Linux Audio

Check our new training course

Loading...
v5.9
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/swap.c
   4 *
   5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6 */
   7
   8/*
   9 * This file contains the default values for the operation of the
  10 * Linux VM subsystem. Fine-tuning documentation can be found in
  11 * Documentation/admin-guide/sysctl/vm.rst.
  12 * Started 18.12.91
  13 * Swap aging added 23.2.95, Stephen Tweedie.
  14 * Buffermem limits added 12.3.98, Rik van Riel.
  15 */
  16
  17#include <linux/mm.h>
  18#include <linux/sched.h>
  19#include <linux/kernel_stat.h>
  20#include <linux/swap.h>
  21#include <linux/mman.h>
  22#include <linux/pagemap.h>
  23#include <linux/pagevec.h>
  24#include <linux/init.h>
  25#include <linux/export.h>
  26#include <linux/mm_inline.h>
  27#include <linux/percpu_counter.h>
  28#include <linux/memremap.h>
  29#include <linux/percpu.h>
  30#include <linux/cpu.h>
  31#include <linux/notifier.h>
  32#include <linux/backing-dev.h>
  33#include <linux/memcontrol.h>
  34#include <linux/gfp.h>
  35#include <linux/uio.h>
  36#include <linux/hugetlb.h>
  37#include <linux/page_idle.h>
  38#include <linux/local_lock.h>
  39
  40#include "internal.h"
  41
  42#define CREATE_TRACE_POINTS
  43#include <trace/events/pagemap.h>
  44
  45/* How many pages do we try to swap or page in/out together? */
  46int page_cluster;
  47
  48/* Protecting only lru_rotate.pvec which requires disabling interrupts */
  49struct lru_rotate {
  50	local_lock_t lock;
  51	struct pagevec pvec;
  52};
  53static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
  54	.lock = INIT_LOCAL_LOCK(lock),
  55};
  56
  57/*
  58 * The following struct pagevec are grouped together because they are protected
  59 * by disabling preemption (and interrupts remain enabled).
  60 */
  61struct lru_pvecs {
  62	local_lock_t lock;
  63	struct pagevec lru_add;
  64	struct pagevec lru_deactivate_file;
  65	struct pagevec lru_deactivate;
  66	struct pagevec lru_lazyfree;
  67#ifdef CONFIG_SMP
  68	struct pagevec activate_page;
  69#endif
  70};
  71static DEFINE_PER_CPU(struct lru_pvecs, lru_pvecs) = {
  72	.lock = INIT_LOCAL_LOCK(lock),
  73};
  74
  75/*
  76 * This path almost never happens for VM activity - pages are normally
  77 * freed via pagevecs.  But it gets used by networking.
  78 */
  79static void __page_cache_release(struct page *page)
  80{
  81	if (PageLRU(page)) {
  82		pg_data_t *pgdat = page_pgdat(page);
  83		struct lruvec *lruvec;
  84		unsigned long flags;
  85
  86		spin_lock_irqsave(&pgdat->lru_lock, flags);
  87		lruvec = mem_cgroup_page_lruvec(page, pgdat);
  88		VM_BUG_ON_PAGE(!PageLRU(page), page);
  89		__ClearPageLRU(page);
  90		del_page_from_lru_list(page, lruvec, page_off_lru(page));
  91		spin_unlock_irqrestore(&pgdat->lru_lock, flags);
  92	}
  93	__ClearPageWaiters(page);
  94}
  95
  96static void __put_single_page(struct page *page)
  97{
  98	__page_cache_release(page);
  99	mem_cgroup_uncharge(page);
 100	free_unref_page(page);
 101}
 102
 103static void __put_compound_page(struct page *page)
 104{
 105	/*
 106	 * __page_cache_release() is supposed to be called for thp, not for
 107	 * hugetlb. This is because hugetlb page does never have PageLRU set
 108	 * (it's never listed to any LRU lists) and no memcg routines should
 109	 * be called for hugetlb (it has a separate hugetlb_cgroup.)
 110	 */
 111	if (!PageHuge(page))
 112		__page_cache_release(page);
 113	destroy_compound_page(page);
 114}
 115
 116void __put_page(struct page *page)
 117{
 118	if (is_zone_device_page(page)) {
 119		put_dev_pagemap(page->pgmap);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 120
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 121		/*
 122		 * The page belongs to the device that created pgmap. Do
 123		 * not return it to page allocator.
 
 
 
 124		 */
 125		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 126	}
 127
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 128	if (unlikely(PageCompound(page)))
 129		__put_compound_page(page);
 130	else
 131		__put_single_page(page);
 132}
 133EXPORT_SYMBOL(__put_page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 134
 135/**
 136 * put_pages_list() - release a list of pages
 137 * @pages: list of pages threaded on page->lru
 138 *
 139 * Release a list of pages which are strung together on page.lru.  Currently
 140 * used by read_cache_pages() and related error recovery code.
 141 */
 142void put_pages_list(struct list_head *pages)
 143{
 144	while (!list_empty(pages)) {
 145		struct page *victim;
 146
 147		victim = lru_to_page(pages);
 148		list_del(&victim->lru);
 149		put_page(victim);
 150	}
 151}
 152EXPORT_SYMBOL(put_pages_list);
 153
 154/*
 155 * get_kernel_pages() - pin kernel pages in memory
 156 * @kiov:	An array of struct kvec structures
 157 * @nr_segs:	number of segments to pin
 158 * @write:	pinning for read/write, currently ignored
 159 * @pages:	array that receives pointers to the pages pinned.
 160 *		Should be at least nr_segs long.
 161 *
 162 * Returns number of pages pinned. This may be fewer than the number
 163 * requested. If nr_pages is 0 or negative, returns 0. If no pages
 164 * were pinned, returns -errno. Each page returned must be released
 165 * with a put_page() call when it is finished with.
 166 */
 167int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 168		struct page **pages)
 169{
 170	int seg;
 171
 172	for (seg = 0; seg < nr_segs; seg++) {
 173		if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 174			return seg;
 175
 176		pages[seg] = kmap_to_page(kiov[seg].iov_base);
 177		get_page(pages[seg]);
 178	}
 179
 180	return seg;
 181}
 182EXPORT_SYMBOL_GPL(get_kernel_pages);
 183
 184/*
 185 * get_kernel_page() - pin a kernel page in memory
 186 * @start:	starting kernel address
 187 * @write:	pinning for read/write, currently ignored
 188 * @pages:	array that receives pointer to the page pinned.
 189 *		Must be at least nr_segs long.
 190 *
 191 * Returns 1 if page is pinned. If the page was not pinned, returns
 192 * -errno. The page returned must be released with a put_page() call
 193 * when it is finished with.
 194 */
 195int get_kernel_page(unsigned long start, int write, struct page **pages)
 196{
 197	const struct kvec kiov = {
 198		.iov_base = (void *)start,
 199		.iov_len = PAGE_SIZE
 200	};
 201
 202	return get_kernel_pages(&kiov, 1, write, pages);
 203}
 204EXPORT_SYMBOL_GPL(get_kernel_page);
 205
 206static void pagevec_lru_move_fn(struct pagevec *pvec,
 207	void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 208	void *arg)
 209{
 210	int i;
 211	struct pglist_data *pgdat = NULL;
 212	struct lruvec *lruvec;
 213	unsigned long flags = 0;
 214
 215	for (i = 0; i < pagevec_count(pvec); i++) {
 216		struct page *page = pvec->pages[i];
 217		struct pglist_data *pagepgdat = page_pgdat(page);
 218
 219		if (pagepgdat != pgdat) {
 220			if (pgdat)
 221				spin_unlock_irqrestore(&pgdat->lru_lock, flags);
 222			pgdat = pagepgdat;
 223			spin_lock_irqsave(&pgdat->lru_lock, flags);
 224		}
 225
 226		lruvec = mem_cgroup_page_lruvec(page, pgdat);
 227		(*move_fn)(page, lruvec, arg);
 228	}
 229	if (pgdat)
 230		spin_unlock_irqrestore(&pgdat->lru_lock, flags);
 231	release_pages(pvec->pages, pvec->nr);
 232	pagevec_reinit(pvec);
 233}
 234
 235static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 236				 void *arg)
 237{
 238	int *pgmoved = arg;
 239
 240	if (PageLRU(page) && !PageUnevictable(page)) {
 241		del_page_from_lru_list(page, lruvec, page_lru(page));
 242		ClearPageActive(page);
 243		add_page_to_lru_list_tail(page, lruvec, page_lru(page));
 244		(*pgmoved) += thp_nr_pages(page);
 245	}
 246}
 247
 248/*
 249 * pagevec_move_tail() must be called with IRQ disabled.
 250 * Otherwise this may cause nasty races.
 251 */
 252static void pagevec_move_tail(struct pagevec *pvec)
 253{
 254	int pgmoved = 0;
 255
 256	pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 257	__count_vm_events(PGROTATED, pgmoved);
 258}
 259
 260/*
 261 * Writeback is about to end against a page which has been marked for immediate
 262 * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 263 * inactive list.
 264 */
 265void rotate_reclaimable_page(struct page *page)
 266{
 267	if (!PageLocked(page) && !PageDirty(page) &&
 268	    !PageUnevictable(page) && PageLRU(page)) {
 269		struct pagevec *pvec;
 270		unsigned long flags;
 271
 272		get_page(page);
 273		local_lock_irqsave(&lru_rotate.lock, flags);
 274		pvec = this_cpu_ptr(&lru_rotate.pvec);
 275		if (!pagevec_add(pvec, page) || PageCompound(page))
 276			pagevec_move_tail(pvec);
 277		local_unlock_irqrestore(&lru_rotate.lock, flags);
 278	}
 279}
 280
 281void lru_note_cost(struct lruvec *lruvec, bool file, unsigned int nr_pages)
 
