avg_write_bandwidth;
 908	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
 909	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
 910	unsigned long wb_thresh = dtc->wb_thresh;
 911	unsigned long x_intercept;
 912	unsigned long setpoint;		/* dirty pages' target balance point */
 913	unsigned long wb_setpoint;
 914	unsigned long span;
 915	long long pos_ratio;		/* for scaling up/down the rate limit */
 916	long x;
 917
 918	dtc->pos_ratio = 0;
 919
 920	if (unlikely(dtc->dirty >= limit))
 921		return;
 922
 923	/*
 924	 * global setpoint
 925	 *
 926	 * See comment for pos_ratio_polynom().
 927	 */
 928	setpoint = (freerun + limit) / 2;
 929	pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
 930
 931	/*
 932	 * The strictlimit feature is a tool preventing mistrusted filesystems
 933	 * from growing a large number of dirty pages before throttling. For
 934	 * such filesystems balance_dirty_pages always checks wb counters
 935	 * against wb limits. Even if global "nr_dirty" is under "freerun".
 936	 * This is especially important for fuse which sets bdi->max_ratio to
 937	 * 1% by default. Without strictlimit feature, fuse writeback may
 938	 * consume arbitrary amount of RAM because it is accounted in
 939	 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
 940	 *
 941	 * Here, in wb_position_ratio(), we calculate pos_ratio based on
 942	 * two values: wb_dirty and wb_thresh. Let's consider an example:
 943	 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
 944	 * limits are set by default to 10% and 20% (background and throttle).
 945	 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
 946	 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
 947	 * about ~6K pages (as the average of background and throttle wb
 948	 * limits). The 3rd order polynomial will provide positive feedback if
 949	 * wb_dirty is under wb_setpoint and vice versa.
 950	 *
 951	 * Note, that we cannot use global counters in these calculations
 952	 * because we want to throttle process writing to a strictlimit wb
 953	 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
 954	 * in the example above).
 955	 */
 956	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
 957		long long wb_pos_ratio;
 958
 959		if (dtc->wb_dirty < 8) {
 960			dtc->pos_ratio = min_t(long long, pos_ratio * 2,
 961					   2 << RATELIMIT_CALC_SHIFT);
 962			return;
 963		}
 964
 965		if (dtc->wb_dirty >= wb_thresh)
 966			return;
 967
 968		wb_setpoint = dirty_freerun_ceiling(wb_thresh,
 969						    dtc->wb_bg_thresh);
 970
 971		if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
 972			return;
 973
 974		wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
 975						 wb_thresh);
 976
 977		/*
 978		 * Typically, for strictlimit case, wb_setpoint << setpoint
 979		 * and pos_ratio >> wb_pos_ratio. In the other words global
 980		 * state ("dirty") is not limiting factor and we have to
 981		 * make decision based on wb counters. But there is an
 982		 * important case when global pos_ratio should get precedence:
 983		 * global limits are exceeded (e.g. due to activities on other
 984		 * wb's) while given strictlimit wb is below limit.
 985		 *
 986		 * "pos_ratio * wb_pos_ratio" would work for the case above,
 987		 * but it would look too non-natural for the case of all
 988		 * activity in the system coming from a single strictlimit wb
 989		 * with bdi->max_ratio == 100%.
 990		 *
 991		 * Note that min() below somewhat changes the dynamics of the
 992		 * control system. Normally, pos_ratio value can be well over 3
 993		 * (when globally we are at freerun and wb is well below wb
 994		 * setpoint). Now the maximum pos_ratio in the same situation
 995		 * is 2. We might want to tweak this if we observe the control
 996		 * system is too slow to adapt.
 997		 */
 998		dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
 999		return;
1000	}
1001
1002	/*
1003	 * We have computed basic pos_ratio above based on global situation. If
1004	 * the wb is over/under its share of dirty pages, we want to scale
1005	 * pos_ratio further down/up. That is done by the following mechanism.
1006	 */
1007
1008	/*
1009	 * wb setpoint
1010	 *
1011	 *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1012	 *
1013	 *                        x_intercept - wb_dirty
1014	 *                     := --------------------------
1015	 *                        x_intercept - wb_setpoint
1016	 *
1017	 * The main wb control line is a linear function that subjects to
1018	 *
1019	 * (1) f(wb_setpoint) = 1.0
1020	 * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1021	 *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1022	 *
1023	 * For single wb case, the dirty pages are observed to fluctuate
1024	 * regularly within range
1025	 *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1026	 * for various filesystems, where (2) can yield in a reasonable 12.5%
1027	 * fluctuation range for pos_ratio.
1028	 *
1029	 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1030	 * own size, so move the slope over accordingly and choose a slope that
1031	 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1032	 */
1033	if (unlikely(wb_thresh > dtc->thresh))
1034		wb_thresh = dtc->thresh;
1035	/*
1036	 * It's very possible that wb_thresh is close to 0 not because the
1037	 * device is slow, but that it has remained inactive for long time.
1038	 * Honour such devices a reasonable good (hopefully IO efficient)
1039	 * threshold, so that the occasional writes won't be blocked and active
1040	 * writes can rampup the threshold quickly.
1041	 */
1042	wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1043	/*
1044	 * scale global setpoint to wb's:
1045	 *	wb_setpoint = setpoint * wb_thresh / thresh
1046	 */
1047	x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1048	wb_setpoint = setpoint * (u64)x >> 16;
1049	/*
1050	 * Use span=(8*write_bw) in single wb case as indicated by
1051	 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1052	 *
1053	 *        wb_thresh                    thresh - wb_thresh
1054	 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1055	 *         thresh                           thresh
1056	 */
1057	span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1058	x_intercept = wb_setpoint + span;
1059
1060	if (dtc->wb_dirty < x_intercept - span / 4) {
1061		pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1062				      (x_intercept - wb_setpoint) | 1);
1063	} else
1064		pos_ratio /= 4;
1065
1066	/*
1067	 * wb reserve area, safeguard against dirty pool underrun and disk idle
1068	 * It may push the desired control point of global dirty pages higher
1069	 * than setpoint.
1070	 */
1071	x_intercept = wb_thresh / 2;
1072	if (dtc->wb_dirty < x_intercept) {
1073		if (dtc->wb_dirty > x_intercept / 8)
1074			pos_ratio = div_u64(pos_ratio * x_intercept,
1075					    dtc->wb_dirty);
1076		else
1077			pos_ratio *= 8;
1078	}
1079
1080	dtc->pos_ratio = pos_ratio;
1081}
1082
1083static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1084				      unsigned long elapsed,
1085				      unsigned long written)
1086{
1087	const unsigned long period = roundup_pow_of_two(3 * HZ);
1088	unsigned long avg = wb->avg_write_bandwidth;
1089	unsigned long old = wb->write_bandwidth;
1090	u64 bw;
1091
1092	/*
1093	 * bw = written * HZ / elapsed
1094	 *
1095	 *                   bw * elapsed + write_bandwidth * (period - elapsed)
1096	 * write_bandwidth = ---------------------------------------------------
1097	 *                                          period
1098	 *
1099	 * @written may have decreased due to account_page_redirty().
1100	 * Avoid underflowing @bw calculation.
1101	 */
1102	bw = written - min(written, wb->written_stamp);
1103	bw *= HZ;
1104	if (unlikely(elapsed > period)) {
1105		do_div(bw, elapsed);
1106		avg = bw;
1107		goto out;
1108	}
1109	bw += (u64)wb->write_bandwidth * (period - elapsed);
1110	bw >>= ilog2(period);
1111
1112	/*
1113	 * one more level of smoothing, for filtering out sudden spikes
1114	 */
1115	if (avg > old && old >= (unsigned long)bw)
1116		avg -= (avg - old) >> 3;
1117
1118	if (avg < old && old <= (unsigned long)bw)
1119		avg += (old - avg) >> 3;
1120
1121out:
1122	/* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1123	avg = max(avg, 1LU);
1124	if (wb_has_dirty_io(wb)) {
1125		long delta = avg - wb->avg_write_bandwidth;
1126		WARN_ON_ONCE(atomic_long_add_return(delta,
1127					&wb->bdi->tot_write_bandwidth) <= 0);
1128	}
1129	wb->write_bandwidth = bw;
1130	wb->avg_write_bandwidth = avg;
1131}
1132
1133static void update_dirty_limit(struct dirty_throttle_control *dtc)
1134{
1135	struct wb_domain *dom = dtc_dom(dtc);
1136	unsigned long thresh = dtc->thresh;
1137	unsigned long limit = dom->dirty_limit;
1138
1139	/*
1140	 * Follow up in one step.
1141	 */
1142	if (limit < thresh) {
1143		limit = thresh;
1144		goto update;
1145	}
1146
1147	/*
1148	 * Follow down slowly. Use the higher one as the target, because thresh
1149	 * may drop below dirty. This is exactly the reason to introduce
1150	 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1151	 */
1152	thresh = max(thresh, dtc->dirty);
1153	if (limit > thresh) {
1154		limit -= (limit - thresh) >> 5;
1155		goto update;
1156	}
1157	return;
1158update:
1159	dom->dirty_limit = limit;
1160}
1161
1162static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1163				    unsigned long now)
1164{
1165	struct wb_domain *dom = dtc_dom(dtc);
1166
1167	/*
1168	 * check locklessly first to optimize away locking for the most time
1169	 */
1170	if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1171		return;
1172
1173	spin_lock(&dom->lock);
1174	if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1175		update_dirty_limit(dtc);
1176		dom->dirty_limit_tstamp = now;
1177	}
1178	spin_unlock(&dom->lock);
1179}
1180
1181/*
1182 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1183 *
1184 * Normal wb tasks will be curbed at or below it in long term.
1185 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1186 */
1187static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1188				      unsigned long dirtied,
1189				      unsigned long elapsed)
1190{
1191	struct bdi_writeback *wb = dtc->wb;
1192	unsigned long dirty = dtc->dirty;
1193	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1194	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1195	unsigned long setpoint = (freerun + limit) / 2;
1196	unsigned long write_bw = wb->avg_write_bandwidth;
1197	unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1198	unsigned long dirty_rate;
1199	unsigned long task_ratelimit;
1200	unsigned long balanced_dirty_ratelimit;
1201	unsigned long step;
1202	unsigned long x;
1203	unsigned long shift;
1204
1205	/*
1206	 * The dirty rate will match the writeout rate in long term, except
1207	 * when dirty pages are truncated by userspace or re-dirtied by FS.
1208	 */
1209	dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1210
1211	/*
1212	 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1213	 */
1214	task_ratelimit = (u64)dirty_ratelimit *
1215					dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1216	task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1217
1218	/*
1219	 * A linear estimation of the "balanced" throttle rate. The theory is,
1220	 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1221	 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1222	 * formula will yield the balanced rate limit (write_bw / N).
1223	 *
1224	 * Note that the expanded form is not a pure rate feedback:
1225	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate)		     (1)
1226	 * but also takes pos_ratio into account:
1227	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1228	 *
1229	 * (1) is not realistic because pos_ratio also takes part in balancing
1230	 * the dirty rate.  Consider the state
1231	 *	pos_ratio = 0.5						     (3)
1232	 *	rate = 2 * (write_bw / N)				     (4)
1233	 * If (1) is used, it will stuck in that state! Because each dd will
1234	 * be throttled at
1235	 *	task_ratelimit = pos_ratio * rate = (write_bw / N)	     (5)
1236	 * yielding
1237	 *	dirty_rate = N * task_ratelimit = write_bw		     (6)
1238	 * put (6) into (1) we get
1239	 *	rate_(i+1) = rate_(i)					     (7)
1240	 *
1241	 * So we end up using (2) to always keep
1242	 *	rate_(i+1) ~= (write_bw / N)				     (8)
1243	 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1244	 * pos_ratio is able to drive itself to 1.0, which is not only where
1245	 * the dirty count meet the setpoint, but also where the slope of
1246	 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1247	 */
1248	balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1249					   dirty_rate | 1);
1250	/*
1251	 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1252	 */
1253	if (unlikely(balanced_dirty_ratelimit > write_bw))
1254		balanced_dirty_ratelimit = write_bw;
1255
1256	/*
1257	 * We could safely do this and return immediately:
1258	 *
1259	 *	wb->dirty_ratelimit = balanced_dirty_ratelimit;
1260	 *
1261	 * However to get a more stable dirty_ratelimit, the below elaborated
1262	 * code makes use of task_ratelimit to filter out singular points and
1263	 * limit the step size.
1264	 *
1265	 * The below code essentially only uses the relative value of
1266	 *
1267	 *	task_ratelimit - dirty_ratelimit
1268	 *	= (pos_ratio - 1) * dirty_ratelimit
1269	 *
1270	 * which reflects the direction and size of dirty position error.
1271	 */
1272
1273	/*
1274	 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1275	 * task_ratelimit is on the same side of dirty_ratelimit, too.
1276	 * For example, when
1277	 * - dirty_ratelimit > balanced_dirty_ratelimit
1278	 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1279	 * lowering dirty_ratelimit will help meet both the position and rate
1280	 * control targets. Otherwise, don't update dirty_ratelimit if it will
1281	 * only help meet the rate target. After all, what the users ultimately
1282	 * feel and care are stable dirty rate and small position error.
1283	 *
1284	 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1285	 * and filter out the singular points of balanced_dirty_ratelimit. Which
1286	 * keeps jumping around randomly and can even leap far away at times
1287	 * due to the small 200ms estimation period of dirty_rate (we want to
1288	 * keep that period small to reduce time lags).
1289	 */
1290	step = 0;
1291
1292	/*
1293	 * For strictlimit case, calculations above were based on wb counters
1294	 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1295	 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1296	 * Hence, to calculate "step" properly, we have to use wb_dirty as
1297	 * "dirty" and wb_setpoint as "setpoint".
1298	 *
1299	 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1300	 * it's possible that wb_thresh is close to zero due to inactivity
1301	 * of backing device.
1302	 */
1303	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1304		dirty = dtc->wb_dirty;
1305		if (dtc->wb_dirty < 8)
1306			setpoint = dtc->wb_dirty + 1;
1307		else
1308			setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1309	}
1310
1311	if (dirty < setpoint) {
1312		x = min3(wb->balanced_dirty_ratelimit,
1313			 balanced_dirty_ratelimit, task_ratelimit);
1314		if (dirty_ratelimit < x)
1315			step = x - dirty_ratelimit;
1316	} else {
1317		x = max3(wb->balanced_dirty_ratelimit,
1318			 balanced_dirty_ratelimit, task_ratelimit);
1319		if (dirty_ratelimit > x)
1320			step = dirty_ratelimit - x;
1321	}
1322
1323	/*
1324	 * Don't pursue 100% rate matching. It's impossible since the balanced
1325	 * rate itself is constantly fluctuating. So decrease the track speed
1326	 * when it gets close to the target. Helps eliminate pointless tremors.
1327	 */
1328	shift = dirty_ratelimit / (2 * step + 1);
1329	if (shift < BITS_PER_LONG)
1330		step = DIV_ROUND_UP(step >> shift, 8);
1331	else
1332		step = 0;
1333
1334	if (dirty_ratelimit < balanced_dirty_ratelimit)
1335		dirty_ratelimit += step;
1336	else
1337		dirty_ratelimit -= step;
1338
1339	wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1340	wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1341
1342	trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1343}
1344
1345static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1346				  struct dirty_throttle_control *mdtc,
1347				  unsigned long start_time,
1348				  bool update_ratelimit)
1349{
1350	struct bdi_writeback *wb = gdtc->wb;
1351	unsigned long now = jiffies;
1352	unsigned long elapsed = now - wb->bw_time_stamp;
1353	unsigned long dirtied;
1354	unsigned long written;
1355
1356	lockdep_assert_held(&wb->list_lock);
1357
1358	/*
1359	 * rate-limit, only update once every 200ms.
 
 
 
1360	 */
1361	if (elapsed < BANDWIDTH_INTERVAL)
1362		return;
1363
1364	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1365	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1366
1367	/*
1368	 * Skip quiet periods when disk bandwidth is under-utilized.
1369	 * (at least 1s idle time between two flusher runs)
1370	 */
1371	if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1372		goto snapshot;
1373
1374	if (update_ratelimit) {
1375		domain_update_bandwidth(gdtc, now);
1376		wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1377
1378		/*
1379		 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1380		 * compiler has no way to figure that out.  Help it.
1381		 */
1382		if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1383			domain_update_bandwidth(mdtc, now);
1384			wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1385		}
1386	}
1387	wb_update_write_bandwidth(wb, elapsed, written);
1388
1389snapshot:
1390	wb->dirtied_stamp = dirtied;
1391	wb->written_stamp = written;
1392	wb->bw_time_stamp = now;
 
1393}
1394
1395void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1396{
1397	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1398
1399	__wb_update_bandwidth(&gdtc, NULL, start_time, false);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1400}
1401
1402/*
1403 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1404 * will look to see if it needs to start dirty throttling.
1405 *
1406 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1407 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1408 * (the number of pages we may dirty without exceeding the dirty limits).
1409 */
1410static unsigned long dirty_poll_interval(unsigned long dirty,
1411					 unsigned long thresh)
1412{
1413	if (thresh > dirty)
1414		return 1UL << (ilog2(thresh - dirty) >> 1);
1415
1416	return 1;
1417}
1418
1419static unsigned long wb_max_pause(struct bdi_writeback *wb,
1420				  unsigned long wb_dirty)
1421{
1422	unsigned long bw = wb->avg_write_bandwidth;
1423	unsigned long t;
1424
1425	/*
1426	 * Limit pause time for small memory systems. If sleeping for too long
1427	 * time, a small pool of dirty/writeback pages may go empty and disk go
1428	 * idle.
1429	 *
1430	 * 8 serves as the safety ratio.
1431	 */
1432	t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1433	t++;
1434
1435	return min_t(unsigned long, t, MAX_PAUSE);
1436}
1437
1438static long wb_min_pause(struct bdi_writeback *wb,
1439			 long max_pause,
1440			 unsigned long task_ratelimit,
1441			 unsigned long dirty_ratelimit,
1442			 int *nr_dirtied_pause)
1443{
1444	long hi = ilog2(wb->avg_write_bandwidth);
1445	long lo = ilog2(wb->dirty_ratelimit);
1446	long t;		/* target pause */
1447	long pause;	/* estimated next pause */
1448	int pages;	/* target nr_dirtied_pause */
1449
1450	/* target for 10ms pause on 1-dd case */
1451	t = max(1, HZ / 100);
1452
1453	/*
1454	 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1455	 * overheads.
1456	 *
1457	 * (N * 10ms) on 2^N concurrent tasks.
1458	 */
1459	if (hi > lo)
1460		t += (hi - lo) * (10 * HZ) / 1024;
1461
1462	/*
1463	 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1464	 * on the much more stable dirty_ratelimit. However the next pause time
1465	 * will be computed based on task_ratelimit and the two rate limits may
1466	 * depart considerably at some time. Especially if task_ratelimit goes
1467	 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1468	 * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1469	 * result task_ratelimit won't be executed faithfully, which could
1470	 * eventually bring down dirty_ratelimit.
1471	 *
1472	 * We apply two rules to fix it up:
1473	 * 1) try to estimate the next pause time and if necessary, use a lower
1474	 *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1475	 *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1476	 * 2) limit the target pause time to max_pause/2, so that the normal
1477	 *    small fluctuations of task_ratelimit won't trigger rule (1) and
1478	 *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1479	 */
1480	t = min(t, 1 + max_pause / 2);
1481	pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1482
1483	/*
1484	 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1485	 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1486	 * When the 16 consecutive reads are often interrupted by some dirty
1487	 * throttling pause during the async writes, cfq will go into idles
1488	 * (deadline is fine). So push nr_dirtied_pause as high as possible
1489	 * until reaches DIRTY_POLL_THRESH=32 pages.
1490	 */
1491	if (pages < DIRTY_POLL_THRESH) {
1492		t = max_pause;
1493		pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1494		if (pages > DIRTY_POLL_THRESH) {
1495			pages = DIRTY_POLL_THRESH;
1496			t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1497		}
1498	}
1499
1500	pause = HZ * pages / (task_ratelimit + 1);
1501	if (pause > max_pause) {
1502		t = max_pause;
1503		pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1504	}
1505
1506	*nr_dirtied_pause = pages;
1507	/*
1508	 * The minimal pause time will normally be half the target pause time.
1509	 */
1510	return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1511}
1512
1513static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1514{
1515	struct bdi_writeback *wb = dtc->wb;
1516	unsigned long wb_reclaimable;
1517
1518	/*
1519	 * wb_thresh is not treated as some limiting factor as
1520	 * dirty_thresh, due to reasons
1521	 * - in JBOD setup, wb_thresh can fluctuate a lot
1522	 * - in a system with HDD and USB key, the USB key may somehow
1523	 *   go into state (wb_dirty >> wb_thresh) either because
1524	 *   wb_dirty starts high, or because wb_thresh drops low.
1525	 *   In this case we don't want to hard throttle the USB key
1526	 *   dirtiers for 100 seconds until wb_dirty drops under
1527	 *   wb_thresh. Instead the auxiliary wb control line in
1528	 *   wb_position_ratio() will let the dirtier task progress
1529	 *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1530	 */
1531	dtc->wb_thresh = __wb_calc_thresh(dtc);
1532	dtc->wb_bg_thresh = dtc->thresh ?
1533		div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1534
1535	/*
1536	 * In order to avoid the stacked BDI deadlock we need
1537	 * to ensure we accurately count the 'dirty' pages when
1538	 * the threshold is low.
1539	 *
1540	 * Otherwise it would be possible to get thresh+n pages
1541	 * reported dirty, even though there are thresh-m pages
1542	 * actually dirty; with m+n sitting in the percpu
1543	 * deltas.
1544	 */
1545	if (dtc->wb_thresh < 2 * wb_stat_error()) {
1546		wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1547		dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1548	} else {
1549		wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1550		dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1551	}
1552}
1553
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1554/*
1555 * balance_dirty_pages() must be called by processes which are generating dirty
1556 * data.  It looks at the number of dirty pages in the machine and will force
1557 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1558 * If we're over `background_thresh' then the writeback threads are woken to
1559 * perform some writeout.
1560 */
1561static void balance_dirty_pages(struct bdi_writeback *wb,
1562				unsigned long pages_dirtied)
1563{
1564	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1565	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1566	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1567	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1568						     &mdtc_stor : NULL;
1569	struct dirty_throttle_control *sdtc;
1570	unsigned long nr_reclaimable;	/* = file_dirty + unstable_nfs */
1571	long period;
1572	long pause;
1573	long max_pause;
1574	long min_pause;
1575	int nr_dirtied_pause;
1576	bool dirty_exceeded = false;
1577	unsigned long task_ratelimit;
1578	unsigned long dirty_ratelimit;
1579	struct backing_dev_info *bdi = wb->bdi;
1580	bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1581	unsigned long start_time = jiffies;
 
