Linux Audio

Check our new training course

Loading...
   1/*
   2 *  linux/mm/vmscan.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 *
   6 *  Swap reorganised 29.12.95, Stephen Tweedie.
   7 *  kswapd added: 7.1.96  sct
   8 *  Removed kswapd_ctl limits, and swap out as many pages as needed
   9 *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
  10 *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  11 *  Multiqueue VM started 5.8.00, Rik van Riel.
  12 */
  13
  14#include <linux/mm.h>
  15#include <linux/module.h>
  16#include <linux/gfp.h>
  17#include <linux/kernel_stat.h>
  18#include <linux/swap.h>
  19#include <linux/pagemap.h>
  20#include <linux/init.h>
  21#include <linux/highmem.h>
  22#include <linux/vmstat.h>
  23#include <linux/file.h>
  24#include <linux/writeback.h>
  25#include <linux/blkdev.h>
  26#include <linux/buffer_head.h>	/* for try_to_release_page(),
  27					buffer_heads_over_limit */
  28#include <linux/mm_inline.h>
  29#include <linux/backing-dev.h>
  30#include <linux/rmap.h>
  31#include <linux/topology.h>
  32#include <linux/cpu.h>
  33#include <linux/cpuset.h>
  34#include <linux/compaction.h>
  35#include <linux/notifier.h>
  36#include <linux/rwsem.h>
  37#include <linux/delay.h>
  38#include <linux/kthread.h>
  39#include <linux/freezer.h>
  40#include <linux/memcontrol.h>
  41#include <linux/delayacct.h>
  42#include <linux/sysctl.h>
  43#include <linux/oom.h>
  44#include <linux/prefetch.h>
  45
  46#include <asm/tlbflush.h>
  47#include <asm/div64.h>
  48
  49#include <linux/swapops.h>
  50
  51#include "internal.h"
  52
  53#define CREATE_TRACE_POINTS
  54#include <trace/events/vmscan.h>
  55
  56struct scan_control {
  57	/* Incremented by the number of inactive pages that were scanned */
  58	unsigned long nr_scanned;
  59
  60	/* Number of pages freed so far during a call to shrink_zones() */
  61	unsigned long nr_reclaimed;
  62
  63	/* How many pages shrink_list() should reclaim */
  64	unsigned long nr_to_reclaim;
  65
  66	unsigned long hibernation_mode;
  67
  68	/* This context's GFP mask */
  69	gfp_t gfp_mask;
  70
  71	int may_writepage;
  72
  73	/* Can mapped pages be reclaimed? */
  74	int may_unmap;
  75
  76	/* Can pages be swapped as part of reclaim? */
  77	int may_swap;
  78
  79	int order;
  80
  81	/* Scan (total_size >> priority) pages at once */
  82	int priority;
  83
  84	/*
  85	 * The memory cgroup that hit its limit and as a result is the
  86	 * primary target of this reclaim invocation.
  87	 */
  88	struct mem_cgroup *target_mem_cgroup;
  89
  90	/*
  91	 * Nodemask of nodes allowed by the caller. If NULL, all nodes
  92	 * are scanned.
  93	 */
  94	nodemask_t	*nodemask;
  95};
  96
  97#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
  98
  99#ifdef ARCH_HAS_PREFETCH
 100#define prefetch_prev_lru_page(_page, _base, _field)			\
 101	do {								\
 102		if ((_page)->lru.prev != _base) {			\
 103			struct page *prev;				\
 104									\
 105			prev = lru_to_page(&(_page->lru));		\
 106			prefetch(&prev->_field);			\
 107		}							\
 108	} while (0)
 109#else
 110#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 111#endif
 112
 113#ifdef ARCH_HAS_PREFETCHW
 114#define prefetchw_prev_lru_page(_page, _base, _field)			\
 115	do {								\
 116		if ((_page)->lru.prev != _base) {			\
 117			struct page *prev;				\
 118									\
 119			prev = lru_to_page(&(_page->lru));		\
 120			prefetchw(&prev->_field);			\
 121		}							\
 122	} while (0)
 123#else
 124#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 125#endif
 126
 127/*
 128 * From 0 .. 100.  Higher means more swappy.
 129 */
 130int vm_swappiness = 60;
 131long vm_total_pages;	/* The total number of pages which the VM controls */
 132
 133static LIST_HEAD(shrinker_list);
 134static DECLARE_RWSEM(shrinker_rwsem);
 135
 136#ifdef CONFIG_CGROUP_MEM_RES_CTLR
 137static bool global_reclaim(struct scan_control *sc)
 138{
 139	return !sc->target_mem_cgroup;
 140}
 141#else
 142static bool global_reclaim(struct scan_control *sc)
 143{
 144	return true;
 145}
 146#endif
 147
 148static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
 149{
 150	if (!mem_cgroup_disabled())
 151		return mem_cgroup_get_lru_size(lruvec, lru);
 152
 153	return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
 154}
 155
 156/*
 157 * Add a shrinker callback to be called from the vm
 158 */
 159void register_shrinker(struct shrinker *shrinker)
 160{
 161	atomic_long_set(&shrinker->nr_in_batch, 0);
 162	down_write(&shrinker_rwsem);
 163	list_add_tail(&shrinker->list, &shrinker_list);
 164	up_write(&shrinker_rwsem);
 165}
 166EXPORT_SYMBOL(register_shrinker);
 167
 168/*
 169 * Remove one
 170 */
 171void unregister_shrinker(struct shrinker *shrinker)
 172{
 173	down_write(&shrinker_rwsem);
 174	list_del(&shrinker->list);
 175	up_write(&shrinker_rwsem);
 176}
 177EXPORT_SYMBOL(unregister_shrinker);
 178
 179static inline int do_shrinker_shrink(struct shrinker *shrinker,
 180				     struct shrink_control *sc,
 181				     unsigned long nr_to_scan)
 182{
 183	sc->nr_to_scan = nr_to_scan;
 184	return (*shrinker->shrink)(shrinker, sc);
 185}
 186
 187#define SHRINK_BATCH 128
 188/*
 189 * Call the shrink functions to age shrinkable caches
 190 *
 191 * Here we assume it costs one seek to replace a lru page and that it also
 192 * takes a seek to recreate a cache object.  With this in mind we age equal
 193 * percentages of the lru and ageable caches.  This should balance the seeks
 194 * generated by these structures.
 195 *
 196 * If the vm encountered mapped pages on the LRU it increase the pressure on
 197 * slab to avoid swapping.
 198 *
 199 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
 200 *
 201 * `lru_pages' represents the number of on-LRU pages in all the zones which
 202 * are eligible for the caller's allocation attempt.  It is used for balancing
 203 * slab reclaim versus page reclaim.
 204 *
 205 * Returns the number of slab objects which we shrunk.
 206 */
 207unsigned long shrink_slab(struct shrink_control *shrink,
 208			  unsigned long nr_pages_scanned,
 209			  unsigned long lru_pages)
 210{
 211	struct shrinker *shrinker;
 212	unsigned long ret = 0;
 213
 214	if (nr_pages_scanned == 0)
 215		nr_pages_scanned = SWAP_CLUSTER_MAX;
 216
 217	if (!down_read_trylock(&shrinker_rwsem)) {
 218		/* Assume we'll be able to shrink next time */
 219		ret = 1;
 220		goto out;
 221	}
 222
 223	list_for_each_entry(shrinker, &shrinker_list, list) {
 224		unsigned long long delta;
 225		long total_scan;
 226		long max_pass;
 227		int shrink_ret = 0;
 228		long nr;
 229		long new_nr;
 230		long batch_size = shrinker->batch ? shrinker->batch
 231						  : SHRINK_BATCH;
 232
 233		max_pass = do_shrinker_shrink(shrinker, shrink, 0);
 234		if (max_pass <= 0)
 235			continue;
 236
 237		/*
 238		 * copy the current shrinker scan count into a local variable
 239		 * and zero it so that other concurrent shrinker invocations
 240		 * don't also do this scanning work.
 241		 */
 242		nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
 243
 244		total_scan = nr;
 245		delta = (4 * nr_pages_scanned) / shrinker->seeks;
 246		delta *= max_pass;
 247		do_div(delta, lru_pages + 1);
 248		total_scan += delta;
 249		if (total_scan < 0) {
 250			printk(KERN_ERR "shrink_slab: %pF negative objects to "
 251			       "delete nr=%ld\n",
 252			       shrinker->shrink, total_scan);
 253			total_scan = max_pass;
 254		}
 255
 256		/*
 257		 * We need to avoid excessive windup on filesystem shrinkers
 258		 * due to large numbers of GFP_NOFS allocations causing the
 259		 * shrinkers to return -1 all the time. This results in a large
 260		 * nr being built up so when a shrink that can do some work
 261		 * comes along it empties the entire cache due to nr >>>
 262		 * max_pass.  This is bad for sustaining a working set in
 263		 * memory.
 264		 *
 265		 * Hence only allow the shrinker to scan the entire cache when
 266		 * a large delta change is calculated directly.
 267		 */
 268		if (delta < max_pass / 4)
 269			total_scan = min(total_scan, max_pass / 2);
 270
 271		/*
 272		 * Avoid risking looping forever due to too large nr value:
 273		 * never try to free more than twice the estimate number of
 274		 * freeable entries.
 275		 */
 276		if (total_scan > max_pass * 2)
 277			total_scan = max_pass * 2;
 278
 279		trace_mm_shrink_slab_start(shrinker, shrink, nr,
 280					nr_pages_scanned, lru_pages,
 281					max_pass, delta, total_scan);
 282
 283		while (total_scan >= batch_size) {
 284			int nr_before;
 285
 286			nr_before = do_shrinker_shrink(shrinker, shrink, 0);
 287			shrink_ret = do_shrinker_shrink(shrinker, shrink,
 288							batch_size);
 289			if (shrink_ret == -1)
 290				break;
 291			if (shrink_ret < nr_before)
 292				ret += nr_before - shrink_ret;
 293			count_vm_events(SLABS_SCANNED, batch_size);
 294			total_scan -= batch_size;
 295
 296			cond_resched();
 297		}
 298
 299		/*
 300		 * move the unused scan count back into the shrinker in a
 301		 * manner that handles concurrent updates. If we exhausted the
 302		 * scan, there is no need to do an update.
 303		 */
 304		if (total_scan > 0)
 305			new_nr = atomic_long_add_return(total_scan,
 306					&shrinker->nr_in_batch);
 307		else
 308			new_nr = atomic_long_read(&shrinker->nr_in_batch);
 309
 310		trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
 311	}
 312	up_read(&shrinker_rwsem);
 313out:
 314	cond_resched();
 315	return ret;
 316}
 317
 318static inline int is_page_cache_freeable(struct page *page)
 319{
 320	/*
 321	 * A freeable page cache page is referenced only by the caller
 322	 * that isolated the page, the page cache radix tree and
 323	 * optional buffer heads at page->private.
 324	 */
 325	return page_count(page) - page_has_private(page) == 2;
 326}
 327
 328static int may_write_to_queue(struct backing_dev_info *bdi,
 329			      struct scan_control *sc)
 330{
 331	if (current->flags & PF_SWAPWRITE)
 332		return 1;
 333	if (!bdi_write_congested(bdi))
 334		return 1;
 335	if (bdi == current->backing_dev_info)
 336		return 1;
 337	return 0;
 338}
 339
 340/*
 341 * We detected a synchronous write error writing a page out.  Probably
 342 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 343 * fsync(), msync() or close().
 344 *
 345 * The tricky part is that after writepage we cannot touch the mapping: nothing
 346 * prevents it from being freed up.  But we have a ref on the page and once
 347 * that page is locked, the mapping is pinned.
 348 *
 349 * We're allowed to run sleeping lock_page() here because we know the caller has
 350 * __GFP_FS.
 351 */
 352static void handle_write_error(struct address_space *mapping,
 353				struct page *page, int error)
 354{
 355	lock_page(page);
 356	if (page_mapping(page) == mapping)
 357		mapping_set_error(mapping, error);
 358	unlock_page(page);
 359}
 360
 361/* possible outcome of pageout() */
 362typedef enum {
 363	/* failed to write page out, page is locked */
 364	PAGE_KEEP,
 365	/* move page to the active list, page is locked */
 366	PAGE_ACTIVATE,
 367	/* page has been sent to the disk successfully, page is unlocked */
 368	PAGE_SUCCESS,
 369	/* page is clean and locked */
 370	PAGE_CLEAN,
 371} pageout_t;
 372
 373/*
 374 * pageout is called by shrink_page_list() for each dirty page.
 375 * Calls ->writepage().
 376 */
 377static pageout_t pageout(struct page *page, struct address_space *mapping,
 378			 struct scan_control *sc)
 379{
 380	/*
 381	 * If the page is dirty, only perform writeback if that write
 382	 * will be non-blocking.  To prevent this allocation from being
 383	 * stalled by pagecache activity.  But note that there may be
 384	 * stalls if we need to run get_block().  We could test
 385	 * PagePrivate for that.
 386	 *
 387	 * If this process is currently in __generic_file_aio_write() against
 388	 * this page's queue, we can perform writeback even if that
 389	 * will block.
 390	 *
 391	 * If the page is swapcache, write it back even if that would
 392	 * block, for some throttling. This happens by accident, because
 393	 * swap_backing_dev_info is bust: it doesn't reflect the
 394	 * congestion state of the swapdevs.  Easy to fix, if needed.
 395	 */
 396	if (!is_page_cache_freeable(page))
 397		return PAGE_KEEP;
 398	if (!mapping) {
 399		/*
 400		 * Some data journaling orphaned pages can have
 401		 * page->mapping == NULL while being dirty with clean buffers.