 282{
 283	do {
 284		unsigned long lrusize;
 285
 286		/* Record cost event */
 287		if (file)
 288			lruvec->file_cost += nr_pages;
 289		else
 290			lruvec->anon_cost += nr_pages;
 291
 292		/*
 293		 * Decay previous events
 294		 *
 295		 * Because workloads change over time (and to avoid
 296		 * overflow) we keep these statistics as a floating
 297		 * average, which ends up weighing recent refaults
 298		 * more than old ones.
 299		 */
 300		lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
 301			  lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
 302			  lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
 303			  lruvec_page_state(lruvec, NR_ACTIVE_FILE);
 304
 305		if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
 306			lruvec->file_cost /= 2;
 307			lruvec->anon_cost /= 2;
 308		}
 309	} while ((lruvec = parent_lruvec(lruvec)));
 310}
 311
 312void lru_note_cost_page(struct page *page)
 313{
 314	lru_note_cost(mem_cgroup_page_lruvec(page, page_pgdat(page)),
 315		      page_is_file_lru(page), thp_nr_pages(page));
 316}
 317
 318static void __activate_page(struct page *page, struct lruvec *lruvec,
 319			    void *arg)
 320{
 321	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 
 322		int lru = page_lru_base_type(page);
 323		int nr_pages = thp_nr_pages(page);
 324
 325		del_page_from_lru_list(page, lruvec, lru);
 326		SetPageActive(page);
 327		lru += LRU_ACTIVE;
 328		add_page_to_lru_list(page, lruvec, lru);
 329		trace_mm_lru_activate(page);
 330
 331		__count_vm_events(PGACTIVATE, nr_pages);
 332		__count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
 333				     nr_pages);
 334	}
 335}
 336
 337#ifdef CONFIG_SMP
 
 
 338static void activate_page_drain(int cpu)
 339{
 340	struct pagevec *pvec = &per_cpu(lru_pvecs.activate_page, cpu);
 341
 342	if (pagevec_count(pvec))
 343		pagevec_lru_move_fn(pvec, __activate_page, NULL);
 344}
 345
 346static bool need_activate_page_drain(int cpu)
 347{
 348	return pagevec_count(&per_cpu(lru_pvecs.activate_page, cpu)) != 0;
 349}
 350
 351void activate_page(struct page *page)
 352{
 353	page = compound_head(page);
 354	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 355		struct pagevec *pvec;
 356
 357		local_lock(&lru_pvecs.lock);
 358		pvec = this_cpu_ptr(&lru_pvecs.activate_page);
 359		get_page(page);
 360		if (!pagevec_add(pvec, page) || PageCompound(page))
 361			pagevec_lru_move_fn(pvec, __activate_page, NULL);
 362		local_unlock(&lru_pvecs.lock);
 363	}
 364}
 365
 366#else
 367static inline void activate_page_drain(int cpu)
 368{
 369}
 370
 
 
 
 
 
 371void activate_page(struct page *page)
 372{
 373	pg_data_t *pgdat = page_pgdat(page);
 374
 375	page = compound_head(page);
 376	spin_lock_irq(&pgdat->lru_lock);
 377	__activate_page(page, mem_cgroup_page_lruvec(page, pgdat), NULL);
 378	spin_unlock_irq(&pgdat->lru_lock);
 379}
 380#endif
 381
 382static void __lru_cache_activate_page(struct page *page)
 383{
 384	struct pagevec *pvec;
 385	int i;
 386
 387	local_lock(&lru_pvecs.lock);
 388	pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 389
 390	/*
 391	 * Search backwards on the optimistic assumption that the page being
 392	 * activated has just been added to this pagevec. Note that only
 393	 * the local pagevec is examined as a !PageLRU page could be in the
 394	 * process of being released, reclaimed, migrated or on a remote
 395	 * pagevec that is currently being drained. Furthermore, marking
 396	 * a remote pagevec's page PageActive potentially hits a race where
 397	 * a page is marked PageActive just after it is added to the inactive
 398	 * list causing accounting errors and BUG_ON checks to trigger.
 399	 */
 400	for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 401		struct page *pagevec_page = pvec->pages[i];
 402
 403		if (pagevec_page == page) {
 404			SetPageActive(page);
 405			break;
 406		}
 407	}
 408
 409	local_unlock(&lru_pvecs.lock);
 410}
 411
 412/*
 413 * Mark a page as having seen activity.
 414 *
 415 * inactive,unreferenced	->	inactive,referenced
 416 * inactive,referenced		->	active,unreferenced
 417 * active,unreferenced		->	active,referenced
 418 *
 419 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
 420 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
 421 */
 422void mark_page_accessed(struct page *page)
 423{
 424	page = compound_head(page);
 
 425
 426	if (!PageReferenced(page)) {
 427		SetPageReferenced(page);
 428	} else if (PageUnevictable(page)) {
 429		/*
 430		 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
 431		 * this list is never rotated or maintained, so marking an
 432		 * evictable page accessed has no effect.
 433		 */
 434	} else if (!PageActive(page)) {
 435		/*
 436		 * If the page is on the LRU, queue it for activation via
 437		 * lru_pvecs.activate_page. Otherwise, assume the page is on a
 438		 * pagevec, mark it active and it'll be moved to the active
 439		 * LRU on the next drain.
 440		 */
 441		if (PageLRU(page))
 442			activate_page(page);
 443		else
 444			__lru_cache_activate_page(page);
 445		ClearPageReferenced(page);
 446		workingset_activation(page);
 
 
 
 447	}
 448	if (page_is_idle(page))
 449		clear_page_idle(page);
 450}
 451EXPORT_SYMBOL(mark_page_accessed);
 452
 453/**
 454 * lru_cache_add - add a page to a page list
 455 * @page: the page to be added to the LRU.
 456 *
 457 * Queue the page for addition to the LRU via pagevec. The decision on whether
 458 * to add the page to the [in]active [file|anon] list is deferred until the
 459 * pagevec is drained. This gives a chance for the caller of lru_cache_add()
 460 * have the page added to the active list using mark_page_accessed().
 461 */
 462void lru_cache_add(struct page *page)
 463{
 464	struct pagevec *pvec;
 
 
 
 
 
 
 
 
 465
 
 
 
 
 
 
 466	VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
 467	VM_BUG_ON_PAGE(PageLRU(page), page);
 468
 469	get_page(page);
 470	local_lock(&lru_pvecs.lock);
 471	pvec = this_cpu_ptr(&lru_pvecs.lru_add);
 472	if (!pagevec_add(pvec, page) || PageCompound(page))
 473		__pagevec_lru_add(pvec);
 474	local_unlock(&lru_pvecs.lock);
 475}
 476EXPORT_SYMBOL(lru_cache_add);
 477
 478/**
 479 * lru_cache_add_inactive_or_unevictable
 480 * @page:  the page to be added to LRU
 481 * @vma:   vma in which page is mapped for determining reclaimability
 482 *
 483 * Place @page on the inactive or unevictable LRU list, depending on its
 484 * evictability.  Note that if the page is not evictable, it goes
 485 * directly back onto it's zone's unevictable list, it does NOT use a
 486 * per cpu pagevec.
 
 487 */
 488void lru_cache_add_inactive_or_unevictable(struct page *page,
 489					 struct vm_area_struct *vma)
 490{
 491	bool unevictable;
 492
 493	VM_BUG_ON_PAGE(PageLRU(page), page);
 494
 495	unevictable = (vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED;
 496	if (unlikely(unevictable) && !TestSetPageMlocked(page)) {
 497		int nr_pages = thp_nr_pages(page);
 498		/*
 499		 * We use the irq-unsafe __mod_zone_page_stat because this
 500		 * counter is not modified from interrupt context, and the pte
 501		 * lock is held(spinlock), which implies preemption disabled.
 502		 */
 503		__mod_zone_page_state(page_zone(page), NR_MLOCK, nr_pages);
 504		count_vm_events(UNEVICTABLE_PGMLOCKED, nr_pages);
 505	}
 506	lru_cache_add(page);
 507}
 508
 509/*
 510 * If the page can not be invalidated, it is moved to the
 511 * inactive list to speed up its reclaim.  It is moved to the
 512 * head of the list, rather than the tail, to give the flusher
 513 * threads some time to write it out, as this is much more
 514 * effective than the single-page writeout from reclaim.
 515 *
 516 * If the page isn't page_mapped and dirty/writeback, the page
 517 * could reclaim asap using PG_reclaim.
 518 *
 519 * 1. active, mapped page -> none
 520 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 521 * 3. inactive, mapped page -> none
 522 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 523 * 5. inactive, clean -> inactive, tail
 524 * 6. Others -> none
 525 *
 526 * In 4, why it moves inactive's head, the VM expects the page would
 527 * be write it out by flusher threads as this is much more effective
 528 * than the single-page writeout from reclaim.
 529 */
 530static void lru_deactivate_file_fn(struct page *page, struct lruvec *lruvec,
 531			      void *arg)
 532{
 533	int lru;
 534	bool active;
 535	int nr_pages = thp_nr_pages(page);
 536
 537	if (!PageLRU(page))
 538		return;
 539
 540	if (PageUnevictable(page))
 541		return;
 542
 543	/* Some processes are using the page */
 544	if (page_mapped(page))
 545		return;
 546
 547	active = PageActive(page);
 
 548	lru = page_lru_base_type(page);
 549
 550	del_page_from_lru_list(page, lruvec, lru + active);
 551	ClearPageActive(page);
 552	ClearPageReferenced(page);
 
