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);
v5.14.15
   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/export.h>
  17#include <linux/spinlock.h>
  18#include <linux/fs.h>
  19#include <linux/mm.h>
  20#include <linux/swap.h>
  21#include <linux/slab.h>
  22#include <linux/pagemap.h>
  23#include <linux/writeback.h>
  24#include <linux/init.h>
  25#include <linux/backing-dev.h>
  26#include <linux/task_io_accounting_ops.h>
  27#include <linux/blkdev.h>
  28#include <linux/mpage.h>
  29#include <linux/rmap.h>
  30#include <linux/percpu.h>
 
  31#include <linux/smp.h>
  32#include <linux/sysctl.h>
  33#include <linux/cpu.h>
  34#include <linux/syscalls.h>
 
  35#include <linux/pagevec.h>
  36#include <linux/timer.h>
  37#include <linux/sched/rt.h>
  38#include <linux/sched/signal.h>
  39#include <linux/mm_inline.h>
  40#include <trace/events/writeback.h>
  41
  42#include "internal.h"
  43
  44/*
  45 * Sleep at most 200ms at a time in balance_dirty_pages().
  46 */
  47#define MAX_PAUSE		max(HZ/5, 1)
  48
  49/*
  50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
  51 * by raising pause time to max_pause when falls below it.
  52 */
  53#define DIRTY_POLL_THRESH	(128 >> (PAGE_SHIFT - 10))
  54
  55/*
  56 * Estimate write bandwidth at 200ms intervals.
  57 */
  58#define BANDWIDTH_INTERVAL	max(HZ/5, 1)
  59
  60#define RATELIMIT_CALC_SHIFT	10
  61
  62/*
  63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  64 * will look to see if it needs to force writeback or throttling.
  65 */
  66static long ratelimit_pages = 32;
  67
  68/* The following parameters are exported via /proc/sys/vm */
  69
  70/*
  71 * Start background writeback (via writeback threads) at this percentage
  72 */
  73int dirty_background_ratio = 10;
  74
  75/*
  76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
  77 * dirty_background_ratio * the amount of dirtyable memory
  78 */
  79unsigned long dirty_background_bytes;
  80
  81/*
  82 * free highmem will not be subtracted from the total free memory
  83 * for calculating free ratios if vm_highmem_is_dirtyable is true
  84 */
  85int vm_highmem_is_dirtyable;
  86
  87/*
  88 * The generator of dirty data starts writeback at this percentage
  89 */
  90int vm_dirty_ratio = 20;
  91
  92/*
  93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
  94 * vm_dirty_ratio * the amount of dirtyable memory
  95 */
  96unsigned long vm_dirty_bytes;
  97
  98/*
  99 * The interval between `kupdate'-style writebacks
 100 */
 101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
 102
 103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
 104
 105/*
 106 * The longest time for which data is allowed to remain dirty
 107 */
 108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
 109
 110/*
 
 
 
 
 
 111 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
 112 * a full sync is triggered after this time elapses without any disk activity.
 113 */
 114int laptop_mode;
 115
 116EXPORT_SYMBOL(laptop_mode);
 117
 118/* End of sysctl-exported parameters */
 119
 120struct wb_domain global_wb_domain;
 121
 122/* consolidated parameters for balance_dirty_pages() and its subroutines */
 123struct dirty_throttle_control {
 124#ifdef CONFIG_CGROUP_WRITEBACK
 125	struct wb_domain	*dom;
 126	struct dirty_throttle_control *gdtc;	/* only set in memcg dtc's */
 127#endif
 128	struct bdi_writeback	*wb;
 129	struct fprop_local_percpu *wb_completions;
 130
 131	unsigned long		avail;		/* dirtyable */
 132	unsigned long		dirty;		/* file_dirty + write + nfs */
 133	unsigned long		thresh;		/* dirty threshold */
 134	unsigned long		bg_thresh;	/* dirty background threshold */
 135
 136	unsigned long		wb_dirty;	/* per-wb counterparts */
 137	unsigned long		wb_thresh;
 138	unsigned long		wb_bg_thresh;
 139
 140	unsigned long		pos_ratio;
 141};
 142
 143/*
 144 * Length of period for aging writeout fractions of bdis. This is an
 145 * arbitrarily chosen number. The longer the period, the slower fractions will
 146 * reflect changes in current writeout rate.
 147 */
 148#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
 149
 150#ifdef CONFIG_CGROUP_WRITEBACK
 151
 152#define GDTC_INIT(__wb)		.wb = (__wb),				\
 153				.dom = &global_wb_domain,		\
 154				.wb_completions = &(__wb)->completions
 155
 156#define GDTC_INIT_NO_WB		.dom = &global_wb_domain
 157
 158#define MDTC_INIT(__wb, __gdtc)	.wb = (__wb),				\
 159				.dom = mem_cgroup_wb_domain(__wb),	\
 160				.wb_completions = &(__wb)->memcg_completions, \
 161				.gdtc = __gdtc
 162
 163static bool mdtc_valid(struct dirty_throttle_control *dtc)
 164{
 165	return dtc->dom;
 166}
 167
 168static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
 169{
 170	return dtc->dom;
 171}
 172
 173static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
 174{
 175	return mdtc->gdtc;
 176}
 177
 178static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
 179{
 180	return &wb->memcg_completions;
 181}
 182
 183static void wb_min_max_ratio(struct bdi_writeback *wb,
 184			     unsigned long *minp, unsigned long *maxp)
 185{
 186	unsigned long this_bw = wb->avg_write_bandwidth;
 187	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
 188	unsigned long long min = wb->bdi->min_ratio;
 189	unsigned long long max = wb->bdi->max_ratio;
 190
 191	/*
 192	 * @wb may already be clean by the time control reaches here and
 193	 * the total may not include its bw.
 194	 */
 195	if (this_bw < tot_bw) {
 196		if (min) {
 197			min *= this_bw;
 198			min = div64_ul(min, tot_bw);
 199		}
 200		if (max < 100) {
 201			max *= this_bw;
 202			max = div64_ul(max, tot_bw);
 203		}
 204	}
 205
 206	*minp = min;
 207	*maxp = max;
 208}
 209
 210#else	/* CONFIG_CGROUP_WRITEBACK */
 211
 212#define GDTC_INIT(__wb)		.wb = (__wb),                           \
 213				.wb_completions = &(__wb)->completions
 214#define GDTC_INIT_NO_WB
 215#define MDTC_INIT(__wb, __gdtc)
 216
 217static bool mdtc_valid(struct dirty_throttle_control *dtc)
 218{
 219	return false;
 220}
 221
 222static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
 223{
 224	return &global_wb_domain;
 225}
 226
 227static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
 228{
 229	return NULL;
 230}
 231
 232static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
 233{
 234	return NULL;
 235}
 236
 237static void wb_min_max_ratio(struct bdi_writeback *wb,
 238			     unsigned long *minp, unsigned long *maxp)
 239{
 240	*minp = wb->bdi->min_ratio;
 241	*maxp = wb->bdi->max_ratio;
 242}
 243
 244#endif	/* CONFIG_CGROUP_WRITEBACK */
 245
 246/*
 247 * In a memory zone, there is a certain amount of pages we consider
 248 * available for the page cache, which is essentially the number of
 249 * free and reclaimable pages, minus some zone reserves to protect
 250 * lowmem and the ability to uphold the zone's watermarks without
 251 * requiring writeback.
 252 *
 253 * This number of dirtyable pages is the base value of which the
 254 * user-configurable dirty ratio is the effective number of pages that
 255 * are allowed to be actually dirtied.  Per individual zone, or
 256 * globally by using the sum of dirtyable pages over all zones.
 257 *
 258 * Because the user is allowed to specify the dirty limit globally as
 259 * absolute number of bytes, calculating the per-zone dirty limit can
 260 * require translating the configured limit into a percentage of
 261 * global dirtyable memory first.
 262 */
 263
 264/**
 265 * node_dirtyable_memory - number of dirtyable pages in a node
 266 * @pgdat: the node
 267 *
 268 * Return: the node's number of pages potentially available for dirty
 269 * page cache.  This is the base value for the per-node dirty limits.
 270 */
 271static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
 272{
 273	unsigned long nr_pages = 0;
 274	int z;
 275
 276	for (z = 0; z < MAX_NR_ZONES; z++) {
 277		struct zone *zone = pgdat->node_zones + z;
 278
 279		if (!populated_zone(zone))
 280			continue;
 281
 282		nr_pages += zone_page_state(zone, NR_FREE_PAGES);
 283	}
 284
 285	/*
 286	 * Pages reserved for the kernel should not be considered
 287	 * dirtyable, to prevent a situation where reclaim has to
 288	 * clean pages in order to balance the zones.
 289	 */
 290	nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
 291
 292	nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
 293	nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
 294
 295	return nr_pages;
 296}
 297
 298static unsigned long highmem_dirtyable_memory(unsigned long total)
 299{
 300#ifdef CONFIG_HIGHMEM
 301	int node;
 302	unsigned long x = 0;
 303	int i;
 304
 305	for_each_node_state(node, N_HIGH_MEMORY) {
 306		for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
 307			struct zone *z;
 308			unsigned long nr_pages;
 309
 310			if (!is_highmem_idx(i))
 311				continue;
 312
 313			z = &NODE_DATA(node)->node_zones[i];
 314			if (!populated_zone(z))
 315				continue;
 316
 317			nr_pages = zone_page_state(z, NR_FREE_PAGES);
 318			/* watch for underflows */
 319			nr_pages -= min(nr_pages, high_wmark_pages(z));
 320			nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
 321			nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
 322			x += nr_pages;
 323		}
 324	}
 325
 326	/*
 327	 * Unreclaimable memory (kernel memory or anonymous memory
 328	 * without swap) can bring down the dirtyable pages below
 329	 * the zone's dirty balance reserve and the above calculation
 330	 * will underflow.  However we still want to add in nodes
 331	 * which are below threshold (negative values) to get a more
 332	 * accurate calculation but make sure that the total never
 333	 * underflows.
 334	 */
 335	if ((long)x < 0)
 336		x = 0;
 337
 338	/*
 339	 * Make sure that the number of highmem pages is never larger
 340	 * than the number of the total dirtyable memory. This can only
 341	 * occur in very strange VM situations but we want to make sure
 342	 * that this does not occur.
 343	 */
 344	return min(x, total);
 345#else
 346	return 0;
 347#endif
 348}
 349
 350/**
 351 * global_dirtyable_memory - number of globally dirtyable pages
 352 *
 353 * Return: the global number of pages potentially available for dirty
 354 * page cache.  This is the base value for the global dirty limits.
 355 */
 356static unsigned long global_dirtyable_memory(void)
 357{
 358	unsigned long x;
 359
 360	x = global_zone_page_state(NR_FREE_PAGES);
 361	/*
 362	 * Pages reserved for the kernel should not be considered
 363	 * dirtyable, to prevent a situation where reclaim has to
 364	 * clean pages in order to balance the zones.
 365	 */
 366	x -= min(x, totalreserve_pages);
 367
 368	x += global_node_page_state(NR_INACTIVE_FILE);
 369	x += global_node_page_state(NR_ACTIVE_FILE);
 370
 371	if (!vm_highmem_is_dirtyable)
 372		x -= highmem_dirtyable_memory(x);
 373
 374	return x + 1;	/* Ensure that we never return 0 */
 375}
 376
 377/**
 378 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
 379 * @dtc: dirty_throttle_control of interest
 380 *
 381 * Calculate @dtc->thresh and ->bg_thresh considering
 382 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}.  The caller
 383 * must ensure that @dtc->avail is set before calling this function.  The
 384 * dirty limits will be lifted by 1/4 for real-time tasks.
 