 402		 */
 403		if (page_has_private(page)) {
 404			if (try_to_free_buffers(page)) {
 405				ClearPageDirty(page);
 406				printk("%s: orphaned page\n", __func__);
 407				return PAGE_CLEAN;
 408			}
 409		}
 410		return PAGE_KEEP;
 411	}
 412	if (mapping->a_ops->writepage == NULL)
 413		return PAGE_ACTIVATE;
 414	if (!may_write_to_queue(mapping->backing_dev_info, sc))
 415		return PAGE_KEEP;
 416
 417	if (clear_page_dirty_for_io(page)) {
 418		int res;
 419		struct writeback_control wbc = {
 420			.sync_mode = WB_SYNC_NONE,
 421			.nr_to_write = SWAP_CLUSTER_MAX,
 422			.range_start = 0,
 423			.range_end = LLONG_MAX,
 424			.for_reclaim = 1,
 425		};
 426
 427		SetPageReclaim(page);
 428		res = mapping->a_ops->writepage(page, &wbc);
 429		if (res < 0)
 430			handle_write_error(mapping, page, res);
 431		if (res == AOP_WRITEPAGE_ACTIVATE) {
 432			ClearPageReclaim(page);
 433			return PAGE_ACTIVATE;
 434		}
 435
 436		if (!PageWriteback(page)) {
 437			/* synchronous write or broken a_ops? */
 438			ClearPageReclaim(page);
 439		}
 440		trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
 441		inc_zone_page_state(page, NR_VMSCAN_WRITE);
 442		return PAGE_SUCCESS;
 443	}
 444
 445	return PAGE_CLEAN;
 446}
 447
 448/*
 449 * Same as remove_mapping, but if the page is removed from the mapping, it
 450 * gets returned with a refcount of 0.
 451 */
 452static int __remove_mapping(struct address_space *mapping, struct page *page)
 453{
 454	BUG_ON(!PageLocked(page));
 455	BUG_ON(mapping != page_mapping(page));
 456
 457	spin_lock_irq(&mapping->tree_lock);
 458	/*
 459	 * The non racy check for a busy page.
 460	 *
 461	 * Must be careful with the order of the tests. When someone has
 462	 * a ref to the page, it may be possible that they dirty it then
 463	 * drop the reference. So if PageDirty is tested before page_count
 464	 * here, then the following race may occur:
 465	 *
 466	 * get_user_pages(&page);
 467	 * [user mapping goes away]
 468	 * write_to(page);
 469	 *				!PageDirty(page)    [good]
 470	 * SetPageDirty(page);
 471	 * put_page(page);
 472	 *				!page_count(page)   [good, discard it]
 473	 *
 474	 * [oops, our write_to data is lost]
 475	 *
 476	 * Reversing the order of the tests ensures such a situation cannot
 477	 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 478	 * load is not satisfied before that of page->_count.
 479	 *
 480	 * Note that if SetPageDirty is always performed via set_page_dirty,
 481	 * and thus under tree_lock, then this ordering is not required.
 482	 */
 483	if (!page_freeze_refs(page, 2))
 484		goto cannot_free;
 485	/* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
 486	if (unlikely(PageDirty(page))) {
 487		page_unfreeze_refs(page, 2);
 488		goto cannot_free;
 489	}
 490
 491	if (PageSwapCache(page)) {
 492		swp_entry_t swap = { .val = page_private(page) };
 493		__delete_from_swap_cache(page);
 494		spin_unlock_irq(&mapping->tree_lock);
 495		swapcache_free(swap, page);
 496	} else {
 497		void (*freepage)(struct page *);
 498
 499		freepage = mapping->a_ops->freepage;
 500
 501		__delete_from_page_cache(page);
 502		spin_unlock_irq(&mapping->tree_lock);
 503		mem_cgroup_uncharge_cache_page(page);
 504
 505		if (freepage != NULL)
 506			freepage(page);
 507	}
 508
 509	return 1;
 510
 511cannot_free:
 512	spin_unlock_irq(&mapping->tree_lock);
 513	return 0;
 514}
 515
 516/*
 517 * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 518 * someone else has a ref on the page, abort and return 0.  If it was
 519 * successfully detached, return 1.  Assumes the caller has a single ref on
 520 * this page.
 521 */
 522int remove_mapping(struct address_space *mapping, struct page *page)
 523{
 524	if (__remove_mapping(mapping, page)) {
 525		/*
 526		 * Unfreezing the refcount with 1 rather than 2 effectively
 527		 * drops the pagecache ref for us without requiring another
 528		 * atomic operation.
 529		 */
 530		page_unfreeze_refs(page, 1);
 531		return 1;
 532	}
 533	return 0;
 534}
 535
 536/**
 537 * putback_lru_page - put previously isolated page onto appropriate LRU list
 538 * @page: page to be put back to appropriate lru list
 539 *
 540 * Add previously isolated @page to appropriate LRU list.
 541 * Page may still be unevictable for other reasons.
 542 *
 543 * lru_lock must not be held, interrupts must be enabled.
 544 */
 545void putback_lru_page(struct page *page)
 546{
 547	int lru;
 548	int active = !!TestClearPageActive(page);
 549	int was_unevictable = PageUnevictable(page);
 550
 551	VM_BUG_ON(PageLRU(page));
 552
 553redo:
 554	ClearPageUnevictable(page);
 555
 556	if (page_evictable(page, NULL)) {
 557		/*
 558		 * For evictable pages, we can use the cache.
 559		 * In event of a race, worst case is we end up with an
 560		 * unevictable page on [in]active list.
 561		 * We know how to handle that.
 562		 */
 563		lru = active + page_lru_base_type(page);
 564		lru_cache_add_lru(page, lru);
 565	} else {
 566		/*
 567		 * Put unevictable pages directly on zone's unevictable
 568		 * list.
 569		 */
 570		lru = LRU_UNEVICTABLE;
 571		add_page_to_unevictable_list(page);
 572		/*
 573		 * When racing with an mlock or AS_UNEVICTABLE clearing
 574		 * (page is unlocked) make sure that if the other thread
 575		 * does not observe our setting of PG_lru and fails
 576		 * isolation/check_move_unevictable_pages,
 577		 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
 578		 * the page back to the evictable list.
 579		 *
 580		 * The other side is TestClearPageMlocked() or shmem_lock().
 581		 */
 582		smp_mb();
 583	}
 584
 585	/*
 586	 * page's status can change while we move it among lru. If an evictable
 587	 * page is on unevictable list, it never be freed. To avoid that,
 588	 * check after we added it to the list, again.
 589	 */
 590	if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
 591		if (!isolate_lru_page(page)) {
 592			put_page(page);
 593			goto redo;
 594		}
 595		/* This means someone else dropped this page from LRU
 596		 * So, it will be freed or putback to LRU again. There is
 597		 * nothing to do here.
 598		 */
 599	}
 600
 601	if (was_unevictable && lru != LRU_UNEVICTABLE)
 602		count_vm_event(UNEVICTABLE_PGRESCUED);
 603	else if (!was_unevictable && lru == LRU_UNEVICTABLE)
 604		count_vm_event(UNEVICTABLE_PGCULLED);
 605
 606	put_page(page);		/* drop ref from isolate */
 607}
 608
 609enum page_references {
 610	PAGEREF_RECLAIM,
 611	PAGEREF_RECLAIM_CLEAN,
 612	PAGEREF_KEEP,
 613	PAGEREF_ACTIVATE,
 614};
 615
 616static enum page_references page_check_references(struct page *page,
 617						  struct scan_control *sc)
 618{
 619	int referenced_ptes, referenced_page;
 620	unsigned long vm_flags;
 621
 622	referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
 623					  &vm_flags);
 624	referenced_page = TestClearPageReferenced(page);
 625
 626	/*
 627	 * Mlock lost the isolation race with us.  Let try_to_unmap()
 628	 * move the page to the unevictable list.
 629	 */
 630	if (vm_flags & VM_LOCKED)
 631		return PAGEREF_RECLAIM;
 632
 633	if (referenced_ptes) {
 634		if (PageSwapBacked(page))
 635			return PAGEREF_ACTIVATE;
 636		/*
 637		 * All mapped pages start out with page table
 638		 * references from the instantiating fault, so we need
 639		 * to look twice if a mapped file page is used more
 640		 * than once.
 641		 *
 642		 * Mark it and spare it for another trip around the
 643		 * inactive list.  Another page table reference will
 644		 * lead to its activation.
 645		 *
 646		 * Note: the mark is set for activated pages as well
 647		 * so that recently deactivated but used pages are
 648		 * quickly recovered.
 649		 */
 650		SetPageReferenced(page);
 651
 652		if (referenced_page || referenced_ptes > 1)
 653			return PAGEREF_ACTIVATE;
 654
 655		/*
 656		 * Activate file-backed executable pages after first usage.
 657		 */
 658		if (vm_flags & VM_EXEC)
 659			return PAGEREF_ACTIVATE;
 660
 661		return PAGEREF_KEEP;
 662	}
 663
 664	/* Reclaim if clean, defer dirty pages to writeback */
 665	if (referenced_page && !PageSwapBacked(page))
 666		return PAGEREF_RECLAIM_CLEAN;
 667
 668	return PAGEREF_RECLAIM;
 669}
 670
 671/*
 672 * shrink_page_list() returns the number of reclaimed pages
 673 */
 674static unsigned long shrink_page_list(struct list_head *page_list,
 675				      struct zone *zone,
 676				      struct scan_control *sc,
 677				      unsigned long *ret_nr_dirty,
 678				      unsigned long *ret_nr_writeback)
 679{
 680	LIST_HEAD(ret_pages);
 681	LIST_HEAD(free_pages);
 682	int pgactivate = 0;
 683	unsigned long nr_dirty = 0;
 684	unsigned long nr_congested = 0;
 685	unsigned long nr_reclaimed = 0;
 686	unsigned long nr_writeback = 0;
 687
 688	cond_resched();
 689
 690	while (!list_empty(page_list)) {
 691		enum page_references references;
 692		struct address_space *mapping;
 693		struct page *page;
 694		int may_enter_fs;
 695
 696		cond_resched();
 697
 698		page = lru_to_page(page_list);
 699		list_del(&page->lru);
 700
 701		if (!trylock_page(page))
 702			goto keep;
 703
 704		VM_BUG_ON(PageActive(page));
 705		VM_BUG_ON(page_zone(page) != zone);
 706
 707		sc->nr_scanned++;
 708
 709		if (unlikely(!page_evictable(page, NULL)))
 710			goto cull_mlocked;
 711
 712		if (!sc->may_unmap && page_mapped(page))
 713			goto keep_locked;
 714
 715		/* Double the slab pressure for mapped and swapcache pages */
 716		if (page_mapped(page) || PageSwapCache(page))
 717			sc->nr_scanned++;
 718
 719		may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 720			(PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 721
 722		if (PageWriteback(page)) {
 723			/*
 724			 * memcg doesn't have any dirty pages throttling so we
 725			 * could easily OOM just because too many pages are in
 726			 * writeback and there is nothing else to reclaim.
 727			 *
 728			 * Check __GFP_IO, certainly because a loop driver
 729			 * thread might enter reclaim, and deadlock if it waits
 730			 * on a page for which it is needed to do the write
 731			 * (loop masks off __GFP_IO|__GFP_FS for this reason);
 732			 * but more thought would probably show more reasons.
 733			 *
 734			 * Don't require __GFP_FS, since we're not going into
 735			 * the FS, just waiting on its writeback completion.
 736			 * Worryingly, ext4 gfs2 and xfs allocate pages with
 737			 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
 738			 * testing may_enter_fs here is liable to OOM on them.
 739			 */
 740			if (global_reclaim(sc) ||
 741			    !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
 742				/*
 743				 * This is slightly racy - end_page_writeback()
 744				 * might have just cleared PageReclaim, then
 745				 * setting PageReclaim here end up interpreted
 746				 * as PageReadahead - but that does not matter
 747				 * enough to care.  What we do want is for this
 748				 * page to have PageReclaim set next time memcg
 749				 * reclaim reaches the tests above, so it will
 750				 * then wait_on_page_writeback() to avoid OOM;
 751				 * and it's also appropriate in global reclaim.
 752				 */
 753				SetPageReclaim(page);
 754				nr_writeback++;
 755				goto keep_locked;
 756			}
 757			wait_on_page_writeback(page);
 758		}
 759
 760		references = page_check_references(page, sc);
 761		switch (references) {
 762		case PAGEREF_ACTIVATE:
 763			goto activate_locked;
 764		case PAGEREF_KEEP:
 765			goto keep_locked;
 766		case PAGEREF_RECLAIM:
 767		case PAGEREF_RECLAIM_CLEAN:
 768			; /* try to reclaim the page below */
 769		}
 770
 771		/*
 772		 * Anonymous process memory has backing store?
 773		 * Try to allocate it some swap space here.
 774		 */
 775		if (PageAnon(page) && !PageSwapCache(page)) {
 776			if (!(sc->gfp_mask & __GFP_IO))
 777				goto keep_locked;
 778			if (!add_to_swap(page))
 779				goto activate_locked;
 780			may_enter_fs = 1;
 781		}
 782
 783		mapping = page_mapping(page);
 784
 785		/*
 786		 * The page is mapped into the page tables of one or more
 787		 * processes. Try to unmap it here.
 788		 */
 789		if (page_mapped(page) && mapping) {
 790			switch (try_to_unmap(page, TTU_UNMAP)) {
 791			case SWAP_FAIL:
 792				goto activate_locked;
 793			case SWAP_AGAIN:
 794				goto keep_locked;
 795			case SWAP_MLOCK:
 796				goto cull_mlocked;
 797			case SWAP_SUCCESS:
 798				; /* try to free the page below */
 799			}
 800		}
 801
 802		if (PageDirty(page)) {
 803			nr_dirty++;
 804
 805			/*
 806			 * Only kswapd can writeback filesystem pages to
 807			 * avoid risk of stack overflow but do not writeback
 808			 * unless under significant pressure.
 809			 */
 810			if (page_is_file_cache(page) &&
 811					(!current_is_kswapd() ||
 812					 sc->priority >= DEF_PRIORITY - 2)) {
 813				/*
 814				 * Immediately reclaim when written back.