 553
 554	if (PageWriteback(page) || PageDirty(page)) {
 555		/*
 556		 * PG_reclaim could be raced with end_page_writeback
 557		 * It can make readahead confusing.  But race window
 558		 * is _really_ small and  it's non-critical problem.
 559		 */
 560		add_page_to_lru_list(page, lruvec, lru);
 561		SetPageReclaim(page);
 562	} else {
 563		/*
 564		 * The page's writeback ends up during pagevec
 565		 * We moves tha page into tail of inactive.
 566		 */
 567		add_page_to_lru_list_tail(page, lruvec, lru);
 568		__count_vm_events(PGROTATED, nr_pages);
 569	}
 570
 571	if (active) {
 572		__count_vm_events(PGDEACTIVATE, nr_pages);
 573		__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 574				     nr_pages);
 575	}
 576}
 577
 578static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 579			    void *arg)
 580{
 581	if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 582		int lru = page_lru_base_type(page);
 583		int nr_pages = thp_nr_pages(page);
 584
 585		del_page_from_lru_list(page, lruvec, lru + LRU_ACTIVE);
 586		ClearPageActive(page);
 587		ClearPageReferenced(page);
 588		add_page_to_lru_list(page, lruvec, lru);
 589
 590		__count_vm_events(PGDEACTIVATE, nr_pages);
 591		__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 592				     nr_pages);
 593	}
 594}
 595
 596static void lru_lazyfree_fn(struct page *page, struct lruvec *lruvec,
 597			    void *arg)
 598{
 599	if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 600	    !PageSwapCache(page) && !PageUnevictable(page)) {
 601		bool active = PageActive(page);
 602		int nr_pages = thp_nr_pages(page);
 603
 604		del_page_from_lru_list(page, lruvec,
 605				       LRU_INACTIVE_ANON + active);
 606		ClearPageActive(page);
 607		ClearPageReferenced(page);
 608		/*
 609		 * Lazyfree pages are clean anonymous pages.  They have
 610		 * PG_swapbacked flag cleared, to distinguish them from normal
 611		 * anonymous pages
 612		 */
 613		ClearPageSwapBacked(page);
 614		add_page_to_lru_list(page, lruvec, LRU_INACTIVE_FILE);
 615
 616		__count_vm_events(PGLAZYFREE, nr_pages);
 617		__count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
 618				     nr_pages);
 619	}
 620}
 621
 622/*
 623 * Drain pages out of the cpu's pagevecs.
 624 * Either "cpu" is the current CPU, and preemption has already been
 625 * disabled; or "cpu" is being hot-unplugged, and is already dead.
 626 */
 627void lru_add_drain_cpu(int cpu)
 628{
 629	struct pagevec *pvec = &per_cpu(lru_pvecs.lru_add, cpu);
 630
 631	if (pagevec_count(pvec))
 632		__pagevec_lru_add(pvec);
 633
 634	pvec = &per_cpu(lru_rotate.pvec, cpu);
 635	/* Disabling interrupts below acts as a compiler barrier. */
 636	if (data_race(pagevec_count(pvec))) {
 637		unsigned long flags;
 638
 639		/* No harm done if a racing interrupt already did this */
 640		local_lock_irqsave(&lru_rotate.lock, flags);
 641		pagevec_move_tail(pvec);
 642		local_unlock_irqrestore(&lru_rotate.lock, flags);
 643	}
 644
 645	pvec = &per_cpu(lru_pvecs.lru_deactivate_file, cpu);
 646	if (pagevec_count(pvec))
 647		pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 648
 649	pvec = &per_cpu(lru_pvecs.lru_deactivate, cpu);
 650	if (pagevec_count(pvec))
 651		pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 652
 653	pvec = &per_cpu(lru_pvecs.lru_lazyfree, cpu);
 654	if (pagevec_count(pvec))
 655		pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 656
 657	activate_page_drain(cpu);
 658}
 659
 660/**
 661 * deactivate_file_page - forcefully deactivate a file page
 662 * @page: page to deactivate
 663 *
 664 * This function hints the VM that @page is a good reclaim candidate,
 665 * for example if its invalidation fails due to the page being dirty
 666 * or under writeback.
 667 */
 668void deactivate_file_page(struct page *page)
 669{
 670	/*
 671	 * In a workload with many unevictable page such as mprotect,
 672	 * unevictable page deactivation for accelerating reclaim is pointless.
 673	 */
 674	if (PageUnevictable(page))
 675		return;
 676
 677	if (likely(get_page_unless_zero(page))) {
 678		struct pagevec *pvec;
 679
 680		local_lock(&lru_pvecs.lock);
 681		pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate_file);
 682
 683		if (!pagevec_add(pvec, page) || PageCompound(page))
 684			pagevec_lru_move_fn(pvec, lru_deactivate_file_fn, NULL);
 685		local_unlock(&lru_pvecs.lock);
 686	}
 687}
 688
 689/*
 690 * deactivate_page - deactivate a page
 691 * @page: page to deactivate
 692 *
 693 * deactivate_page() moves @page to the inactive list if @page was on the active
 694 * list and was not an unevictable page.  This is done to accelerate the reclaim
 695 * of @page.
 696 */
 697void deactivate_page(struct page *page)
 698{
 699	if (PageLRU(page) && PageActive(page) && !PageUnevictable(page)) {
 700		struct pagevec *pvec;
 701
 702		local_lock(&lru_pvecs.lock);
 703		pvec = this_cpu_ptr(&lru_pvecs.lru_deactivate);
 704		get_page(page);
 705		if (!pagevec_add(pvec, page) || PageCompound(page))
 706			pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 707		local_unlock(&lru_pvecs.lock);
 708	}
 709}
 710
 711/**
 712 * mark_page_lazyfree - make an anon page lazyfree
 713 * @page: page to deactivate
 714 *
 715 * mark_page_lazyfree() moves @page to the inactive file list.
 716 * This is done to accelerate the reclaim of @page.
 717 */
 718void mark_page_lazyfree(struct page *page)
 719{
 720	if (PageLRU(page) && PageAnon(page) && PageSwapBacked(page) &&
 721	    !PageSwapCache(page) && !PageUnevictable(page)) {
 722		struct pagevec *pvec;
 723
 724		local_lock(&lru_pvecs.lock);
 725		pvec = this_cpu_ptr(&lru_pvecs.lru_lazyfree);
 726		get_page(page);
 727		if (!pagevec_add(pvec, page) || PageCompound(page))
 728			pagevec_lru_move_fn(pvec, lru_lazyfree_fn, NULL);
 729		local_unlock(&lru_pvecs.lock);
 730	}
 731}
 732
 733void lru_add_drain(void)
 734{
 735	local_lock(&lru_pvecs.lock);
 736	lru_add_drain_cpu(smp_processor_id());
 737	local_unlock(&lru_pvecs.lock);
 738}
 739
 740void lru_add_drain_cpu_zone(struct zone *zone)
 741{
 742	local_lock(&lru_pvecs.lock);
 743	lru_add_drain_cpu(smp_processor_id());
 744	drain_local_pages(zone);
 745	local_unlock(&lru_pvecs.lock);
 746}
 747
 748#ifdef CONFIG_SMP
 749
 750static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 751
 752static void lru_add_drain_per_cpu(struct work_struct *dummy)
 753{
 754	lru_add_drain();
 755}
 756
 757/*
 758 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 759 * kworkers being shut down before our page_alloc_cpu_dead callback is
 760 * executed on the offlined cpu.
 761 * Calling this function with cpu hotplug locks held can actually lead
 762 * to obscure indirect dependencies via WQ context.
 763 */
 764void lru_add_drain_all(void)
 765{
 766	static seqcount_t seqcount = SEQCNT_ZERO(seqcount);
 767	static DEFINE_MUTEX(lock);
 768	static struct cpumask has_work;
 769	int cpu, seq;
 770
 771	/*
 772	 * Make sure nobody triggers this path before mm_percpu_wq is fully
 773	 * initialized.
 774	 */
 775	if (WARN_ON(!mm_percpu_wq))
 776		return;
 777
 778	seq = raw_read_seqcount_latch(&seqcount);
 779
 780	mutex_lock(&lock);
 781
 782	/*
 783	 * Piggyback on drain started and finished while we waited for lock:
 784	 * all pages pended at the time of our enter were drained from vectors.
 785	 */
 786	if (__read_seqcount_retry(&seqcount, seq))
 787		goto done;
 788
 789	raw_write_seqcount_latch(&seqcount);
 790
 791	cpumask_clear(&has_work);
 792
 793	for_each_online_cpu(cpu) {
 794		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 795
 796		if (pagevec_count(&per_cpu(lru_pvecs.lru_add, cpu)) ||
 797		    data_race(pagevec_count(&per_cpu(lru_rotate.pvec, cpu))) ||
 798		    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
 799		    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
 800		    pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
 801		    need_activate_page_drain(cpu)) {
 802			INIT_WORK(work, lru_add_drain_per_cpu);
 803			queue_work_on(cpu, mm_percpu_wq, work);
 804			cpumask_set_cpu(cpu, &has_work);
 805		}
 806	}
 807
 808	for_each_cpu(cpu, &has_work)
 809		flush_work(&per_cpu(lru_add_drain_work, cpu));
 810
 811done:
 812	mutex_unlock(&lock);
 813}
 814#else
 815void lru_add_drain_all(void)
 816{
 817	lru_add_drain();
 818}
 819#endif
 820
 821/**
 822 * release_pages - batched put_page()
 823 * @pages: array of pages to release
 824 * @nr: number of pages
 825 *
 826 * Decrement the reference count on all the pages in @pages.  If it
 827 * fell to zero, remove the page from the LRU and free it.
 
 
 
 
 
 828 */
 829void release_pages(struct page **pages, int nr)
 830{
 831	int i;
 832	LIST_HEAD(pages_to_free);
 833	struct pglist_data *locked_pgdat = NULL;
 834	struct lruvec *lruvec;
 835	unsigned long flags;
 836	unsigned int lock_batch;
 837
 838	for (i = 0; i < nr; i++) {
 839		struct page *page = pages[i];
 840
 841		/*
 842		 * Make sure the IRQ-safe lock-holding time does not get
 843		 * excessive with a continuous string of pages from the
 844		 * same pgdat. The lock is held only if pgdat != NULL.
 845		 */
 846		if (locked_pgdat && ++lock_batch == SWAP_CLUSTER_MAX) {
 847			spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 848			locked_pgdat = NULL;
 849		}
 850
 851		if (is_huge_zero_page(page))
 852			continue;
 853
 854		if (is_zone_device_page(page)) {
 855			if (locked_pgdat) {
 856				spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 857						       flags);
 858				locked_pgdat = NULL;
 859			}
 860			/*
 861			 * ZONE_DEVICE pages that return 'false' from
 862			 * put_devmap_managed_page() do not require special
 863			 * processing, and instead, expect a call to
 864			 * put_page_testzero().
 865			 */
 866			if (page_is_devmap_managed(page)) {
 867				put_devmap_managed_page(page);
 868				continue;
 869			}
 