1582
1583	for (;;) {
1584		unsigned long now = jiffies;
1585		unsigned long dirty, thresh, bg_thresh;
1586		unsigned long m_dirty = 0;	/* stop bogus uninit warnings */
1587		unsigned long m_thresh = 0;
1588		unsigned long m_bg_thresh = 0;
1589
1590		/*
1591		 * Unstable writes are a feature of certain networked
1592		 * filesystems (i.e. NFS) in which data may have been
1593		 * written to the server's write cache, but has not yet
1594		 * been flushed to permanent storage.
1595		 */
1596		nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1597					global_node_page_state(NR_UNSTABLE_NFS);
1598		gdtc->avail = global_dirtyable_memory();
1599		gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1600
1601		domain_dirty_limits(gdtc);
1602
1603		if (unlikely(strictlimit)) {
1604			wb_dirty_limits(gdtc);
1605
1606			dirty = gdtc->wb_dirty;
1607			thresh = gdtc->wb_thresh;
1608			bg_thresh = gdtc->wb_bg_thresh;
1609		} else {
1610			dirty = gdtc->dirty;
1611			thresh = gdtc->thresh;
1612			bg_thresh = gdtc->bg_thresh;
1613		}
1614
1615		if (mdtc) {
1616			unsigned long filepages, headroom, writeback;
1617
 
 
1618			/*
1619			 * If @wb belongs to !root memcg, repeat the same
1620			 * basic calculations for the memcg domain.
1621			 */
1622			mem_cgroup_wb_stats(wb, &filepages, &headroom,
1623					    &mdtc->dirty, &writeback);
1624			mdtc->dirty += writeback;
1625			mdtc_calc_avail(mdtc, filepages, headroom);
1626
1627			domain_dirty_limits(mdtc);
1628
1629			if (unlikely(strictlimit)) {
1630				wb_dirty_limits(mdtc);
1631				m_dirty = mdtc->wb_dirty;
1632				m_thresh = mdtc->wb_thresh;
1633				m_bg_thresh = mdtc->wb_bg_thresh;
1634			} else {
1635				m_dirty = mdtc->dirty;
1636				m_thresh = mdtc->thresh;
1637				m_bg_thresh = mdtc->bg_thresh;
1638			}
1639		}
1640
1641		/*
1642		 * Throttle it only when the background writeback cannot
1643		 * catch-up. This avoids (excessively) small writeouts
1644		 * when the wb limits are ramping up in case of !strictlimit.
1645		 *
1646		 * In strictlimit case make decision based on the wb counters
1647		 * and limits. Small writeouts when the wb limits are ramping
1648		 * up are the price we consciously pay for strictlimit-ing.
1649		 *
 
 
 
 
 
 
 
 
1650		 * If memcg domain is in effect, @dirty should be under
1651		 * both global and memcg freerun ceilings.
1652		 */
1653		if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1654		    (!mdtc ||
1655		     m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1656			unsigned long intv = dirty_poll_interval(dirty, thresh);
1657			unsigned long m_intv = ULONG_MAX;
 
 
1658
1659			current->dirty_paused_when = now;
1660			current->nr_dirtied = 0;
1661			if (mdtc)
1662				m_intv = dirty_poll_interval(m_dirty, m_thresh);
1663			current->nr_dirtied_pause = min(intv, m_intv);
1664			break;
1665		}
1666
 
1667		if (unlikely(!writeback_in_progress(wb)))
1668			wb_start_background_writeback(wb);
1669
 
 
1670		/*
1671		 * Calculate global domain's pos_ratio and select the
1672		 * global dtc by default.
1673		 */
1674		if (!strictlimit)
1675			wb_dirty_limits(gdtc);
1676
1677		dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1678			((gdtc->dirty > gdtc->thresh) || strictlimit);
1679
1680		wb_position_ratio(gdtc);
1681		sdtc = gdtc;
1682
1683		if (mdtc) {
1684			/*
1685			 * If memcg domain is in effect, calculate its
1686			 * pos_ratio.  @wb should satisfy constraints from
1687			 * both global and memcg domains.  Choose the one
1688			 * w/ lower pos_ratio.
1689			 */
1690			if (!strictlimit)
1691				wb_dirty_limits(mdtc);
1692
1693			dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1694				((mdtc->dirty > mdtc->thresh) || strictlimit);
1695
1696			wb_position_ratio(mdtc);
1697			if (mdtc->pos_ratio < gdtc->pos_ratio)
1698				sdtc = mdtc;
1699		}
1700
1701		if (dirty_exceeded && !wb->dirty_exceeded)
1702			wb->dirty_exceeded = 1;
1703
1704		if (time_is_before_jiffies(wb->bw_time_stamp +
1705					   BANDWIDTH_INTERVAL)) {
1706			spin_lock(&wb->list_lock);
1707			__wb_update_bandwidth(gdtc, mdtc, start_time, true);
1708			spin_unlock(&wb->list_lock);
1709		}
1710
1711		/* throttle according to the chosen dtc */
1712		dirty_ratelimit = wb->dirty_ratelimit;
1713		task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1714							RATELIMIT_CALC_SHIFT;
1715		max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1716		min_pause = wb_min_pause(wb, max_pause,
1717					 task_ratelimit, dirty_ratelimit,
1718					 &nr_dirtied_pause);
1719
1720		if (unlikely(task_ratelimit == 0)) {
1721			period = max_pause;
1722			pause = max_pause;
1723			goto pause;
1724		}
1725		period = HZ * pages_dirtied / task_ratelimit;
1726		pause = period;
1727		if (current->dirty_paused_when)
1728			pause -= now - current->dirty_paused_when;
1729		/*
1730		 * For less than 1s think time (ext3/4 may block the dirtier
1731		 * for up to 800ms from time to time on 1-HDD; so does xfs,
1732		 * however at much less frequency), try to compensate it in
1733		 * future periods by updating the virtual time; otherwise just
1734		 * do a reset, as it may be a light dirtier.
1735		 */
1736		if (pause < min_pause) {
1737			trace_balance_dirty_pages(wb,
1738						  sdtc->thresh,
1739						  sdtc->bg_thresh,
1740						  sdtc->dirty,
1741						  sdtc->wb_thresh,
1742						  sdtc->wb_dirty,
1743						  dirty_ratelimit,
1744						  task_ratelimit,
1745						  pages_dirtied,
1746						  period,
1747						  min(pause, 0L),
1748						  start_time);
1749			if (pause < -HZ) {
1750				current->dirty_paused_when = now;
1751				current->nr_dirtied = 0;
1752			} else if (period) {
1753				current->dirty_paused_when += period;
1754				current->nr_dirtied = 0;
1755			} else if (current->nr_dirtied_pause <= pages_dirtied)
1756				current->nr_dirtied_pause += pages_dirtied;
1757			break;
1758		}
1759		if (unlikely(pause > max_pause)) {
1760			/* for occasional dropped task_ratelimit */
1761			now += min(pause - max_pause, max_pause);
1762			pause = max_pause;
1763		}
1764
1765pause:
1766		trace_balance_dirty_pages(wb,
1767					  sdtc->thresh,
1768					  sdtc->bg_thresh,
1769					  sdtc->dirty,
1770					  sdtc->wb_thresh,
1771					  sdtc->wb_dirty,
1772					  dirty_ratelimit,
1773					  task_ratelimit,
1774					  pages_dirtied,
1775					  period,
1776					  pause,
1777					  start_time);
 
 
 
 
1778		__set_current_state(TASK_KILLABLE);
1779		wb->dirty_sleep = now;
1780		io_schedule_timeout(pause);
1781
1782		current->dirty_paused_when = now + pause;
1783		current->nr_dirtied = 0;
1784		current->nr_dirtied_pause = nr_dirtied_pause;
1785
1786		/*
1787		 * This is typically equal to (dirty < thresh) and can also
1788		 * keep "1000+ dd on a slow USB stick" under control.
1789		 */
1790		if (task_ratelimit)
1791			break;
1792
1793		/*
1794		 * In the case of an unresponding NFS server and the NFS dirty
1795		 * pages exceeds dirty_thresh, give the other good wb's a pipe
1796		 * to go through, so that tasks on them still remain responsive.
1797		 *
1798		 * In theory 1 page is enough to keep the consumer-producer
1799		 * pipe going: the flusher cleans 1 page => the task dirties 1
1800		 * more page. However wb_dirty has accounting errors.  So use
1801		 * the larger and more IO friendly wb_stat_error.
1802		 */
1803		if (sdtc->wb_dirty <= wb_stat_error())
1804			break;
1805
1806		if (fatal_signal_pending(current))
1807			break;
1808	}
1809
1810	if (!dirty_exceeded && wb->dirty_exceeded)
1811		wb->dirty_exceeded = 0;
1812
1813	if (writeback_in_progress(wb))
1814		return;
1815
1816	/*
1817	 * In laptop mode, we wait until hitting the higher threshold before
1818	 * starting background writeout, and then write out all the way down
1819	 * to the lower threshold.  So slow writers cause minimal disk activity.
1820	 *
1821	 * In normal mode, we start background writeout at the lower
1822	 * background_thresh, to keep the amount of dirty memory low.
1823	 */
1824	if (laptop_mode)
1825		return;
1826
1827	if (nr_reclaimable > gdtc->bg_thresh)
1828		wb_start_background_writeback(wb);
1829}
1830
1831static DEFINE_PER_CPU(int, bdp_ratelimits);
1832
1833/*
1834 * Normal tasks are throttled by
1835 *	loop {
1836 *		dirty tsk->nr_dirtied_pause pages;
1837 *		take a snap in balance_dirty_pages();
1838 *	}
1839 * However there is a worst case. If every task exit immediately when dirtied
1840 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1841 * called to throttle the page dirties. The solution is to save the not yet
1842 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1843 * randomly into the running tasks. This works well for the above worst case,
1844 * as the new task will pick up and accumulate the old task's leaked dirty
1845 * count and eventually get throttled.
1846 */
1847DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1848
1849/**
1850 * balance_dirty_pages_ratelimited - balance dirty memory state
1851 * @mapping: address_space which was dirtied
 
1852 *
1853 * Processes which are dirtying memory should call in here once for each page
1854 * which was newly dirtied.  The function will periodically check the system's
1855 * dirty state and will initiate writeback if needed.
1856 *
1857 * On really big machines, get_writeback_state is expensive, so try to avoid
1858 * calling it too often (ratelimiting).  But once we're over the dirty memory
1859 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1860 * from overshooting the limit by (ratelimit_pages) each.
 
 
 
1861 */
1862void balance_dirty_pages_ratelimited(struct address_space *mapping)
 
1863{
1864	struct inode *inode = mapping->host;
1865	struct backing_dev_info *bdi = inode_to_bdi(inode);
1866	struct bdi_writeback *wb = NULL;
1867	int ratelimit;
 
1868	int *p;
1869
1870	if (!bdi_cap_account_dirty(bdi))
1871		return;
1872
1873	if (inode_cgwb_enabled(inode))
1874		wb = wb_get_create_current(bdi, GFP_KERNEL);
1875	if (!wb)
1876		wb = &bdi->wb;
1877
1878	ratelimit = current->nr_dirtied_pause;
1879	if (wb->dirty_exceeded)
1880		ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1881
1882	preempt_disable();
1883	/*
1884	 * This prevents one CPU to accumulate too many dirtied pages without
1885	 * calling into balance_dirty_pages(), which can happen when there are
1886	 * 1000+ tasks, all of them start dirtying pages at exactly the same
1887	 * time, hence all honoured too large initial task->nr_dirtied_pause.
1888	 */
1889	p =  this_cpu_ptr(&bdp_ratelimits);
1890	if (unlikely(current->nr_dirtied >= ratelimit))
1891		*p = 0;
1892	else if (unlikely(*p >= ratelimit_pages)) {
1893		*p = 0;
1894		ratelimit = 0;
1895	}
1896	/*
1897	 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1898	 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1899	 * the dirty throttling and livelock other long-run dirtiers.
1900	 */
1901	p = this_cpu_ptr(&dirty_throttle_leaks);
1902	if (*p > 0 && current->nr_dirtied < ratelimit) {
1903		unsigned long nr_pages_dirtied;
1904		nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1905		*p -= nr_pages_dirtied;
1906		current->nr_dirtied += nr_pages_dirtied;
1907	}
1908	preempt_enable();
1909
1910	if (unlikely(current->nr_dirtied >= ratelimit))
1911		balance_dirty_pages(wb, current->nr_dirtied);
1912
1913	wb_put(wb);
 
1914}
1915EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1916
1917/**
1918 * wb_over_bg_thresh - does @wb need to be written back?
1919 * @wb: bdi_writeback of interest
1920 *
1921 * Determines whether background writeback should keep writing @wb or it's
1922 * clean enough.  Returns %true if writeback should continue.
 
 
 
 
 
1923 */
1924bool wb_over_bg_thresh(struct bdi_writeback *wb)
1925{
1926	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1927	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1928	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1929	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1930						     &mdtc_stor : NULL;
1931
1932	/*
1933	 * Similar to balance_dirty_pages() but ignores pages being written
1934	 * as we're trying to decide whether to put more under writeback.
1935	 */
1936	gdtc->avail = global_dirtyable_memory();
1937	gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1938		      global_node_page_state(NR_UNSTABLE_NFS);
1939	domain_dirty_limits(gdtc);
1940
1941	if (gdtc->dirty > gdtc->bg_thresh)
 
 
 
 
 
 
 
 
 
 
 
1942		return true;
1943
1944	if (wb_stat(wb, WB_RECLAIMABLE) >
1945	    wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
1946		return true;
1947
1948	if (mdtc) {
1949		unsigned long filepages, headroom, writeback;
1950
1951		mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1952				    &writeback);
1953		mdtc_calc_avail(mdtc, filepages, headroom);
1954		domain_dirty_limits(mdtc);	/* ditto, ignore writeback */
 
 
 
 
 
 
 
 
 
1955
1956		if (mdtc->dirty > mdtc->bg_thresh)
1957			return true;
1958
1959		if (wb_stat(wb, WB_RECLAIMABLE) >
1960		    wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
1961			return true;
1962	}
1963
1964	return false;
1965}
1966
 
1967/*
1968 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1969 */
1970int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1971	void __user *buffer, size_t *length, loff_t *ppos)
1972{
1973	unsigned int old_interval = dirty_writeback_interval;
1974	int ret;
1975
1976	ret = proc_dointvec(table, write, buffer, length, ppos);
1977
1978	/*
1979	 * Writing 0 to dirty_writeback_interval will disable periodic writeback
1980	 * and a different non-zero value will wakeup the writeback threads.
1981	 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1982	 * iterate over all bdis and wbs.
1983	 * The reason we do this is to make the change take effect immediately.
1984	 */
1985	if (!ret && write && dirty_writeback_interval &&
1986		dirty_writeback_interval != old_interval)
1987		wakeup_flusher_threads(WB_REASON_PERIODIC);
1988
1989	return ret;
1990}
 
1991
1992#ifdef CONFIG_BLOCK
1993void laptop_mode_timer_fn(struct timer_list *t)
1994{
1995	struct backing_dev_info *backing_dev_info =
1996		from_timer(backing_dev_info, t, laptop_mode_wb_timer);
1997
1998	wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
1999}
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 */
2006void laptop_io_completion(struct backing_dev_info *info)
2007{
2008	mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2009}
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{
2018	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();
2026}
2027#endif
2028
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
2037 * thresholds.
2038 */
2039
2040void writeback_set_ratelimit(void)
2041{
2042	struct wb_domain *dom = &global_wb_domain;
2043	unsigned long background_thresh;
2044	unsigned long dirty_thresh;
2045
2046	global_dirty_limits(&background_thresh, &dirty_thresh);
2047	dom->dirty_limit = dirty_thresh;
2048	ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2049	if (ratelimit_pages < 16)
2050		ratelimit_pages = 16;
2051}
2052
2053static int page_writeback_cpu_online(unsigned int cpu)
2054{
2055	writeback_set_ratelimit();
2056	return 0;
2057}
2058
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2059/*
2060 * 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..
2076 */
2077void __init page_writeback_init(void)
2078{
2079	BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2080
2081	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);
 
 
 
2085}
2086
2087/**
2088 * 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 the i_pages lock
2103 * latency.
2104 */
2105void tag_pages_for_writeback(struct address_space *mapping,
2106			     pgoff_t start, pgoff_t end)
2107{
2108#define WRITEBACK_TAG_BATCH 4096
2109	unsigned long tagged = 0;
2110	struct radix_tree_iter iter;
2111	void **slot;
2112
2113	xa_lock_irq(&mapping->i_pages);
2114	radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start,
2115							PAGECACHE_TAG_DIRTY) {
2116		if (iter.index > end)
2117			break;
2118		radix_tree_iter_tag_set(&mapping->i_pages, &iter,
2119							PAGECACHE_TAG_TOWRITE);
2120		tagged++;
2121		if ((tagged % WRITEBACK_TAG_BATCH) != 0)
2122			continue;
2123		slot = radix_tree_iter_resume(slot, &iter);
2124		xa_unlock_irq(&mapping->i_pages);
 
2125		cond_resched();
2126		xa_lock_irq(&mapping->i_pages);
2127	}
2128	xa_unlock_irq(&mapping->i_pages);
2129}
2130EXPORT_SYMBOL(tag_pages_for_writeback);
2131
2132/**
2133 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2134 * @mapping: address space structure to write
2135 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2136 * @writepage: function called for each page
2137 * @data: data passed to writepage function
2138 *
2139 * If a page is already under I/O, write_cache_pages() skips it, even
2140 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2141 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2142 * and msync() need to guarantee that all the data which was dirty at the time
2143 * the call was made get new I/O started against them.  If wbc->sync_mode is
2144 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2145 * existing IO to complete.
2146 *
2147 * To avoid livelocks (when other process dirties new pages), we first tag
2148 * pages which should be written back with TOWRITE tag and only then start
2149 * writing them. For data-integrity sync we have to be careful so that we do
2150 * not miss some pages (e.g., because some other process has cleared TOWRITE
2151 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2152 * by the process clearing the DIRTY tag (and submitting the page for IO).
2153 */
2154int write_cache_pages(struct address_space *mapping,
2155		      struct writeback_control *wbc, writepage_t writepage,
2156		      void *data)
2157{
2158	int ret = 0;
2159	int done = 0;
2160	struct pagevec pvec;
2161	int nr_pages;
2162	pgoff_t uninitialized_var(writeback_index);
2163	pgoff_t index;
2164	pgoff_t end;		/* Inclusive */
2165	pgoff_t done_index;
2166	int cycled;
2167	int range_whole = 0;
2168	int tag;
2169
2170	pagevec_init(&pvec);
2171	if (wbc->range_cyclic) {
2172		writeback_index = mapping->writeback_index; /* prev offset */
2173		index = writeback_index;
2174		if (index == 0)
2175			cycled = 1;
2176		else
2177			cycled = 0;
2178		end = -1;
2179	} else {
2180		index = wbc->range_start >> PAGE_SHIFT;
2181		end = wbc->range_end >> PAGE_SHIFT;
2182		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2183			range_whole = 1;
2184		cycled = 1; /* ignore range_cyclic tests */
2185	}
2186	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2187		tag = PAGECACHE_TAG_TOWRITE;
2188	else
2189		tag = PAGECACHE_TAG_DIRTY;
2190retry:
2191	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2192		tag_pages_for_writeback(mapping, index, end);
2193	done_index = index;
2194	while (!done && (index <= end)) {
2195		int i;
2196
2197		nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2198				tag);
2199		if (nr_pages == 0)
2200			break;
2201
2202		for (i = 0; i < nr_pages; i++) {
2203			struct page *page = pvec.pages[i];
 
 
 
2204
2205			done_index = page->index;
 
 
 
 
 
2206
2207			lock_page(page);
 
2208
2209			/*
2210			 * Page truncated or invalidated. We can freely skip it
2211			 * then, even for data integrity operations: the page
2212			 * has disappeared concurrently, so there could be no
2213			 * real expectation of this data interity operation
2214			 * even if there is now a new, dirty page at the same
2215			 * pagecache address.
2216			 */
2217			if (unlikely(page->mapping != mapping)) {
2218continue_unlock:
2219				unlock_page(page);
2220				continue;
2221			}
2222
2223			if (!PageDirty(page)) {
2224				/* someone wrote it for us */
2225				goto continue_unlock;
2226			}
 
 
2227
2228			if (PageWriteback(page)) {
2229				if (wbc->sync_mode != WB_SYNC_NONE)
2230					wait_on_page_writeback(page);
2231				else
2232					goto continue_unlock;
2233			}
2234
2235			BUG_ON(PageWriteback(page));
2236			if (!clear_page_dirty_for_io(page))
2237				goto continue_unlock;
2238
2239			trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2240			ret = (*writepage)(page, wbc, data);
2241			if (unlikely(ret)) {
2242				if (ret == AOP_WRITEPAGE_ACTIVATE) {
2243					unlock_page(page);
2244					ret = 0;
2245				} else {
2246					/*
2247					 * done_index is set past this page,
2248					 * so media errors will not choke
2249					 * background writeout for the entire
2250					 * file. This has consequences for
2251					 * range_cyclic semantics (ie. it may
2252					 * not be suitable for data integrity
2253					 * writeout).
2254					 */
2255					done_index = page->index + 1;
2256					done = 1;
2257					break;
2258				}
2259			}
2260
2261			/*
2262			 * We stop writing back only if we are not doing
2263			 * integrity sync. In case of integrity sync we have to
2264			 * keep going until we have written all the pages
2265			 * we tagged for writeback prior to entering this loop.
2266			 */
2267			if (--wbc->nr_to_write <= 0 &&
2268			    wbc->sync_mode == WB_SYNC_NONE) {
2269				done = 1;
2270				break;
2271			}
2272		}
2273		pagevec_release(&pvec);
2274		cond_resched();
 
 
 
 
 
2275	}
2276	if (!cycled && !done) {
2277		/*
2278		 * range_cyclic:
2279		 * We hit the last page and there is more work to be done: wrap
2280		 * back to the start of the file
2281		 */
2282		cycled = 1;
2283		index = 0;
2284		end = writeback_index - 1;
2285		goto retry;
2286	}
2287	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2288		mapping->writeback_index = done_index;
2289
2290	return ret;
 
2291}
2292EXPORT_SYMBOL(write_cache_pages);
2293
2294/*
2295 * Function used by generic_writepages to call the real writepage
2296 * function and set the mapping flags on error
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2297 */
2298static int __writepage(struct page *page, struct writeback_control *wbc,
2299		       void *data)
2300{
2301	struct address_space *mapping = data;
2302	int ret = mapping->a_ops->writepage(page, wbc);
2303	mapping_set_error(mapping, ret);
2304	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2305}
 
2306
2307/**
2308 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2309 * @mapping: address space structure to write
2310 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 
 
2311 *
2312 * This is a library function, which implements the writepages()
2313 * address_space_operation.
 
