Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  linux/mm/vmscan.c
   4 *
   5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6 *
   7 *  Swap reorganised 29.12.95, Stephen Tweedie.
   8 *  kswapd added: 7.1.96  sct
   9 *  Removed kswapd_ctl limits, and swap out as many pages as needed
  10 *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
  11 *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  12 *  Multiqueue VM started 5.8.00, Rik van Riel.
  13 */
  14
  15#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  16
  17#include <linux/mm.h>
  18#include <linux/sched/mm.h>
  19#include <linux/module.h>
  20#include <linux/gfp.h>
  21#include <linux/kernel_stat.h>
  22#include <linux/swap.h>
  23#include <linux/pagemap.h>
  24#include <linux/init.h>
  25#include <linux/highmem.h>
  26#include <linux/vmpressure.h>
  27#include <linux/vmstat.h>
  28#include <linux/file.h>
  29#include <linux/writeback.h>
  30#include <linux/blkdev.h>
  31#include <linux/buffer_head.h>	/* for try_to_release_page(),
  32					buffer_heads_over_limit */
  33#include <linux/mm_inline.h>
  34#include <linux/backing-dev.h>
  35#include <linux/rmap.h>
  36#include <linux/topology.h>
  37#include <linux/cpu.h>
  38#include <linux/cpuset.h>
  39#include <linux/compaction.h>
  40#include <linux/notifier.h>
  41#include <linux/rwsem.h>
  42#include <linux/delay.h>
  43#include <linux/kthread.h>
  44#include <linux/freezer.h>
  45#include <linux/memcontrol.h>
  46#include <linux/delayacct.h>
  47#include <linux/sysctl.h>
  48#include <linux/oom.h>
  49#include <linux/pagevec.h>
  50#include <linux/prefetch.h>
  51#include <linux/printk.h>
  52#include <linux/dax.h>
  53#include <linux/psi.h>
  54
  55#include <asm/tlbflush.h>
  56#include <asm/div64.h>
  57
  58#include <linux/swapops.h>
  59#include <linux/balloon_compaction.h>
  60
  61#include "internal.h"
  62
  63#define CREATE_TRACE_POINTS
  64#include <trace/events/vmscan.h>
  65
  66struct scan_control {
  67	/* How many pages shrink_list() should reclaim */
  68	unsigned long nr_to_reclaim;
  69
 
 
 
 
 
 
  70	/*
  71	 * Nodemask of nodes allowed by the caller. If NULL, all nodes
  72	 * are scanned.
  73	 */
  74	nodemask_t	*nodemask;
  75
  76	/*
  77	 * The memory cgroup that hit its limit and as a result is the
  78	 * primary target of this reclaim invocation.
  79	 */
  80	struct mem_cgroup *target_mem_cgroup;
  81
 
 
 
 
 
 
  82	/* Writepage batching in laptop mode; RECLAIM_WRITE */
  83	unsigned int may_writepage:1;
  84
  85	/* Can mapped pages be reclaimed? */
  86	unsigned int may_unmap:1;
  87
  88	/* Can pages be swapped as part of reclaim? */
  89	unsigned int may_swap:1;
  90
  91	/*
  92	 * Cgroups are not reclaimed below their configured memory.low,
  93	 * unless we threaten to OOM. If any cgroups are skipped due to
  94	 * memory.low and nothing was reclaimed, go back for memory.low.
  95	 */
  96	unsigned int memcg_low_reclaim:1;
  97	unsigned int memcg_low_skipped:1;
  98
  99	unsigned int hibernation_mode:1;
 100
 101	/* One of the zones is ready for compaction */
 102	unsigned int compaction_ready:1;
 103
 104	/* Allocation order */
 105	s8 order;
 106
 107	/* Scan (total_size >> priority) pages at once */
 108	s8 priority;
 109
 110	/* The highest zone to isolate pages for reclaim from */
 111	s8 reclaim_idx;
 112
 113	/* This context's GFP mask */
 114	gfp_t gfp_mask;
 115
 116	/* Incremented by the number of inactive pages that were scanned */
 117	unsigned long nr_scanned;
 118
 119	/* Number of pages freed so far during a call to shrink_zones() */
 120	unsigned long nr_reclaimed;
 121
 122	struct {
 123		unsigned int dirty;
 124		unsigned int unqueued_dirty;
 125		unsigned int congested;
 126		unsigned int writeback;
 127		unsigned int immediate;
 128		unsigned int file_taken;
 129		unsigned int taken;
 130	} nr;
 131
 132	/* for recording the reclaimed slab by now */
 133	struct reclaim_state reclaim_state;
 134};
 135
 136#ifdef ARCH_HAS_PREFETCH
 137#define prefetch_prev_lru_page(_page, _base, _field)			\
 138	do {								\
 139		if ((_page)->lru.prev != _base) {			\
 140			struct page *prev;				\
 141									\
 142			prev = lru_to_page(&(_page->lru));		\
 143			prefetch(&prev->_field);			\
 144		}							\
 145	} while (0)
 146#else
 147#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 148#endif
 149
 150#ifdef ARCH_HAS_PREFETCHW
 151#define prefetchw_prev_lru_page(_page, _base, _field)			\
 152	do {								\
 153		if ((_page)->lru.prev != _base) {			\
 154			struct page *prev;				\
 155									\
 156			prev = lru_to_page(&(_page->lru));		\
 157			prefetchw(&prev->_field);			\
 158		}							\
 159	} while (0)
 160#else
 161#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 162#endif
 163
 164/*
 165 * From 0 .. 100.  Higher means more swappy.
 166 */
 167int vm_swappiness = 60;
 168/*
 169 * The total number of pages which are beyond the high watermark within all
 170 * zones.
 171 */
 172unsigned long vm_total_pages;
 173
 174static void set_task_reclaim_state(struct task_struct *task,
 175				   struct reclaim_state *rs)
 176{
 177	/* Check for an overwrite */
 178	WARN_ON_ONCE(rs && task->reclaim_state);
 179
 180	/* Check for the nulling of an already-nulled member */
 181	WARN_ON_ONCE(!rs && !task->reclaim_state);
 182
 183	task->reclaim_state = rs;
 184}
 185
 186static LIST_HEAD(shrinker_list);
 187static DECLARE_RWSEM(shrinker_rwsem);
 188
 189#ifdef CONFIG_MEMCG
 190/*
 191 * We allow subsystems to populate their shrinker-related
 192 * LRU lists before register_shrinker_prepared() is called
 193 * for the shrinker, since we don't want to impose
 194 * restrictions on their internal registration order.
 195 * In this case shrink_slab_memcg() may find corresponding
 196 * bit is set in the shrinkers map.
 197 *
 198 * This value is used by the function to detect registering
 199 * shrinkers and to skip do_shrink_slab() calls for them.
 200 */
 201#define SHRINKER_REGISTERING ((struct shrinker *)~0UL)
 202
 203static DEFINE_IDR(shrinker_idr);
 204static int shrinker_nr_max;
 205
 206static int prealloc_memcg_shrinker(struct shrinker *shrinker)
 207{
 208	int id, ret = -ENOMEM;
 209
 210	down_write(&shrinker_rwsem);
 211	/* This may call shrinker, so it must use down_read_trylock() */
 212	id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL);
 213	if (id < 0)
 214		goto unlock;
 215
 216	if (id >= shrinker_nr_max) {
 217		if (memcg_expand_shrinker_maps(id)) {
 218			idr_remove(&shrinker_idr, id);
 219			goto unlock;
 220		}
 221
 222		shrinker_nr_max = id + 1;
 223	}
 224	shrinker->id = id;
 225	ret = 0;
 226unlock:
 227	up_write(&shrinker_rwsem);
 228	return ret;
 229}
 230
 231static void unregister_memcg_shrinker(struct shrinker *shrinker)
 232{
 233	int id = shrinker->id;
 234
 235	BUG_ON(id < 0);
 236
 237	down_write(&shrinker_rwsem);
 238	idr_remove(&shrinker_idr, id);
 239	up_write(&shrinker_rwsem);
 240}
 241
 242static bool global_reclaim(struct scan_control *sc)
 243{
 244	return !sc->target_mem_cgroup;
 245}
 246
 247/**
 248 * sane_reclaim - is the usual dirty throttling mechanism operational?
 249 * @sc: scan_control in question
 250 *
 251 * The normal page dirty throttling mechanism in balance_dirty_pages() is
 252 * completely broken with the legacy memcg and direct stalling in
 253 * shrink_page_list() is used for throttling instead, which lacks all the
 254 * niceties such as fairness, adaptive pausing, bandwidth proportional
 255 * allocation and configurability.
 256 *
 257 * This function tests whether the vmscan currently in progress can assume
 258 * that the normal dirty throttling mechanism is operational.
 259 */
 260static bool sane_reclaim(struct scan_control *sc)
 261{
 262	struct mem_cgroup *memcg = sc->target_mem_cgroup;
 263
 264	if (!memcg)
 265		return true;
 266#ifdef CONFIG_CGROUP_WRITEBACK
 267	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
 268		return true;
 269#endif
 270	return false;
 271}
 272
 273static void set_memcg_congestion(pg_data_t *pgdat,
 274				struct mem_cgroup *memcg,
 275				bool congested)
 276{
 277	struct mem_cgroup_per_node *mn;
 278
 279	if (!memcg)
 280		return;
 281
 282	mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
 283	WRITE_ONCE(mn->congested, congested);
 284}
 285
 286static bool memcg_congested(pg_data_t *pgdat,
 287			struct mem_cgroup *memcg)
 288{
 289	struct mem_cgroup_per_node *mn;
 290
 291	mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
 292	return READ_ONCE(mn->congested);
 293
 294}
 295#else
 296static int prealloc_memcg_shrinker(struct shrinker *shrinker)
 297{
 298	return 0;
 299}
 300
 301static void unregister_memcg_shrinker(struct shrinker *shrinker)
 302{
 303}
 304
 305static bool global_reclaim(struct scan_control *sc)
 306{
 307	return true;
 308}
 309
 310static bool sane_reclaim(struct scan_control *sc)
 311{
 312	return true;
 313}
 314
 315static inline void set_memcg_congestion(struct pglist_data *pgdat,
 316				struct mem_cgroup *memcg, bool congested)
 317{
 318}
 319
 320static inline bool memcg_congested(struct pglist_data *pgdat,
 321			struct mem_cgroup *memcg)
 322{
 323	return false;
 324
 325}
 326#endif
 327
 328/*
 329 * This misses isolated pages which are not accounted for to save counters.
 330 * As the data only determines if reclaim or compaction continues, it is
 331 * not expected that isolated pages will be a dominating factor.
 332 */
 333unsigned long zone_reclaimable_pages(struct zone *zone)
 334{
 335	unsigned long nr;
 336
 337	nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
 338		zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
 339	if (get_nr_swap_pages() > 0)
 340		nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
 341			zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
 342
 343	return nr;
 344}
 345
 346/**
 347 * lruvec_lru_size -  Returns the number of pages on the given LRU list.
 348 * @lruvec: lru vector
 349 * @lru: lru to use
 350 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
 351 */
 352unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
 353{
 354	unsigned long lru_size = 0;
 355	int zid;
 356
 357	if (!mem_cgroup_disabled()) {
 358		for (zid = 0; zid < MAX_NR_ZONES; zid++)
 359			lru_size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
 360	} else
 361		lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
 362
 363	for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
 364		struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
 365		unsigned long size;
 366
 367		if (!managed_zone(zone))
 368			continue;
 369
 370		if (!mem_cgroup_disabled())
 371			size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
 372		else
 373			size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
 374				       NR_ZONE_LRU_BASE + lru);
 375		lru_size -= min(size, lru_size);
 376	}
 377
 378	return lru_size;
 379
 380}
 381
 382/*
 383 * Add a shrinker callback to be called from the vm.
 384 */
 385int prealloc_shrinker(struct shrinker *shrinker)
 386{
 387	unsigned int size = sizeof(*shrinker->nr_deferred);
 388
 389	if (shrinker->flags & SHRINKER_NUMA_AWARE)
 390		size *= nr_node_ids;
 391
 392	shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
 393	if (!shrinker->nr_deferred)
 394		return -ENOMEM;
 395
 396	if (shrinker->flags & SHRINKER_MEMCG_AWARE) {
 397		if (prealloc_memcg_shrinker(shrinker))
 398			goto free_deferred;
 399	}
 400
 401	return 0;
 402
 403free_deferred:
 404	kfree(shrinker->nr_deferred);
 405	shrinker->nr_deferred = NULL;
 406	return -ENOMEM;
 407}
 408
 409void free_prealloced_shrinker(struct shrinker *shrinker)
 410{
 411	if (!shrinker->nr_deferred)
 412		return;
 413
 414	if (shrinker->flags & SHRINKER_MEMCG_AWARE)
 415		unregister_memcg_shrinker(shrinker);
 416
 417	kfree(shrinker->nr_deferred);
 418	shrinker->nr_deferred = NULL;
 419}
 420
 421void register_shrinker_prepared(struct shrinker *shrinker)
 422{
 423	down_write(&shrinker_rwsem);
 424	list_add_tail(&shrinker->list, &shrinker_list);
 425#ifdef CONFIG_MEMCG_KMEM
 426	if (shrinker->flags & SHRINKER_MEMCG_AWARE)
 427		idr_replace(&shrinker_idr, shrinker, shrinker->id);
 428#endif
 429	up_write(&shrinker_rwsem);
 430}
 431
 432int register_shrinker(struct shrinker *shrinker)
 433{
 434	int err = prealloc_shrinker(shrinker);
 435
 436	if (err)
 437		return err;
 438	register_shrinker_prepared(shrinker);
 439	return 0;
 440}
 441EXPORT_SYMBOL(register_shrinker);
 442
 443/*
 444 * Remove one
 445 */
 446void unregister_shrinker(struct shrinker *shrinker)
 447{
 448	if (!shrinker->nr_deferred)
 449		return;
 450	if (shrinker->flags & SHRINKER_MEMCG_AWARE)
 451		unregister_memcg_shrinker(shrinker);
 452	down_write(&shrinker_rwsem);
 453	list_del(&shrinker->list);
 454	up_write(&shrinker_rwsem);
 455	kfree(shrinker->nr_deferred);
 456	shrinker->nr_deferred = NULL;
 457}
 458EXPORT_SYMBOL(unregister_shrinker);
 459
 460#define SHRINK_BATCH 128
 461
 462static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
 463				    struct shrinker *shrinker, int priority)
 464{
 465	unsigned long freed = 0;
 466	unsigned long long delta;
 467	long total_scan;
 468	long freeable;
 469	long nr;
 470	long new_nr;
 471	int nid = shrinkctl->nid;
 472	long batch_size = shrinker->batch ? shrinker->batch
 473					  : SHRINK_BATCH;
 474	long scanned = 0, next_deferred;
 475
 476	if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
 477		nid = 0;
 478
 479	freeable = shrinker->count_objects(shrinker, shrinkctl);
 480	if (freeable == 0 || freeable == SHRINK_EMPTY)
 481		return freeable;
 482
 483	/*
 484	 * copy the current shrinker scan count into a local variable
 485	 * and zero it so that other concurrent shrinker invocations
 486	 * don't also do this scanning work.
 487	 */
 488	nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
 489
 490	total_scan = nr;
 491	if (shrinker->seeks) {
 492		delta = freeable >> priority;
 493		delta *= 4;
 494		do_div(delta, shrinker->seeks);
 495	} else {
 496		/*
 497		 * These objects don't require any IO to create. Trim
 498		 * them aggressively under memory pressure to keep
 499		 * them from causing refetches in the IO caches.
 500		 */
 501		delta = freeable / 2;
 502	}
 503
 504	total_scan += delta;
 505	if (total_scan < 0) {
 506		pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
 507		       shrinker->scan_objects, total_scan);
 508		total_scan = freeable;
 509		next_deferred = nr;
 510	} else
 511		next_deferred = total_scan;
 512
 513	/*
 514	 * We need to avoid excessive windup on filesystem shrinkers
 515	 * due to large numbers of GFP_NOFS allocations causing the
 516	 * shrinkers to return -1 all the time. This results in a large
 517	 * nr being built up so when a shrink that can do some work
 518	 * comes along it empties the entire cache due to nr >>>
 519	 * freeable. This is bad for sustaining a working set in
 520	 * memory.
 521	 *
 522	 * Hence only allow the shrinker to scan the entire cache when
 523	 * a large delta change is calculated directly.
 524	 */
 525	if (delta < freeable / 4)
 526		total_scan = min(total_scan, freeable / 2);
 527
 528	/*
 529	 * Avoid risking looping forever due to too large nr value:
 530	 * never try to free more than twice the estimate number of
 531	 * freeable entries.
 532	 */
 533	if (total_scan > freeable * 2)
 534		total_scan = freeable * 2;
 535
 536	trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
 537				   freeable, delta, total_scan, priority);
 538
 539	/*
 540	 * Normally, we should not scan less than batch_size objects in one
 541	 * pass to avoid too frequent shrinker calls, but if the slab has less
 542	 * than batch_size objects in total and we are really tight on memory,
 543	 * we will try to reclaim all available objects, otherwise we can end
 544	 * up failing allocations although there are plenty of reclaimable
 545	 * objects spread over several slabs with usage less than the
 546	 * batch_size.
 547	 *
 548	 * We detect the "tight on memory" situations by looking at the total
 549	 * number of objects we want to scan (total_scan). If it is greater
 550	 * than the total number of objects on slab (freeable), we must be
 551	 * scanning at high prio and therefore should try to reclaim as much as
 552	 * possible.
 553	 */
 554	while (total_scan >= batch_size ||
 555	       total_scan >= freeable) {
 556		unsigned long ret;
 557		unsigned long nr_to_scan = min(batch_size, total_scan);
 558
 559		shrinkctl->nr_to_scan = nr_to_scan;
 560		shrinkctl->nr_scanned = nr_to_scan;
 561		ret = shrinker->scan_objects(shrinker, shrinkctl);
 562		if (ret == SHRINK_STOP)
 563			break;
 564		freed += ret;
 565
 566		count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
 567		total_scan -= shrinkctl->nr_scanned;
 568		scanned += shrinkctl->nr_scanned;
 569
 570		cond_resched();
 571	}
 572
 573	if (next_deferred >= scanned)
 574		next_deferred -= scanned;
 575	else
 576		next_deferred = 0;
 577	/*
 578	 * move the unused scan count back into the shrinker in a
 579	 * manner that handles concurrent updates. If we exhausted the
 580	 * scan, there is no need to do an update.
 581	 */
 582	if (next_deferred > 0)
 583		new_nr = atomic_long_add_return(next_deferred,
 584						&shrinker->nr_deferred[nid]);
 585	else
 586		new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
 587
 588	trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
 589	return freed;
 590}
 591
 592#ifdef CONFIG_MEMCG
 593static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
 594			struct mem_cgroup *memcg, int priority)
 595{
 596	struct memcg_shrinker_map *map;
 597	unsigned long ret, freed = 0;
 598	int i;
 599
 600	if (!mem_cgroup_online(memcg))
 601		return 0;
 602
 603	if (!down_read_trylock(&shrinker_rwsem))
 604		return 0;
 605
 606	map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map,
 607					true);
 608	if (unlikely(!map))
 609		goto unlock;
 610
 611	for_each_set_bit(i, map->map, shrinker_nr_max) {
 612		struct shrink_control sc = {
 613			.gfp_mask = gfp_mask,
 614			.nid = nid,
 615			.memcg = memcg,
 616		};
 617		struct shrinker *shrinker;
 618
 619		shrinker = idr_find(&shrinker_idr, i);
 620		if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) {
 621			if (!shrinker)
 622				clear_bit(i, map->map);
 623			continue;
 624		}
 625
 626		/* Call non-slab shrinkers even though kmem is disabled */
 627		if (!memcg_kmem_enabled() &&
 628		    !(shrinker->flags & SHRINKER_NONSLAB))
 629			continue;
 630
 631		ret = do_shrink_slab(&sc, shrinker, priority);
 632		if (ret == SHRINK_EMPTY) {
 633			clear_bit(i, map->map);
 634			/*
 635			 * After the shrinker reported that it had no objects to
 636			 * free, but before we cleared the corresponding bit in
 637			 * the memcg shrinker map, a new object might have been
 638			 * added. To make sure, we have the bit set in this
 639			 * case, we invoke the shrinker one more time and reset
 640			 * the bit if it reports that it is not empty anymore.
 641			 * The memory barrier here pairs with the barrier in
 642			 * memcg_set_shrinker_bit():
 643			 *
 644			 * list_lru_add()     shrink_slab_memcg()
 645			 *   list_add_tail()    clear_bit()
 646			 *   <MB>               <MB>
 647			 *   set_bit()          do_shrink_slab()
 648			 */
 649			smp_mb__after_atomic();
 650			ret = do_shrink_slab(&sc, shrinker, priority);
 651			if (ret == SHRINK_EMPTY)
 652				ret = 0;
 653			else
 654				memcg_set_shrinker_bit(memcg, nid, i);
 655		}
 656		freed += ret;
 657
 658		if (rwsem_is_contended(&shrinker_rwsem)) {
 659			freed = freed ? : 1;
 660			break;
 661		}
 662	}
 663unlock:
 664	up_read(&shrinker_rwsem);
 665	return freed;
 666}
 667#else /* CONFIG_MEMCG */
 668static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid,
 669			struct mem_cgroup *memcg, int priority)
 670{
 671	return 0;
 672}
 673#endif /* CONFIG_MEMCG */
 674
 675/**
 676 * shrink_slab - shrink slab caches
 677 * @gfp_mask: allocation context
 678 * @nid: node whose slab caches to target
 679 * @memcg: memory cgroup whose slab caches to target
 680 * @priority: the reclaim priority
 681 *
 682 * Call the shrink functions to age shrinkable caches.
 683 *
 684 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
 685 * unaware shrinkers will receive a node id of 0 instead.
 686 *
 687 * @memcg specifies the memory cgroup to target. Unaware shrinkers
 688 * are called only if it is the root cgroup.
 
 
 689 *
 690 * @priority is sc->priority, we take the number of objects and >> by priority
 691 * in order to get the scan target.
 692 *
 693 * Returns the number of reclaimed slab objects.
 694 */
 695static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
 696				 struct mem_cgroup *memcg,
 697				 int priority)
 698{
 699	unsigned long ret, freed = 0;
 700	struct shrinker *shrinker;
 
 701
 702	/*
 703	 * The root memcg might be allocated even though memcg is disabled
 704	 * via "cgroup_disable=memory" boot parameter.  This could make
 705	 * mem_cgroup_is_root() return false, then just run memcg slab
 706	 * shrink, but skip global shrink.  This may result in premature
 707	 * oom.
 708	 */
 709	if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg))
 710		return shrink_slab_memcg(gfp_mask, nid, memcg, priority);
 711
 712	if (!down_read_trylock(&shrinker_rwsem))
 713		goto out;
 714
 715	list_for_each_entry(shrinker, &shrinker_list, list) {
 716		struct shrink_control sc = {
 717			.gfp_mask = gfp_mask,
 718			.nid = nid,
 719			.memcg = memcg,
 720		};
 721
 722		ret = do_shrink_slab(&sc, shrinker, priority);
 723		if (ret == SHRINK_EMPTY)
 724			ret = 0;
 725		freed += ret;
 
 
 
 
 
 
 
 
 
 726		/*
 727		 * Bail out if someone want to register a new shrinker to
 728		 * prevent the regsitration from being stalled for long periods
 729		 * by parallel ongoing shrinking.
 730		 */
 731		if (rwsem_is_contended(&shrinker_rwsem)) {
 732			freed = freed ? : 1;
 733			break;
 734		}
 735	}
 736
 737	up_read(&shrinker_rwsem);
 738out:
 739	cond_resched();
 740	return freed;
 741}
 742
 743void drop_slab_node(int nid)
 744{
 745	unsigned long freed;
 746
 747	do {
 748		struct mem_cgroup *memcg = NULL;
 749
 750		freed = 0;
 751		memcg = mem_cgroup_iter(NULL, NULL, NULL);
 752		do {
 753			freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
 754		} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
 755	} while (freed > 10);
 756}
 757
 758void drop_slab(void)
 759{
 760	int nid;
 761
 762	for_each_online_node(nid)
 763		drop_slab_node(nid);
 764}
 765
 766static inline int is_page_cache_freeable(struct page *page)
 767{
 768	/*
 769	 * A freeable page cache page is referenced only by the caller
 770	 * that isolated the page, the page cache and optional buffer
 771	 * heads at page->private.
 772	 */
 773	int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ?
 774		HPAGE_PMD_NR : 1;
 775	return page_count(page) - page_has_private(page) == 1 + page_cache_pins;
 776}
 777
 778static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
 779{
 780	if (current->flags & PF_SWAPWRITE)
 781		return 1;
 782	if (!inode_write_congested(inode))
 783		return 1;
 784	if (inode_to_bdi(inode) == current->backing_dev_info)
 785		return 1;
 786	return 0;
 787}
 788
 789/*
 790 * We detected a synchronous write error writing a page out.  Probably
 791 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 792 * fsync(), msync() or close().
 793 *
 794 * The tricky part is that after writepage we cannot touch the mapping: nothing
 795 * prevents it from being freed up.  But we have a ref on the page and once
 796 * that page is locked, the mapping is pinned.
 797 *
 798 * We're allowed to run sleeping lock_page() here because we know the caller has
 799 * __GFP_FS.
 800 */
 801static void handle_write_error(struct address_space *mapping,
 802				struct page *page, int error)
 803{
 804	lock_page(page);
 805	if (page_mapping(page) == mapping)
 806		mapping_set_error(mapping, error);
 807	unlock_page(page);
 808}
 809
 810/* possible outcome of pageout() */
 811typedef enum {
 812	/* failed to write page out, page is locked */
 813	PAGE_KEEP,
 814	/* move page to the active list, page is locked */
 815	PAGE_ACTIVATE,
 816	/* page has been sent to the disk successfully, page is unlocked */
 817	PAGE_SUCCESS,
 818	/* page is clean and locked */
 819	PAGE_CLEAN,
 820} pageout_t;
 821
 822/*
 823 * pageout is called by shrink_page_list() for each dirty page.
 824 * Calls ->writepage().
 825 */
 826static pageout_t pageout(struct page *page, struct address_space *mapping,
 827			 struct scan_control *sc)
 828{
 829	/*
 830	 * If the page is dirty, only perform writeback if that write
 831	 * will be non-blocking.  To prevent this allocation from being
 832	 * stalled by pagecache activity.  But note that there may be
 833	 * stalls if we need to run get_block().  We could test
 834	 * PagePrivate for that.
 835	 *
 836	 * If this process is currently in __generic_file_write_iter() against
 837	 * this page's queue, we can perform writeback even if that
 838	 * will block.
 839	 *
 840	 * If the page is swapcache, write it back even if that would
 841	 * block, for some throttling. This happens by accident, because
 842	 * swap_backing_dev_info is bust: it doesn't reflect the
 843	 * congestion state of the swapdevs.  Easy to fix, if needed.
 844	 */
 845	if (!is_page_cache_freeable(page))
 846		return PAGE_KEEP;
 847	if (!mapping) {
 848		/*
 849		 * Some data journaling orphaned pages can have
 850		 * page->mapping == NULL while being dirty with clean buffers.
 851		 */
 852		if (page_has_private(page)) {
 853			if (try_to_free_buffers(page)) {
 854				ClearPageDirty(page);
 855				pr_info("%s: orphaned page\n", __func__);
 856				return PAGE_CLEAN;
 857			}
 858		}
 859		return PAGE_KEEP;
 860	}
 861	if (mapping->a_ops->writepage == NULL)
 862		return PAGE_ACTIVATE;
 863	if (!may_write_to_inode(mapping->host, sc))
 864		return PAGE_KEEP;
 865
 866	if (clear_page_dirty_for_io(page)) {
 867		int res;
 868		struct writeback_control wbc = {
 869			.sync_mode = WB_SYNC_NONE,
 870			.nr_to_write = SWAP_CLUSTER_MAX,
 871			.range_start = 0,
 872			.range_end = LLONG_MAX,
 873			.for_reclaim = 1,
 874		};
 875
 876		SetPageReclaim(page);
 877		res = mapping->a_ops->writepage(page, &wbc);
 878		if (res < 0)
 879			handle_write_error(mapping, page, res);
 880		if (res == AOP_WRITEPAGE_ACTIVATE) {
 881			ClearPageReclaim(page);
 882			return PAGE_ACTIVATE;
 883		}
 884
 885		if (!PageWriteback(page)) {
 886			/* synchronous write or broken a_ops? */
 887			ClearPageReclaim(page);
 888		}
 889		trace_mm_vmscan_writepage(page);
 890		inc_node_page_state(page, NR_VMSCAN_WRITE);
 891		return PAGE_SUCCESS;
 892	}
 893
 894	return PAGE_CLEAN;
 895}
 896
 897/*
 898 * Same as remove_mapping, but if the page is removed from the mapping, it
 899 * gets returned with a refcount of 0.
 900 */
 901static int __remove_mapping(struct address_space *mapping, struct page *page,
 902			    bool reclaimed)
 903{
 904	unsigned long flags;
 905	int refcount;
 906
 907	BUG_ON(!PageLocked(page));
 908	BUG_ON(mapping != page_mapping(page));
 909
 910	xa_lock_irqsave(&mapping->i_pages, flags);
 911	/*
 912	 * The non racy check for a busy page.
 913	 *
 914	 * Must be careful with the order of the tests. When someone has
 915	 * a ref to the page, it may be possible that they dirty it then
 916	 * drop the reference. So if PageDirty is tested before page_count
 917	 * here, then the following race may occur:
 918	 *
 919	 * get_user_pages(&page);
 920	 * [user mapping goes away]
 921	 * write_to(page);
 922	 *				!PageDirty(page)    [good]
 923	 * SetPageDirty(page);
 924	 * put_page(page);
 925	 *				!page_count(page)   [good, discard it]
 926	 *
 927	 * [oops, our write_to data is lost]
 928	 *
 929	 * Reversing the order of the tests ensures such a situation cannot
 930	 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 931	 * load is not satisfied before that of page->_refcount.
 932	 *
 933	 * Note that if SetPageDirty is always performed via set_page_dirty,
 934	 * and thus under the i_pages lock, then this ordering is not required.
 935	 */
 936	refcount = 1 + compound_nr(page);
 