 
 870		}
 871
 872		page = compound_head(page);
 873		if (!put_page_testzero(page))
 874			continue;
 875
 876		if (PageCompound(page)) {
 877			if (locked_pgdat) {
 878				spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 879				locked_pgdat = NULL;
 880			}
 881			__put_compound_page(page);
 882			continue;
 883		}
 884
 885		if (PageLRU(page)) {
 886			struct pglist_data *pgdat = page_pgdat(page);
 887
 888			if (pgdat != locked_pgdat) {
 889				if (locked_pgdat)
 890					spin_unlock_irqrestore(&locked_pgdat->lru_lock,
 891									flags);
 892				lock_batch = 0;
 893				locked_pgdat = pgdat;
 894				spin_lock_irqsave(&locked_pgdat->lru_lock, flags);
 895			}
 896
 897			lruvec = mem_cgroup_page_lruvec(page, locked_pgdat);
 898			VM_BUG_ON_PAGE(!PageLRU(page), page);
 899			__ClearPageLRU(page);
 900			del_page_from_lru_list(page, lruvec, page_off_lru(page));
 901		}
 902
 903		/* Clear Active bit in case of parallel mark_page_accessed */
 904		__ClearPageActive(page);
 905		__ClearPageWaiters(page);
 906
 907		list_add(&page->lru, &pages_to_free);
 908	}
 909	if (locked_pgdat)
 910		spin_unlock_irqrestore(&locked_pgdat->lru_lock, flags);
 911
 912	mem_cgroup_uncharge_list(&pages_to_free);
 913	free_unref_page_list(&pages_to_free);
 914}
 915EXPORT_SYMBOL(release_pages);
 916
 917/*
 918 * The pages which we're about to release may be in the deferred lru-addition
 919 * queues.  That would prevent them from really being freed right now.  That's
 920 * OK from a correctness point of view but is inefficient - those pages may be
 921 * cache-warm and we want to give them back to the page allocator ASAP.
 922 *
 923 * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 924 * and __pagevec_lru_add_active() call release_pages() directly to avoid
 925 * mutual recursion.
 926 */
 927void __pagevec_release(struct pagevec *pvec)
 928{
 929	if (!pvec->percpu_pvec_drained) {
 930		lru_add_drain();
 931		pvec->percpu_pvec_drained = true;
 932	}
 933	release_pages(pvec->pages, pagevec_count(pvec));
 934	pagevec_reinit(pvec);
 935}
 936EXPORT_SYMBOL(__pagevec_release);
 937
 938#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 939/* used by __split_huge_page_refcount() */
 940void lru_add_page_tail(struct page *page, struct page *page_tail,
 941		       struct lruvec *lruvec, struct list_head *list)
 942{
 
 
 943	VM_BUG_ON_PAGE(!PageHead(page), page);
 944	VM_BUG_ON_PAGE(PageCompound(page_tail), page);
 945	VM_BUG_ON_PAGE(PageLRU(page_tail), page);
 946	lockdep_assert_held(&lruvec_pgdat(lruvec)->lru_lock);
 
 947
 948	if (!list)
 949		SetPageLRU(page_tail);
 950
 951	if (likely(PageLRU(page)))
 952		list_add_tail(&page_tail->lru, &page->lru);
 953	else if (list) {
 954		/* page reclaim is reclaiming a huge page */
 955		get_page(page_tail);
 956		list_add_tail(&page_tail->lru, list);
 957	} else {
 
 958		/*
 959		 * Head page has not yet been counted, as an hpage,
 960		 * so we must account for each subpage individually.
 961		 *
 962		 * Put page_tail on the list at the correct position
 963		 * so they all end up in order.
 964		 */
 965		add_page_to_lru_list_tail(page_tail, lruvec,
 966					  page_lru(page_tail));
 
 967	}
 
 
 
 968}
 969#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 970
 971static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 972				 void *arg)
 973{
 974	enum lru_list lru;
 975	int was_unevictable = TestClearPageUnevictable(page);
 976	int nr_pages = thp_nr_pages(page);
 977
 978	VM_BUG_ON_PAGE(PageLRU(page), page);
 979
 980	/*
 981	 * Page becomes evictable in two ways:
 982	 * 1) Within LRU lock [munlock_vma_page() and __munlock_pagevec()].
 983	 * 2) Before acquiring LRU lock to put the page to correct LRU and then
 984	 *   a) do PageLRU check with lock [check_move_unevictable_pages]
 985	 *   b) do PageLRU check before lock [clear_page_mlock]
 986	 *
 987	 * (1) & (2a) are ok as LRU lock will serialize them. For (2b), we need
 988	 * following strict ordering:
 989	 *
 990	 * #0: __pagevec_lru_add_fn		#1: clear_page_mlock
 991	 *
 992	 * SetPageLRU()				TestClearPageMlocked()
 993	 * smp_mb() // explicit ordering	// above provides strict
 994	 *					// ordering
 995	 * PageMlocked()			PageLRU()
 996	 *
 997	 *
 998	 * if '#1' does not observe setting of PG_lru by '#0' and fails
 999	 * isolation, the explicit barrier will make sure that page_evictable
1000	 * check will put the page in correct LRU. Without smp_mb(), SetPageLRU
1001	 * can be reordered after PageMlocked check and can make '#1' to fail
1002	 * the isolation of the page whose Mlocked bit is cleared (#0 is also
1003	 * looking at the same page) and the evictable page will be stranded
1004	 * in an unevictable LRU.
1005	 */
1006	SetPageLRU(page);
1007	smp_mb__after_atomic();
1008
1009	if (page_evictable(page)) {
1010		lru = page_lru(page);
1011		if (was_unevictable)
1012			__count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
1013	} else {
1014		lru = LRU_UNEVICTABLE;
1015		ClearPageActive(page);
1016		SetPageUnevictable(page);
1017		if (!was_unevictable)
1018			__count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
1019	}
1020
1021	add_page_to_lru_list(page, lruvec, lru);
1022	trace_mm_lru_insertion(page, lru);
 
1023}
1024
1025/*
1026 * Add the passed pages to the LRU, then drop the caller's refcount
1027 * on them.  Reinitialises the caller's pagevec.
1028 */
1029void __pagevec_lru_add(struct pagevec *pvec)
1030{
1031	pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
1032}
 
1033
1034/**
1035 * pagevec_lookup_entries - gang pagecache lookup
1036 * @pvec:	Where the resulting entries are placed
1037 * @mapping:	The address_space to search
1038 * @start:	The starting entry index
1039 * @nr_entries:	The maximum number of pages
1040 * @indices:	The cache indices corresponding to the entries in @pvec
1041 *
1042 * pagevec_lookup_entries() will search for and return a group of up
1043 * to @nr_pages pages and shadow entries in the mapping.  All
1044 * entries are placed in @pvec.  pagevec_lookup_entries() takes a
1045 * reference against actual pages in @pvec.
1046 *
1047 * The search returns a group of mapping-contiguous entries with
1048 * ascending indexes.  There may be holes in the indices due to
1049 * not-present entries.
1050 *
1051 * Only one subpage of a Transparent Huge Page is returned in one call:
1052 * allowing truncate_inode_pages_range() to evict the whole THP without
1053 * cycling through a pagevec of extra references.
1054 *
1055 * pagevec_lookup_entries() returns the number of entries which were
1056 * found.
1057 */
1058unsigned pagevec_lookup_entries(struct pagevec *pvec,
1059				struct address_space *mapping,
1060				pgoff_t start, unsigned nr_entries,
1061				pgoff_t *indices)
1062{
1063	pvec->nr = find_get_entries(mapping, start, nr_entries,
1064				    pvec->pages, indices);
1065	return pagevec_count(pvec);
1066}
1067
1068/**
1069 * pagevec_remove_exceptionals - pagevec exceptionals pruning
1070 * @pvec:	The pagevec to prune
1071 *
1072 * pagevec_lookup_entries() fills both pages and exceptional radix
1073 * tree entries into the pagevec.  This function prunes all
1074 * exceptionals from @pvec without leaving holes, so that it can be
1075 * passed on to page-only pagevec operations.
1076 */
1077void pagevec_remove_exceptionals(struct pagevec *pvec)
1078{
1079	int i, j;
1080
1081	for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
1082		struct page *page = pvec->pages[i];
1083		if (!xa_is_value(page))
1084			pvec->pages[j++] = page;
1085	}
1086	pvec->nr = j;
1087}
1088
1089/**
1090 * pagevec_lookup_range - gang pagecache lookup
1091 * @pvec:	Where the resulting pages are placed
1092 * @mapping:	The address_space to search
1093 * @start:	The starting page index
1094 * @end:	The final page index
1095 *
1096 * pagevec_lookup_range() will search for & return a group of up to PAGEVEC_SIZE
1097 * pages in the mapping starting from index @start and upto index @end
1098 * (inclusive).  The pages are placed in @pvec.  pagevec_lookup() takes a
1099 * reference against the pages in @pvec.
1100 *
1101 * The search returns a group of mapping-contiguous pages with ascending
1102 * indexes.  There may be holes in the indices due to not-present pages. We
1103 * also update @start to index the next page for the traversal.
1104 *
1105 * pagevec_lookup_range() returns the number of pages which were found. If this
1106 * number is smaller than PAGEVEC_SIZE, the end of specified range has been
1107 * reached.
1108 */
1109unsigned pagevec_lookup_range(struct pagevec *pvec,
1110		struct address_space *mapping, pgoff_t *start, pgoff_t end)
1111{
1112	pvec->nr = find_get_pages_range(mapping, start, end, PAGEVEC_SIZE,
1113					pvec->pages);
1114	return pagevec_count(pvec);
1115}
1116EXPORT_SYMBOL(pagevec_lookup_range);
1117
1118unsigned pagevec_lookup_range_tag(struct pagevec *pvec,
1119		struct address_space *mapping, pgoff_t *index, pgoff_t end,
1120		xa_mark_t tag)
1121{
1122	pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1123					PAGEVEC_SIZE, pvec->pages);
1124	return pagevec_count(pvec);
1125}
1126EXPORT_SYMBOL(pagevec_lookup_range_tag);
1127
1128unsigned pagevec_lookup_range_nr_tag(struct pagevec *pvec,
1129		struct address_space *mapping, pgoff_t *index, pgoff_t end,
1130		xa_mark_t tag, unsigned max_pages)
1131{
1132	pvec->nr = find_get_pages_range_tag(mapping, index, end, tag,
1133		min_t(unsigned int, max_pages, PAGEVEC_SIZE), pvec->pages);
1134	return pagevec_count(pvec);
1135}
1136EXPORT_SYMBOL(pagevec_lookup_range_nr_tag);
1137/*
1138 * Perform any setup for the swap system
1139 */
1140void __init swap_setup(void)
1141{
1142	unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
 