2314 */
2315int generic_writepages(struct address_space *mapping,
2316		       struct writeback_control *wbc)
 
2317{
2318	struct blk_plug plug;
2319	int ret;
2320
2321	/* deal with chardevs and other special file */
2322	if (!mapping->a_ops->writepage)
2323		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2324
2325	blk_start_plug(&plug);
2326	ret = write_cache_pages(mapping, wbc, __writepage, mapping);
 
 
 
 
 
 
 
2327	blk_finish_plug(&plug);
2328	return ret;
2329}
2330
2331EXPORT_SYMBOL(generic_writepages);
 
2332
2333int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2334{
2335	int ret;
 
2336
2337	if (wbc->nr_to_write <= 0)
2338		return 0;
 
 
2339	while (1) {
2340		if (mapping->a_ops->writepages)
2341			ret = mapping->a_ops->writepages(mapping, wbc);
2342		else
2343			ret = generic_writepages(mapping, wbc);
2344		if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
 
 
 
 
2345			break;
2346		cond_resched();
2347		congestion_wait(BLK_RW_ASYNC, HZ/50);
2348	}
2349	return ret;
2350}
2351
2352/**
2353 * write_one_page - write out a single page and wait on I/O
2354 * @page: the page to write
2355 *
2356 * The page must be locked by the caller and will be unlocked upon return.
2357 *
2358 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2359 * function returns.
2360 */
2361int write_one_page(struct page *page)
2362{
2363	struct address_space *mapping = page->mapping;
2364	int ret = 0;
2365	struct writeback_control wbc = {
2366		.sync_mode = WB_SYNC_ALL,
2367		.nr_to_write = 1,
2368	};
2369
2370	BUG_ON(!PageLocked(page));
2371
2372	wait_on_page_writeback(page);
2373
2374	if (clear_page_dirty_for_io(page)) {
2375		get_page(page);
2376		ret = mapping->a_ops->writepage(page, &wbc);
2377		if (ret == 0)
2378			wait_on_page_writeback(page);
2379		put_page(page);
2380	} else {
2381		unlock_page(page);
2382	}
2383
2384	if (!ret)
2385		ret = filemap_check_errors(mapping);
 
 
 
 
 
2386	return ret;
2387}
2388EXPORT_SYMBOL(write_one_page);
2389
2390/*
2391 * For address_spaces which do not use buffers nor write back.
2392 */
2393int __set_page_dirty_no_writeback(struct page *page)
2394{
2395	if (!PageDirty(page))
2396		return !TestSetPageDirty(page);
2397	return 0;
2398}
 
2399
2400/*
2401 * Helper function for set_page_dirty family.
2402 *
2403 * Caller must hold lock_page_memcg().
2404 *
2405 * NOTE: This relies on being atomic wrt interrupts.
2406 */
2407void account_page_dirtied(struct page *page, struct address_space *mapping)
 
2408{
2409	struct inode *inode = mapping->host;
2410
2411	trace_writeback_dirty_page(page, mapping);
2412
2413	if (mapping_cap_account_dirty(mapping)) {
2414		struct bdi_writeback *wb;
 
2415
2416		inode_attach_wb(inode, page);
2417		wb = inode_to_wb(inode);
2418
2419		__inc_lruvec_page_state(page, NR_FILE_DIRTY);
2420		__inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2421		__inc_node_page_state(page, NR_DIRTIED);
2422		inc_wb_stat(wb, WB_RECLAIMABLE);
2423		inc_wb_stat(wb, WB_DIRTIED);
2424		task_io_account_write(PAGE_SIZE);
2425		current->nr_dirtied++;
2426		this_cpu_inc(bdp_ratelimits);
 
 
2427	}
2428}
2429EXPORT_SYMBOL(account_page_dirtied);
2430
2431/*
2432 * Helper function for deaccounting dirty page without writeback.
2433 *
2434 * Caller must hold lock_page_memcg().
2435 */
2436void account_page_cleaned(struct page *page, struct address_space *mapping,
2437			  struct bdi_writeback *wb)
2438{
2439	if (mapping_cap_account_dirty(mapping)) {
2440		dec_lruvec_page_state(page, NR_FILE_DIRTY);
2441		dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2442		dec_wb_stat(wb, WB_RECLAIMABLE);
2443		task_io_account_cancelled_write(PAGE_SIZE);
2444	}
2445}
2446
2447/*
2448 * For address_spaces which do not use buffers.  Just tag the page as dirty in
2449 * its radix tree.
2450 *
2451 * This is also used when a single buffer is being dirtied: we want to set the
2452 * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2453 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2454 *
2455 * The caller must ensure this doesn't race with truncation.  Most will simply
2456 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2457 * the pte lock held, which also locks out truncation.
 
 
 
 
 
2458 */
2459int __set_page_dirty_nobuffers(struct page *page)
 
2460{
2461	lock_page_memcg(page);
2462	if (!TestSetPageDirty(page)) {
2463		struct address_space *mapping = page_mapping(page);
2464		unsigned long flags;
2465
2466		if (!mapping) {
2467			unlock_page_memcg(page);
2468			return 1;
2469		}
2470
2471		xa_lock_irqsave(&mapping->i_pages, flags);
2472		BUG_ON(page_mapping(page) != mapping);
2473		WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2474		account_page_dirtied(page, mapping);
2475		radix_tree_tag_set(&mapping->i_pages, page_index(page),
2476				   PAGECACHE_TAG_DIRTY);
2477		xa_unlock_irqrestore(&mapping->i_pages, flags);
2478		unlock_page_memcg(page);
2479
2480		if (mapping->host) {
2481			/* !PageAnon && !swapper_space */
2482			__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2483		}
2484		return 1;
 
2485	}
2486	unlock_page_memcg(page);
2487	return 0;
2488}
2489EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2490
2491/*
2492 * Call this whenever redirtying a page, to de-account the dirty counters
2493 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2494 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2495 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2496 * control.
 
 
 
 
 
 
 
 
 
 
 
 
2497 */
2498void account_page_redirty(struct page *page)
2499{
2500	struct address_space *mapping = page->mapping;
 
 
 
2501
2502	if (mapping && mapping_cap_account_dirty(mapping)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2503		struct inode *inode = mapping->host;
2504		struct bdi_writeback *wb;
2505		struct wb_lock_cookie cookie = {};
2506
2507		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2508		current->nr_dirtied--;
2509		dec_node_page_state(page, NR_DIRTIED);
2510		dec_wb_stat(wb, WB_DIRTIED);
2511		unlocked_inode_to_wb_end(inode, &cookie);
2512	}
2513}
2514EXPORT_SYMBOL(account_page_redirty);
2515
2516/*
2517 * When a writepage implementation decides that it doesn't want to write this
2518 * page for some reason, it should redirty the locked page via
2519 * redirty_page_for_writepage() and it should then unlock the page and return 0
2520 */
2521int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2522{
2523	int ret;
2524
2525	wbc->pages_skipped++;
2526	ret = __set_page_dirty_nobuffers(page);
2527	account_page_redirty(page);
2528	return ret;
2529}
2530EXPORT_SYMBOL(redirty_page_for_writepage);
2531
2532/*
2533 * Dirty a page.
 
2534 *
2535 * For pages with a mapping this should be done under the page lock
2536 * for the benefit of asynchronous memory errors who prefer a consistent
2537 * dirty state. This rule can be broken in some special cases,
2538 * but should be better not to.
 
 
2539 *
2540 * If the mapping doesn't provide a set_page_dirty a_op, then
2541 * just fall through and assume that it wants buffer_heads.
2542 */
2543int set_page_dirty(struct page *page)
2544{
2545	struct address_space *mapping = page_mapping(page);
2546
2547	page = compound_head(page);
2548	if (likely(mapping)) {
2549		int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2550		/*
2551		 * readahead/lru_deactivate_page could remain
2552		 * PG_readahead/PG_reclaim due to race with end_page_writeback
2553		 * About readahead, if the page is written, the flags would be
2554		 * reset. So no problem.
2555		 * About lru_deactivate_page, if the page is redirty, the flag
2556		 * will be reset. So no problem. but if the page is used by readahead
2557		 * it will confuse readahead and make it restart the size rampup
2558		 * process. But it's a trivial problem.
2559		 */
2560		if (PageReclaim(page))
2561			ClearPageReclaim(page);
2562#ifdef CONFIG_BLOCK
2563		if (!spd)
2564			spd = __set_page_dirty_buffers;
2565#endif
2566		return (*spd)(page);
2567	}
2568	if (!PageDirty(page)) {
2569		if (!TestSetPageDirty(page))
2570			return 1;
2571	}
2572	return 0;
 
2573}
2574EXPORT_SYMBOL(set_page_dirty);
2575
2576/*
2577 * set_page_dirty() is racy if the caller has no reference against
2578 * page->mapping->host, and if the page is unlocked.  This is because another
2579 * CPU could truncate the page off the mapping and then free the mapping.
2580 *
2581 * Usually, the page _is_ locked, or the caller is a user-space process which
2582 * holds a reference on the inode by having an open file.
2583 *
2584 * In other cases, the page should be locked before running set_page_dirty().
2585 */
2586int set_page_dirty_lock(struct page *page)
2587{
2588	int ret;
2589
2590	lock_page(page);
2591	ret = set_page_dirty(page);
2592	unlock_page(page);
2593	return ret;
2594}
2595EXPORT_SYMBOL(set_page_dirty_lock);
2596
2597/*
2598 * This cancels just the dirty bit on the kernel page itself, it does NOT
2599 * actually remove dirty bits on any mmap's that may be around. It also
2600 * leaves the page tagged dirty, so any sync activity will still find it on
2601 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2602 * look at the dirty bits in the VM.
2603 *
2604 * Doing this should *normally* only ever be done when a page is truncated,
2605 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2606 * this when it notices that somebody has cleaned out all the buffers on a
2607 * page without actually doing it through the VM. Can you say "ext3 is
2608 * horribly ugly"? Thought you could.
2609 */
2610void __cancel_dirty_page(struct page *page)
2611{
2612	struct address_space *mapping = page_mapping(page);
2613
2614	if (mapping_cap_account_dirty(mapping)) {
2615		struct inode *inode = mapping->host;
2616		struct bdi_writeback *wb;
2617		struct wb_lock_cookie cookie = {};
2618
2619		lock_page_memcg(page);
2620		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2621
2622		if (TestClearPageDirty(page))
2623			account_page_cleaned(page, mapping, wb);
2624
2625		unlocked_inode_to_wb_end(inode, &cookie);
2626		unlock_page_memcg(page);
2627	} else {
2628		ClearPageDirty(page);
2629	}
2630}
2631EXPORT_SYMBOL(__cancel_dirty_page);
2632
2633/*
2634 * Clear a page's dirty flag, while caring for dirty memory accounting.
2635 * Returns true if the page was previously dirty.
2636 *
2637 * This is for preparing to put the page under writeout.  We leave the page
2638 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2639 * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2640 * implementation will run either set_page_writeback() or set_page_dirty(),
2641 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2642 * back into sync.
2643 *
2644 * This incoherency between the page's dirty flag and radix-tree tag is
2645 * unfortunate, but it only exists while the page is locked.
2646 */
2647int clear_page_dirty_for_io(struct page *page)
2648{
2649	struct address_space *mapping = page_mapping(page);
2650	int ret = 0;
2651
2652	BUG_ON(!PageLocked(page));
2653
2654	if (mapping && mapping_cap_account_dirty(mapping)) {
2655		struct inode *inode = mapping->host;
2656		struct bdi_writeback *wb;
2657		struct wb_lock_cookie cookie = {};
2658
2659		/*
2660		 * Yes, Virginia, this is indeed insane.
2661		 *
2662		 * We use this sequence to make sure that
2663		 *  (a) we account for dirty stats properly
2664		 *  (b) we tell the low-level filesystem to
2665		 *      mark the whole page dirty if it was
2666		 *      dirty in a pagetable. Only to then
2667		 *  (c) clean the page again and return 1 to
2668		 *      cause the writeback.
2669		 *
2670		 * This way we avoid all nasty races with the
2671		 * dirty bit in multiple places and clearing
2672		 * them concurrently from different threads.
2673		 *
2674		 * Note! Normally the "set_page_dirty(page)"
2675		 * has no effect on the actual dirty bit - since
2676		 * that will already usually be set. But we
2677		 * need the side effects, and it can help us
2678		 * avoid races.
2679		 *
2680		 * We basically use the page "master dirty bit"
2681		 * as a serialization point for all the different
2682		 * threads doing their things.
2683		 */
2684		if (page_mkclean(page))
2685			set_page_dirty(page);
2686		/*
2687		 * We carefully synchronise fault handlers against
2688		 * installing a dirty pte and marking the page dirty
2689		 * at this point.  We do this by having them hold the
2690		 * page lock while dirtying the page, and pages are
2691		 * always locked coming in here, so we get the desired
2692		 * exclusion.
2693		 */
2694		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2695		if (TestClearPageDirty(page)) {
2696			dec_lruvec_page_state(page, NR_FILE_DIRTY);
2697			dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2698			dec_wb_stat(wb, WB_RECLAIMABLE);
2699			ret = 1;
 
2700		}
2701		unlocked_inode_to_wb_end(inode, &cookie);
2702		return ret;
2703	}
2704	return TestClearPageDirty(page);
2705}
2706EXPORT_SYMBOL(clear_page_dirty_for_io);
2707
2708int test_clear_page_writeback(struct page *page)
2709{
2710	struct address_space *mapping = page_mapping(page);
2711	struct mem_cgroup *memcg;
2712	struct lruvec *lruvec;
2713	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2714
2715	memcg = lock_page_memcg(page);
2716	lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
2717	if (mapping && mapping_use_writeback_tags(mapping)) {
2718		struct inode *inode = mapping->host;
2719		struct backing_dev_info *bdi = inode_to_bdi(inode);
2720		unsigned long flags;
2721
2722		xa_lock_irqsave(&mapping->i_pages, flags);
2723		ret = TestClearPageWriteback(page);
2724		if (ret) {
2725			radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2726						PAGECACHE_TAG_WRITEBACK);
2727			if (bdi_cap_account_writeback(bdi)) {
2728				struct bdi_writeback *wb = inode_to_wb(inode);
2729
2730				dec_wb_stat(wb, WB_WRITEBACK);
2731				__wb_writeout_inc(wb);
2732			}
2733		}
2734
2735		if (mapping->host && !mapping_tagged(mapping,
2736						     PAGECACHE_TAG_WRITEBACK))
2737			sb_clear_inode_writeback(mapping->host);
2738
2739		xa_unlock_irqrestore(&mapping->i_pages, flags);
2740	} else {
2741		ret = TestClearPageWriteback(page);
2742	}
2743	/*
2744	 * NOTE: Page might be free now! Writeback doesn't hold a page
2745	 * reference on its own, it relies on truncation to wait for
2746	 * the clearing of PG_writeback. The below can only access
2747	 * page state that is static across allocation cycles.
2748	 */
2749	if (ret) {
2750		dec_lruvec_state(lruvec, NR_WRITEBACK);
2751		dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2752		inc_node_page_state(page, NR_WRITTEN);
2753	}
2754	__unlock_page_memcg(memcg);
2755	return ret;
2756}
2757
2758int __test_set_page_writeback(struct page *page, bool keep_write)
2759{
2760	struct address_space *mapping = page_mapping(page);
2761	int ret;
 
 
 
2762
2763	lock_page_memcg(page);
2764	if (mapping && mapping_use_writeback_tags(mapping)) {
 
2765		struct inode *inode = mapping->host;
2766		struct backing_dev_info *bdi = inode_to_bdi(inode);
2767		unsigned long flags;
 
2768
2769		xa_lock_irqsave(&mapping->i_pages, flags);
2770		ret = TestSetPageWriteback(page);
2771		if (!ret) {
2772			bool on_wblist;
2773
2774			on_wblist = mapping_tagged(mapping,
2775						   PAGECACHE_TAG_WRITEBACK);
2776
2777			radix_tree_tag_set(&mapping->i_pages, page_index(page),
2778						PAGECACHE_TAG_WRITEBACK);
2779			if (bdi_cap_account_writeback(bdi))
2780				inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2781
2782			/*
2783			 * We can come through here when swapping anonymous
2784			 * pages, so we don't necessarily have an inode to track
2785			 * for sync.
2786			 */
2787			if (mapping->host && !on_wblist)
2788				sb_mark_inode_writeback(mapping->host);
 
 
2789		}
2790		if (!PageDirty(page))
2791			radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2792						PAGECACHE_TAG_DIRTY);
 
 
 
 
 
 
 
2793		if (!keep_write)
2794			radix_tree_tag_clear(&mapping->i_pages, page_index(page),
2795						PAGECACHE_TAG_TOWRITE);
2796		xa_unlock_irqrestore(&mapping->i_pages, flags);
2797	} else {
2798		ret = TestSetPageWriteback(page);
2799	}
2800	if (!ret) {
2801		inc_lruvec_page_state(page, NR_WRITEBACK);
2802		inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2803	}
2804	unlock_page_memcg(page);
2805	return ret;
2806
 
 
 
 
 
 
 
 
 
2807}
2808EXPORT_SYMBOL(__test_set_page_writeback);
2809
2810/*
2811 * Return true if any of the pages in the mapping are marked with the
2812 * passed tag.
 
 
 
 
 
 
 
 
2813 */
2814int mapping_tagged(struct address_space *mapping, int tag)
2815{
2816	return radix_tree_tagged(&mapping->i_pages, tag);
 
 
 
2817}
2818EXPORT_SYMBOL(mapping_tagged);
2819
2820/**
2821 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2822 * @page:	The page to wait on.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2823 *
2824 * This function determines if the given page is related to a backing device
2825 * that requires page contents to be held stable during writeback.  If so, then
2826 * it will wait for any pending writeback to complete.
 