 385 */
 386static void domain_dirty_limits(struct dirty_throttle_control *dtc)
 387{
 388	const unsigned long available_memory = dtc->avail;
 389	struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
 390	unsigned long bytes = vm_dirty_bytes;
 391	unsigned long bg_bytes = dirty_background_bytes;
 392	/* convert ratios to per-PAGE_SIZE for higher precision */
 393	unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
 394	unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
 395	unsigned long thresh;
 396	unsigned long bg_thresh;
 397	struct task_struct *tsk;
 398
 399	/* gdtc is !NULL iff @dtc is for memcg domain */
 400	if (gdtc) {
 401		unsigned long global_avail = gdtc->avail;
 402
 403		/*
 404		 * The byte settings can't be applied directly to memcg
 405		 * domains.  Convert them to ratios by scaling against
 406		 * globally available memory.  As the ratios are in
 407		 * per-PAGE_SIZE, they can be obtained by dividing bytes by
 408		 * number of pages.
 409		 */
 410		if (bytes)
 411			ratio = min(DIV_ROUND_UP(bytes, global_avail),
 412				    PAGE_SIZE);
 413		if (bg_bytes)
 414			bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
 415				       PAGE_SIZE);
 416		bytes = bg_bytes = 0;
 417	}
 418
 419	if (bytes)
 420		thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
 421	else
 422		thresh = (ratio * available_memory) / PAGE_SIZE;
 423
 424	if (bg_bytes)
 425		bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
 426	else
 427		bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
 428
 429	if (bg_thresh >= thresh)
 430		bg_thresh = thresh / 2;
 431	tsk = current;
 432	if (rt_task(tsk)) {
 433		bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
 434		thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
 435	}
 436	dtc->thresh = thresh;
 437	dtc->bg_thresh = bg_thresh;
 438
 439	/* we should eventually report the domain in the TP */
 440	if (!gdtc)
 441		trace_global_dirty_state(bg_thresh, thresh);
 442}
 443
 444/**
 445 * global_dirty_limits - background-writeback and dirty-throttling thresholds
 446 * @pbackground: out parameter for bg_thresh
 447 * @pdirty: out parameter for thresh
 448 *
 449 * Calculate bg_thresh and thresh for global_wb_domain.  See
 450 * domain_dirty_limits() for details.
 451 */
 452void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
 453{
 454	struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
 455
 456	gdtc.avail = global_dirtyable_memory();
 457	domain_dirty_limits(&gdtc);
 458
 459	*pbackground = gdtc.bg_thresh;
 460	*pdirty = gdtc.thresh;
 461}
 462
 463/**
 464 * node_dirty_limit - maximum number of dirty pages allowed in a node
 465 * @pgdat: the node
 466 *
 467 * Return: the maximum number of dirty pages allowed in a node, based
 468 * on the node's dirtyable memory.
 469 */
 470static unsigned long node_dirty_limit(struct pglist_data *pgdat)
 471{
 472	unsigned long node_memory = node_dirtyable_memory(pgdat);
 473	struct task_struct *tsk = current;
 474	unsigned long dirty;
 475
 476	if (vm_dirty_bytes)
 477		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
 478			node_memory / global_dirtyable_memory();
 479	else
 480		dirty = vm_dirty_ratio * node_memory / 100;
 481
 482	if (rt_task(tsk))
 483		dirty += dirty / 4;
 484
 485	return dirty;
 486}
 487
 488/**
 489 * node_dirty_ok - tells whether a node is within its dirty limits
 490 * @pgdat: the node to check
 491 *
 492 * Return: %true when the dirty pages in @pgdat are within the node's
 493 * dirty limit, %false if the limit is exceeded.
 494 */
 495bool node_dirty_ok(struct pglist_data *pgdat)
 496{
 497	unsigned long limit = node_dirty_limit(pgdat);
 498	unsigned long nr_pages = 0;
 499
 500	nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
 
 501	nr_pages += node_page_state(pgdat, NR_WRITEBACK);
 502
 503	return nr_pages <= limit;
 504}
 505
 506int dirty_background_ratio_handler(struct ctl_table *table, int write,
 507		void *buffer, size_t *lenp, loff_t *ppos)
 
 508{
 509	int ret;
 510
 511	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 512	if (ret == 0 && write)
 513		dirty_background_bytes = 0;
 514	return ret;
 515}
 516
 517int dirty_background_bytes_handler(struct ctl_table *table, int write,
 518		void *buffer, size_t *lenp, loff_t *ppos)
 
 519{
 520	int ret;
 521
 522	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 523	if (ret == 0 && write)
 524		dirty_background_ratio = 0;
 525	return ret;
 526}
 527
 528int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer,
 529		size_t *lenp, loff_t *ppos)
 
 530{
 531	int old_ratio = vm_dirty_ratio;
 532	int ret;
 533
 534	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 535	if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
 536		writeback_set_ratelimit();
 537		vm_dirty_bytes = 0;
 538	}
 539	return ret;
 540}
 541
 542int dirty_bytes_handler(struct ctl_table *table, int write,
 543		void *buffer, size_t *lenp, loff_t *ppos)
 
 544{
 545	unsigned long old_bytes = vm_dirty_bytes;
 546	int ret;
 547
 548	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 549	if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
 550		writeback_set_ratelimit();
 551		vm_dirty_ratio = 0;
 552	}
 553	return ret;
 554}
 555
 556static unsigned long wp_next_time(unsigned long cur_time)
 557{
 558	cur_time += VM_COMPLETIONS_PERIOD_LEN;
 559	/* 0 has a special meaning... */
 560	if (!cur_time)
 561		return 1;
 562	return cur_time;
 563}
 564
 565static void wb_domain_writeout_inc(struct wb_domain *dom,
 566				   struct fprop_local_percpu *completions,
 567				   unsigned int max_prop_frac)
 568{
 569	__fprop_inc_percpu_max(&dom->completions, completions,
 570			       max_prop_frac);
 571	/* First event after period switching was turned off? */
 572	if (unlikely(!dom->period_time)) {
 573		/*
 574		 * We can race with other __bdi_writeout_inc calls here but
 575		 * it does not cause any harm since the resulting time when
 576		 * timer will fire and what is in writeout_period_time will be
 577		 * roughly the same.
 578		 */
 579		dom->period_time = wp_next_time(jiffies);
 580		mod_timer(&dom->period_timer, dom->period_time);
 581	}
 582}
 583
 584/*
 585 * Increment @wb's writeout completion count and the global writeout
 586 * completion count. Called from test_clear_page_writeback().
 587 */
 588static inline void __wb_writeout_inc(struct bdi_writeback *wb)
 589{
 590	struct wb_domain *cgdom;
 591
 592	inc_wb_stat(wb, WB_WRITTEN);
 593	wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
 594			       wb->bdi->max_prop_frac);
 595
 596	cgdom = mem_cgroup_wb_domain(wb);
 597	if (cgdom)
 598		wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
 599				       wb->bdi->max_prop_frac);
 600}
 601
 602void wb_writeout_inc(struct bdi_writeback *wb)
 603{
 604	unsigned long flags;
 605
 606	local_irq_save(flags);
 607	__wb_writeout_inc(wb);
 608	local_irq_restore(flags);
 609}
 610EXPORT_SYMBOL_GPL(wb_writeout_inc);
 611
 612/*
 613 * On idle system, we can be called long after we scheduled because we use
 614 * deferred timers so count with missed periods.
 615 */
 616static void writeout_period(struct timer_list *t)
 617{
 618	struct wb_domain *dom = from_timer(dom, t, period_timer);
 619	int miss_periods = (jiffies - dom->period_time) /
 620						 VM_COMPLETIONS_PERIOD_LEN;
 621
 622	if (fprop_new_period(&dom->completions, miss_periods + 1)) {
 623		dom->period_time = wp_next_time(dom->period_time +
 624				miss_periods * VM_COMPLETIONS_PERIOD_LEN);
 625		mod_timer(&dom->period_timer, dom->period_time);
 626	} else {
 627		/*
 628		 * Aging has zeroed all fractions. Stop wasting CPU on period
 629		 * updates.
 630		 */
 631		dom->period_time = 0;
 632	}
 633}
 634
 635int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
 636{
 637	memset(dom, 0, sizeof(*dom));
 638
 639	spin_lock_init(&dom->lock);
 640
 641	timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
 642
 643	dom->dirty_limit_tstamp = jiffies;
 644
 645	return fprop_global_init(&dom->completions, gfp);
 646}
 647
 648#ifdef CONFIG_CGROUP_WRITEBACK
 649void wb_domain_exit(struct wb_domain *dom)
 650{
 651	del_timer_sync(&dom->period_timer);
 652	fprop_global_destroy(&dom->completions);
 653}
 654#endif
 655
 656/*
 657 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
 658 * registered backing devices, which, for obvious reasons, can not
 659 * exceed 100%.
 660 */
 661static unsigned int bdi_min_ratio;
 662
 663int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 664{
 665	int ret = 0;
 666
 667	spin_lock_bh(&bdi_lock);
 668	if (min_ratio > bdi->max_ratio) {
 669		ret = -EINVAL;
 670	} else {
 671		min_ratio -= bdi->min_ratio;
 672		if (bdi_min_ratio + min_ratio < 100) {
 673			bdi_min_ratio += min_ratio;
 674			bdi->min_ratio += min_ratio;
 675		} else {
 676			ret = -EINVAL;
 677		}
 678	}
 679	spin_unlock_bh(&bdi_lock);
 680
 681	return ret;
 682}
 683
 684int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
 685{
 686	int ret = 0;
 687
 688	if (max_ratio > 100)
 689		return -EINVAL;
 690
 691	spin_lock_bh(&bdi_lock);
 692	if (bdi->min_ratio > max_ratio) {
 693		ret = -EINVAL;
 694	} else {
 695		bdi->max_ratio = max_ratio;
 696		bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
 697	}
 698	spin_unlock_bh(&bdi_lock);
 699
 700	return ret;
 701}
 702EXPORT_SYMBOL(bdi_set_max_ratio);
 703
 704static unsigned long dirty_freerun_ceiling(unsigned long thresh,
 705					   unsigned long bg_thresh)
 706{
 707	return (thresh + bg_thresh) / 2;
 708}
 709
 710static unsigned long hard_dirty_limit(struct wb_domain *dom,
 711				      unsigned long thresh)
 712{
 713	return max(thresh, dom->dirty_limit);
 714}
 715
 716/*
 717 * Memory which can be further allocated to a memcg domain is capped by
 718 * system-wide clean memory excluding the amount being used in the domain.
 719 */
 720static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
 721			    unsigned long filepages, unsigned long headroom)
 722{
 723	struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
 724	unsigned long clean = filepages - min(filepages, mdtc->dirty);
 725	unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
 726	unsigned long other_clean = global_clean - min(global_clean, clean);
 727
 728	mdtc->avail = filepages + min(headroom, other_clean);
 729}
 730
 731/**
 732 * __wb_calc_thresh - @wb's share of dirty throttling threshold
 733 * @dtc: dirty_throttle_context of interest
 734 *
 
 
 