 
 
 
 
 
 
 
 
 
1143
1144	/* Use a smaller cluster for small-memory machines */
1145	if (megs < 16)
1146		page_cluster = 2;
1147	else
1148		page_cluster = 3;
1149	/*
1150	 * Right now other parts of the system means that we
1151	 * _really_ don't want to cluster much more
1152	 */
1153}
1154
1155#ifdef CONFIG_DEV_PAGEMAP_OPS
1156void put_devmap_managed_page(struct page *page)
1157{
1158	int count;
1159
1160	if (WARN_ON_ONCE(!page_is_devmap_managed(page)))
1161		return;
1162
1163	count = page_ref_dec_return(page);
1164
1165	/*
1166	 * devmap page refcounts are 1-based, rather than 0-based: if
1167	 * refcount is 1, then the page is free and the refcount is
1168	 * stable because nobody holds a reference on the page.
1169	 */
1170	if (count == 1)
1171		free_devmap_managed_page(page);
1172	else if (!count)
1173		__put_page(page);
1174}
1175EXPORT_SYMBOL(put_devmap_managed_page);
1176#endif
v3.15
 
   1/*
   2 *  linux/mm/swap.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 */
   6
   7/*
   8 * This file contains the default values for the operation of the
   9 * Linux VM subsystem. Fine-tuning documentation can be found in
  10 * Documentation/sysctl/vm.txt.
  11 * Started 18.12.91
  12 * Swap aging added 23.2.95, Stephen Tweedie.
  13 * Buffermem limits added 12.3.98, Rik van Riel.
  14 */
  15
  16#include <linux/mm.h>
  17#include <linux/sched.h>
  18#include <linux/kernel_stat.h>
  19#include <linux/swap.h>
  20#include <linux/mman.h>
  21#include <linux/pagemap.h>
  22#include <linux/pagevec.h>
  23#include <linux/init.h>
  24#include <linux/export.h>
  25#include <linux/mm_inline.h>
  26#include <linux/percpu_counter.h>
 
  27#include <linux/percpu.h>
  28#include <linux/cpu.h>
  29#include <linux/notifier.h>
  30#include <linux/backing-dev.h>
  31#include <linux/memcontrol.h>
  32#include <linux/gfp.h>
  33#include <linux/uio.h>
 
 
 
  34
  35#include "internal.h"
  36
  37#define CREATE_TRACE_POINTS
  38#include <trace/events/pagemap.h>
  39
  40/* How many pages do we try to swap or page in/out together? */
  41int page_cluster;
  42
  43static DEFINE_PER_CPU(struct pagevec, lru_add_pvec);
  44static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
  45static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  46
  47/*
  48 * This path almost never happens for VM activity - pages are normally
  49 * freed via pagevecs.  But it gets used by networking.
  50 */
  51static void __page_cache_release(struct page *page)
  52{
  53	if (PageLRU(page)) {
  54		struct zone *zone = page_zone(page);
  55		struct lruvec *lruvec;
  56		unsigned long flags;
  57
  58		spin_lock_irqsave(&zone->lru_lock, flags);
  59		lruvec = mem_cgroup_page_lruvec(page, zone);
  60		VM_BUG_ON_PAGE(!PageLRU(page), page);
  61		__ClearPageLRU(page);
  62		del_page_from_lru_list(page, lruvec, page_off_lru(page));
  63		spin_unlock_irqrestore(&zone->lru_lock, flags);
  64	}
 
  65}
  66
  67static void __put_single_page(struct page *page)
  68{
  69	__page_cache_release(page);
  70	free_hot_cold_page(page, 0);
 
  71}
  72
  73static void __put_compound_page(struct page *page)
  74{
  75	compound_page_dtor *dtor;
  76
  77	__page_cache_release(page);
  78	dtor = get_compound_page_dtor(page);
  79	(*dtor)(page);
 
 
 