2827 */
2828void wait_for_stable_page(struct page *page)
2829{
2830	if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2831		wait_on_page_writeback(page);
2832}
2833EXPORT_SYMBOL_GPL(wait_for_stable_page);
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * mm/page-writeback.c
   4 *
   5 * Copyright (C) 2002, Linus Torvalds.
   6 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
   7 *
   8 * Contains functions related to writing back dirty pages at the
   9 * address_space level.
  10 *
  11 * 10Apr2002	Andrew Morton
  12 *		Initial version
  13 */
  14
  15#include <linux/kernel.h>
  16#include <linux/math64.h>
  17#include <linux/export.h>
  18#include <linux/spinlock.h>
  19#include <linux/fs.h>
  20#include <linux/mm.h>
  21#include <linux/swap.h>
  22#include <linux/slab.h>
  23#include <linux/pagemap.h>
  24#include <linux/writeback.h>
  25#include <linux/init.h>
  26#include <linux/backing-dev.h>
  27#include <linux/task_io_accounting_ops.h>
  28#include <linux/blkdev.h>
  29#include <linux/mpage.h>
  30#include <linux/rmap.h>
  31#include <linux/percpu.h>
 
  32#include <linux/smp.h>
  33#include <linux/sysctl.h>
  34#include <linux/cpu.h>
  35#include <linux/syscalls.h>
 
  36#include <linux/pagevec.h>
  37#include <linux/timer.h>
  38#include <linux/sched/rt.h>
  39#include <linux/sched/signal.h>
  40#include <linux/mm_inline.h>
  41#include <trace/events/writeback.h>
  42
  43#include "internal.h"
  44
  45/*
  46 * Sleep at most 200ms at a time in balance_dirty_pages().
  47 */
  48#define MAX_PAUSE		max(HZ/5, 1)
  49
  50/*
  51 * Try to keep balance_dirty_pages() call intervals higher than this many pages
  52 * by raising pause time to max_pause when falls below it.
  53 */
  54#define DIRTY_POLL_THRESH	(128 >> (PAGE_SHIFT - 10))
  55
  56/*
  57 * Estimate write bandwidth or update dirty limit at 200ms intervals.
  58 */
  59#define BANDWIDTH_INTERVAL	max(HZ/5, 1)
  60
  61#define RATELIMIT_CALC_SHIFT	10
  62
  63/*
  64 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  65 * will look to see if it needs to force writeback or throttling.
  66 */
  67static long ratelimit_pages = 32;
  68
  69/* The following parameters are exported via /proc/sys/vm */
  70
  71/*
  72 * Start background writeback (via writeback threads) at this percentage
  73 */
  74static int dirty_background_ratio = 10;
  75
  76/*
  77 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
  78 * dirty_background_ratio * the amount of dirtyable memory
  79 */
  80static unsigned long dirty_background_bytes;
  81
  82/*
  83 * free highmem will not be subtracted from the total free memory
  84 * for calculating free ratios if vm_highmem_is_dirtyable is true
  85 */
  86static int vm_highmem_is_dirtyable;
  87
  88/*
  89 * The generator of dirty data starts writeback at this percentage
  90 */
  91static int vm_dirty_ratio = 20;
  92
  93/*
  94 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
  95 * vm_dirty_ratio * the amount of dirtyable memory
  96 */
  97static unsigned long vm_dirty_bytes;
  98
  99/*
 100 * The interval between `kupdate'-style writebacks
 101 */
 102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
 103
 104EXPORT_SYMBOL_GPL(dirty_writeback_interval);
 105
 106/*
 107 * The longest time for which data is allowed to remain dirty
 108 */
 109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
 110
 111/*
 
 
 
 
 
 112 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
 113 * a full sync is triggered after this time elapses without any disk activity.
 114 */
 115int laptop_mode;
 116
 117EXPORT_SYMBOL(laptop_mode);
 118
 119/* End of sysctl-exported parameters */
 120
 121struct wb_domain global_wb_domain;
 122
 123/* consolidated parameters for balance_dirty_pages() and its subroutines */
 124struct dirty_throttle_control {
 125#ifdef CONFIG_CGROUP_WRITEBACK
 126	struct wb_domain	*dom;
 127	struct dirty_throttle_control *gdtc;	/* only set in memcg dtc's */
 128#endif
 129	struct bdi_writeback	*wb;
 130	struct fprop_local_percpu *wb_completions;
 131
 132	unsigned long		avail;		/* dirtyable */
 133	unsigned long		dirty;		/* file_dirty + write + nfs */
 134	unsigned long		thresh;		/* dirty threshold */
 135	unsigned long		bg_thresh;	/* dirty background threshold */
 136
 137	unsigned long		wb_dirty;	/* per-wb counterparts */
 138	unsigned long		wb_thresh;
 139	unsigned long		wb_bg_thresh;
 140
 141	unsigned long		pos_ratio;
 142	bool			freerun;
 143	bool			dirty_exceeded;
 144};
 145
 146/*
 147 * Length of period for aging writeout fractions of bdis. This is an
 148 * arbitrarily chosen number. The longer the period, the slower fractions will
 149 * reflect changes in current writeout rate.
 150 */
 151#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
 152
 153#ifdef CONFIG_CGROUP_WRITEBACK
 154
 155#define GDTC_INIT(__wb)		.wb = (__wb),				\
 156				.dom = &global_wb_domain,		\
 157				.wb_completions = &(__wb)->completions
 158
 159#define GDTC_INIT_NO_WB		.dom = &global_wb_domain
 160
 161#define MDTC_INIT(__wb, __gdtc)	.wb = (__wb),				\
 162				.dom = mem_cgroup_wb_domain(__wb),	\
 163				.wb_completions = &(__wb)->memcg_completions, \
 164				.gdtc = __gdtc
 165
 166static bool mdtc_valid(struct dirty_throttle_control *dtc)
 167{
 168	return dtc->dom;
 169}
 170
 171static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
 172{
 173	return dtc->dom;
 174}
 175
 176static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
 177{
 178	return mdtc->gdtc;
 179}
 180
 181static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
 182{
 183	return &wb->memcg_completions;
 184}
 185
 186static void wb_min_max_ratio(struct bdi_writeback *wb,
 187			     unsigned long *minp, unsigned long *maxp)
 188{
 189	unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
 190	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
 191	unsigned long long min = wb->bdi->min_ratio;
 192	unsigned long long max = wb->bdi->max_ratio;
 193
 194	/*
 195	 * @wb may already be clean by the time control reaches here and
 196	 * the total may not include its bw.
 197	 */
 198	if (this_bw < tot_bw) {
 199		if (min) {
 200			min *= this_bw;
 201			min = div64_ul(min, tot_bw);
 202		}
 203		if (max < 100 * BDI_RATIO_SCALE) {
 204			max *= this_bw;
 205			max = div64_ul(max, tot_bw);
 206		}
 207	}
 208
 209	*minp = min;
 210	*maxp = max;
 211}
 212
 213#else	/* CONFIG_CGROUP_WRITEBACK */
 214
 215#define GDTC_INIT(__wb)		.wb = (__wb),                           \
 216				.wb_completions = &(__wb)->completions
 217#define GDTC_INIT_NO_WB
 218#define MDTC_INIT(__wb, __gdtc)
 219
 220static bool mdtc_valid(struct dirty_throttle_control *dtc)
 221{
 222	return false;
 223}
 224
 225static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
 226{
 227	return &global_wb_domain;
 228}
 229
 230static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
 231{
 232	return NULL;
 233}
 234
 235static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
 236{
 237	return NULL;
 238}
 239
 240static void wb_min_max_ratio(struct bdi_writeback *wb,
 241			     unsigned long *minp, unsigned long *maxp)
 242{
 243	*minp = wb->bdi->min_ratio;
 244	*maxp = wb->bdi->max_ratio;
 245}
 246
 247#endif	/* CONFIG_CGROUP_WRITEBACK */
 248
 249/*
 250 * In a memory zone, there is a certain amount of pages we consider
 251 * available for the page cache, which is essentially the number of
 252 * free and reclaimable pages, minus some zone reserves to protect
 253 * lowmem and the ability to uphold the zone's watermarks without
 254 * requiring writeback.
 255 *
 256 * This number of dirtyable pages is the base value of which the
 257 * user-configurable dirty ratio is the effective number of pages that
 258 * are allowed to be actually dirtied.  Per individual zone, or
 259 * globally by using the sum of dirtyable pages over all zones.
 260 *
 261 * Because the user is allowed to specify the dirty limit globally as
 262 * absolute number of bytes, calculating the per-zone dirty limit can
 263 * require translating the configured limit into a percentage of
 264 * global dirtyable memory first.
 265 */
 266
 267/**
 268 * node_dirtyable_memory - number of dirtyable pages in a node
 269 * @pgdat: the node
 270 *
 271 * Return: the node's number of pages potentially available for dirty
 272 * page cache.  This is the base value for the per-node dirty limits.
 273 */
 274static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
 275{
 276	unsigned long nr_pages = 0;
 277	int z;
 278
 279	for (z = 0; z < MAX_NR_ZONES; z++) {
 280		struct zone *zone = pgdat->node_zones + z;
 281
 282		if (!populated_zone(zone))
 283			continue;
 284
 285		nr_pages += zone_page_state(zone, NR_FREE_PAGES);
 286	}
 287
 288	/*
 289	 * Pages reserved for the kernel should not be considered
 290	 * dirtyable, to prevent a situation where reclaim has to
 291	 * clean pages in order to balance the zones.
 292	 */
 293	nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
 294
 295	nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
 296	nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
 297
 298	return nr_pages;
 299}
 300
 301static unsigned long highmem_dirtyable_memory(unsigned long total)
 302{
 303#ifdef CONFIG_HIGHMEM
 304	int node;
 305	unsigned long x = 0;
 306	int i;
 307
 308	for_each_node_state(node, N_HIGH_MEMORY) {
 309		for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
 310			struct zone *z;
 311			unsigned long nr_pages;
 312
 313			if (!is_highmem_idx(i))
 314				continue;
 315
 316			z = &NODE_DATA(node)->node_zones[i];
 317			if (!populated_zone(z))
 318				continue;
 319
 320			nr_pages = zone_page_state(z, NR_FREE_PAGES);
 321			/* watch for underflows */
 322			nr_pages -= min(nr_pages, high_wmark_pages(z));
 323			nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
 324			nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
 325			x += nr_pages;
 326		}
 327	}
 328
 329	/*
 
 
 
 
 
 
 
 
 
 
 
 
 330	 * Make sure that the number of highmem pages is never larger
 331	 * than the number of the total dirtyable memory. This can only
 332	 * occur in very strange VM situations but we want to make sure
 333	 * that this does not occur.
 334	 */
 335	return min(x, total);
 336#else
 337	return 0;
 338#endif
 339}
 340
 341/**
 342 * global_dirtyable_memory - number of globally dirtyable pages
 343 *
 344 * Return: the global number of pages potentially available for dirty
 345 * page cache.  This is the base value for the global dirty limits.
 346 */
 347static unsigned long global_dirtyable_memory(void)
 348{
 349	unsigned long x;
 350
 351	x = global_zone_page_state(NR_FREE_PAGES);
 352	/*
 353	 * Pages reserved for the kernel should not be considered
 354	 * dirtyable, to prevent a situation where reclaim has to
 355	 * clean pages in order to balance the zones.
 356	 */
 357	x -= min(x, totalreserve_pages);
 358
 359	x += global_node_page_state(NR_INACTIVE_FILE);
 360	x += global_node_page_state(NR_ACTIVE_FILE);
 361
 362	if (!vm_highmem_is_dirtyable)
 363		x -= highmem_dirtyable_memory(x);
 364
 365	return x + 1;	/* Ensure that we never return 0 */
 366}
 367
 368/**
 369 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
 370 * @dtc: dirty_throttle_control of interest
 371 *
 372 * Calculate @dtc->thresh and ->bg_thresh considering
 373 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
 374 * must ensure that @dtc->avail is set before calling this function.  The
 375 * dirty limits will be lifted by 1/4 for real-time tasks.
 
 376 */
 377static void domain_dirty_limits(struct dirty_throttle_control *dtc)
 378{
 379	const unsigned long available_memory = dtc->avail;
 380	struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
 381	unsigned long bytes = vm_dirty_bytes;
 382	unsigned long bg_bytes = dirty_background_bytes;
 383	/* convert ratios to per-PAGE_SIZE for higher precision */
 384	unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
 385	unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
 386	unsigned long thresh;
 387	unsigned long bg_thresh;
 388	struct task_struct *tsk;
 389
 390	/* gdtc is !NULL iff @dtc is for memcg domain */
 391	if (gdtc) {
 392		unsigned long global_avail = gdtc->avail;
 393
 394		/*
 395		 * The byte settings can't be applied directly to memcg
 396		 * domains.  Convert them to ratios by scaling against
 397		 * globally available memory.  As the ratios are in
 398		 * per-PAGE_SIZE, they can be obtained by dividing bytes by
 399		 * number of pages.
 400		 */
 401		if (bytes)
 402			ratio = min(DIV_ROUND_UP(bytes, global_avail),
 403				    PAGE_SIZE);
 404		if (bg_bytes)
 405			bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
 406				       PAGE_SIZE);
 407		bytes = bg_bytes = 0;
 408	}
 409
 410	if (bytes)
 411		thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
 412	else
 413		thresh = (ratio * available_memory) / PAGE_SIZE;
 414
 415	if (bg_bytes)
 416		bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
 417	else
 418		bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
 419
 
 
 420	tsk = current;
 421	if (rt_or_dl_task(tsk)) {
 422		bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
 423		thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
 424	}
 425	/*
 426	 * Dirty throttling logic assumes the limits in page units fit into
 427	 * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
 428	 */
 429	if (thresh > UINT_MAX)
 430		thresh = UINT_MAX;
 431	/* This makes sure bg_thresh is within 32-bits as well */
 432	if (bg_thresh >= thresh)
 433		bg_thresh = thresh / 2;
 434	dtc->thresh = thresh;
 435	dtc->bg_thresh = bg_thresh;
 436
 437	/* we should eventually report the domain in the TP */
 438	if (!gdtc)
 439		trace_global_dirty_state(bg_thresh, thresh);
 440}
 441
 442/**
 443 * global_dirty_limits - background-writeback and dirty-throttling thresholds
 444 * @pbackground: out parameter for bg_thresh
 445 * @pdirty: out parameter for thresh
 446 *
 447 * Calculate bg_thresh and thresh for global_wb_domain.  See
 448 * domain_dirty_limits() for details.
 449 */
 450void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
 451{
 452	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
 453
 454	gdtc.avail = global_dirtyable_memory();
 455	domain_dirty_limits(&gdtc);
 456
 457	*pbackground = gdtc.bg_thresh;
 458	*pdirty = gdtc.thresh;
 459}
 460
 461/**
 462 * node_dirty_limit - maximum number of dirty pages allowed in a node
 463 * @pgdat: the node
 464 *
 465 * Return: the maximum number of dirty pages allowed in a node, based
 466 * on the node's dirtyable memory.
 467 */
 468static unsigned long node_dirty_limit(struct pglist_data *pgdat)
 469{
 470	unsigned long node_memory = node_dirtyable_memory(pgdat);
 471	struct task_struct *tsk = current;
 472	unsigned long dirty;
 473
 474	if (vm_dirty_bytes)
 475		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
 476			node_memory / global_dirtyable_memory();
 477	else
 478		dirty = vm_dirty_ratio * node_memory / 100;
 479
 480	if (rt_or_dl_task(tsk))
 481		dirty += dirty / 4;
 482
 483	/*
 484	 * Dirty throttling logic assumes the limits in page units fit into
 485	 * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
 486	 */
 487	return min_t(unsigned long, dirty, UINT_MAX);
 488}
 489
 490/**
 491 * node_dirty_ok - tells whether a node is within its dirty limits
 492 * @pgdat: the node to check
 493 *
 494 * Return: %true when the dirty pages in @pgdat are within the node's
 495 * dirty limit, %false if the limit is exceeded.
 496 */
 497bool node_dirty_ok(struct pglist_data *pgdat)
 498{
 499	unsigned long limit = node_dirty_limit(pgdat);
 500	unsigned long nr_pages = 0;
 501
 502	nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
 
 503	nr_pages += node_page_state(pgdat, NR_WRITEBACK);
 504
 505	return nr_pages <= limit;
 506}
 507
 508#ifdef CONFIG_SYSCTL
 509static int dirty_background_ratio_handler(const struct ctl_table *table, int write,
 510		void *buffer, size_t *lenp, loff_t *ppos)
 511{
 512	int ret;
 513
 514	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 515	if (ret == 0 && write)
 516		dirty_background_bytes = 0;
 517	return ret;
 518}
 519
 520static int dirty_background_bytes_handler(const struct ctl_table *table, int write,
 521		void *buffer, size_t *lenp, loff_t *ppos)
 
 522{
 523	int ret;
 524	unsigned long old_bytes = dirty_background_bytes;
 525
 526	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 527	if (ret == 0 && write) {
 528		if (DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE) >
 529								UINT_MAX) {
 530			dirty_background_bytes = old_bytes;
 531			return -ERANGE;
 532		}
 533		dirty_background_ratio = 0;
 534	}
 535	return ret;
 536}
 537
 538static int dirty_ratio_handler(const struct ctl_table *table, int write, void *buffer,
 539		size_t *lenp, loff_t *ppos)
 
 540{
 541	int old_ratio = vm_dirty_ratio;
 542	int ret;
 543
 544	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 545	if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
 546		writeback_set_ratelimit();
 547		vm_dirty_bytes = 0;
 548	}
 549	return ret;
 550}
 551
 552static int dirty_bytes_handler(const struct ctl_table *table, int write,
 553		void *buffer, size_t *lenp, loff_t *ppos)
 
 554{
 555	unsigned long old_bytes = vm_dirty_bytes;
 556	int ret;
 557
 558	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 559	if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
 560		if (DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) > UINT_MAX) {
 561			vm_dirty_bytes = old_bytes;
 562			return -ERANGE;
 563		}
 564		writeback_set_ratelimit();
 565		vm_dirty_ratio = 0;
 566	}
 567	return ret;
 568}
 569#endif
 570
 571static unsigned long wp_next_time(unsigned long cur_time)
 572{
 573	cur_time += VM_COMPLETIONS_PERIOD_LEN;
 574	/* 0 has a special meaning... */
 575	if (!cur_time)
 576		return 1;
 577	return cur_time;
 578}
 579
 580static void wb_domain_writeout_add(struct wb_domain *dom,
 581				   struct fprop_local_percpu *completions,
 582				   unsigned int max_prop_frac, long nr)
 583{
 584	__fprop_add_percpu_max(&dom->completions, completions,
 585			       max_prop_frac, nr);
 586	/* First event after period switching was turned off? */
 587	if (unlikely(!dom->period_time)) {
 588		/*
 589		 * We can race with other wb_domain_writeout_add calls here but
 590		 * it does not cause any harm since the resulting time when
 591		 * timer will fire and what is in writeout_period_time will be
 592		 * roughly the same.
 593		 */
 594		dom->period_time = wp_next_time(jiffies);
 595		mod_timer(&dom->period_timer, dom->period_time);
 596	}
 597}
 598
 599/*
 600 * Increment @wb's writeout completion count and the global writeout
 601 * completion count. Called from __folio_end_writeback().
 602 */
 603static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
 604{
 605	struct wb_domain *cgdom;
 606
 607	wb_stat_mod(wb, WB_WRITTEN, nr);
 608	wb_domain_writeout_add(&global_wb_domain, &wb->completions,
 609			       wb->bdi->max_prop_frac, nr);
 610
 611	cgdom = mem_cgroup_wb_domain(wb);
 612	if (cgdom)
 613		wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
 614				       wb->bdi->max_prop_frac, nr);
 615}
 616
 617void wb_writeout_inc(struct bdi_writeback *wb)
 618{
 619	unsigned long flags;
 620
 621	local_irq_save(flags);
 622	__wb_writeout_add(wb, 1);
 623	local_irq_restore(flags);
 624}
 625EXPORT_SYMBOL_GPL(wb_writeout_inc);
 626
 627/*
 628 * On idle system, we can be called long after we scheduled because we use
 629 * deferred timers so count with missed periods.
 630 */
 631static void writeout_period(struct timer_list *t)
 632{
 633	struct wb_domain *dom = from_timer(dom, t, period_timer);
 634	int miss_periods = (jiffies - dom->period_time) /
 635						 VM_COMPLETIONS_PERIOD_LEN;
 636
 637	if (fprop_new_period(&dom->completions, miss_periods + 1)) {
 638		dom->period_time = wp_next_time(dom->period_time +
 639				miss_periods * VM_COMPLETIONS_PERIOD_LEN);
 640		mod_timer(&dom->period_timer, dom->period_time);
 641	} else {
 642		/*
 643		 * Aging has zeroed all fractions. Stop wasting CPU on period
 644		 * updates.
 645		 */
 646		dom->period_time = 0;
 647	}
 648}
 649
 650int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
 651{
 652	memset(dom, 0, sizeof(*dom));
 653
 654	spin_lock_init(&dom->lock);
 655
 656	timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
 657
 658	dom->dirty_limit_tstamp = jiffies;
 659
 660	return fprop_global_init(&dom->completions, gfp);
 661}
 662
 663#ifdef CONFIG_CGROUP_WRITEBACK
 664void wb_domain_exit(struct wb_domain *dom)
 665{
 666	del_timer_sync(&dom->period_timer);
 667	fprop_global_destroy(&dom->completions);
 668}
 669#endif
 670
 671/*
 672 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
 673 * registered backing devices, which, for obvious reasons, can not
 674 * exceed 100%.
 675 */
 676static unsigned int bdi_min_ratio;
 677
 678static int bdi_check_pages_limit(unsigned long pages)
 679{
 680	unsigned long max_dirty_pages = global_dirtyable_memory();
 681
 682	if (pages > max_dirty_pages)
 683		return -EINVAL;
 684
 685	return 0;
 686}
 687
 688static unsigned long bdi_ratio_from_pages(unsigned long pages)
 689{
 690	unsigned long background_thresh;
 691	unsigned long dirty_thresh;
 692	unsigned long ratio;
 693
 694	global_dirty_limits(&background_thresh, &dirty_thresh);
 695	if (!dirty_thresh)
 696		return -EINVAL;
 697	ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh);
 698
 699	return ratio;
 700}
 701
 702static u64 bdi_get_bytes(unsigned int ratio)
 703{
 704	unsigned long background_thresh;
 705	unsigned long dirty_thresh;
 706	u64 bytes;
 707
 708	global_dirty_limits(&background_thresh, &dirty_thresh);
 709	bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
 710
 711	return bytes;
 712}
 713
 714static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 715{
 716	unsigned int delta;
 717	int ret = 0;
 718
 719	if (min_ratio > 100 * BDI_RATIO_SCALE)
 720		return -EINVAL;
 721
 722	spin_lock_bh(&bdi_lock);
 723	if (min_ratio > bdi->max_ratio) {
 724		ret = -EINVAL;
 725	} else {
 726		if (min_ratio < bdi->min_ratio) {
 727			delta = bdi->min_ratio - min_ratio;
 728			bdi_min_ratio -= delta;
 729			bdi->min_ratio = min_ratio;
 730		} else {
 731			delta = min_ratio - bdi->min_ratio;
 732			if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
 733				bdi_min_ratio += delta;
 734				bdi->min_ratio = min_ratio;
 735			} else {
 736				ret = -EINVAL;
 737			}
 738		}
 739	}
 740	spin_unlock_bh(&bdi_lock);
 741
 742	return ret;
 743}
 744
 745static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
 746{
 747	int ret = 0;
 748
 749	if (max_ratio > 100 * BDI_RATIO_SCALE)
 750		return -EINVAL;
 751
 752	spin_lock_bh(&bdi_lock);
 753	if (bdi->min_ratio > max_ratio) {
 754		ret = -EINVAL;
 755	} else {
 756		bdi->max_ratio = max_ratio;
 757		bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) /
 758						(100 * BDI_RATIO_SCALE);
 759	}
 760	spin_unlock_bh(&bdi_lock);
 761
 762	return ret;
 763}
 764
 765int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio)
 766{
 767	return __bdi_set_min_ratio(bdi, min_ratio);
 768}
 769
 770int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio)
 771{
 772	return __bdi_set_max_ratio(bdi, max_ratio);
 773}
 774
 775int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 776{
 777	return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE);
 778}
 779
 780int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
 781{
 782	return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
 783}
 784EXPORT_SYMBOL(bdi_set_max_ratio);
 785
 786u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
 787{
 788	return bdi_get_bytes(bdi->min_ratio);
 789}
 790
 791int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes)
 792{
 793	int ret;
 794	unsigned long pages = min_bytes >> PAGE_SHIFT;
 795	long min_ratio;
 796
 797	ret = bdi_check_pages_limit(pages);
 798	if (ret)
 799		return ret;
 800
 801	min_ratio = bdi_ratio_from_pages(pages);
 802	if (min_ratio < 0)
 803		return min_ratio;
 804	return __bdi_set_min_ratio(bdi, min_ratio);
 805}
 806
 807u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
 808{
 809	return bdi_get_bytes(bdi->max_ratio);
 810}
 811
 812int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
 813{
 814	int ret;
 815	unsigned long pages = max_bytes >> PAGE_SHIFT;
 816	long max_ratio;
 817
 818	ret = bdi_check_pages_limit(pages);
 819	if (ret)
 820		return ret;
 821
 822	max_ratio = bdi_ratio_from_pages(pages);
 823	if (max_ratio < 0)
 824		return max_ratio;
 825	return __bdi_set_max_ratio(bdi, max_ratio);
 826}
 827
 828int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
 829{
 830	if (strict_limit > 1)
 831		return -EINVAL;
 832
 833	spin_lock_bh(&bdi_lock);
 834	if (strict_limit)
 835		bdi->capabilities |= BDI_CAP_STRICTLIMIT;
 836	else
 837		bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
 838	spin_unlock_bh(&bdi_lock);
 839
 840	return 0;
 841}
 842
 843static unsigned long dirty_freerun_ceiling(unsigned long thresh,
 844					   unsigned long bg_thresh)
 845{
 846	return (thresh + bg_thresh) / 2;
 847}
 848
 849static unsigned long hard_dirty_limit(struct wb_domain *dom,
 850				      unsigned long thresh)
 851{
 852	return max(thresh, dom->dirty_limit);
 853}
 854
 855/*
 856 * Memory which can be further allocated to a memcg domain is capped by
 857 * system-wide clean memory excluding the amount being used in the domain.
 858 */
 859static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
 860			    unsigned long filepages, unsigned long headroom)
 861{
 862	struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
 863	unsigned long clean = filepages - min(filepages, mdtc->dirty);
 864	unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
 865	unsigned long other_clean = global_clean - min(global_clean, clean);
 866
 867	mdtc->avail = filepages + min(headroom, other_clean);
 868}
 869
 870static inline bool dtc_is_global(struct dirty_throttle_control *dtc)
 871{
 872	return mdtc_gdtc(dtc) == NULL;
 873}
 874
 875/*
 876 * Dirty background will ignore pages being written as we're trying to
 877 * decide whether to put more under writeback.
 878 */
 879static void domain_dirty_avail(struct dirty_throttle_control *dtc,
 880			       bool include_writeback)
 881{
 882	if (dtc_is_global(dtc)) {
 883		dtc->avail = global_dirtyable_memory();
 884		dtc->dirty = global_node_page_state(NR_FILE_DIRTY);
 885		if (include_writeback)
 886			dtc->dirty += global_node_page_state(NR_WRITEBACK);
 887	} else {
 888		unsigned long filepages = 0, headroom = 0, writeback = 0;
 889
 890		mem_cgroup_wb_stats(dtc->wb, &filepages, &headroom, &dtc->dirty,
 891				    &writeback);
 892		if (include_writeback)
 893			dtc->dirty += writeback;
 894		mdtc_calc_avail(dtc, filepages, headroom);
 895	}
 896}
 897
 898/**
 899 * __wb_calc_thresh - @wb's share of dirty threshold
 900 * @dtc: dirty_throttle_context of interest
 901 * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc
 902 *
 903 * Note that balance_dirty_pages() will only seriously take dirty throttling
 904 * threshold as a hard limit when sleeping max_pause per page is not enough
 905 * to keep the dirty pages under control. For example, when the device is
 906 * completely stalled due to some error conditions, or when there are 1000
 907 * dd tasks writing to a slow 10MB/s USB key.
 