 815				 * Similar in principal to deactivate_page()
 816				 * except we already have the page isolated
 817				 * and know it's dirty
 818				 */
 819				inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
 820				SetPageReclaim(page);
 821
 822				goto keep_locked;
 823			}
 824
 825			if (references == PAGEREF_RECLAIM_CLEAN)
 826				goto keep_locked;
 827			if (!may_enter_fs)
 828				goto keep_locked;
 829			if (!sc->may_writepage)
 830				goto keep_locked;
 831
 832			/* Page is dirty, try to write it out here */
 833			switch (pageout(page, mapping, sc)) {
 834			case PAGE_KEEP:
 835				nr_congested++;
 836				goto keep_locked;
 837			case PAGE_ACTIVATE:
 838				goto activate_locked;
 839			case PAGE_SUCCESS:
 840				if (PageWriteback(page))
 841					goto keep;
 842				if (PageDirty(page))
 843					goto keep;
 844
 845				/*
 846				 * A synchronous write - probably a ramdisk.  Go
 847				 * ahead and try to reclaim the page.
 848				 */
 849				if (!trylock_page(page))
 850					goto keep;
 851				if (PageDirty(page) || PageWriteback(page))
 852					goto keep_locked;
 853				mapping = page_mapping(page);
 854			case PAGE_CLEAN:
 855				; /* try to free the page below */
 856			}
 857		}
 858
 859		/*
 860		 * If the page has buffers, try to free the buffer mappings
 861		 * associated with this page. If we succeed we try to free
 862		 * the page as well.
 863		 *
 864		 * We do this even if the page is PageDirty().
 865		 * try_to_release_page() does not perform I/O, but it is
 866		 * possible for a page to have PageDirty set, but it is actually
 867		 * clean (all its buffers are clean).  This happens if the
 868		 * buffers were written out directly, with submit_bh(). ext3
 869		 * will do this, as well as the blockdev mapping.
 870		 * try_to_release_page() will discover that cleanness and will
 871		 * drop the buffers and mark the page clean - it can be freed.
 872		 *
 873		 * Rarely, pages can have buffers and no ->mapping.  These are
 874		 * the pages which were not successfully invalidated in
 875		 * truncate_complete_page().  We try to drop those buffers here
 876		 * and if that worked, and the page is no longer mapped into
 877		 * process address space (page_count == 1) it can be freed.
 878		 * Otherwise, leave the page on the LRU so it is swappable.
 879		 */
 880		if (page_has_private(page)) {
 881			if (!try_to_release_page(page, sc->gfp_mask))
 882				goto activate_locked;
 883			if (!mapping && page_count(page) == 1) {
 884				unlock_page(page);
 885				if (put_page_testzero(page))
 886					goto free_it;
 887				else {
 888					/*
 889					 * rare race with speculative reference.
 890					 * the speculative reference will free
 891					 * this page shortly, so we may
 892					 * increment nr_reclaimed here (and
 893					 * leave it off the LRU).
 894					 */
 895					nr_reclaimed++;
 896					continue;
 897				}
 898			}
 899		}
 900
 901		if (!mapping || !__remove_mapping(mapping, page))
 902			goto keep_locked;
 903
 904		/*
 905		 * At this point, we have no other references and there is
 906		 * no way to pick any more up (removed from LRU, removed
 907		 * from pagecache). Can use non-atomic bitops now (and
 908		 * we obviously don't have to worry about waking up a process
 909		 * waiting on the page lock, because there are no references.
 910		 */
 911		__clear_page_locked(page);
 912free_it:
 913		nr_reclaimed++;
 914
 915		/*
 916		 * Is there need to periodically free_page_list? It would
 917		 * appear not as the counts should be low
 918		 */
 919		list_add(&page->lru, &free_pages);
 920		continue;
 921
 922cull_mlocked:
 923		if (PageSwapCache(page))
 924			try_to_free_swap(page);
 925		unlock_page(page);
 926		putback_lru_page(page);
 927		continue;
 928
 929activate_locked:
 930		/* Not a candidate for swapping, so reclaim swap space. */
 931		if (PageSwapCache(page) && vm_swap_full())
 932			try_to_free_swap(page);
 933		VM_BUG_ON(PageActive(page));
 934		SetPageActive(page);
 935		pgactivate++;
 936keep_locked:
 937		unlock_page(page);
 938keep:
 939		list_add(&page->lru, &ret_pages);
 940		VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
 941	}
 942
 943	/*
 944	 * Tag a zone as congested if all the dirty pages encountered were
 945	 * backed by a congested BDI. In this case, reclaimers should just
 946	 * back off and wait for congestion to clear because further reclaim
 947	 * will encounter the same problem
 948	 */
 949	if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
 950		zone_set_flag(zone, ZONE_CONGESTED);
 951
 952	free_hot_cold_page_list(&free_pages, 1);
 953
 954	list_splice(&ret_pages, page_list);
 955	count_vm_events(PGACTIVATE, pgactivate);
 956	*ret_nr_dirty += nr_dirty;
 957	*ret_nr_writeback += nr_writeback;
 958	return nr_reclaimed;
 959}
 960
 961/*
 962 * Attempt to remove the specified page from its LRU.  Only take this page
 963 * if it is of the appropriate PageActive status.  Pages which are being
 964 * freed elsewhere are also ignored.
 965 *
 966 * page:	page to consider
 967 * mode:	one of the LRU isolation modes defined above
 968 *
 969 * returns 0 on success, -ve errno on failure.
 970 */
 971int __isolate_lru_page(struct page *page, isolate_mode_t mode)
 972{
 973	int ret = -EINVAL;
 974
 975	/* Only take pages on the LRU. */
 976	if (!PageLRU(page))
 977		return ret;
 978
 979	/* Do not give back unevictable pages for compaction */
 980	if (PageUnevictable(page))
 981		return ret;
 982
 983	ret = -EBUSY;
 984
 985	/*
 986	 * To minimise LRU disruption, the caller can indicate that it only
 987	 * wants to isolate pages it will be able to operate on without
 988	 * blocking - clean pages for the most part.
 989	 *
 990	 * ISOLATE_CLEAN means that only clean pages should be isolated. This
 991	 * is used by reclaim when it is cannot write to backing storage
 992	 *
 993	 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
 994	 * that it is possible to migrate without blocking
 995	 */
 996	if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
 997		/* All the caller can do on PageWriteback is block */
 998		if (PageWriteback(page))
 999			return ret;
1000
1001		if (PageDirty(page)) {
1002			struct address_space *mapping;
1003
1004			/* ISOLATE_CLEAN means only clean pages */
1005			if (mode & ISOLATE_CLEAN)
1006				return ret;
1007
1008			/*
1009			 * Only pages without mappings or that have a
1010			 * ->migratepage callback are possible to migrate
1011			 * without blocking
1012			 */
1013			mapping = page_mapping(page);
1014			if (mapping && !mapping->a_ops->migratepage)
1015				return ret;
1016		}
1017	}
1018
1019	if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1020		return ret;
1021
1022	if (likely(get_page_unless_zero(page))) {
1023		/*
1024		 * Be careful not to clear PageLRU until after we're
1025		 * sure the page is not being freed elsewhere -- the
1026		 * page release code relies on it.
1027		 */
1028		ClearPageLRU(page);
1029		ret = 0;
1030	}
1031
1032	return ret;
1033}
1034
1035/*
1036 * zone->lru_lock is heavily contended.  Some of the functions that
1037 * shrink the lists perform better by taking out a batch of pages
1038 * and working on them outside the LRU lock.
1039 *
1040 * For pagecache intensive workloads, this function is the hottest
1041 * spot in the kernel (apart from copy_*_user functions).
1042 *
1043 * Appropriate locks must be held before calling this function.
1044 *
1045 * @nr_to_scan:	The number of pages to look through on the list.
1046 * @lruvec:	The LRU vector to pull pages from.
1047 * @dst:	The temp list to put pages on to.
1048 * @nr_scanned:	The number of pages that were scanned.
1049 * @sc:		The scan_control struct for this reclaim session
1050 * @mode:	One of the LRU isolation modes
1051 * @lru:	LRU list id for isolating
1052 *
1053 * returns how many pages were moved onto *@dst.
1054 */
1055static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1056		struct lruvec *lruvec, struct list_head *dst,
1057		unsigned long *nr_scanned, struct scan_control *sc,
1058		isolate_mode_t mode, enum lru_list lru)
1059{
1060	struct list_head *src = &lruvec->lists[lru];
1061	unsigned long nr_taken = 0;
1062	unsigned long scan;
1063
1064	for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1065		struct page *page;
1066		int nr_pages;
1067
1068		page = lru_to_page(src);
1069		prefetchw_prev_lru_page(page, src, flags);
1070
1071		VM_BUG_ON(!PageLRU(page));
1072
1073		switch (__isolate_lru_page(page, mode)) {
1074		case 0:
1075			nr_pages = hpage_nr_pages(page);
1076			mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1077			list_move(&page->lru, dst);
1078			nr_taken += nr_pages;
1079			break;
1080
1081		case -EBUSY:
1082			/* else it is being freed elsewhere */
1083			list_move(&page->lru, src);
1084			continue;
1085
1086		default:
1087			BUG();
1088		}
1089	}
1090
1091	*nr_scanned = scan;
1092	trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1093				    nr_taken, mode, is_file_lru(lru));
1094	return nr_taken;
1095}
1096
1097/**
1098 * isolate_lru_page - tries to isolate a page from its LRU list
1099 * @page: page to isolate from its LRU list
1100 *
1101 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1102 * vmstat statistic corresponding to whatever LRU list the page was on.
1103 *
1104 * Returns 0 if the page was removed from an LRU list.
1105 * Returns -EBUSY if the page was not on an LRU list.
1106 *
1107 * The returned page will have PageLRU() cleared.  If it was found on
1108 * the active list, it will have PageActive set.  If it was found on
1109 * the unevictable list, it will have the PageUnevictable bit set. That flag
1110 * may need to be cleared by the caller before letting the page go.
1111 *
1112 * The vmstat statistic corresponding to the list on which the page was
1113 * found will be decremented.
1114 *
1115 * Restrictions:
1116 * (1) Must be called with an elevated refcount on the page. This is a
1117 *     fundamentnal difference from isolate_lru_pages (which is called
1118 *     without a stable reference).
1119 * (2) the lru_lock must not be held.
1120 * (3) interrupts must be enabled.
1121 */
1122int isolate_lru_page(struct page *page)
1123{
1124	int ret = -EBUSY;
1125
1126	VM_BUG_ON(!page_count(page));
1127
1128	if (PageLRU(page)) {
1129		struct zone *zone = page_zone(page);
1130		struct lruvec *lruvec;
1131
1132		spin_lock_irq(&zone->lru_lock);
1133		lruvec = mem_cgroup_page_lruvec(page, zone);
1134		if (PageLRU(page)) {
1135			int lru = page_lru(page);
1136			get_page(page);
1137			ClearPageLRU(page);
1138			del_page_from_lru_list(page, lruvec, lru);
1139			ret = 0;
1140		}
1141		spin_unlock_irq(&zone->lru_lock);
1142	}
1143	return ret;
1144}
1145
1146/*
1147 * Are there way too many processes in the direct reclaim path already?
1148 */
1149static int too_many_isolated(struct zone *zone, int file,
1150		struct scan_control *sc)
1151{
1152	unsigned long inactive, isolated;
1153
1154	if (current_is_kswapd())
1155		return 0;
1156
1157	if (!global_reclaim(sc))
1158		return 0;
1159
1160	if (file) {
1161		inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1162		isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1163	} else {
1164		inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1165		isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1166	}
1167
1168	return isolated > inactive;
1169}
1170
1171static noinline_for_stack void
1172putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1173{
1174	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1175	struct zone *zone = lruvec_zone(lruvec);
1176	LIST_HEAD(pages_to_free);
1177
1178	/*
1179	 * Put back any unfreeable pages.
1180	 */
1181	while (!list_empty(page_list)) {
1182		struct page *page = lru_to_page(page_list);
1183		int lru;
1184
1185		VM_BUG_ON(PageLRU(page));
1186		list_del(&page->lru);
1187		if (unlikely(!page_evictable(page, NULL))) {
1188			spin_unlock_irq(&zone->lru_lock);
1189			putback_lru_page(page);
1190			spin_lock_irq(&zone->lru_lock);
1191			continue;
1192		}
1193
1194		lruvec = mem_cgroup_page_lruvec(page, zone);
1195
1196		SetPageLRU(page);
1197		lru = page_lru(page);
1198		add_page_to_lru_list(page, lruvec, lru);
1199
1200		if (is_active_lru(lru)) {
1201			int file = is_file_lru(lru);
1202			int numpages = hpage_nr_pages(page);
1203			reclaim_stat->recent_rotated[file] += numpages;
1204		}
1205		if (put_page_testzero(page)) {
1206			__ClearPageLRU(page);
1207			__ClearPageActive(page);
1208			del_page_from_lru_list(page, lruvec, lru);
1209
1210			if (unlikely(PageCompound(page))) {
1211				spin_unlock_irq(&zone->lru_lock);
1212				(*get_compound_page_dtor(page))(page);
1213				spin_lock_irq(&zone->lru_lock);
1214			} else
1215				list_add(&page->lru, &pages_to_free);
1216		}
1217	}
1218
1219	/*
1220	 * To save our caller's stack, now use input list for pages to free.