 
  80}
  81
  82static void put_compound_page(struct page *page)
  83{
  84	struct page *page_head;
  85
  86	if (likely(!PageTail(page))) {
  87		if (put_page_testzero(page)) {
  88			/*
  89			 * By the time all refcounts have been released
  90			 * split_huge_page cannot run anymore from under us.
  91			 */
  92			if (PageHead(page))
  93				__put_compound_page(page);
  94			else
  95				__put_single_page(page);
  96		}
  97		return;
  98	}
  99
 100	/* __split_huge_page_refcount can run under us */
 101	page_head = compound_head(page);
 102
 103	/*
 104	 * THP can not break up slab pages so avoid taking
 105	 * compound_lock() and skip the tail page refcounting (in
 106	 * _mapcount) too. Slab performs non-atomic bit ops on
 107	 * page->flags for better performance. In particular
 108	 * slab_unlock() in slub used to be a hot path. It is still
 109	 * hot on arches that do not support
 110	 * this_cpu_cmpxchg_double().
 111	 *
 112	 * If "page" is part of a slab or hugetlbfs page it cannot be
 113	 * splitted and the head page cannot change from under us. And
 114	 * if "page" is part of a THP page under splitting, if the
 115	 * head page pointed by the THP tail isn't a THP head anymore,
 116	 * we'll find PageTail clear after smp_rmb() and we'll treat
 117	 * it as a single page.
 118	 */
 119	if (!__compound_tail_refcounted(page_head)) {
 120		/*
 121		 * If "page" is a THP tail, we must read the tail page
 122		 * flags after the head page flags. The
 123		 * split_huge_page side enforces write memory barriers
 124		 * between clearing PageTail and before the head page
 125		 * can be freed and reallocated.
 126		 */
 127		smp_rmb();
 128		if (likely(PageTail(page))) {
 129			/*
 130			 * __split_huge_page_refcount cannot race
 131			 * here.
 132			 */
 133			VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
 134			VM_BUG_ON_PAGE(page_mapcount(page) != 0, page);
 135			if (put_page_testzero(page_head)) {
 136				/*
 137				 * If this is the tail of a slab
 138				 * compound page, the tail pin must
 139				 * not be the last reference held on
 140				 * the page, because the PG_slab
 141				 * cannot be cleared before all tail
 142				 * pins (which skips the _mapcount
 143				 * tail refcounting) have been
 144				 * released. For hugetlbfs the tail
 145				 * pin may be the last reference on
 146				 * the page instead, because
 147				 * PageHeadHuge will not go away until
 148				 * the compound page enters the buddy
 149				 * allocator.
 150				 */
 151				VM_BUG_ON_PAGE(PageSlab(page_head), page_head);
 152				__put_compound_page(page_head);
 153			}
 154			return;
 155		} else
 156			/*
 157			 * __split_huge_page_refcount run before us,
 158			 * "page" was a THP tail. The split page_head
 159			 * has been freed and reallocated as slab or
 160			 * hugetlbfs page of smaller order (only
 161			 * possible if reallocated as slab on x86).
 162			 */
 163			goto out_put_single;
 164	}
 165
 166	if (likely(page != page_head && get_page_unless_zero(page_head))) {
 167		unsigned long flags;
 168
 169		/*
 170		 * page_head wasn't a dangling pointer but it may not
 171		 * be a head page anymore by the time we obtain the
 172		 * lock. That is ok as long as it can't be freed from
 173		 * under us.
 174		 */
 175		flags = compound_lock_irqsave(page_head);
 176		if (unlikely(!PageTail(page))) {
 177			/* __split_huge_page_refcount run before us */
 178			compound_unlock_irqrestore(page_head, flags);
 179			if (put_page_testzero(page_head)) {
 180				/*
 181				 * The head page may have been freed
 182				 * and reallocated as a compound page
 183				 * of smaller order and then freed
 184				 * again.  All we know is that it
 185				 * cannot have become: a THP page, a
 186				 * compound page of higher order, a
 187				 * tail page.  That is because we
 188				 * still hold the refcount of the
 189				 * split THP tail and page_head was
 190				 * the THP head before the split.
 191				 */
 192				if (PageHead(page_head))
 193					__put_compound_page(page_head);
 194				else
 195					__put_single_page(page_head);
 196			}
 197out_put_single:
 198			if (put_page_testzero(page))
 199				__put_single_page(page);
 200			return;
 201		}
 202		VM_BUG_ON_PAGE(page_head != page->first_page, page);
 203		/*
 204		 * We can release the refcount taken by
 205		 * get_page_unless_zero() now that
 206		 * __split_huge_page_refcount() is blocked on the
 207		 * compound_lock.
 208		 */
 209		if (put_page_testzero(page_head))
 210			VM_BUG_ON_PAGE(1, page_head);
 211		/* __split_huge_page_refcount will wait now */
 212		VM_BUG_ON_PAGE(page_mapcount(page) <= 0, page);
 213		atomic_dec(&page->_mapcount);
 214		VM_BUG_ON_PAGE(atomic_read(&page_head->_count) <= 0, page_head);
 215		VM_BUG_ON_PAGE(atomic_read(&page->_count) != 0, page);
 216		compound_unlock_irqrestore(page_head, flags);
 217
 218		if (put_page_testzero(page_head)) {
 219			if (PageHead(page_head))
 220				__put_compound_page(page_head);
 221			else
 222				__put_single_page(page_head);
 223		}
 224	} else {
 225		/* page_head is a dangling pointer */
 226		VM_BUG_ON_PAGE(PageTail(page), page);
 227		goto out_put_single;
 228	}
 229}
 230
 231void put_page(struct page *page)
 232{
 233	if (unlikely(PageCompound(page)))
 234		put_compound_page(page);
 235	else if (put_page_testzero(page))
 236		__put_single_page(page);
 237}
 238EXPORT_SYMBOL(put_page);
 239
 240/*
 241 * This function is exported but must not be called by anything other
 242 * than get_page(). It implements the slow path of get_page().
 243 */
 244bool __get_page_tail(struct page *page)
 245{
 246	/*
 247	 * This takes care of get_page() if run on a tail page
 248	 * returned by one of the get_user_pages/follow_page variants.
 249	 * get_user_pages/follow_page itself doesn't need the compound
 250	 * lock because it runs __get_page_tail_foll() under the
 251	 * proper PT lock that already serializes against
 252	 * split_huge_page().
 253	 */
 254	unsigned long flags;
 255	bool got;
 256	struct page *page_head = compound_head(page);
 257
 258	/* Ref to put_compound_page() comment. */
 259	if (!__compound_tail_refcounted(page_head)) {
 260		smp_rmb();
 261		if (likely(PageTail(page))) {
 262			/*
 263			 * This is a hugetlbfs page or a slab
 264			 * page. __split_huge_page_refcount
 265			 * cannot race here.
 266			 */
 267			VM_BUG_ON_PAGE(!PageHead(page_head), page_head);
 268			__get_page_tail_foll(page, true);
 269			return true;
 270		} else {
 271			/*
 272			 * __split_huge_page_refcount run
 273			 * before us, "page" was a THP
 274			 * tail. The split page_head has been
 275			 * freed and reallocated as slab or
 276			 * hugetlbfs page of smaller order
 277			 * (only possible if reallocated as
 278			 * slab on x86).
 279			 */
 280			return false;
 281		}
 282	}
 283
 284	got = false;
 285	if (likely(page != page_head && get_page_unless_zero(page_head))) {
 286		/*
 287		 * page_head wasn't a dangling pointer but it
 288		 * may not be a head page anymore by the time
 289		 * we obtain the lock. That is ok as long as it
 290		 * can't be freed from under us.
 291		 */
 292		flags = compound_lock_irqsave(page_head);
 293		/* here __split_huge_page_refcount won't run anymore */
 294		if (likely(PageTail(page))) {
 295			__get_page_tail_foll(page, false);
 296			got = true;
 297		}
 298		compound_unlock_irqrestore(page_head, flags);
 299		if (unlikely(!got))
 300			put_page(page_head);
 301	}
 302	return got;
 303}
 304EXPORT_SYMBOL(__get_page_tail);
 305
 306/**
 307 * put_pages_list() - release a list of pages
 308 * @pages: list of pages threaded on page->lru
 309 *
 310 * Release a list of pages which are strung together on page.lru.  Currently
 311 * used by read_cache_pages() and related error recovery code.
 312 */
 313void put_pages_list(struct list_head *pages)
 314{
 315	while (!list_empty(pages)) {
 316		struct page *victim;
 317
 318		victim = list_entry(pages->prev, struct page, lru);
 319		list_del(&victim->lru);
 320		page_cache_release(victim);
 321	}
 322}
 323EXPORT_SYMBOL(put_pages_list);
 324
 325/*
 326 * get_kernel_pages() - pin kernel pages in memory
 327 * @kiov:	An array of struct kvec structures
 328 * @nr_segs:	number of segments to pin
 329 * @write:	pinning for read/write, currently ignored
 330 * @pages:	array that receives pointers to the pages pinned.
 331 *		Should be at least nr_segs long.
 332 *
 333 * Returns number of pages pinned. This may be fewer than the number
 334 * requested. If nr_pages is 0 or negative, returns 0. If no pages
 335 * were pinned, returns -errno. Each page returned must be released
 336 * with a put_page() call when it is finished with.
 337 */
 338int get_kernel_pages(const struct kvec *kiov, int nr_segs, int write,
 339		struct page **pages)
 340{
 341	int seg;
 342
 343	for (seg = 0; seg < nr_segs; seg++) {
 344		if (WARN_ON(kiov[seg].iov_len != PAGE_SIZE))
 345			return seg;
 346
 347		pages[seg] = kmap_to_page(kiov[seg].iov_base);
 348		page_cache_get(pages[seg]);
 349	}
 350
 351	return seg;
 352}
 353EXPORT_SYMBOL_GPL(get_kernel_pages);
 354
 355/*
 356 * get_kernel_page() - pin a kernel page in memory
 357 * @start:	starting kernel address
 358 * @write:	pinning for read/write, currently ignored
 359 * @pages:	array that receives pointer to the page pinned.
 360 *		Must be at least nr_segs long.
 361 *
 362 * Returns 1 if page is pinned. If the page was not pinned, returns
 363 * -errno. The page returned must be released with a put_page() call
 364 * when it is finished with.
 365 */
 366int get_kernel_page(unsigned long start, int write, struct page **pages)
 367{
 368	const struct kvec kiov = {
 369		.iov_base = (void *)start,
 370		.iov_len = PAGE_SIZE
 371	};
 372
 373	return get_kernel_pages(&kiov, 1, write, pages);
 374}
 375EXPORT_SYMBOL_GPL(get_kernel_page);
 376
 377static void pagevec_lru_move_fn(struct pagevec *pvec,
 378	void (*move_fn)(struct page *page, struct lruvec *lruvec, void *arg),
 379	void *arg)
 380{
 381	int i;
 382	struct zone *zone = NULL;
 383	struct lruvec *lruvec;
 384	unsigned long flags = 0;
 385
 386	for (i = 0; i < pagevec_count(pvec); i++) {
 387		struct page *page = pvec->pages[i];
 388		struct zone *pagezone = page_zone(page);
 389
 390		if (pagezone != zone) {
 391			if (zone)
 392				spin_unlock_irqrestore(&zone->lru_lock, flags);
 393			zone = pagezone;
 394			spin_lock_irqsave(&zone->lru_lock, flags);
 395		}
 396
 397		lruvec = mem_cgroup_page_lruvec(page, zone);
 398		(*move_fn)(page, lruvec, arg);
 399	}
 400	if (zone)
 401		spin_unlock_irqrestore(&zone->lru_lock, flags);
 402	release_pages(pvec->pages, pvec->nr, pvec->cold);
 403	pagevec_reinit(pvec);
 404}
 405
 406static void pagevec_move_tail_fn(struct page *page, struct lruvec *lruvec,
 407				 void *arg)
 408{
 409	int *pgmoved = arg;
 410
 411	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 412		enum lru_list lru = page_lru_base_type(page);
 413		list_move_tail(&page->lru, &lruvec->lists[lru]);
 414		(*pgmoved)++;
 
 415	}
 416}
 417
 418/*
 419 * pagevec_move_tail() must be called with IRQ disabled.
 420 * Otherwise this may cause nasty races.
 421 */
 422static void pagevec_move_tail(struct pagevec *pvec)
 423{
 424	int pgmoved = 0;
 425
 426	pagevec_lru_move_fn(pvec, pagevec_move_tail_fn, &pgmoved);
 427	__count_vm_events(PGROTATED, pgmoved);
 428}
 429
 430/*
 431 * Writeback is about to end against a page which has been marked for immediate
 432 * reclaim.  If it still appears to be reclaimable, move it to the tail of the
 433 * inactive list.
 434 */
 435void rotate_reclaimable_page(struct page *page)
 436{
 437	if (!PageLocked(page) && !PageDirty(page) && !PageActive(page) &&
 438	    !PageUnevictable(page) && PageLRU(page)) {
 439		struct pagevec *pvec;
 440		unsigned long flags;
 441
 442		page_cache_get(page);
 443		local_irq_save(flags);
 444		pvec = &__get_cpu_var(lru_rotate_pvecs);
 445		if (!pagevec_add(pvec, page))
 446			pagevec_move_tail(pvec);
 447		local_irq_restore(flags);
 448	}
 449}
 450
 451static void update_page_reclaim_stat(struct lruvec *lruvec,
 452				     int file, int rotated)
 453{
 454	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 455
 456	reclaim_stat->recent_scanned[file]++;
 457	if (rotated)
 458		reclaim_stat->recent_rotated[file]++;
 
 459}
 460
 461static void __activate_page(struct page *page, struct lruvec *lruvec,
 462			    void *arg)
 463{
 464	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 465		int file = page_is_file_cache(page);
 466		int lru = page_lru_base_type(page);
 
 467
 468		del_page_from_lru_list(page, lruvec, lru);
 469		SetPageActive(page);
 470		lru += LRU_ACTIVE;
 471		add_page_to_lru_list(page, lruvec, lru);
 472		trace_mm_lru_activate(page, page_to_pfn(page));
 473
 474		__count_vm_event(PGACTIVATE);
 475		update_page_reclaim_stat(lruvec, file, 1);
 
 476	}
 477}
 478
 479#ifdef CONFIG_SMP
 480static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
 481
 482static void activate_page_drain(int cpu)
 483{
 484	struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
 485
 486	if (pagevec_count(pvec))
 487		pagevec_lru_move_fn(pvec, __activate_page, NULL);
 488}
 489
 490static bool need_activate_page_drain(int cpu)
 491{
 492	return pagevec_count(&per_cpu(activate_page_pvecs, cpu)) != 0;
 493}
 494
 495void activate_page(struct page *page)
 496{
 
 497	if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
 498		struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
 499
 500		page_cache_get(page);
 501		if (!pagevec_add(pvec, page))
 
 
 502			pagevec_lru_move_fn(pvec, __activate_page, NULL);
 503		put_cpu_var(activate_page_pvecs);
 504	}
 505}
 506
 507#else
 508static inline void activate_page_drain(int cpu)
 509{
 510}
 511
 512static bool need_activate_page_drain(int cpu)
 513{
 514	return false;
 515}
 516
 517void activate_page(struct page *page)
 518{
 519	struct zone *zone = page_zone(page);
 520
 521	spin_lock_irq(&zone->lru_lock);
 522	__activate_page(page, mem_cgroup_page_lruvec(page, zone), NULL);
 523	spin_unlock_irq(&zone->lru_lock);
 
 524}
 525#endif
 526
 527static void __lru_cache_activate_page(struct page *page)
 528{
 529	struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 530	int i;
 531
 
 
 