 
 908 * In the other normal situations, it acts more gently by throttling the tasks
 909 * more (rather than completely block them) when the wb dirty pages go high.
 910 *
 911 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
 912 * - starving fast devices
 913 * - piling up dirty pages (that will take long time to sync) on slow devices
 914 *
 915 * The wb's share of dirty limit will be adapting to its throughput and
 916 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
 917 *
 918 * Return: @wb's dirty limit in pages. For dirty throttling limit, the term
 919 * "dirty" in the context of dirty balancing includes all PG_dirty and
 920 * PG_writeback pages.
 921 */
 922static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc,
 923				      unsigned long thresh)
 924{
 925	struct wb_domain *dom = dtc_dom(dtc);
 926	struct bdi_writeback *wb = dtc->wb;
 927	u64 wb_thresh;
 928	u64 wb_max_thresh;
 929	unsigned long numerator, denominator;
 930	unsigned long wb_min_ratio, wb_max_ratio;
 931
 932	/*
 933	 * Calculate this wb's share of the thresh ratio.
 934	 */
 935	fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
 936			      &numerator, &denominator);
 937
 938	wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
 939	wb_thresh *= numerator;
 940	wb_thresh = div64_ul(wb_thresh, denominator);
 941
 942	wb_min_max_ratio(wb, &wb_min_ratio, &wb_max_ratio);
 943
 944	wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
 945	wb_max_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
 946	if (wb_thresh > wb_max_thresh)
 947		wb_thresh = wb_max_thresh;
 948
 949	/*
 950	 * With strictlimit flag, the wb_thresh is treated as
 951	 * a hard limit in balance_dirty_pages() and wb_position_ratio().
 952	 * It's possible that wb_thresh is close to zero, not because
 953	 * the device is slow, but because it has been inactive.
 954	 * To prevent occasional writes from being blocked, we raise wb_thresh.
 955	 */
 956	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
 957		unsigned long limit = hard_dirty_limit(dom, dtc->thresh);
 958		u64 wb_scale_thresh = 0;
 959
 960		if (limit > dtc->dirty)
 961			wb_scale_thresh = (limit - dtc->dirty) / 100;
 962		wb_thresh = max(wb_thresh, min(wb_scale_thresh, wb_max_thresh / 4));
 963	}
 964
 965	return wb_thresh;
 966}
 967
 968unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
 969{
 970	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
 971
 972	return __wb_calc_thresh(&gdtc, thresh);
 973}
 974
 975unsigned long cgwb_calc_thresh(struct bdi_writeback *wb)
 976{
 977	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
 978	struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
 979
 980	domain_dirty_avail(&gdtc, true);
 981	domain_dirty_avail(&mdtc, true);
 982	domain_dirty_limits(&mdtc);
 983
 984	return __wb_calc_thresh(&mdtc, mdtc.thresh);
 985}
 986
 987/*
 988 *                           setpoint - dirty 3
 989 *        f(dirty) := 1.0 + (----------------)
 990 *                           limit - setpoint
 991 *
 992 * it's a 3rd order polynomial that subjects to
 993 *
 994 * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
 995 * (2) f(setpoint) = 1.0 => the balance point
 996 * (3) f(limit)    = 0   => the hard limit
 997 * (4) df/dx      <= 0	 => negative feedback control
 998 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
 999 *     => fast response on large errors; small oscillation near setpoint
1000 */
1001static long long pos_ratio_polynom(unsigned long setpoint,
1002					  unsigned long dirty,
1003					  unsigned long limit)
1004{
1005	long long pos_ratio;
1006	long x;
1007
1008	x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
1009		      (limit - setpoint) | 1);
1010	pos_ratio = x;
1011	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
1012	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
1013	pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
1014
1015	return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
1016}
1017
1018/*
1019 * Dirty position control.
1020 *
1021 * (o) global/bdi setpoints
1022 *
1023 * We want the dirty pages be balanced around the global/wb setpoints.
1024 * When the number of dirty pages is higher/lower than the setpoint, the
1025 * dirty position control ratio (and hence task dirty ratelimit) will be
1026 * decreased/increased to bring the dirty pages back to the setpoint.
1027 *
1028 *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
1029 *
1030 *     if (dirty < setpoint) scale up   pos_ratio
1031 *     if (dirty > setpoint) scale down pos_ratio
1032 *
1033 *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
1034 *     if (wb_dirty > wb_setpoint) scale down pos_ratio
1035 *
1036 *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
1037 *
1038 * (o) global control line
1039 *
1040 *     ^ pos_ratio
1041 *     |
1042 *     |            |<===== global dirty control scope ======>|
1043 * 2.0  * * * * * * *
1044 *     |            .*
1045 *     |            . *
1046 *     |            .   *
1047 *     |            .     *
1048 *     |            .        *
1049 *     |            .            *
1050 * 1.0 ................................*
1051 *     |            .                  .     *
1052 *     |            .                  .          *
1053 *     |            .                  .              *
1054 *     |            .                  .                 *
1055 *     |            .                  .                    *
1056 *   0 +------------.------------------.----------------------*------------->
1057 *           freerun^          setpoint^                 limit^   dirty pages
1058 *
1059 * (o) wb control line
1060 *
1061 *     ^ pos_ratio
1062 *     |
1063 *     |            *
1064 *     |              *
1065 *     |                *
1066 *     |                  *
1067 *     |                    * |<=========== span ============>|
1068 * 1.0 .......................*
1069 *     |                      . *
1070 *     |                      .   *
1071 *     |                      .     *
1072 *     |                      .       *
1073 *     |                      .         *
1074 *     |                      .           *
1075 *     |                      .             *
1076 *     |                      .               *
1077 *     |                      .                 *
1078 *     |                      .                   *
1079 *     |                      .                     *
1080 * 1/4 ...............................................* * * * * * * * * * * *
1081 *     |                      .                         .
1082 *     |                      .                           .
1083 *     |                      .                             .
1084 *   0 +----------------------.-------------------------------.------------->
1085 *                wb_setpoint^                    x_intercept^
1086 *
1087 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
1088 * be smoothly throttled down to normal if it starts high in situations like
1089 * - start writing to a slow SD card and a fast disk at the same time. The SD
1090 *   card's wb_dirty may rush to many times higher than wb_setpoint.
1091 * - the wb dirty thresh drops quickly due to change of JBOD workload
1092 */
1093static void wb_position_ratio(struct dirty_throttle_control *dtc)
1094{
1095	struct bdi_writeback *wb = dtc->wb;
1096	unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
1097	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1098	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1099	unsigned long wb_thresh = dtc->wb_thresh;
1100	unsigned long x_intercept;
1101	unsigned long setpoint;		/* dirty pages' target balance point */
1102	unsigned long wb_setpoint;
1103	unsigned long span;
1104	long long pos_ratio;		/* for scaling up/down the rate limit */
1105	long x;
1106
1107	dtc->pos_ratio = 0;
1108
1109	if (unlikely(dtc->dirty >= limit))
1110		return;
1111
1112	/*
1113	 * global setpoint
1114	 *
1115	 * See comment for pos_ratio_polynom().
1116	 */
1117	setpoint = (freerun + limit) / 2;
1118	pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
1119
1120	/*
1121	 * The strictlimit feature is a tool preventing mistrusted filesystems
1122	 * from growing a large number of dirty pages before throttling. For
1123	 * such filesystems balance_dirty_pages always checks wb counters
1124	 * against wb limits. Even if global "nr_dirty" is under "freerun".
1125	 * This is especially important for fuse which sets bdi->max_ratio to
1126	 * 1% by default. Without strictlimit feature, fuse writeback may
1127	 * consume arbitrary amount of RAM because it is accounted in
1128	 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
1129	 *
1130	 * Here, in wb_position_ratio(), we calculate pos_ratio based on
1131	 * two values: wb_dirty and wb_thresh. Let's consider an example:
1132	 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
1133	 * limits are set by default to 10% and 20% (background and throttle).
1134	 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
1135	 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
1136	 * about ~6K pages (as the average of background and throttle wb
1137	 * limits). The 3rd order polynomial will provide positive feedback if
1138	 * wb_dirty is under wb_setpoint and vice versa.
1139	 *
1140	 * Note, that we cannot use global counters in these calculations
1141	 * because we want to throttle process writing to a strictlimit wb
1142	 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
1143	 * in the example above).
1144	 */
1145	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1146		long long wb_pos_ratio;
1147
1148		if (dtc->wb_dirty < 8) {
1149			dtc->pos_ratio = min_t(long long, pos_ratio * 2,
1150					   2 << RATELIMIT_CALC_SHIFT);
1151			return;
1152		}
1153
1154		if (dtc->wb_dirty >= wb_thresh)
1155			return;
1156
1157		wb_setpoint = dirty_freerun_ceiling(wb_thresh,
1158						    dtc->wb_bg_thresh);
1159
1160		if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
1161			return;
1162
1163		wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
1164						 wb_thresh);
1165
1166		/*
1167		 * Typically, for strictlimit case, wb_setpoint << setpoint
1168		 * and pos_ratio >> wb_pos_ratio. In the other words global
1169		 * state ("dirty") is not limiting factor and we have to
1170		 * make decision based on wb counters. But there is an
1171		 * important case when global pos_ratio should get precedence:
1172		 * global limits are exceeded (e.g. due to activities on other
1173		 * wb's) while given strictlimit wb is below limit.
1174		 *
1175		 * "pos_ratio * wb_pos_ratio" would work for the case above,
1176		 * but it would look too non-natural for the case of all
1177		 * activity in the system coming from a single strictlimit wb
1178		 * with bdi->max_ratio == 100%.
1179		 *
1180		 * Note that min() below somewhat changes the dynamics of the
1181		 * control system. Normally, pos_ratio value can be well over 3
1182		 * (when globally we are at freerun and wb is well below wb
1183		 * setpoint). Now the maximum pos_ratio in the same situation
1184		 * is 2. We might want to tweak this if we observe the control
1185		 * system is too slow to adapt.
1186		 */
1187		dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
1188		return;
1189	}
1190
1191	/*
1192	 * We have computed basic pos_ratio above based on global situation. If
1193	 * the wb is over/under its share of dirty pages, we want to scale
1194	 * pos_ratio further down/up. That is done by the following mechanism.
1195	 */
1196
1197	/*
1198	 * wb setpoint
1199	 *
1200	 *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1201	 *
1202	 *                        x_intercept - wb_dirty
1203	 *                     := --------------------------
1204	 *                        x_intercept - wb_setpoint
1205	 *
1206	 * The main wb control line is a linear function that subjects to
1207	 *
1208	 * (1) f(wb_setpoint) = 1.0
1209	 * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1210	 *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1211	 *
1212	 * For single wb case, the dirty pages are observed to fluctuate
1213	 * regularly within range
1214	 *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1215	 * for various filesystems, where (2) can yield in a reasonable 12.5%
1216	 * fluctuation range for pos_ratio.
1217	 *
1218	 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1219	 * own size, so move the slope over accordingly and choose a slope that
1220	 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1221	 */
1222	if (unlikely(wb_thresh > dtc->thresh))
1223		wb_thresh = dtc->thresh;
1224	/*
1225	 * It's very possible that wb_thresh is close to 0 not because the
1226	 * device is slow, but that it has remained inactive for long time.
1227	 * Honour such devices a reasonable good (hopefully IO efficient)
1228	 * threshold, so that the occasional writes won't be blocked and active
1229	 * writes can rampup the threshold quickly.
1230	 */
1231	wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1232	/*
1233	 * scale global setpoint to wb's:
1234	 *	wb_setpoint = setpoint * wb_thresh / thresh
1235	 */
1236	x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1237	wb_setpoint = setpoint * (u64)x >> 16;
1238	/*
1239	 * Use span=(8*write_bw) in single wb case as indicated by
1240	 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1241	 *
1242	 *        wb_thresh                    thresh - wb_thresh
1243	 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1244	 *         thresh                           thresh
1245	 */
1246	span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1247	x_intercept = wb_setpoint + span;
1248
1249	if (dtc->wb_dirty < x_intercept - span / 4) {
1250		pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1251				      (x_intercept - wb_setpoint) | 1);
1252	} else
1253		pos_ratio /= 4;
1254
1255	/*
1256	 * wb reserve area, safeguard against dirty pool underrun and disk idle
1257	 * It may push the desired control point of global dirty pages higher
1258	 * than setpoint.
1259	 */
1260	x_intercept = wb_thresh / 2;
1261	if (dtc->wb_dirty < x_intercept) {
1262		if (dtc->wb_dirty > x_intercept / 8)
1263			pos_ratio = div_u64(pos_ratio * x_intercept,
1264					    dtc->wb_dirty);
1265		else
1266			pos_ratio *= 8;
1267	}
1268
1269	dtc->pos_ratio = pos_ratio;
1270}
1271
1272static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1273				      unsigned long elapsed,
1274				      unsigned long written)
1275{
1276	const unsigned long period = roundup_pow_of_two(3 * HZ);
1277	unsigned long avg = wb->avg_write_bandwidth;
1278	unsigned long old = wb->write_bandwidth;
1279	u64 bw;
1280
1281	/*
1282	 * bw = written * HZ / elapsed
1283	 *
1284	 *                   bw * elapsed + write_bandwidth * (period - elapsed)
1285	 * write_bandwidth = ---------------------------------------------------
1286	 *                                          period
1287	 *
1288	 * @written may have decreased due to folio_redirty_for_writepage().
1289	 * Avoid underflowing @bw calculation.
1290	 */
1291	bw = written - min(written, wb->written_stamp);
1292	bw *= HZ;
1293	if (unlikely(elapsed > period)) {
1294		bw = div64_ul(bw, elapsed);
1295		avg = bw;
1296		goto out;
1297	}
1298	bw += (u64)wb->write_bandwidth * (period - elapsed);
1299	bw >>= ilog2(period);
1300
1301	/*
1302	 * one more level of smoothing, for filtering out sudden spikes
1303	 */
1304	if (avg > old && old >= (unsigned long)bw)
1305		avg -= (avg - old) >> 3;
1306
1307	if (avg < old && old <= (unsigned long)bw)
1308		avg += (old - avg) >> 3;
1309
1310out:
1311	/* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1312	avg = max(avg, 1LU);
1313	if (wb_has_dirty_io(wb)) {
1314		long delta = avg - wb->avg_write_bandwidth;
1315		WARN_ON_ONCE(atomic_long_add_return(delta,
1316					&wb->bdi->tot_write_bandwidth) <= 0);
1317	}
1318	wb->write_bandwidth = bw;
1319	WRITE_ONCE(wb->avg_write_bandwidth, avg);
1320}
1321
1322static void update_dirty_limit(struct dirty_throttle_control *dtc)
1323{
1324	struct wb_domain *dom = dtc_dom(dtc);
1325	unsigned long thresh = dtc->thresh;
1326	unsigned long limit = dom->dirty_limit;
1327
1328	/*
1329	 * Follow up in one step.
1330	 */
1331	if (limit < thresh) {
1332		limit = thresh;
1333		goto update;
1334	}
1335
1336	/*
1337	 * Follow down slowly. Use the higher one as the target, because thresh
1338	 * may drop below dirty. This is exactly the reason to introduce
1339	 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1340	 */
1341	thresh = max(thresh, dtc->dirty);
1342	if (limit > thresh) {
1343		limit -= (limit - thresh) >> 5;
1344		goto update;
1345	}
1346	return;
1347update:
1348	dom->dirty_limit = limit;
1349}
1350
1351static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
1352				      unsigned long now)
1353{
1354	struct wb_domain *dom = dtc_dom(dtc);
1355
1356	/*
1357	 * check locklessly first to optimize away locking for the most time
1358	 */
1359	if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1360		return;
1361
1362	spin_lock(&dom->lock);
1363	if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1364		update_dirty_limit(dtc);
1365		dom->dirty_limit_tstamp = now;
1366	}
1367	spin_unlock(&dom->lock);
1368}
1369
1370/*
1371 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1372 *
1373 * Normal wb tasks will be curbed at or below it in long term.
1374 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1375 */
1376static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1377				      unsigned long dirtied,
1378				      unsigned long elapsed)
1379{
1380	struct bdi_writeback *wb = dtc->wb;
1381	unsigned long dirty = dtc->dirty;
1382	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1383	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1384	unsigned long setpoint = (freerun + limit) / 2;
1385	unsigned long write_bw = wb->avg_write_bandwidth;
1386	unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1387	unsigned long dirty_rate;
1388	unsigned long task_ratelimit;
1389	unsigned long balanced_dirty_ratelimit;
1390	unsigned long step;
1391	unsigned long x;
1392	unsigned long shift;
1393
1394	/*
1395	 * The dirty rate will match the writeout rate in long term, except
1396	 * when dirty pages are truncated by userspace or re-dirtied by FS.
1397	 */
1398	dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1399
1400	/*
1401	 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1402	 */
1403	task_ratelimit = (u64)dirty_ratelimit *
1404					dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1405	task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1406
1407	/*
1408	 * A linear estimation of the "balanced" throttle rate. The theory is,
1409	 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1410	 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1411	 * formula will yield the balanced rate limit (write_bw / N).
1412	 *
1413	 * Note that the expanded form is not a pure rate feedback:
1414	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate)		     (1)
1415	 * but also takes pos_ratio into account:
1416	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1417	 *
1418	 * (1) is not realistic because pos_ratio also takes part in balancing
1419	 * the dirty rate.  Consider the state
1420	 *	pos_ratio = 0.5						     (3)
1421	 *	rate = 2 * (write_bw / N)				     (4)
1422	 * If (1) is used, it will stuck in that state! Because each dd will
1423	 * be throttled at
1424	 *	task_ratelimit = pos_ratio * rate = (write_bw / N)	     (5)
1425	 * yielding
1426	 *	dirty_rate = N * task_ratelimit = write_bw		     (6)
1427	 * put (6) into (1) we get
1428	 *	rate_(i+1) = rate_(i)					     (7)
1429	 *
1430	 * So we end up using (2) to always keep
1431	 *	rate_(i+1) ~= (write_bw / N)				     (8)
1432	 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1433	 * pos_ratio is able to drive itself to 1.0, which is not only where
1434	 * the dirty count meet the setpoint, but also where the slope of
1435	 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1436	 */
1437	balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1438					   dirty_rate | 1);
1439	/*
1440	 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1441	 */
1442	if (unlikely(balanced_dirty_ratelimit > write_bw))
1443		balanced_dirty_ratelimit = write_bw;
1444
1445	/*
1446	 * We could safely do this and return immediately:
1447	 *
1448	 *	wb->dirty_ratelimit = balanced_dirty_ratelimit;
1449	 *
1450	 * However to get a more stable dirty_ratelimit, the below elaborated
1451	 * code makes use of task_ratelimit to filter out singular points and
1452	 * limit the step size.
1453	 *
1454	 * The below code essentially only uses the relative value of
1455	 *
1456	 *	task_ratelimit - dirty_ratelimit
1457	 *	= (pos_ratio - 1) * dirty_ratelimit
1458	 *
1459	 * which reflects the direction and size of dirty position error.
1460	 */
1461
1462	/*
1463	 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1464	 * task_ratelimit is on the same side of dirty_ratelimit, too.
1465	 * For example, when
1466	 * - dirty_ratelimit > balanced_dirty_ratelimit
1467	 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1468	 * lowering dirty_ratelimit will help meet both the position and rate
1469	 * control targets. Otherwise, don't update dirty_ratelimit if it will
1470	 * only help meet the rate target. After all, what the users ultimately
1471	 * feel and care are stable dirty rate and small position error.
1472	 *
1473	 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1474	 * and filter out the singular points of balanced_dirty_ratelimit. Which
1475	 * keeps jumping around randomly and can even leap far away at times
1476	 * due to the small 200ms estimation period of dirty_rate (we want to
1477	 * keep that period small to reduce time lags).
1478	 */
1479	step = 0;
1480
1481	/*
1482	 * For strictlimit case, calculations above were based on wb counters
1483	 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1484	 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1485	 * Hence, to calculate "step" properly, we have to use wb_dirty as
1486	 * "dirty" and wb_setpoint as "setpoint".
1487	 *
1488	 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1489	 * it's possible that wb_thresh is close to zero due to inactivity
1490	 * of backing device.
1491	 */
1492	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1493		dirty = dtc->wb_dirty;
1494		if (dtc->wb_dirty < 8)
1495			setpoint = dtc->wb_dirty + 1;
1496		else
1497			setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1498	}
1499
1500	if (dirty < setpoint) {
1501		x = min3(wb->balanced_dirty_ratelimit,
1502			 balanced_dirty_ratelimit, task_ratelimit);
1503		if (dirty_ratelimit < x)
1504			step = x - dirty_ratelimit;
1505	} else {
1506		x = max3(wb->balanced_dirty_ratelimit,
1507			 balanced_dirty_ratelimit, task_ratelimit);
1508		if (dirty_ratelimit > x)
1509			step = dirty_ratelimit - x;
1510	}
1511
1512	/*
1513	 * Don't pursue 100% rate matching. It's impossible since the balanced
1514	 * rate itself is constantly fluctuating. So decrease the track speed
1515	 * when it gets close to the target. Helps eliminate pointless tremors.
1516	 */
1517	shift = dirty_ratelimit / (2 * step + 1);
1518	if (shift < BITS_PER_LONG)
1519		step = DIV_ROUND_UP(step >> shift, 8);
1520	else
1521		step = 0;
1522
1523	if (dirty_ratelimit < balanced_dirty_ratelimit)
1524		dirty_ratelimit += step;
1525	else
1526		dirty_ratelimit -= step;
1527
1528	WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
1529	wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1530
1531	trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1532}
1533
1534static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1535				  struct dirty_throttle_control *mdtc,
 