 
 
 937	if (!page_ref_freeze(page, refcount))
 938		goto cannot_free;
 939	/* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */
 940	if (unlikely(PageDirty(page))) {
 941		page_ref_unfreeze(page, refcount);
 942		goto cannot_free;
 943	}
 944
 945	if (PageSwapCache(page)) {
 946		swp_entry_t swap = { .val = page_private(page) };
 947		mem_cgroup_swapout(page, swap);
 948		__delete_from_swap_cache(page, swap);
 949		xa_unlock_irqrestore(&mapping->i_pages, flags);
 950		put_swap_page(page, swap);
 951	} else {
 952		void (*freepage)(struct page *);
 953		void *shadow = NULL;
 954
 955		freepage = mapping->a_ops->freepage;
 956		/*
 957		 * Remember a shadow entry for reclaimed file cache in
 958		 * order to detect refaults, thus thrashing, later on.
 959		 *
 960		 * But don't store shadows in an address space that is
 961		 * already exiting.  This is not just an optizimation,
 962		 * inode reclaim needs to empty out the radix tree or
 963		 * the nodes are lost.  Don't plant shadows behind its
 964		 * back.
 965		 *
 966		 * We also don't store shadows for DAX mappings because the
 967		 * only page cache pages found in these are zero pages
 968		 * covering holes, and because we don't want to mix DAX
 969		 * exceptional entries and shadow exceptional entries in the
 970		 * same address_space.
 971		 */
 972		if (reclaimed && page_is_file_cache(page) &&
 973		    !mapping_exiting(mapping) && !dax_mapping(mapping))
 974			shadow = workingset_eviction(page);
 975		__delete_from_page_cache(page, shadow);
 976		xa_unlock_irqrestore(&mapping->i_pages, flags);
 977
 978		if (freepage != NULL)
 979			freepage(page);
 980	}
 981
 982	return 1;
 983
 984cannot_free:
 985	xa_unlock_irqrestore(&mapping->i_pages, flags);
 986	return 0;
 987}
 988
 989/*
 990 * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 991 * someone else has a ref on the page, abort and return 0.  If it was
 992 * successfully detached, return 1.  Assumes the caller has a single ref on
 993 * this page.
 994 */
 995int remove_mapping(struct address_space *mapping, struct page *page)
 996{
 997	if (__remove_mapping(mapping, page, false)) {
 998		/*
 999		 * Unfreezing the refcount with 1 rather than 2 effectively
1000		 * drops the pagecache ref for us without requiring another
1001		 * atomic operation.
1002		 */
1003		page_ref_unfreeze(page, 1);
1004		return 1;
1005	}
1006	return 0;
1007}
1008
1009/**
1010 * putback_lru_page - put previously isolated page onto appropriate LRU list
1011 * @page: page to be put back to appropriate lru list
1012 *
1013 * Add previously isolated @page to appropriate LRU list.
1014 * Page may still be unevictable for other reasons.
1015 *
1016 * lru_lock must not be held, interrupts must be enabled.
1017 */
1018void putback_lru_page(struct page *page)
1019{
1020	lru_cache_add(page);
1021	put_page(page);		/* drop ref from isolate */
1022}
1023
1024enum page_references {
1025	PAGEREF_RECLAIM,
1026	PAGEREF_RECLAIM_CLEAN,
1027	PAGEREF_KEEP,
1028	PAGEREF_ACTIVATE,
1029};
1030
1031static enum page_references page_check_references(struct page *page,
1032						  struct scan_control *sc)
1033{
1034	int referenced_ptes, referenced_page;
1035	unsigned long vm_flags;
1036
1037	referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
1038					  &vm_flags);
1039	referenced_page = TestClearPageReferenced(page);
1040
1041	/*
1042	 * Mlock lost the isolation race with us.  Let try_to_unmap()
1043	 * move the page to the unevictable list.
1044	 */
1045	if (vm_flags & VM_LOCKED)
1046		return PAGEREF_RECLAIM;
1047
1048	if (referenced_ptes) {
1049		if (PageSwapBacked(page))
1050			return PAGEREF_ACTIVATE;
1051		/*
1052		 * All mapped pages start out with page table
1053		 * references from the instantiating fault, so we need
1054		 * to look twice if a mapped file page is used more
1055		 * than once.
1056		 *
1057		 * Mark it and spare it for another trip around the
1058		 * inactive list.  Another page table reference will
1059		 * lead to its activation.
1060		 *
1061		 * Note: the mark is set for activated pages as well
1062		 * so that recently deactivated but used pages are
1063		 * quickly recovered.
1064		 */
1065		SetPageReferenced(page);
1066
1067		if (referenced_page || referenced_ptes > 1)
1068			return PAGEREF_ACTIVATE;
1069
1070		/*
1071		 * Activate file-backed executable pages after first usage.
1072		 */
1073		if (vm_flags & VM_EXEC)
1074			return PAGEREF_ACTIVATE;
1075
1076		return PAGEREF_KEEP;
1077	}
1078
1079	/* Reclaim if clean, defer dirty pages to writeback */
1080	if (referenced_page && !PageSwapBacked(page))
1081		return PAGEREF_RECLAIM_CLEAN;
1082
1083	return PAGEREF_RECLAIM;
1084}
1085
1086/* Check if a page is dirty or under writeback */
1087static void page_check_dirty_writeback(struct page *page,
1088				       bool *dirty, bool *writeback)
1089{
1090	struct address_space *mapping;
1091
1092	/*
1093	 * Anonymous pages are not handled by flushers and must be written
1094	 * from reclaim context. Do not stall reclaim based on them
1095	 */
1096	if (!page_is_file_cache(page) ||
1097	    (PageAnon(page) && !PageSwapBacked(page))) {
1098		*dirty = false;
1099		*writeback = false;
1100		return;
1101	}
1102
1103	/* By default assume that the page flags are accurate */
1104	*dirty = PageDirty(page);
1105	*writeback = PageWriteback(page);
1106
1107	/* Verify dirty/writeback state if the filesystem supports it */
1108	if (!page_has_private(page))
1109		return;
1110
1111	mapping = page_mapping(page);
1112	if (mapping && mapping->a_ops->is_dirty_writeback)
1113		mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
1114}
1115
1116/*
1117 * shrink_page_list() returns the number of reclaimed pages
1118 */
1119static unsigned long shrink_page_list(struct list_head *page_list,
1120				      struct pglist_data *pgdat,
1121				      struct scan_control *sc,
1122				      enum ttu_flags ttu_flags,
1123				      struct reclaim_stat *stat,
1124				      bool ignore_references)
1125{
1126	LIST_HEAD(ret_pages);
1127	LIST_HEAD(free_pages);
 
 
 
 
1128	unsigned nr_reclaimed = 0;
1129	unsigned pgactivate = 0;
 
 
 
1130
1131	memset(stat, 0, sizeof(*stat));
1132	cond_resched();
1133
1134	while (!list_empty(page_list)) {
1135		struct address_space *mapping;
1136		struct page *page;
1137		int may_enter_fs;
1138		enum page_references references = PAGEREF_RECLAIM;
1139		bool dirty, writeback;
1140		unsigned int nr_pages;
1141
1142		cond_resched();
1143
1144		page = lru_to_page(page_list);
1145		list_del(&page->lru);
1146
1147		if (!trylock_page(page))
1148			goto keep;
1149
1150		VM_BUG_ON_PAGE(PageActive(page), page);
1151
1152		nr_pages = compound_nr(page);
1153
1154		/* Account the number of base pages even though THP */
1155		sc->nr_scanned += nr_pages;
1156
1157		if (unlikely(!page_evictable(page)))
1158			goto activate_locked;
1159
1160		if (!sc->may_unmap && page_mapped(page))
1161			goto keep_locked;
1162
 
 
 
 
 
1163		may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
1164			(PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
1165
1166		/*
1167		 * The number of dirty pages determines if a node is marked
1168		 * reclaim_congested which affects wait_iff_congested. kswapd
1169		 * will stall and start writing pages if the tail of the LRU
1170		 * is all dirty unqueued pages.
1171		 */
1172		page_check_dirty_writeback(page, &dirty, &writeback);
1173		if (dirty || writeback)
1174			stat->nr_dirty++;
1175
1176		if (dirty && !writeback)
1177			stat->nr_unqueued_dirty++;
1178
1179		/*
1180		 * Treat this page as congested if the underlying BDI is or if
1181		 * pages are cycling through the LRU so quickly that the
1182		 * pages marked for immediate reclaim are making it to the
1183		 * end of the LRU a second time.
1184		 */
1185		mapping = page_mapping(page);
1186		if (((dirty || writeback) && mapping &&
1187		     inode_write_congested(mapping->host)) ||
1188		    (writeback && PageReclaim(page)))
1189			stat->nr_congested++;
1190
1191		/*
1192		 * If a page at the tail of the LRU is under writeback, there
1193		 * are three cases to consider.
1194		 *
1195		 * 1) If reclaim is encountering an excessive number of pages
1196		 *    under writeback and this page is both under writeback and
1197		 *    PageReclaim then it indicates that pages are being queued
1198		 *    for IO but are being recycled through the LRU before the
1199		 *    IO can complete. Waiting on the page itself risks an
1200		 *    indefinite stall if it is impossible to writeback the
1201		 *    page due to IO error or disconnected storage so instead
1202		 *    note that the LRU is being scanned too quickly and the
1203		 *    caller can stall after page list has been processed.
1204		 *
1205		 * 2) Global or new memcg reclaim encounters a page that is
1206		 *    not marked for immediate reclaim, or the caller does not
1207		 *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1208		 *    not to fs). In this case mark the page for immediate
1209		 *    reclaim and continue scanning.
1210		 *
1211		 *    Require may_enter_fs because we would wait on fs, which
1212		 *    may not have submitted IO yet. And the loop driver might
1213		 *    enter reclaim, and deadlock if it waits on a page for
1214		 *    which it is needed to do the write (loop masks off
1215		 *    __GFP_IO|__GFP_FS for this reason); but more thought
1216		 *    would probably show more reasons.
1217		 *
1218		 * 3) Legacy memcg encounters a page that is already marked
1219		 *    PageReclaim. memcg does not have any dirty pages
1220		 *    throttling so we could easily OOM just because too many
1221		 *    pages are in writeback and there is nothing else to
1222		 *    reclaim. Wait for the writeback to complete.
1223		 *
1224		 * In cases 1) and 2) we activate the pages to get them out of
1225		 * the way while we continue scanning for clean pages on the
1226		 * inactive list and refilling from the active list. The
1227		 * observation here is that waiting for disk writes is more
1228		 * expensive than potentially causing reloads down the line.
1229		 * Since they're marked for immediate reclaim, they won't put
1230		 * memory pressure on the cache working set any longer than it
1231		 * takes to write them to disk.
1232		 */
1233		if (PageWriteback(page)) {
1234			/* Case 1 above */
1235			if (current_is_kswapd() &&
1236			    PageReclaim(page) &&
1237			    test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1238				stat->nr_immediate++;
1239				goto activate_locked;
1240
1241			/* Case 2 above */
1242			} else if (sane_reclaim(sc) ||
1243			    !PageReclaim(page) || !may_enter_fs) {
1244				/*
1245				 * This is slightly racy - end_page_writeback()
1246				 * might have just cleared PageReclaim, then
1247				 * setting PageReclaim here end up interpreted
1248				 * as PageReadahead - but that does not matter
1249				 * enough to care.  What we do want is for this
1250				 * page to have PageReclaim set next time memcg
1251				 * reclaim reaches the tests above, so it will
1252				 * then wait_on_page_writeback() to avoid OOM;
1253				 * and it's also appropriate in global reclaim.
1254				 */
1255				SetPageReclaim(page);
1256				stat->nr_writeback++;
1257				goto activate_locked;
1258
1259			/* Case 3 above */
1260			} else {
1261				unlock_page(page);
1262				wait_on_page_writeback(page);
1263				/* then go back and try same page again */
1264				list_add_tail(&page->lru, page_list);
1265				continue;
1266			}
1267		}
1268
1269		if (!ignore_references)
1270			references = page_check_references(page, sc);
1271
1272		switch (references) {
1273		case PAGEREF_ACTIVATE:
1274			goto activate_locked;
1275		case PAGEREF_KEEP:
1276			stat->nr_ref_keep += nr_pages;
1277			goto keep_locked;
1278		case PAGEREF_RECLAIM:
1279		case PAGEREF_RECLAIM_CLEAN:
1280			; /* try to reclaim the page below */
1281		}
1282
1283		/*
1284		 * Anonymous process memory has backing store?
1285		 * Try to allocate it some swap space here.
1286		 * Lazyfree page could be freed directly
1287		 */
1288		if (PageAnon(page) && PageSwapBacked(page)) {
1289			if (!PageSwapCache(page)) {
1290				if (!(sc->gfp_mask & __GFP_IO))
1291					goto keep_locked;
1292				if (PageTransHuge(page)) {
1293					/* cannot split THP, skip it */
1294					if (!can_split_huge_page(page, NULL))
1295						goto activate_locked;
1296					/*
1297					 * Split pages without a PMD map right
1298					 * away. Chances are some or all of the
1299					 * tail pages can be freed without IO.
1300					 */
1301					if (!compound_mapcount(page) &&
1302					    split_huge_page_to_list(page,
1303								    page_list))
1304						goto activate_locked;
1305				}
1306				if (!add_to_swap(page)) {
1307					if (!PageTransHuge(page))
1308						goto activate_locked_split;
1309					/* Fallback to swap normal pages */
1310					if (split_huge_page_to_list(page,
1311								    page_list))
1312						goto activate_locked;
1313#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1314					count_vm_event(THP_SWPOUT_FALLBACK);
1315#endif
1316					if (!add_to_swap(page))
1317						goto activate_locked_split;
1318				}
1319
1320				may_enter_fs = 1;
1321
1322				/* Adding to swap updated mapping */
1323				mapping = page_mapping(page);
1324			}
1325		} else if (unlikely(PageTransHuge(page))) {
1326			/* Split file THP */
1327			if (split_huge_page_to_list(page, page_list))
1328				goto keep_locked;
1329		}
1330
1331		/*
1332		 * THP may get split above, need minus tail pages and update
1333		 * nr_pages to avoid accounting tail pages twice.
1334		 *
1335		 * The tail pages that are added into swap cache successfully
1336		 * reach here.
1337		 */
1338		if ((nr_pages > 1) && !PageTransHuge(page)) {
1339			sc->nr_scanned -= (nr_pages - 1);
1340			nr_pages = 1;
1341		}
1342
1343		/*
1344		 * The page is mapped into the page tables of one or more
1345		 * processes. Try to unmap it here.
1346		 */
1347		if (page_mapped(page)) {
1348			enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1349
1350			if (unlikely(PageTransHuge(page)))
1351				flags |= TTU_SPLIT_HUGE_PMD;
1352			if (!try_to_unmap(page, flags)) {
1353				stat->nr_unmap_fail += nr_pages;
1354				goto activate_locked;
1355			}
1356		}
1357
1358		if (PageDirty(page)) {
1359			/*
1360			 * Only kswapd can writeback filesystem pages
1361			 * to avoid risk of stack overflow. But avoid
1362			 * injecting inefficient single-page IO into
1363			 * flusher writeback as much as possible: only
1364			 * write pages when we've encountered many
1365			 * dirty pages, and when we've already scanned
1366			 * the rest of the LRU for clean pages and see
1367			 * the same dirty pages again (PageReclaim).
1368			 */
1369			if (page_is_file_cache(page) &&
1370			    (!current_is_kswapd() || !PageReclaim(page) ||
1371			     !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1372				/*
1373				 * Immediately reclaim when written back.
1374				 * Similar in principal to deactivate_page()
1375				 * except we already have the page isolated
1376				 * and know it's dirty
1377				 */
1378				inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1379				SetPageReclaim(page);
1380
1381				goto activate_locked;
1382			}
1383
1384			if (references == PAGEREF_RECLAIM_CLEAN)
1385				goto keep_locked;
1386			if (!may_enter_fs)
1387				goto keep_locked;
1388			if (!sc->may_writepage)
1389				goto keep_locked;
1390
1391			/*
1392			 * Page is dirty. Flush the TLB if a writable entry
1393			 * potentially exists to avoid CPU writes after IO
1394			 * starts and then write it out here.
1395			 */
1396			try_to_unmap_flush_dirty();
1397			switch (pageout(page, mapping, sc)) {
1398			case PAGE_KEEP:
1399				goto keep_locked;
1400			case PAGE_ACTIVATE:
1401				goto activate_locked;
1402			case PAGE_SUCCESS:
1403				if (PageWriteback(page))
1404					goto keep;
1405				if (PageDirty(page))
1406					goto keep;
1407
1408				/*
1409				 * A synchronous write - probably a ramdisk.  Go
1410				 * ahead and try to reclaim the page.
1411				 */
1412				if (!trylock_page(page))
1413					goto keep;
1414				if (PageDirty(page) || PageWriteback(page))
1415					goto keep_locked;
1416				mapping = page_mapping(page);
1417			case PAGE_CLEAN:
1418				; /* try to free the page below */
1419			}
1420		}
1421
1422		/*
1423		 * If the page has buffers, try to free the buffer mappings
1424		 * associated with this page. If we succeed we try to free
1425		 * the page as well.
1426		 *
1427		 * We do this even if the page is PageDirty().
1428		 * try_to_release_page() does not perform I/O, but it is
1429		 * possible for a page to have PageDirty set, but it is actually
1430		 * clean (all its buffers are clean).  This happens if the
1431		 * buffers were written out directly, with submit_bh(). ext3
1432		 * will do this, as well as the blockdev mapping.
1433		 * try_to_release_page() will discover that cleanness and will
1434		 * drop the buffers and mark the page clean - it can be freed.
1435		 *
1436		 * Rarely, pages can have buffers and no ->mapping.  These are
1437		 * the pages which were not successfully invalidated in
1438		 * truncate_complete_page().  We try to drop those buffers here
1439		 * and if that worked, and the page is no longer mapped into
1440		 * process address space (page_count == 1) it can be freed.
1441		 * Otherwise, leave the page on the LRU so it is swappable.
1442		 */
1443		if (page_has_private(page)) {
1444			if (!try_to_release_page(page, sc->gfp_mask))
1445				goto activate_locked;
1446			if (!mapping && page_count(page) == 1) {
1447				unlock_page(page);
1448				if (put_page_testzero(page))
1449					goto free_it;
1450				else {
1451					/*
1452					 * rare race with speculative reference.
1453					 * the speculative reference will free
1454					 * this page shortly, so we may
1455					 * increment nr_reclaimed here (and
1456					 * leave it off the LRU).
1457					 */
1458					nr_reclaimed++;
1459					continue;
1460				}
1461			}
1462		}
1463
1464		if (PageAnon(page) && !PageSwapBacked(page)) {
1465			/* follow __remove_mapping for reference */
1466			if (!page_ref_freeze(page, 1))
1467				goto keep_locked;
1468			if (PageDirty(page)) {
1469				page_ref_unfreeze(page, 1);
1470				goto keep_locked;
1471			}
1472
1473			count_vm_event(PGLAZYFREED);
1474			count_memcg_page_event(page, PGLAZYFREED);
1475		} else if (!mapping || !__remove_mapping(mapping, page, true))
1476			goto keep_locked;
1477
1478		unlock_page(page);
1479free_it:
1480		/*
1481		 * THP may get swapped out in a whole, need account
1482		 * all base pages.
 
 
 
1483		 */
1484		nr_reclaimed += nr_pages;
 
 
1485
1486		/*
1487		 * Is there need to periodically free_page_list? It would
1488		 * appear not as the counts should be low
1489		 */
1490		if (unlikely(PageTransHuge(page)))
 
1491			(*get_compound_page_dtor(page))(page);
1492		else
1493			list_add(&page->lru, &free_pages);
1494		continue;
1495
1496activate_locked_split:
1497		/*
1498		 * The tail pages that are failed to add into swap cache
1499		 * reach here.  Fixup nr_scanned and nr_pages.
1500		 */
1501		if (nr_pages > 1) {
1502			sc->nr_scanned -= (nr_pages - 1);
1503			nr_pages = 1;
1504		}
1505activate_locked:
1506		/* Not a candidate for swapping, so reclaim swap space. */
1507		if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1508						PageMlocked(page)))
1509			try_to_free_swap(page);
1510		VM_BUG_ON_PAGE(PageActive(page), page);
1511		if (!PageMlocked(page)) {
1512			int type = page_is_file_cache(page);
1513			SetPageActive(page);
1514			stat->nr_activate[type] += nr_pages;
1515			count_memcg_page_event(page, PGACTIVATE);
1516		}
1517keep_locked:
1518		unlock_page(page);
1519keep:
1520		list_add(&page->lru, &ret_pages);
1521		VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1522	}
1523
1524	pgactivate = stat->nr_activate[0] + stat->nr_activate[1];
1525
1526	mem_cgroup_uncharge_list(&free_pages);
1527	try_to_unmap_flush();
1528	free_unref_page_list(&free_pages);
1529
1530	list_splice(&ret_pages, page_list);
1531	count_vm_events(PGACTIVATE, pgactivate);
1532
 
 
 
 
 
 
 
 
 
 
1533	return nr_reclaimed;
1534}
1535
1536unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1537					    struct list_head *page_list)
1538{
1539	struct scan_control sc = {
1540		.gfp_mask = GFP_KERNEL,
1541		.priority = DEF_PRIORITY,
1542		.may_unmap = 1,
1543	};
1544	struct reclaim_stat dummy_stat;
1545	unsigned long ret;
1546	struct page *page, *next;
1547	LIST_HEAD(clean_pages);
1548
1549	list_for_each_entry_safe(page, next, page_list, lru) {
1550		if (page_is_file_cache(page) && !PageDirty(page) &&
1551		    !__PageMovable(page) && !PageUnevictable(page)) {
1552			ClearPageActive(page);
1553			list_move(&page->lru, &clean_pages);
1554		}
1555	}
1556
1557	ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1558			TTU_IGNORE_ACCESS, &dummy_stat, true);
1559	list_splice(&clean_pages, page_list);
1560	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1561	return ret;
1562}
1563
1564/*
1565 * Attempt to remove the specified page from its LRU.  Only take this page
1566 * if it is of the appropriate PageActive status.  Pages which are being
1567 * freed elsewhere are also ignored.
1568 *
1569 * page:	page to consider
1570 * mode:	one of the LRU isolation modes defined above
1571 *
1572 * returns 0 on success, -ve errno on failure.
1573 */
1574int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1575{
1576	int ret = -EINVAL;
1577
1578	/* Only take pages on the LRU. */
1579	if (!PageLRU(page))
1580		return ret;
1581
1582	/* Compaction should not handle unevictable pages but CMA can do so */
1583	if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1584		return ret;
1585
1586	ret = -EBUSY;
1587
1588	/*
1589	 * To minimise LRU disruption, the caller can indicate that it only
1590	 * wants to isolate pages it will be able to operate on without
1591	 * blocking - clean pages for the most part.
1592	 *
1593	 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1594	 * that it is possible to migrate without blocking
1595	 */
1596	if (mode & ISOLATE_ASYNC_MIGRATE) {
1597		/* All the caller can do on PageWriteback is block */
1598		if (PageWriteback(page))
1599			return ret;
1600
1601		if (PageDirty(page)) {
1602			struct address_space *mapping;
1603			bool migrate_dirty;
1604
1605			/*
1606			 * Only pages without mappings or that have a
1607			 * ->migratepage callback are possible to migrate
1608			 * without blocking. However, we can be racing with
1609			 * truncation so it's necessary to lock the page
1610			 * to stabilise the mapping as truncation holds
1611			 * the page lock until after the page is removed
1612			 * from the page cache.
1613			 */
1614			if (!trylock_page(page))
1615				return ret;
1616
1617			mapping = page_mapping(page);
1618			migrate_dirty = !mapping || mapping->a_ops->migratepage;
1619			unlock_page(page);
1620			if (!migrate_dirty)
1621				return ret;
1622		}
1623	}
1624
1625	if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1626		return ret;
1627
1628	if (likely(get_page_unless_zero(page))) {
1629		/*
1630		 * Be careful not to clear PageLRU until after we're
1631		 * sure the page is not being freed elsewhere -- the
1632		 * page release code relies on it.
1633		 */
1634		ClearPageLRU(page);
1635		ret = 0;
1636	}
1637
1638	return ret;
1639}
1640
1641
1642/*
1643 * Update LRU sizes after isolating pages. The LRU size updates must
1644 * be complete before mem_cgroup_update_lru_size due to a santity check.
1645 */
1646static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1647			enum lru_list lru, unsigned long *nr_zone_taken)
1648{
1649	int zid;
1650
1651	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1652		if (!nr_zone_taken[zid])
1653			continue;
1654
1655		__update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1656#ifdef CONFIG_MEMCG
1657		mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1658#endif
1659	}
1660
1661}
1662
1663/**
1664 * pgdat->lru_lock is heavily contended.  Some of the functions that
1665 * shrink the lists perform better by taking out a batch of pages
1666 * and working on them outside the LRU lock.
1667 *
1668 * For pagecache intensive workloads, this function is the hottest
1669 * spot in the kernel (apart from copy_*_user functions).
1670 *
1671 * Appropriate locks must be held before calling this function.
1672 *
1673 * @nr_to_scan:	The number of eligible pages to look through on the list.
1674 * @lruvec:	The LRU vector to pull pages from.
1675 * @dst:	The temp list to put pages on to.
1676 * @nr_scanned:	The number of pages that were scanned.
1677 * @sc:		The scan_control struct for this reclaim session
1678 * @mode:	One of the LRU isolation modes
1679 * @lru:	LRU list id for isolating
1680 *
1681 * returns how many pages were moved onto *@dst.
1682 */
1683static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1684		struct lruvec *lruvec, struct list_head *dst,
1685		unsigned long *nr_scanned, struct scan_control *sc,
1686		enum lru_list lru)
1687{
1688	struct list_head *src = &lruvec->lists[lru];
1689	unsigned long nr_taken = 0;
1690	unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1691	unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1692	unsigned long skipped = 0;
1693	unsigned long scan, total_scan, nr_pages;
1694	LIST_HEAD(pages_skipped);
1695	isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED);
1696
1697	total_scan = 0;
1698	scan = 0;
1699	while (scan < nr_to_scan && !list_empty(src)) {
 
 
1700		struct page *page;
1701
1702		page = lru_to_page(src);
1703		prefetchw_prev_lru_page(page, src, flags);
1704
1705		VM_BUG_ON_PAGE(!PageLRU(page), page);
1706
1707		nr_pages = compound_nr(page);
1708		total_scan += nr_pages;
1709
1710		if (page_zonenum(page) > sc->reclaim_idx) {
1711			list_move(&page->lru, &pages_skipped);
1712			nr_skipped[page_zonenum(page)] += nr_pages;
1713			continue;
1714		}
1715
1716		/*
1717		 * Do not count skipped pages because that makes the function
1718		 * return with no isolated pages if the LRU mostly contains
1719		 * ineligible pages.  This causes the VM to not reclaim any
1720		 * pages, triggering a premature OOM.
1721		 *
1722		 * Account all tail pages of THP.  This would not cause
1723		 * premature OOM since __isolate_lru_page() returns -EBUSY
1724		 * only when the page is being freed somewhere else.
1725		 */
1726		scan += nr_pages;
1727		switch (__isolate_lru_page(page, mode)) {
1728		case 0:
 