1221	 */
1222	list_splice(&pages_to_free, page_list);
1223}
1224
1225/*
1226 * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1227 * of reclaimed pages
1228 */
1229static noinline_for_stack unsigned long
1230shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1231		     struct scan_control *sc, enum lru_list lru)
1232{
1233	LIST_HEAD(page_list);
1234	unsigned long nr_scanned;
1235	unsigned long nr_reclaimed = 0;
1236	unsigned long nr_taken;
1237	unsigned long nr_dirty = 0;
1238	unsigned long nr_writeback = 0;
1239	isolate_mode_t isolate_mode = 0;
1240	int file = is_file_lru(lru);
1241	struct zone *zone = lruvec_zone(lruvec);
1242	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1243
1244	while (unlikely(too_many_isolated(zone, file, sc))) {
1245		congestion_wait(BLK_RW_ASYNC, HZ/10);
1246
1247		/* We are about to die and free our memory. Return now. */
1248		if (fatal_signal_pending(current))
1249			return SWAP_CLUSTER_MAX;
1250	}
1251
1252	lru_add_drain();
1253
1254	if (!sc->may_unmap)
1255		isolate_mode |= ISOLATE_UNMAPPED;
1256	if (!sc->may_writepage)
1257		isolate_mode |= ISOLATE_CLEAN;
1258
1259	spin_lock_irq(&zone->lru_lock);
1260
1261	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1262				     &nr_scanned, sc, isolate_mode, lru);
1263
1264	__mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1265	__mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1266
1267	if (global_reclaim(sc)) {
1268		zone->pages_scanned += nr_scanned;
1269		if (current_is_kswapd())
1270			__count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1271		else
1272			__count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1273	}
1274	spin_unlock_irq(&zone->lru_lock);
1275
1276	if (nr_taken == 0)
1277		return 0;
1278
1279	nr_reclaimed = shrink_page_list(&page_list, zone, sc,
1280						&nr_dirty, &nr_writeback);
1281
1282	spin_lock_irq(&zone->lru_lock);
1283
1284	reclaim_stat->recent_scanned[file] += nr_taken;
1285
1286	if (global_reclaim(sc)) {
1287		if (current_is_kswapd())
1288			__count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1289					       nr_reclaimed);
1290		else
1291			__count_zone_vm_events(PGSTEAL_DIRECT, zone,
1292					       nr_reclaimed);
1293	}
1294
1295	putback_inactive_pages(lruvec, &page_list);
1296
1297	__mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1298
1299	spin_unlock_irq(&zone->lru_lock);
1300
1301	free_hot_cold_page_list(&page_list, 1);
1302
1303	/*
1304	 * If reclaim is isolating dirty pages under writeback, it implies
1305	 * that the long-lived page allocation rate is exceeding the page
1306	 * laundering rate. Either the global limits are not being effective
1307	 * at throttling processes due to the page distribution throughout
1308	 * zones or there is heavy usage of a slow backing device. The
1309	 * only option is to throttle from reclaim context which is not ideal
1310	 * as there is no guarantee the dirtying process is throttled in the
1311	 * same way balance_dirty_pages() manages.
1312	 *
1313	 * This scales the number of dirty pages that must be under writeback
1314	 * before throttling depending on priority. It is a simple backoff
1315	 * function that has the most effect in the range DEF_PRIORITY to
1316	 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1317	 * in trouble and reclaim is considered to be in trouble.
1318	 *
1319	 * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1320	 * DEF_PRIORITY-1  50% must be PageWriteback
1321	 * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1322	 * ...
1323	 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1324	 *                     isolated page is PageWriteback
1325	 */
1326	if (nr_writeback && nr_writeback >=
1327			(nr_taken >> (DEF_PRIORITY - sc->priority)))
1328		wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1329
1330	trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1331		zone_idx(zone),
1332		nr_scanned, nr_reclaimed,
1333		sc->priority,
1334		trace_shrink_flags(file));
1335	return nr_reclaimed;
1336}
1337
1338/*
1339 * This moves pages from the active list to the inactive list.
1340 *
1341 * We move them the other way if the page is referenced by one or more
1342 * processes, from rmap.
1343 *
1344 * If the pages are mostly unmapped, the processing is fast and it is
1345 * appropriate to hold zone->lru_lock across the whole operation.  But if
1346 * the pages are mapped, the processing is slow (page_referenced()) so we
1347 * should drop zone->lru_lock around each page.  It's impossible to balance
1348 * this, so instead we remove the pages from the LRU while processing them.
1349 * It is safe to rely on PG_active against the non-LRU pages in here because
1350 * nobody will play with that bit on a non-LRU page.
1351 *
1352 * The downside is that we have to touch page->_count against each page.
1353 * But we had to alter page->flags anyway.
1354 */
1355
1356static void move_active_pages_to_lru(struct lruvec *lruvec,
1357				     struct list_head *list,
1358				     struct list_head *pages_to_free,
1359				     enum lru_list lru)
1360{
1361	struct zone *zone = lruvec_zone(lruvec);
1362	unsigned long pgmoved = 0;
1363	struct page *page;
1364	int nr_pages;
1365
1366	while (!list_empty(list)) {
1367		page = lru_to_page(list);
1368		lruvec = mem_cgroup_page_lruvec(page, zone);
1369
1370		VM_BUG_ON(PageLRU(page));
1371		SetPageLRU(page);
1372
1373		nr_pages = hpage_nr_pages(page);
1374		mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1375		list_move(&page->lru, &lruvec->lists[lru]);
1376		pgmoved += nr_pages;
1377
1378		if (put_page_testzero(page)) {
1379			__ClearPageLRU(page);
1380			__ClearPageActive(page);
1381			del_page_from_lru_list(page, lruvec, lru);
1382
1383			if (unlikely(PageCompound(page))) {
1384				spin_unlock_irq(&zone->lru_lock);
1385				(*get_compound_page_dtor(page))(page);
1386				spin_lock_irq(&zone->lru_lock);
1387			} else
1388				list_add(&page->lru, pages_to_free);
1389		}
1390	}
1391	__mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1392	if (!is_active_lru(lru))
1393		__count_vm_events(PGDEACTIVATE, pgmoved);
1394}
1395
1396static void shrink_active_list(unsigned long nr_to_scan,
1397			       struct lruvec *lruvec,
1398			       struct scan_control *sc,
1399			       enum lru_list lru)
1400{
1401	unsigned long nr_taken;
1402	unsigned long nr_scanned;
1403	unsigned long vm_flags;
1404	LIST_HEAD(l_hold);	/* The pages which were snipped off */
1405	LIST_HEAD(l_active);
1406	LIST_HEAD(l_inactive);
1407	struct page *page;
1408	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1409	unsigned long nr_rotated = 0;
1410	isolate_mode_t isolate_mode = 0;
1411	int file = is_file_lru(lru);
1412	struct zone *zone = lruvec_zone(lruvec);
1413
1414	lru_add_drain();
1415
1416	if (!sc->may_unmap)
1417		isolate_mode |= ISOLATE_UNMAPPED;
1418	if (!sc->may_writepage)
1419		isolate_mode |= ISOLATE_CLEAN;
1420
1421	spin_lock_irq(&zone->lru_lock);
1422
1423	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1424				     &nr_scanned, sc, isolate_mode, lru);
1425	if (global_reclaim(sc))
1426		zone->pages_scanned += nr_scanned;
1427
1428	reclaim_stat->recent_scanned[file] += nr_taken;
1429
1430	__count_zone_vm_events(PGREFILL, zone, nr_scanned);
1431	__mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1432	__mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1433	spin_unlock_irq(&zone->lru_lock);
1434
1435	while (!list_empty(&l_hold)) {
1436		cond_resched();
1437		page = lru_to_page(&l_hold);
1438		list_del(&page->lru);
1439
1440		if (unlikely(!page_evictable(page, NULL))) {
1441			putback_lru_page(page);
1442			continue;
1443		}
1444
1445		if (unlikely(buffer_heads_over_limit)) {
1446			if (page_has_private(page) && trylock_page(page)) {
1447				if (page_has_private(page))
1448					try_to_release_page(page, 0);
1449				unlock_page(page);
1450			}
1451		}
1452
1453		if (page_referenced(page, 0, sc->target_mem_cgroup,
1454				    &vm_flags)) {
1455			nr_rotated += hpage_nr_pages(page);
1456			/*
1457			 * Identify referenced, file-backed active pages and
1458			 * give them one more trip around the active list. So
1459			 * that executable code get better chances to stay in
1460			 * memory under moderate memory pressure.  Anon pages
1461			 * are not likely to be evicted by use-once streaming
1462			 * IO, plus JVM can create lots of anon VM_EXEC pages,
1463			 * so we ignore them here.
1464			 */
1465			if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1466				list_add(&page->lru, &l_active);
1467				continue;
1468			}
1469		}
1470
1471		ClearPageActive(page);	/* we are de-activating */
1472		list_add(&page->lru, &l_inactive);
1473	}
1474
1475	/*
1476	 * Move pages back to the lru list.
1477	 */
1478	spin_lock_irq(&zone->lru_lock);
1479	/*
1480	 * Count referenced pages from currently used mappings as rotated,
1481	 * even though only some of them are actually re-activated.  This
1482	 * helps balance scan pressure between file and anonymous pages in
1483	 * get_scan_ratio.
1484	 */
1485	reclaim_stat->recent_rotated[file] += nr_rotated;
1486
1487	move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1488	move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1489	__mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1490	spin_unlock_irq(&zone->lru_lock);
1491
1492	free_hot_cold_page_list(&l_hold, 1);
1493}
1494
1495#ifdef CONFIG_SWAP
1496static int inactive_anon_is_low_global(struct zone *zone)
1497{
1498	unsigned long active, inactive;
1499
1500	active = zone_page_state(zone, NR_ACTIVE_ANON);
1501	inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1502
1503	if (inactive * zone->inactive_ratio < active)
1504		return 1;
1505
1506	return 0;
1507}
1508
1509/**
1510 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1511 * @lruvec: LRU vector to check
1512 *
1513 * Returns true if the zone does not have enough inactive anon pages,
1514 * meaning some active anon pages need to be deactivated.
1515 */
1516static int inactive_anon_is_low(struct lruvec *lruvec)
1517{
1518	/*
1519	 * If we don't have swap space, anonymous page deactivation
1520	 * is pointless.
1521	 */
1522	if (!total_swap_pages)
1523		return 0;
1524
1525	if (!mem_cgroup_disabled())
1526		return mem_cgroup_inactive_anon_is_low(lruvec);
1527
1528	return inactive_anon_is_low_global(lruvec_zone(lruvec));
1529}
1530#else
1531static inline int inactive_anon_is_low(struct lruvec *lruvec)
1532{
1533	return 0;
1534}
1535#endif
1536
1537static int inactive_file_is_low_global(struct zone *zone)
1538{
1539	unsigned long active, inactive;
1540
1541	active = zone_page_state(zone, NR_ACTIVE_FILE);
1542	inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1543
1544	return (active > inactive);
1545}
1546
1547/**
1548 * inactive_file_is_low - check if file pages need to be deactivated
1549 * @lruvec: LRU vector to check
1550 *
1551 * When the system is doing streaming IO, memory pressure here
1552 * ensures that active file pages get deactivated, until more
1553 * than half of the file pages are on the inactive list.
1554 *
1555 * Once we get to that situation, protect the system's working
1556 * set from being evicted by disabling active file page aging.
1557 *
1558 * This uses a different ratio than the anonymous pages, because
1559 * the page cache uses a use-once replacement algorithm.
1560 */
1561static int inactive_file_is_low(struct lruvec *lruvec)
1562{
1563	if (!mem_cgroup_disabled())
1564		return mem_cgroup_inactive_file_is_low(lruvec);
1565
1566	return inactive_file_is_low_global(lruvec_zone(lruvec));
1567}
1568
1569static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1570{
1571	if (is_file_lru(lru))
1572		return inactive_file_is_low(lruvec);
1573	else
1574		return inactive_anon_is_low(lruvec);
1575}
1576
1577static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1578				 struct lruvec *lruvec, struct scan_control *sc)
1579{
1580	if (is_active_lru(lru)) {
1581		if (inactive_list_is_low(lruvec, lru))
1582			shrink_active_list(nr_to_scan, lruvec, sc, lru);
1583		return 0;
1584	}
1585
1586	return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1587}
1588
1589static int vmscan_swappiness(struct scan_control *sc)
1590{
1591	if (global_reclaim(sc))
1592		return vm_swappiness;
1593	return mem_cgroup_swappiness(sc->target_mem_cgroup);
1594}
1595
1596/*
1597 * Determine how aggressively the anon and file LRU lists should be
1598 * scanned.  The relative value of each set of LRU lists is determined
1599 * by looking at the fraction of the pages scanned we did rotate back
1600 * onto the active list instead of evict.
1601 *
1602 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1603 */
1604static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1605			   unsigned long *nr)
1606{
1607	unsigned long anon, file, free;
1608	unsigned long anon_prio, file_prio;
1609	unsigned long ap, fp;
1610	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1611	u64 fraction[2], denominator;
1612	enum lru_list lru;
1613	int noswap = 0;
1614	bool force_scan = false;
1615	struct zone *zone = lruvec_zone(lruvec);
1616
1617	/*
1618	 * If the zone or memcg is small, nr[l] can be 0.  This
1619	 * results in no scanning on this priority and a potential
1620	 * priority drop.  Global direct reclaim can go to the next
1621	 * zone and tends to have no problems. Global kswapd is for
1622	 * zone balancing and it needs to scan a minimum amount. When
1623	 * reclaiming for a memcg, a priority drop can cause high
1624	 * latencies, so it's better to scan a minimum amount there as
1625	 * well.
1626	 */
1627	if (current_is_kswapd() && zone->all_unreclaimable)
1628		force_scan = true;
1629	if (!global_reclaim(sc))
1630		force_scan = true;
1631
1632	/* If we have no swap space, do not bother scanning anon pages. */
1633	if (!sc->may_swap || (nr_swap_pages <= 0)) {
1634		noswap = 1;
1635		fraction[0] = 0;
1636		fraction[1] = 1;
1637		denominator = 1;
1638		goto out;
1639	}
1640
1641	anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1642		get_lru_size(lruvec, LRU_INACTIVE_ANON);
1643	file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1644		get_lru_size(lruvec, LRU_INACTIVE_FILE);
1645
1646	if (global_reclaim(sc)) {
1647		free  = zone_page_state(zone, NR_FREE_PAGES);
1648		/* If we have very few page cache pages,
1649		   force-scan anon pages. */
1650		if (unlikely(file + free <= high_wmark_pages(zone))) {
1651			fraction[0] = 1;
1652			fraction[1] = 0;
1653			denominator = 1;
1654			goto out;
1655		}
1656	}
1657
1658	/*
1659	 * With swappiness at 100, anonymous and file have the same priority.