 735 * Note that balance_dirty_pages() will only seriously take it as a hard limit
 736 * when sleeping max_pause per page is not enough to keep the dirty pages under
 737 * control. For example, when the device is completely stalled due to some error
 738 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
 739 * In the other normal situations, it acts more gently by throttling the tasks
 740 * more (rather than completely block them) when the wb dirty pages go high.
 741 *
 742 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
 743 * - starving fast devices
 744 * - piling up dirty pages (that will take long time to sync) on slow devices
 745 *
 746 * The wb's share of dirty limit will be adapting to its throughput and
 747 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
 748 *
 749 * Return: @wb's dirty limit in pages. The term "dirty" in the context of
 750 * dirty balancing includes all PG_dirty and PG_writeback pages.
 751 */
 752static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
 753{
 754	struct wb_domain *dom = dtc_dom(dtc);
 755	unsigned long thresh = dtc->thresh;
 756	u64 wb_thresh;
 757	unsigned long numerator, denominator;
 758	unsigned long wb_min_ratio, wb_max_ratio;
 759
 760	/*
 761	 * Calculate this BDI's share of the thresh ratio.
 762	 */
 763	fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
 764			      &numerator, &denominator);
 765
 766	wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
 767	wb_thresh *= numerator;
 768	wb_thresh = div64_ul(wb_thresh, denominator);
 769
 770	wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
 771
 772	wb_thresh += (thresh * wb_min_ratio) / 100;
 773	if (wb_thresh > (thresh * wb_max_ratio) / 100)
 774		wb_thresh = thresh * wb_max_ratio / 100;
 775
 776	return wb_thresh;
 777}
 778
 779unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
 780{
 781	struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
 782					       .thresh = thresh };
 783	return __wb_calc_thresh(&gdtc);
 784}
 785
 786/*
 787 *                           setpoint - dirty 3
 788 *        f(dirty) := 1.0 + (----------------)
 789 *                           limit - setpoint
 790 *
 791 * it's a 3rd order polynomial that subjects to
 792 *
 793 * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
 794 * (2) f(setpoint) = 1.0 => the balance point
 795 * (3) f(limit)    = 0   => the hard limit
 796 * (4) df/dx      <= 0	 => negative feedback control
 797 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
 798 *     => fast response on large errors; small oscillation near setpoint
 799 */
 800static long long pos_ratio_polynom(unsigned long setpoint,
 801					  unsigned long dirty,
 802					  unsigned long limit)
 803{
 804	long long pos_ratio;
 805	long x;
 806
 807	x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
 808		      (limit - setpoint) | 1);
 809	pos_ratio = x;
 810	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 811	pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 812	pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
 813
 814	return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
 815}
 816
 817/*
 818 * Dirty position control.
 819 *
 820 * (o) global/bdi setpoints
 821 *
 822 * We want the dirty pages be balanced around the global/wb setpoints.
 823 * When the number of dirty pages is higher/lower than the setpoint, the
 824 * dirty position control ratio (and hence task dirty ratelimit) will be
 825 * decreased/increased to bring the dirty pages back to the setpoint.
 826 *
 827 *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
 828 *
 829 *     if (dirty < setpoint) scale up   pos_ratio
 830 *     if (dirty > setpoint) scale down pos_ratio
 831 *
 832 *     if (wb_dirty < wb_setpoint) scale up   pos_ratio
 833 *     if (wb_dirty > wb_setpoint) scale down pos_ratio
 834 *
 835 *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
 836 *
 837 * (o) global control line
 838 *
 839 *     ^ pos_ratio
 840 *     |
 841 *     |            |<===== global dirty control scope ======>|
 842 * 2.0  * * * * * * *
 843 *     |            .*
 844 *     |            . *
 845 *     |            .   *
 846 *     |            .     *
 847 *     |            .        *
 848 *     |            .            *
 849 * 1.0 ................................*
 850 *     |            .                  .     *
 851 *     |            .                  .          *
 852 *     |            .                  .              *
 853 *     |            .                  .                 *
 854 *     |            .                  .                    *
 855 *   0 +------------.------------------.----------------------*------------->
 856 *           freerun^          setpoint^                 limit^   dirty pages
 857 *
 858 * (o) wb control line
 859 *
 860 *     ^ pos_ratio
 861 *     |
 862 *     |            *
 863 *     |              *
 864 *     |                *
 865 *     |                  *
 866 *     |                    * |<=========== span ============>|
 867 * 1.0 .......................*
 868 *     |                      . *
 869 *     |                      .   *
 870 *     |                      .     *
 871 *     |                      .       *
 872 *     |                      .         *
 873 *     |                      .           *
 874 *     |                      .             *
 875 *     |                      .               *
 876 *     |                      .                 *
 877 *     |                      .                   *
 878 *     |                      .                     *
 879 * 1/4 ...............................................* * * * * * * * * * * *
 880 *     |                      .                         .
 881 *     |                      .                           .
 882 *     |                      .                             .
 883 *   0 +----------------------.-------------------------------.------------->
 884 *                wb_setpoint^                    x_intercept^
 885 *
 886 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
 887 * be smoothly throttled down to normal if it starts high in situations like
 888 * - start writing to a slow SD card and a fast disk at the same time. The SD
 889 *   card's wb_dirty may rush to many times higher than wb_setpoint.
 890 * - the wb dirty thresh drops quickly due to change of JBOD workload
 891 */
 892static void wb_position_ratio(struct dirty_throttle_control *dtc)
 893{
 894	struct bdi_writeback *wb = dtc->wb;
 895	unsigned long write_bw = wb->avg_write_bandwidth;
 896	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
 897	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
 898	unsigned long wb_thresh = dtc->wb_thresh;
 899	unsigned long x_intercept;
 900	unsigned long setpoint;		/* dirty pages' target balance point */
 901	unsigned long wb_setpoint;
 902	unsigned long span;
 903	long long pos_ratio;		/* for scaling up/down the rate limit */
 904	long x;
 905
 906	dtc->pos_ratio = 0;
 907
 908	if (unlikely(dtc->dirty >= limit))
 909		return;
 910
 911	/*
 912	 * global setpoint
 913	 *
 914	 * See comment for pos_ratio_polynom().
 915	 */
 916	setpoint = (freerun + limit) / 2;
 917	pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
 918
 919	/*
 920	 * The strictlimit feature is a tool preventing mistrusted filesystems
 921	 * from growing a large number of dirty pages before throttling. For
 922	 * such filesystems balance_dirty_pages always checks wb counters
 923	 * against wb limits. Even if global "nr_dirty" is under "freerun".
 924	 * This is especially important for fuse which sets bdi->max_ratio to
 925	 * 1% by default. Without strictlimit feature, fuse writeback may
 926	 * consume arbitrary amount of RAM because it is accounted in
 927	 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
 928	 *
 929	 * Here, in wb_position_ratio(), we calculate pos_ratio based on
 930	 * two values: wb_dirty and wb_thresh. Let's consider an example:
 931	 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
 932	 * limits are set by default to 10% and 20% (background and throttle).
 933	 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
 934	 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
 935	 * about ~6K pages (as the average of background and throttle wb
 936	 * limits). The 3rd order polynomial will provide positive feedback if
 937	 * wb_dirty is under wb_setpoint and vice versa.
 938	 *
 939	 * Note, that we cannot use global counters in these calculations
 940	 * because we want to throttle process writing to a strictlimit wb
 941	 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
 942	 * in the example above).
 943	 */
 944	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
 945		long long wb_pos_ratio;
 946
 947		if (dtc->wb_dirty < 8) {
 948			dtc->pos_ratio = min_t(long long, pos_ratio * 2,
 949					   2 << RATELIMIT_CALC_SHIFT);
 950			return;
 951		}
 952
 953		if (dtc->wb_dirty >= wb_thresh)
 954			return;
 955
 956		wb_setpoint = dirty_freerun_ceiling(wb_thresh,
 957						    dtc->wb_bg_thresh);
 958
 959		if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
 960			return;
 961
 962		wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
 963						 wb_thresh);
 964
 965		/*
 966		 * Typically, for strictlimit case, wb_setpoint << setpoint
 967		 * and pos_ratio >> wb_pos_ratio. In the other words global
 968		 * state ("dirty") is not limiting factor and we have to
 969		 * make decision based on wb counters. But there is an
 970		 * important case when global pos_ratio should get precedence:
 971		 * global limits are exceeded (e.g. due to activities on other
 972		 * wb's) while given strictlimit wb is below limit.
 973		 *
 974		 * "pos_ratio * wb_pos_ratio" would work for the case above,
 975		 * but it would look too non-natural for the case of all
 976		 * activity in the system coming from a single strictlimit wb
 977		 * with bdi->max_ratio == 100%.
 978		 *
 979		 * Note that min() below somewhat changes the dynamics of the
 980		 * control system. Normally, pos_ratio value can be well over 3
 981		 * (when globally we are at freerun and wb is well below wb
 982		 * setpoint). Now the maximum pos_ratio in the same situation
 983		 * is 2. We might want to tweak this if we observe the control
 984		 * system is too slow to adapt.
 985		 */
 986		dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
 987		return;
 988	}
 989
 990	/*
 991	 * We have computed basic pos_ratio above based on global situation. If
 992	 * the wb is over/under its share of dirty pages, we want to scale
 993	 * pos_ratio further down/up. That is done by the following mechanism.
 994	 */
 995
 996	/*
 997	 * wb setpoint
 998	 *
 999	 *        f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
1000	 *
1001	 *                        x_intercept - wb_dirty
1002	 *                     := --------------------------
1003	 *                        x_intercept - wb_setpoint
1004	 *
1005	 * The main wb control line is a linear function that subjects to
1006	 *
1007	 * (1) f(wb_setpoint) = 1.0
1008	 * (2) k = - 1 / (8 * write_bw)  (in single wb case)
1009	 *     or equally: x_intercept = wb_setpoint + 8 * write_bw
1010	 *
1011	 * For single wb case, the dirty pages are observed to fluctuate
1012	 * regularly within range
1013	 *        [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
1014	 * for various filesystems, where (2) can yield in a reasonable 12.5%
1015	 * fluctuation range for pos_ratio.
1016	 *
1017	 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
1018	 * own size, so move the slope over accordingly and choose a slope that
1019	 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
1020	 */
1021	if (unlikely(wb_thresh > dtc->thresh))
1022		wb_thresh = dtc->thresh;
1023	/*
1024	 * It's very possible that wb_thresh is close to 0 not because the
1025	 * device is slow, but that it has remained inactive for long time.
1026	 * Honour such devices a reasonable good (hopefully IO efficient)
1027	 * threshold, so that the occasional writes won't be blocked and active
1028	 * writes can rampup the threshold quickly.
1029	 */
1030	wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
1031	/*
1032	 * scale global setpoint to wb's:
1033	 *	wb_setpoint = setpoint * wb_thresh / thresh
1034	 */
1035	x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
1036	wb_setpoint = setpoint * (u64)x >> 16;
1037	/*
1038	 * Use span=(8*write_bw) in single wb case as indicated by
1039	 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
1040	 *
1041	 *        wb_thresh                    thresh - wb_thresh
1042	 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1043	 *         thresh                           thresh
1044	 */
1045	span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
1046	x_intercept = wb_setpoint + span;
1047
1048	if (dtc->wb_dirty < x_intercept - span / 4) {
1049		pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
1050				      (x_intercept - wb_setpoint) | 1);
1051	} else
1052		pos_ratio /= 4;
1053
1054	/*
1055	 * wb reserve area, safeguard against dirty pool underrun and disk idle
1056	 * It may push the desired control point of global dirty pages higher
1057	 * than setpoint.
1058	 */
1059	x_intercept = wb_thresh / 2;
1060	if (dtc->wb_dirty < x_intercept) {
1061		if (dtc->wb_dirty > x_intercept / 8)
1062			pos_ratio = div_u64(pos_ratio * x_intercept,
1063					    dtc->wb_dirty);
1064		else
1065			pos_ratio *= 8;
1066	}
1067
1068	dtc->pos_ratio = pos_ratio;
1069}
1070
1071static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1072				      unsigned long elapsed,
1073				      unsigned long written)
1074{
1075	const unsigned long period = roundup_pow_of_two(3 * HZ);
1076	unsigned long avg = wb->avg_write_bandwidth;
1077	unsigned long old = wb->write_bandwidth;
1078	u64 bw;
1079
1080	/*
1081	 * bw = written * HZ / elapsed
1082	 *
1083	 *                   bw * elapsed + write_bandwidth * (period - elapsed)
1084	 * write_bandwidth = ---------------------------------------------------
1085	 *                                          period
1086	 *
1087	 * @written may have decreased due to account_page_redirty().
1088	 * Avoid underflowing @bw calculation.
1089	 */
1090	bw = written - min(written, wb->written_stamp);
1091	bw *= HZ;
1092	if (unlikely(elapsed > period)) {
1093		bw = div64_ul(bw, elapsed);
1094		avg = bw;
1095		goto out;
1096	}
1097	bw += (u64)wb->write_bandwidth * (period - elapsed);
1098	bw >>= ilog2(period);
1099
1100	/*
1101	 * one more level of smoothing, for filtering out sudden spikes
1102	 */
1103	if (avg > old && old >= (unsigned long)bw)
1104		avg -= (avg - old) >> 3;
1105
1106	if (avg < old && old <= (unsigned long)bw)
1107		avg += (old - avg) >> 3;
1108
1109out:
1110	/* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1111	avg = max(avg, 1LU);
1112	if (wb_has_dirty_io(wb)) {
1113		long delta = avg - wb->avg_write_bandwidth;
1114		WARN_ON_ONCE(atomic_long_add_return(delta,
1115					&wb->bdi->tot_write_bandwidth) <= 0);
1116	}
1117	wb->write_bandwidth = bw;
1118	wb->avg_write_bandwidth = avg;
1119}
1120
1121static void update_dirty_limit(struct dirty_throttle_control *dtc)
1122{
1123	struct wb_domain *dom = dtc_dom(dtc);
1124	unsigned long thresh = dtc->thresh;
1125	unsigned long limit = dom->dirty_limit;
1126
1127	/*
1128	 * Follow up in one step.
1129	 */
1130	if (limit < thresh) {
1131		limit = thresh;
1132		goto update;
1133	}
1134
1135	/*
1136	 * Follow down slowly. Use the higher one as the target, because thresh
1137	 * may drop below dirty. This is exactly the reason to introduce
1138	 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
1139	 */
1140	thresh = max(thresh, dtc->dirty);
1141	if (limit > thresh) {
1142		limit -= (limit - thresh) >> 5;
1143		goto update;
1144	}
1145	return;