1729			nr_taken += nr_pages;
1730			nr_zone_taken[page_zonenum(page)] += nr_pages;
1731			list_move(&page->lru, dst);
1732			break;
1733
1734		case -EBUSY:
1735			/* else it is being freed elsewhere */
1736			list_move(&page->lru, src);
1737			continue;
1738
1739		default:
1740			BUG();
1741		}
1742	}
1743
1744	/*
1745	 * Splice any skipped pages to the start of the LRU list. Note that
1746	 * this disrupts the LRU order when reclaiming for lower zones but
1747	 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1748	 * scanning would soon rescan the same pages to skip and put the
1749	 * system at risk of premature OOM.
1750	 */
1751	if (!list_empty(&pages_skipped)) {
1752		int zid;
1753
1754		list_splice(&pages_skipped, src);
1755		for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1756			if (!nr_skipped[zid])
1757				continue;
1758
1759			__count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1760			skipped += nr_skipped[zid];
1761		}
1762	}
1763	*nr_scanned = total_scan;
1764	trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1765				    total_scan, skipped, nr_taken, mode, lru);
1766	update_lru_sizes(lruvec, lru, nr_zone_taken);
1767	return nr_taken;
1768}
1769
1770/**
1771 * isolate_lru_page - tries to isolate a page from its LRU list
1772 * @page: page to isolate from its LRU list
1773 *
1774 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1775 * vmstat statistic corresponding to whatever LRU list the page was on.
1776 *
1777 * Returns 0 if the page was removed from an LRU list.
1778 * Returns -EBUSY if the page was not on an LRU list.
1779 *
1780 * The returned page will have PageLRU() cleared.  If it was found on
1781 * the active list, it will have PageActive set.  If it was found on
1782 * the unevictable list, it will have the PageUnevictable bit set. That flag
1783 * may need to be cleared by the caller before letting the page go.
1784 *
1785 * The vmstat statistic corresponding to the list on which the page was
1786 * found will be decremented.
1787 *
1788 * Restrictions:
1789 *
1790 * (1) Must be called with an elevated refcount on the page. This is a
1791 *     fundamentnal difference from isolate_lru_pages (which is called
1792 *     without a stable reference).
1793 * (2) the lru_lock must not be held.
1794 * (3) interrupts must be enabled.
1795 */
1796int isolate_lru_page(struct page *page)
1797{
1798	int ret = -EBUSY;
1799
1800	VM_BUG_ON_PAGE(!page_count(page), page);
1801	WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1802
1803	if (PageLRU(page)) {
1804		pg_data_t *pgdat = page_pgdat(page);
1805		struct lruvec *lruvec;
1806
1807		spin_lock_irq(&pgdat->lru_lock);
1808		lruvec = mem_cgroup_page_lruvec(page, pgdat);
1809		if (PageLRU(page)) {
1810			int lru = page_lru(page);
1811			get_page(page);
1812			ClearPageLRU(page);
1813			del_page_from_lru_list(page, lruvec, lru);
1814			ret = 0;
1815		}
1816		spin_unlock_irq(&pgdat->lru_lock);
1817	}
1818	return ret;
1819}
1820
1821/*
1822 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1823 * then get resheduled. When there are massive number of tasks doing page
1824 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1825 * the LRU list will go small and be scanned faster than necessary, leading to
1826 * unnecessary swapping, thrashing and OOM.
1827 */
1828static int too_many_isolated(struct pglist_data *pgdat, int file,
1829		struct scan_control *sc)
1830{
1831	unsigned long inactive, isolated;
1832
1833	if (current_is_kswapd())
1834		return 0;
1835
1836	if (!sane_reclaim(sc))
1837		return 0;
1838
1839	if (file) {
1840		inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1841		isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1842	} else {
1843		inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1844		isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1845	}
1846
1847	/*
1848	 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1849	 * won't get blocked by normal direct-reclaimers, forming a circular
1850	 * deadlock.
1851	 */
1852	if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1853		inactive >>= 3;
1854
1855	return isolated > inactive;
1856}
1857
1858/*
1859 * This moves pages from @list to corresponding LRU list.
1860 *
1861 * We move them the other way if the page is referenced by one or more
1862 * processes, from rmap.
1863 *
1864 * If the pages are mostly unmapped, the processing is fast and it is
1865 * appropriate to hold zone_lru_lock across the whole operation.  But if
1866 * the pages are mapped, the processing is slow (page_referenced()) so we
1867 * should drop zone_lru_lock around each page.  It's impossible to balance
1868 * this, so instead we remove the pages from the LRU while processing them.
1869 * It is safe to rely on PG_active against the non-LRU pages in here because
1870 * nobody will play with that bit on a non-LRU page.
1871 *
1872 * The downside is that we have to touch page->_refcount against each page.
1873 * But we had to alter page->flags anyway.
1874 *
1875 * Returns the number of pages moved to the given lruvec.
1876 */
1877
1878static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec,
1879						     struct list_head *list)
1880{
 
1881	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1882	int nr_pages, nr_moved = 0;
1883	LIST_HEAD(pages_to_free);
1884	struct page *page;
1885	enum lru_list lru;
1886
1887	while (!list_empty(list)) {
1888		page = lru_to_page(list);
 
 
 
 
 
1889		VM_BUG_ON_PAGE(PageLRU(page), page);
 
1890		if (unlikely(!page_evictable(page))) {
1891			list_del(&page->lru);
1892			spin_unlock_irq(&pgdat->lru_lock);
1893			putback_lru_page(page);
1894			spin_lock_irq(&pgdat->lru_lock);
1895			continue;
1896		}
 
1897		lruvec = mem_cgroup_page_lruvec(page, pgdat);
1898
1899		SetPageLRU(page);
1900		lru = page_lru(page);
 
1901
1902		nr_pages = hpage_nr_pages(page);
1903		update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1904		list_move(&page->lru, &lruvec->lists[lru]);
1905
 
1906		if (put_page_testzero(page)) {
1907			__ClearPageLRU(page);
1908			__ClearPageActive(page);
1909			del_page_from_lru_list(page, lruvec, lru);
1910
1911			if (unlikely(PageCompound(page))) {
1912				spin_unlock_irq(&pgdat->lru_lock);
 
1913				(*get_compound_page_dtor(page))(page);
1914				spin_lock_irq(&pgdat->lru_lock);
1915			} else
1916				list_add(&page->lru, &pages_to_free);
1917		} else {
1918			nr_moved += nr_pages;
1919		}
1920	}
1921
1922	/*
1923	 * To save our caller's stack, now use input list for pages to free.
1924	 */
1925	list_splice(&pages_to_free, list);
1926
1927	return nr_moved;
1928}
1929
1930/*
1931 * If a kernel thread (such as nfsd for loop-back mounts) services
1932 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1933 * In that case we should only throttle if the backing device it is
1934 * writing to is congested.  In other cases it is safe to throttle.
1935 */
1936static int current_may_throttle(void)
1937{
1938	return !(current->flags & PF_LESS_THROTTLE) ||
1939		current->backing_dev_info == NULL ||
1940		bdi_write_congested(current->backing_dev_info);
1941}
1942
1943/*
1944 * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1945 * of reclaimed pages
1946 */
1947static noinline_for_stack unsigned long
1948shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1949		     struct scan_control *sc, enum lru_list lru)
1950{
1951	LIST_HEAD(page_list);
1952	unsigned long nr_scanned;
1953	unsigned long nr_reclaimed = 0;
1954	unsigned long nr_taken;
1955	struct reclaim_stat stat;
 
1956	int file = is_file_lru(lru);
1957	enum vm_event_item item;
1958	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1959	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1960	bool stalled = false;
1961
1962	while (unlikely(too_many_isolated(pgdat, file, sc))) {
1963		if (stalled)
1964			return 0;
1965
1966		/* wait a bit for the reclaimer. */
1967		msleep(100);
1968		stalled = true;
1969
1970		/* We are about to die and free our memory. Return now. */
1971		if (fatal_signal_pending(current))
1972			return SWAP_CLUSTER_MAX;
1973	}
1974
1975	lru_add_drain();
1976
 
 
 
1977	spin_lock_irq(&pgdat->lru_lock);
1978
1979	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1980				     &nr_scanned, sc, lru);
1981
1982	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1983	reclaim_stat->recent_scanned[file] += nr_taken;
1984
1985	item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT;
1986	if (global_reclaim(sc))
1987		__count_vm_events(item, nr_scanned);
1988	__count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
 
 
 
 
 
 
 
1989	spin_unlock_irq(&pgdat->lru_lock);
1990
1991	if (nr_taken == 0)
1992		return 0;
1993
1994	nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1995				&stat, false);
1996
1997	spin_lock_irq(&pgdat->lru_lock);
1998
1999	item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT;
2000	if (global_reclaim(sc))
2001		__count_vm_events(item, nr_reclaimed);
2002	__count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
2003	reclaim_stat->recent_rotated[0] += stat.nr_activate[0];
2004	reclaim_stat->recent_rotated[1] += stat.nr_activate[1];
 
 
 
 
 
2005
2006	move_pages_to_lru(lruvec, &page_list);
2007
2008	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2009
2010	spin_unlock_irq(&pgdat->lru_lock);
2011
2012	mem_cgroup_uncharge_list(&page_list);
2013	free_unref_page_list(&page_list);
2014
2015	/*
2016	 * If dirty pages are scanned that are not queued for IO, it
2017	 * implies that flushers are not doing their job. This can
2018	 * happen when memory pressure pushes dirty pages to the end of
2019	 * the LRU before the dirty limits are breached and the dirty
2020	 * data has expired. It can also happen when the proportion of
2021	 * dirty pages grows not through writes but through memory
2022	 * pressure reclaiming all the clean cache. And in some cases,
2023	 * the flushers simply cannot keep up with the allocation
2024	 * rate. Nudge the flusher threads in case they are asleep.
2025	 */
2026	if (stat.nr_unqueued_dirty == nr_taken)
2027		wakeup_flusher_threads(WB_REASON_VMSCAN);
2028
2029	sc->nr.dirty += stat.nr_dirty;
2030	sc->nr.congested += stat.nr_congested;
2031	sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
2032	sc->nr.writeback += stat.nr_writeback;
2033	sc->nr.immediate += stat.nr_immediate;
2034	sc->nr.taken += nr_taken;
2035	if (file)
2036		sc->nr.file_taken += nr_taken;
2037
2038	trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
2039			nr_scanned, nr_reclaimed, &stat, sc->priority, file);
2040	return nr_reclaimed;
2041}
2042
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2043static void shrink_active_list(unsigned long nr_to_scan,
2044			       struct lruvec *lruvec,
2045			       struct scan_control *sc,
2046			       enum lru_list lru)
2047{
2048	unsigned long nr_taken;
2049	unsigned long nr_scanned;
2050	unsigned long vm_flags;
2051	LIST_HEAD(l_hold);	/* The pages which were snipped off */
2052	LIST_HEAD(l_active);
2053	LIST_HEAD(l_inactive);
2054	struct page *page;
2055	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2056	unsigned nr_deactivate, nr_activate;
2057	unsigned nr_rotated = 0;
 
2058	int file = is_file_lru(lru);
2059	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2060
2061	lru_add_drain();
2062
 
 
 
2063	spin_lock_irq(&pgdat->lru_lock);
2064
2065	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
2066				     &nr_scanned, sc, lru);
2067
2068	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
2069	reclaim_stat->recent_scanned[file] += nr_taken;
2070
2071	__count_vm_events(PGREFILL, nr_scanned);
2072	__count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
2073
2074	spin_unlock_irq(&pgdat->lru_lock);
2075
2076	while (!list_empty(&l_hold)) {
2077		cond_resched();
2078		page = lru_to_page(&l_hold);
2079		list_del(&page->lru);
2080
2081		if (unlikely(!page_evictable(page))) {
2082			putback_lru_page(page);
2083			continue;
2084		}
2085
2086		if (unlikely(buffer_heads_over_limit)) {
2087			if (page_has_private(page) && trylock_page(page)) {
2088				if (page_has_private(page))
2089					try_to_release_page(page, 0);
2090				unlock_page(page);
2091			}
2092		}
2093
2094		if (page_referenced(page, 0, sc->target_mem_cgroup,
2095				    &vm_flags)) {
2096			nr_rotated += hpage_nr_pages(page);
2097			/*
2098			 * Identify referenced, file-backed active pages and
2099			 * give them one more trip around the active list. So
2100			 * that executable code get better chances to stay in
2101			 * memory under moderate memory pressure.  Anon pages
2102			 * are not likely to be evicted by use-once streaming
2103			 * IO, plus JVM can create lots of anon VM_EXEC pages,
2104			 * so we ignore them here.
2105			 */
2106			if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
2107				list_add(&page->lru, &l_active);
2108				continue;
2109			}
2110		}
2111
2112		ClearPageActive(page);	/* we are de-activating */
2113		SetPageWorkingset(page);
2114		list_add(&page->lru, &l_inactive);
2115	}
2116
2117	/*
2118	 * Move pages back to the lru list.
2119	 */
2120	spin_lock_irq(&pgdat->lru_lock);
2121	/*
2122	 * Count referenced pages from currently used mappings as rotated,
2123	 * even though only some of them are actually re-activated.  This
2124	 * helps balance scan pressure between file and anonymous pages in
2125	 * get_scan_count.
2126	 */
2127	reclaim_stat->recent_rotated[file] += nr_rotated;
2128
2129	nr_activate = move_pages_to_lru(lruvec, &l_active);
2130	nr_deactivate = move_pages_to_lru(lruvec, &l_inactive);
2131	/* Keep all free pages in l_active list */
2132	list_splice(&l_inactive, &l_active);
2133
2134	__count_vm_events(PGDEACTIVATE, nr_deactivate);
2135	__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate);
2136
2137	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
2138	spin_unlock_irq(&pgdat->lru_lock);
2139
2140	mem_cgroup_uncharge_list(&l_active);
2141	free_unref_page_list(&l_active);
2142	trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2143			nr_deactivate, nr_rotated, sc->priority, file);
2144}
2145
2146unsigned long reclaim_pages(struct list_head *page_list)
2147{
2148	int nid = -1;
2149	unsigned long nr_reclaimed = 0;
2150	LIST_HEAD(node_page_list);
2151	struct reclaim_stat dummy_stat;
2152	struct page *page;
2153	struct scan_control sc = {
2154		.gfp_mask = GFP_KERNEL,
2155		.priority = DEF_PRIORITY,
2156		.may_writepage = 1,
2157		.may_unmap = 1,
2158		.may_swap = 1,
2159	};
2160
2161	while (!list_empty(page_list)) {
2162		page = lru_to_page(page_list);
2163		if (nid == -1) {
2164			nid = page_to_nid(page);
2165			INIT_LIST_HEAD(&node_page_list);
2166		}
2167
2168		if (nid == page_to_nid(page)) {
2169			ClearPageActive(page);
2170			list_move(&page->lru, &node_page_list);
2171			continue;
2172		}
2173
2174		nr_reclaimed += shrink_page_list(&node_page_list,
2175						NODE_DATA(nid),
2176						&sc, 0,
2177						&dummy_stat, false);
2178		while (!list_empty(&node_page_list)) {
2179			page = lru_to_page(&node_page_list);
2180			list_del(&page->lru);
2181			putback_lru_page(page);
2182		}
2183
2184		nid = -1;
2185	}
2186
2187	if (!list_empty(&node_page_list)) {
2188		nr_reclaimed += shrink_page_list(&node_page_list,
2189						NODE_DATA(nid),
2190						&sc, 0,
2191						&dummy_stat, false);
2192		while (!list_empty(&node_page_list)) {
2193			page = lru_to_page(&node_page_list);
2194			list_del(&page->lru);
2195			putback_lru_page(page);
2196		}
2197	}
2198
2199	return nr_reclaimed;
2200}
2201
2202/*
2203 * The inactive anon list should be small enough that the VM never has
2204 * to do too much work.
2205 *
2206 * The inactive file list should be small enough to leave most memory
2207 * to the established workingset on the scan-resistant active list,
2208 * but large enough to avoid thrashing the aggregate readahead window.
2209 *
2210 * Both inactive lists should also be large enough that each inactive
2211 * page has a chance to be referenced again before it is reclaimed.
2212 *
2213 * If that fails and refaulting is observed, the inactive list grows.
2214 *
2215 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2216 * on this LRU, maintained by the pageout code. An inactive_ratio
2217 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2218 *
2219 * total     target    max
2220 * memory    ratio     inactive
2221 * -------------------------------------
2222 *   10MB       1         5MB
2223 *  100MB       1        50MB
2224 *    1GB       3       250MB
2225 *   10GB      10       0.9GB
2226 *  100GB      31         3GB
2227 *    1TB     101        10GB
2228 *   10TB     320        32GB
2229 */
2230static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2231				 struct scan_control *sc, bool trace)
 
2232{
2233	enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2234	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2235	enum lru_list inactive_lru = file * LRU_FILE;
2236	unsigned long inactive, active;
2237	unsigned long inactive_ratio;
2238	unsigned long refaults;
2239	unsigned long gb;
2240
2241	/*
2242	 * If we don't have swap space, anonymous page deactivation
2243	 * is pointless.
2244	 */
2245	if (!file && !total_swap_pages)
2246		return false;
2247
2248	inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2249	active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2250
 
 
 
 
 
2251	/*
2252	 * When refaults are being observed, it means a new workingset
2253	 * is being established. Disable active list protection to get
2254	 * rid of the stale workingset quickly.
2255	 */
2256	refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
2257	if (file && lruvec->refaults != refaults) {
2258		inactive_ratio = 0;
2259	} else {
2260		gb = (inactive + active) >> (30 - PAGE_SHIFT);
2261		if (gb)
2262			inactive_ratio = int_sqrt(10 * gb);
2263		else
2264			inactive_ratio = 1;
2265	}
2266
2267	if (trace)
2268		trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2269			lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2270			lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2271			inactive_ratio, file);
2272
2273	return inactive * inactive_ratio < active;
2274}
2275
2276static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2277				 struct lruvec *lruvec, struct scan_control *sc)
 
2278{
2279	if (is_active_lru(lru)) {
2280		if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true))
 
2281			shrink_active_list(nr_to_scan, lruvec, sc, lru);
2282		return 0;
2283	}
2284
2285	return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2286}
2287
2288enum scan_balance {
2289	SCAN_EQUAL,
2290	SCAN_FRACT,
2291	SCAN_ANON,
2292	SCAN_FILE,
2293};
2294
2295/*
2296 * Determine how aggressively the anon and file LRU lists should be
2297 * scanned.  The relative value of each set of LRU lists is determined
2298 * by looking at the fraction of the pages scanned we did rotate back
2299 * onto the active list instead of evict.
2300 *
2301 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2302 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2303 */
2304static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2305			   struct scan_control *sc, unsigned long *nr,
2306			   unsigned long *lru_pages)
2307{
2308	int swappiness = mem_cgroup_swappiness(memcg);
2309	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2310	u64 fraction[2];
2311	u64 denominator = 0;	/* gcc */
2312	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2313	unsigned long anon_prio, file_prio;
2314	enum scan_balance scan_balance;
2315	unsigned long anon, file;
2316	unsigned long ap, fp;
2317	enum lru_list lru;
2318
2319	/* If we have no swap space, do not bother scanning anon pages. */
2320	if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2321		scan_balance = SCAN_FILE;
2322		goto out;
2323	}
2324
2325	/*
2326	 * Global reclaim will swap to prevent OOM even with no
2327	 * swappiness, but memcg users want to use this knob to
2328	 * disable swapping for individual groups completely when
2329	 * using the memory controller's swap limit feature would be
2330	 * too expensive.
2331	 */
2332	if (!global_reclaim(sc) && !swappiness) {
2333		scan_balance = SCAN_FILE;
2334		goto out;
2335	}
2336
2337	/*
2338	 * Do not apply any pressure balancing cleverness when the
2339	 * system is close to OOM, scan both anon and file equally
2340	 * (unless the swappiness setting disagrees with swapping).
2341	 */
2342	if (!sc->priority && swappiness) {
2343		scan_balance = SCAN_EQUAL;
2344		goto out;
2345	}
2346
2347	/*
2348	 * Prevent the reclaimer from falling into the cache trap: as
2349	 * cache pages start out inactive, every cache fault will tip
2350	 * the scan balance towards the file LRU.  And as the file LRU
2351	 * shrinks, so does the window for rotation from references.
2352	 * This means we have a runaway feedback loop where a tiny
2353	 * thrashing file LRU becomes infinitely more attractive than
2354	 * anon pages.  Try to detect this based on file LRU size.
2355	 */
2356	if (global_reclaim(sc)) {
2357		unsigned long pgdatfile;
2358		unsigned long pgdatfree;
2359		int z;
2360		unsigned long total_high_wmark = 0;
2361
2362		pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2363		pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2364			   node_page_state(pgdat, NR_INACTIVE_FILE);
2365
2366		for (z = 0; z < MAX_NR_ZONES; z++) {
2367			struct zone *zone = &pgdat->node_zones[z];
2368			if (!managed_zone(zone))
2369				continue;
2370
2371			total_high_wmark += high_wmark_pages(zone);
2372		}
2373
2374		if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2375			/*
2376			 * Force SCAN_ANON if there are enough inactive
2377			 * anonymous pages on the LRU in eligible zones.
2378			 * Otherwise, the small LRU gets thrashed.
2379			 */
2380			if (!inactive_list_is_low(lruvec, false, sc, false) &&
2381			    lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2382					>> sc->priority) {
2383				scan_balance = SCAN_ANON;
2384				goto out;
2385			}
2386		}
2387	}
2388
2389	/*
2390	 * If there is enough inactive page cache, i.e. if the size of the
2391	 * inactive list is greater than that of the active list *and* the
2392	 * inactive list actually has some pages to scan on this priority, we
2393	 * do not reclaim anything from the anonymous working set right now.
2394	 * Without the second condition we could end up never scanning an
2395	 * lruvec even if it has plenty of old anonymous pages unless the
2396	 * system is under heavy pressure.
2397	 */
2398	if (!inactive_list_is_low(lruvec, true, sc, false) &&
2399	    lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2400		scan_balance = SCAN_FILE;
2401		goto out;
2402	}
2403
2404	scan_balance = SCAN_FRACT;
2405
2406	/*
2407	 * With swappiness at 100, anonymous and file have the same priority.
2408	 * This scanning priority is essentially the inverse of IO cost.
2409	 */
2410	anon_prio = swappiness;
2411	file_prio = 200 - anon_prio;
2412
2413	/*
2414	 * OK, so we have swap space and a fair amount of page cache
2415	 * pages.  We use the recently rotated / recently scanned
2416	 * ratios to determine how valuable each cache is.
2417	 *
2418	 * Because workloads change over time (and to avoid overflow)
2419	 * we keep these statistics as a floating average, which ends
2420	 * up weighing recent references more than old ones.
2421	 *
2422	 * anon in [0], file in [1]
2423	 */
2424
2425	anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2426		lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2427	file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2428		lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2429
2430	spin_lock_irq(&pgdat->lru_lock);
2431	if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2432		reclaim_stat->recent_scanned[0] /= 2;
2433		reclaim_stat->recent_rotated[0] /= 2;
2434	}
2435
2436	if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2437		reclaim_stat->recent_scanned[1] /= 2;
2438		reclaim_stat->recent_rotated[1] /= 2;
2439	}
2440
2441	/*
2442	 * The amount of pressure on anon vs file pages is inversely
2443	 * proportional to the fraction of recently scanned pages on
2444	 * each list that were recently referenced and in active use.
2445	 */
2446	ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2447	ap /= reclaim_stat->recent_rotated[0] + 1;
2448
2449	fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2450	fp /= reclaim_stat->recent_rotated[1] + 1;
2451	spin_unlock_irq(&pgdat->lru_lock);
2452
2453	fraction[0] = ap;
2454	fraction[1] = fp;
2455	denominator = ap + fp + 1;
2456out:
2457	*lru_pages = 0;
2458	for_each_evictable_lru(lru) {
2459		int file = is_file_lru(lru);
2460		unsigned long lruvec_size;
2461		unsigned long scan;
2462		unsigned long protection;
2463
2464		lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2465		protection = mem_cgroup_protection(memcg,
2466						   sc->memcg_low_reclaim);
2467
2468		if (protection) {
2469			/*
2470			 * Scale a cgroup's reclaim pressure by proportioning
2471			 * its current usage to its memory.low or memory.min
2472			 * setting.
2473			 *
2474			 * This is important, as otherwise scanning aggression
2475			 * becomes extremely binary -- from nothing as we
2476			 * approach the memory protection threshold, to totally
2477			 * nominal as we exceed it.  This results in requiring
2478			 * setting extremely liberal protection thresholds. It
2479			 * also means we simply get no protection at all if we
2480			 * set it too low, which is not ideal.
2481			 *
2482			 * If there is any protection in place, we reduce scan
2483			 * pressure by how much of the total memory used is
2484			 * within protection thresholds.
2485			 *
2486			 * There is one special case: in the first reclaim pass,
2487			 * we skip over all groups that are within their low
2488			 * protection. If that fails to reclaim enough pages to
2489			 * satisfy the reclaim goal, we come back and override
2490			 * the best-effort low protection. However, we still
2491			 * ideally want to honor how well-behaved groups are in
2492			 * that case instead of simply punishing them all
2493			 * equally. As such, we reclaim them based on how much
2494			 * memory they are using, reducing the scan pressure
2495			 * again by how much of the total memory used is under
2496			 * hard protection.
2497			 */
2498			unsigned long cgroup_size = mem_cgroup_size(memcg);
2499
2500			/* Avoid TOCTOU with earlier protection check */
2501			cgroup_size = max(cgroup_size, protection);
2502
2503			scan = lruvec_size - lruvec_size * protection /
2504				cgroup_size;
2505
2506			/*
2507			 * Minimally target SWAP_CLUSTER_MAX pages to keep
2508			 * reclaim moving forwards, avoiding decremeting
2509			 * sc->priority further than desirable.
2510			 */
2511			scan = max(scan, SWAP_CLUSTER_MAX);
2512		} else {
2513			scan = lruvec_size;
2514		}
2515
2516		scan >>= sc->priority;
2517
 
 
2518		/*
2519		 * If the cgroup's already been deleted, make sure to
2520		 * scrape out the remaining cache.
2521		 */
2522		if (!scan && !mem_cgroup_online(memcg))
2523			scan = min(lruvec_size, SWAP_CLUSTER_MAX);
2524
2525		switch (scan_balance) {
2526		case SCAN_EQUAL:
2527			/* Scan lists relative to size */
2528			break;
2529		case SCAN_FRACT:
2530			/*
2531			 * Scan types proportional to swappiness and
2532			 * their relative recent reclaim efficiency.
2533			 * Make sure we don't miss the last page
2534			 * because of a round-off error.
2535			 */
2536			scan = DIV64_U64_ROUND_UP(scan * fraction[file],
2537						  denominator);
2538			break;
2539		case SCAN_FILE:
2540		case SCAN_ANON:
2541			/* Scan one type exclusively */
2542			if ((scan_balance == SCAN_FILE) != file) {
2543				lruvec_size = 0;
2544				scan = 0;
2545			}
2546			break;
2547		default:
2548			/* Look ma, no brain */
2549			BUG();
2550		}
2551
2552		*lru_pages += lruvec_size;
2553		nr[lru] = scan;
2554	}
2555}
2556
2557/*
2558 * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2559 */
2560static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2561			      struct scan_control *sc, unsigned long *lru_pages)
2562{
2563	struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2564	unsigned long nr[NR_LRU_LISTS];
2565	unsigned long targets[NR_LRU_LISTS];
2566	unsigned long nr_to_scan;
2567	enum lru_list lru;
2568	unsigned long nr_reclaimed = 0;
2569	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2570	struct blk_plug plug;
2571	bool scan_adjusted;
2572
2573	get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2574
2575	/* Record the original scan target for proportional adjustments later */
2576	memcpy(targets, nr, sizeof(nr));
2577
2578	/*
2579	 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2580	 * event that can occur when there is little memory pressure e.g.
2581	 * multiple streaming readers/writers. Hence, we do not abort scanning
2582	 * when the requested number of pages are reclaimed when scanning at
2583	 * DEF_PRIORITY on the assumption that the fact we are direct
2584	 * reclaiming implies that kswapd is not keeping up and it is best to
2585	 * do a batch of work at once. For memcg reclaim one check is made to
2586	 * abort proportional reclaim if either the file or anon lru has already
2587	 * dropped to zero at the first pass.
2588	 */
2589	scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2590			 sc->priority == DEF_PRIORITY);
2591
2592	blk_start_plug(&plug);
2593	while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2594					nr[LRU_INACTIVE_FILE]) {
2595		unsigned long nr_anon, nr_file, percentage;
2596		unsigned long nr_scanned;
2597
2598		for_each_evictable_lru(lru) {
2599			if (nr[lru]) {
2600				nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2601				nr[lru] -= nr_to_scan;
2602
2603				nr_reclaimed += shrink_list(lru, nr_to_scan,
2604							    lruvec, sc);
2605			}
2606		}
2607
2608		cond_resched();
2609
2610		if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2611			continue;
2612
2613		/*
2614		 * For kswapd and memcg, reclaim at least the number of pages
2615		 * requested. Ensure that the anon and file LRUs are scanned
2616		 * proportionally what was requested by get_scan_count(). We
2617		 * stop reclaiming one LRU and reduce the amount scanning
2618		 * proportional to the original scan target.
2619		 */
2620		nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2621		nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2622
2623		/*
2624		 * It's just vindictive to attack the larger once the smaller
2625		 * has gone to zero.  And given the way we stop scanning the
2626		 * smaller below, this makes sure that we only make one nudge
2627		 * towards proportionality once we've got nr_to_reclaim.
2628		 */
2629		if (!nr_file || !nr_anon)
2630			break;
2631
2632		if (nr_file > nr_anon) {
2633			unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2634						targets[LRU_ACTIVE_ANON] + 1;
2635			lru = LRU_BASE;
2636			percentage = nr_anon * 100 / scan_target;
2637		} else {
2638			unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2639						targets[LRU_ACTIVE_FILE] + 1;
2640			lru = LRU_FILE;
2641			percentage = nr_file * 100 / scan_target;
2642		}
2643
2644		/* Stop scanning the smaller of the LRU */
2645		nr[lru] = 0;
2646		nr[lru + LRU_ACTIVE] = 0;
2647
2648		/*
2649		 * Recalculate the other LRU scan count based on its original
2650		 * scan target and the percentage scanning already complete
2651		 */
2652		lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2653		nr_scanned = targets[lru] - nr[lru];
2654		nr[lru] = targets[lru] * (100 - percentage) / 100;
2655		nr[lru] -= min(nr[lru], nr_scanned);
2656
2657		lru += LRU_ACTIVE;
2658		nr_scanned = targets[lru] - nr[lru];
2659		nr[lru] = targets[lru] * (100 - percentage) / 100;
2660		nr[lru] -= min(nr[lru], nr_scanned);
2661
2662		scan_adjusted = true;
2663	}
2664	blk_finish_plug(&plug);
2665	sc->nr_reclaimed += nr_reclaimed;
2666
2667	/*
2668	 * Even if we did not try to evict anon pages at all, we want to
2669	 * rebalance the anon lru active/inactive ratio.
2670	 */
2671	if (inactive_list_is_low(lruvec, false, sc, true))
2672		shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2673				   sc, LRU_ACTIVE_ANON);
2674}
2675
2676/* Use reclaim/compaction for costly allocs or under memory pressure */
2677static bool in_reclaim_compaction(struct scan_control *sc)
2678{
2679	if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2680			(sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2681			 sc->priority < DEF_PRIORITY - 2))
2682		return true;
2683
2684	return false;
2685}
2686
2687/*
2688 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2689 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2690 * true if more pages should be reclaimed such that when the page allocator
2691 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2692 * It will give up earlier than that if there is difficulty reclaiming pages.
2693 */
2694static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2695					unsigned long nr_reclaimed,
 