1660	 * This scanning priority is essentially the inverse of IO cost.
1661	 */
1662	anon_prio = vmscan_swappiness(sc);
1663	file_prio = 200 - anon_prio;
1664
1665	/*
1666	 * OK, so we have swap space and a fair amount of page cache
1667	 * pages.  We use the recently rotated / recently scanned
1668	 * ratios to determine how valuable each cache is.
1669	 *
1670	 * Because workloads change over time (and to avoid overflow)
1671	 * we keep these statistics as a floating average, which ends
1672	 * up weighing recent references more than old ones.
1673	 *
1674	 * anon in [0], file in [1]
1675	 */
1676	spin_lock_irq(&zone->lru_lock);
1677	if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1678		reclaim_stat->recent_scanned[0] /= 2;
1679		reclaim_stat->recent_rotated[0] /= 2;
1680	}
1681
1682	if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1683		reclaim_stat->recent_scanned[1] /= 2;
1684		reclaim_stat->recent_rotated[1] /= 2;
1685	}
1686
1687	/*
1688	 * The amount of pressure on anon vs file pages is inversely
1689	 * proportional to the fraction of recently scanned pages on
1690	 * each list that were recently referenced and in active use.
1691	 */
1692	ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1693	ap /= reclaim_stat->recent_rotated[0] + 1;
1694
1695	fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1696	fp /= reclaim_stat->recent_rotated[1] + 1;
1697	spin_unlock_irq(&zone->lru_lock);
1698
1699	fraction[0] = ap;
1700	fraction[1] = fp;
1701	denominator = ap + fp + 1;
1702out:
1703	for_each_evictable_lru(lru) {
1704		int file = is_file_lru(lru);
1705		unsigned long scan;
1706
1707		scan = get_lru_size(lruvec, lru);
1708		if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1709			scan >>= sc->priority;
1710			if (!scan && force_scan)
1711				scan = SWAP_CLUSTER_MAX;
1712			scan = div64_u64(scan * fraction[file], denominator);
1713		}
1714		nr[lru] = scan;
1715	}
1716}
1717
1718/* Use reclaim/compaction for costly allocs or under memory pressure */
1719static bool in_reclaim_compaction(struct scan_control *sc)
1720{
1721	if (COMPACTION_BUILD && sc->order &&
1722			(sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1723			 sc->priority < DEF_PRIORITY - 2))
1724		return true;
1725
1726	return false;
1727}
1728
1729/*
1730 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1731 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1732 * true if more pages should be reclaimed such that when the page allocator
1733 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1734 * It will give up earlier than that if there is difficulty reclaiming pages.
1735 */
1736static inline bool should_continue_reclaim(struct lruvec *lruvec,
1737					unsigned long nr_reclaimed,
1738					unsigned long nr_scanned,
1739					struct scan_control *sc)
1740{
1741	unsigned long pages_for_compaction;
1742	unsigned long inactive_lru_pages;
1743
1744	/* If not in reclaim/compaction mode, stop */
1745	if (!in_reclaim_compaction(sc))
1746		return false;
1747
1748	/* Consider stopping depending on scan and reclaim activity */
1749	if (sc->gfp_mask & __GFP_REPEAT) {
1750		/*
1751		 * For __GFP_REPEAT allocations, stop reclaiming if the
1752		 * full LRU list has been scanned and we are still failing
1753		 * to reclaim pages. This full LRU scan is potentially
1754		 * expensive but a __GFP_REPEAT caller really wants to succeed
1755		 */
1756		if (!nr_reclaimed && !nr_scanned)
1757			return false;
1758	} else {
1759		/*
1760		 * For non-__GFP_REPEAT allocations which can presumably
1761		 * fail without consequence, stop if we failed to reclaim
1762		 * any pages from the last SWAP_CLUSTER_MAX number of
1763		 * pages that were scanned. This will return to the
1764		 * caller faster at the risk reclaim/compaction and
1765		 * the resulting allocation attempt fails
1766		 */
1767		if (!nr_reclaimed)
1768			return false;
1769	}
1770
1771	/*
1772	 * If we have not reclaimed enough pages for compaction and the
1773	 * inactive lists are large enough, continue reclaiming
1774	 */
1775	pages_for_compaction = (2UL << sc->order);
1776	inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1777	if (nr_swap_pages > 0)
1778		inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON);
1779	if (sc->nr_reclaimed < pages_for_compaction &&
1780			inactive_lru_pages > pages_for_compaction)
1781		return true;
1782
1783	/* If compaction would go ahead or the allocation would succeed, stop */
1784	switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
1785	case COMPACT_PARTIAL:
1786	case COMPACT_CONTINUE:
1787		return false;
1788	default:
1789		return true;
1790	}
1791}
1792
1793/*
1794 * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1795 */
1796static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1797{
1798	unsigned long nr[NR_LRU_LISTS];
1799	unsigned long nr_to_scan;
1800	enum lru_list lru;
1801	unsigned long nr_reclaimed, nr_scanned;
1802	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1803	struct blk_plug plug;
1804
1805restart:
1806	nr_reclaimed = 0;
1807	nr_scanned = sc->nr_scanned;
1808	get_scan_count(lruvec, sc, nr);
1809
1810	blk_start_plug(&plug);
1811	while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1812					nr[LRU_INACTIVE_FILE]) {
1813		for_each_evictable_lru(lru) {
1814			if (nr[lru]) {
1815				nr_to_scan = min_t(unsigned long,
1816						   nr[lru], SWAP_CLUSTER_MAX);
1817				nr[lru] -= nr_to_scan;
1818
1819				nr_reclaimed += shrink_list(lru, nr_to_scan,
1820							    lruvec, sc);
1821			}
1822		}
1823		/*
1824		 * On large memory systems, scan >> priority can become
1825		 * really large. This is fine for the starting priority;
1826		 * we want to put equal scanning pressure on each zone.
1827		 * However, if the VM has a harder time of freeing pages,
1828		 * with multiple processes reclaiming pages, the total
1829		 * freeing target can get unreasonably large.
1830		 */
1831		if (nr_reclaimed >= nr_to_reclaim &&
1832		    sc->priority < DEF_PRIORITY)
1833			break;
1834	}
1835	blk_finish_plug(&plug);
1836	sc->nr_reclaimed += nr_reclaimed;
1837
1838	/*
1839	 * Even if we did not try to evict anon pages at all, we want to
1840	 * rebalance the anon lru active/inactive ratio.
1841	 */
1842	if (inactive_anon_is_low(lruvec))
1843		shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1844				   sc, LRU_ACTIVE_ANON);
1845
1846	/* reclaim/compaction might need reclaim to continue */
1847	if (should_continue_reclaim(lruvec, nr_reclaimed,
1848				    sc->nr_scanned - nr_scanned, sc))
1849		goto restart;
1850
1851	throttle_vm_writeout(sc->gfp_mask);
1852}
1853
1854static void shrink_zone(struct zone *zone, struct scan_control *sc)
1855{
1856	struct mem_cgroup *root = sc->target_mem_cgroup;
1857	struct mem_cgroup_reclaim_cookie reclaim = {
1858		.zone = zone,
1859		.priority = sc->priority,
1860	};
1861	struct mem_cgroup *memcg;
1862
1863	memcg = mem_cgroup_iter(root, NULL, &reclaim);
1864	do {
1865		struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1866
1867		shrink_lruvec(lruvec, sc);
1868
1869		/*
1870		 * Limit reclaim has historically picked one memcg and
1871		 * scanned it with decreasing priority levels until
1872		 * nr_to_reclaim had been reclaimed.  This priority
1873		 * cycle is thus over after a single memcg.
1874		 *
1875		 * Direct reclaim and kswapd, on the other hand, have
1876		 * to scan all memory cgroups to fulfill the overall
1877		 * scan target for the zone.
1878		 */
1879		if (!global_reclaim(sc)) {
1880			mem_cgroup_iter_break(root, memcg);
1881			break;
1882		}
1883		memcg = mem_cgroup_iter(root, memcg, &reclaim);
1884	} while (memcg);
1885}
1886
1887/* Returns true if compaction should go ahead for a high-order request */
1888static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1889{
1890	unsigned long balance_gap, watermark;
1891	bool watermark_ok;
1892
1893	/* Do not consider compaction for orders reclaim is meant to satisfy */
1894	if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1895		return false;
1896
1897	/*
1898	 * Compaction takes time to run and there are potentially other
1899	 * callers using the pages just freed. Continue reclaiming until
1900	 * there is a buffer of free pages available to give compaction
1901	 * a reasonable chance of completing and allocating the page
1902	 */
1903	balance_gap = min(low_wmark_pages(zone),
1904		(zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1905			KSWAPD_ZONE_BALANCE_GAP_RATIO);
1906	watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1907	watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1908
1909	/*
1910	 * If compaction is deferred, reclaim up to a point where
1911	 * compaction will have a chance of success when re-enabled
1912	 */
1913	if (compaction_deferred(zone, sc->order))
1914		return watermark_ok;
1915
1916	/* If compaction is not ready to start, keep reclaiming */
1917	if (!compaction_suitable(zone, sc->order))
1918		return false;
1919
1920	return watermark_ok;
1921}
1922
1923/*
1924 * This is the direct reclaim path, for page-allocating processes.  We only
1925 * try to reclaim pages from zones which will satisfy the caller's allocation
1926 * request.
1927 *
1928 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1929 * Because:
1930 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1931 *    allocation or
1932 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1933 *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1934 *    zone defense algorithm.
1935 *
1936 * If a zone is deemed to be full of pinned pages then just give it a light
1937 * scan then give up on it.
1938 *
1939 * This function returns true if a zone is being reclaimed for a costly
1940 * high-order allocation and compaction is ready to begin. This indicates to
1941 * the caller that it should consider retrying the allocation instead of
1942 * further reclaim.
1943 */
1944static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1945{
1946	struct zoneref *z;
1947	struct zone *zone;
1948	unsigned long nr_soft_reclaimed;
1949	unsigned long nr_soft_scanned;
1950	bool aborted_reclaim = false;
1951
1952	/*
1953	 * If the number of buffer_heads in the machine exceeds the maximum
1954	 * allowed level, force direct reclaim to scan the highmem zone as
1955	 * highmem pages could be pinning lowmem pages storing buffer_heads
1956	 */
1957	if (buffer_heads_over_limit)
1958		sc->gfp_mask |= __GFP_HIGHMEM;
1959
1960	for_each_zone_zonelist_nodemask(zone, z, zonelist,
1961					gfp_zone(sc->gfp_mask), sc->nodemask) {
1962		if (!populated_zone(zone))
1963			continue;
1964		/*
1965		 * Take care memory controller reclaiming has small influence
1966		 * to global LRU.
1967		 */
1968		if (global_reclaim(sc)) {
1969			if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1970				continue;
1971			if (zone->all_unreclaimable &&
1972					sc->priority != DEF_PRIORITY)
1973				continue;	/* Let kswapd poll it */
1974			if (COMPACTION_BUILD) {
1975				/*
1976				 * If we already have plenty of memory free for
1977				 * compaction in this zone, don't free any more.
1978				 * Even though compaction is invoked for any
1979				 * non-zero order, only frequent costly order
1980				 * reclamation is disruptive enough to become a
1981				 * noticeable problem, like transparent huge
1982				 * page allocations.
1983				 */
1984				if (compaction_ready(zone, sc)) {
1985					aborted_reclaim = true;
1986					continue;
1987				}
1988			}
1989			/*
1990			 * This steals pages from memory cgroups over softlimit
1991			 * and returns the number of reclaimed pages and
1992			 * scanned pages. This works for global memory pressure
1993			 * and balancing, not for a memcg's limit.
1994			 */
1995			nr_soft_scanned = 0;
1996			nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
1997						sc->order, sc->gfp_mask,
1998						&nr_soft_scanned);
1999			sc->nr_reclaimed += nr_soft_reclaimed;
2000			sc->nr_scanned += nr_soft_scanned;
2001			/* need some check for avoid more shrink_zone() */
2002		}
2003
2004		shrink_zone(zone, sc);
2005	}
2006
2007	return aborted_reclaim;
2008}
2009
2010static bool zone_reclaimable(struct zone *zone)
2011{
2012	return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2013}
2014
2015/* All zones in zonelist are unreclaimable? */
2016static bool all_unreclaimable(struct zonelist *zonelist,
2017		struct scan_control *sc)
2018{
2019	struct zoneref *z;
2020	struct zone *zone;
2021
2022	for_each_zone_zonelist_nodemask(zone, z, zonelist,
2023			gfp_zone(sc->gfp_mask), sc->nodemask) {
2024		if (!populated_zone(zone))
2025			continue;
2026		if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2027			continue;
2028		if (!zone->all_unreclaimable)
2029			return false;
2030	}
2031
2032	return true;
2033}
2034
2035/*
2036 * This is the main entry point to direct page reclaim.
2037 *
2038 * If a full scan of the inactive list fails to free enough memory then we
2039 * are "out of memory" and something needs to be killed.
2040 *
2041 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2042 * high - the zone may be full of dirty or under-writeback pages, which this
2043 * caller can't do much about.  We kick the writeback threads and take explicit
2044 * naps in the hope that some of these pages can be written.  But if the
2045 * allocating task holds filesystem locks which prevent writeout this might not
2046 * work, and the allocation attempt will fail.