 532	/*
 533	 * Search backwards on the optimistic assumption that the page being
 534	 * activated has just been added to this pagevec. Note that only
 535	 * the local pagevec is examined as a !PageLRU page could be in the
 536	 * process of being released, reclaimed, migrated or on a remote
 537	 * pagevec that is currently being drained. Furthermore, marking
 538	 * a remote pagevec's page PageActive potentially hits a race where
 539	 * a page is marked PageActive just after it is added to the inactive
 540	 * list causing accounting errors and BUG_ON checks to trigger.
 541	 */
 542	for (i = pagevec_count(pvec) - 1; i >= 0; i--) {
 543		struct page *pagevec_page = pvec->pages[i];
 544
 545		if (pagevec_page == page) {
 546			SetPageActive(page);
 547			break;
 548		}
 549	}
 550
 551	put_cpu_var(lru_add_pvec);
 552}
 553
 554/*
 555 * Mark a page as having seen activity.
 556 *
 557 * inactive,unreferenced	->	inactive,referenced
 558 * inactive,referenced		->	active,unreferenced
 559 * active,unreferenced		->	active,referenced
 
 
 
 560 */
 561void mark_page_accessed(struct page *page)
 562{
 563	if (!PageActive(page) && !PageUnevictable(page) &&
 564			PageReferenced(page)) {
 565
 
 
 
 
 
 
 
 
 
 566		/*
 567		 * If the page is on the LRU, queue it for activation via
 568		 * activate_page_pvecs. Otherwise, assume the page is on a
 569		 * pagevec, mark it active and it'll be moved to the active
 570		 * LRU on the next drain.
 571		 */
 572		if (PageLRU(page))
 573			activate_page(page);
 574		else
 575			__lru_cache_activate_page(page);
 576		ClearPageReferenced(page);
 577		if (page_is_file_cache(page))
 578			workingset_activation(page);
 579	} else if (!PageReferenced(page)) {
 580		SetPageReferenced(page);
 581	}
 
 
 582}
 583EXPORT_SYMBOL(mark_page_accessed);
 584
 585/*
 
 
 
 586 * Queue the page for addition to the LRU via pagevec. The decision on whether
 587 * to add the page to the [in]active [file|anon] list is deferred until the
 588 * pagevec is drained. This gives a chance for the caller of __lru_cache_add()
 589 * have the page added to the active list using mark_page_accessed().
 590 */
 591void __lru_cache_add(struct page *page)
 592{
 593	struct pagevec *pvec = &get_cpu_var(lru_add_pvec);
 594
 595	page_cache_get(page);
 596	if (!pagevec_space(pvec))
 597		__pagevec_lru_add(pvec);
 598	pagevec_add(pvec, page);
 599	put_cpu_var(lru_add_pvec);
 600}
 601EXPORT_SYMBOL(__lru_cache_add);
 602
 603/**
 604 * lru_cache_add - add a page to a page list
 605 * @page: the page to be added to the LRU.
 606 */
 607void lru_cache_add(struct page *page)
 608{
 609	VM_BUG_ON_PAGE(PageActive(page) && PageUnevictable(page), page);
 610	VM_BUG_ON_PAGE(PageLRU(page), page);
 611	__lru_cache_add(page);
 
 
 
 
 
 
 612}
 
 613
 614/**
 615 * add_page_to_unevictable_list - add a page to the unevictable list
 616 * @page:  the page to be added to the unevictable list
 
 617 *
 618 * Add page directly to its zone's unevictable list.  To avoid races with
 619 * tasks that might be making the page evictable, through eg. munlock,
 620 * munmap or exit, while it's not on the lru, we want to add the page
 621 * while it's locked or otherwise "invisible" to other tasks.  This is
 622 * difficult to do when using the pagevec cache, so bypass that.
 623 */
 624void add_page_to_unevictable_list(struct page *page)
 
 625{
 626	struct zone *zone = page_zone(page);
 627	struct lruvec *lruvec;
 
 628
 629	spin_lock_irq(&zone->lru_lock);
 630	lruvec = mem_cgroup_page_lruvec(page, zone);
 631	ClearPageActive(page);
 632	SetPageUnevictable(page);
 633	SetPageLRU(page);
 634	add_page_to_lru_list(page, lruvec, LRU_UNEVICTABLE);
 635	spin_unlock_irq(&zone->lru_lock);
 
 
 
 
 
 636}
 637
 638/*
 639 * If the page can not be invalidated, it is moved to the
 640 * inactive list to speed up its reclaim.  It is moved to the
 641 * head of the list, rather than the tail, to give the flusher
 642 * threads some time to write it out, as this is much more
 643 * effective than the single-page writeout from reclaim.
 644 *
 645 * If the page isn't page_mapped and dirty/writeback, the page
 646 * could reclaim asap using PG_reclaim.
 647 *
 648 * 1. active, mapped page -> none
 649 * 2. active, dirty/writeback page -> inactive, head, PG_reclaim
 650 * 3. inactive, mapped page -> none
 651 * 4. inactive, dirty/writeback page -> inactive, head, PG_reclaim
 652 * 5. inactive, clean -> inactive, tail
 653 * 6. Others -> none
 654 *
 655 * In 4, why it moves inactive's head, the VM expects the page would
 656 * be write it out by flusher threads as this is much more effective
 657 * than the single-page writeout from reclaim.
 658 */
 659static void lru_deactivate_fn(struct page *page, struct lruvec *lruvec,
 660			      void *arg)
 661{
 662	int lru, file;
 663	bool active;
 
 664
 665	if (!PageLRU(page))
 666		return;
 667
 668	if (PageUnevictable(page))
 669		return;
 670
 671	/* Some processes are using the page */
 672	if (page_mapped(page))
 673		return;
 674
 675	active = PageActive(page);
 676	file = page_is_file_cache(page);
 677	lru = page_lru_base_type(page);
 678
 679	del_page_from_lru_list(page, lruvec, lru + active);
 680	ClearPageActive(page);
 681	ClearPageReferenced(page);
 682	add_page_to_lru_list(page, lruvec, lru);
 683
 684	if (PageWriteback(page) || PageDirty(page)) {
 685		/*
 686		 * PG_reclaim could be raced with end_page_writeback
 687		 * It can make readahead confusing.  But race window
 688		 * is _really_ small and  it's non-critical problem.
 689		 */
 
 690		SetPageReclaim(page);
 691	} else {
 692		/*
 693		 * The page's writeback ends up during pagevec
 694		 * We moves tha page into tail of inactive.
 695		 */
 696		list_move_tail(&page->lru, &lruvec->lists[lru]);
 697		__count_vm_event(PGROTATED);
 698	}
 699
 700	if (active)
 701		__count_vm_event(PGDEACTIVATE);
 702	update_page_reclaim_stat(lruvec, file, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 703}
 704
 705/*
 706 * Drain pages out of the cpu's pagevecs.
 707 * Either "cpu" is the current CPU, and preemption has already been
 708 * disabled; or "cpu" is being hot-unplugged, and is already dead.
 709 */
 710void lru_add_drain_cpu(int cpu)
 711{
 712	struct pagevec *pvec = &per_cpu(lru_add_pvec, cpu);
 713
 714	if (pagevec_count(pvec))
 715		__pagevec_lru_add(pvec);
 716
 717	pvec = &per_cpu(lru_rotate_pvecs, cpu);
 718	if (pagevec_count(pvec)) {
 
 719		unsigned long flags;
 720
 721		/* No harm done if a racing interrupt already did this */
 722		local_irq_save(flags);
 723		pagevec_move_tail(pvec);
 724		local_irq_restore(flags);
 725	}
 726
 727	pvec = &per_cpu(lru_deactivate_pvecs, cpu);
 
 
 
 
 728	if (pagevec_count(pvec))
 729		pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 730
 
 
 
 
 731	activate_page_drain(cpu);
 732}
 733
 734/**
 735 * deactivate_page - forcefully deactivate a page
 736 * @page: page to deactivate
 737 *
 738 * This function hints the VM that @page is a good reclaim candidate,
 739 * for example if its invalidation fails due to the page being dirty
 740 * or under writeback.
 741 */
 742void deactivate_page(struct page *page)
 743{
 744	/*
 745	 * In a workload with many unevictable page such as mprotect, unevictable
 746	 * page deactivation for accelerating reclaim is pointless.
 747	 */
 748	if (PageUnevictable(page))
 749		return;
 750
 751	if (likely(get_page_unless_zero(page))) {
 752		struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 753
 754		if (!pagevec_add(pvec, page))
 
 
 
 755			pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
 756		put_cpu_var(lru_deactivate_pvecs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 757	}
 758}
 759
 760void lru_add_drain(void)
 761{
 762	lru_add_drain_cpu(get_cpu());
 763	put_cpu();
 
 764}
 765
 
 
 
 
 
 
 
 
 
 
 
 
 766static void lru_add_drain_per_cpu(struct work_struct *dummy)
 767{
 768	lru_add_drain();
 769}
 770
 771static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 772
 
 
 
 
 
 773void lru_add_drain_all(void)
 774{
 
 775	static DEFINE_MUTEX(lock);
 776	static struct cpumask has_work;
 777	int cpu;
 
 
 
 
 
 
 
 
 
 778
 779	mutex_lock(&lock);
 780	get_online_cpus();
 
 
 
 
 
 
 
 
 
 781	cpumask_clear(&has_work);
 782
 783	for_each_online_cpu(cpu) {
 784		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 785
 786		if (pagevec_count(&per_cpu(lru_add_pvec, cpu)) ||
 787		    pagevec_count(&per_cpu(lru_rotate_pvecs, cpu)) ||
 788		    pagevec_count(&per_cpu(lru_deactivate_pvecs, cpu)) ||
 
 
 789		    need_activate_page_drain(cpu)) {
 790			INIT_WORK(work, lru_add_drain_per_cpu);
 791			schedule_work_on(cpu, work);
 792			cpumask_set_cpu(cpu, &has_work);
 793		}
 794	}
 795
 796	for_each_cpu(cpu, &has_work)
 797		flush_work(&per_cpu(lru_add_drain_work, cpu));
 798
 799	put_online_cpus();
 800	mutex_unlock(&lock);
 801}
 
 
 
 
 