2696					struct scan_control *sc)
2697{
2698	unsigned long pages_for_compaction;
2699	unsigned long inactive_lru_pages;
2700	int z;
2701
2702	/* If not in reclaim/compaction mode, stop */
2703	if (!in_reclaim_compaction(sc))
2704		return false;
2705
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2706	/*
2707	 * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX
2708	 * number of pages that were scanned. This will return to the caller
2709	 * with the risk reclaim/compaction and the resulting allocation attempt
2710	 * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL
2711	 * allocations through requiring that the full LRU list has been scanned
2712	 * first, by assuming that zero delta of sc->nr_scanned means full LRU
2713	 * scan, but that approximation was wrong, and there were corner cases
2714	 * where always a non-zero amount of pages were scanned.
2715	 */
2716	if (!nr_reclaimed)
2717		return false;
 
 
 
 
 
2718
2719	/* If compaction would go ahead or the allocation would succeed, stop */
2720	for (z = 0; z <= sc->reclaim_idx; z++) {
2721		struct zone *zone = &pgdat->node_zones[z];
2722		if (!managed_zone(zone))
2723			continue;
2724
2725		switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2726		case COMPACT_SUCCESS:
2727		case COMPACT_CONTINUE:
2728			return false;
2729		default:
2730			/* check next zone */
2731			;
2732		}
2733	}
2734
2735	/*
2736	 * If we have not reclaimed enough pages for compaction and the
2737	 * inactive lists are large enough, continue reclaiming
2738	 */
2739	pages_for_compaction = compact_gap(sc->order);
2740	inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2741	if (get_nr_swap_pages() > 0)
2742		inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2743
2744	return inactive_lru_pages > pages_for_compaction;
2745}
2746
2747static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2748{
2749	return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2750		(memcg && memcg_congested(pgdat, memcg));
2751}
2752
2753static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2754{
2755	struct reclaim_state *reclaim_state = current->reclaim_state;
2756	unsigned long nr_reclaimed, nr_scanned;
2757	bool reclaimable = false;
2758
2759	do {
2760		struct mem_cgroup *root = sc->target_mem_cgroup;
 
 
 
 
2761		unsigned long node_lru_pages = 0;
2762		struct mem_cgroup *memcg;
2763
2764		memset(&sc->nr, 0, sizeof(sc->nr));
2765
2766		nr_reclaimed = sc->nr_reclaimed;
2767		nr_scanned = sc->nr_scanned;
2768
2769		memcg = mem_cgroup_iter(root, NULL, NULL);
2770		do {
2771			unsigned long lru_pages;
2772			unsigned long reclaimed;
2773			unsigned long scanned;
2774
2775			switch (mem_cgroup_protected(root, memcg)) {
2776			case MEMCG_PROT_MIN:
2777				/*
2778				 * Hard protection.
2779				 * If there is no reclaimable memory, OOM.
2780				 */
2781				continue;
2782			case MEMCG_PROT_LOW:
2783				/*
2784				 * Soft protection.
2785				 * Respect the protection only as long as
2786				 * there is an unprotected supply
2787				 * of reclaimable memory from other cgroups.
2788				 */
2789				if (!sc->memcg_low_reclaim) {
2790					sc->memcg_low_skipped = 1;
2791					continue;
2792				}
2793				memcg_memory_event(memcg, MEMCG_LOW);
2794				break;
2795			case MEMCG_PROT_NONE:
2796				/*
2797				 * All protection thresholds breached. We may
2798				 * still choose to vary the scan pressure
2799				 * applied based on by how much the cgroup in
2800				 * question has exceeded its protection
2801				 * thresholds (see get_scan_count).
2802				 */
2803				break;
2804			}
2805
2806			reclaimed = sc->nr_reclaimed;
2807			scanned = sc->nr_scanned;
2808			shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2809			node_lru_pages += lru_pages;
2810
2811			shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
2812					sc->priority);
 
2813
2814			/* Record the group's reclaim efficiency */
2815			vmpressure(sc->gfp_mask, memcg, false,
2816				   sc->nr_scanned - scanned,
2817				   sc->nr_reclaimed - reclaimed);
2818
2819		} while ((memcg = mem_cgroup_iter(root, memcg, NULL)));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2820
2821		if (reclaim_state) {
2822			sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2823			reclaim_state->reclaimed_slab = 0;
2824		}
2825
2826		/* Record the subtree's reclaim efficiency */
2827		vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2828			   sc->nr_scanned - nr_scanned,
2829			   sc->nr_reclaimed - nr_reclaimed);
2830
2831		if (sc->nr_reclaimed - nr_reclaimed)
2832			reclaimable = true;
2833
2834		if (current_is_kswapd()) {
2835			/*
2836			 * If reclaim is isolating dirty pages under writeback,
2837			 * it implies that the long-lived page allocation rate
2838			 * is exceeding the page laundering rate. Either the
2839			 * global limits are not being effective at throttling
2840			 * processes due to the page distribution throughout
2841			 * zones or there is heavy usage of a slow backing
2842			 * device. The only option is to throttle from reclaim
2843			 * context which is not ideal as there is no guarantee
2844			 * the dirtying process is throttled in the same way
2845			 * balance_dirty_pages() manages.
2846			 *
2847			 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2848			 * count the number of pages under pages flagged for
2849			 * immediate reclaim and stall if any are encountered
2850			 * in the nr_immediate check below.
2851			 */
2852			if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2853				set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2854
2855			/*
2856			 * Tag a node as congested if all the dirty pages
2857			 * scanned were backed by a congested BDI and
2858			 * wait_iff_congested will stall.
2859			 */
2860			if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2861				set_bit(PGDAT_CONGESTED, &pgdat->flags);
2862
2863			/* Allow kswapd to start writing pages during reclaim.*/
2864			if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2865				set_bit(PGDAT_DIRTY, &pgdat->flags);
2866
2867			/*
2868			 * If kswapd scans pages marked marked for immediate
2869			 * reclaim and under writeback (nr_immediate), it
2870			 * implies that pages are cycling through the LRU
2871			 * faster than they are written so also forcibly stall.
2872			 */
2873			if (sc->nr.immediate)
2874				congestion_wait(BLK_RW_ASYNC, HZ/10);
2875		}
2876
2877		/*
2878		 * Legacy memcg will stall in page writeback so avoid forcibly
2879		 * stalling in wait_iff_congested().
2880		 */
2881		if (!global_reclaim(sc) && sane_reclaim(sc) &&
2882		    sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2883			set_memcg_congestion(pgdat, root, true);
2884
2885		/*
2886		 * Stall direct reclaim for IO completions if underlying BDIs
2887		 * and node is congested. Allow kswapd to continue until it
2888		 * starts encountering unqueued dirty pages or cycling through
2889		 * the LRU too quickly.
2890		 */
2891		if (!sc->hibernation_mode && !current_is_kswapd() &&
2892		   current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2893			wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2894
2895	} while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2896					 sc));
2897
2898	/*
2899	 * Kswapd gives up on balancing particular nodes after too
2900	 * many failures to reclaim anything from them and goes to
2901	 * sleep. On reclaim progress, reset the failure counter. A
2902	 * successful direct reclaim run will revive a dormant kswapd.
2903	 */
2904	if (reclaimable)
2905		pgdat->kswapd_failures = 0;
2906
2907	return reclaimable;
2908}
2909
2910/*
2911 * Returns true if compaction should go ahead for a costly-order request, or
2912 * the allocation would already succeed without compaction. Return false if we
2913 * should reclaim first.
2914 */
2915static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2916{
2917	unsigned long watermark;
2918	enum compact_result suitable;
2919
2920	suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2921	if (suitable == COMPACT_SUCCESS)
2922		/* Allocation should succeed already. Don't reclaim. */
2923		return true;
2924	if (suitable == COMPACT_SKIPPED)
2925		/* Compaction cannot yet proceed. Do reclaim. */
2926		return false;
2927
2928	/*
2929	 * Compaction is already possible, but it takes time to run and there
2930	 * are potentially other callers using the pages just freed. So proceed
2931	 * with reclaim to make a buffer of free pages available to give
2932	 * compaction a reasonable chance of completing and allocating the page.
2933	 * Note that we won't actually reclaim the whole buffer in one attempt
2934	 * as the target watermark in should_continue_reclaim() is lower. But if
2935	 * we are already above the high+gap watermark, don't reclaim at all.
2936	 */
2937	watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2938
2939	return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2940}
2941
2942/*
2943 * This is the direct reclaim path, for page-allocating processes.  We only
2944 * try to reclaim pages from zones which will satisfy the caller's allocation
2945 * request.
2946 *
2947 * If a zone is deemed to be full of pinned pages then just give it a light
2948 * scan then give up on it.
2949 */
2950static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2951{
2952	struct zoneref *z;
2953	struct zone *zone;
2954	unsigned long nr_soft_reclaimed;
2955	unsigned long nr_soft_scanned;
2956	gfp_t orig_mask;
2957	pg_data_t *last_pgdat = NULL;
2958
2959	/*
2960	 * If the number of buffer_heads in the machine exceeds the maximum
2961	 * allowed level, force direct reclaim to scan the highmem zone as
2962	 * highmem pages could be pinning lowmem pages storing buffer_heads
2963	 */
2964	orig_mask = sc->gfp_mask;
2965	if (buffer_heads_over_limit) {
2966		sc->gfp_mask |= __GFP_HIGHMEM;
2967		sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2968	}
2969
2970	for_each_zone_zonelist_nodemask(zone, z, zonelist,
2971					sc->reclaim_idx, sc->nodemask) {
2972		/*
2973		 * Take care memory controller reclaiming has small influence
2974		 * to global LRU.
2975		 */
2976		if (global_reclaim(sc)) {
2977			if (!cpuset_zone_allowed(zone,
2978						 GFP_KERNEL | __GFP_HARDWALL))
2979				continue;
2980
2981			/*
2982			 * If we already have plenty of memory free for
2983			 * compaction in this zone, don't free any more.
2984			 * Even though compaction is invoked for any
2985			 * non-zero order, only frequent costly order
2986			 * reclamation is disruptive enough to become a
2987			 * noticeable problem, like transparent huge
2988			 * page allocations.
2989			 */
2990			if (IS_ENABLED(CONFIG_COMPACTION) &&
2991			    sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2992			    compaction_ready(zone, sc)) {
2993				sc->compaction_ready = true;
2994				continue;
2995			}
2996
2997			/*
2998			 * Shrink each node in the zonelist once. If the
2999			 * zonelist is ordered by zone (not the default) then a
3000			 * node may be shrunk multiple times but in that case
3001			 * the user prefers lower zones being preserved.
3002			 */
3003			if (zone->zone_pgdat == last_pgdat)
3004				continue;
3005
3006			/*
3007			 * This steals pages from memory cgroups over softlimit
3008			 * and returns the number of reclaimed pages and
3009			 * scanned pages. This works for global memory pressure
3010			 * and balancing, not for a memcg's limit.
3011			 */
3012			nr_soft_scanned = 0;
3013			nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
3014						sc->order, sc->gfp_mask,
3015						&nr_soft_scanned);
3016			sc->nr_reclaimed += nr_soft_reclaimed;
3017			sc->nr_scanned += nr_soft_scanned;
3018			/* need some check for avoid more shrink_zone() */
3019		}
3020
3021		/* See comment about same check for global reclaim above */
3022		if (zone->zone_pgdat == last_pgdat)
3023			continue;
3024		last_pgdat = zone->zone_pgdat;
3025		shrink_node(zone->zone_pgdat, sc);
3026	}
3027
3028	/*
3029	 * Restore to original mask to avoid the impact on the caller if we
3030	 * promoted it to __GFP_HIGHMEM.
3031	 */
3032	sc->gfp_mask = orig_mask;
3033}
3034
3035static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
3036{
3037	struct mem_cgroup *memcg;
3038
3039	memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
3040	do {
3041		unsigned long refaults;
3042		struct lruvec *lruvec;
3043
 
 
 
 
 
3044		lruvec = mem_cgroup_lruvec(pgdat, memcg);
3045		refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE);
3046		lruvec->refaults = refaults;
3047	} while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
3048}
3049
3050/*
3051 * This is the main entry point to direct page reclaim.
3052 *
3053 * If a full scan of the inactive list fails to free enough memory then we
3054 * are "out of memory" and something needs to be killed.
3055 *
3056 * If the caller is !__GFP_FS then the probability of a failure is reasonably
3057 * high - the zone may be full of dirty or under-writeback pages, which this
3058 * caller can't do much about.  We kick the writeback threads and take explicit
3059 * naps in the hope that some of these pages can be written.  But if the
3060 * allocating task holds filesystem locks which prevent writeout this might not
3061 * work, and the allocation attempt will fail.
3062 *
3063 * returns:	0, if no pages reclaimed
3064 * 		else, the number of pages reclaimed
3065 */
3066static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3067					  struct scan_control *sc)
3068{
3069	int initial_priority = sc->priority;
3070	pg_data_t *last_pgdat;
3071	struct zoneref *z;
3072	struct zone *zone;
3073retry:
3074	delayacct_freepages_start();
3075
3076	if (global_reclaim(sc))
3077		__count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
3078
3079	do {
3080		vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
3081				sc->priority);
3082		sc->nr_scanned = 0;
3083		shrink_zones(zonelist, sc);
3084
3085		if (sc->nr_reclaimed >= sc->nr_to_reclaim)
3086			break;
3087
3088		if (sc->compaction_ready)
3089			break;
3090
3091		/*
3092		 * If we're getting trouble reclaiming, start doing
3093		 * writepage even in laptop mode.
3094		 */
3095		if (sc->priority < DEF_PRIORITY - 2)
3096			sc->may_writepage = 1;
3097	} while (--sc->priority >= 0);
3098
3099	last_pgdat = NULL;
3100	for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
3101					sc->nodemask) {
3102		if (zone->zone_pgdat == last_pgdat)
3103			continue;
3104		last_pgdat = zone->zone_pgdat;
3105		snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
3106		set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
3107	}
3108
3109	delayacct_freepages_end();
3110
3111	if (sc->nr_reclaimed)
3112		return sc->nr_reclaimed;
3113
3114	/* Aborted reclaim to try compaction? don't OOM, then */
3115	if (sc->compaction_ready)
3116		return 1;
3117
3118	/* Untapped cgroup reserves?  Don't OOM, retry. */
3119	if (sc->memcg_low_skipped) {
3120		sc->priority = initial_priority;
3121		sc->memcg_low_reclaim = 1;
3122		sc->memcg_low_skipped = 0;
3123		goto retry;
3124	}
3125
3126	return 0;
3127}
3128
3129static bool allow_direct_reclaim(pg_data_t *pgdat)
3130{
3131	struct zone *zone;
3132	unsigned long pfmemalloc_reserve = 0;
3133	unsigned long free_pages = 0;
3134	int i;
3135	bool wmark_ok;
3136
3137	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3138		return true;
3139
3140	for (i = 0; i <= ZONE_NORMAL; i++) {
3141		zone = &pgdat->node_zones[i];
3142		if (!managed_zone(zone))
3143			continue;
3144
3145		if (!zone_reclaimable_pages(zone))
3146			continue;
3147
3148		pfmemalloc_reserve += min_wmark_pages(zone);
3149		free_pages += zone_page_state(zone, NR_FREE_PAGES);
3150	}
3151
3152	/* If there are no reserves (unexpected config) then do not throttle */
3153	if (!pfmemalloc_reserve)
3154		return true;
3155
3156	wmark_ok = free_pages > pfmemalloc_reserve / 2;
3157
3158	/* kswapd must be awake if processes are being throttled */
3159	if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
3160		pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
3161						(enum zone_type)ZONE_NORMAL);
3162		wake_up_interruptible(&pgdat->kswapd_wait);
3163	}
3164
3165	return wmark_ok;
3166}
3167
3168/*
3169 * Throttle direct reclaimers if backing storage is backed by the network
3170 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
3171 * depleted. kswapd will continue to make progress and wake the processes
3172 * when the low watermark is reached.
3173 *
3174 * Returns true if a fatal signal was delivered during throttling. If this
3175 * happens, the page allocator should not consider triggering the OOM killer.
3176 */
3177static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
3178					nodemask_t *nodemask)
3179{
3180	struct zoneref *z;
3181	struct zone *zone;
3182	pg_data_t *pgdat = NULL;
3183
3184	/*
3185	 * Kernel threads should not be throttled as they may be indirectly
3186	 * responsible for cleaning pages necessary for reclaim to make forward
3187	 * progress. kjournald for example may enter direct reclaim while
3188	 * committing a transaction where throttling it could forcing other
3189	 * processes to block on log_wait_commit().
3190	 */
3191	if (current->flags & PF_KTHREAD)
3192		goto out;
3193
3194	/*
3195	 * If a fatal signal is pending, this process should not throttle.
3196	 * It should return quickly so it can exit and free its memory
3197	 */
3198	if (fatal_signal_pending(current))
3199		goto out;
3200
3201	/*
3202	 * Check if the pfmemalloc reserves are ok by finding the first node
3203	 * with a usable ZONE_NORMAL or lower zone. The expectation is that
3204	 * GFP_KERNEL will be required for allocating network buffers when
3205	 * swapping over the network so ZONE_HIGHMEM is unusable.
3206	 *
3207	 * Throttling is based on the first usable node and throttled processes
3208	 * wait on a queue until kswapd makes progress and wakes them. There
3209	 * is an affinity then between processes waking up and where reclaim
3210	 * progress has been made assuming the process wakes on the same node.
3211	 * More importantly, processes running on remote nodes will not compete
3212	 * for remote pfmemalloc reserves and processes on different nodes
3213	 * should make reasonable progress.
3214	 */
3215	for_each_zone_zonelist_nodemask(zone, z, zonelist,
3216					gfp_zone(gfp_mask), nodemask) {
3217		if (zone_idx(zone) > ZONE_NORMAL)
3218			continue;
3219
3220		/* Throttle based on the first usable node */
3221		pgdat = zone->zone_pgdat;
3222		if (allow_direct_reclaim(pgdat))
3223			goto out;
3224		break;
3225	}
3226
3227	/* If no zone was usable by the allocation flags then do not throttle */
3228	if (!pgdat)
3229		goto out;
3230
3231	/* Account for the throttling */
3232	count_vm_event(PGSCAN_DIRECT_THROTTLE);
3233
3234	/*
3235	 * If the caller cannot enter the filesystem, it's possible that it
3236	 * is due to the caller holding an FS lock or performing a journal
3237	 * transaction in the case of a filesystem like ext[3|4]. In this case,
3238	 * it is not safe to block on pfmemalloc_wait as kswapd could be
3239	 * blocked waiting on the same lock. Instead, throttle for up to a
3240	 * second before continuing.
3241	 */
3242	if (!(gfp_mask & __GFP_FS)) {
3243		wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3244			allow_direct_reclaim(pgdat), HZ);
3245
3246		goto check_pending;
3247	}
3248
3249	/* Throttle until kswapd wakes the process */
3250	wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3251		allow_direct_reclaim(pgdat));
3252
3253check_pending:
3254	if (fatal_signal_pending(current))
3255		return true;
3256
3257out:
3258	return false;
3259}
3260
3261unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3262				gfp_t gfp_mask, nodemask_t *nodemask)
3263{
3264	unsigned long nr_reclaimed;
3265	struct scan_control sc = {
3266		.nr_to_reclaim = SWAP_CLUSTER_MAX,
3267		.gfp_mask = current_gfp_context(gfp_mask),
3268		.reclaim_idx = gfp_zone(gfp_mask),
3269		.order = order,
3270		.nodemask = nodemask,
3271		.priority = DEF_PRIORITY,
3272		.may_writepage = !laptop_mode,
3273		.may_unmap = 1,
3274		.may_swap = 1,
3275	};
3276
3277	/*
3278	 * scan_control uses s8 fields for order, priority, and reclaim_idx.
3279	 * Confirm they are large enough for max values.
3280	 */
3281	BUILD_BUG_ON(MAX_ORDER > S8_MAX);
3282	BUILD_BUG_ON(DEF_PRIORITY > S8_MAX);
3283	BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX);
3284
3285	/*
3286	 * Do not enter reclaim if fatal signal was delivered while throttled.
3287	 * 1 is returned so that the page allocator does not OOM kill at this
3288	 * point.
3289	 */
3290	if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3291		return 1;
3292
3293	set_task_reclaim_state(current, &sc.reclaim_state);
3294	trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask);
 
 
3295
3296	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3297
3298	trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
3299	set_task_reclaim_state(current, NULL);
3300
3301	return nr_reclaimed;
3302}
3303
3304#ifdef CONFIG_MEMCG
3305
3306/* Only used by soft limit reclaim. Do not reuse for anything else. */
3307unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3308						gfp_t gfp_mask, bool noswap,
3309						pg_data_t *pgdat,
3310						unsigned long *nr_scanned)
3311{
3312	struct scan_control sc = {
3313		.nr_to_reclaim = SWAP_CLUSTER_MAX,
3314		.target_mem_cgroup = memcg,
3315		.may_writepage = !laptop_mode,
3316		.may_unmap = 1,
3317		.reclaim_idx = MAX_NR_ZONES - 1,
3318		.may_swap = !noswap,
3319	};
3320	unsigned long lru_pages;
3321
3322	WARN_ON_ONCE(!current->reclaim_state);
3323
3324	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3325			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3326
3327	trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3328						      sc.gfp_mask);
 
 
3329
3330	/*
3331	 * NOTE: Although we can get the priority field, using it
3332	 * here is not a good idea, since it limits the pages we can scan.
3333	 * if we don't reclaim here, the shrink_node from balance_pgdat
3334	 * will pick up pages from other mem cgroup's as well. We hack
3335	 * the priority and make it zero.
3336	 */
3337	shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3338
3339	trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3340
3341	*nr_scanned = sc.nr_scanned;
3342
3343	return sc.nr_reclaimed;
3344}
3345
3346unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3347					   unsigned long nr_pages,
3348					   gfp_t gfp_mask,
3349					   bool may_swap)
3350{
3351	struct zonelist *zonelist;
3352	unsigned long nr_reclaimed;
3353	unsigned long pflags;
3354	int nid;
3355	unsigned int noreclaim_flag;
3356	struct scan_control sc = {
3357		.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3358		.gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3359				(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3360		.reclaim_idx = MAX_NR_ZONES - 1,
3361		.target_mem_cgroup = memcg,
3362		.priority = DEF_PRIORITY,
3363		.may_writepage = !laptop_mode,
3364		.may_unmap = 1,
3365		.may_swap = may_swap,
3366	};
3367
3368	set_task_reclaim_state(current, &sc.reclaim_state);
3369	/*
3370	 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3371	 * take care of from where we get pages. So the node where we start the
3372	 * scan does not need to be the current node.
3373	 */
3374	nid = mem_cgroup_select_victim_node(memcg);
3375
3376	zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3377
3378	trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask);
 
 
 
3379
3380	psi_memstall_enter(&pflags);
3381	noreclaim_flag = memalloc_noreclaim_save();
3382
3383	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3384
3385	memalloc_noreclaim_restore(noreclaim_flag);
3386	psi_memstall_leave(&pflags);
3387
3388	trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3389	set_task_reclaim_state(current, NULL);
3390
3391	return nr_reclaimed;
3392}
3393#endif
3394
3395static void age_active_anon(struct pglist_data *pgdat,
3396				struct scan_control *sc)
3397{
3398	struct mem_cgroup *memcg;
3399
3400	if (!total_swap_pages)
3401		return;
3402
3403	memcg = mem_cgroup_iter(NULL, NULL, NULL);
3404	do {
3405		struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3406
3407		if (inactive_list_is_low(lruvec, false, sc, true))
3408			shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3409					   sc, LRU_ACTIVE_ANON);
3410
3411		memcg = mem_cgroup_iter(NULL, memcg, NULL);
3412	} while (memcg);
3413}
3414
3415static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx)
3416{
3417	int i;
3418	struct zone *zone;
3419
3420	/*
3421	 * Check for watermark boosts top-down as the higher zones
3422	 * are more likely to be boosted. Both watermarks and boosts
3423	 * should not be checked at the time time as reclaim would
3424	 * start prematurely when there is no boosting and a lower
3425	 * zone is balanced.
3426	 */
3427	for (i = classzone_idx; i >= 0; i--) {
3428		zone = pgdat->node_zones + i;
3429		if (!managed_zone(zone))
3430			continue;
3431
3432		if (zone->watermark_boost)
3433			return true;
3434	}
3435
3436	return false;
3437}
3438
3439/*
3440 * Returns true if there is an eligible zone balanced for the request order
3441 * and classzone_idx
3442 */
3443static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3444{
3445	int i;
3446	unsigned long mark = -1;
3447	struct zone *zone;
3448
3449	/*
3450	 * Check watermarks bottom-up as lower zones are more likely to
3451	 * meet watermarks.
3452	 */
3453	for (i = 0; i <= classzone_idx; i++) {
3454		zone = pgdat->node_zones + i;
3455
3456		if (!managed_zone(zone))
3457			continue;
3458
3459		mark = high_wmark_pages(zone);
3460		if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3461			return true;
3462	}
3463
3464	/*
3465	 * If a node has no populated zone within classzone_idx, it does not
3466	 * need balancing by definition. This can happen if a zone-restricted
3467	 * allocation tries to wake a remote kswapd.
3468	 */
3469	if (mark == -1)
3470		return true;
3471
3472	return false;
3473}
3474
3475/* Clear pgdat state for congested, dirty or under writeback. */
3476static void clear_pgdat_congested(pg_data_t *pgdat)
3477{
3478	clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3479	clear_bit(PGDAT_DIRTY, &pgdat->flags);
3480	clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3481}
3482
3483/*
3484 * Prepare kswapd for sleeping. This verifies that there are no processes
3485 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3486 *
3487 * Returns true if kswapd is ready to sleep
3488 */
3489static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3490{
3491	/*
3492	 * The throttled processes are normally woken up in balance_pgdat() as
3493	 * soon as allow_direct_reclaim() is true. But there is a potential
3494	 * race between when kswapd checks the watermarks and a process gets
3495	 * throttled. There is also a potential race if processes get
3496	 * throttled, kswapd wakes, a large process exits thereby balancing the
3497	 * zones, which causes kswapd to exit balance_pgdat() before reaching
3498	 * the wake up checks. If kswapd is going to sleep, no process should
3499	 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3500	 * the wake up is premature, processes will wake kswapd and get
3501	 * throttled again. The difference from wake ups in balance_pgdat() is
3502	 * that here we are under prepare_to_wait().
3503	 */
3504	if (waitqueue_active(&pgdat->pfmemalloc_wait))
3505		wake_up_all(&pgdat->pfmemalloc_wait);
3506
3507	/* Hopeless node, leave it to direct reclaim */
3508	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3509		return true;
3510
3511	if (pgdat_balanced(pgdat, order, classzone_idx)) {
3512		clear_pgdat_congested(pgdat);
3513		return true;
3514	}
3515
3516	return false;
3517}
3518
3519/*
3520 * kswapd shrinks a node of pages that are at or below the highest usable
3521 * zone that is currently unbalanced.
3522 *
3523 * Returns true if kswapd scanned at least the requested number of pages to
3524 * reclaim or if the lack of progress was due to pages under writeback.
3525 * This is used to determine if the scanning priority needs to be raised.
3526 */
3527static bool kswapd_shrink_node(pg_data_t *pgdat,
3528			       struct scan_control *sc)
3529{
3530	struct zone *zone;
3531	int z;
3532
3533	/* Reclaim a number of pages proportional to the number of zones */
3534	sc->nr_to_reclaim = 0;
3535	for (z = 0; z <= sc->reclaim_idx; z++) {
3536		zone = pgdat->node_zones + z;
3537		if (!managed_zone(zone))
3538			continue;
3539
3540		sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3541	}
3542
3543	/*
3544	 * Historically care was taken to put equal pressure on all zones but
3545	 * now pressure is applied based on node LRU order.
3546	 */
3547	shrink_node(pgdat, sc);
3548
3549	/*
3550	 * Fragmentation may mean that the system cannot be rebalanced for
3551	 * high-order allocations. If twice the allocation size has been
3552	 * reclaimed then recheck watermarks only at order-0 to prevent
3553	 * excessive reclaim. Assume that a process requested a high-order
3554	 * can direct reclaim/compact.
3555	 */
3556	if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3557		sc->order = 0;
3558
3559	return sc->nr_scanned >= sc->nr_to_reclaim;
3560}
3561
3562/*
3563 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3564 * that are eligible for use by the caller until at least one zone is
3565 * balanced.
3566 *
3567 * Returns the order kswapd finished reclaiming at.
3568 *
3569 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
3570 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3571 * found to have free_pages <= high_wmark_pages(zone), any page in that zone
3572 * or lower is eligible for reclaim until at least one usable zone is
3573 * balanced.
3574 */
3575static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3576{
3577	int i;
3578	unsigned long nr_soft_reclaimed;
3579	unsigned long nr_soft_scanned;
3580	unsigned long pflags;
3581	unsigned long nr_boost_reclaim;
3582	unsigned long zone_boosts[MAX_NR_ZONES] = { 0, };
3583	bool boosted;
3584	struct zone *zone;
3585	struct scan_control sc = {
3586		.gfp_mask = GFP_KERNEL,
3587		.order = order,
 