1536				  bool update_ratelimit)
1537{
1538	struct bdi_writeback *wb = gdtc->wb;
1539	unsigned long now = jiffies;
1540	unsigned long elapsed;
1541	unsigned long dirtied;
1542	unsigned long written;
1543
1544	spin_lock(&wb->list_lock);
1545
1546	/*
1547	 * Lockless checks for elapsed time are racy and delayed update after
1548	 * IO completion doesn't do it at all (to make sure written pages are
1549	 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
1550	 * division errors.
1551	 */
1552	elapsed = max(now - wb->bw_time_stamp, 1UL);
 
 
1553	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1554	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1555
 
 
 
 
 
 
 
1556	if (update_ratelimit) {
1557		domain_update_dirty_limit(gdtc, now);
1558		wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1559
1560		/*
1561		 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1562		 * compiler has no way to figure that out.  Help it.
1563		 */
1564		if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1565			domain_update_dirty_limit(mdtc, now);
1566			wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1567		}
1568	}
1569	wb_update_write_bandwidth(wb, elapsed, written);
1570
 
1571	wb->dirtied_stamp = dirtied;
1572	wb->written_stamp = written;
1573	WRITE_ONCE(wb->bw_time_stamp, now);
1574	spin_unlock(&wb->list_lock);
1575}
1576
1577void wb_update_bandwidth(struct bdi_writeback *wb)
1578{
1579	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1580
1581	__wb_update_bandwidth(&gdtc, NULL, false);
1582}
1583
1584/* Interval after which we consider wb idle and don't estimate bandwidth */
1585#define WB_BANDWIDTH_IDLE_JIF (HZ)
1586
1587static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
1588{
1589	unsigned long now = jiffies;
1590	unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
1591
1592	if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
1593	    !atomic_read(&wb->writeback_inodes)) {
1594		spin_lock(&wb->list_lock);
1595		wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
1596		wb->written_stamp = wb_stat(wb, WB_WRITTEN);
1597		WRITE_ONCE(wb->bw_time_stamp, now);
1598		spin_unlock(&wb->list_lock);
1599	}
1600}
1601
1602/*
1603 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1604 * will look to see if it needs to start dirty throttling.
1605 *
1606 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1607 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1608 * (the number of pages we may dirty without exceeding the dirty limits).
1609 */
1610static unsigned long dirty_poll_interval(unsigned long dirty,
1611					 unsigned long thresh)
1612{
1613	if (thresh > dirty)
1614		return 1UL << (ilog2(thresh - dirty) >> 1);
1615
1616	return 1;
1617}
1618
1619static unsigned long wb_max_pause(struct bdi_writeback *wb,
1620				  unsigned long wb_dirty)
1621{
1622	unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
1623	unsigned long t;
1624
1625	/*
1626	 * Limit pause time for small memory systems. If sleeping for too long
1627	 * time, a small pool of dirty/writeback pages may go empty and disk go
1628	 * idle.
1629	 *
1630	 * 8 serves as the safety ratio.
1631	 */
1632	t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1633	t++;
1634
1635	return min_t(unsigned long, t, MAX_PAUSE);
1636}
1637
1638static long wb_min_pause(struct bdi_writeback *wb,
1639			 long max_pause,
1640			 unsigned long task_ratelimit,
1641			 unsigned long dirty_ratelimit,
1642			 int *nr_dirtied_pause)
1643{
1644	long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
1645	long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
1646	long t;		/* target pause */
1647	long pause;	/* estimated next pause */
1648	int pages;	/* target nr_dirtied_pause */
1649
1650	/* target for 10ms pause on 1-dd case */
1651	t = max(1, HZ / 100);
1652
1653	/*
1654	 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1655	 * overheads.
1656	 *
1657	 * (N * 10ms) on 2^N concurrent tasks.
1658	 */
1659	if (hi > lo)
1660		t += (hi - lo) * (10 * HZ) / 1024;
1661
1662	/*
1663	 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1664	 * on the much more stable dirty_ratelimit. However the next pause time
1665	 * will be computed based on task_ratelimit and the two rate limits may
1666	 * depart considerably at some time. Especially if task_ratelimit goes
1667	 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1668	 * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1669	 * result task_ratelimit won't be executed faithfully, which could
1670	 * eventually bring down dirty_ratelimit.
1671	 *
1672	 * We apply two rules to fix it up:
1673	 * 1) try to estimate the next pause time and if necessary, use a lower
1674	 *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1675	 *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1676	 * 2) limit the target pause time to max_pause/2, so that the normal
1677	 *    small fluctuations of task_ratelimit won't trigger rule (1) and
1678	 *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1679	 */
1680	t = min(t, 1 + max_pause / 2);
1681	pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1682
1683	/*
1684	 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1685	 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1686	 * When the 16 consecutive reads are often interrupted by some dirty
1687	 * throttling pause during the async writes, cfq will go into idles
1688	 * (deadline is fine). So push nr_dirtied_pause as high as possible
1689	 * until reaches DIRTY_POLL_THRESH=32 pages.
1690	 */
1691	if (pages < DIRTY_POLL_THRESH) {
1692		t = max_pause;
1693		pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1694		if (pages > DIRTY_POLL_THRESH) {
1695			pages = DIRTY_POLL_THRESH;
1696			t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1697		}
1698	}
1699
1700	pause = HZ * pages / (task_ratelimit + 1);
1701	if (pause > max_pause) {
1702		t = max_pause;
1703		pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1704	}
1705
1706	*nr_dirtied_pause = pages;
1707	/*
1708	 * The minimal pause time will normally be half the target pause time.
1709	 */
1710	return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1711}
1712
1713static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1714{
1715	struct bdi_writeback *wb = dtc->wb;
1716	unsigned long wb_reclaimable;
1717
1718	/*
1719	 * wb_thresh is not treated as some limiting factor as
1720	 * dirty_thresh, due to reasons
1721	 * - in JBOD setup, wb_thresh can fluctuate a lot
1722	 * - in a system with HDD and USB key, the USB key may somehow
1723	 *   go into state (wb_dirty >> wb_thresh) either because
1724	 *   wb_dirty starts high, or because wb_thresh drops low.
1725	 *   In this case we don't want to hard throttle the USB key
1726	 *   dirtiers for 100 seconds until wb_dirty drops under
1727	 *   wb_thresh. Instead the auxiliary wb control line in
1728	 *   wb_position_ratio() will let the dirtier task progress
1729	 *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1730	 */
1731	dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh);
1732	dtc->wb_bg_thresh = dtc->thresh ?
1733		div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1734
1735	/*
1736	 * In order to avoid the stacked BDI deadlock we need
1737	 * to ensure we accurately count the 'dirty' pages when
1738	 * the threshold is low.
1739	 *
1740	 * Otherwise it would be possible to get thresh+n pages
1741	 * reported dirty, even though there are thresh-m pages
1742	 * actually dirty; with m+n sitting in the percpu
1743	 * deltas.
1744	 */
1745	if (dtc->wb_thresh < 2 * wb_stat_error()) {
1746		wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1747		dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1748	} else {
1749		wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1750		dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1751	}
1752}
1753
1754static unsigned long domain_poll_intv(struct dirty_throttle_control *dtc,
1755				      bool strictlimit)
1756{
1757	unsigned long dirty, thresh;
1758
1759	if (strictlimit) {
1760		dirty = dtc->wb_dirty;
1761		thresh = dtc->wb_thresh;
1762	} else {
1763		dirty = dtc->dirty;
1764		thresh = dtc->thresh;
1765	}
1766
1767	return dirty_poll_interval(dirty, thresh);
1768}
1769
1770/*
1771 * Throttle it only when the background writeback cannot catch-up. This avoids
1772 * (excessively) small writeouts when the wb limits are ramping up in case of
1773 * !strictlimit.
1774 *
1775 * In strictlimit case make decision based on the wb counters and limits. Small
1776 * writeouts when the wb limits are ramping up are the price we consciously pay
1777 * for strictlimit-ing.
1778 */
1779static void domain_dirty_freerun(struct dirty_throttle_control *dtc,
1780				 bool strictlimit)
1781{
1782	unsigned long dirty, thresh, bg_thresh;
1783
1784	if (unlikely(strictlimit)) {
1785		wb_dirty_limits(dtc);
1786		dirty = dtc->wb_dirty;
1787		thresh = dtc->wb_thresh;
1788		bg_thresh = dtc->wb_bg_thresh;
1789	} else {
1790		dirty = dtc->dirty;
1791		thresh = dtc->thresh;
1792		bg_thresh = dtc->bg_thresh;
1793	}
1794	dtc->freerun = dirty <= dirty_freerun_ceiling(thresh, bg_thresh);
1795}
1796
1797static void balance_domain_limits(struct dirty_throttle_control *dtc,
1798				  bool strictlimit)
1799{
1800	domain_dirty_avail(dtc, true);
1801	domain_dirty_limits(dtc);
1802	domain_dirty_freerun(dtc, strictlimit);
1803}
1804
1805static void wb_dirty_freerun(struct dirty_throttle_control *dtc,
1806			     bool strictlimit)
1807{
1808	dtc->freerun = false;
1809
1810	/* was already handled in domain_dirty_freerun */
1811	if (strictlimit)
1812		return;
1813
1814	wb_dirty_limits(dtc);
1815	/*
1816	 * LOCAL_THROTTLE tasks must not be throttled when below the per-wb
1817	 * freerun ceiling.
1818	 */
1819	if (!(current->flags & PF_LOCAL_THROTTLE))
1820		return;
1821
1822	dtc->freerun = dtc->wb_dirty <
1823		       dirty_freerun_ceiling(dtc->wb_thresh, dtc->wb_bg_thresh);
1824}
1825
1826static inline void wb_dirty_exceeded(struct dirty_throttle_control *dtc,
1827				     bool strictlimit)
1828{
1829	dtc->dirty_exceeded = (dtc->wb_dirty > dtc->wb_thresh) &&
1830		((dtc->dirty > dtc->thresh) || strictlimit);
1831}
1832
1833/*
1834 * The limits fields dirty_exceeded and pos_ratio won't be updated if wb is
1835 * in freerun state. Please don't use these invalid fields in freerun case.
1836 */
1837static void balance_wb_limits(struct dirty_throttle_control *dtc,
1838			      bool strictlimit)
1839{
1840	wb_dirty_freerun(dtc, strictlimit);
1841	if (dtc->freerun)
1842		return;
1843
1844	wb_dirty_exceeded(dtc, strictlimit);
1845	wb_position_ratio(dtc);
1846}
1847
1848/*
1849 * balance_dirty_pages() must be called by processes which are generating dirty
1850 * data.  It looks at the number of dirty pages in the machine and will force
1851 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1852 * If we're over `background_thresh' then the writeback threads are woken to
1853 * perform some writeout.
1854 */
1855static int balance_dirty_pages(struct bdi_writeback *wb,
1856			       unsigned long pages_dirtied, unsigned int flags)
1857{
1858	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1859	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1860	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1861	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1862						     &mdtc_stor : NULL;
1863	struct dirty_throttle_control *sdtc;
1864	unsigned long nr_dirty;
1865	long period;
1866	long pause;
1867	long max_pause;
1868	long min_pause;
1869	int nr_dirtied_pause;
 
1870	unsigned long task_ratelimit;
1871	unsigned long dirty_ratelimit;
1872	struct backing_dev_info *bdi = wb->bdi;
1873	bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1874	unsigned long start_time = jiffies;
1875	int ret = 0;
1876
1877	for (;;) {
1878		unsigned long now = jiffies;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1879
1880		nr_dirty = global_node_page_state(NR_FILE_DIRTY);
 
1881
1882		balance_domain_limits(gdtc, strictlimit);
1883		if (mdtc) {
1884			/*
1885			 * If @wb belongs to !root memcg, repeat the same
1886			 * basic calculations for the memcg domain.
1887			 */
1888			balance_domain_limits(mdtc, strictlimit);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1889		}
1890
1891		/*
1892		 * In laptop mode, we wait until hitting the higher threshold
1893		 * before starting background writeout, and then write out all
1894		 * the way down to the lower threshold.  So slow writers cause
1895		 * minimal disk activity.
 
 
 
1896		 *
1897		 * In normal mode, we start background writeout at the lower
1898		 * background_thresh, to keep the amount of dirty memory low.
1899		 */
1900		if (!laptop_mode && nr_dirty > gdtc->bg_thresh &&
1901		    !writeback_in_progress(wb))
1902			wb_start_background_writeback(wb);
1903
1904		/*
1905		 * If memcg domain is in effect, @dirty should be under
1906		 * both global and memcg freerun ceilings.
1907		 */
1908		if (gdtc->freerun && (!mdtc || mdtc->freerun)) {
1909			unsigned long intv;
1910			unsigned long m_intv;
1911
1912free_running:
1913			intv = domain_poll_intv(gdtc, strictlimit);
1914			m_intv = ULONG_MAX;
1915
1916			current->dirty_paused_when = now;
1917			current->nr_dirtied = 0;
1918			if (mdtc)
1919				m_intv = domain_poll_intv(mdtc, strictlimit);
1920			current->nr_dirtied_pause = min(intv, m_intv);
1921			break;
1922		}
1923
1924		/* Start writeback even when in laptop mode */
1925		if (unlikely(!writeback_in_progress(wb)))
1926			wb_start_background_writeback(wb);
1927
1928		mem_cgroup_flush_foreign(wb);
1929
1930		/*
1931		 * Calculate global domain's pos_ratio and select the
1932		 * global dtc by default.
1933		 */
1934		balance_wb_limits(gdtc, strictlimit);
1935		if (gdtc->freerun)
1936			goto free_running;
 
 
 
 
1937		sdtc = gdtc;
1938
1939		if (mdtc) {
1940			/*
1941			 * If memcg domain is in effect, calculate its
1942			 * pos_ratio.  @wb should satisfy constraints from
1943			 * both global and memcg domains.  Choose the one
1944			 * w/ lower pos_ratio.
1945			 */
1946			balance_wb_limits(mdtc, strictlimit);
1947			if (mdtc->freerun)
1948				goto free_running;
 
 
 
 
1949			if (mdtc->pos_ratio < gdtc->pos_ratio)
1950				sdtc = mdtc;
1951		}
1952
1953		wb->dirty_exceeded = gdtc->dirty_exceeded ||
1954				     (mdtc && mdtc->dirty_exceeded);
1955		if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
1956					   BANDWIDTH_INTERVAL))
1957			__wb_update_bandwidth(gdtc, mdtc, true);
 
 
 