1146update:
1147	dom->dirty_limit = limit;
1148}
1149
1150static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
1151				    unsigned long now)
1152{
1153	struct wb_domain *dom = dtc_dom(dtc);
1154
1155	/*
1156	 * check locklessly first to optimize away locking for the most time
1157	 */
1158	if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
1159		return;
1160
1161	spin_lock(&dom->lock);
1162	if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
1163		update_dirty_limit(dtc);
1164		dom->dirty_limit_tstamp = now;
1165	}
1166	spin_unlock(&dom->lock);
1167}
1168
1169/*
1170 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
1171 *
1172 * Normal wb tasks will be curbed at or below it in long term.
1173 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1174 */
1175static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
1176				      unsigned long dirtied,
1177				      unsigned long elapsed)
1178{
1179	struct bdi_writeback *wb = dtc->wb;
1180	unsigned long dirty = dtc->dirty;
1181	unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
1182	unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
1183	unsigned long setpoint = (freerun + limit) / 2;
1184	unsigned long write_bw = wb->avg_write_bandwidth;
1185	unsigned long dirty_ratelimit = wb->dirty_ratelimit;
1186	unsigned long dirty_rate;
1187	unsigned long task_ratelimit;
1188	unsigned long balanced_dirty_ratelimit;
1189	unsigned long step;
1190	unsigned long x;
1191	unsigned long shift;
1192
1193	/*
1194	 * The dirty rate will match the writeout rate in long term, except
1195	 * when dirty pages are truncated by userspace or re-dirtied by FS.
1196	 */
1197	dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
1198
1199	/*
1200	 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1201	 */
1202	task_ratelimit = (u64)dirty_ratelimit *
1203					dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
1204	task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1205
1206	/*
1207	 * A linear estimation of the "balanced" throttle rate. The theory is,
1208	 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
1209	 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1210	 * formula will yield the balanced rate limit (write_bw / N).
1211	 *
1212	 * Note that the expanded form is not a pure rate feedback:
1213	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate)		     (1)
1214	 * but also takes pos_ratio into account:
1215	 *	rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1216	 *
1217	 * (1) is not realistic because pos_ratio also takes part in balancing
1218	 * the dirty rate.  Consider the state
1219	 *	pos_ratio = 0.5						     (3)
1220	 *	rate = 2 * (write_bw / N)				     (4)
1221	 * If (1) is used, it will stuck in that state! Because each dd will
1222	 * be throttled at
1223	 *	task_ratelimit = pos_ratio * rate = (write_bw / N)	     (5)
1224	 * yielding
1225	 *	dirty_rate = N * task_ratelimit = write_bw		     (6)
1226	 * put (6) into (1) we get
1227	 *	rate_(i+1) = rate_(i)					     (7)
1228	 *
1229	 * So we end up using (2) to always keep
1230	 *	rate_(i+1) ~= (write_bw / N)				     (8)
1231	 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1232	 * pos_ratio is able to drive itself to 1.0, which is not only where
1233	 * the dirty count meet the setpoint, but also where the slope of
1234	 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1235	 */
1236	balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1237					   dirty_rate | 1);
1238	/*
1239	 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1240	 */
1241	if (unlikely(balanced_dirty_ratelimit > write_bw))
1242		balanced_dirty_ratelimit = write_bw;
1243
1244	/*
1245	 * We could safely do this and return immediately:
1246	 *
1247	 *	wb->dirty_ratelimit = balanced_dirty_ratelimit;
1248	 *
1249	 * However to get a more stable dirty_ratelimit, the below elaborated
1250	 * code makes use of task_ratelimit to filter out singular points and
1251	 * limit the step size.
1252	 *
1253	 * The below code essentially only uses the relative value of
1254	 *
1255	 *	task_ratelimit - dirty_ratelimit
1256	 *	= (pos_ratio - 1) * dirty_ratelimit
1257	 *
1258	 * which reflects the direction and size of dirty position error.
1259	 */
1260
1261	/*
1262	 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1263	 * task_ratelimit is on the same side of dirty_ratelimit, too.
1264	 * For example, when
1265	 * - dirty_ratelimit > balanced_dirty_ratelimit
1266	 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1267	 * lowering dirty_ratelimit will help meet both the position and rate
1268	 * control targets. Otherwise, don't update dirty_ratelimit if it will
1269	 * only help meet the rate target. After all, what the users ultimately
1270	 * feel and care are stable dirty rate and small position error.
1271	 *
1272	 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1273	 * and filter out the singular points of balanced_dirty_ratelimit. Which
1274	 * keeps jumping around randomly and can even leap far away at times
1275	 * due to the small 200ms estimation period of dirty_rate (we want to
1276	 * keep that period small to reduce time lags).
1277	 */
1278	step = 0;
1279
1280	/*
1281	 * For strictlimit case, calculations above were based on wb counters
1282	 * and limits (starting from pos_ratio = wb_position_ratio() and up to
1283	 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1284	 * Hence, to calculate "step" properly, we have to use wb_dirty as
1285	 * "dirty" and wb_setpoint as "setpoint".
1286	 *
1287	 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1288	 * it's possible that wb_thresh is close to zero due to inactivity
1289	 * of backing device.
1290	 */
1291	if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1292		dirty = dtc->wb_dirty;
1293		if (dtc->wb_dirty < 8)
1294			setpoint = dtc->wb_dirty + 1;
1295		else
1296			setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
1297	}
1298
1299	if (dirty < setpoint) {
1300		x = min3(wb->balanced_dirty_ratelimit,
1301			 balanced_dirty_ratelimit, task_ratelimit);
1302		if (dirty_ratelimit < x)
1303			step = x - dirty_ratelimit;
1304	} else {
1305		x = max3(wb->balanced_dirty_ratelimit,
1306			 balanced_dirty_ratelimit, task_ratelimit);
1307		if (dirty_ratelimit > x)
1308			step = dirty_ratelimit - x;
1309	}
1310
1311	/*
1312	 * Don't pursue 100% rate matching. It's impossible since the balanced
1313	 * rate itself is constantly fluctuating. So decrease the track speed
1314	 * when it gets close to the target. Helps eliminate pointless tremors.
1315	 */
1316	shift = dirty_ratelimit / (2 * step + 1);
1317	if (shift < BITS_PER_LONG)
1318		step = DIV_ROUND_UP(step >> shift, 8);
1319	else
1320		step = 0;
1321
1322	if (dirty_ratelimit < balanced_dirty_ratelimit)
1323		dirty_ratelimit += step;
1324	else
1325		dirty_ratelimit -= step;
1326
1327	wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1328	wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1329
1330	trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
1331}
1332
1333static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1334				  struct dirty_throttle_control *mdtc,
1335				  unsigned long start_time,
1336				  bool update_ratelimit)
1337{
1338	struct bdi_writeback *wb = gdtc->wb;
1339	unsigned long now = jiffies;
1340	unsigned long elapsed = now - wb->bw_time_stamp;
1341	unsigned long dirtied;
1342	unsigned long written;
1343
1344	lockdep_assert_held(&wb->list_lock);
1345
1346	/*
1347	 * rate-limit, only update once every 200ms.
1348	 */
1349	if (elapsed < BANDWIDTH_INTERVAL)
1350		return;
1351
1352	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1353	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
1354
1355	/*
1356	 * Skip quiet periods when disk bandwidth is under-utilized.
1357	 * (at least 1s idle time between two flusher runs)
1358	 */
1359	if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
1360		goto snapshot;
1361
1362	if (update_ratelimit) {
1363		domain_update_bandwidth(gdtc, now);
1364		wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1365
1366		/*
1367		 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1368		 * compiler has no way to figure that out.  Help it.
1369		 */
1370		if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1371			domain_update_bandwidth(mdtc, now);
1372			wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1373		}
1374	}
1375	wb_update_write_bandwidth(wb, elapsed, written);
1376
1377snapshot:
1378	wb->dirtied_stamp = dirtied;
1379	wb->written_stamp = written;
1380	wb->bw_time_stamp = now;
1381}
1382
1383void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
1384{
1385	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1386
1387	__wb_update_bandwidth(&gdtc, NULL, start_time, false);
1388}
1389
1390/*
1391 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1392 * will look to see if it needs to start dirty throttling.
1393 *
1394 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1395 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
1396 * (the number of pages we may dirty without exceeding the dirty limits).
1397 */
1398static unsigned long dirty_poll_interval(unsigned long dirty,
1399					 unsigned long thresh)
1400{
1401	if (thresh > dirty)
1402		return 1UL << (ilog2(thresh - dirty) >> 1);
1403
1404	return 1;
1405}
1406
1407static unsigned long wb_max_pause(struct bdi_writeback *wb,
1408				  unsigned long wb_dirty)
1409{
1410	unsigned long bw = wb->avg_write_bandwidth;
1411	unsigned long t;
1412
1413	/*
1414	 * Limit pause time for small memory systems. If sleeping for too long
1415	 * time, a small pool of dirty/writeback pages may go empty and disk go
1416	 * idle.
1417	 *
1418	 * 8 serves as the safety ratio.
1419	 */
1420	t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1421	t++;
1422
1423	return min_t(unsigned long, t, MAX_PAUSE);
1424}
1425
1426static long wb_min_pause(struct bdi_writeback *wb,
1427			 long max_pause,
1428			 unsigned long task_ratelimit,
1429			 unsigned long dirty_ratelimit,
1430			 int *nr_dirtied_pause)
1431{
1432	long hi = ilog2(wb->avg_write_bandwidth);
1433	long lo = ilog2(wb->dirty_ratelimit);
1434	long t;		/* target pause */
1435	long pause;	/* estimated next pause */
1436	int pages;	/* target nr_dirtied_pause */
1437
1438	/* target for 10ms pause on 1-dd case */
1439	t = max(1, HZ / 100);
1440
1441	/*
1442	 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1443	 * overheads.
1444	 *
1445	 * (N * 10ms) on 2^N concurrent tasks.
1446	 */
1447	if (hi > lo)
1448		t += (hi - lo) * (10 * HZ) / 1024;
1449
1450	/*
1451	 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1452	 * on the much more stable dirty_ratelimit. However the next pause time
1453	 * will be computed based on task_ratelimit and the two rate limits may
1454	 * depart considerably at some time. Especially if task_ratelimit goes
1455	 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1456	 * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1457	 * result task_ratelimit won't be executed faithfully, which could
1458	 * eventually bring down dirty_ratelimit.
1459	 *
1460	 * We apply two rules to fix it up:
1461	 * 1) try to estimate the next pause time and if necessary, use a lower
1462	 *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1463	 *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1464	 * 2) limit the target pause time to max_pause/2, so that the normal
1465	 *    small fluctuations of task_ratelimit won't trigger rule (1) and
1466	 *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1467	 */
1468	t = min(t, 1 + max_pause / 2);
1469	pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1470
1471	/*
1472	 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1473	 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1474	 * When the 16 consecutive reads are often interrupted by some dirty
1475	 * throttling pause during the async writes, cfq will go into idles
1476	 * (deadline is fine). So push nr_dirtied_pause as high as possible
1477	 * until reaches DIRTY_POLL_THRESH=32 pages.
1478	 */
1479	if (pages < DIRTY_POLL_THRESH) {
1480		t = max_pause;
1481		pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1482		if (pages > DIRTY_POLL_THRESH) {
1483			pages = DIRTY_POLL_THRESH;
1484			t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1485		}
1486	}
1487
1488	pause = HZ * pages / (task_ratelimit + 1);
1489	if (pause > max_pause) {
1490		t = max_pause;
1491		pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1492	}
1493
1494	*nr_dirtied_pause = pages;
1495	/*
1496	 * The minimal pause time will normally be half the target pause time.
1497	 */
1498	return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1499}
1500
1501static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
1502{
1503	struct bdi_writeback *wb = dtc->wb;
1504	unsigned long wb_reclaimable;
1505
1506	/*
1507	 * wb_thresh is not treated as some limiting factor as
1508	 * dirty_thresh, due to reasons
1509	 * - in JBOD setup, wb_thresh can fluctuate a lot
1510	 * - in a system with HDD and USB key, the USB key may somehow
1511	 *   go into state (wb_dirty >> wb_thresh) either because
1512	 *   wb_dirty starts high, or because wb_thresh drops low.
1513	 *   In this case we don't want to hard throttle the USB key
1514	 *   dirtiers for 100 seconds until wb_dirty drops under
1515	 *   wb_thresh. Instead the auxiliary wb control line in
1516	 *   wb_position_ratio() will let the dirtier task progress
1517	 *   at some rate <= (write_bw / 2) for bringing down wb_dirty.
1518	 */
1519	dtc->wb_thresh = __wb_calc_thresh(dtc);
1520	dtc->wb_bg_thresh = dtc->thresh ?
1521		div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
1522
1523	/*
1524	 * In order to avoid the stacked BDI deadlock we need
1525	 * to ensure we accurately count the 'dirty' pages when
1526	 * the threshold is low.
1527	 *
1528	 * Otherwise it would be possible to get thresh+n pages
1529	 * reported dirty, even though there are thresh-m pages
1530	 * actually dirty; with m+n sitting in the percpu
1531	 * deltas.
1532	 */
1533	if (dtc->wb_thresh < 2 * wb_stat_error()) {
1534		wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1535		dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
1536	} else {
1537		wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1538		dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
1539	}
1540}
1541
1542/*
1543 * balance_dirty_pages() must be called by processes which are generating dirty
1544 * data.  It looks at the number of dirty pages in the machine and will force
1545 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1546 * If we're over `background_thresh' then the writeback threads are woken to
1547 * perform some writeout.
1548 */
1549static void balance_dirty_pages(struct bdi_writeback *wb,
1550				unsigned long pages_dirtied)
1551{
1552	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1553	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1554	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1555	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1556						     &mdtc_stor : NULL;
1557	struct dirty_throttle_control *sdtc;
1558	unsigned long nr_reclaimable;	/* = file_dirty */
1559	long period;
1560	long pause;
1561	long max_pause;
1562	long min_pause;
1563	int nr_dirtied_pause;
1564	bool dirty_exceeded = false;
1565	unsigned long task_ratelimit;
1566	unsigned long dirty_ratelimit;
1567	struct backing_dev_info *bdi = wb->bdi;
1568	bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1569	unsigned long start_time = jiffies;
1570
1571	for (;;) {
1572		unsigned long now = jiffies;
1573		unsigned long dirty, thresh, bg_thresh;
1574		unsigned long m_dirty = 0;	/* stop bogus uninit warnings */
1575		unsigned long m_thresh = 0;
1576		unsigned long m_bg_thresh = 0;
1577
1578		nr_reclaimable = global_node_page_state(NR_FILE_DIRTY);
 