 
3588		.may_unmap = 1,
 
3589	};
3590
3591	set_task_reclaim_state(current, &sc.reclaim_state);
3592	psi_memstall_enter(&pflags);
3593	__fs_reclaim_acquire();
3594
3595	count_vm_event(PAGEOUTRUN);
3596
3597	/*
3598	 * Account for the reclaim boost. Note that the zone boost is left in
3599	 * place so that parallel allocations that are near the watermark will
3600	 * stall or direct reclaim until kswapd is finished.
3601	 */
3602	nr_boost_reclaim = 0;
3603	for (i = 0; i <= classzone_idx; i++) {
3604		zone = pgdat->node_zones + i;
3605		if (!managed_zone(zone))
3606			continue;
3607
3608		nr_boost_reclaim += zone->watermark_boost;
3609		zone_boosts[i] = zone->watermark_boost;
3610	}
3611	boosted = nr_boost_reclaim;
3612
3613restart:
3614	sc.priority = DEF_PRIORITY;
3615	do {
3616		unsigned long nr_reclaimed = sc.nr_reclaimed;
3617		bool raise_priority = true;
3618		bool balanced;
3619		bool ret;
3620
3621		sc.reclaim_idx = classzone_idx;
3622
3623		/*
3624		 * If the number of buffer_heads exceeds the maximum allowed
3625		 * then consider reclaiming from all zones. This has a dual
3626		 * purpose -- on 64-bit systems it is expected that
3627		 * buffer_heads are stripped during active rotation. On 32-bit
3628		 * systems, highmem pages can pin lowmem memory and shrinking
3629		 * buffers can relieve lowmem pressure. Reclaim may still not
3630		 * go ahead if all eligible zones for the original allocation
3631		 * request are balanced to avoid excessive reclaim from kswapd.
3632		 */
3633		if (buffer_heads_over_limit) {
3634			for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3635				zone = pgdat->node_zones + i;
3636				if (!managed_zone(zone))
3637					continue;
3638
3639				sc.reclaim_idx = i;
3640				break;
3641			}
3642		}
3643
3644		/*
3645		 * If the pgdat is imbalanced then ignore boosting and preserve
3646		 * the watermarks for a later time and restart. Note that the
3647		 * zone watermarks will be still reset at the end of balancing
3648		 * on the grounds that the normal reclaim should be enough to
3649		 * re-evaluate if boosting is required when kswapd next wakes.
3650		 */
3651		balanced = pgdat_balanced(pgdat, sc.order, classzone_idx);
3652		if (!balanced && nr_boost_reclaim) {
3653			nr_boost_reclaim = 0;
3654			goto restart;
3655		}
3656
3657		/*
3658		 * If boosting is not active then only reclaim if there are no
3659		 * eligible zones. Note that sc.reclaim_idx is not used as
3660		 * buffer_heads_over_limit may have adjusted it.
3661		 */
3662		if (!nr_boost_reclaim && balanced)
3663			goto out;
3664
3665		/* Limit the priority of boosting to avoid reclaim writeback */
3666		if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2)
3667			raise_priority = false;
3668
3669		/*
3670		 * Do not writeback or swap pages for boosted reclaim. The
3671		 * intent is to relieve pressure not issue sub-optimal IO
3672		 * from reclaim context. If no pages are reclaimed, the
3673		 * reclaim will be aborted.
3674		 */
3675		sc.may_writepage = !laptop_mode && !nr_boost_reclaim;
3676		sc.may_swap = !nr_boost_reclaim;
3677
3678		/*
3679		 * Do some background aging of the anon list, to give
3680		 * pages a chance to be referenced before reclaiming. All
3681		 * pages are rotated regardless of classzone as this is
3682		 * about consistent aging.
3683		 */
3684		age_active_anon(pgdat, &sc);
3685
3686		/*
3687		 * If we're getting trouble reclaiming, start doing writepage
3688		 * even in laptop mode.
3689		 */
3690		if (sc.priority < DEF_PRIORITY - 2)
3691			sc.may_writepage = 1;
3692
3693		/* Call soft limit reclaim before calling shrink_node. */
3694		sc.nr_scanned = 0;
3695		nr_soft_scanned = 0;
3696		nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3697						sc.gfp_mask, &nr_soft_scanned);
3698		sc.nr_reclaimed += nr_soft_reclaimed;
3699
3700		/*
3701		 * There should be no need to raise the scanning priority if
3702		 * enough pages are already being scanned that that high
3703		 * watermark would be met at 100% efficiency.
3704		 */
3705		if (kswapd_shrink_node(pgdat, &sc))
3706			raise_priority = false;
3707
3708		/*
3709		 * If the low watermark is met there is no need for processes
3710		 * to be throttled on pfmemalloc_wait as they should not be
3711		 * able to safely make forward progress. Wake them
3712		 */
3713		if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3714				allow_direct_reclaim(pgdat))
3715			wake_up_all(&pgdat->pfmemalloc_wait);
3716
3717		/* Check if kswapd should be suspending */
3718		__fs_reclaim_release();
3719		ret = try_to_freeze();
3720		__fs_reclaim_acquire();
3721		if (ret || kthread_should_stop())
3722			break;
3723
3724		/*
3725		 * Raise priority if scanning rate is too low or there was no
3726		 * progress in reclaiming pages
3727		 */
3728		nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
3729		nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed);
3730
3731		/*
3732		 * If reclaim made no progress for a boost, stop reclaim as
3733		 * IO cannot be queued and it could be an infinite loop in
3734		 * extreme circumstances.
3735		 */
3736		if (nr_boost_reclaim && !nr_reclaimed)
3737			break;
3738
3739		if (raise_priority || !nr_reclaimed)
3740			sc.priority--;
3741	} while (sc.priority >= 1);
3742
3743	if (!sc.nr_reclaimed)
3744		pgdat->kswapd_failures++;
3745
3746out:
3747	/* If reclaim was boosted, account for the reclaim done in this pass */
3748	if (boosted) {
3749		unsigned long flags;
3750
3751		for (i = 0; i <= classzone_idx; i++) {
3752			if (!zone_boosts[i])
3753				continue;
3754
3755			/* Increments are under the zone lock */
3756			zone = pgdat->node_zones + i;
3757			spin_lock_irqsave(&zone->lock, flags);
3758			zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]);
3759			spin_unlock_irqrestore(&zone->lock, flags);
3760		}
3761
3762		/*
3763		 * As there is now likely space, wakeup kcompact to defragment
3764		 * pageblocks.
3765		 */
3766		wakeup_kcompactd(pgdat, pageblock_order, classzone_idx);
3767	}
3768
3769	snapshot_refaults(NULL, pgdat);
3770	__fs_reclaim_release();
3771	psi_memstall_leave(&pflags);
3772	set_task_reclaim_state(current, NULL);
3773
3774	/*
3775	 * Return the order kswapd stopped reclaiming at as
3776	 * prepare_kswapd_sleep() takes it into account. If another caller
3777	 * entered the allocator slow path while kswapd was awake, order will
3778	 * remain at the higher level.
3779	 */
3780	return sc.order;
3781}
3782
3783/*
3784 * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be
3785 * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not
3786 * a valid index then either kswapd runs for first time or kswapd couldn't sleep
3787 * after previous reclaim attempt (node is still unbalanced). In that case
3788 * return the zone index of the previous kswapd reclaim cycle.
3789 */
3790static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3791					   enum zone_type prev_classzone_idx)
3792{
3793	if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3794		return prev_classzone_idx;
3795	return pgdat->kswapd_classzone_idx;
 
3796}
3797
3798static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3799				unsigned int classzone_idx)
3800{
3801	long remaining = 0;
3802	DEFINE_WAIT(wait);
3803
3804	if (freezing(current) || kthread_should_stop())
3805		return;
3806
3807	prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3808
3809	/*
3810	 * Try to sleep for a short interval. Note that kcompactd will only be
3811	 * woken if it is possible to sleep for a short interval. This is
3812	 * deliberate on the assumption that if reclaim cannot keep an
3813	 * eligible zone balanced that it's also unlikely that compaction will
3814	 * succeed.
3815	 */
3816	if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3817		/*
3818		 * Compaction records what page blocks it recently failed to
3819		 * isolate pages from and skips them in the future scanning.
3820		 * When kswapd is going to sleep, it is reasonable to assume
3821		 * that pages and compaction may succeed so reset the cache.
3822		 */
3823		reset_isolation_suitable(pgdat);
3824
3825		/*
3826		 * We have freed the memory, now we should compact it to make
3827		 * allocation of the requested order possible.
3828		 */
3829		wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3830
3831		remaining = schedule_timeout(HZ/10);
3832
3833		/*
3834		 * If woken prematurely then reset kswapd_classzone_idx and
3835		 * order. The values will either be from a wakeup request or
3836		 * the previous request that slept prematurely.
3837		 */
3838		if (remaining) {
3839			pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3840			pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3841		}
3842
3843		finish_wait(&pgdat->kswapd_wait, &wait);
3844		prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3845	}
3846
3847	/*
3848	 * After a short sleep, check if it was a premature sleep. If not, then
3849	 * go fully to sleep until explicitly woken up.
3850	 */
3851	if (!remaining &&
3852	    prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3853		trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3854
3855		/*
3856		 * vmstat counters are not perfectly accurate and the estimated
3857		 * value for counters such as NR_FREE_PAGES can deviate from the
3858		 * true value by nr_online_cpus * threshold. To avoid the zone
3859		 * watermarks being breached while under pressure, we reduce the
3860		 * per-cpu vmstat threshold while kswapd is awake and restore
3861		 * them before going back to sleep.
3862		 */
3863		set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3864
3865		if (!kthread_should_stop())
3866			schedule();
3867
3868		set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3869	} else {
3870		if (remaining)
3871			count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3872		else
3873			count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3874	}
3875	finish_wait(&pgdat->kswapd_wait, &wait);
3876}
3877
3878/*
3879 * The background pageout daemon, started as a kernel thread
3880 * from the init process.
3881 *
3882 * This basically trickles out pages so that we have _some_
3883 * free memory available even if there is no other activity
3884 * that frees anything up. This is needed for things like routing
3885 * etc, where we otherwise might have all activity going on in
3886 * asynchronous contexts that cannot page things out.
3887 *
3888 * If there are applications that are active memory-allocators
3889 * (most normal use), this basically shouldn't matter.
3890 */
3891static int kswapd(void *p)
3892{
3893	unsigned int alloc_order, reclaim_order;
3894	unsigned int classzone_idx = MAX_NR_ZONES - 1;
3895	pg_data_t *pgdat = (pg_data_t*)p;
3896	struct task_struct *tsk = current;
 
 
 
 
3897	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3898
3899	if (!cpumask_empty(cpumask))
3900		set_cpus_allowed_ptr(tsk, cpumask);
 
3901
3902	/*
3903	 * Tell the memory management that we're a "memory allocator",
3904	 * and that if we need more memory we should get access to it
3905	 * regardless (see "__alloc_pages()"). "kswapd" should
3906	 * never get caught in the normal page freeing logic.
3907	 *
3908	 * (Kswapd normally doesn't need memory anyway, but sometimes
3909	 * you need a small amount of memory in order to be able to
3910	 * page out something else, and this flag essentially protects
3911	 * us from recursively trying to free more memory as we're
3912	 * trying to free the first piece of memory in the first place).
3913	 */
3914	tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3915	set_freezable();
3916
3917	pgdat->kswapd_order = 0;
3918	pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3919	for ( ; ; ) {
3920		bool ret;
3921
3922		alloc_order = reclaim_order = pgdat->kswapd_order;
3923		classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3924
3925kswapd_try_sleep:
3926		kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3927					classzone_idx);
3928
3929		/* Read the new order and classzone_idx */
3930		alloc_order = reclaim_order = pgdat->kswapd_order;
3931		classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3932		pgdat->kswapd_order = 0;
3933		pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3934
3935		ret = try_to_freeze();
3936		if (kthread_should_stop())
3937			break;
3938
3939		/*
3940		 * We can speed up thawing tasks if we don't call balance_pgdat
3941		 * after returning from the refrigerator
3942		 */
3943		if (ret)
3944			continue;
3945
3946		/*
3947		 * Reclaim begins at the requested order but if a high-order
3948		 * reclaim fails then kswapd falls back to reclaiming for
3949		 * order-0. If that happens, kswapd will consider sleeping
3950		 * for the order it finished reclaiming at (reclaim_order)
3951		 * but kcompactd is woken to compact for the original
3952		 * request (alloc_order).
3953		 */
3954		trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3955						alloc_order);
 
3956		reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
 
3957		if (reclaim_order < alloc_order)
3958			goto kswapd_try_sleep;
3959	}
3960
3961	tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
 
3962
3963	return 0;
3964}
3965
3966/*
3967 * A zone is low on free memory or too fragmented for high-order memory.  If
3968 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3969 * pgdat.  It will wake up kcompactd after reclaiming memory.  If kswapd reclaim
3970 * has failed or is not needed, still wake up kcompactd if only compaction is
3971 * needed.
3972 */
3973void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3974		   enum zone_type classzone_idx)
3975{
3976	pg_data_t *pgdat;
3977
3978	if (!managed_zone(zone))
3979		return;
3980
3981	if (!cpuset_zone_allowed(zone, gfp_flags))
3982		return;
3983	pgdat = zone->zone_pgdat;
3984
3985	if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3986		pgdat->kswapd_classzone_idx = classzone_idx;
3987	else
3988		pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
3989						  classzone_idx);
3990	pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3991	if (!waitqueue_active(&pgdat->kswapd_wait))
3992		return;
3993
3994	/* Hopeless node, leave it to direct reclaim if possible */
3995	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3996	    (pgdat_balanced(pgdat, order, classzone_idx) &&
3997	     !pgdat_watermark_boosted(pgdat, classzone_idx))) {
3998		/*
3999		 * There may be plenty of free memory available, but it's too
4000		 * fragmented for high-order allocations.  Wake up kcompactd
4001		 * and rely on compaction_suitable() to determine if it's
4002		 * needed.  If it fails, it will defer subsequent attempts to
4003		 * ratelimit its work.
4004		 */
4005		if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
4006			wakeup_kcompactd(pgdat, order, classzone_idx);
4007		return;
4008	}
4009
4010	trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
4011				      gfp_flags);
4012	wake_up_interruptible(&pgdat->kswapd_wait);
4013}
4014
4015#ifdef CONFIG_HIBERNATION
4016/*
4017 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
4018 * freed pages.
4019 *
4020 * Rather than trying to age LRUs the aim is to preserve the overall
4021 * LRU order by reclaiming preferentially
4022 * inactive > active > active referenced > active mapped
4023 */
4024unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
4025{
 
4026	struct scan_control sc = {
4027		.nr_to_reclaim = nr_to_reclaim,
4028		.gfp_mask = GFP_HIGHUSER_MOVABLE,
4029		.reclaim_idx = MAX_NR_ZONES - 1,
4030		.priority = DEF_PRIORITY,
4031		.may_writepage = 1,
4032		.may_unmap = 1,
4033		.may_swap = 1,
4034		.hibernation_mode = 1,
4035	};
4036	struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
 
4037	unsigned long nr_reclaimed;
4038	unsigned int noreclaim_flag;
4039
4040	fs_reclaim_acquire(sc.gfp_mask);
4041	noreclaim_flag = memalloc_noreclaim_save();
4042	set_task_reclaim_state(current, &sc.reclaim_state);
 
 
4043
4044	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
4045
4046	set_task_reclaim_state(current, NULL);
4047	memalloc_noreclaim_restore(noreclaim_flag);
4048	fs_reclaim_release(sc.gfp_mask);
 
4049
4050	return nr_reclaimed;
4051}
4052#endif /* CONFIG_HIBERNATION */
4053
4054/* It's optimal to keep kswapds on the same CPUs as their memory, but
4055   not required for correctness.  So if the last cpu in a node goes
4056   away, we get changed to run anywhere: as the first one comes back,
4057   restore their cpu bindings. */
4058static int kswapd_cpu_online(unsigned int cpu)
4059{
4060	int nid;
4061
4062	for_each_node_state(nid, N_MEMORY) {
4063		pg_data_t *pgdat = NODE_DATA(nid);
4064		const struct cpumask *mask;
4065
4066		mask = cpumask_of_node(pgdat->node_id);
4067
4068		if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
4069			/* One of our CPUs online: restore mask */
4070			set_cpus_allowed_ptr(pgdat->kswapd, mask);
4071	}
4072	return 0;
4073}
4074
4075/*
4076 * This kswapd start function will be called by init and node-hot-add.
4077 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
4078 */
4079int kswapd_run(int nid)
4080{
4081	pg_data_t *pgdat = NODE_DATA(nid);
4082	int ret = 0;
4083
4084	if (pgdat->kswapd)
4085		return 0;
4086
4087	pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
4088	if (IS_ERR(pgdat->kswapd)) {
4089		/* failure at boot is fatal */
4090		BUG_ON(system_state < SYSTEM_RUNNING);
4091		pr_err("Failed to start kswapd on node %d\n", nid);
4092		ret = PTR_ERR(pgdat->kswapd);
4093		pgdat->kswapd = NULL;
4094	}
4095	return ret;
4096}
4097
4098/*
4099 * Called by memory hotplug when all memory in a node is offlined.  Caller must
4100 * hold mem_hotplug_begin/end().
4101 */
4102void kswapd_stop(int nid)
4103{
4104	struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
4105
4106	if (kswapd) {
4107		kthread_stop(kswapd);
4108		NODE_DATA(nid)->kswapd = NULL;
4109	}
4110}
4111
4112static int __init kswapd_init(void)
4113{
4114	int nid, ret;
4115
4116	swap_setup();
4117	for_each_node_state(nid, N_MEMORY)
4118 		kswapd_run(nid);
4119	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
4120					"mm/vmscan:online", kswapd_cpu_online,
4121					NULL);
4122	WARN_ON(ret < 0);
4123	return 0;
4124}
4125
4126module_init(kswapd_init)
4127
4128#ifdef CONFIG_NUMA
4129/*
4130 * Node reclaim mode
4131 *
4132 * If non-zero call node_reclaim when the number of free pages falls below
4133 * the watermarks.
4134 */
4135int node_reclaim_mode __read_mostly;
4136
4137#define RECLAIM_OFF 0
4138#define RECLAIM_ZONE (1<<0)	/* Run shrink_inactive_list on the zone */
4139#define RECLAIM_WRITE (1<<1)	/* Writeout pages during reclaim */
4140#define RECLAIM_UNMAP (1<<2)	/* Unmap pages during reclaim */
4141
4142/*
4143 * Priority for NODE_RECLAIM. This determines the fraction of pages
4144 * of a node considered for each zone_reclaim. 4 scans 1/16th of
4145 * a zone.
4146 */
4147#define NODE_RECLAIM_PRIORITY 4
4148
4149/*
4150 * Percentage of pages in a zone that must be unmapped for node_reclaim to
4151 * occur.
4152 */
4153int sysctl_min_unmapped_ratio = 1;
4154
4155/*
4156 * If the number of slab pages in a zone grows beyond this percentage then
4157 * slab reclaim needs to occur.
4158 */
4159int sysctl_min_slab_ratio = 5;
4160
4161static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
4162{
4163	unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
4164	unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
4165		node_page_state(pgdat, NR_ACTIVE_FILE);
4166
4167	/*
4168	 * It's possible for there to be more file mapped pages than
4169	 * accounted for by the pages on the file LRU lists because
4170	 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
4171	 */
4172	return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
4173}
4174
4175/* Work out how many page cache pages we can reclaim in this reclaim_mode */
4176static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
4177{
4178	unsigned long nr_pagecache_reclaimable;
4179	unsigned long delta = 0;
4180
4181	/*
4182	 * If RECLAIM_UNMAP is set, then all file pages are considered
4183	 * potentially reclaimable. Otherwise, we have to worry about
4184	 * pages like swapcache and node_unmapped_file_pages() provides
4185	 * a better estimate
4186	 */
4187	if (node_reclaim_mode & RECLAIM_UNMAP)
4188		nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
4189	else
4190		nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
4191
4192	/* If we can't clean pages, remove dirty pages from consideration */
4193	if (!(node_reclaim_mode & RECLAIM_WRITE))
4194		delta += node_page_state(pgdat, NR_FILE_DIRTY);
4195
4196	/* Watch for any possible underflows due to delta */
4197	if (unlikely(delta > nr_pagecache_reclaimable))
4198		delta = nr_pagecache_reclaimable;
4199
4200	return nr_pagecache_reclaimable - delta;
4201}
4202
4203/*
4204 * Try to free up some pages from this node through reclaim.
4205 */
4206static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4207{
4208	/* Minimum pages needed in order to stay on node */
4209	const unsigned long nr_pages = 1 << order;
4210	struct task_struct *p = current;
 
4211	unsigned int noreclaim_flag;
4212	struct scan_control sc = {
4213		.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
4214		.gfp_mask = current_gfp_context(gfp_mask),
4215		.order = order,
4216		.priority = NODE_RECLAIM_PRIORITY,
4217		.may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
4218		.may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
4219		.may_swap = 1,
4220		.reclaim_idx = gfp_zone(gfp_mask),
4221	};
4222
4223	trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
4224					   sc.gfp_mask);
4225
4226	cond_resched();
4227	fs_reclaim_acquire(sc.gfp_mask);
4228	/*
4229	 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
4230	 * and we also need to be able to write out pages for RECLAIM_WRITE
4231	 * and RECLAIM_UNMAP.
4232	 */
4233	noreclaim_flag = memalloc_noreclaim_save();
4234	p->flags |= PF_SWAPWRITE;
4235	set_task_reclaim_state(p, &sc.reclaim_state);
 
 
4236
4237	if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
4238		/*
4239		 * Free memory by calling shrink node with increasing
4240		 * priorities until we have enough memory freed.
4241		 */
4242		do {
4243			shrink_node(pgdat, &sc);
4244		} while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
4245	}
4246
4247	set_task_reclaim_state(p, NULL);
 
4248	current->flags &= ~PF_SWAPWRITE;
4249	memalloc_noreclaim_restore(noreclaim_flag);
4250	fs_reclaim_release(sc.gfp_mask);
4251
4252	trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed);
4253
4254	return sc.nr_reclaimed >= nr_pages;
4255}
4256
4257int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
4258{
4259	int ret;
4260
4261	/*
4262	 * Node reclaim reclaims unmapped file backed pages and
4263	 * slab pages if we are over the defined limits.
4264	 *
4265	 * A small portion of unmapped file backed pages is needed for
4266	 * file I/O otherwise pages read by file I/O will be immediately
4267	 * thrown out if the node is overallocated. So we do not reclaim
4268	 * if less than a specified percentage of the node is used by
4269	 * unmapped file backed pages.
4270	 */
4271	if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
4272	    node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
4273		return NODE_RECLAIM_FULL;
4274
4275	/*
4276	 * Do not scan if the allocation should not be delayed.
4277	 */
4278	if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
4279		return NODE_RECLAIM_NOSCAN;
4280
4281	/*
4282	 * Only run node reclaim on the local node or on nodes that do not
4283	 * have associated processors. This will favor the local processor
4284	 * over remote processors and spread off node memory allocations
4285	 * as wide as possible.
4286	 */
4287	if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
4288		return NODE_RECLAIM_NOSCAN;
4289
4290	if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
4291		return NODE_RECLAIM_NOSCAN;
4292
4293	ret = __node_reclaim(pgdat, gfp_mask, order);
4294	clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
4295
4296	if (!ret)
4297		count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
4298
4299	return ret;
4300}
4301#endif
4302
4303/*
4304 * page_evictable - test whether a page is evictable
4305 * @page: the page to test
4306 *
4307 * Test whether page is evictable--i.e., should be placed on active/inactive
4308 * lists vs unevictable list.
4309 *
4310 * Reasons page might not be evictable:
4311 * (1) page's mapping marked unevictable
4312 * (2) page is part of an mlocked VMA
4313 *
4314 */
4315int page_evictable(struct page *page)
4316{
4317	int ret;
4318
4319	/* Prevent address_space of inode and swap cache from being freed */
4320	rcu_read_lock();
4321	ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
4322	rcu_read_unlock();
4323	return ret;
4324}
4325
 
4326/**
4327 * check_move_unevictable_pages - check pages for evictability and move to
4328 * appropriate zone lru list
4329 * @pvec: pagevec with lru pages to check
4330 *
4331 * Checks pages for evictability, if an evictable page is in the unevictable
4332 * lru list, moves it to the appropriate evictable lru list. This function
4333 * should be only used for lru pages.
4334 */
4335void check_move_unevictable_pages(struct pagevec *pvec)
4336{
4337	struct lruvec *lruvec;
4338	struct pglist_data *pgdat = NULL;
4339	int pgscanned = 0;
4340	int pgrescued = 0;
4341	int i;
4342
4343	for (i = 0; i < pvec->nr; i++) {
4344		struct page *page = pvec->pages[i];
4345		struct pglist_data *pagepgdat = page_pgdat(page);
4346
4347		pgscanned++;
4348		if (pagepgdat != pgdat) {
4349			if (pgdat)
4350				spin_unlock_irq(&pgdat->lru_lock);
4351			pgdat = pagepgdat;
4352			spin_lock_irq(&pgdat->lru_lock);
4353		}
4354		lruvec = mem_cgroup_page_lruvec(page, pgdat);
4355
4356		if (!PageLRU(page) || !PageUnevictable(page))
4357			continue;
4358
4359		if (page_evictable(page)) {
4360			enum lru_list lru = page_lru_base_type(page);
4361
4362			VM_BUG_ON_PAGE(PageActive(page), page);
4363			ClearPageUnevictable(page);
4364			del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4365			add_page_to_lru_list(page, lruvec, lru);
4366			pgrescued++;
4367		}
4368	}
4369
4370	if (pgdat) {
4371		__count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4372		__count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4373		spin_unlock_irq(&pgdat->lru_lock);
4374	}
4375}
4376EXPORT_SYMBOL_GPL(check_move_unevictable_pages);
v4.17
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  linux/mm/vmscan.c
   4 *
   5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6 *
   7 *  Swap reorganised 29.12.95, Stephen Tweedie.
   8 *  kswapd added: 7.1.96  sct
   9 *  Removed kswapd_ctl limits, and swap out as many pages as needed
  10 *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
  11 *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
  12 *  Multiqueue VM started 5.8.00, Rik van Riel.
  13 */
  14
  15#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  16
  17#include <linux/mm.h>
  18#include <linux/sched/mm.h>
  19#include <linux/module.h>
  20#include <linux/gfp.h>
  21#include <linux/kernel_stat.h>
  22#include <linux/swap.h>
  23#include <linux/pagemap.h>
  24#include <linux/init.h>
  25#include <linux/highmem.h>
  26#include <linux/vmpressure.h>
  27#include <linux/vmstat.h>
  28#include <linux/file.h>
  29#include <linux/writeback.h>
  30#include <linux/blkdev.h>
  31#include <linux/buffer_head.h>	/* for try_to_release_page(),
  32					buffer_heads_over_limit */
  33#include <linux/mm_inline.h>
  34#include <linux/backing-dev.h>
  35#include <linux/rmap.h>
  36#include <linux/topology.h>
  37#include <linux/cpu.h>
  38#include <linux/cpuset.h>
  39#include <linux/compaction.h>
  40#include <linux/notifier.h>
  41#include <linux/rwsem.h>
  42#include <linux/delay.h>
  43#include <linux/kthread.h>
  44#include <linux/freezer.h>
  45#include <linux/memcontrol.h>
  46#include <linux/delayacct.h>
  47#include <linux/sysctl.h>
  48#include <linux/oom.h>
 