 
1958
1959		/* throttle according to the chosen dtc */
1960		dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
1961		task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1962							RATELIMIT_CALC_SHIFT;
1963		max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1964		min_pause = wb_min_pause(wb, max_pause,
1965					 task_ratelimit, dirty_ratelimit,
1966					 &nr_dirtied_pause);
1967
1968		if (unlikely(task_ratelimit == 0)) {
1969			period = max_pause;
1970			pause = max_pause;
1971			goto pause;
1972		}
1973		period = HZ * pages_dirtied / task_ratelimit;
1974		pause = period;
1975		if (current->dirty_paused_when)
1976			pause -= now - current->dirty_paused_when;
1977		/*
1978		 * For less than 1s think time (ext3/4 may block the dirtier
1979		 * for up to 800ms from time to time on 1-HDD; so does xfs,
1980		 * however at much less frequency), try to compensate it in
1981		 * future periods by updating the virtual time; otherwise just
1982		 * do a reset, as it may be a light dirtier.
1983		 */
1984		if (pause < min_pause) {
1985			trace_balance_dirty_pages(wb,
1986						  sdtc->thresh,
1987						  sdtc->bg_thresh,
1988						  sdtc->dirty,
1989						  sdtc->wb_thresh,
1990						  sdtc->wb_dirty,
1991						  dirty_ratelimit,
1992						  task_ratelimit,
1993						  pages_dirtied,
1994						  period,
1995						  min(pause, 0L),
1996						  start_time);
1997			if (pause < -HZ) {
1998				current->dirty_paused_when = now;
1999				current->nr_dirtied = 0;
2000			} else if (period) {
2001				current->dirty_paused_when += period;
2002				current->nr_dirtied = 0;
2003			} else if (current->nr_dirtied_pause <= pages_dirtied)
2004				current->nr_dirtied_pause += pages_dirtied;
2005			break;
2006		}
2007		if (unlikely(pause > max_pause)) {
2008			/* for occasional dropped task_ratelimit */
2009			now += min(pause - max_pause, max_pause);
2010			pause = max_pause;
2011		}
2012
2013pause:
2014		trace_balance_dirty_pages(wb,
2015					  sdtc->thresh,
2016					  sdtc->bg_thresh,
2017					  sdtc->dirty,
2018					  sdtc->wb_thresh,
2019					  sdtc->wb_dirty,
2020					  dirty_ratelimit,
2021					  task_ratelimit,
2022					  pages_dirtied,
2023					  period,
2024					  pause,
2025					  start_time);
2026		if (flags & BDP_ASYNC) {
2027			ret = -EAGAIN;
2028			break;
2029		}
2030		__set_current_state(TASK_KILLABLE);
2031		bdi->last_bdp_sleep = jiffies;
2032		io_schedule_timeout(pause);
2033
2034		current->dirty_paused_when = now + pause;
2035		current->nr_dirtied = 0;
2036		current->nr_dirtied_pause = nr_dirtied_pause;
2037
2038		/*
2039		 * This is typically equal to (dirty < thresh) and can also
2040		 * keep "1000+ dd on a slow USB stick" under control.
2041		 */
2042		if (task_ratelimit)
2043			break;
2044
2045		/*
2046		 * In the case of an unresponsive NFS server and the NFS dirty
2047		 * pages exceeds dirty_thresh, give the other good wb's a pipe
2048		 * to go through, so that tasks on them still remain responsive.
2049		 *
2050		 * In theory 1 page is enough to keep the consumer-producer
2051		 * pipe going: the flusher cleans 1 page => the task dirties 1
2052		 * more page. However wb_dirty has accounting errors.  So use
2053		 * the larger and more IO friendly wb_stat_error.
2054		 */
2055		if (sdtc->wb_dirty <= wb_stat_error())
2056			break;
2057
2058		if (fatal_signal_pending(current))
2059			break;
2060	}
2061	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2062}
2063
2064static DEFINE_PER_CPU(int, bdp_ratelimits);
2065
2066/*
2067 * Normal tasks are throttled by
2068 *	loop {
2069 *		dirty tsk->nr_dirtied_pause pages;
2070 *		take a snap in balance_dirty_pages();
2071 *	}
2072 * However there is a worst case. If every task exit immediately when dirtied
2073 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
2074 * called to throttle the page dirties. The solution is to save the not yet
2075 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
2076 * randomly into the running tasks. This works well for the above worst case,
2077 * as the new task will pick up and accumulate the old task's leaked dirty
2078 * count and eventually get throttled.
2079 */
2080DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
2081
2082/**
2083 * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
2084 * @mapping: address_space which was dirtied.
2085 * @flags: BDP flags.
2086 *
2087 * Processes which are dirtying memory should call in here once for each page
2088 * which was newly dirtied.  The function will periodically check the system's
2089 * dirty state and will initiate writeback if needed.
2090 *
2091 * See balance_dirty_pages_ratelimited() for details.
2092 *
2093 * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
2094 * indicate that memory is out of balance and the caller must wait
2095 * for I/O to complete.  Otherwise, it will return 0 to indicate
2096 * that either memory was already in balance, or it was able to sleep
2097 * until the amount of dirty memory returned to balance.
2098 */
2099int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
2100					unsigned int flags)
2101{
2102	struct inode *inode = mapping->host;
2103	struct backing_dev_info *bdi = inode_to_bdi(inode);
2104	struct bdi_writeback *wb = NULL;
2105	int ratelimit;
2106	int ret = 0;
2107	int *p;
2108
2109	if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
2110		return ret;
2111
2112	if (inode_cgwb_enabled(inode))
2113		wb = wb_get_create_current(bdi, GFP_KERNEL);
2114	if (!wb)
2115		wb = &bdi->wb;
2116
2117	ratelimit = current->nr_dirtied_pause;
2118	if (wb->dirty_exceeded)
2119		ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
2120
2121	preempt_disable();
2122	/*
2123	 * This prevents one CPU to accumulate too many dirtied pages without
2124	 * calling into balance_dirty_pages(), which can happen when there are
2125	 * 1000+ tasks, all of them start dirtying pages at exactly the same
2126	 * time, hence all honoured too large initial task->nr_dirtied_pause.
2127	 */
2128	p =  this_cpu_ptr(&bdp_ratelimits);
2129	if (unlikely(current->nr_dirtied >= ratelimit))
2130		*p = 0;
2131	else if (unlikely(*p >= ratelimit_pages)) {
2132		*p = 0;
2133		ratelimit = 0;
2134	}
2135	/*
2136	 * Pick up the dirtied pages by the exited tasks. This avoids lots of
2137	 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
2138	 * the dirty throttling and livelock other long-run dirtiers.
2139	 */
2140	p = this_cpu_ptr(&dirty_throttle_leaks);
2141	if (*p > 0 && current->nr_dirtied < ratelimit) {
2142		unsigned long nr_pages_dirtied;
2143		nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
2144		*p -= nr_pages_dirtied;
2145		current->nr_dirtied += nr_pages_dirtied;
2146	}
2147	preempt_enable();
2148
2149	if (unlikely(current->nr_dirtied >= ratelimit))
2150		ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
2151
2152	wb_put(wb);
2153	return ret;
2154}
2155EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
2156
2157/**
2158 * balance_dirty_pages_ratelimited - balance dirty memory state.
2159 * @mapping: address_space which was dirtied.
2160 *
2161 * Processes which are dirtying memory should call in here once for each page
2162 * which was newly dirtied.  The function will periodically check the system's
2163 * dirty state and will initiate writeback if needed.
2164 *
2165 * Once we're over the dirty memory limit we decrease the ratelimiting
2166 * by a lot, to prevent individual processes from overshooting the limit
2167 * by (ratelimit_pages) each.
2168 */
2169void balance_dirty_pages_ratelimited(struct address_space *mapping)
2170{
2171	balance_dirty_pages_ratelimited_flags(mapping, 0);
2172}
2173EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
 
 
2174
2175/*
2176 * Similar to wb_dirty_limits, wb_bg_dirty_limits also calculates dirty
2177 * and thresh, but it's for background writeback.
2178 */
2179static void wb_bg_dirty_limits(struct dirty_throttle_control *dtc)
2180{
2181	struct bdi_writeback *wb = dtc->wb;
 
2182
2183	dtc->wb_bg_thresh = __wb_calc_thresh(dtc, dtc->bg_thresh);
2184	if (dtc->wb_bg_thresh < 2 * wb_stat_error())
2185		dtc->wb_dirty = wb_stat_sum(wb, WB_RECLAIMABLE);
2186	else
2187		dtc->wb_dirty = wb_stat(wb, WB_RECLAIMABLE);
2188}
2189
2190static bool domain_over_bg_thresh(struct dirty_throttle_control *dtc)
2191{
2192	domain_dirty_avail(dtc, false);
2193	domain_dirty_limits(dtc);
2194	if (dtc->dirty > dtc->bg_thresh)
2195		return true;
2196
2197	wb_bg_dirty_limits(dtc);
2198	if (dtc->wb_dirty > dtc->wb_bg_thresh)
2199		return true;
2200
2201	return false;
2202}
2203
2204/**
2205 * wb_over_bg_thresh - does @wb need to be written back?
2206 * @wb: bdi_writeback of interest
2207 *
2208 * Determines whether background writeback should keep writing @wb or it's
2209 * clean enough.
2210 *
2211 * Return: %true if writeback should continue.
2212 */
2213bool wb_over_bg_thresh(struct bdi_writeback *wb)
2214{
2215	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
2216	struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
2217
2218	if (domain_over_bg_thresh(&gdtc))
2219		return true;
2220
2221	if (mdtc_valid(&mdtc))
2222		return domain_over_bg_thresh(&mdtc);
 
 
2223
2224	return false;
2225}
2226
2227#ifdef CONFIG_SYSCTL
2228/*
2229 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
2230 */
2231static int dirty_writeback_centisecs_handler(const struct ctl_table *table, int write,
2232		void *buffer, size_t *length, loff_t *ppos)
2233{
2234	unsigned int old_interval = dirty_writeback_interval;
2235	int ret;
2236
2237	ret = proc_dointvec(table, write, buffer, length, ppos);
2238
2239	/*
2240	 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2241	 * and a different non-zero value will wakeup the writeback threads.
2242	 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2243	 * iterate over all bdis and wbs.
2244	 * The reason we do this is to make the change take effect immediately.
2245	 */
2246	if (!ret && write && dirty_writeback_interval &&
2247		dirty_writeback_interval != old_interval)
2248		wakeup_flusher_threads(WB_REASON_PERIODIC);
2249
2250	return ret;
2251}
2252#endif
2253
 
2254void laptop_mode_timer_fn(struct timer_list *t)
2255{
2256	struct backing_dev_info *backing_dev_info =
2257		from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2258
2259	wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2260}
2261
2262/*
2263 * We've spun up the disk and we're in laptop mode: schedule writeback
2264 * of all dirty data a few seconds from now.  If the flush is already scheduled
2265 * then push it back - the user is still using the disk.
2266 */
2267void laptop_io_completion(struct backing_dev_info *info)
2268{
2269	mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2270}
2271
2272/*
2273 * We're in laptop mode and we've just synced. The sync's writes will have
2274 * caused another writeback to be scheduled by laptop_io_completion.
2275 * Nothing needs to be written back anymore, so we unschedule the writeback.
2276 */
2277void laptop_sync_completion(void)
2278{
2279	struct backing_dev_info *bdi;
2280
2281	rcu_read_lock();
2282
2283	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2284		del_timer(&bdi->laptop_mode_wb_timer);
2285
2286	rcu_read_unlock();
2287}
 
2288
2289/*
2290 * If ratelimit_pages is too high then we can get into dirty-data overload
2291 * if a large number of processes all perform writes at the same time.
 
 
2292 *
2293 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2294 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2295 * thresholds.
2296 */
2297
2298void writeback_set_ratelimit(void)
2299{
2300	struct wb_domain *dom = &global_wb_domain;
2301	unsigned long background_thresh;
2302	unsigned long dirty_thresh;
2303
2304	global_dirty_limits(&background_thresh, &dirty_thresh);
2305	dom->dirty_limit = dirty_thresh;
2306	ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2307	if (ratelimit_pages < 16)
2308		ratelimit_pages = 16;
2309}
2310
2311static int page_writeback_cpu_online(unsigned int cpu)
2312{
2313	writeback_set_ratelimit();
2314	return 0;
2315}
2316
2317#ifdef CONFIG_SYSCTL
2318
2319/* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
2320static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
2321
2322static struct ctl_table vm_page_writeback_sysctls[] = {
2323	{
2324		.procname   = "dirty_background_ratio",
2325		.data       = &dirty_background_ratio,
2326		.maxlen     = sizeof(dirty_background_ratio),
2327		.mode       = 0644,
2328		.proc_handler   = dirty_background_ratio_handler,
2329		.extra1     = SYSCTL_ZERO,
2330		.extra2     = SYSCTL_ONE_HUNDRED,
2331	},
2332	{
2333		.procname   = "dirty_background_bytes",
2334		.data       = &dirty_background_bytes,
2335		.maxlen     = sizeof(dirty_background_bytes),
2336		.mode       = 0644,
2337		.proc_handler   = dirty_background_bytes_handler,
2338		.extra1     = SYSCTL_LONG_ONE,
2339	},
2340	{
2341		.procname   = "dirty_ratio",
2342		.data       = &vm_dirty_ratio,
2343		.maxlen     = sizeof(vm_dirty_ratio),
2344		.mode       = 0644,
2345		.proc_handler   = dirty_ratio_handler,
2346		.extra1     = SYSCTL_ZERO,
2347		.extra2     = SYSCTL_ONE_HUNDRED,
2348	},
2349	{
2350		.procname   = "dirty_bytes",
2351		.data       = &vm_dirty_bytes,
2352		.maxlen     = sizeof(vm_dirty_bytes),
2353		.mode       = 0644,
2354		.proc_handler   = dirty_bytes_handler,
2355		.extra1     = (void *)&dirty_bytes_min,
2356	},
2357	{
2358		.procname   = "dirty_writeback_centisecs",
2359		.data       = &dirty_writeback_interval,
2360		.maxlen     = sizeof(dirty_writeback_interval),
2361		.mode       = 0644,
2362		.proc_handler   = dirty_writeback_centisecs_handler,
2363	},
2364	{
2365		.procname   = "dirty_expire_centisecs",
2366		.data       = &dirty_expire_interval,
2367		.maxlen     = sizeof(dirty_expire_interval),
2368		.mode       = 0644,
2369		.proc_handler   = proc_dointvec_minmax,
2370		.extra1     = SYSCTL_ZERO,
2371	},
2372#ifdef CONFIG_HIGHMEM
2373	{
2374		.procname	= "highmem_is_dirtyable",
2375		.data		= &vm_highmem_is_dirtyable,
2376		.maxlen		= sizeof(vm_highmem_is_dirtyable),
2377		.mode		= 0644,
2378		.proc_handler	= proc_dointvec_minmax,
2379		.extra1		= SYSCTL_ZERO,
2380		.extra2		= SYSCTL_ONE,
2381	},
2382#endif
2383	{
2384		.procname	= "laptop_mode",
2385		.data		= &laptop_mode,
2386		.maxlen		= sizeof(laptop_mode),
2387		.mode		= 0644,
2388		.proc_handler	= proc_dointvec_jiffies,
2389	},
2390};
2391#endif
2392
2393/*
2394 * Called early on to tune the page writeback dirty limits.
2395 *
2396 * We used to scale dirty pages according to how total memory
2397 * related to pages that could be allocated for buffers.
 
2398 *
2399 * However, that was when we used "dirty_ratio" to scale with
2400 * all memory, and we don't do that any more. "dirty_ratio"
2401 * is now applied to total non-HIGHPAGE memory, and as such we can't
 
2402 * get into the old insane situation any more where we had
2403 * large amounts of dirty pages compared to a small amount of
2404 * non-HIGHMEM memory.
2405 *
2406 * But we might still want to scale the dirty_ratio by how
2407 * much memory the box has..
2408 */
2409void __init page_writeback_init(void)
2410{
2411	BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2412
2413	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2414			  page_writeback_cpu_online, NULL);
2415	cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2416			  page_writeback_cpu_online);
2417#ifdef CONFIG_SYSCTL
2418	register_sysctl_init("vm", vm_page_writeback_sysctls);
2419#endif
2420}
2421
2422/**
2423 * tag_pages_for_writeback - tag pages to be written by writeback
2424 * @mapping: address space structure to write
2425 * @start: starting page index
2426 * @end: ending page index (inclusive)
2427 *
2428 * This function scans the page range from @start to @end (inclusive) and tags
2429 * all pages that have DIRTY tag set with a special TOWRITE tag.  The caller
2430 * can then use the TOWRITE tag to identify pages eligible for writeback.
2431 * This mechanism is used to avoid livelocking of writeback by a process
2432 * steadily creating new dirty pages in the file (thus it is important for this
2433 * function to be quick so that it can tag pages faster than a dirtying process
2434 * can create them).
 
 
 
 
2435 */
2436void tag_pages_for_writeback(struct address_space *mapping,
2437			     pgoff_t start, pgoff_t end)
2438{
2439	XA_STATE(xas, &mapping->i_pages, start);
2440	unsigned int tagged = 0;
2441	void *page;
2442
2443	xas_lock_irq(&xas);
2444	xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2445		xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2446		if (++tagged % XA_CHECK_SCHED)
 
 
 
 
 
 
2447			continue;
2448
2449		xas_pause(&xas);
2450		xas_unlock_irq(&xas);
2451		cond_resched();
2452		xas_lock_irq(&xas);
2453	}
2454	xas_unlock_irq(&xas);
2455}
2456EXPORT_SYMBOL(tag_pages_for_writeback);
2457
2458static bool folio_prepare_writeback(struct address_space *mapping,
2459		struct writeback_control *wbc, struct folio *folio)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2460{
2461	/*
2462	 * Folio truncated or invalidated. We can freely skip it then,
2463	 * even for data integrity operations: the folio has disappeared
2464	 * concurrently, so there could be no real expectation of this
2465	 * data integrity operation even if there is now a new, dirty
2466	 * folio at the same pagecache index.
2467	 */
2468	if (unlikely(folio->mapping != mapping))
2469		return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2470
2471	/*
2472	 * Did somebody else write it for us?
2473	 */
2474	if (!folio_test_dirty(folio))
2475		return false;
2476
2477	if (folio_test_writeback(folio)) {
2478		if (wbc->sync_mode == WB_SYNC_NONE)
2479			return false;
2480		folio_wait_writeback(folio);
2481	}
2482	BUG_ON(folio_test_writeback(folio));
2483
2484	if (!folio_clear_dirty_for_io(folio))
2485		return false;
2486
2487	return true;
2488}
 
 
 
 
 
 
 
 
 
 
 
2489
2490static xa_mark_t wbc_to_tag(struct writeback_control *wbc)
2491{
2492	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2493		return PAGECACHE_TAG_TOWRITE;
2494	return PAGECACHE_TAG_DIRTY;
2495}
2496
2497static pgoff_t wbc_end(struct writeback_control *wbc)
2498{
2499	if (wbc->range_cyclic)
2500		return -1;
2501	return wbc->range_end >> PAGE_SHIFT;
2502}
2503
2504static struct folio *writeback_get_folio(struct address_space *mapping,
2505		struct writeback_control *wbc)
2506{
2507	struct folio *folio;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2508
2509retry:
2510	folio = folio_batch_next(&wbc->fbatch);
2511	if (!folio) {
2512		folio_batch_release(&wbc->fbatch);
 
 
 
 
 
 
 
 
 
2513		cond_resched();
2514		filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc),
2515				wbc_to_tag(wbc), &wbc->fbatch);
2516		folio = folio_batch_next(&wbc->fbatch);
2517		if (!folio)
2518			return NULL;
2519	}
2520
2521	folio_lock(folio);
2522	if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) {
2523		folio_unlock(folio);
 
 
 
 
 
2524		goto retry;
2525	}
 
 
2526
2527	trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2528	return folio;
2529}
 
2530
2531/**
2532 * writeback_iter - iterate folio of a mapping for writeback
2533 * @mapping: address space structure to write
2534 * @wbc: writeback context
2535 * @folio: previously iterated folio (%NULL to start)
2536 * @error: in-out pointer for writeback errors (see below)
2537 *
2538 * This function returns the next folio for the writeback operation described by
2539 * @wbc on @mapping and  should be called in a while loop in the ->writepages
2540 * implementation.
2541 *
2542 * To start the writeback operation, %NULL is passed in the @folio argument, and
2543 * for every subsequent iteration the folio returned previously should be passed
2544 * back in.
2545 *
2546 * If there was an error in the per-folio writeback inside the writeback_iter()
2547 * loop, @error should be set to the error value.
2548 *
2549 * Once the writeback described in @wbc has finished, this function will return
2550 * %NULL and if there was an error in any iteration restore it to @error.
2551 *
2552 * Note: callers should not manually break out of the loop using break or goto
2553 * but must keep calling writeback_iter() until it returns %NULL.
2554 *
2555 * Return: the folio to write or %NULL if the loop is done.
2556 */
2557struct folio *writeback_iter(struct address_space *mapping,
2558		struct writeback_control *wbc, struct folio *folio, int *error)
2559{
2560	if (!folio) {
2561		folio_batch_init(&wbc->fbatch);
2562		wbc->saved_err = *error = 0;
2563
2564		/*
2565		 * For range cyclic writeback we remember where we stopped so
2566		 * that we can continue where we stopped.
2567		 *
2568		 * For non-cyclic writeback we always start at the beginning of
2569		 * the passed in range.
2570		 */
2571		if (wbc->range_cyclic)
2572			wbc->index = mapping->writeback_index;
2573		else
2574			wbc->index = wbc->range_start >> PAGE_SHIFT;
2575
2576		/*
2577		 * To avoid livelocks when other processes dirty new pages, we
2578		 * first tag pages which should be written back and only then
2579		 * start writing them.
2580		 *
2581		 * For data-integrity writeback we have to be careful so that we
2582		 * do not miss some pages (e.g., because some other process has
2583		 * cleared the TOWRITE tag we set).  The rule we follow is that
2584		 * TOWRITE tag can be cleared only by the process clearing the
2585		 * DIRTY tag (and submitting the page for I/O).
2586		 */
2587		if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2588			tag_pages_for_writeback(mapping, wbc->index,
2589					wbc_end(wbc));
2590	} else {
2591		wbc->nr_to_write -= folio_nr_pages(folio);
2592
2593		WARN_ON_ONCE(*error > 0);
2594
2595		/*
2596		 * For integrity writeback we have to keep going until we have
2597		 * written all the folios we tagged for writeback above, even if
2598		 * we run past wbc->nr_to_write or encounter errors.
2599		 * We stash away the first error we encounter in wbc->saved_err
2600		 * so that it can be retrieved when we're done.  This is because
2601		 * the file system may still have state to clear for each folio.
2602		 *
2603		 * For background writeback we exit as soon as we run past
2604		 * wbc->nr_to_write or encounter the first error.
2605		 */
2606		if (wbc->sync_mode == WB_SYNC_ALL) {
2607			if (*error && !wbc->saved_err)
2608				wbc->saved_err = *error;
2609		} else {
2610			if (*error || wbc->nr_to_write <= 0)
2611				goto done;
2612		}
2613	}
2614
2615	folio = writeback_get_folio(mapping, wbc);
2616	if (!folio) {
2617		/*
2618		 * To avoid deadlocks between range_cyclic writeback and callers
2619		 * that hold pages in PageWriteback to aggregate I/O until
2620		 * the writeback iteration finishes, we do not loop back to the
2621		 * start of the file.  Doing so causes a page lock/page
2622		 * writeback access order inversion - we should only ever lock
2623		 * multiple pages in ascending page->index order, and looping
2624		 * back to the start of the file violates that rule and causes
2625		 * deadlocks.
2626		 */
2627		if (wbc->range_cyclic)
2628			mapping->writeback_index = 0;
2629
2630		/*
2631		 * Return the first error we encountered (if there was any) to
2632		 * the caller.
2633		 */
2634		*error = wbc->saved_err;
2635	}
2636	return folio;
2637
2638done:
2639	if (wbc->range_cyclic)
2640		mapping->writeback_index = folio_next_index(folio);
2641	folio_batch_release(&wbc->fbatch);
2642	return NULL;
2643}
2644EXPORT_SYMBOL_GPL(writeback_iter);
2645
2646/**
2647 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2648 * @mapping: address space structure to write
2649 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2650 * @writepage: function called for each page
2651 * @data: data passed to writepage function
2652 *
2653 * Return: %0 on success, negative error code otherwise
2654 *
2655 * Note: please use writeback_iter() instead.
2656 */
2657int write_cache_pages(struct address_space *mapping,
2658		      struct writeback_control *wbc, writepage_t writepage,
2659		      void *data)
2660{
2661	struct folio *folio = NULL;
2662	int error;
2663
2664	while ((folio = writeback_iter(mapping, wbc, folio, &error))) {
2665		error = writepage(folio, wbc, data);
2666		if (error == AOP_WRITEPAGE_ACTIVATE) {
2667			folio_unlock(folio);
2668			error = 0;
2669		}
2670	}
2671
2672	return error;
2673}
2674EXPORT_SYMBOL(write_cache_pages);
2675
2676static int writeback_use_writepage(struct address_space *mapping,
2677		struct writeback_control *wbc)
2678{
2679	struct folio *folio = NULL;
2680	struct blk_plug plug;
2681	int err;
2682
2683	blk_start_plug(&plug);
2684	while ((folio = writeback_iter(mapping, wbc, folio, &err))) {
2685		err = mapping->a_ops->writepage(&folio->page, wbc);
2686		if (err == AOP_WRITEPAGE_ACTIVATE) {
2687			folio_unlock(folio);
2688			err = 0;
2689		}
2690		mapping_set_error(mapping, err);
2691	}
2692	blk_finish_plug(&plug);
 