 
 
 
 
 
 
1579		gdtc->avail = global_dirtyable_memory();
1580		gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
1581
1582		domain_dirty_limits(gdtc);
1583
1584		if (unlikely(strictlimit)) {
1585			wb_dirty_limits(gdtc);
1586
1587			dirty = gdtc->wb_dirty;
1588			thresh = gdtc->wb_thresh;
1589			bg_thresh = gdtc->wb_bg_thresh;
1590		} else {
1591			dirty = gdtc->dirty;
1592			thresh = gdtc->thresh;
1593			bg_thresh = gdtc->bg_thresh;
1594		}
1595
1596		if (mdtc) {
1597			unsigned long filepages, headroom, writeback;
1598
1599			/*
1600			 * If @wb belongs to !root memcg, repeat the same
1601			 * basic calculations for the memcg domain.
1602			 */
1603			mem_cgroup_wb_stats(wb, &filepages, &headroom,
1604					    &mdtc->dirty, &writeback);
1605			mdtc->dirty += writeback;
1606			mdtc_calc_avail(mdtc, filepages, headroom);
1607
1608			domain_dirty_limits(mdtc);
1609
1610			if (unlikely(strictlimit)) {
1611				wb_dirty_limits(mdtc);
1612				m_dirty = mdtc->wb_dirty;
1613				m_thresh = mdtc->wb_thresh;
1614				m_bg_thresh = mdtc->wb_bg_thresh;
1615			} else {
1616				m_dirty = mdtc->dirty;
1617				m_thresh = mdtc->thresh;
1618				m_bg_thresh = mdtc->bg_thresh;
1619			}
1620		}
1621
1622		/*
1623		 * Throttle it only when the background writeback cannot
1624		 * catch-up. This avoids (excessively) small writeouts
1625		 * when the wb limits are ramping up in case of !strictlimit.
1626		 *
1627		 * In strictlimit case make decision based on the wb counters
1628		 * and limits. Small writeouts when the wb limits are ramping
1629		 * up are the price we consciously pay for strictlimit-ing.
1630		 *
1631		 * If memcg domain is in effect, @dirty should be under
1632		 * both global and memcg freerun ceilings.
1633		 */
1634		if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1635		    (!mdtc ||
1636		     m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1637			unsigned long intv;
1638			unsigned long m_intv;
1639
1640free_running:
1641			intv = dirty_poll_interval(dirty, thresh);
1642			m_intv = ULONG_MAX;
1643
1644			current->dirty_paused_when = now;
1645			current->nr_dirtied = 0;
1646			if (mdtc)
1647				m_intv = dirty_poll_interval(m_dirty, m_thresh);
1648			current->nr_dirtied_pause = min(intv, m_intv);
1649			break;
1650		}
1651
1652		if (unlikely(!writeback_in_progress(wb)))
1653			wb_start_background_writeback(wb);
1654
1655		mem_cgroup_flush_foreign(wb);
1656
1657		/*
1658		 * Calculate global domain's pos_ratio and select the
1659		 * global dtc by default.
1660		 */
1661		if (!strictlimit) {
1662			wb_dirty_limits(gdtc);
1663
1664			if ((current->flags & PF_LOCAL_THROTTLE) &&
1665			    gdtc->wb_dirty <
1666			    dirty_freerun_ceiling(gdtc->wb_thresh,
1667						  gdtc->wb_bg_thresh))
1668				/*
1669				 * LOCAL_THROTTLE tasks must not be throttled
1670				 * when below the per-wb freerun ceiling.
1671				 */
1672				goto free_running;
1673		}
1674
1675		dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1676			((gdtc->dirty > gdtc->thresh) || strictlimit);
1677
1678		wb_position_ratio(gdtc);
1679		sdtc = gdtc;
1680
1681		if (mdtc) {
1682			/*
1683			 * If memcg domain is in effect, calculate its
1684			 * pos_ratio.  @wb should satisfy constraints from
1685			 * both global and memcg domains.  Choose the one
1686			 * w/ lower pos_ratio.
1687			 */
1688			if (!strictlimit) {
1689				wb_dirty_limits(mdtc);
1690
1691				if ((current->flags & PF_LOCAL_THROTTLE) &&
1692				    mdtc->wb_dirty <
1693				    dirty_freerun_ceiling(mdtc->wb_thresh,
1694							  mdtc->wb_bg_thresh))
1695					/*
1696					 * LOCAL_THROTTLE tasks must not be
1697					 * throttled when below the per-wb
1698					 * freerun ceiling.
1699					 */
1700					goto free_running;
1701			}
1702			dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1703				((mdtc->dirty > mdtc->thresh) || strictlimit);
1704
1705			wb_position_ratio(mdtc);
1706			if (mdtc->pos_ratio < gdtc->pos_ratio)
1707				sdtc = mdtc;
1708		}
1709
1710		if (dirty_exceeded && !wb->dirty_exceeded)
1711			wb->dirty_exceeded = 1;
1712
1713		if (time_is_before_jiffies(wb->bw_time_stamp +
1714					   BANDWIDTH_INTERVAL)) {
1715			spin_lock(&wb->list_lock);
1716			__wb_update_bandwidth(gdtc, mdtc, start_time, true);
1717			spin_unlock(&wb->list_lock);
1718		}
1719
1720		/* throttle according to the chosen dtc */
1721		dirty_ratelimit = wb->dirty_ratelimit;
1722		task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
1723							RATELIMIT_CALC_SHIFT;
1724		max_pause = wb_max_pause(wb, sdtc->wb_dirty);
1725		min_pause = wb_min_pause(wb, max_pause,
1726					 task_ratelimit, dirty_ratelimit,
1727					 &nr_dirtied_pause);
1728
1729		if (unlikely(task_ratelimit == 0)) {
1730			period = max_pause;
1731			pause = max_pause;
1732			goto pause;
1733		}
1734		period = HZ * pages_dirtied / task_ratelimit;
1735		pause = period;
1736		if (current->dirty_paused_when)
1737			pause -= now - current->dirty_paused_when;
1738		/*
1739		 * For less than 1s think time (ext3/4 may block the dirtier
1740		 * for up to 800ms from time to time on 1-HDD; so does xfs,
1741		 * however at much less frequency), try to compensate it in
1742		 * future periods by updating the virtual time; otherwise just
1743		 * do a reset, as it may be a light dirtier.
1744		 */
1745		if (pause < min_pause) {
1746			trace_balance_dirty_pages(wb,
1747						  sdtc->thresh,
1748						  sdtc->bg_thresh,
1749						  sdtc->dirty,
1750						  sdtc->wb_thresh,
1751						  sdtc->wb_dirty,
1752						  dirty_ratelimit,
1753						  task_ratelimit,
1754						  pages_dirtied,
1755						  period,
1756						  min(pause, 0L),
1757						  start_time);
1758			if (pause < -HZ) {
1759				current->dirty_paused_when = now;
1760				current->nr_dirtied = 0;
1761			} else if (period) {
1762				current->dirty_paused_when += period;
1763				current->nr_dirtied = 0;
1764			} else if (current->nr_dirtied_pause <= pages_dirtied)
1765				current->nr_dirtied_pause += pages_dirtied;
1766			break;
1767		}
1768		if (unlikely(pause > max_pause)) {
1769			/* for occasional dropped task_ratelimit */
1770			now += min(pause - max_pause, max_pause);
1771			pause = max_pause;
1772		}
1773
1774pause:
1775		trace_balance_dirty_pages(wb,
1776					  sdtc->thresh,
1777					  sdtc->bg_thresh,
1778					  sdtc->dirty,
1779					  sdtc->wb_thresh,
1780					  sdtc->wb_dirty,
1781					  dirty_ratelimit,
1782					  task_ratelimit,
1783					  pages_dirtied,
1784					  period,
1785					  pause,
1786					  start_time);
1787		__set_current_state(TASK_KILLABLE);
1788		wb->dirty_sleep = now;
1789		io_schedule_timeout(pause);
1790
1791		current->dirty_paused_when = now + pause;
1792		current->nr_dirtied = 0;
1793		current->nr_dirtied_pause = nr_dirtied_pause;
1794
1795		/*
1796		 * This is typically equal to (dirty < thresh) and can also
1797		 * keep "1000+ dd on a slow USB stick" under control.
1798		 */
1799		if (task_ratelimit)
1800			break;
1801
1802		/*
1803		 * In the case of an unresponsive NFS server and the NFS dirty
1804		 * pages exceeds dirty_thresh, give the other good wb's a pipe
1805		 * to go through, so that tasks on them still remain responsive.
1806		 *
1807		 * In theory 1 page is enough to keep the consumer-producer
1808		 * pipe going: the flusher cleans 1 page => the task dirties 1
1809		 * more page. However wb_dirty has accounting errors.  So use
1810		 * the larger and more IO friendly wb_stat_error.
1811		 */
1812		if (sdtc->wb_dirty <= wb_stat_error())
1813			break;
1814
1815		if (fatal_signal_pending(current))
1816			break;
1817	}
1818
1819	if (!dirty_exceeded && wb->dirty_exceeded)
1820		wb->dirty_exceeded = 0;
1821
1822	if (writeback_in_progress(wb))
1823		return;
1824
1825	/*
1826	 * In laptop mode, we wait until hitting the higher threshold before
1827	 * starting background writeout, and then write out all the way down
1828	 * to the lower threshold.  So slow writers cause minimal disk activity.
1829	 *
1830	 * In normal mode, we start background writeout at the lower
1831	 * background_thresh, to keep the amount of dirty memory low.
1832	 */
1833	if (laptop_mode)
1834		return;
1835
1836	if (nr_reclaimable > gdtc->bg_thresh)
1837		wb_start_background_writeback(wb);
1838}
1839
1840static DEFINE_PER_CPU(int, bdp_ratelimits);
1841
1842/*
1843 * Normal tasks are throttled by
1844 *	loop {
1845 *		dirty tsk->nr_dirtied_pause pages;
1846 *		take a snap in balance_dirty_pages();
1847 *	}
1848 * However there is a worst case. If every task exit immediately when dirtied
1849 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1850 * called to throttle the page dirties. The solution is to save the not yet
1851 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1852 * randomly into the running tasks. This works well for the above worst case,
1853 * as the new task will pick up and accumulate the old task's leaked dirty
1854 * count and eventually get throttled.
1855 */
1856DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1857
1858/**
1859 * balance_dirty_pages_ratelimited - balance dirty memory state
1860 * @mapping: address_space which was dirtied
1861 *
1862 * Processes which are dirtying memory should call in here once for each page
1863 * which was newly dirtied.  The function will periodically check the system's
1864 * dirty state and will initiate writeback if needed.
1865 *
1866 * Once we're over the dirty memory limit we decrease the ratelimiting
1867 * by a lot, to prevent individual processes from overshooting the limit
1868 * by (ratelimit_pages) each.
 