  49#include <linux/prefetch.h>
  50#include <linux/printk.h>
  51#include <linux/dax.h>
 
  52
  53#include <asm/tlbflush.h>
  54#include <asm/div64.h>
  55
  56#include <linux/swapops.h>
  57#include <linux/balloon_compaction.h>
  58
  59#include "internal.h"
  60
  61#define CREATE_TRACE_POINTS
  62#include <trace/events/vmscan.h>
  63
  64struct scan_control {
  65	/* How many pages shrink_list() should reclaim */
  66	unsigned long nr_to_reclaim;
  67
  68	/* This context's GFP mask */
  69	gfp_t gfp_mask;
  70
  71	/* Allocation order */
  72	int order;
  73
  74	/*
  75	 * Nodemask of nodes allowed by the caller. If NULL, all nodes
  76	 * are scanned.
  77	 */
  78	nodemask_t	*nodemask;
  79
  80	/*
  81	 * The memory cgroup that hit its limit and as a result is the
  82	 * primary target of this reclaim invocation.
  83	 */
  84	struct mem_cgroup *target_mem_cgroup;
  85
  86	/* Scan (total_size >> priority) pages at once */
  87	int priority;
  88
  89	/* The highest zone to isolate pages for reclaim from */
  90	enum zone_type reclaim_idx;
  91
  92	/* Writepage batching in laptop mode; RECLAIM_WRITE */
  93	unsigned int may_writepage:1;
  94
  95	/* Can mapped pages be reclaimed? */
  96	unsigned int may_unmap:1;
  97
  98	/* Can pages be swapped as part of reclaim? */
  99	unsigned int may_swap:1;
 100
 101	/*
 102	 * Cgroups are not reclaimed below their configured memory.low,
 103	 * unless we threaten to OOM. If any cgroups are skipped due to
 104	 * memory.low and nothing was reclaimed, go back for memory.low.
 105	 */
 106	unsigned int memcg_low_reclaim:1;
 107	unsigned int memcg_low_skipped:1;
 108
 109	unsigned int hibernation_mode:1;
 110
 111	/* One of the zones is ready for compaction */
 112	unsigned int compaction_ready:1;
 113
 
 
 
 
 
 
 
 
 
 
 
 
 114	/* Incremented by the number of inactive pages that were scanned */
 115	unsigned long nr_scanned;
 116
 117	/* Number of pages freed so far during a call to shrink_zones() */
 118	unsigned long nr_reclaimed;
 119
 120	struct {
 121		unsigned int dirty;
 122		unsigned int unqueued_dirty;
 123		unsigned int congested;
 124		unsigned int writeback;
 125		unsigned int immediate;
 126		unsigned int file_taken;
 127		unsigned int taken;
 128	} nr;
 
 
 
 129};
 130
 131#ifdef ARCH_HAS_PREFETCH
 132#define prefetch_prev_lru_page(_page, _base, _field)			\
 133	do {								\
 134		if ((_page)->lru.prev != _base) {			\
 135			struct page *prev;				\
 136									\
 137			prev = lru_to_page(&(_page->lru));		\
 138			prefetch(&prev->_field);			\
 139		}							\
 140	} while (0)
 141#else
 142#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
 143#endif
 144
 145#ifdef ARCH_HAS_PREFETCHW
 146#define prefetchw_prev_lru_page(_page, _base, _field)			\
 147	do {								\
 148		if ((_page)->lru.prev != _base) {			\
 149			struct page *prev;				\
 150									\
 151			prev = lru_to_page(&(_page->lru));		\
 152			prefetchw(&prev->_field);			\
 153		}							\
 154	} while (0)
 155#else
 156#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
 157#endif
 158
 159/*
 160 * From 0 .. 100.  Higher means more swappy.
 161 */
 162int vm_swappiness = 60;
 163/*
 164 * The total number of pages which are beyond the high watermark within all
 165 * zones.
 166 */
 167unsigned long vm_total_pages;
 168
 
 
 
 
 
 
 
 
 
 
 
 
 169static LIST_HEAD(shrinker_list);
 170static DECLARE_RWSEM(shrinker_rwsem);
 171
 172#ifdef CONFIG_MEMCG
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 173static bool global_reclaim(struct scan_control *sc)
 174{
 175	return !sc->target_mem_cgroup;
 176}
 177
 178/**
 179 * sane_reclaim - is the usual dirty throttling mechanism operational?
 180 * @sc: scan_control in question
 181 *
 182 * The normal page dirty throttling mechanism in balance_dirty_pages() is
 183 * completely broken with the legacy memcg and direct stalling in
 184 * shrink_page_list() is used for throttling instead, which lacks all the
 185 * niceties such as fairness, adaptive pausing, bandwidth proportional
 186 * allocation and configurability.
 187 *
 188 * This function tests whether the vmscan currently in progress can assume
 189 * that the normal dirty throttling mechanism is operational.
 190 */
 191static bool sane_reclaim(struct scan_control *sc)
 192{
 193	struct mem_cgroup *memcg = sc->target_mem_cgroup;
 194
 195	if (!memcg)
 196		return true;
 197#ifdef CONFIG_CGROUP_WRITEBACK
 198	if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
 199		return true;
 200#endif
 201	return false;
 202}
 203
 204static void set_memcg_congestion(pg_data_t *pgdat,
 205				struct mem_cgroup *memcg,
 206				bool congested)
 207{
 208	struct mem_cgroup_per_node *mn;
 209
 210	if (!memcg)
 211		return;
 212
 213	mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
 214	WRITE_ONCE(mn->congested, congested);
 215}
 216
 217static bool memcg_congested(pg_data_t *pgdat,
 218			struct mem_cgroup *memcg)
 219{
 220	struct mem_cgroup_per_node *mn;
 221
 222	mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
 223	return READ_ONCE(mn->congested);
 224
 225}
 226#else
 
 
 
 
 
 
 
 
 
 227static bool global_reclaim(struct scan_control *sc)
 228{
 229	return true;
 230}
 231
 232static bool sane_reclaim(struct scan_control *sc)
 233{
 234	return true;
 235}
 236
 237static inline void set_memcg_congestion(struct pglist_data *pgdat,
 238				struct mem_cgroup *memcg, bool congested)
 239{
 240}
 241
 242static inline bool memcg_congested(struct pglist_data *pgdat,
 243			struct mem_cgroup *memcg)
 244{
 245	return false;
 246
 247}
 248#endif
 249
 250/*
 251 * This misses isolated pages which are not accounted for to save counters.
 252 * As the data only determines if reclaim or compaction continues, it is
 253 * not expected that isolated pages will be a dominating factor.
 254 */
 255unsigned long zone_reclaimable_pages(struct zone *zone)
 256{
 257	unsigned long nr;
 258
 259	nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
 260		zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
 261	if (get_nr_swap_pages() > 0)
 262		nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
 263			zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
 264
 265	return nr;
 266}
 267
 268/**
 269 * lruvec_lru_size -  Returns the number of pages on the given LRU list.
 270 * @lruvec: lru vector
 271 * @lru: lru to use
 272 * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list)
 273 */
 274unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx)
 275{
 276	unsigned long lru_size;
 277	int zid;
 278
 279	if (!mem_cgroup_disabled())
 280		lru_size = mem_cgroup_get_lru_size(lruvec, lru);
 281	else
 
 282		lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
 283
 284	for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) {
 285		struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid];
 286		unsigned long size;
 287
 288		if (!managed_zone(zone))
 289			continue;
 290
 291		if (!mem_cgroup_disabled())
 292			size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid);
 293		else
 294			size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid],
 295				       NR_ZONE_LRU_BASE + lru);
 296		lru_size -= min(size, lru_size);
 297	}
 298
 299	return lru_size;
 300
 301}
 302
 303/*
 304 * Add a shrinker callback to be called from the vm.
 305 */
 306int prealloc_shrinker(struct shrinker *shrinker)
 307{
 308	size_t size = sizeof(*shrinker->nr_deferred);
 309
 310	if (shrinker->flags & SHRINKER_NUMA_AWARE)
 311		size *= nr_node_ids;
 312
 313	shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
 314	if (!shrinker->nr_deferred)
 315		return -ENOMEM;
 
 
 
 
 
 
 316	return 0;
 
 
 
 
 
 317}
 318
 319void free_prealloced_shrinker(struct shrinker *shrinker)
 320{
 
 
 
 
 
 
 321	kfree(shrinker->nr_deferred);
 322	shrinker->nr_deferred = NULL;
 323}
 324
 325void register_shrinker_prepared(struct shrinker *shrinker)
 326{
 327	down_write(&shrinker_rwsem);
 328	list_add_tail(&shrinker->list, &shrinker_list);
 
 
 
 
 329	up_write(&shrinker_rwsem);
 330}
 331
 332int register_shrinker(struct shrinker *shrinker)
 333{
 334	int err = prealloc_shrinker(shrinker);
 335
 336	if (err)
 337		return err;
 338	register_shrinker_prepared(shrinker);
 339	return 0;
 340}
 341EXPORT_SYMBOL(register_shrinker);
 342
 343/*
 344 * Remove one
 345 */
 346void unregister_shrinker(struct shrinker *shrinker)
 347{
 348	if (!shrinker->nr_deferred)
 349		return;
 
 
 350	down_write(&shrinker_rwsem);
 351	list_del(&shrinker->list);
 352	up_write(&shrinker_rwsem);
 353	kfree(shrinker->nr_deferred);
 354	shrinker->nr_deferred = NULL;
 355}
 356EXPORT_SYMBOL(unregister_shrinker);
 357
 358#define SHRINK_BATCH 128
 359
 360static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
 361				    struct shrinker *shrinker, int priority)
 362{
 363	unsigned long freed = 0;
 364	unsigned long long delta;
 365	long total_scan;
 366	long freeable;
 367	long nr;
 368	long new_nr;
 369	int nid = shrinkctl->nid;
 370	long batch_size = shrinker->batch ? shrinker->batch
 371					  : SHRINK_BATCH;
 372	long scanned = 0, next_deferred;
 373
 
 
 
 374	freeable = shrinker->count_objects(shrinker, shrinkctl);
 375	if (freeable == 0)
 376		return 0;
 377
 378	/*
 379	 * copy the current shrinker scan count into a local variable
 380	 * and zero it so that other concurrent shrinker invocations
 381	 * don't also do this scanning work.
 382	 */
 383	nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
 384
 385	total_scan = nr;
 386	delta = freeable >> priority;
 387	delta *= 4;
 388	do_div(delta, shrinker->seeks);
 
 
 
 
 
 
 
 
 
 
 389	total_scan += delta;
 390	if (total_scan < 0) {
 391		pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
 392		       shrinker->scan_objects, total_scan);
 393		total_scan = freeable;
 394		next_deferred = nr;
 395	} else
 396		next_deferred = total_scan;
 397
 398	/*
 399	 * We need to avoid excessive windup on filesystem shrinkers
 400	 * due to large numbers of GFP_NOFS allocations causing the
 401	 * shrinkers to return -1 all the time. This results in a large
 402	 * nr being built up so when a shrink that can do some work
 403	 * comes along it empties the entire cache due to nr >>>
 404	 * freeable. This is bad for sustaining a working set in
 405	 * memory.
 406	 *
 407	 * Hence only allow the shrinker to scan the entire cache when
 408	 * a large delta change is calculated directly.
 409	 */
 410	if (delta < freeable / 4)
 411		total_scan = min(total_scan, freeable / 2);
 412
 413	/*
 414	 * Avoid risking looping forever due to too large nr value:
 415	 * never try to free more than twice the estimate number of
 416	 * freeable entries.
 417	 */
 418	if (total_scan > freeable * 2)
 419		total_scan = freeable * 2;
 420
 421	trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
 422				   freeable, delta, total_scan, priority);
 423
 424	/*
 425	 * Normally, we should not scan less than batch_size objects in one
 426	 * pass to avoid too frequent shrinker calls, but if the slab has less
 427	 * than batch_size objects in total and we are really tight on memory,
 428	 * we will try to reclaim all available objects, otherwise we can end
 429	 * up failing allocations although there are plenty of reclaimable
 430	 * objects spread over several slabs with usage less than the
 431	 * batch_size.
 432	 *
 433	 * We detect the "tight on memory" situations by looking at the total
 434	 * number of objects we want to scan (total_scan). If it is greater
 435	 * than the total number of objects on slab (freeable), we must be
 436	 * scanning at high prio and therefore should try to reclaim as much as
 437	 * possible.
 438	 */
 439	while (total_scan >= batch_size ||
 440	       total_scan >= freeable) {
 441		unsigned long ret;
 442		unsigned long nr_to_scan = min(batch_size, total_scan);
 443
 444		shrinkctl->nr_to_scan = nr_to_scan;
 445		shrinkctl->nr_scanned = nr_to_scan;
 446		ret = shrinker->scan_objects(shrinker, shrinkctl);
 447		if (ret == SHRINK_STOP)
 448			break;
 449		freed += ret;
 450
 451		count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
 452		total_scan -= shrinkctl->nr_scanned;
 453		scanned += shrinkctl->nr_scanned;
 454
 455		cond_resched();
 456	}
 457
 458	if (next_deferred >= scanned)
 459		next_deferred -= scanned;
 460	else
 461		next_deferred = 0;
 462	/*
 463	 * move the unused scan count back into the shrinker in a
 464	 * manner that handles concurrent updates. If we exhausted the
 465	 * scan, there is no need to do an update.
 466	 */
 467	if (next_deferred > 0)
 468		new_nr = atomic_long_add_return(next_deferred,
 469						&shrinker->nr_deferred[nid]);
 470	else
 471		new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
 472
 473	trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
 474	return freed;
 475}
 476
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 477/**
 478 * shrink_slab - shrink slab caches
 479 * @gfp_mask: allocation context
 480 * @nid: node whose slab caches to target
 481 * @memcg: memory cgroup whose slab caches to target
 482 * @priority: the reclaim priority
 483 *
 484 * Call the shrink functions to age shrinkable caches.
 485 *
 486 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
 487 * unaware shrinkers will receive a node id of 0 instead.
 488 *
 489 * @memcg specifies the memory cgroup to target. If it is not NULL,
 490 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
 491 * objects from the memory cgroup specified. Otherwise, only unaware
 492 * shrinkers are called.
 493 *
 494 * @priority is sc->priority, we take the number of objects and >> by priority
 495 * in order to get the scan target.
 496 *
 497 * Returns the number of reclaimed slab objects.
 498 */
 499static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
 500				 struct mem_cgroup *memcg,
 501				 int priority)
 502{
 
 503	struct shrinker *shrinker;
 504	unsigned long freed = 0;
 505
 506	if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
 507		return 0;
 
 
 
 
 
 
 
 508
 509	if (!down_read_trylock(&shrinker_rwsem))
 510		goto out;
 511
 512	list_for_each_entry(shrinker, &shrinker_list, list) {
 513		struct shrink_control sc = {
 514			.gfp_mask = gfp_mask,
 515			.nid = nid,
 516			.memcg = memcg,
 517		};
 518
 519		/*
 520		 * If kernel memory accounting is disabled, we ignore
 521		 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
 522		 * passing NULL for memcg.
 523		 */
 524		if (memcg_kmem_enabled() &&
 525		    !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
 526			continue;
 527
 528		if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
 529			sc.nid = 0;
 530
 531		freed += do_shrink_slab(&sc, shrinker, priority);
 532		/*
 533		 * Bail out if someone want to register a new shrinker to
 534		 * prevent the regsitration from being stalled for long periods
 535		 * by parallel ongoing shrinking.
 536		 */
 537		if (rwsem_is_contended(&shrinker_rwsem)) {
 538			freed = freed ? : 1;
 539			break;
 540		}
 541	}
 542
 543	up_read(&shrinker_rwsem);
 544out:
 545	cond_resched();
 546	return freed;
 547}
 548
 549void drop_slab_node(int nid)
 550{
 551	unsigned long freed;
 552
 553	do {
 554		struct mem_cgroup *memcg = NULL;
 555
 556		freed = 0;
 
 557		do {
 558			freed += shrink_slab(GFP_KERNEL, nid, memcg, 0);
 559		} while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
 560	} while (freed > 10);
 561}
 562
 563void drop_slab(void)
 564{
 565	int nid;
 566
 567	for_each_online_node(nid)
 568		drop_slab_node(nid);
 569}
 570
 571static inline int is_page_cache_freeable(struct page *page)
 572{
 573	/*
 574	 * A freeable page cache page is referenced only by the caller
 575	 * that isolated the page, the page cache radix tree and
 576	 * optional buffer heads at page->private.
 577	 */
 578	int radix_pins = PageTransHuge(page) && PageSwapCache(page) ?
 579		HPAGE_PMD_NR : 1;
 580	return page_count(page) - page_has_private(page) == 1 + radix_pins;
 581}
 582
 583static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
 584{
 585	if (current->flags & PF_SWAPWRITE)
 586		return 1;
 587	if (!inode_write_congested(inode))
 588		return 1;
 589	if (inode_to_bdi(inode) == current->backing_dev_info)
 590		return 1;
 591	return 0;
 592}
 593
 594/*
 595 * We detected a synchronous write error writing a page out.  Probably
 596 * -ENOSPC.  We need to propagate that into the address_space for a subsequent
 597 * fsync(), msync() or close().
 598 *
 599 * The tricky part is that after writepage we cannot touch the mapping: nothing
 600 * prevents it from being freed up.  But we have a ref on the page and once
 601 * that page is locked, the mapping is pinned.
 602 *
 603 * We're allowed to run sleeping lock_page() here because we know the caller has
 604 * __GFP_FS.
 605 */
 606static void handle_write_error(struct address_space *mapping,
 607				struct page *page, int error)
 608{
 609	lock_page(page);
 610	if (page_mapping(page) == mapping)
 611		mapping_set_error(mapping, error);
 612	unlock_page(page);
 613}
 614
 615/* possible outcome of pageout() */
 616typedef enum {
 617	/* failed to write page out, page is locked */
 618	PAGE_KEEP,
 619	/* move page to the active list, page is locked */
 620	PAGE_ACTIVATE,
 621	/* page has been sent to the disk successfully, page is unlocked */
 622	PAGE_SUCCESS,
 623	/* page is clean and locked */
 624	PAGE_CLEAN,
 625} pageout_t;
 626
 627/*
 628 * pageout is called by shrink_page_list() for each dirty page.
 629 * Calls ->writepage().
 630 */
 631static pageout_t pageout(struct page *page, struct address_space *mapping,
 632			 struct scan_control *sc)
 633{
 634	/*
 635	 * If the page is dirty, only perform writeback if that write
 636	 * will be non-blocking.  To prevent this allocation from being
 637	 * stalled by pagecache activity.  But note that there may be
 638	 * stalls if we need to run get_block().  We could test
 639	 * PagePrivate for that.
 640	 *
 641	 * If this process is currently in __generic_file_write_iter() against
 642	 * this page's queue, we can perform writeback even if that
 643	 * will block.
 644	 *
 645	 * If the page is swapcache, write it back even if that would
 646	 * block, for some throttling. This happens by accident, because
 647	 * swap_backing_dev_info is bust: it doesn't reflect the
 648	 * congestion state of the swapdevs.  Easy to fix, if needed.
 649	 */
 650	if (!is_page_cache_freeable(page))
 651		return PAGE_KEEP;
 652	if (!mapping) {
 653		/*
 654		 * Some data journaling orphaned pages can have
 655		 * page->mapping == NULL while being dirty with clean buffers.
 656		 */
 657		if (page_has_private(page)) {
 658			if (try_to_free_buffers(page)) {
 659				ClearPageDirty(page);
 660				pr_info("%s: orphaned page\n", __func__);
 661				return PAGE_CLEAN;
 662			}
 663		}
 664		return PAGE_KEEP;
 665	}
 666	if (mapping->a_ops->writepage == NULL)
 667		return PAGE_ACTIVATE;
 668	if (!may_write_to_inode(mapping->host, sc))
 669		return PAGE_KEEP;
 670
 671	if (clear_page_dirty_for_io(page)) {
 672		int res;
 673		struct writeback_control wbc = {
 674			.sync_mode = WB_SYNC_NONE,
 675			.nr_to_write = SWAP_CLUSTER_MAX,
 676			.range_start = 0,
 677			.range_end = LLONG_MAX,
 678			.for_reclaim = 1,
 679		};
 680
 681		SetPageReclaim(page);
 682		res = mapping->a_ops->writepage(page, &wbc);
 683		if (res < 0)
 684			handle_write_error(mapping, page, res);
 685		if (res == AOP_WRITEPAGE_ACTIVATE) {
 686			ClearPageReclaim(page);
 687			return PAGE_ACTIVATE;
 688		}
 689
 690		if (!PageWriteback(page)) {
 691			/* synchronous write or broken a_ops? */
 692			ClearPageReclaim(page);
 693		}
 694		trace_mm_vmscan_writepage(page);
 695		inc_node_page_state(page, NR_VMSCAN_WRITE);
 696		return PAGE_SUCCESS;
 697	}
 698
 699	return PAGE_CLEAN;
 700}
 701
 702/*
 703 * Same as remove_mapping, but if the page is removed from the mapping, it
 704 * gets returned with a refcount of 0.
 705 */
 706static int __remove_mapping(struct address_space *mapping, struct page *page,
 707			    bool reclaimed)
 708{
 709	unsigned long flags;
 710	int refcount;
 711
 712	BUG_ON(!PageLocked(page));
 713	BUG_ON(mapping != page_mapping(page));
 714
 715	xa_lock_irqsave(&mapping->i_pages, flags);
 716	/*
 717	 * The non racy check for a busy page.
 718	 *
 719	 * Must be careful with the order of the tests. When someone has
 720	 * a ref to the page, it may be possible that they dirty it then
 721	 * drop the reference. So if PageDirty is tested before page_count
 722	 * here, then the following race may occur:
 723	 *
 724	 * get_user_pages(&page);
 725	 * [user mapping goes away]
 726	 * write_to(page);
 727	 *				!PageDirty(page)    [good]
 728	 * SetPageDirty(page);
 729	 * put_page(page);
 730	 *				!page_count(page)   [good, discard it]
 731	 *
 732	 * [oops, our write_to data is lost]
 733	 *
 734	 * Reversing the order of the tests ensures such a situation cannot
 735	 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
 736	 * load is not satisfied before that of page->_refcount.
 737	 *
 738	 * Note that if SetPageDirty is always performed via set_page_dirty,
 739	 * and thus under the i_pages lock, then this ordering is not required.
 740	 */
 741	if (unlikely(PageTransHuge(page)) && PageSwapCache(page))
 742		refcount = 1 + HPAGE_PMD_NR;
 743	else
 744		refcount = 2;
 745	if (!page_ref_freeze(page, refcount))
 746		goto cannot_free;
 747	/* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
 748	if (unlikely(PageDirty(page))) {
 749		page_ref_unfreeze(page, refcount);
 750		goto cannot_free;
 751	}
 752
 753	if (PageSwapCache(page)) {
 754		swp_entry_t swap = { .val = page_private(page) };
 755		mem_cgroup_swapout(page, swap);
 756		__delete_from_swap_cache(page);
 757		xa_unlock_irqrestore(&mapping->i_pages, flags);
 758		put_swap_page(page, swap);
 759	} else {
 760		void (*freepage)(struct page *);
 761		void *shadow = NULL;
 762
 763		freepage = mapping->a_ops->freepage;
 764		/*
 765		 * Remember a shadow entry for reclaimed file cache in
 766		 * order to detect refaults, thus thrashing, later on.
 767		 *
 768		 * But don't store shadows in an address space that is
 769		 * already exiting.  This is not just an optizimation,
 770		 * inode reclaim needs to empty out the radix tree or
 771		 * the nodes are lost.  Don't plant shadows behind its
 772		 * back.
 773		 *
 774		 * We also don't store shadows for DAX mappings because the
 775		 * only page cache pages found in these are zero pages
 776		 * covering holes, and because we don't want to mix DAX
 777		 * exceptional entries and shadow exceptional entries in the
 778		 * same address_space.
 779		 */
 780		if (reclaimed && page_is_file_cache(page) &&
 781		    !mapping_exiting(mapping) && !dax_mapping(mapping))
 782			shadow = workingset_eviction(mapping, page);
 783		__delete_from_page_cache(page, shadow);
 784		xa_unlock_irqrestore(&mapping->i_pages, flags);
 785
 786		if (freepage != NULL)
 787			freepage(page);
 788	}
 789
 790	return 1;
 791
 792cannot_free:
 793	xa_unlock_irqrestore(&mapping->i_pages, flags);
 794	return 0;
 795}
 796
 797/*
 798 * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
 799 * someone else has a ref on the page, abort and return 0.  If it was
 800 * successfully detached, return 1.  Assumes the caller has a single ref on
 801 * this page.
 802 */
 803int remove_mapping(struct address_space *mapping, struct page *page)
 804{
 805	if (__remove_mapping(mapping, page, false)) {
 806		/*
 807		 * Unfreezing the refcount with 1 rather than 2 effectively
 808		 * drops the pagecache ref for us without requiring another
 809		 * atomic operation.
 810		 */
 811		page_ref_unfreeze(page, 1);
 812		return 1;
 813	}
 814	return 0;
 815}
 816
 817/**
 818 * putback_lru_page - put previously isolated page onto appropriate LRU list
 819 * @page: page to be put back to appropriate lru list
 820 *
 821 * Add previously isolated @page to appropriate LRU list.
 822 * Page may still be unevictable for other reasons.
 823 *
 824 * lru_lock must not be held, interrupts must be enabled.
 825 */
 826void putback_lru_page(struct page *page)
 827{
 828	lru_cache_add(page);
 829	put_page(page);		/* drop ref from isolate */
 830}
 831
 832enum page_references {
 833	PAGEREF_RECLAIM,
 834	PAGEREF_RECLAIM_CLEAN,
 835	PAGEREF_KEEP,
 836	PAGEREF_ACTIVATE,
 837};
 838
 839static enum page_references page_check_references(struct page *page,
 840						  struct scan_control *sc)
 841{
 842	int referenced_ptes, referenced_page;
 843	unsigned long vm_flags;
 844
 845	referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
 846					  &vm_flags);
 847	referenced_page = TestClearPageReferenced(page);
 848
 849	/*
 850	 * Mlock lost the isolation race with us.  Let try_to_unmap()
 851	 * move the page to the unevictable list.
 852	 */
 853	if (vm_flags & VM_LOCKED)
 854		return PAGEREF_RECLAIM;
 855
 856	if (referenced_ptes) {
 857		if (PageSwapBacked(page))
 858			return PAGEREF_ACTIVATE;
 859		/*
 860		 * All mapped pages start out with page table
 861		 * references from the instantiating fault, so we need
 862		 * to look twice if a mapped file page is used more
 863		 * than once.
 864		 *
 865		 * Mark it and spare it for another trip around the
 866		 * inactive list.  Another page table reference will
 867		 * lead to its activation.
 868		 *
 869		 * Note: the mark is set for activated pages as well
 870		 * so that recently deactivated but used pages are
 871		 * quickly recovered.
 872		 */
 873		SetPageReferenced(page);
 874
 875		if (referenced_page || referenced_ptes > 1)
 876			return PAGEREF_ACTIVATE;
 877
 878		/*
 879		 * Activate file-backed executable pages after first usage.
 880		 */
 881		if (vm_flags & VM_EXEC)
 882			return PAGEREF_ACTIVATE;
 883
 884		return PAGEREF_KEEP;
 885	}
 886
 887	/* Reclaim if clean, defer dirty pages to writeback */
 888	if (referenced_page && !PageSwapBacked(page))
 889		return PAGEREF_RECLAIM_CLEAN;
 890
 891	return PAGEREF_RECLAIM;
 892}
 893
 894/* Check if a page is dirty or under writeback */
 895static void page_check_dirty_writeback(struct page *page,
 896				       bool *dirty, bool *writeback)
 897{
 898	struct address_space *mapping;
 899
 900	/*
 901	 * Anonymous pages are not handled by flushers and must be written
 902	 * from reclaim context. Do not stall reclaim based on them
 903	 */
 904	if (!page_is_file_cache(page) ||
 905	    (PageAnon(page) && !PageSwapBacked(page))) {
 906		*dirty = false;
 907		*writeback = false;
 908		return;
 909	}
 910
 911	/* By default assume that the page flags are accurate */
 912	*dirty = PageDirty(page);
 913	*writeback = PageWriteback(page);
 914
 915	/* Verify dirty/writeback state if the filesystem supports it */
 916	if (!page_has_private(page))
 917		return;
 918
 919	mapping = page_mapping(page);
 920	if (mapping && mapping->a_ops->is_dirty_writeback)
 921		mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
 922}
 923
 924/*
 925 * shrink_page_list() returns the number of reclaimed pages
 926 */
 927static unsigned long shrink_page_list(struct list_head *page_list,
 928				      struct pglist_data *pgdat,
 929				      struct scan_control *sc,
 930				      enum ttu_flags ttu_flags,
 931				      struct reclaim_stat *stat,
 932				      bool force_reclaim)
 933{
 934	LIST_HEAD(ret_pages);
 935	LIST_HEAD(free_pages);
 936	int pgactivate = 0;
 937	unsigned nr_unqueued_dirty = 0;
 938	unsigned nr_dirty = 0;
 939	unsigned nr_congested = 0;
 940	unsigned nr_reclaimed = 0;
 941	unsigned nr_writeback = 0;
 942	unsigned nr_immediate = 0;
 943	unsigned nr_ref_keep = 0;
 944	unsigned nr_unmap_fail = 0;
 945
 