 
2693
2694	return err;
2695}
2696
2697int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2698{
2699	int ret;
2700	struct bdi_writeback *wb;
2701
2702	if (wbc->nr_to_write <= 0)
2703		return 0;
2704	wb = inode_to_wb_wbc(mapping->host, wbc);
2705	wb_bandwidth_estimate_start(wb);
2706	while (1) {
2707		if (mapping->a_ops->writepages) {
2708			ret = mapping->a_ops->writepages(mapping, wbc);
2709		} else if (mapping->a_ops->writepage) {
2710			ret = writeback_use_writepage(mapping, wbc);
2711		} else {
2712			/* deal with chardevs and other special files */
2713			ret = 0;
2714		}
2715		if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL)
2716			break;
 
 
 
 
 
2717
2718		/*
2719		 * Lacking an allocation context or the locality or writeback
2720		 * state of any of the inode's pages, throttle based on
2721		 * writeback activity on the local node. It's as good a
2722		 * guess as any.
2723		 */
2724		reclaim_throttle(NODE_DATA(numa_node_id()),
2725			VMSCAN_THROTTLE_WRITEBACK);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2726	}
2727	/*
2728	 * Usually few pages are written by now from those we've just submitted
2729	 * but if there's constant writeback being submitted, this makes sure
2730	 * writeback bandwidth is updated once in a while.
2731	 */
2732	if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
2733				   BANDWIDTH_INTERVAL))
2734		wb_update_bandwidth(wb);
2735	return ret;
2736}
 
2737
2738/*
2739 * For address_spaces which do not use buffers nor write back.
2740 */
2741bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
2742{
2743	if (!folio_test_dirty(folio))
2744		return !folio_test_set_dirty(folio);
2745	return false;
2746}
2747EXPORT_SYMBOL(noop_dirty_folio);
2748
2749/*
2750 * Helper function for set_page_dirty family.
2751 *
 
 
2752 * NOTE: This relies on being atomic wrt interrupts.
2753 */
2754static void folio_account_dirtied(struct folio *folio,
2755		struct address_space *mapping)
2756{
2757	struct inode *inode = mapping->host;
2758
2759	trace_writeback_dirty_folio(folio, mapping);
2760
2761	if (mapping_can_writeback(mapping)) {
2762		struct bdi_writeback *wb;
2763		long nr = folio_nr_pages(folio);
2764
2765		inode_attach_wb(inode, folio);
2766		wb = inode_to_wb(inode);
2767
2768		__lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
2769		__zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
2770		__node_stat_mod_folio(folio, NR_DIRTIED, nr);
2771		wb_stat_mod(wb, WB_RECLAIMABLE, nr);
2772		wb_stat_mod(wb, WB_DIRTIED, nr);
2773		task_io_account_write(nr * PAGE_SIZE);
2774		current->nr_dirtied += nr;
2775		__this_cpu_add(bdp_ratelimits, nr);
2776
2777		mem_cgroup_track_foreign_dirty(folio, wb);
2778	}
2779}
 
2780
2781/*
2782 * Helper function for deaccounting dirty page without writeback.
2783 *
 
2784 */
2785void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
 
2786{
2787	long nr = folio_nr_pages(folio);
2788
2789	lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
2790	zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
2791	wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
2792	task_io_account_cancelled_write(nr * PAGE_SIZE);
2793}
2794
2795/*
2796 * Mark the folio dirty, and set it dirty in the page cache.
 
2797 *
2798 * If warn is true, then emit a warning if the folio is not uptodate and has
2799 * not been truncated.
 
2800 *
2801 * It is the caller's responsibility to prevent the folio from being truncated
2802 * while this function is in progress, although it may have been truncated
2803 * before this function is called.  Most callers have the folio locked.
2804 * A few have the folio blocked from truncation through other means (e.g.
2805 * zap_vma_pages() has it mapped and is holding the page table lock).
2806 * When called from mark_buffer_dirty(), the filesystem should hold a
2807 * reference to the buffer_head that is being marked dirty, which causes
2808 * try_to_free_buffers() to fail.
2809 */
2810void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
2811			     int warn)
2812{
2813	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2814
2815	xa_lock_irqsave(&mapping->i_pages, flags);
2816	if (folio->mapping) {	/* Race with truncate? */
2817		WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
2818		folio_account_dirtied(folio, mapping);
2819		__xa_set_mark(&mapping->i_pages, folio_index(folio),
2820				PAGECACHE_TAG_DIRTY);
2821	}
2822	xa_unlock_irqrestore(&mapping->i_pages, flags);
 
2823}
 
2824
2825/**
2826 * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
2827 * @mapping: Address space this folio belongs to.
2828 * @folio: Folio to be marked as dirty.
2829 *
2830 * Filesystems which do not use buffer heads should call this function
2831 * from their dirty_folio address space operation.  It ignores the
2832 * contents of folio_get_private(), so if the filesystem marks individual
2833 * blocks as dirty, the filesystem should handle that itself.
2834 *
2835 * This is also sometimes used by filesystems which use buffer_heads when
2836 * a single buffer is being dirtied: we want to set the folio dirty in
2837 * that case, but not all the buffers.  This is a "bottom-up" dirtying,
2838 * whereas block_dirty_folio() is a "top-down" dirtying.
2839 *
2840 * The caller must ensure this doesn't race with truncation.  Most will
2841 * simply hold the folio lock, but e.g. zap_pte_range() calls with the
2842 * folio mapped and the pte lock held, which also locks out truncation.
2843 */
2844bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
2845{
2846	if (folio_test_set_dirty(folio))
2847		return false;
2848
2849	__folio_mark_dirty(folio, mapping, !folio_test_private(folio));
2850
2851	if (mapping->host) {
2852		/* !PageAnon && !swapper_space */
2853		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2854	}
2855	return true;
2856}
2857EXPORT_SYMBOL(filemap_dirty_folio);
2858
2859/**
2860 * folio_redirty_for_writepage - Decline to write a dirty folio.
2861 * @wbc: The writeback control.
2862 * @folio: The folio.
2863 *
2864 * When a writepage implementation decides that it doesn't want to write
2865 * @folio for some reason, it should call this function, unlock @folio and
2866 * return 0.
2867 *
2868 * Return: True if we redirtied the folio.  False if someone else dirtied
2869 * it first.
2870 */
2871bool folio_redirty_for_writepage(struct writeback_control *wbc,
2872		struct folio *folio)
2873{
2874	struct address_space *mapping = folio->mapping;
2875	long nr = folio_nr_pages(folio);
2876	bool ret;
2877
2878	wbc->pages_skipped += nr;
2879	ret = filemap_dirty_folio(mapping, folio);
2880	if (mapping && mapping_can_writeback(mapping)) {
2881		struct inode *inode = mapping->host;
2882		struct bdi_writeback *wb;
2883		struct wb_lock_cookie cookie = {};
2884
2885		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2886		current->nr_dirtied -= nr;
2887		node_stat_mod_folio(folio, NR_DIRTIED, -nr);
2888		wb_stat_mod(wb, WB_DIRTIED, -nr);
2889		unlocked_inode_to_wb_end(inode, &cookie);
2890	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2891	return ret;
2892}
2893EXPORT_SYMBOL(folio_redirty_for_writepage);
2894
2895/**
2896 * folio_mark_dirty - Mark a folio as being modified.
2897 * @folio: The folio.
2898 *
2899 * The folio may not be truncated while this function is running.
2900 * Holding the folio lock is sufficient to prevent truncation, but some
2901 * callers cannot acquire a sleeping lock.  These callers instead hold
2902 * the page table lock for a page table which contains at least one page
2903 * in this folio.  Truncation will block on the page table lock as it
2904 * unmaps pages before removing the folio from its mapping.
2905 *
2906 * Return: True if the folio was newly dirtied, false if it was already dirty.
 
2907 */
2908bool folio_mark_dirty(struct folio *folio)
2909{
2910	struct address_space *mapping = folio_mapping(folio);
2911
 
2912	if (likely(mapping)) {
 
2913		/*
2914		 * readahead/folio_deactivate could remain
2915		 * PG_readahead/PG_reclaim due to race with folio_end_writeback
2916		 * About readahead, if the folio is written, the flags would be
2917		 * reset. So no problem.
2918		 * About folio_deactivate, if the folio is redirtied,
2919		 * the flag will be reset. So no problem. but if the
2920		 * folio is used by readahead it will confuse readahead
2921		 * and make it restart the size rampup process. But it's
2922		 * a trivial problem.
2923		 */
2924		if (folio_test_reclaim(folio))
2925			folio_clear_reclaim(folio);
2926		return mapping->a_ops->dirty_folio(mapping, folio);
 
 
 
 
 
 
 
2927	}
2928
2929	return noop_dirty_folio(mapping, folio);
2930}
2931EXPORT_SYMBOL(folio_mark_dirty);
2932
2933/*
2934 * folio_mark_dirty() is racy if the caller has no reference against
2935 * folio->mapping->host, and if the folio is unlocked.  This is because another
2936 * CPU could truncate the folio off the mapping and then free the mapping.
2937 *
2938 * Usually, the folio _is_ locked, or the caller is a user-space process which
2939 * holds a reference on the inode by having an open file.
2940 *
2941 * In other cases, the folio should be locked before running folio_mark_dirty().
2942 */
2943bool folio_mark_dirty_lock(struct folio *folio)
2944{
2945	bool ret;
2946
2947	folio_lock(folio);
2948	ret = folio_mark_dirty(folio);
2949	folio_unlock(folio);
2950	return ret;
2951}
2952EXPORT_SYMBOL(folio_mark_dirty_lock);
2953
2954/*
2955 * This cancels just the dirty bit on the kernel page itself, it does NOT
2956 * actually remove dirty bits on any mmap's that may be around. It also
2957 * leaves the page tagged dirty, so any sync activity will still find it on
2958 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2959 * look at the dirty bits in the VM.
2960 *
2961 * Doing this should *normally* only ever be done when a page is truncated,
2962 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2963 * this when it notices that somebody has cleaned out all the buffers on a
2964 * page without actually doing it through the VM. Can you say "ext3 is
2965 * horribly ugly"? Thought you could.
2966 */
2967void __folio_cancel_dirty(struct folio *folio)
2968{
2969	struct address_space *mapping = folio_mapping(folio);
2970
2971	if (mapping_can_writeback(mapping)) {
2972		struct inode *inode = mapping->host;
2973		struct bdi_writeback *wb;
2974		struct wb_lock_cookie cookie = {};
2975
 
2976		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2977
2978		if (folio_test_clear_dirty(folio))
2979			folio_account_cleaned(folio, wb);
2980
2981		unlocked_inode_to_wb_end(inode, &cookie);
 
2982	} else {
2983		folio_clear_dirty(folio);
2984	}
2985}
2986EXPORT_SYMBOL(__folio_cancel_dirty);
2987
2988/*
2989 * Clear a folio's dirty flag, while caring for dirty memory accounting.
2990 * Returns true if the folio was previously dirty.
2991 *
2992 * This is for preparing to put the folio under writeout.  We leave
2993 * the folio tagged as dirty in the xarray so that a concurrent
2994 * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
2995 * The ->writepage implementation will run either folio_start_writeback()
2996 * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
2997 * and xarray dirty tag back into sync.
2998 *
2999 * This incoherency between the folio's dirty flag and xarray tag is
3000 * unfortunate, but it only exists while the folio is locked.
3001 */
3002bool folio_clear_dirty_for_io(struct folio *folio)
3003{
3004	struct address_space *mapping = folio_mapping(folio);
3005	bool ret = false;
3006
3007	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
3008
3009	if (mapping && mapping_can_writeback(mapping)) {
3010		struct inode *inode = mapping->host;
3011		struct bdi_writeback *wb;
3012		struct wb_lock_cookie cookie = {};
3013
3014		/*
3015		 * Yes, Virginia, this is indeed insane.
3016		 *
3017		 * We use this sequence to make sure that
3018		 *  (a) we account for dirty stats properly
3019		 *  (b) we tell the low-level filesystem to
3020		 *      mark the whole folio dirty if it was
3021		 *      dirty in a pagetable. Only to then
3022		 *  (c) clean the folio again and return 1 to
3023		 *      cause the writeback.
3024		 *
3025		 * This way we avoid all nasty races with the
3026		 * dirty bit in multiple places and clearing
3027		 * them concurrently from different threads.
3028		 *
3029		 * Note! Normally the "folio_mark_dirty(folio)"
3030		 * has no effect on the actual dirty bit - since
3031		 * that will already usually be set. But we
3032		 * need the side effects, and it can help us
3033		 * avoid races.
3034		 *
3035		 * We basically use the folio "master dirty bit"
3036		 * as a serialization point for all the different
3037		 * threads doing their things.
3038		 */
3039		if (folio_mkclean(folio))
3040			folio_mark_dirty(folio);
3041		/*
3042		 * We carefully synchronise fault handlers against
3043		 * installing a dirty pte and marking the folio dirty
3044		 * at this point.  We do this by having them hold the
3045		 * page lock while dirtying the folio, and folios are
3046		 * always locked coming in here, so we get the desired
3047		 * exclusion.
3048		 */
3049		wb = unlocked_inode_to_wb_begin(inode, &cookie);
3050		if (folio_test_clear_dirty(folio)) {
3051			long nr = folio_nr_pages(folio);
3052			lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
3053			zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3054			wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
3055			ret = true;
3056		}
3057		unlocked_inode_to_wb_end(inode, &cookie);
3058		return ret;
3059	}
3060	return folio_test_clear_dirty(folio);
3061}
3062EXPORT_SYMBOL(folio_clear_dirty_for_io);
3063
3064static void wb_inode_writeback_start(struct bdi_writeback *wb)
3065{
3066	atomic_inc(&wb->writeback_inodes);
3067}
3068
3069static void wb_inode_writeback_end(struct bdi_writeback *wb)
3070{
3071	unsigned long flags;
3072	atomic_dec(&wb->writeback_inodes);
3073	/*
3074	 * Make sure estimate of writeback throughput gets updated after
3075	 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
3076	 * (which is the interval other bandwidth updates use for batching) so
3077	 * that if multiple inodes end writeback at a similar time, they get
3078	 * batched into one bandwidth update.
3079	 */
3080	spin_lock_irqsave(&wb->work_lock, flags);
3081	if (test_bit(WB_registered, &wb->state))
3082		queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
3083	spin_unlock_irqrestore(&wb->work_lock, flags);
3084}
3085
3086bool __folio_end_writeback(struct folio *folio)
3087{
3088	long nr = folio_nr_pages(folio);
3089	struct address_space *mapping = folio_mapping(folio);
3090	bool ret;
3091
 
 
3092	if (mapping && mapping_use_writeback_tags(mapping)) {
3093		struct inode *inode = mapping->host;
3094		struct backing_dev_info *bdi = inode_to_bdi(inode);
3095		unsigned long flags;
3096
3097		xa_lock_irqsave(&mapping->i_pages, flags);
3098		ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3099		__xa_clear_mark(&mapping->i_pages, folio_index(folio),
3100					PAGECACHE_TAG_WRITEBACK);
3101		if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3102			struct bdi_writeback *wb = inode_to_wb(inode);
3103
3104			wb_stat_mod(wb, WB_WRITEBACK, -nr);
3105			__wb_writeout_add(wb, nr);
3106			if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
3107				wb_inode_writeback_end(wb);
3108		}
3109
3110		if (mapping->host && !mapping_tagged(mapping,
3111						     PAGECACHE_TAG_WRITEBACK))
3112			sb_clear_inode_writeback(mapping->host);
3113
3114		xa_unlock_irqrestore(&mapping->i_pages, flags);
3115	} else {
3116		ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
3117	}
3118
3119	lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
3120	zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
3121	node_stat_mod_folio(folio, NR_WRITTEN, nr);
3122
 
 
 
 
 
 
 
3123	return ret;
3124}
3125
3126void __folio_start_writeback(struct folio *folio, bool keep_write)
3127{
3128	long nr = folio_nr_pages(folio);
3129	struct address_space *mapping = folio_mapping(folio);
3130	int access_ret;
3131
3132	VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
3133
 
3134	if (mapping && mapping_use_writeback_tags(mapping)) {
3135		XA_STATE(xas, &mapping->i_pages, folio_index(folio));
3136		struct inode *inode = mapping->host;
3137		struct backing_dev_info *bdi = inode_to_bdi(inode);
3138		unsigned long flags;
3139		bool on_wblist;
3140
3141		xas_lock_irqsave(&xas, flags);
3142		xas_load(&xas);
3143		folio_test_set_writeback(folio);
 
 
 
 
 
 
 
 
 
3144
3145		on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
3146
3147		xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
3148		if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
3149			struct bdi_writeback *wb = inode_to_wb(inode);
3150
3151			wb_stat_mod(wb, WB_WRITEBACK, nr);
3152			if (!on_wblist)
3153				wb_inode_writeback_start(wb);
3154		}
3155
3156		/*
3157		 * We can come through here when swapping anonymous
3158		 * folios, so we don't necessarily have an inode to
3159		 * track for sync.
3160		 */
3161		if (mapping->host && !on_wblist)
3162			sb_mark_inode_writeback(mapping->host);
3163		if (!folio_test_dirty(folio))
3164			xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
3165		if (!keep_write)
3166			xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
3167		xas_unlock_irqrestore(&xas, flags);
 
3168	} else {
3169		folio_test_set_writeback(folio);
 
 
 
 
3170	}
 
 
3171
3172	lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
3173	zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
3174
3175	access_ret = arch_make_folio_accessible(folio);
3176	/*
3177	 * If writeback has been triggered on a page that cannot be made
3178	 * accessible, it is too late to recover here.
3179	 */
3180	VM_BUG_ON_FOLIO(access_ret != 0, folio);
3181}
3182EXPORT_SYMBOL(__folio_start_writeback);
3183
3184/**
3185 * folio_wait_writeback - Wait for a folio to finish writeback.
3186 * @folio: The folio to wait for.
3187 *
3188 * If the folio is currently being written back to storage, wait for the
3189 * I/O to complete.
3190 *
3191 * Context: Sleeps.  Must be called in process context and with
3192 * no spinlocks held.  Caller should hold a reference on the folio.
3193 * If the folio is not locked, writeback may start again after writeback
3194 * has finished.
3195 */
3196void folio_wait_writeback(struct folio *folio)
3197{
3198	while (folio_test_writeback(folio)) {
3199		trace_folio_wait_writeback(folio, folio_mapping(folio));
3200		folio_wait_bit(folio, PG_writeback);
3201	}
3202}
3203EXPORT_SYMBOL_GPL(folio_wait_writeback);
3204
3205/**
3206 * folio_wait_writeback_killable - Wait for a folio to finish writeback.
3207 * @folio: The folio to wait for.
3208 *
3209 * If the folio is currently being written back to storage, wait for the
3210 * I/O to complete or a fatal signal to arrive.
3211 *
3212 * Context: Sleeps.  Must be called in process context and with
3213 * no spinlocks held.  Caller should hold a reference on the folio.
3214 * If the folio is not locked, writeback may start again after writeback
3215 * has finished.
3216 * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
3217 */
3218int folio_wait_writeback_killable(struct folio *folio)
3219{
3220	while (folio_test_writeback(folio)) {
3221		trace_folio_wait_writeback(folio, folio_mapping(folio));
3222		if (folio_wait_bit_killable(folio, PG_writeback))
3223			return -EINTR;
3224	}
3225
3226	return 0;
3227}
3228EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
3229
3230/**
3231 * folio_wait_stable() - wait for writeback to finish, if necessary.
3232 * @folio: The folio to wait on.
3233 *
3234 * This function determines if the given folio is related to a backing
3235 * device that requires folio contents to be held stable during writeback.
3236 * If so, then it will wait for any pending writeback to complete.
3237 *
3238 * Context: Sleeps.  Must be called in process context and with
3239 * no spinlocks held.  Caller should hold a reference on the folio.
3240 * If the folio is not locked, writeback may start again after writeback
3241 * has finished.
3242 */
3243void folio_wait_stable(struct folio *folio)
3244{
3245	if (mapping_stable_writes(folio_mapping(folio)))
3246		folio_wait_writeback(folio);
3247}
3248EXPORT_SYMBOL_GPL(folio_wait_stable);