1869 */
1870void balance_dirty_pages_ratelimited(struct address_space *mapping)
1871{
1872	struct inode *inode = mapping->host;
1873	struct backing_dev_info *bdi = inode_to_bdi(inode);
1874	struct bdi_writeback *wb = NULL;
1875	int ratelimit;
1876	int *p;
1877
1878	if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
1879		return;
1880
1881	if (inode_cgwb_enabled(inode))
1882		wb = wb_get_create_current(bdi, GFP_KERNEL);
1883	if (!wb)
1884		wb = &bdi->wb;
1885
1886	ratelimit = current->nr_dirtied_pause;
1887	if (wb->dirty_exceeded)
1888		ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1889
1890	preempt_disable();
1891	/*
1892	 * This prevents one CPU to accumulate too many dirtied pages without
1893	 * calling into balance_dirty_pages(), which can happen when there are
1894	 * 1000+ tasks, all of them start dirtying pages at exactly the same
1895	 * time, hence all honoured too large initial task->nr_dirtied_pause.
1896	 */
1897	p =  this_cpu_ptr(&bdp_ratelimits);
1898	if (unlikely(current->nr_dirtied >= ratelimit))
1899		*p = 0;
1900	else if (unlikely(*p >= ratelimit_pages)) {
1901		*p = 0;
1902		ratelimit = 0;
1903	}
1904	/*
1905	 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1906	 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1907	 * the dirty throttling and livelock other long-run dirtiers.
1908	 */
1909	p = this_cpu_ptr(&dirty_throttle_leaks);
1910	if (*p > 0 && current->nr_dirtied < ratelimit) {
1911		unsigned long nr_pages_dirtied;
1912		nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1913		*p -= nr_pages_dirtied;
1914		current->nr_dirtied += nr_pages_dirtied;
1915	}
1916	preempt_enable();
1917
1918	if (unlikely(current->nr_dirtied >= ratelimit))
1919		balance_dirty_pages(wb, current->nr_dirtied);
1920
1921	wb_put(wb);
1922}
1923EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1924
1925/**
1926 * wb_over_bg_thresh - does @wb need to be written back?
1927 * @wb: bdi_writeback of interest
1928 *
1929 * Determines whether background writeback should keep writing @wb or it's
1930 * clean enough.
1931 *
1932 * Return: %true if writeback should continue.
1933 */
1934bool wb_over_bg_thresh(struct bdi_writeback *wb)
1935{
1936	struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
1937	struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
1938	struct dirty_throttle_control * const gdtc = &gdtc_stor;
1939	struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1940						     &mdtc_stor : NULL;
1941	unsigned long reclaimable;
1942	unsigned long thresh;
1943
1944	/*
1945	 * Similar to balance_dirty_pages() but ignores pages being written
1946	 * as we're trying to decide whether to put more under writeback.
1947	 */
1948	gdtc->avail = global_dirtyable_memory();
1949	gdtc->dirty = global_node_page_state(NR_FILE_DIRTY);
 
1950	domain_dirty_limits(gdtc);
1951
1952	if (gdtc->dirty > gdtc->bg_thresh)
1953		return true;
1954
1955	thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh);
1956	if (thresh < 2 * wb_stat_error())
1957		reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1958	else
1959		reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1960
1961	if (reclaimable > thresh)
1962		return true;
1963
1964	if (mdtc) {
1965		unsigned long filepages, headroom, writeback;
1966
1967		mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1968				    &writeback);
1969		mdtc_calc_avail(mdtc, filepages, headroom);
1970		domain_dirty_limits(mdtc);	/* ditto, ignore writeback */
1971
1972		if (mdtc->dirty > mdtc->bg_thresh)
1973			return true;
1974
1975		thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh);
1976		if (thresh < 2 * wb_stat_error())
1977			reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
1978		else
1979			reclaimable = wb_stat(wb, WB_RECLAIMABLE);
1980
1981		if (reclaimable > thresh)
1982			return true;
1983	}
1984
1985	return false;
1986}
1987
1988/*
1989 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1990 */
1991int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
1992		void *buffer, size_t *length, loff_t *ppos)
1993{
1994	unsigned int old_interval = dirty_writeback_interval;
1995	int ret;
1996
1997	ret = proc_dointvec(table, write, buffer, length, ppos);
1998
1999	/*
2000	 * Writing 0 to dirty_writeback_interval will disable periodic writeback
2001	 * and a different non-zero value will wakeup the writeback threads.
2002	 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
2003	 * iterate over all bdis and wbs.
2004	 * The reason we do this is to make the change take effect immediately.
2005	 */
2006	if (!ret && write && dirty_writeback_interval &&
2007		dirty_writeback_interval != old_interval)
2008		wakeup_flusher_threads(WB_REASON_PERIODIC);
2009
2010	return ret;
2011}
2012
2013#ifdef CONFIG_BLOCK
2014void laptop_mode_timer_fn(struct timer_list *t)
2015{
2016	struct backing_dev_info *backing_dev_info =
2017		from_timer(backing_dev_info, t, laptop_mode_wb_timer);
2018
2019	wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
2020}
2021
2022/*
2023 * We've spun up the disk and we're in laptop mode: schedule writeback
2024 * of all dirty data a few seconds from now.  If the flush is already scheduled
2025 * then push it back - the user is still using the disk.
2026 */
2027void laptop_io_completion(struct backing_dev_info *info)
2028{
2029	mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
2030}
2031
2032/*
2033 * We're in laptop mode and we've just synced. The sync's writes will have
2034 * caused another writeback to be scheduled by laptop_io_completion.
2035 * Nothing needs to be written back anymore, so we unschedule the writeback.
2036 */
2037void laptop_sync_completion(void)
2038{
2039	struct backing_dev_info *bdi;
2040
2041	rcu_read_lock();
2042
2043	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2044		del_timer(&bdi->laptop_mode_wb_timer);
2045
2046	rcu_read_unlock();
2047}
2048#endif
2049
2050/*
2051 * If ratelimit_pages is too high then we can get into dirty-data overload
2052 * if a large number of processes all perform writes at the same time.
 
 
2053 *
2054 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2055 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
2056 * thresholds.
2057 */
2058
2059void writeback_set_ratelimit(void)
2060{
2061	struct wb_domain *dom = &global_wb_domain;
2062	unsigned long background_thresh;
2063	unsigned long dirty_thresh;
2064
2065	global_dirty_limits(&background_thresh, &dirty_thresh);
2066	dom->dirty_limit = dirty_thresh;
2067	ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
2068	if (ratelimit_pages < 16)
2069		ratelimit_pages = 16;
2070}
2071
2072static int page_writeback_cpu_online(unsigned int cpu)
2073{
2074	writeback_set_ratelimit();
2075	return 0;
2076}
2077
2078/*
2079 * Called early on to tune the page writeback dirty limits.
2080 *
2081 * We used to scale dirty pages according to how total memory
2082 * related to pages that could be allocated for buffers.
 
2083 *
2084 * However, that was when we used "dirty_ratio" to scale with
2085 * all memory, and we don't do that any more. "dirty_ratio"
2086 * is now applied to total non-HIGHPAGE memory, and as such we can't
 
2087 * get into the old insane situation any more where we had
2088 * large amounts of dirty pages compared to a small amount of
2089 * non-HIGHMEM memory.
2090 *
2091 * But we might still want to scale the dirty_ratio by how
2092 * much memory the box has..
2093 */
2094void __init page_writeback_init(void)
2095{
2096	BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2097
2098	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2099			  page_writeback_cpu_online, NULL);
2100	cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2101			  page_writeback_cpu_online);
2102}
2103
2104/**
2105 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2106 * @mapping: address space structure to write
2107 * @start: starting page index
2108 * @end: ending page index (inclusive)
2109 *
2110 * This function scans the page range from @start to @end (inclusive) and tags
2111 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2112 * that write_cache_pages (or whoever calls this function) will then use
2113 * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
2114 * used to avoid livelocking of writeback by a process steadily creating new
2115 * dirty pages in the file (thus it is important for this function to be quick
2116 * so that it can tag pages faster than a dirtying process can create them).
2117 */
 
 
 
 
2118void tag_pages_for_writeback(struct address_space *mapping,
2119			     pgoff_t start, pgoff_t end)
2120{
2121	XA_STATE(xas, &mapping->i_pages, start);
2122	unsigned int tagged = 0;
2123	void *page;
2124
2125	xas_lock_irq(&xas);
2126	xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
2127		xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
2128		if (++tagged % XA_CHECK_SCHED)
 
 
 
 
 
 
2129			continue;
2130
2131		xas_pause(&xas);
2132		xas_unlock_irq(&xas);
2133		cond_resched();
2134		xas_lock_irq(&xas);
2135	}
2136	xas_unlock_irq(&xas);
2137}
2138EXPORT_SYMBOL(tag_pages_for_writeback);
2139
2140/**
2141 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2142 * @mapping: address space structure to write
2143 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2144 * @writepage: function called for each page
2145 * @data: data passed to writepage function
2146 *
2147 * If a page is already under I/O, write_cache_pages() skips it, even
2148 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
2149 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
2150 * and msync() need to guarantee that all the data which was dirty at the time
2151 * the call was made get new I/O started against them.  If wbc->sync_mode is
2152 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2153 * existing IO to complete.
2154 *
2155 * To avoid livelocks (when other process dirties new pages), we first tag
2156 * pages which should be written back with TOWRITE tag and only then start
2157 * writing them. For data-integrity sync we have to be careful so that we do
2158 * not miss some pages (e.g., because some other process has cleared TOWRITE
2159 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2160 * by the process clearing the DIRTY tag (and submitting the page for IO).
2161 *
2162 * To avoid deadlocks between range_cyclic writeback and callers that hold
2163 * pages in PageWriteback to aggregate IO until write_cache_pages() returns,
2164 * we do not loop back to the start of the file. Doing so causes a page
2165 * lock/page writeback access order inversion - we should only ever lock
2166 * multiple pages in ascending page->index order, and looping back to the start
2167 * of the file violates that rule and causes deadlocks.
2168 *
2169 * Return: %0 on success, negative error code otherwise
2170 */
2171int write_cache_pages(struct address_space *mapping,
2172		      struct writeback_control *wbc, writepage_t writepage,
2173		      void *data)
2174{
2175	int ret = 0;
2176	int done = 0;
2177	int error;
2178	struct pagevec pvec;
2179	int nr_pages;
 