2047 *
2048 * returns:	0, if no pages reclaimed
2049 * 		else, the number of pages reclaimed
2050 */
2051static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2052					struct scan_control *sc,
2053					struct shrink_control *shrink)
2054{
2055	unsigned long total_scanned = 0;
2056	struct reclaim_state *reclaim_state = current->reclaim_state;
2057	struct zoneref *z;
2058	struct zone *zone;
2059	unsigned long writeback_threshold;
2060	bool aborted_reclaim;
2061
2062	delayacct_freepages_start();
2063
2064	if (global_reclaim(sc))
2065		count_vm_event(ALLOCSTALL);
2066
2067	do {
2068		sc->nr_scanned = 0;
2069		aborted_reclaim = shrink_zones(zonelist, sc);
2070
2071		/*
2072		 * Don't shrink slabs when reclaiming memory from
2073		 * over limit cgroups
2074		 */
2075		if (global_reclaim(sc)) {
2076			unsigned long lru_pages = 0;
2077			for_each_zone_zonelist(zone, z, zonelist,
2078					gfp_zone(sc->gfp_mask)) {
2079				if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2080					continue;
2081
2082				lru_pages += zone_reclaimable_pages(zone);
2083			}
2084
2085			shrink_slab(shrink, sc->nr_scanned, lru_pages);
2086			if (reclaim_state) {
2087				sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2088				reclaim_state->reclaimed_slab = 0;
2089			}
2090		}
2091		total_scanned += sc->nr_scanned;
2092		if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2093			goto out;
2094
2095		/*
2096		 * Try to write back as many pages as we just scanned.  This
2097		 * tends to cause slow streaming writers to write data to the
2098		 * disk smoothly, at the dirtying rate, which is nice.   But
2099		 * that's undesirable in laptop mode, where we *want* lumpy
2100		 * writeout.  So in laptop mode, write out the whole world.
2101		 */
2102		writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2103		if (total_scanned > writeback_threshold) {
2104			wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2105						WB_REASON_TRY_TO_FREE_PAGES);
2106			sc->may_writepage = 1;
2107		}
2108
2109		/* Take a nap, wait for some writeback to complete */
2110		if (!sc->hibernation_mode && sc->nr_scanned &&
2111		    sc->priority < DEF_PRIORITY - 2) {
2112			struct zone *preferred_zone;
2113
2114			first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2115						&cpuset_current_mems_allowed,
2116						&preferred_zone);
2117			wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2118		}
2119	} while (--sc->priority >= 0);
2120
2121out:
2122	delayacct_freepages_end();
2123
2124	if (sc->nr_reclaimed)
2125		return sc->nr_reclaimed;
2126
2127	/*
2128	 * As hibernation is going on, kswapd is freezed so that it can't mark
2129	 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2130	 * check.
2131	 */
2132	if (oom_killer_disabled)
2133		return 0;
2134
2135	/* Aborted reclaim to try compaction? don't OOM, then */
2136	if (aborted_reclaim)
2137		return 1;
2138
2139	/* top priority shrink_zones still had more to do? don't OOM, then */
2140	if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2141		return 1;
2142
2143	return 0;
2144}
2145
2146unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2147				gfp_t gfp_mask, nodemask_t *nodemask)
2148{
2149	unsigned long nr_reclaimed;
2150	struct scan_control sc = {
2151		.gfp_mask = gfp_mask,
2152		.may_writepage = !laptop_mode,
2153		.nr_to_reclaim = SWAP_CLUSTER_MAX,
2154		.may_unmap = 1,
2155		.may_swap = 1,
2156		.order = order,
2157		.priority = DEF_PRIORITY,
2158		.target_mem_cgroup = NULL,
2159		.nodemask = nodemask,
2160	};
2161	struct shrink_control shrink = {
2162		.gfp_mask = sc.gfp_mask,
2163	};
2164
2165	trace_mm_vmscan_direct_reclaim_begin(order,
2166				sc.may_writepage,
2167				gfp_mask);
2168
2169	nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2170
2171	trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2172
2173	return nr_reclaimed;
2174}
2175
2176#ifdef CONFIG_CGROUP_MEM_RES_CTLR
2177
2178unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2179						gfp_t gfp_mask, bool noswap,
2180						struct zone *zone,
2181						unsigned long *nr_scanned)
2182{
2183	struct scan_control sc = {
2184		.nr_scanned = 0,
2185		.nr_to_reclaim = SWAP_CLUSTER_MAX,
2186		.may_writepage = !laptop_mode,
2187		.may_unmap = 1,
2188		.may_swap = !noswap,
2189		.order = 0,
2190		.priority = 0,
2191		.target_mem_cgroup = memcg,
2192	};
2193	struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2194
2195	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2196			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2197
2198	trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2199						      sc.may_writepage,
2200						      sc.gfp_mask);
2201
2202	/*
2203	 * NOTE: Although we can get the priority field, using it
2204	 * here is not a good idea, since it limits the pages we can scan.
2205	 * if we don't reclaim here, the shrink_zone from balance_pgdat
2206	 * will pick up pages from other mem cgroup's as well. We hack
2207	 * the priority and make it zero.
2208	 */
2209	shrink_lruvec(lruvec, &sc);
2210
2211	trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2212
2213	*nr_scanned = sc.nr_scanned;
2214	return sc.nr_reclaimed;
2215}
2216
2217unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2218					   gfp_t gfp_mask,
2219					   bool noswap)
2220{
2221	struct zonelist *zonelist;
2222	unsigned long nr_reclaimed;
2223	int nid;
2224	struct scan_control sc = {
2225		.may_writepage = !laptop_mode,
2226		.may_unmap = 1,
2227		.may_swap = !noswap,
2228		.nr_to_reclaim = SWAP_CLUSTER_MAX,
2229		.order = 0,
2230		.priority = DEF_PRIORITY,
2231		.target_mem_cgroup = memcg,
2232		.nodemask = NULL, /* we don't care the placement */
2233		.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2234				(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2235	};
2236	struct shrink_control shrink = {
2237		.gfp_mask = sc.gfp_mask,
2238	};
2239
2240	/*
2241	 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2242	 * take care of from where we get pages. So the node where we start the
2243	 * scan does not need to be the current node.
2244	 */
2245	nid = mem_cgroup_select_victim_node(memcg);
2246
2247	zonelist = NODE_DATA(nid)->node_zonelists;
2248
2249	trace_mm_vmscan_memcg_reclaim_begin(0,
2250					    sc.may_writepage,
2251					    sc.gfp_mask);
2252
2253	nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2254
2255	trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2256
2257	return nr_reclaimed;
2258}
2259#endif
2260
2261static void age_active_anon(struct zone *zone, struct scan_control *sc)
2262{
2263	struct mem_cgroup *memcg;
2264
2265	if (!total_swap_pages)
2266		return;
2267
2268	memcg = mem_cgroup_iter(NULL, NULL, NULL);
2269	do {
2270		struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2271
2272		if (inactive_anon_is_low(lruvec))
2273			shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2274					   sc, LRU_ACTIVE_ANON);
2275
2276		memcg = mem_cgroup_iter(NULL, memcg, NULL);
2277	} while (memcg);
2278}
2279
2280/*
2281 * pgdat_balanced is used when checking if a node is balanced for high-order
2282 * allocations. Only zones that meet watermarks and are in a zone allowed
2283 * by the callers classzone_idx are added to balanced_pages. The total of
2284 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2285 * for the node to be considered balanced. Forcing all zones to be balanced
2286 * for high orders can cause excessive reclaim when there are imbalanced zones.
2287 * The choice of 25% is due to
2288 *   o a 16M DMA zone that is balanced will not balance a zone on any
2289 *     reasonable sized machine
2290 *   o On all other machines, the top zone must be at least a reasonable
2291 *     percentage of the middle zones. For example, on 32-bit x86, highmem
2292 *     would need to be at least 256M for it to be balance a whole node.
2293 *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2294 *     to balance a node on its own. These seemed like reasonable ratios.
2295 */
2296static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2297						int classzone_idx)
2298{
2299	unsigned long present_pages = 0;
2300	int i;
2301
2302	for (i = 0; i <= classzone_idx; i++)
2303		present_pages += pgdat->node_zones[i].present_pages;
2304
2305	/* A special case here: if zone has no page, we think it's balanced */
2306	return balanced_pages >= (present_pages >> 2);
2307}
2308
2309/* is kswapd sleeping prematurely? */
2310static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2311					int classzone_idx)
2312{
2313	int i;
2314	unsigned long balanced = 0;
2315	bool all_zones_ok = true;
2316
2317	/* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2318	if (remaining)
2319		return true;
2320
2321	/* Check the watermark levels */
2322	for (i = 0; i <= classzone_idx; i++) {
2323		struct zone *zone = pgdat->node_zones + i;
2324
2325		if (!populated_zone(zone))
2326			continue;
2327
2328		/*
2329		 * balance_pgdat() skips over all_unreclaimable after
2330		 * DEF_PRIORITY. Effectively, it considers them balanced so
2331		 * they must be considered balanced here as well if kswapd
2332		 * is to sleep
2333		 */
2334		if (zone->all_unreclaimable) {
2335			balanced += zone->present_pages;
2336			continue;
2337		}
2338
2339		if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2340							i, 0))
2341			all_zones_ok = false;
2342		else
2343			balanced += zone->present_pages;
2344	}
2345
2346	/*
2347	 * For high-order requests, the balanced zones must contain at least
2348	 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2349	 * must be balanced
2350	 */
2351	if (order)
2352		return !pgdat_balanced(pgdat, balanced, classzone_idx);
2353	else
2354		return !all_zones_ok;
2355}
2356
2357/*
2358 * For kswapd, balance_pgdat() will work across all this node's zones until
2359 * they are all at high_wmark_pages(zone).
2360 *
2361 * Returns the final order kswapd was reclaiming at
2362 *
2363 * There is special handling here for zones which are full of pinned pages.
2364 * This can happen if the pages are all mlocked, or if they are all used by
2365 * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2366 * What we do is to detect the case where all pages in the zone have been
2367 * scanned twice and there has been zero successful reclaim.  Mark the zone as
2368 * dead and from now on, only perform a short scan.  Basically we're polling
2369 * the zone for when the problem goes away.
2370 *
2371 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2372 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2373 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2374 * lower zones regardless of the number of free pages in the lower zones. This
2375 * interoperates with the page allocator fallback scheme to ensure that aging
2376 * of pages is balanced across the zones.
2377 */
2378static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2379							int *classzone_idx)
2380{
2381	int all_zones_ok;
2382	unsigned long balanced;
2383	int i;
2384	int end_zone = 0;	/* Inclusive.  0 = ZONE_DMA */
2385	unsigned long total_scanned;
2386	struct reclaim_state *reclaim_state = current->reclaim_state;
2387	unsigned long nr_soft_reclaimed;
2388	unsigned long nr_soft_scanned;
2389	struct scan_control sc = {
2390		.gfp_mask = GFP_KERNEL,
2391		.may_unmap = 1,
2392		.may_swap = 1,
2393		/*
2394		 * kswapd doesn't want to be bailed out while reclaim. because
2395		 * we want to put equal scanning pressure on each zone.
2396		 */
2397		.nr_to_reclaim = ULONG_MAX,
2398		.order = order,
2399		.target_mem_cgroup = NULL,
2400	};
2401	struct shrink_control shrink = {
2402		.gfp_mask = sc.gfp_mask,
2403	};
2404loop_again:
2405	total_scanned = 0;
2406	sc.priority = DEF_PRIORITY;
2407	sc.nr_reclaimed = 0;
2408	sc.may_writepage = !laptop_mode;
2409	count_vm_event(PAGEOUTRUN);
2410
2411	do {
2412		unsigned long lru_pages = 0;
2413		int has_under_min_watermark_zone = 0;
2414
2415		all_zones_ok = 1;
2416		balanced = 0;
2417
2418		/*
2419		 * Scan in the highmem->dma direction for the highest
2420		 * zone which needs scanning
2421		 */
2422		for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2423			struct zone *zone = pgdat->node_zones + i;
2424
2425			if (!populated_zone(zone))
2426				continue;
2427
2428			if (zone->all_unreclaimable &&
2429			    sc.priority != DEF_PRIORITY)
2430				continue;
2431
2432			/*
2433			 * Do some background aging of the anon list, to give
2434			 * pages a chance to be referenced before reclaiming.
2435			 */
2436			age_active_anon(zone, &sc);
2437
2438			/*
2439			 * If the number of buffer_heads in the machine
2440			 * exceeds the maximum allowed level and this node
2441			 * has a highmem zone, force kswapd to reclaim from
2442			 * it to relieve lowmem pressure.
2443			 */
2444			if (buffer_heads_over_limit && is_highmem_idx(i)) {
2445				end_zone = i;
2446				break;
2447			}
2448
2449			if (!zone_watermark_ok_safe(zone, order,
2450					high_wmark_pages(zone), 0, 0)) {
2451				end_zone = i;
2452				break;
2453			} else {
2454				/* If balanced, clear the congested flag */
2455				zone_clear_flag(zone, ZONE_CONGESTED);
2456			}
2457		}
2458		if (i < 0)
2459			goto out;
2460
2461		for (i = 0; i <= end_zone; i++) {
2462			struct zone *zone = pgdat->node_zones + i;
2463
2464			lru_pages += zone_reclaimable_pages(zone);
2465		}
2466
2467		/*
2468		 * Now scan the zone in the dma->highmem direction, stopping
2469		 * at the last zone which needs scanning.
2470		 *
2471		 * We do this because the page allocator works in the opposite
2472		 * direction.  This prevents the page allocator from allocating
2473		 * pages behind kswapd's direction of progress, which would
2474		 * cause too much scanning of the lower zones.
2475		 */
2476		for (i = 0; i <= end_zone; i++) {
2477			struct zone *zone = pgdat->node_zones + i;
2478			int nr_slab, testorder;
2479			unsigned long balance_gap;
2480
2481			if (!populated_zone(zone))
2482				continue;
2483
2484			if (zone->all_unreclaimable &&
2485			    sc.priority != DEF_PRIORITY)
2486				continue;
2487
2488			sc.nr_scanned = 0;
2489
2490			nr_soft_scanned = 0;
2491			/*
2492			 * Call soft limit reclaim before calling shrink_zone.
2493			 */
2494			nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2495							order, sc.gfp_mask,
2496							&nr_soft_scanned);
2497			sc.nr_reclaimed += nr_soft_reclaimed;
2498			total_scanned += nr_soft_scanned;
2499
2500			/*
2501			 * We put equal pressure on every zone, unless
2502			 * one zone has way too many pages free
2503			 * already. The "too many pages" is defined
2504			 * as the high wmark plus a "gap" where the
2505			 * gap is either the low watermark or 1%
2506			 * of the zone, whichever is smaller.