 
 802
 803/*
 804 * Batched page_cache_release().  Decrement the reference count on all the
 805 * passed pages.  If it fell to zero then remove the page from the LRU and
 806 * free it.
 807 *
 808 * Avoid taking zone->lru_lock if possible, but if it is taken, retain it
 809 * for the remainder of the operation.
 810 *
 811 * The locking in this function is against shrink_inactive_list(): we recheck
 812 * the page count inside the lock to see whether shrink_inactive_list()
 813 * grabbed the page via the LRU.  If it did, give up: shrink_inactive_list()
 814 * will free it.
 815 */
 816void release_pages(struct page **pages, int nr, int cold)
 817{
 818	int i;
 819	LIST_HEAD(pages_to_free);
 820	struct zone *zone = NULL;
 821	struct lruvec *lruvec;
 822	unsigned long uninitialized_var(flags);
 
 823
 824	for (i = 0; i < nr; i++) {
 825		struct page *page = pages[i];
 826
 827		if (unlikely(PageCompound(page))) {
 828			if (zone) {
 829				spin_unlock_irqrestore(&zone->lru_lock, flags);
 830				zone = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 831			}
 832			put_compound_page(page);
 833			continue;
 834		}
 835
 
 836		if (!put_page_testzero(page))
 837			continue;
 838
 
 
 
 
 
 
 
 
 
 839		if (PageLRU(page)) {
 840			struct zone *pagezone = page_zone(page);
 841
 842			if (pagezone != zone) {
 843				if (zone)
 844					spin_unlock_irqrestore(&zone->lru_lock,
 845									flags);
 846				zone = pagezone;
 847				spin_lock_irqsave(&zone->lru_lock, flags);
 
 848			}
 849
 850			lruvec = mem_cgroup_page_lruvec(page, zone);
 851			VM_BUG_ON_PAGE(!PageLRU(page), page);
 852			__ClearPageLRU(page);
 853			del_page_from_lru_list(page, lruvec, page_off_lru(page));
 854		}
 855
 856		/* Clear Active bit in case of parallel mark_page_accessed */
 857		ClearPageActive(page);
 
 858
 859		list_add(&page->lru, &pages_to_free);
 860	}
 861	if (zone)
 862		spin_unlock_irqrestore(&zone->lru_lock, flags);
 863
 864	free_hot_cold_page_list(&pages_to_free, cold);
 
 865}
 866EXPORT_SYMBOL(release_pages);
 867
 868/*
 869 * The pages which we're about to release may be in the deferred lru-addition
 870 * queues.  That would prevent them from really being freed right now.  That's
 871 * OK from a correctness point of view but is inefficient - those pages may be
 872 * cache-warm and we want to give them back to the page allocator ASAP.
 873 *
 874 * So __pagevec_release() will drain those queues here.  __pagevec_lru_add()
 875 * and __pagevec_lru_add_active() call release_pages() directly to avoid
 876 * mutual recursion.
 877 */
 878void __pagevec_release(struct pagevec *pvec)
 879{
 880	lru_add_drain();
 881	release_pages(pvec->pages, pagevec_count(pvec), pvec->cold);
 
 
 
 882	pagevec_reinit(pvec);
 883}
 884EXPORT_SYMBOL(__pagevec_release);
 885
 886#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 887/* used by __split_huge_page_refcount() */
 888void lru_add_page_tail(struct page *page, struct page *page_tail,
 889		       struct lruvec *lruvec, struct list_head *list)
 890{
 891	const int file = 0;
 892
 893	VM_BUG_ON_PAGE(!PageHead(page), page);
 894	VM_BUG_ON_PAGE(PageCompound(page_tail), page);
 895	VM_BUG_ON_PAGE(PageLRU(page_tail), page);
 896	VM_BUG_ON(NR_CPUS != 1 &&
 897		  !spin_is_locked(&lruvec_zone(lruvec)->lru_lock));
 898
 899	if (!list)
 900		SetPageLRU(page_tail);
 901
 902	if (likely(PageLRU(page)))
 903		list_add_tail(&page_tail->lru, &page->lru);
 904	else if (list) {
 905		/* page reclaim is reclaiming a huge page */
 906		get_page(page_tail);
 907		list_add_tail(&page_tail->lru, list);
 908	} else {
 909		struct list_head *list_head;
 910		/*
 911		 * Head page has not yet been counted, as an hpage,
 912		 * so we must account for each subpage individually.
 913		 *
 914		 * Use the standard add function to put page_tail on the list,
 915		 * but then correct its position so they all end up in order.
 916		 */
 917		add_page_to_lru_list(page_tail, lruvec, page_lru(page_tail));
 918		list_head = page_tail->lru.prev;
 919		list_move_tail(&page_tail->lru, list_head);
 920	}
 921
 922	if (!PageUnevictable(page))
 923		update_page_reclaim_stat(lruvec, file, PageActive(page_tail));
 924}
 925#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 926
 927static void __pagevec_lru_add_fn(struct page *page, struct lruvec *lruvec,
 928				 void *arg)
 929{
 930	int file = page_is_file_cache(page);
 931	int active = PageActive(page);
 932	enum lru_list lru = page_lru(page);
 933
 934	VM_BUG_ON_PAGE(PageLRU(page), page);
 935
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 936	SetPageLRU(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 937	add_page_to_lru_list(page, lruvec, lru);
 938	update_page_reclaim_stat(lruvec, file, active);
 939	trace_mm_lru_insertion(page, page_to_pfn(page), lru, trace_pagemap_flags(page));
 940}
 941
 942/*
 943 * Add the passed pages to the LRU, then drop the caller's refcount
 944 * on them.  Reinitialises the caller's pagevec.
 945 */
 946void __pagevec_lru_add(struct pagevec *pvec)
 947{
 948	pagevec_lru_move_fn(pvec, __pagevec_lru_add_fn, NULL);
 949}
 950EXPORT_SYMBOL(__pagevec_lru_add);
 951
 952/**
 953 * pagevec_lookup_entries - gang pagecache lookup
 954 * @pvec:	Where the resulting entries are placed
 955 * @mapping:	The address_space to search
 956 * @start:	The starting entry index
 957 * @nr_entries:	The maximum number of entries
 958 * @indices:	The cache indices corresponding to the entries in @pvec
 959 *
 960 * pagevec_lookup_entries() will search for and return a group of up
 961 * to @nr_entries pages and shadow entries in the mapping.  All
 962 * entries are placed in @pvec.  pagevec_lookup_entries() takes a
 963 * reference against actual pages in @pvec.
 964 *
 965 * The search returns a group of mapping-contiguous entries with
 966 * ascending indexes.  There may be holes in the indices due to
 967 * not-present entries.
 968 *
 
 
 
 
 969 * pagevec_lookup_entries() returns the number of entries which were
 970 * found.
 971 */
 972unsigned pagevec_lookup_entries(struct pagevec *pvec,
 973				struct address_space *mapping,
 974				pgoff_t start, unsigned nr_pages,
 975				pgoff_t *indices)
 976{
 977	pvec->nr = find_get_entries(mapping, start, nr_pages,
 978				    pvec->pages, indices);
 979	return pagevec_count(pvec);
 980}
 981
 982/**
 983 * pagevec_remove_exceptionals - pagevec exceptionals pruning
 984 * @pvec:	The pagevec to prune
 985 *
 986 * pagevec_lookup_entries() fills both pages and exceptional radix
 987 * tree entries into the pagevec.  This function prunes all
 988 * exceptionals from @pvec without leaving holes, so that it can be
 989 * passed on to page-only pagevec operations.
 990 */
 991void pagevec_remove_exceptionals(struct pagevec *pvec)
 992{
 993	int i, j;
 994
 995	for (i = 0, j = 0; i < pagevec_count(pvec); i++) {
 996		struct page *page = pvec->pages[i];
 997		if (!radix_tree_exceptional_entry(page))
 998			pvec->pages[j++] = page;
 999	}
1000	pvec->nr = j;
1001}
1002
1003/**
1004 * pagevec_lookup - gang pagecache lookup
1005 * @pvec:	Where the resulting pages are placed
1006 * @mapping:	The address_space to search
1007 * @start:	The starting page index
1008 * @nr_pages:	The maximum number of pages
1009 *
1010 * pagevec_lookup() will search for and return a group of up to @nr_pages pages
1011 * in the mapping.  The pages are placed in @pvec.  pagevec_lookup() takes a
 
1012 * reference against the pages in @pvec.
1013 *
1014 * The search returns a group of mapping-contiguous pages with ascending
1015 * indexes.  There may be holes in the indices due to not-present pages.
 
1016 *
1017 * pagevec_lookup() returns the number of pages which were found.
 
 
1018 */
1019unsigned pagevec_lookup(struct pagevec *pvec, struct address_space *mapping,
1020		pgoff_t start, unsigned nr_pages)
1021{
1022	pvec->nr = find_get_pages(mapping, start, nr_pages, pvec->pages);
 
1023	return pagevec_count(pvec);
1024}
1025EXPORT_SYMBOL(pagevec_lookup);
1026
1027unsigned pagevec_lookup_tag(struct pagevec *pvec, struct address_space *mapping,
1028		pgoff_t *index, int tag, unsigned nr_pages)
 
1029{
1030	pvec->nr = find_get_pages_tag(mapping, index, tag,
1031					nr_pages, pvec->pages);
1032	return pagevec_count(pvec);
1033}
1034EXPORT_SYMBOL(pagevec_lookup_tag);
1035
 
 
 
 
 
 
 
 
 
1036/*
1037 * Perform any setup for the swap system
1038 */
1039void __init swap_setup(void)
1040{
1041	unsigned long megs = totalram_pages >> (20 - PAGE_SHIFT);
1042#ifdef CONFIG_SWAP
1043	int i;
1044
1045	if (bdi_init(swapper_spaces[0].backing_dev_info))
1046		panic("Failed to init swap bdi");
1047	for (i = 0; i < MAX_SWAPFILES; i++) {
1048		spin_lock_init(&swapper_spaces[i].tree_lock);
1049		INIT_LIST_HEAD(&swapper_spaces[i].i_mmap_nonlinear);
1050	}
1051#endif
1052
1053	/* Use a smaller cluster for small-memory machines */
1054	if (megs < 16)
1055		page_cluster = 2;
1056	else
1057		page_cluster = 3;
1058	/*
1059	 * Right now other parts of the system means that we
1060	 * _really_ don't want to cluster much more
1061	 */
1062}