2180	pgoff_t index;
2181	pgoff_t end;		/* Inclusive */
2182	pgoff_t done_index;
 
2183	int range_whole = 0;
2184	xa_mark_t tag;
2185
2186	pagevec_init(&pvec);
2187	if (wbc->range_cyclic) {
2188		index = mapping->writeback_index; /* prev offset */
 
 
 
 
 
2189		end = -1;
2190	} else {
2191		index = wbc->range_start >> PAGE_SHIFT;
2192		end = wbc->range_end >> PAGE_SHIFT;
2193		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2194			range_whole = 1;
 
2195	}
2196	if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) {
2197		tag_pages_for_writeback(mapping, index, end);
2198		tag = PAGECACHE_TAG_TOWRITE;
2199	} else {
2200		tag = PAGECACHE_TAG_DIRTY;
2201	}
 
 
2202	done_index = index;
2203	while (!done && (index <= end)) {
2204		int i;
2205
2206		nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2207				tag);
2208		if (nr_pages == 0)
2209			break;
2210
2211		for (i = 0; i < nr_pages; i++) {
2212			struct page *page = pvec.pages[i];
2213
2214			done_index = page->index;
2215
2216			lock_page(page);
2217
2218			/*
2219			 * Page truncated or invalidated. We can freely skip it
2220			 * then, even for data integrity operations: the page
2221			 * has disappeared concurrently, so there could be no
2222			 * real expectation of this data integrity operation
2223			 * even if there is now a new, dirty page at the same
2224			 * pagecache address.
2225			 */
2226			if (unlikely(page->mapping != mapping)) {
2227continue_unlock:
2228				unlock_page(page);
2229				continue;
2230			}
2231
2232			if (!PageDirty(page)) {
2233				/* someone wrote it for us */
2234				goto continue_unlock;
2235			}
2236
2237			if (PageWriteback(page)) {
2238				if (wbc->sync_mode != WB_SYNC_NONE)
2239					wait_on_page_writeback(page);
2240				else
2241					goto continue_unlock;
2242			}
2243
2244			BUG_ON(PageWriteback(page));
2245			if (!clear_page_dirty_for_io(page))
2246				goto continue_unlock;
2247
2248			trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
2249			error = (*writepage)(page, wbc, data);
2250			if (unlikely(error)) {
2251				/*
2252				 * Handle errors according to the type of
2253				 * writeback. There's no need to continue for
2254				 * background writeback. Just push done_index
2255				 * past this page so media errors won't choke
2256				 * writeout for the entire file. For integrity
2257				 * writeback, we must process the entire dirty
2258				 * set regardless of errors because the fs may
2259				 * still have state to clear for each page. In
2260				 * that case we continue processing and return
2261				 * the first error.
2262				 */
2263				if (error == AOP_WRITEPAGE_ACTIVATE) {
2264					unlock_page(page);
2265					error = 0;
2266				} else if (wbc->sync_mode != WB_SYNC_ALL) {
2267					ret = error;
 
 
 
 
 
 
 
 
2268					done_index = page->index + 1;
2269					done = 1;
2270					break;
2271				}
2272				if (!ret)
2273					ret = error;
2274			}
2275
2276			/*
2277			 * We stop writing back only if we are not doing
2278			 * integrity sync. In case of integrity sync we have to
2279			 * keep going until we have written all the pages
2280			 * we tagged for writeback prior to entering this loop.
2281			 */
2282			if (--wbc->nr_to_write <= 0 &&
2283			    wbc->sync_mode == WB_SYNC_NONE) {
2284				done = 1;
2285				break;
2286			}
2287		}
2288		pagevec_release(&pvec);
2289		cond_resched();
2290	}
2291
2292	/*
2293	 * If we hit the last page and there is more work to be done: wrap
2294	 * back the index back to the start of the file for the next
2295	 * time we are called.
2296	 */
2297	if (wbc->range_cyclic && !done)
2298		done_index = 0;
 
 
 
2299	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2300		mapping->writeback_index = done_index;
2301
2302	return ret;
2303}
2304EXPORT_SYMBOL(write_cache_pages);
2305
2306/*
2307 * Function used by generic_writepages to call the real writepage
2308 * function and set the mapping flags on error
2309 */
2310static int __writepage(struct page *page, struct writeback_control *wbc,
2311		       void *data)
2312{
2313	struct address_space *mapping = data;
2314	int ret = mapping->a_ops->writepage(page, wbc);
2315	mapping_set_error(mapping, ret);
2316	return ret;
2317}
2318
2319/**
2320 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2321 * @mapping: address space structure to write
2322 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2323 *
2324 * This is a library function, which implements the writepages()
2325 * address_space_operation.
2326 *
2327 * Return: %0 on success, negative error code otherwise
2328 */
2329int generic_writepages(struct address_space *mapping,
2330		       struct writeback_control *wbc)
2331{
2332	struct blk_plug plug;
2333	int ret;
2334
2335	/* deal with chardevs and other special file */
2336	if (!mapping->a_ops->writepage)
2337		return 0;
2338
2339	blk_start_plug(&plug);
2340	ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2341	blk_finish_plug(&plug);
2342	return ret;
2343}
2344
2345EXPORT_SYMBOL(generic_writepages);
2346
2347int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2348{
2349	int ret;
2350
2351	if (wbc->nr_to_write <= 0)
2352		return 0;
2353	while (1) {
2354		if (mapping->a_ops->writepages)
2355			ret = mapping->a_ops->writepages(mapping, wbc);
2356		else
2357			ret = generic_writepages(mapping, wbc);
2358		if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2359			break;
2360		cond_resched();
2361		congestion_wait(BLK_RW_ASYNC, HZ/50);
2362	}
2363	return ret;
2364}
2365
2366/**
2367 * write_one_page - write out a single page and wait on I/O
2368 * @page: the page to write
2369 *
2370 * The page must be locked by the caller and will be unlocked upon return.
2371 *
2372 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2373 * function returns.
2374 *
2375 * Return: %0 on success, negative error code otherwise
2376 */
2377int write_one_page(struct page *page)
2378{
2379	struct address_space *mapping = page->mapping;
2380	int ret = 0;
2381	struct writeback_control wbc = {
2382		.sync_mode = WB_SYNC_ALL,
2383		.nr_to_write = 1,
2384	};
2385
2386	BUG_ON(!PageLocked(page));
2387
2388	wait_on_page_writeback(page);
2389
2390	if (clear_page_dirty_for_io(page)) {
2391		get_page(page);
2392		ret = mapping->a_ops->writepage(page, &wbc);
2393		if (ret == 0)
2394			wait_on_page_writeback(page);
2395		put_page(page);
2396	} else {
2397		unlock_page(page);
2398	}
2399
2400	if (!ret)
2401		ret = filemap_check_errors(mapping);
2402	return ret;
2403}
2404EXPORT_SYMBOL(write_one_page);
2405
2406/*
2407 * For address_spaces which do not use buffers nor write back.
2408 */
2409int __set_page_dirty_no_writeback(struct page *page)
2410{
2411	if (!PageDirty(page))
2412		return !TestSetPageDirty(page);
2413	return 0;
2414}
2415EXPORT_SYMBOL(__set_page_dirty_no_writeback);
2416
2417/*
2418 * Helper function for set_page_dirty family.
2419 *
2420 * Caller must hold lock_page_memcg().
2421 *
2422 * NOTE: This relies on being atomic wrt interrupts.
2423 */
2424static void account_page_dirtied(struct page *page,
2425		struct address_space *mapping)
2426{
2427	struct inode *inode = mapping->host;
2428
2429	trace_writeback_dirty_page(page, mapping);
2430
2431	if (mapping_can_writeback(mapping)) {
2432		struct bdi_writeback *wb;
2433
2434		inode_attach_wb(inode, page);
2435		wb = inode_to_wb(inode);
2436
2437		__inc_lruvec_page_state(page, NR_FILE_DIRTY);
2438		__inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2439		__inc_node_page_state(page, NR_DIRTIED);
2440		inc_wb_stat(wb, WB_RECLAIMABLE);
2441		inc_wb_stat(wb, WB_DIRTIED);
2442		task_io_account_write(PAGE_SIZE);
2443		current->nr_dirtied++;
2444		__this_cpu_inc(bdp_ratelimits);
2445
2446		mem_cgroup_track_foreign_dirty(page, wb);
2447	}
2448}
 
2449
2450/*
2451 * Helper function for deaccounting dirty page without writeback.
2452 *
2453 * Caller must hold lock_page_memcg().
2454 */
2455void account_page_cleaned(struct page *page, struct address_space *mapping,
2456			  struct bdi_writeback *wb)
2457{
2458	if (mapping_can_writeback(mapping)) {
2459		dec_lruvec_page_state(page, NR_FILE_DIRTY);
2460		dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2461		dec_wb_stat(wb, WB_RECLAIMABLE);
2462		task_io_account_cancelled_write(PAGE_SIZE);
2463	}
2464}
2465
2466/*
2467 * Mark the page dirty, and set it dirty in the page cache, and mark the inode
2468 * dirty.
2469 *
2470 * If warn is true, then emit a warning if the page is not uptodate and has
2471 * not been truncated.
2472 *
2473 * The caller must hold lock_page_memcg().
2474 */
2475void __set_page_dirty(struct page *page, struct address_space *mapping,
2476			     int warn)
2477{
2478	unsigned long flags;
2479
2480	xa_lock_irqsave(&mapping->i_pages, flags);
2481	if (page->mapping) {	/* Race with truncate? */
2482		WARN_ON_ONCE(warn && !PageUptodate(page));
2483		account_page_dirtied(page, mapping);
2484		__xa_set_mark(&mapping->i_pages, page_index(page),
2485				PAGECACHE_TAG_DIRTY);
2486	}
2487	xa_unlock_irqrestore(&mapping->i_pages, flags);
2488}
2489
2490/*
2491 * For address_spaces which do not use buffers.  Just tag the page as dirty in
2492 * the xarray.
2493 *
2494 * This is also used when a single buffer is being dirtied: we want to set the
2495 * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2496 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2497 *
2498 * The caller must ensure this doesn't race with truncation.  Most will simply
2499 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2500 * the pte lock held, which also locks out truncation.
2501 */
2502int __set_page_dirty_nobuffers(struct page *page)
2503{
2504	lock_page_memcg(page);
2505	if (!TestSetPageDirty(page)) {
2506		struct address_space *mapping = page_mapping(page);
 
2507
2508		if (!mapping) {
2509			unlock_page_memcg(page);
2510			return 1;
2511		}
2512		__set_page_dirty(page, mapping, !PagePrivate(page));
 
 
 
 
 
 
 
2513		unlock_page_memcg(page);
2514
2515		if (mapping->host) {
2516			/* !PageAnon && !swapper_space */
2517			__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2518		}
2519		return 1;
2520	}
2521	unlock_page_memcg(page);
2522	return 0;
2523}
2524EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2525
2526/*
2527 * Call this whenever redirtying a page, to de-account the dirty counters
2528 * (NR_DIRTIED, WB_DIRTIED, tsk->nr_dirtied), so that they match the written
2529 * counters (NR_WRITTEN, WB_WRITTEN) in long term. The mismatches will lead to
2530 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2531 * control.
2532 */
2533void account_page_redirty(struct page *page)
2534{
2535	struct address_space *mapping = page->mapping;
2536
2537	if (mapping && mapping_can_writeback(mapping)) {
2538		struct inode *inode = mapping->host;
2539		struct bdi_writeback *wb;
2540		struct wb_lock_cookie cookie = {};
2541
2542		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2543		current->nr_dirtied--;
2544		dec_node_page_state(page, NR_DIRTIED);
2545		dec_wb_stat(wb, WB_DIRTIED);
2546		unlocked_inode_to_wb_end(inode, &cookie);
2547	}
2548}
2549EXPORT_SYMBOL(account_page_redirty);
2550
2551/*
2552 * When a writepage implementation decides that it doesn't want to write this
2553 * page for some reason, it should redirty the locked page via
2554 * redirty_page_for_writepage() and it should then unlock the page and return 0
2555 */
2556int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2557{
2558	int ret;
2559
2560	wbc->pages_skipped++;
2561	ret = __set_page_dirty_nobuffers(page);
2562	account_page_redirty(page);
2563	return ret;
2564}
2565EXPORT_SYMBOL(redirty_page_for_writepage);
2566
2567/*
2568 * Dirty a page.
2569 *
2570 * For pages with a mapping this should be done under the page lock for the
2571 * benefit of asynchronous memory errors who prefer a consistent dirty state.
2572 * This rule can be broken in some special cases, but should be better not to.
 