2507			 */
2508			balance_gap = min(low_wmark_pages(zone),
2509				(zone->present_pages +
2510					KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2511				KSWAPD_ZONE_BALANCE_GAP_RATIO);
2512			/*
2513			 * Kswapd reclaims only single pages with compaction
2514			 * enabled. Trying too hard to reclaim until contiguous
2515			 * free pages have become available can hurt performance
2516			 * by evicting too much useful data from memory.
2517			 * Do not reclaim more than needed for compaction.
2518			 */
2519			testorder = order;
2520			if (COMPACTION_BUILD && order &&
2521					compaction_suitable(zone, order) !=
2522						COMPACT_SKIPPED)
2523				testorder = 0;
2524
2525			if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2526				    !zone_watermark_ok_safe(zone, testorder,
2527					high_wmark_pages(zone) + balance_gap,
2528					end_zone, 0)) {
2529				shrink_zone(zone, &sc);
2530
2531				reclaim_state->reclaimed_slab = 0;
2532				nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2533				sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2534				total_scanned += sc.nr_scanned;
2535
2536				if (nr_slab == 0 && !zone_reclaimable(zone))
2537					zone->all_unreclaimable = 1;
2538			}
2539
2540			/*
2541			 * If we've done a decent amount of scanning and
2542			 * the reclaim ratio is low, start doing writepage
2543			 * even in laptop mode
2544			 */
2545			if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2546			    total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2547				sc.may_writepage = 1;
2548
2549			if (zone->all_unreclaimable) {
2550				if (end_zone && end_zone == i)
2551					end_zone--;
2552				continue;
2553			}
2554
2555			if (!zone_watermark_ok_safe(zone, testorder,
2556					high_wmark_pages(zone), end_zone, 0)) {
2557				all_zones_ok = 0;
2558				/*
2559				 * We are still under min water mark.  This
2560				 * means that we have a GFP_ATOMIC allocation
2561				 * failure risk. Hurry up!
2562				 */
2563				if (!zone_watermark_ok_safe(zone, order,
2564					    min_wmark_pages(zone), end_zone, 0))
2565					has_under_min_watermark_zone = 1;
2566			} else {
2567				/*
2568				 * If a zone reaches its high watermark,
2569				 * consider it to be no longer congested. It's
2570				 * possible there are dirty pages backed by
2571				 * congested BDIs but as pressure is relieved,
2572				 * spectulatively avoid congestion waits
2573				 */
2574				zone_clear_flag(zone, ZONE_CONGESTED);
2575				if (i <= *classzone_idx)
2576					balanced += zone->present_pages;
2577			}
2578
2579		}
2580		if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2581			break;		/* kswapd: all done */
2582		/*
2583		 * OK, kswapd is getting into trouble.  Take a nap, then take
2584		 * another pass across the zones.
2585		 */
2586		if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2587			if (has_under_min_watermark_zone)
2588				count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2589			else
2590				congestion_wait(BLK_RW_ASYNC, HZ/10);
2591		}
2592
2593		/*
2594		 * We do this so kswapd doesn't build up large priorities for
2595		 * example when it is freeing in parallel with allocators. It
2596		 * matches the direct reclaim path behaviour in terms of impact
2597		 * on zone->*_priority.
2598		 */
2599		if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2600			break;
2601	} while (--sc.priority >= 0);
2602out:
2603
2604	/*
2605	 * order-0: All zones must meet high watermark for a balanced node
2606	 * high-order: Balanced zones must make up at least 25% of the node
2607	 *             for the node to be balanced
2608	 */
2609	if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2610		cond_resched();
2611
2612		try_to_freeze();
2613
2614		/*
2615		 * Fragmentation may mean that the system cannot be
2616		 * rebalanced for high-order allocations in all zones.
2617		 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2618		 * it means the zones have been fully scanned and are still
2619		 * not balanced. For high-order allocations, there is
2620		 * little point trying all over again as kswapd may
2621		 * infinite loop.
2622		 *
2623		 * Instead, recheck all watermarks at order-0 as they
2624		 * are the most important. If watermarks are ok, kswapd will go
2625		 * back to sleep. High-order users can still perform direct
2626		 * reclaim if they wish.
2627		 */
2628		if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2629			order = sc.order = 0;
2630
2631		goto loop_again;
2632	}
2633
2634	/*
2635	 * If kswapd was reclaiming at a higher order, it has the option of
2636	 * sleeping without all zones being balanced. Before it does, it must
2637	 * ensure that the watermarks for order-0 on *all* zones are met and
2638	 * that the congestion flags are cleared. The congestion flag must
2639	 * be cleared as kswapd is the only mechanism that clears the flag
2640	 * and it is potentially going to sleep here.
2641	 */
2642	if (order) {
2643		int zones_need_compaction = 1;
2644
2645		for (i = 0; i <= end_zone; i++) {
2646			struct zone *zone = pgdat->node_zones + i;
2647
2648			if (!populated_zone(zone))
2649				continue;
2650
2651			if (zone->all_unreclaimable &&
2652			    sc.priority != DEF_PRIORITY)
2653				continue;
2654
2655			/* Would compaction fail due to lack of free memory? */
2656			if (COMPACTION_BUILD &&
2657			    compaction_suitable(zone, order) == COMPACT_SKIPPED)
2658				goto loop_again;
2659
2660			/* Confirm the zone is balanced for order-0 */
2661			if (!zone_watermark_ok(zone, 0,
2662					high_wmark_pages(zone), 0, 0)) {
2663				order = sc.order = 0;
2664				goto loop_again;
2665			}
2666
2667			/* Check if the memory needs to be defragmented. */
2668			if (zone_watermark_ok(zone, order,
2669				    low_wmark_pages(zone), *classzone_idx, 0))
2670				zones_need_compaction = 0;
2671
2672			/* If balanced, clear the congested flag */
2673			zone_clear_flag(zone, ZONE_CONGESTED);
2674		}
2675
2676		if (zones_need_compaction)
2677			compact_pgdat(pgdat, order);
2678	}
2679
2680	/*
2681	 * Return the order we were reclaiming at so sleeping_prematurely()
2682	 * makes a decision on the order we were last reclaiming at. However,
2683	 * if another caller entered the allocator slow path while kswapd
2684	 * was awake, order will remain at the higher level
2685	 */
2686	*classzone_idx = end_zone;
2687	return order;
2688}
2689
2690static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2691{
2692	long remaining = 0;
2693	DEFINE_WAIT(wait);
2694
2695	if (freezing(current) || kthread_should_stop())
2696		return;
2697
2698	prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2699
2700	/* Try to sleep for a short interval */
2701	if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2702		remaining = schedule_timeout(HZ/10);
2703		finish_wait(&pgdat->kswapd_wait, &wait);
2704		prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2705	}
2706
2707	/*
2708	 * After a short sleep, check if it was a premature sleep. If not, then
2709	 * go fully to sleep until explicitly woken up.
2710	 */
2711	if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2712		trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2713
2714		/*
2715		 * vmstat counters are not perfectly accurate and the estimated
2716		 * value for counters such as NR_FREE_PAGES can deviate from the
2717		 * true value by nr_online_cpus * threshold. To avoid the zone
2718		 * watermarks being breached while under pressure, we reduce the
2719		 * per-cpu vmstat threshold while kswapd is awake and restore
2720		 * them before going back to sleep.
2721		 */
2722		set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2723
2724		if (!kthread_should_stop())
2725			schedule();
2726
2727		set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2728	} else {
2729		if (remaining)
2730			count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2731		else
2732			count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2733	}
2734	finish_wait(&pgdat->kswapd_wait, &wait);
2735}
2736
2737/*
2738 * The background pageout daemon, started as a kernel thread
2739 * from the init process.
2740 *
2741 * This basically trickles out pages so that we have _some_
2742 * free memory available even if there is no other activity
2743 * that frees anything up. This is needed for things like routing
2744 * etc, where we otherwise might have all activity going on in
2745 * asynchronous contexts that cannot page things out.
2746 *
2747 * If there are applications that are active memory-allocators
2748 * (most normal use), this basically shouldn't matter.
2749 */
2750static int kswapd(void *p)
2751{
2752	unsigned long order, new_order;
2753	unsigned balanced_order;
2754	int classzone_idx, new_classzone_idx;
2755	int balanced_classzone_idx;
2756	pg_data_t *pgdat = (pg_data_t*)p;
2757	struct task_struct *tsk = current;
2758
2759	struct reclaim_state reclaim_state = {
2760		.reclaimed_slab = 0,
2761	};
2762	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2763
2764	lockdep_set_current_reclaim_state(GFP_KERNEL);
2765
2766	if (!cpumask_empty(cpumask))
2767		set_cpus_allowed_ptr(tsk, cpumask);
2768	current->reclaim_state = &reclaim_state;
2769
2770	/*
2771	 * Tell the memory management that we're a "memory allocator",
2772	 * and that if we need more memory we should get access to it
2773	 * regardless (see "__alloc_pages()"). "kswapd" should
2774	 * never get caught in the normal page freeing logic.
2775	 *
2776	 * (Kswapd normally doesn't need memory anyway, but sometimes
2777	 * you need a small amount of memory in order to be able to
2778	 * page out something else, and this flag essentially protects
2779	 * us from recursively trying to free more memory as we're
2780	 * trying to free the first piece of memory in the first place).
2781	 */
2782	tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2783	set_freezable();
2784
2785	order = new_order = 0;
2786	balanced_order = 0;
2787	classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2788	balanced_classzone_idx = classzone_idx;
2789	for ( ; ; ) {
2790		int ret;
2791
2792		/*
2793		 * If the last balance_pgdat was unsuccessful it's unlikely a
2794		 * new request of a similar or harder type will succeed soon
2795		 * so consider going to sleep on the basis we reclaimed at
2796		 */
2797		if (balanced_classzone_idx >= new_classzone_idx &&
2798					balanced_order == new_order) {
2799			new_order = pgdat->kswapd_max_order;
2800			new_classzone_idx = pgdat->classzone_idx;
2801			pgdat->kswapd_max_order =  0;
2802			pgdat->classzone_idx = pgdat->nr_zones - 1;
2803		}
2804
2805		if (order < new_order || classzone_idx > new_classzone_idx) {
2806			/*
2807			 * Don't sleep if someone wants a larger 'order'
2808			 * allocation or has tigher zone constraints
2809			 */
2810			order = new_order;
2811			classzone_idx = new_classzone_idx;
2812		} else {
2813			kswapd_try_to_sleep(pgdat, balanced_order,
2814						balanced_classzone_idx);
2815			order = pgdat->kswapd_max_order;
2816			classzone_idx = pgdat->classzone_idx;
2817			new_order = order;
2818			new_classzone_idx = classzone_idx;
2819			pgdat->kswapd_max_order = 0;
2820			pgdat->classzone_idx = pgdat->nr_zones - 1;
2821		}
2822
2823		ret = try_to_freeze();
2824		if (kthread_should_stop())
2825			break;
2826
2827		/*
2828		 * We can speed up thawing tasks if we don't call balance_pgdat
2829		 * after returning from the refrigerator
2830		 */
2831		if (!ret) {
2832			trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2833			balanced_classzone_idx = classzone_idx;
2834			balanced_order = balance_pgdat(pgdat, order,
2835						&balanced_classzone_idx);
2836		}
2837	}
2838	return 0;
2839}
2840
2841/*
2842 * A zone is low on free memory, so wake its kswapd task to service it.
2843 */
2844void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2845{
2846	pg_data_t *pgdat;
2847
2848	if (!populated_zone(zone))
2849		return;
2850
2851	if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2852		return;
2853	pgdat = zone->zone_pgdat;
2854	if (pgdat->kswapd_max_order < order) {
2855		pgdat->kswapd_max_order = order;
2856		pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2857	}
2858	if (!waitqueue_active(&pgdat->kswapd_wait))
2859		return;
2860	if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2861		return;
2862
2863	trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2864	wake_up_interruptible(&pgdat->kswapd_wait);
2865}
2866
2867/*
2868 * The reclaimable count would be mostly accurate.
2869 * The less reclaimable pages may be
2870 * - mlocked pages, which will be moved to unevictable list when encountered
2871 * - mapped pages, which may require several travels to be reclaimed
2872 * - dirty pages, which is not "instantly" reclaimable
2873 */
2874unsigned long global_reclaimable_pages(void)
2875{
2876	int nr;
2877
2878	nr = global_page_state(NR_ACTIVE_FILE) +
2879	     global_page_state(NR_INACTIVE_FILE);
2880
2881	if (nr_swap_pages > 0)
2882		nr += global_page_state(NR_ACTIVE_ANON) +
2883		      global_page_state(NR_INACTIVE_ANON);
2884
2885	return nr;
2886}
2887
2888unsigned long zone_reclaimable_pages(struct zone *zone)
2889{
2890	int nr;
2891
2892	nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2893	     zone_page_state(zone, NR_INACTIVE_FILE);
2894
2895	if (nr_swap_pages > 0)
2896		nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2897		      zone_page_state(zone, NR_INACTIVE_ANON);
2898
2899	return nr;
2900}
2901
2902#ifdef CONFIG_HIBERNATION
2903/*
2904 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2905 * freed pages.