 946	cond_resched();
 947
 948	while (!list_empty(page_list)) {
 949		struct address_space *mapping;
 950		struct page *page;
 951		int may_enter_fs;
 952		enum page_references references = PAGEREF_RECLAIM_CLEAN;
 953		bool dirty, writeback;
 
 954
 955		cond_resched();
 956
 957		page = lru_to_page(page_list);
 958		list_del(&page->lru);
 959
 960		if (!trylock_page(page))
 961			goto keep;
 962
 963		VM_BUG_ON_PAGE(PageActive(page), page);
 964
 965		sc->nr_scanned++;
 
 
 
 966
 967		if (unlikely(!page_evictable(page)))
 968			goto activate_locked;
 969
 970		if (!sc->may_unmap && page_mapped(page))
 971			goto keep_locked;
 972
 973		/* Double the slab pressure for mapped and swapcache pages */
 974		if ((page_mapped(page) || PageSwapCache(page)) &&
 975		    !(PageAnon(page) && !PageSwapBacked(page)))
 976			sc->nr_scanned++;
 977
 978		may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
 979			(PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
 980
 981		/*
 982		 * The number of dirty pages determines if a node is marked
 983		 * reclaim_congested which affects wait_iff_congested. kswapd
 984		 * will stall and start writing pages if the tail of the LRU
 985		 * is all dirty unqueued pages.
 986		 */
 987		page_check_dirty_writeback(page, &dirty, &writeback);
 988		if (dirty || writeback)
 989			nr_dirty++;
 990
 991		if (dirty && !writeback)
 992			nr_unqueued_dirty++;
 993
 994		/*
 995		 * Treat this page as congested if the underlying BDI is or if
 996		 * pages are cycling through the LRU so quickly that the
 997		 * pages marked for immediate reclaim are making it to the
 998		 * end of the LRU a second time.
 999		 */
1000		mapping = page_mapping(page);
1001		if (((dirty || writeback) && mapping &&
1002		     inode_write_congested(mapping->host)) ||
1003		    (writeback && PageReclaim(page)))
1004			nr_congested++;
1005
1006		/*
1007		 * If a page at the tail of the LRU is under writeback, there
1008		 * are three cases to consider.
1009		 *
1010		 * 1) If reclaim is encountering an excessive number of pages
1011		 *    under writeback and this page is both under writeback and
1012		 *    PageReclaim then it indicates that pages are being queued
1013		 *    for IO but are being recycled through the LRU before the
1014		 *    IO can complete. Waiting on the page itself risks an
1015		 *    indefinite stall if it is impossible to writeback the
1016		 *    page due to IO error or disconnected storage so instead
1017		 *    note that the LRU is being scanned too quickly and the
1018		 *    caller can stall after page list has been processed.
1019		 *
1020		 * 2) Global or new memcg reclaim encounters a page that is
1021		 *    not marked for immediate reclaim, or the caller does not
1022		 *    have __GFP_FS (or __GFP_IO if it's simply going to swap,
1023		 *    not to fs). In this case mark the page for immediate
1024		 *    reclaim and continue scanning.
1025		 *
1026		 *    Require may_enter_fs because we would wait on fs, which
1027		 *    may not have submitted IO yet. And the loop driver might
1028		 *    enter reclaim, and deadlock if it waits on a page for
1029		 *    which it is needed to do the write (loop masks off
1030		 *    __GFP_IO|__GFP_FS for this reason); but more thought
1031		 *    would probably show more reasons.
1032		 *
1033		 * 3) Legacy memcg encounters a page that is already marked
1034		 *    PageReclaim. memcg does not have any dirty pages
1035		 *    throttling so we could easily OOM just because too many
1036		 *    pages are in writeback and there is nothing else to
1037		 *    reclaim. Wait for the writeback to complete.
1038		 *
1039		 * In cases 1) and 2) we activate the pages to get them out of
1040		 * the way while we continue scanning for clean pages on the
1041		 * inactive list and refilling from the active list. The
1042		 * observation here is that waiting for disk writes is more
1043		 * expensive than potentially causing reloads down the line.
1044		 * Since they're marked for immediate reclaim, they won't put
1045		 * memory pressure on the cache working set any longer than it
1046		 * takes to write them to disk.
1047		 */
1048		if (PageWriteback(page)) {
1049			/* Case 1 above */
1050			if (current_is_kswapd() &&
1051			    PageReclaim(page) &&
1052			    test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
1053				nr_immediate++;
1054				goto activate_locked;
1055
1056			/* Case 2 above */
1057			} else if (sane_reclaim(sc) ||
1058			    !PageReclaim(page) || !may_enter_fs) {
1059				/*
1060				 * This is slightly racy - end_page_writeback()
1061				 * might have just cleared PageReclaim, then
1062				 * setting PageReclaim here end up interpreted
1063				 * as PageReadahead - but that does not matter
1064				 * enough to care.  What we do want is for this
1065				 * page to have PageReclaim set next time memcg
1066				 * reclaim reaches the tests above, so it will
1067				 * then wait_on_page_writeback() to avoid OOM;
1068				 * and it's also appropriate in global reclaim.
1069				 */
1070				SetPageReclaim(page);
1071				nr_writeback++;
1072				goto activate_locked;
1073
1074			/* Case 3 above */
1075			} else {
1076				unlock_page(page);
1077				wait_on_page_writeback(page);
1078				/* then go back and try same page again */
1079				list_add_tail(&page->lru, page_list);
1080				continue;
1081			}
1082		}
1083
1084		if (!force_reclaim)
1085			references = page_check_references(page, sc);
1086
1087		switch (references) {
1088		case PAGEREF_ACTIVATE:
1089			goto activate_locked;
1090		case PAGEREF_KEEP:
1091			nr_ref_keep++;
1092			goto keep_locked;
1093		case PAGEREF_RECLAIM:
1094		case PAGEREF_RECLAIM_CLEAN:
1095			; /* try to reclaim the page below */
1096		}
1097
1098		/*
1099		 * Anonymous process memory has backing store?
1100		 * Try to allocate it some swap space here.
1101		 * Lazyfree page could be freed directly
1102		 */
1103		if (PageAnon(page) && PageSwapBacked(page)) {
1104			if (!PageSwapCache(page)) {
1105				if (!(sc->gfp_mask & __GFP_IO))
1106					goto keep_locked;
1107				if (PageTransHuge(page)) {
1108					/* cannot split THP, skip it */
1109					if (!can_split_huge_page(page, NULL))
1110						goto activate_locked;
1111					/*
1112					 * Split pages without a PMD map right
1113					 * away. Chances are some or all of the
1114					 * tail pages can be freed without IO.
1115					 */
1116					if (!compound_mapcount(page) &&
1117					    split_huge_page_to_list(page,
1118								    page_list))
1119						goto activate_locked;
1120				}
1121				if (!add_to_swap(page)) {
1122					if (!PageTransHuge(page))
1123						goto activate_locked;
1124					/* Fallback to swap normal pages */
1125					if (split_huge_page_to_list(page,
1126								    page_list))
1127						goto activate_locked;
1128#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1129					count_vm_event(THP_SWPOUT_FALLBACK);
1130#endif
1131					if (!add_to_swap(page))
1132						goto activate_locked;
1133				}
1134
1135				may_enter_fs = 1;
1136
1137				/* Adding to swap updated mapping */
1138				mapping = page_mapping(page);
1139			}
1140		} else if (unlikely(PageTransHuge(page))) {
1141			/* Split file THP */
1142			if (split_huge_page_to_list(page, page_list))
1143				goto keep_locked;
1144		}
1145
1146		/*
 
 
 
 
 
 
 
 
 
 
 
 
1147		 * The page is mapped into the page tables of one or more
1148		 * processes. Try to unmap it here.
1149		 */
1150		if (page_mapped(page)) {
1151			enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH;
1152
1153			if (unlikely(PageTransHuge(page)))
1154				flags |= TTU_SPLIT_HUGE_PMD;
1155			if (!try_to_unmap(page, flags)) {
1156				nr_unmap_fail++;
1157				goto activate_locked;
1158			}
1159		}
1160
1161		if (PageDirty(page)) {
1162			/*
1163			 * Only kswapd can writeback filesystem pages
1164			 * to avoid risk of stack overflow. But avoid
1165			 * injecting inefficient single-page IO into
1166			 * flusher writeback as much as possible: only
1167			 * write pages when we've encountered many
1168			 * dirty pages, and when we've already scanned
1169			 * the rest of the LRU for clean pages and see
1170			 * the same dirty pages again (PageReclaim).
1171			 */
1172			if (page_is_file_cache(page) &&
1173			    (!current_is_kswapd() || !PageReclaim(page) ||
1174			     !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
1175				/*
1176				 * Immediately reclaim when written back.
1177				 * Similar in principal to deactivate_page()
1178				 * except we already have the page isolated
1179				 * and know it's dirty
1180				 */
1181				inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
1182				SetPageReclaim(page);
1183
1184				goto activate_locked;
1185			}
1186
1187			if (references == PAGEREF_RECLAIM_CLEAN)
1188				goto keep_locked;
1189			if (!may_enter_fs)
1190				goto keep_locked;
1191			if (!sc->may_writepage)
1192				goto keep_locked;
1193
1194			/*
1195			 * Page is dirty. Flush the TLB if a writable entry
1196			 * potentially exists to avoid CPU writes after IO
1197			 * starts and then write it out here.
1198			 */
1199			try_to_unmap_flush_dirty();
1200			switch (pageout(page, mapping, sc)) {
1201			case PAGE_KEEP:
1202				goto keep_locked;
1203			case PAGE_ACTIVATE:
1204				goto activate_locked;
1205			case PAGE_SUCCESS:
1206				if (PageWriteback(page))
1207					goto keep;
1208				if (PageDirty(page))
1209					goto keep;
1210
1211				/*
1212				 * A synchronous write - probably a ramdisk.  Go
1213				 * ahead and try to reclaim the page.
1214				 */
1215				if (!trylock_page(page))
1216					goto keep;
1217				if (PageDirty(page) || PageWriteback(page))
1218					goto keep_locked;
1219				mapping = page_mapping(page);
1220			case PAGE_CLEAN:
1221				; /* try to free the page below */
1222			}
1223		}
1224
1225		/*
1226		 * If the page has buffers, try to free the buffer mappings
1227		 * associated with this page. If we succeed we try to free
1228		 * the page as well.
1229		 *
1230		 * We do this even if the page is PageDirty().
1231		 * try_to_release_page() does not perform I/O, but it is
1232		 * possible for a page to have PageDirty set, but it is actually
1233		 * clean (all its buffers are clean).  This happens if the
1234		 * buffers were written out directly, with submit_bh(). ext3
1235		 * will do this, as well as the blockdev mapping.
1236		 * try_to_release_page() will discover that cleanness and will
1237		 * drop the buffers and mark the page clean - it can be freed.
1238		 *
1239		 * Rarely, pages can have buffers and no ->mapping.  These are
1240		 * the pages which were not successfully invalidated in
1241		 * truncate_complete_page().  We try to drop those buffers here
1242		 * and if that worked, and the page is no longer mapped into
1243		 * process address space (page_count == 1) it can be freed.
1244		 * Otherwise, leave the page on the LRU so it is swappable.
1245		 */
1246		if (page_has_private(page)) {
1247			if (!try_to_release_page(page, sc->gfp_mask))
1248				goto activate_locked;
1249			if (!mapping && page_count(page) == 1) {
1250				unlock_page(page);
1251				if (put_page_testzero(page))
1252					goto free_it;
1253				else {
1254					/*
1255					 * rare race with speculative reference.
1256					 * the speculative reference will free
1257					 * this page shortly, so we may
1258					 * increment nr_reclaimed here (and
1259					 * leave it off the LRU).
1260					 */
1261					nr_reclaimed++;
1262					continue;
1263				}
1264			}
1265		}
1266
1267		if (PageAnon(page) && !PageSwapBacked(page)) {
1268			/* follow __remove_mapping for reference */
1269			if (!page_ref_freeze(page, 1))
1270				goto keep_locked;
1271			if (PageDirty(page)) {
1272				page_ref_unfreeze(page, 1);
1273				goto keep_locked;
1274			}
1275
1276			count_vm_event(PGLAZYFREED);
1277			count_memcg_page_event(page, PGLAZYFREED);
1278		} else if (!mapping || !__remove_mapping(mapping, page, true))
1279			goto keep_locked;
 
 
 
1280		/*
1281		 * At this point, we have no other references and there is
1282		 * no way to pick any more up (removed from LRU, removed
1283		 * from pagecache). Can use non-atomic bitops now (and
1284		 * we obviously don't have to worry about waking up a process
1285		 * waiting on the page lock, because there are no references.
1286		 */
1287		__ClearPageLocked(page);
1288free_it:
1289		nr_reclaimed++;
1290
1291		/*
1292		 * Is there need to periodically free_page_list? It would
1293		 * appear not as the counts should be low
1294		 */
1295		if (unlikely(PageTransHuge(page))) {
1296			mem_cgroup_uncharge(page);
1297			(*get_compound_page_dtor(page))(page);
1298		} else
1299			list_add(&page->lru, &free_pages);
1300		continue;
1301
 
 
 
 
 
 
 
 
 
1302activate_locked:
1303		/* Not a candidate for swapping, so reclaim swap space. */
1304		if (PageSwapCache(page) && (mem_cgroup_swap_full(page) ||
1305						PageMlocked(page)))
1306			try_to_free_swap(page);
1307		VM_BUG_ON_PAGE(PageActive(page), page);
1308		if (!PageMlocked(page)) {
 
1309			SetPageActive(page);
1310			pgactivate++;
1311			count_memcg_page_event(page, PGACTIVATE);
1312		}
1313keep_locked:
1314		unlock_page(page);
1315keep:
1316		list_add(&page->lru, &ret_pages);
1317		VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1318	}
1319
 
 
1320	mem_cgroup_uncharge_list(&free_pages);
1321	try_to_unmap_flush();
1322	free_unref_page_list(&free_pages);
1323
1324	list_splice(&ret_pages, page_list);
1325	count_vm_events(PGACTIVATE, pgactivate);
1326
1327	if (stat) {
1328		stat->nr_dirty = nr_dirty;
1329		stat->nr_congested = nr_congested;
1330		stat->nr_unqueued_dirty = nr_unqueued_dirty;
1331		stat->nr_writeback = nr_writeback;
1332		stat->nr_immediate = nr_immediate;
1333		stat->nr_activate = pgactivate;
1334		stat->nr_ref_keep = nr_ref_keep;
1335		stat->nr_unmap_fail = nr_unmap_fail;
1336	}
1337	return nr_reclaimed;
1338}
1339
1340unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1341					    struct list_head *page_list)
1342{
1343	struct scan_control sc = {
1344		.gfp_mask = GFP_KERNEL,
1345		.priority = DEF_PRIORITY,
1346		.may_unmap = 1,
1347	};
 
1348	unsigned long ret;
1349	struct page *page, *next;
1350	LIST_HEAD(clean_pages);
1351
1352	list_for_each_entry_safe(page, next, page_list, lru) {
1353		if (page_is_file_cache(page) && !PageDirty(page) &&
1354		    !__PageMovable(page)) {
1355			ClearPageActive(page);
1356			list_move(&page->lru, &clean_pages);
1357		}
1358	}
1359
1360	ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
1361			TTU_IGNORE_ACCESS, NULL, true);
1362	list_splice(&clean_pages, page_list);
1363	mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
1364	return ret;
1365}
1366
1367/*
1368 * Attempt to remove the specified page from its LRU.  Only take this page
1369 * if it is of the appropriate PageActive status.  Pages which are being
1370 * freed elsewhere are also ignored.
1371 *
1372 * page:	page to consider
1373 * mode:	one of the LRU isolation modes defined above
1374 *
1375 * returns 0 on success, -ve errno on failure.
1376 */
1377int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1378{
1379	int ret = -EINVAL;
1380
1381	/* Only take pages on the LRU. */
1382	if (!PageLRU(page))
1383		return ret;
1384
1385	/* Compaction should not handle unevictable pages but CMA can do so */
1386	if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1387		return ret;
1388
1389	ret = -EBUSY;
1390
1391	/*
1392	 * To minimise LRU disruption, the caller can indicate that it only
1393	 * wants to isolate pages it will be able to operate on without
1394	 * blocking - clean pages for the most part.
1395	 *
1396	 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1397	 * that it is possible to migrate without blocking
1398	 */
1399	if (mode & ISOLATE_ASYNC_MIGRATE) {
1400		/* All the caller can do on PageWriteback is block */
1401		if (PageWriteback(page))
1402			return ret;
1403
1404		if (PageDirty(page)) {
1405			struct address_space *mapping;
1406			bool migrate_dirty;
1407
1408			/*
1409			 * Only pages without mappings or that have a
1410			 * ->migratepage callback are possible to migrate
1411			 * without blocking. However, we can be racing with
1412			 * truncation so it's necessary to lock the page
1413			 * to stabilise the mapping as truncation holds
1414			 * the page lock until after the page is removed
1415			 * from the page cache.
1416			 */
1417			if (!trylock_page(page))
1418				return ret;
1419
1420			mapping = page_mapping(page);
1421			migrate_dirty = !mapping || mapping->a_ops->migratepage;
1422			unlock_page(page);
1423			if (!migrate_dirty)
1424				return ret;
1425		}
1426	}
1427
1428	if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1429		return ret;
1430
1431	if (likely(get_page_unless_zero(page))) {
1432		/*
1433		 * Be careful not to clear PageLRU until after we're
1434		 * sure the page is not being freed elsewhere -- the
1435		 * page release code relies on it.
1436		 */
1437		ClearPageLRU(page);
1438		ret = 0;
1439	}
1440
1441	return ret;
1442}
1443
1444
1445/*
1446 * Update LRU sizes after isolating pages. The LRU size updates must
1447 * be complete before mem_cgroup_update_lru_size due to a santity check.
1448 */
1449static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1450			enum lru_list lru, unsigned long *nr_zone_taken)
1451{
1452	int zid;
1453
1454	for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1455		if (!nr_zone_taken[zid])
1456			continue;
1457
1458		__update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1459#ifdef CONFIG_MEMCG
1460		mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1461#endif
1462	}
1463
1464}
1465
1466/*
1467 * zone_lru_lock is heavily contended.  Some of the functions that
1468 * shrink the lists perform better by taking out a batch of pages
1469 * and working on them outside the LRU lock.
1470 *
1471 * For pagecache intensive workloads, this function is the hottest
1472 * spot in the kernel (apart from copy_*_user functions).
1473 *
1474 * Appropriate locks must be held before calling this function.
1475 *
1476 * @nr_to_scan:	The number of eligible pages to look through on the list.
1477 * @lruvec:	The LRU vector to pull pages from.
1478 * @dst:	The temp list to put pages on to.
1479 * @nr_scanned:	The number of pages that were scanned.
1480 * @sc:		The scan_control struct for this reclaim session
1481 * @mode:	One of the LRU isolation modes
1482 * @lru:	LRU list id for isolating
1483 *
1484 * returns how many pages were moved onto *@dst.
1485 */
1486static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1487		struct lruvec *lruvec, struct list_head *dst,
1488		unsigned long *nr_scanned, struct scan_control *sc,
1489		isolate_mode_t mode, enum lru_list lru)
1490{
1491	struct list_head *src = &lruvec->lists[lru];
1492	unsigned long nr_taken = 0;
1493	unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
1494	unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
1495	unsigned long skipped = 0;
1496	unsigned long scan, total_scan, nr_pages;
1497	LIST_HEAD(pages_skipped);
 
1498
 
1499	scan = 0;
1500	for (total_scan = 0;
1501	     scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src);
1502	     total_scan++) {
1503		struct page *page;
1504
1505		page = lru_to_page(src);
1506		prefetchw_prev_lru_page(page, src, flags);
1507
1508		VM_BUG_ON_PAGE(!PageLRU(page), page);
1509
 
 
 
1510		if (page_zonenum(page) > sc->reclaim_idx) {
1511			list_move(&page->lru, &pages_skipped);
1512			nr_skipped[page_zonenum(page)]++;
1513			continue;
1514		}
1515
1516		/*
1517		 * Do not count skipped pages because that makes the function
1518		 * return with no isolated pages if the LRU mostly contains
1519		 * ineligible pages.  This causes the VM to not reclaim any
1520		 * pages, triggering a premature OOM.
 
 
 
 
1521		 */
1522		scan++;
1523		switch (__isolate_lru_page(page, mode)) {
1524		case 0:
1525			nr_pages = hpage_nr_pages(page);
1526			nr_taken += nr_pages;
1527			nr_zone_taken[page_zonenum(page)] += nr_pages;
1528			list_move(&page->lru, dst);
1529			break;
1530
1531		case -EBUSY:
1532			/* else it is being freed elsewhere */
1533			list_move(&page->lru, src);
1534			continue;
1535
1536		default:
1537			BUG();
1538		}
1539	}
1540
1541	/*
1542	 * Splice any skipped pages to the start of the LRU list. Note that
1543	 * this disrupts the LRU order when reclaiming for lower zones but
1544	 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1545	 * scanning would soon rescan the same pages to skip and put the
1546	 * system at risk of premature OOM.
1547	 */
1548	if (!list_empty(&pages_skipped)) {
1549		int zid;
1550
1551		list_splice(&pages_skipped, src);
1552		for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1553			if (!nr_skipped[zid])
1554				continue;
1555
1556			__count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
1557			skipped += nr_skipped[zid];
1558		}
1559	}
1560	*nr_scanned = total_scan;
1561	trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
1562				    total_scan, skipped, nr_taken, mode, lru);
1563	update_lru_sizes(lruvec, lru, nr_zone_taken);
1564	return nr_taken;
1565}
1566
1567/**
1568 * isolate_lru_page - tries to isolate a page from its LRU list
1569 * @page: page to isolate from its LRU list
1570 *
1571 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1572 * vmstat statistic corresponding to whatever LRU list the page was on.
1573 *
1574 * Returns 0 if the page was removed from an LRU list.
1575 * Returns -EBUSY if the page was not on an LRU list.
1576 *
1577 * The returned page will have PageLRU() cleared.  If it was found on
1578 * the active list, it will have PageActive set.  If it was found on
1579 * the unevictable list, it will have the PageUnevictable bit set. That flag
1580 * may need to be cleared by the caller before letting the page go.
1581 *
1582 * The vmstat statistic corresponding to the list on which the page was
1583 * found will be decremented.
1584 *
1585 * Restrictions:
1586 *
1587 * (1) Must be called with an elevated refcount on the page. This is a
1588 *     fundamentnal difference from isolate_lru_pages (which is called
1589 *     without a stable reference).
1590 * (2) the lru_lock must not be held.
1591 * (3) interrupts must be enabled.
1592 */
1593int isolate_lru_page(struct page *page)
1594{
1595	int ret = -EBUSY;
1596
1597	VM_BUG_ON_PAGE(!page_count(page), page);
1598	WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
1599
1600	if (PageLRU(page)) {
1601		struct zone *zone = page_zone(page);
1602		struct lruvec *lruvec;
1603
1604		spin_lock_irq(zone_lru_lock(zone));
1605		lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
1606		if (PageLRU(page)) {
1607			int lru = page_lru(page);
1608			get_page(page);
1609			ClearPageLRU(page);
1610			del_page_from_lru_list(page, lruvec, lru);
1611			ret = 0;
1612		}
1613		spin_unlock_irq(zone_lru_lock(zone));
1614	}
1615	return ret;
1616}
1617
1618/*
1619 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1620 * then get resheduled. When there are massive number of tasks doing page
1621 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1622 * the LRU list will go small and be scanned faster than necessary, leading to
1623 * unnecessary swapping, thrashing and OOM.
1624 */
1625static int too_many_isolated(struct pglist_data *pgdat, int file,
1626		struct scan_control *sc)
1627{
1628	unsigned long inactive, isolated;
1629
1630	if (current_is_kswapd())
1631		return 0;
1632
1633	if (!sane_reclaim(sc))
1634		return 0;
1635
1636	if (file) {
1637		inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1638		isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
1639	} else {
1640		inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1641		isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
1642	}
1643
1644	/*
1645	 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1646	 * won't get blocked by normal direct-reclaimers, forming a circular
1647	 * deadlock.
1648	 */
1649	if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
1650		inactive >>= 3;
1651
1652	return isolated > inactive;
1653}
1654
1655static noinline_for_stack void
1656putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1657{
1658	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1659	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
 
1660	LIST_HEAD(pages_to_free);
 
 
1661
1662	/*
1663	 * Put back any unfreeable pages.
1664	 */
1665	while (!list_empty(page_list)) {
1666		struct page *page = lru_to_page(page_list);
1667		int lru;
1668
1669		VM_BUG_ON_PAGE(PageLRU(page), page);
1670		list_del(&page->lru);
1671		if (unlikely(!page_evictable(page))) {
 
1672			spin_unlock_irq(&pgdat->lru_lock);
1673			putback_lru_page(page);
1674			spin_lock_irq(&pgdat->lru_lock);
1675			continue;
1676		}
1677
1678		lruvec = mem_cgroup_page_lruvec(page, pgdat);
1679
1680		SetPageLRU(page);
1681		lru = page_lru(page);
1682		add_page_to_lru_list(page, lruvec, lru);
1683
1684		if (is_active_lru(lru)) {
1685			int file = is_file_lru(lru);
1686			int numpages = hpage_nr_pages(page);
1687			reclaim_stat->recent_rotated[file] += numpages;
1688		}
1689		if (put_page_testzero(page)) {
1690			__ClearPageLRU(page);
1691			__ClearPageActive(page);
1692			del_page_from_lru_list(page, lruvec, lru);
1693
1694			if (unlikely(PageCompound(page))) {
1695				spin_unlock_irq(&pgdat->lru_lock);
1696				mem_cgroup_uncharge(page);
1697				(*get_compound_page_dtor(page))(page);
1698				spin_lock_irq(&pgdat->lru_lock);
1699			} else
1700				list_add(&page->lru, &pages_to_free);
 
 
1701		}
1702	}
1703
1704	/*
1705	 * To save our caller's stack, now use input list for pages to free.
1706	 */
1707	list_splice(&pages_to_free, page_list);
 
 
1708}
1709
1710/*
1711 * If a kernel thread (such as nfsd for loop-back mounts) services
1712 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1713 * In that case we should only throttle if the backing device it is
1714 * writing to is congested.  In other cases it is safe to throttle.
1715 */
1716static int current_may_throttle(void)
1717{
1718	return !(current->flags & PF_LESS_THROTTLE) ||
1719		current->backing_dev_info == NULL ||
1720		bdi_write_congested(current->backing_dev_info);
1721}
1722
1723/*
1724 * shrink_inactive_list() is a helper for shrink_node().  It returns the number
1725 * of reclaimed pages
1726 */
1727static noinline_for_stack unsigned long
1728shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1729		     struct scan_control *sc, enum lru_list lru)
1730{
1731	LIST_HEAD(page_list);
1732	unsigned long nr_scanned;
1733	unsigned long nr_reclaimed = 0;
1734	unsigned long nr_taken;
1735	struct reclaim_stat stat = {};
1736	isolate_mode_t isolate_mode = 0;
1737	int file = is_file_lru(lru);
 