 
 
 
2573 */
2574int set_page_dirty(struct page *page)
2575{
2576	struct address_space *mapping = page_mapping(page);
2577
2578	page = compound_head(page);
2579	if (likely(mapping)) {
 
2580		/*
2581		 * readahead/lru_deactivate_page could remain
2582		 * PG_readahead/PG_reclaim due to race with end_page_writeback
2583		 * About readahead, if the page is written, the flags would be
2584		 * reset. So no problem.
2585		 * About lru_deactivate_page, if the page is redirty, the flag
2586		 * will be reset. So no problem. but if the page is used by readahead
2587		 * it will confuse readahead and make it restart the size rampup
2588		 * process. But it's a trivial problem.
2589		 */
2590		if (PageReclaim(page))
2591			ClearPageReclaim(page);
2592		return mapping->a_ops->set_page_dirty(page);
 
 
 
 
2593	}
2594	if (!PageDirty(page)) {
2595		if (!TestSetPageDirty(page))
2596			return 1;
2597	}
2598	return 0;
2599}
2600EXPORT_SYMBOL(set_page_dirty);
2601
2602/*
2603 * set_page_dirty() is racy if the caller has no reference against
2604 * page->mapping->host, and if the page is unlocked.  This is because another
2605 * CPU could truncate the page off the mapping and then free the mapping.
2606 *
2607 * Usually, the page _is_ locked, or the caller is a user-space process which
2608 * holds a reference on the inode by having an open file.
2609 *
2610 * In other cases, the page should be locked before running set_page_dirty().
2611 */
2612int set_page_dirty_lock(struct page *page)
2613{
2614	int ret;
2615
2616	lock_page(page);
2617	ret = set_page_dirty(page);
2618	unlock_page(page);
2619	return ret;
2620}
2621EXPORT_SYMBOL(set_page_dirty_lock);
2622
2623/*
2624 * This cancels just the dirty bit on the kernel page itself, it does NOT
2625 * actually remove dirty bits on any mmap's that may be around. It also
2626 * leaves the page tagged dirty, so any sync activity will still find it on
2627 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2628 * look at the dirty bits in the VM.
2629 *
2630 * Doing this should *normally* only ever be done when a page is truncated,
2631 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2632 * this when it notices that somebody has cleaned out all the buffers on a
2633 * page without actually doing it through the VM. Can you say "ext3 is
2634 * horribly ugly"? Thought you could.
2635 */
2636void __cancel_dirty_page(struct page *page)
2637{
2638	struct address_space *mapping = page_mapping(page);
2639
2640	if (mapping_can_writeback(mapping)) {
2641		struct inode *inode = mapping->host;
2642		struct bdi_writeback *wb;
2643		struct wb_lock_cookie cookie = {};
2644
2645		lock_page_memcg(page);
2646		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2647
2648		if (TestClearPageDirty(page))
2649			account_page_cleaned(page, mapping, wb);
2650
2651		unlocked_inode_to_wb_end(inode, &cookie);
2652		unlock_page_memcg(page);
2653	} else {
2654		ClearPageDirty(page);
2655	}
2656}
2657EXPORT_SYMBOL(__cancel_dirty_page);
2658
2659/*
2660 * Clear a page's dirty flag, while caring for dirty memory accounting.
2661 * Returns true if the page was previously dirty.
2662 *
2663 * This is for preparing to put the page under writeout.  We leave the page
2664 * tagged as dirty in the xarray so that a concurrent write-for-sync
2665 * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2666 * implementation will run either set_page_writeback() or set_page_dirty(),
2667 * at which stage we bring the page's dirty flag and xarray dirty tag
2668 * back into sync.
2669 *
2670 * This incoherency between the page's dirty flag and xarray tag is
2671 * unfortunate, but it only exists while the page is locked.
2672 */
2673int clear_page_dirty_for_io(struct page *page)
2674{
2675	struct address_space *mapping = page_mapping(page);
2676	int ret = 0;
2677
2678	VM_BUG_ON_PAGE(!PageLocked(page), page);
2679
2680	if (mapping && mapping_can_writeback(mapping)) {
2681		struct inode *inode = mapping->host;
2682		struct bdi_writeback *wb;
2683		struct wb_lock_cookie cookie = {};
2684
2685		/*
2686		 * Yes, Virginia, this is indeed insane.
2687		 *
2688		 * We use this sequence to make sure that
2689		 *  (a) we account for dirty stats properly
2690		 *  (b) we tell the low-level filesystem to
2691		 *      mark the whole page dirty if it was
2692		 *      dirty in a pagetable. Only to then
2693		 *  (c) clean the page again and return 1 to
2694		 *      cause the writeback.
2695		 *
2696		 * This way we avoid all nasty races with the
2697		 * dirty bit in multiple places and clearing
2698		 * them concurrently from different threads.
2699		 *
2700		 * Note! Normally the "set_page_dirty(page)"
2701		 * has no effect on the actual dirty bit - since
2702		 * that will already usually be set. But we
2703		 * need the side effects, and it can help us
2704		 * avoid races.
2705		 *
2706		 * We basically use the page "master dirty bit"
2707		 * as a serialization point for all the different
2708		 * threads doing their things.
2709		 */
2710		if (page_mkclean(page))
2711			set_page_dirty(page);
2712		/*
2713		 * We carefully synchronise fault handlers against
2714		 * installing a dirty pte and marking the page dirty
2715		 * at this point.  We do this by having them hold the
2716		 * page lock while dirtying the page, and pages are
2717		 * always locked coming in here, so we get the desired
2718		 * exclusion.
2719		 */
2720		wb = unlocked_inode_to_wb_begin(inode, &cookie);
2721		if (TestClearPageDirty(page)) {
2722			dec_lruvec_page_state(page, NR_FILE_DIRTY);
2723			dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2724			dec_wb_stat(wb, WB_RECLAIMABLE);
2725			ret = 1;
2726		}
2727		unlocked_inode_to_wb_end(inode, &cookie);
2728		return ret;
2729	}
2730	return TestClearPageDirty(page);
2731}
2732EXPORT_SYMBOL(clear_page_dirty_for_io);
2733
2734int test_clear_page_writeback(struct page *page)
2735{
2736	struct address_space *mapping = page_mapping(page);
 
 
2737	int ret;
2738
2739	lock_page_memcg(page);
 
2740	if (mapping && mapping_use_writeback_tags(mapping)) {
2741		struct inode *inode = mapping->host;
2742		struct backing_dev_info *bdi = inode_to_bdi(inode);
2743		unsigned long flags;
2744
2745		xa_lock_irqsave(&mapping->i_pages, flags);
2746		ret = TestClearPageWriteback(page);
2747		if (ret) {
2748			__xa_clear_mark(&mapping->i_pages, page_index(page),
2749						PAGECACHE_TAG_WRITEBACK);
2750			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) {
2751				struct bdi_writeback *wb = inode_to_wb(inode);
2752
2753				dec_wb_stat(wb, WB_WRITEBACK);
2754				__wb_writeout_inc(wb);
2755			}
2756		}
2757
2758		if (mapping->host && !mapping_tagged(mapping,
2759						     PAGECACHE_TAG_WRITEBACK))
2760			sb_clear_inode_writeback(mapping->host);
2761
2762		xa_unlock_irqrestore(&mapping->i_pages, flags);
2763	} else {
2764		ret = TestClearPageWriteback(page);
2765	}
 
 
 
 
 
 
2766	if (ret) {
2767		dec_lruvec_page_state(page, NR_WRITEBACK);
2768		dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2769		inc_node_page_state(page, NR_WRITTEN);
2770	}
2771	unlock_page_memcg(page);
2772	return ret;
2773}
2774
2775int __test_set_page_writeback(struct page *page, bool keep_write)
2776{
2777	struct address_space *mapping = page_mapping(page);
2778	int ret, access_ret;
2779
2780	lock_page_memcg(page);
2781	if (mapping && mapping_use_writeback_tags(mapping)) {
2782		XA_STATE(xas, &mapping->i_pages, page_index(page));
2783		struct inode *inode = mapping->host;
2784		struct backing_dev_info *bdi = inode_to_bdi(inode);
2785		unsigned long flags;
2786
2787		xas_lock_irqsave(&xas, flags);
2788		xas_load(&xas);
2789		ret = TestSetPageWriteback(page);
2790		if (!ret) {
2791			bool on_wblist;
2792
2793			on_wblist = mapping_tagged(mapping,
2794						   PAGECACHE_TAG_WRITEBACK);
2795
2796			xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
2797			if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT)
 
2798				inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
2799
2800			/*
2801			 * We can come through here when swapping anonymous
2802			 * pages, so we don't necessarily have an inode to track
2803			 * for sync.
2804			 */
2805			if (mapping->host && !on_wblist)
2806				sb_mark_inode_writeback(mapping->host);
2807		}
2808		if (!PageDirty(page))
2809			xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
 
2810		if (!keep_write)
2811			xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
2812		xas_unlock_irqrestore(&xas, flags);
 
2813	} else {
2814		ret = TestSetPageWriteback(page);
2815	}
2816	if (!ret) {
2817		inc_lruvec_page_state(page, NR_WRITEBACK);
2818		inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
2819	}
2820	unlock_page_memcg(page);
2821	access_ret = arch_make_page_accessible(page);
2822	/*
2823	 * If writeback has been triggered on a page that cannot be made
2824	 * accessible, it is too late to recover here.
2825	 */
2826	VM_BUG_ON_PAGE(access_ret != 0, page);
2827
2828	return ret;
2829
2830}
2831EXPORT_SYMBOL(__test_set_page_writeback);
2832
2833/*
2834 * Wait for a page to complete writeback
2835 */
2836void wait_on_page_writeback(struct page *page)
2837{
2838	while (PageWriteback(page)) {
2839		trace_wait_on_page_writeback(page, page_mapping(page));
2840		wait_on_page_bit(page, PG_writeback);
2841	}
2842}
2843EXPORT_SYMBOL_GPL(wait_on_page_writeback);
2844
2845/*
2846 * Wait for a page to complete writeback.  Returns -EINTR if we get a
2847 * fatal signal while waiting.
2848 */
2849int wait_on_page_writeback_killable(struct page *page)
2850{
2851	while (PageWriteback(page)) {
2852		trace_wait_on_page_writeback(page, page_mapping(page));
2853		if (wait_on_page_bit_killable(page, PG_writeback))
2854			return -EINTR;
2855	}
2856
2857	return 0;
2858}
2859EXPORT_SYMBOL_GPL(wait_on_page_writeback_killable);
2860
2861/**
2862 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2863 * @page:	The page to wait on.
2864 *
2865 * This function determines if the given page is related to a backing device
2866 * that requires page contents to be held stable during writeback.  If so, then
2867 * it will wait for any pending writeback to complete.
2868 */
2869void wait_for_stable_page(struct page *page)
2870{
2871	page = thp_head(page);
2872	if (page->mapping->host->i_sb->s_iflags & SB_I_STABLE_WRITES)
2873		wait_on_page_writeback(page);
2874}
2875EXPORT_SYMBOL_GPL(wait_for_stable_page);