2906 *
2907 * Rather than trying to age LRUs the aim is to preserve the overall
2908 * LRU order by reclaiming preferentially
2909 * inactive > active > active referenced > active mapped
2910 */
2911unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2912{
2913	struct reclaim_state reclaim_state;
2914	struct scan_control sc = {
2915		.gfp_mask = GFP_HIGHUSER_MOVABLE,
2916		.may_swap = 1,
2917		.may_unmap = 1,
2918		.may_writepage = 1,
2919		.nr_to_reclaim = nr_to_reclaim,
2920		.hibernation_mode = 1,
2921		.order = 0,
2922		.priority = DEF_PRIORITY,
2923	};
2924	struct shrink_control shrink = {
2925		.gfp_mask = sc.gfp_mask,
2926	};
2927	struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2928	struct task_struct *p = current;
2929	unsigned long nr_reclaimed;
2930
2931	p->flags |= PF_MEMALLOC;
2932	lockdep_set_current_reclaim_state(sc.gfp_mask);
2933	reclaim_state.reclaimed_slab = 0;
2934	p->reclaim_state = &reclaim_state;
2935
2936	nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2937
2938	p->reclaim_state = NULL;
2939	lockdep_clear_current_reclaim_state();
2940	p->flags &= ~PF_MEMALLOC;
2941
2942	return nr_reclaimed;
2943}
2944#endif /* CONFIG_HIBERNATION */
2945
2946/* It's optimal to keep kswapds on the same CPUs as their memory, but
2947   not required for correctness.  So if the last cpu in a node goes
2948   away, we get changed to run anywhere: as the first one comes back,
2949   restore their cpu bindings. */
2950static int __devinit cpu_callback(struct notifier_block *nfb,
2951				  unsigned long action, void *hcpu)
2952{
2953	int nid;
2954
2955	if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2956		for_each_node_state(nid, N_HIGH_MEMORY) {
2957			pg_data_t *pgdat = NODE_DATA(nid);
2958			const struct cpumask *mask;
2959
2960			mask = cpumask_of_node(pgdat->node_id);
2961
2962			if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2963				/* One of our CPUs online: restore mask */
2964				set_cpus_allowed_ptr(pgdat->kswapd, mask);
2965		}
2966	}
2967	return NOTIFY_OK;
2968}
2969
2970/*
2971 * This kswapd start function will be called by init and node-hot-add.
2972 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2973 */
2974int kswapd_run(int nid)
2975{
2976	pg_data_t *pgdat = NODE_DATA(nid);
2977	int ret = 0;
2978
2979	if (pgdat->kswapd)
2980		return 0;
2981
2982	pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2983	if (IS_ERR(pgdat->kswapd)) {
2984		/* failure at boot is fatal */
2985		BUG_ON(system_state == SYSTEM_BOOTING);
2986		printk("Failed to start kswapd on node %d\n",nid);
2987		ret = -1;
2988	}
2989	return ret;
2990}
2991
2992/*
2993 * Called by memory hotplug when all memory in a node is offlined.  Caller must
2994 * hold lock_memory_hotplug().
2995 */
2996void kswapd_stop(int nid)
2997{
2998	struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2999
3000	if (kswapd) {
3001		kthread_stop(kswapd);
3002		NODE_DATA(nid)->kswapd = NULL;
3003	}
3004}
3005
3006static int __init kswapd_init(void)
3007{
3008	int nid;
3009
3010	swap_setup();
3011	for_each_node_state(nid, N_HIGH_MEMORY)
3012 		kswapd_run(nid);
3013	hotcpu_notifier(cpu_callback, 0);
3014	return 0;
3015}
3016
3017module_init(kswapd_init)
3018
3019#ifdef CONFIG_NUMA
3020/*
3021 * Zone reclaim mode
3022 *
3023 * If non-zero call zone_reclaim when the number of free pages falls below
3024 * the watermarks.
3025 */
3026int zone_reclaim_mode __read_mostly;
3027
3028#define RECLAIM_OFF 0
3029#define RECLAIM_ZONE (1<<0)	/* Run shrink_inactive_list on the zone */
3030#define RECLAIM_WRITE (1<<1)	/* Writeout pages during reclaim */
3031#define RECLAIM_SWAP (1<<2)	/* Swap pages out during reclaim */
3032
3033/*
3034 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3035 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3036 * a zone.
3037 */
3038#define ZONE_RECLAIM_PRIORITY 4
3039
3040/*
3041 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3042 * occur.
3043 */
3044int sysctl_min_unmapped_ratio = 1;
3045
3046/*
3047 * If the number of slab pages in a zone grows beyond this percentage then
3048 * slab reclaim needs to occur.
3049 */
3050int sysctl_min_slab_ratio = 5;
3051
3052static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3053{
3054	unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3055	unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3056		zone_page_state(zone, NR_ACTIVE_FILE);
3057
3058	/*
3059	 * It's possible for there to be more file mapped pages than
3060	 * accounted for by the pages on the file LRU lists because
3061	 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3062	 */
3063	return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3064}
3065
3066/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3067static long zone_pagecache_reclaimable(struct zone *zone)
3068{
3069	long nr_pagecache_reclaimable;
3070	long delta = 0;
3071
3072	/*
3073	 * If RECLAIM_SWAP is set, then all file pages are considered
3074	 * potentially reclaimable. Otherwise, we have to worry about
3075	 * pages like swapcache and zone_unmapped_file_pages() provides
3076	 * a better estimate
3077	 */
3078	if (zone_reclaim_mode & RECLAIM_SWAP)
3079		nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3080	else
3081		nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3082
3083	/* If we can't clean pages, remove dirty pages from consideration */
3084	if (!(zone_reclaim_mode & RECLAIM_WRITE))
3085		delta += zone_page_state(zone, NR_FILE_DIRTY);
3086
3087	/* Watch for any possible underflows due to delta */
3088	if (unlikely(delta > nr_pagecache_reclaimable))
3089		delta = nr_pagecache_reclaimable;
3090
3091	return nr_pagecache_reclaimable - delta;
3092}
3093
3094/*
3095 * Try to free up some pages from this zone through reclaim.
3096 */
3097static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3098{
3099	/* Minimum pages needed in order to stay on node */
3100	const unsigned long nr_pages = 1 << order;
3101	struct task_struct *p = current;
3102	struct reclaim_state reclaim_state;
3103	struct scan_control sc = {
3104		.may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3105		.may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3106		.may_swap = 1,
3107		.nr_to_reclaim = max_t(unsigned long, nr_pages,
3108				       SWAP_CLUSTER_MAX),
3109		.gfp_mask = gfp_mask,
3110		.order = order,
3111		.priority = ZONE_RECLAIM_PRIORITY,
3112	};
3113	struct shrink_control shrink = {
3114		.gfp_mask = sc.gfp_mask,
3115	};
3116	unsigned long nr_slab_pages0, nr_slab_pages1;
3117
3118	cond_resched();
3119	/*
3120	 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3121	 * and we also need to be able to write out pages for RECLAIM_WRITE
3122	 * and RECLAIM_SWAP.
3123	 */
3124	p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3125	lockdep_set_current_reclaim_state(gfp_mask);
3126	reclaim_state.reclaimed_slab = 0;
3127	p->reclaim_state = &reclaim_state;
3128
3129	if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3130		/*
3131		 * Free memory by calling shrink zone with increasing
3132		 * priorities until we have enough memory freed.
3133		 */
3134		do {
3135			shrink_zone(zone, &sc);
3136		} while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3137	}
3138
3139	nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3140	if (nr_slab_pages0 > zone->min_slab_pages) {
3141		/*
3142		 * shrink_slab() does not currently allow us to determine how
3143		 * many pages were freed in this zone. So we take the current
3144		 * number of slab pages and shake the slab until it is reduced
3145		 * by the same nr_pages that we used for reclaiming unmapped
3146		 * pages.
3147		 *
3148		 * Note that shrink_slab will free memory on all zones and may
3149		 * take a long time.
3150		 */
3151		for (;;) {
3152			unsigned long lru_pages = zone_reclaimable_pages(zone);
3153
3154			/* No reclaimable slab or very low memory pressure */
3155			if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3156				break;
3157
3158			/* Freed enough memory */
3159			nr_slab_pages1 = zone_page_state(zone,
3160							NR_SLAB_RECLAIMABLE);
3161			if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3162				break;
3163		}
3164
3165		/*
3166		 * Update nr_reclaimed by the number of slab pages we
3167		 * reclaimed from this zone.
3168		 */
3169		nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3170		if (nr_slab_pages1 < nr_slab_pages0)
3171			sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3172	}
3173
3174	p->reclaim_state = NULL;
3175	current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3176	lockdep_clear_current_reclaim_state();
3177	return sc.nr_reclaimed >= nr_pages;
3178}
3179
3180int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3181{
3182	int node_id;
3183	int ret;
3184
3185	/*
3186	 * Zone reclaim reclaims unmapped file backed pages and
3187	 * slab pages if we are over the defined limits.
3188	 *
3189	 * A small portion of unmapped file backed pages is needed for
3190	 * file I/O otherwise pages read by file I/O will be immediately
3191	 * thrown out if the zone is overallocated. So we do not reclaim
3192	 * if less than a specified percentage of the zone is used by
3193	 * unmapped file backed pages.
3194	 */
3195	if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3196	    zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3197		return ZONE_RECLAIM_FULL;
3198
3199	if (zone->all_unreclaimable)
3200		return ZONE_RECLAIM_FULL;
3201
3202	/*
3203	 * Do not scan if the allocation should not be delayed.
3204	 */
3205	if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3206		return ZONE_RECLAIM_NOSCAN;
3207
3208	/*
3209	 * Only run zone reclaim on the local zone or on zones that do not
3210	 * have associated processors. This will favor the local processor
3211	 * over remote processors and spread off node memory allocations
3212	 * as wide as possible.
3213	 */
3214	node_id = zone_to_nid(zone);
3215	if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3216		return ZONE_RECLAIM_NOSCAN;
3217
3218	if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3219		return ZONE_RECLAIM_NOSCAN;
3220
3221	ret = __zone_reclaim(zone, gfp_mask, order);
3222	zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3223
3224	if (!ret)
3225		count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3226
3227	return ret;
3228}
3229#endif
3230
3231/*
3232 * page_evictable - test whether a page is evictable
3233 * @page: the page to test
3234 * @vma: the VMA in which the page is or will be mapped, may be NULL
3235 *
3236 * Test whether page is evictable--i.e., should be placed on active/inactive
3237 * lists vs unevictable list.  The vma argument is !NULL when called from the
3238 * fault path to determine how to instantate a new page.
3239 *
3240 * Reasons page might not be evictable:
3241 * (1) page's mapping marked unevictable
3242 * (2) page is part of an mlocked VMA
3243 *
3244 */
3245int page_evictable(struct page *page, struct vm_area_struct *vma)
3246{
3247
3248	if (mapping_unevictable(page_mapping(page)))
3249		return 0;
3250
3251	if (PageMlocked(page) || (vma && mlocked_vma_newpage(vma, page)))
3252		return 0;
3253
3254	return 1;
3255}
3256
3257#ifdef CONFIG_SHMEM
3258/**
3259 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3260 * @pages:	array of pages to check
3261 * @nr_pages:	number of pages to check
3262 *
3263 * Checks pages for evictability and moves them to the appropriate lru list.
3264 *
3265 * This function is only used for SysV IPC SHM_UNLOCK.
3266 */
3267void check_move_unevictable_pages(struct page **pages, int nr_pages)
3268{
3269	struct lruvec *lruvec;
3270	struct zone *zone = NULL;
3271	int pgscanned = 0;
3272	int pgrescued = 0;
3273	int i;
3274
3275	for (i = 0; i < nr_pages; i++) {
3276		struct page *page = pages[i];
3277		struct zone *pagezone;
3278
3279		pgscanned++;
3280		pagezone = page_zone(page);
3281		if (pagezone != zone) {
3282			if (zone)
3283				spin_unlock_irq(&zone->lru_lock);
3284			zone = pagezone;
3285			spin_lock_irq(&zone->lru_lock);
3286		}
3287		lruvec = mem_cgroup_page_lruvec(page, zone);
3288
3289		if (!PageLRU(page) || !PageUnevictable(page))
3290			continue;
3291
3292		if (page_evictable(page, NULL)) {
3293			enum lru_list lru = page_lru_base_type(page);
3294
3295			VM_BUG_ON(PageActive(page));
3296			ClearPageUnevictable(page);
3297			del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3298			add_page_to_lru_list(page, lruvec, lru);
3299			pgrescued++;
3300		}
3301	}
3302
3303	if (zone) {
3304		__count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3305		__count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3306		spin_unlock_irq(&zone->lru_lock);
3307	}
3308}
3309#endif /* CONFIG_SHMEM */
3310
3311static void warn_scan_unevictable_pages(void)
3312{
3313	printk_once(KERN_WARNING
3314		    "%s: The scan_unevictable_pages sysctl/node-interface has been "
3315		    "disabled for lack of a legitimate use case.  If you have "
3316		    "one, please send an email to linux-mm@kvack.org.\n",
3317		    current->comm);
3318}
3319
3320/*
3321 * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3322 * all nodes' unevictable lists for evictable pages
3323 */
3324unsigned long scan_unevictable_pages;
3325
3326int scan_unevictable_handler(struct ctl_table *table, int write,
3327			   void __user *buffer,
3328			   size_t *length, loff_t *ppos)
3329{
3330	warn_scan_unevictable_pages();
3331	proc_doulongvec_minmax(table, write, buffer, length, ppos);
3332	scan_unevictable_pages = 0;
3333	return 0;
3334}
3335
3336#ifdef CONFIG_NUMA
3337/*
3338 * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3339 * a specified node's per zone unevictable lists for evictable pages.
3340 */
3341
3342static ssize_t read_scan_unevictable_node(struct device *dev,
3343					  struct device_attribute *attr,
3344					  char *buf)
3345{
3346	warn_scan_unevictable_pages();
3347	return sprintf(buf, "0\n");	/* always zero; should fit... */
3348}
3349
3350static ssize_t write_scan_unevictable_node(struct device *dev,
3351					   struct device_attribute *attr,
3352					const char *buf, size_t count)
3353{
3354	warn_scan_unevictable_pages();
3355	return 1;
3356}
3357
3358
3359static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3360			read_scan_unevictable_node,
3361			write_scan_unevictable_node);
3362
3363int scan_unevictable_register_node(struct node *node)
3364{
3365	return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3366}
3367
3368void scan_unevictable_unregister_node(struct node *node)
3369{
3370	device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3371}
3372#endif