1738	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1739	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1740	bool stalled = false;
1741
1742	while (unlikely(too_many_isolated(pgdat, file, sc))) {
1743		if (stalled)
1744			return 0;
1745
1746		/* wait a bit for the reclaimer. */
1747		msleep(100);
1748		stalled = true;
1749
1750		/* We are about to die and free our memory. Return now. */
1751		if (fatal_signal_pending(current))
1752			return SWAP_CLUSTER_MAX;
1753	}
1754
1755	lru_add_drain();
1756
1757	if (!sc->may_unmap)
1758		isolate_mode |= ISOLATE_UNMAPPED;
1759
1760	spin_lock_irq(&pgdat->lru_lock);
1761
1762	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1763				     &nr_scanned, sc, isolate_mode, lru);
1764
1765	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1766	reclaim_stat->recent_scanned[file] += nr_taken;
1767
1768	if (current_is_kswapd()) {
1769		if (global_reclaim(sc))
1770			__count_vm_events(PGSCAN_KSWAPD, nr_scanned);
1771		count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD,
1772				   nr_scanned);
1773	} else {
1774		if (global_reclaim(sc))
1775			__count_vm_events(PGSCAN_DIRECT, nr_scanned);
1776		count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT,
1777				   nr_scanned);
1778	}
1779	spin_unlock_irq(&pgdat->lru_lock);
1780
1781	if (nr_taken == 0)
1782		return 0;
1783
1784	nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0,
1785				&stat, false);
1786
1787	spin_lock_irq(&pgdat->lru_lock);
1788
1789	if (current_is_kswapd()) {
1790		if (global_reclaim(sc))
1791			__count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
1792		count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD,
1793				   nr_reclaimed);
1794	} else {
1795		if (global_reclaim(sc))
1796			__count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
1797		count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT,
1798				   nr_reclaimed);
1799	}
1800
1801	putback_inactive_pages(lruvec, &page_list);
1802
1803	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1804
1805	spin_unlock_irq(&pgdat->lru_lock);
1806
1807	mem_cgroup_uncharge_list(&page_list);
1808	free_unref_page_list(&page_list);
1809
1810	/*
1811	 * If dirty pages are scanned that are not queued for IO, it
1812	 * implies that flushers are not doing their job. This can
1813	 * happen when memory pressure pushes dirty pages to the end of
1814	 * the LRU before the dirty limits are breached and the dirty
1815	 * data has expired. It can also happen when the proportion of
1816	 * dirty pages grows not through writes but through memory
1817	 * pressure reclaiming all the clean cache. And in some cases,
1818	 * the flushers simply cannot keep up with the allocation
1819	 * rate. Nudge the flusher threads in case they are asleep.
1820	 */
1821	if (stat.nr_unqueued_dirty == nr_taken)
1822		wakeup_flusher_threads(WB_REASON_VMSCAN);
1823
1824	sc->nr.dirty += stat.nr_dirty;
1825	sc->nr.congested += stat.nr_congested;
1826	sc->nr.unqueued_dirty += stat.nr_unqueued_dirty;
1827	sc->nr.writeback += stat.nr_writeback;
1828	sc->nr.immediate += stat.nr_immediate;
1829	sc->nr.taken += nr_taken;
1830	if (file)
1831		sc->nr.file_taken += nr_taken;
1832
1833	trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1834			nr_scanned, nr_reclaimed, &stat, sc->priority, file);
1835	return nr_reclaimed;
1836}
1837
1838/*
1839 * This moves pages from the active list to the inactive list.
1840 *
1841 * We move them the other way if the page is referenced by one or more
1842 * processes, from rmap.
1843 *
1844 * If the pages are mostly unmapped, the processing is fast and it is
1845 * appropriate to hold zone_lru_lock across the whole operation.  But if
1846 * the pages are mapped, the processing is slow (page_referenced()) so we
1847 * should drop zone_lru_lock around each page.  It's impossible to balance
1848 * this, so instead we remove the pages from the LRU while processing them.
1849 * It is safe to rely on PG_active against the non-LRU pages in here because
1850 * nobody will play with that bit on a non-LRU page.
1851 *
1852 * The downside is that we have to touch page->_refcount against each page.
1853 * But we had to alter page->flags anyway.
1854 *
1855 * Returns the number of pages moved to the given lru.
1856 */
1857
1858static unsigned move_active_pages_to_lru(struct lruvec *lruvec,
1859				     struct list_head *list,
1860				     struct list_head *pages_to_free,
1861				     enum lru_list lru)
1862{
1863	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1864	struct page *page;
1865	int nr_pages;
1866	int nr_moved = 0;
1867
1868	while (!list_empty(list)) {
1869		page = lru_to_page(list);
1870		lruvec = mem_cgroup_page_lruvec(page, pgdat);
1871
1872		VM_BUG_ON_PAGE(PageLRU(page), page);
1873		SetPageLRU(page);
1874
1875		nr_pages = hpage_nr_pages(page);
1876		update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
1877		list_move(&page->lru, &lruvec->lists[lru]);
1878
1879		if (put_page_testzero(page)) {
1880			__ClearPageLRU(page);
1881			__ClearPageActive(page);
1882			del_page_from_lru_list(page, lruvec, lru);
1883
1884			if (unlikely(PageCompound(page))) {
1885				spin_unlock_irq(&pgdat->lru_lock);
1886				mem_cgroup_uncharge(page);
1887				(*get_compound_page_dtor(page))(page);
1888				spin_lock_irq(&pgdat->lru_lock);
1889			} else
1890				list_add(&page->lru, pages_to_free);
1891		} else {
1892			nr_moved += nr_pages;
1893		}
1894	}
1895
1896	if (!is_active_lru(lru)) {
1897		__count_vm_events(PGDEACTIVATE, nr_moved);
1898		count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
1899				   nr_moved);
1900	}
1901
1902	return nr_moved;
1903}
1904
1905static void shrink_active_list(unsigned long nr_to_scan,
1906			       struct lruvec *lruvec,
1907			       struct scan_control *sc,
1908			       enum lru_list lru)
1909{
1910	unsigned long nr_taken;
1911	unsigned long nr_scanned;
1912	unsigned long vm_flags;
1913	LIST_HEAD(l_hold);	/* The pages which were snipped off */
1914	LIST_HEAD(l_active);
1915	LIST_HEAD(l_inactive);
1916	struct page *page;
1917	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1918	unsigned nr_deactivate, nr_activate;
1919	unsigned nr_rotated = 0;
1920	isolate_mode_t isolate_mode = 0;
1921	int file = is_file_lru(lru);
1922	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1923
1924	lru_add_drain();
1925
1926	if (!sc->may_unmap)
1927		isolate_mode |= ISOLATE_UNMAPPED;
1928
1929	spin_lock_irq(&pgdat->lru_lock);
1930
1931	nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1932				     &nr_scanned, sc, isolate_mode, lru);
1933
1934	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
1935	reclaim_stat->recent_scanned[file] += nr_taken;
1936
1937	__count_vm_events(PGREFILL, nr_scanned);
1938	count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned);
1939
1940	spin_unlock_irq(&pgdat->lru_lock);
1941
1942	while (!list_empty(&l_hold)) {
1943		cond_resched();
1944		page = lru_to_page(&l_hold);
1945		list_del(&page->lru);
1946
1947		if (unlikely(!page_evictable(page))) {
1948			putback_lru_page(page);
1949			continue;
1950		}
1951
1952		if (unlikely(buffer_heads_over_limit)) {
1953			if (page_has_private(page) && trylock_page(page)) {
1954				if (page_has_private(page))
1955					try_to_release_page(page, 0);
1956				unlock_page(page);
1957			}
1958		}
1959
1960		if (page_referenced(page, 0, sc->target_mem_cgroup,
1961				    &vm_flags)) {
1962			nr_rotated += hpage_nr_pages(page);
1963			/*
1964			 * Identify referenced, file-backed active pages and
1965			 * give them one more trip around the active list. So
1966			 * that executable code get better chances to stay in
1967			 * memory under moderate memory pressure.  Anon pages
1968			 * are not likely to be evicted by use-once streaming
1969			 * IO, plus JVM can create lots of anon VM_EXEC pages,
1970			 * so we ignore them here.
1971			 */
1972			if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1973				list_add(&page->lru, &l_active);
1974				continue;
1975			}
1976		}
1977
1978		ClearPageActive(page);	/* we are de-activating */
 
1979		list_add(&page->lru, &l_inactive);
1980	}
1981
1982	/*
1983	 * Move pages back to the lru list.
1984	 */
1985	spin_lock_irq(&pgdat->lru_lock);
1986	/*
1987	 * Count referenced pages from currently used mappings as rotated,
1988	 * even though only some of them are actually re-activated.  This
1989	 * helps balance scan pressure between file and anonymous pages in
1990	 * get_scan_count.
1991	 */
1992	reclaim_stat->recent_rotated[file] += nr_rotated;
1993
1994	nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1995	nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
 
 
 
 
 
 
1996	__mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1997	spin_unlock_irq(&pgdat->lru_lock);
1998
1999	mem_cgroup_uncharge_list(&l_hold);
2000	free_unref_page_list(&l_hold);
2001	trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate,
2002			nr_deactivate, nr_rotated, sc->priority, file);
2003}
2004
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2005/*
2006 * The inactive anon list should be small enough that the VM never has
2007 * to do too much work.
2008 *
2009 * The inactive file list should be small enough to leave most memory
2010 * to the established workingset on the scan-resistant active list,
2011 * but large enough to avoid thrashing the aggregate readahead window.
2012 *
2013 * Both inactive lists should also be large enough that each inactive
2014 * page has a chance to be referenced again before it is reclaimed.
2015 *
2016 * If that fails and refaulting is observed, the inactive list grows.
2017 *
2018 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
2019 * on this LRU, maintained by the pageout code. An inactive_ratio
2020 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
2021 *
2022 * total     target    max
2023 * memory    ratio     inactive
2024 * -------------------------------------
2025 *   10MB       1         5MB
2026 *  100MB       1        50MB
2027 *    1GB       3       250MB
2028 *   10GB      10       0.9GB
2029 *  100GB      31         3GB
2030 *    1TB     101        10GB
2031 *   10TB     320        32GB
2032 */
2033static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2034				 struct mem_cgroup *memcg,
2035				 struct scan_control *sc, bool actual_reclaim)
2036{
2037	enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE;
2038	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2039	enum lru_list inactive_lru = file * LRU_FILE;
2040	unsigned long inactive, active;
2041	unsigned long inactive_ratio;
2042	unsigned long refaults;
2043	unsigned long gb;
2044
2045	/*
2046	 * If we don't have swap space, anonymous page deactivation
2047	 * is pointless.
2048	 */
2049	if (!file && !total_swap_pages)
2050		return false;
2051
2052	inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx);
2053	active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx);
2054
2055	if (memcg)
2056		refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2057	else
2058		refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2059
2060	/*
2061	 * When refaults are being observed, it means a new workingset
2062	 * is being established. Disable active list protection to get
2063	 * rid of the stale workingset quickly.
2064	 */
2065	if (file && actual_reclaim && lruvec->refaults != refaults) {
 
2066		inactive_ratio = 0;
2067	} else {
2068		gb = (inactive + active) >> (30 - PAGE_SHIFT);
2069		if (gb)
2070			inactive_ratio = int_sqrt(10 * gb);
2071		else
2072			inactive_ratio = 1;
2073	}
2074
2075	if (actual_reclaim)
2076		trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,
2077			lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,
2078			lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active,
2079			inactive_ratio, file);
2080
2081	return inactive * inactive_ratio < active;
2082}
2083
2084static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
2085				 struct lruvec *lruvec, struct mem_cgroup *memcg,
2086				 struct scan_control *sc)
2087{
2088	if (is_active_lru(lru)) {
2089		if (inactive_list_is_low(lruvec, is_file_lru(lru),
2090					 memcg, sc, true))
2091			shrink_active_list(nr_to_scan, lruvec, sc, lru);
2092		return 0;
2093	}
2094
2095	return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
2096}
2097
2098enum scan_balance {
2099	SCAN_EQUAL,
2100	SCAN_FRACT,
2101	SCAN_ANON,
2102	SCAN_FILE,
2103};
2104
2105/*
2106 * Determine how aggressively the anon and file LRU lists should be
2107 * scanned.  The relative value of each set of LRU lists is determined
2108 * by looking at the fraction of the pages scanned we did rotate back
2109 * onto the active list instead of evict.
2110 *
2111 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2112 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
2113 */
2114static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
2115			   struct scan_control *sc, unsigned long *nr,
2116			   unsigned long *lru_pages)
2117{
2118	int swappiness = mem_cgroup_swappiness(memcg);
2119	struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2120	u64 fraction[2];
2121	u64 denominator = 0;	/* gcc */
2122	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2123	unsigned long anon_prio, file_prio;
2124	enum scan_balance scan_balance;
2125	unsigned long anon, file;
2126	unsigned long ap, fp;
2127	enum lru_list lru;
2128
2129	/* If we have no swap space, do not bother scanning anon pages. */
2130	if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
2131		scan_balance = SCAN_FILE;
2132		goto out;
2133	}
2134
2135	/*
2136	 * Global reclaim will swap to prevent OOM even with no
2137	 * swappiness, but memcg users want to use this knob to
2138	 * disable swapping for individual groups completely when
2139	 * using the memory controller's swap limit feature would be
2140	 * too expensive.
2141	 */
2142	if (!global_reclaim(sc) && !swappiness) {
2143		scan_balance = SCAN_FILE;
2144		goto out;
2145	}
2146
2147	/*
2148	 * Do not apply any pressure balancing cleverness when the
2149	 * system is close to OOM, scan both anon and file equally
2150	 * (unless the swappiness setting disagrees with swapping).
2151	 */
2152	if (!sc->priority && swappiness) {
2153		scan_balance = SCAN_EQUAL;
2154		goto out;
2155	}
2156
2157	/*
2158	 * Prevent the reclaimer from falling into the cache trap: as
2159	 * cache pages start out inactive, every cache fault will tip
2160	 * the scan balance towards the file LRU.  And as the file LRU
2161	 * shrinks, so does the window for rotation from references.
2162	 * This means we have a runaway feedback loop where a tiny
2163	 * thrashing file LRU becomes infinitely more attractive than
2164	 * anon pages.  Try to detect this based on file LRU size.
2165	 */
2166	if (global_reclaim(sc)) {
2167		unsigned long pgdatfile;
2168		unsigned long pgdatfree;
2169		int z;
2170		unsigned long total_high_wmark = 0;
2171
2172		pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2173		pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2174			   node_page_state(pgdat, NR_INACTIVE_FILE);
2175
2176		for (z = 0; z < MAX_NR_ZONES; z++) {
2177			struct zone *zone = &pgdat->node_zones[z];
2178			if (!managed_zone(zone))
2179				continue;
2180
2181			total_high_wmark += high_wmark_pages(zone);
2182		}
2183
2184		if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
2185			/*
2186			 * Force SCAN_ANON if there are enough inactive
2187			 * anonymous pages on the LRU in eligible zones.
2188			 * Otherwise, the small LRU gets thrashed.
2189			 */
2190			if (!inactive_list_is_low(lruvec, false, memcg, sc, false) &&
2191			    lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx)
2192					>> sc->priority) {
2193				scan_balance = SCAN_ANON;
2194				goto out;
2195			}
2196		}
2197	}
2198
2199	/*
2200	 * If there is enough inactive page cache, i.e. if the size of the
2201	 * inactive list is greater than that of the active list *and* the
2202	 * inactive list actually has some pages to scan on this priority, we
2203	 * do not reclaim anything from the anonymous working set right now.
2204	 * Without the second condition we could end up never scanning an
2205	 * lruvec even if it has plenty of old anonymous pages unless the
2206	 * system is under heavy pressure.
2207	 */
2208	if (!inactive_list_is_low(lruvec, true, memcg, sc, false) &&
2209	    lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) {
2210		scan_balance = SCAN_FILE;
2211		goto out;
2212	}
2213
2214	scan_balance = SCAN_FRACT;
2215
2216	/*
2217	 * With swappiness at 100, anonymous and file have the same priority.
2218	 * This scanning priority is essentially the inverse of IO cost.
2219	 */
2220	anon_prio = swappiness;
2221	file_prio = 200 - anon_prio;
2222
2223	/*
2224	 * OK, so we have swap space and a fair amount of page cache
2225	 * pages.  We use the recently rotated / recently scanned
2226	 * ratios to determine how valuable each cache is.
2227	 *
2228	 * Because workloads change over time (and to avoid overflow)
2229	 * we keep these statistics as a floating average, which ends
2230	 * up weighing recent references more than old ones.
2231	 *
2232	 * anon in [0], file in [1]
2233	 */
2234
2235	anon  = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) +
2236		lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES);
2237	file  = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) +
2238		lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES);
2239
2240	spin_lock_irq(&pgdat->lru_lock);
2241	if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
2242		reclaim_stat->recent_scanned[0] /= 2;
2243		reclaim_stat->recent_rotated[0] /= 2;
2244	}
2245
2246	if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
2247		reclaim_stat->recent_scanned[1] /= 2;
2248		reclaim_stat->recent_rotated[1] /= 2;
2249	}
2250
2251	/*
2252	 * The amount of pressure on anon vs file pages is inversely
2253	 * proportional to the fraction of recently scanned pages on
2254	 * each list that were recently referenced and in active use.
2255	 */
2256	ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
2257	ap /= reclaim_stat->recent_rotated[0] + 1;
2258
2259	fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
2260	fp /= reclaim_stat->recent_rotated[1] + 1;
2261	spin_unlock_irq(&pgdat->lru_lock);
2262
2263	fraction[0] = ap;
2264	fraction[1] = fp;
2265	denominator = ap + fp + 1;
2266out:
2267	*lru_pages = 0;
2268	for_each_evictable_lru(lru) {
2269		int file = is_file_lru(lru);
2270		unsigned long size;
2271		unsigned long scan;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2272
2273		size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx);
2274		scan = size >> sc->priority;
2275		/*
2276		 * If the cgroup's already been deleted, make sure to
2277		 * scrape out the remaining cache.
2278		 */
2279		if (!scan && !mem_cgroup_online(memcg))
2280			scan = min(size, SWAP_CLUSTER_MAX);
2281
2282		switch (scan_balance) {
2283		case SCAN_EQUAL:
2284			/* Scan lists relative to size */
2285			break;
2286		case SCAN_FRACT:
2287			/*
2288			 * Scan types proportional to swappiness and
2289			 * their relative recent reclaim efficiency.
 
 
2290			 */
2291			scan = div64_u64(scan * fraction[file],
2292					 denominator);
2293			break;
2294		case SCAN_FILE:
2295		case SCAN_ANON:
2296			/* Scan one type exclusively */
2297			if ((scan_balance == SCAN_FILE) != file) {
2298				size = 0;
2299				scan = 0;
2300			}
2301			break;
2302		default:
2303			/* Look ma, no brain */
2304			BUG();
2305		}
2306
2307		*lru_pages += size;
2308		nr[lru] = scan;
2309	}
2310}
2311
2312/*
2313 * This is a basic per-node page freer.  Used by both kswapd and direct reclaim.
2314 */
2315static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
2316			      struct scan_control *sc, unsigned long *lru_pages)
2317{
2318	struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
2319	unsigned long nr[NR_LRU_LISTS];
2320	unsigned long targets[NR_LRU_LISTS];
2321	unsigned long nr_to_scan;
2322	enum lru_list lru;
2323	unsigned long nr_reclaimed = 0;
2324	unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2325	struct blk_plug plug;
2326	bool scan_adjusted;
2327
2328	get_scan_count(lruvec, memcg, sc, nr, lru_pages);
2329
2330	/* Record the original scan target for proportional adjustments later */
2331	memcpy(targets, nr, sizeof(nr));
2332
2333	/*
2334	 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2335	 * event that can occur when there is little memory pressure e.g.
2336	 * multiple streaming readers/writers. Hence, we do not abort scanning
2337	 * when the requested number of pages are reclaimed when scanning at
2338	 * DEF_PRIORITY on the assumption that the fact we are direct
2339	 * reclaiming implies that kswapd is not keeping up and it is best to
2340	 * do a batch of work at once. For memcg reclaim one check is made to
2341	 * abort proportional reclaim if either the file or anon lru has already
2342	 * dropped to zero at the first pass.
2343	 */
2344	scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2345			 sc->priority == DEF_PRIORITY);
2346
2347	blk_start_plug(&plug);
2348	while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2349					nr[LRU_INACTIVE_FILE]) {
2350		unsigned long nr_anon, nr_file, percentage;
2351		unsigned long nr_scanned;
2352
2353		for_each_evictable_lru(lru) {
2354			if (nr[lru]) {
2355				nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2356				nr[lru] -= nr_to_scan;
2357
2358				nr_reclaimed += shrink_list(lru, nr_to_scan,
2359							    lruvec, memcg, sc);
2360			}
2361		}
2362
2363		cond_resched();
2364
2365		if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2366			continue;
2367
2368		/*
2369		 * For kswapd and memcg, reclaim at least the number of pages
2370		 * requested. Ensure that the anon and file LRUs are scanned
2371		 * proportionally what was requested by get_scan_count(). We
2372		 * stop reclaiming one LRU and reduce the amount scanning
2373		 * proportional to the original scan target.
2374		 */
2375		nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2376		nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2377
2378		/*
2379		 * It's just vindictive to attack the larger once the smaller
2380		 * has gone to zero.  And given the way we stop scanning the
2381		 * smaller below, this makes sure that we only make one nudge
2382		 * towards proportionality once we've got nr_to_reclaim.
2383		 */
2384		if (!nr_file || !nr_anon)
2385			break;
2386
2387		if (nr_file > nr_anon) {
2388			unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2389						targets[LRU_ACTIVE_ANON] + 1;
2390			lru = LRU_BASE;
2391			percentage = nr_anon * 100 / scan_target;
2392		} else {
2393			unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2394						targets[LRU_ACTIVE_FILE] + 1;
2395			lru = LRU_FILE;
2396			percentage = nr_file * 100 / scan_target;
2397		}
2398
2399		/* Stop scanning the smaller of the LRU */
2400		nr[lru] = 0;
2401		nr[lru + LRU_ACTIVE] = 0;
2402
2403		/*
2404		 * Recalculate the other LRU scan count based on its original
2405		 * scan target and the percentage scanning already complete
2406		 */
2407		lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2408		nr_scanned = targets[lru] - nr[lru];
2409		nr[lru] = targets[lru] * (100 - percentage) / 100;
2410		nr[lru] -= min(nr[lru], nr_scanned);
2411
2412		lru += LRU_ACTIVE;
2413		nr_scanned = targets[lru] - nr[lru];
2414		nr[lru] = targets[lru] * (100 - percentage) / 100;
2415		nr[lru] -= min(nr[lru], nr_scanned);
2416
2417		scan_adjusted = true;
2418	}
2419	blk_finish_plug(&plug);
2420	sc->nr_reclaimed += nr_reclaimed;
2421
2422	/*
2423	 * Even if we did not try to evict anon pages at all, we want to
2424	 * rebalance the anon lru active/inactive ratio.
2425	 */
2426	if (inactive_list_is_low(lruvec, false, memcg, sc, true))
2427		shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2428				   sc, LRU_ACTIVE_ANON);
2429}
2430
2431/* Use reclaim/compaction for costly allocs or under memory pressure */
2432static bool in_reclaim_compaction(struct scan_control *sc)
2433{
2434	if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2435			(sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2436			 sc->priority < DEF_PRIORITY - 2))
2437		return true;
2438
2439	return false;
2440}
2441
2442/*
2443 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2444 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2445 * true if more pages should be reclaimed such that when the page allocator
2446 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2447 * It will give up earlier than that if there is difficulty reclaiming pages.
2448 */
2449static inline bool should_continue_reclaim(struct pglist_data *pgdat,
2450					unsigned long nr_reclaimed,
2451					unsigned long nr_scanned,
2452					struct scan_control *sc)
2453{
2454	unsigned long pages_for_compaction;
2455	unsigned long inactive_lru_pages;
2456	int z;
2457
2458	/* If not in reclaim/compaction mode, stop */
2459	if (!in_reclaim_compaction(sc))
2460		return false;
2461
2462	/* Consider stopping depending on scan and reclaim activity */
2463	if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) {
2464		/*
2465		 * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the
2466		 * full LRU list has been scanned and we are still failing
2467		 * to reclaim pages. This full LRU scan is potentially
2468		 * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed
2469		 */
2470		if (!nr_reclaimed && !nr_scanned)
2471			return false;
2472	} else {
2473		/*
2474		 * For non-__GFP_RETRY_MAYFAIL allocations which can presumably
2475		 * fail without consequence, stop if we failed to reclaim
2476		 * any pages from the last SWAP_CLUSTER_MAX number of
2477		 * pages that were scanned. This will return to the
2478		 * caller faster at the risk reclaim/compaction and
2479		 * the resulting allocation attempt fails
2480		 */
2481		if (!nr_reclaimed)
2482			return false;
2483	}
2484
2485	/*
2486	 * If we have not reclaimed enough pages for compaction and the
2487	 * inactive lists are large enough, continue reclaiming
 
 
 
 
 
 
2488	 */
2489	pages_for_compaction = compact_gap(sc->order);
2490	inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
2491	if (get_nr_swap_pages() > 0)
2492		inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
2493	if (sc->nr_reclaimed < pages_for_compaction &&
2494			inactive_lru_pages > pages_for_compaction)
2495		return true;
2496
2497	/* If compaction would go ahead or the allocation would succeed, stop */
2498	for (z = 0; z <= sc->reclaim_idx; z++) {
2499		struct zone *zone = &pgdat->node_zones[z];
2500		if (!managed_zone(zone))
2501			continue;
2502
2503		switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
2504		case COMPACT_SUCCESS:
2505		case COMPACT_CONTINUE:
2506			return false;
2507		default:
2508			/* check next zone */
2509			;
2510		}
2511	}
2512	return true;
 
 
 
 
 
 
 
 
 
 
2513}
2514
2515static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg)
2516{
2517	return test_bit(PGDAT_CONGESTED, &pgdat->flags) ||
2518		(memcg && memcg_congested(pgdat, memcg));
2519}
2520
2521static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
2522{
2523	struct reclaim_state *reclaim_state = current->reclaim_state;
2524	unsigned long nr_reclaimed, nr_scanned;
2525	bool reclaimable = false;
2526
2527	do {
2528		struct mem_cgroup *root = sc->target_mem_cgroup;
2529		struct mem_cgroup_reclaim_cookie reclaim = {
2530			.pgdat = pgdat,
2531			.priority = sc->priority,
2532		};
2533		unsigned long node_lru_pages = 0;
2534		struct mem_cgroup *memcg;
2535
2536		memset(&sc->nr, 0, sizeof(sc->nr));
2537
2538		nr_reclaimed = sc->nr_reclaimed;
2539		nr_scanned = sc->nr_scanned;
2540
2541		memcg = mem_cgroup_iter(root, NULL, &reclaim);
2542		do {
2543			unsigned long lru_pages;
2544			unsigned long reclaimed;
2545			unsigned long scanned;
2546
2547			if (mem_cgroup_low(root, memcg)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
2548				if (!sc->memcg_low_reclaim) {
2549					sc->memcg_low_skipped = 1;
2550					continue;
2551				}
2552				memcg_memory_event(memcg, MEMCG_LOW);
 
 
 
 
 
 
 
 
 
 
2553			}
2554
2555			reclaimed = sc->nr_reclaimed;
2556			scanned = sc->nr_scanned;
2557			shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2558			node_lru_pages += lru_pages;
2559
2560			if (memcg)
2561				shrink_slab(sc->gfp_mask, pgdat->node_id,
2562					    memcg, sc->priority);
2563
2564			/* Record the group's reclaim efficiency */
2565			vmpressure(sc->gfp_mask, memcg, false,
2566				   sc->nr_scanned - scanned,
2567				   sc->nr_reclaimed - reclaimed);
2568
2569			/*
2570			 * Direct reclaim and kswapd have to scan all memory
2571			 * cgroups to fulfill the overall scan target for the
2572			 * node.
2573			 *
2574			 * Limit reclaim, on the other hand, only cares about
2575			 * nr_to_reclaim pages to be reclaimed and it will
2576			 * retry with decreasing priority if one round over the
2577			 * whole hierarchy is not sufficient.
2578			 */
2579			if (!global_reclaim(sc) &&
2580					sc->nr_reclaimed >= sc->nr_to_reclaim) {
2581				mem_cgroup_iter_break(root, memcg);
2582				break;
2583			}
2584		} while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
2585
2586		if (global_reclaim(sc))
2587			shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
2588				    sc->priority);
2589
2590		if (reclaim_state) {
2591			sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2592			reclaim_state->reclaimed_slab = 0;
2593		}
2594
2595		/* Record the subtree's reclaim efficiency */
2596		vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
2597			   sc->nr_scanned - nr_scanned,
2598			   sc->nr_reclaimed - nr_reclaimed);
2599
2600		if (sc->nr_reclaimed - nr_reclaimed)
2601			reclaimable = true;
2602
2603		if (current_is_kswapd()) {
2604			/*
2605			 * If reclaim is isolating dirty pages under writeback,
2606			 * it implies that the long-lived page allocation rate
2607			 * is exceeding the page laundering rate. Either the
2608			 * global limits are not being effective at throttling
2609			 * processes due to the page distribution throughout
2610			 * zones or there is heavy usage of a slow backing
2611			 * device. The only option is to throttle from reclaim
2612			 * context which is not ideal as there is no guarantee
2613			 * the dirtying process is throttled in the same way
2614			 * balance_dirty_pages() manages.
2615			 *
2616			 * Once a node is flagged PGDAT_WRITEBACK, kswapd will
2617			 * count the number of pages under pages flagged for
2618			 * immediate reclaim and stall if any are encountered
2619			 * in the nr_immediate check below.
2620			 */
2621			if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken)
2622				set_bit(PGDAT_WRITEBACK, &pgdat->flags);
2623
2624			/*
2625			 * Tag a node as congested if all the dirty pages
2626			 * scanned were backed by a congested BDI and
2627			 * wait_iff_congested will stall.
2628			 */
2629			if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2630				set_bit(PGDAT_CONGESTED, &pgdat->flags);
2631
2632			/* Allow kswapd to start writing pages during reclaim.*/
2633			if (sc->nr.unqueued_dirty == sc->nr.file_taken)
2634				set_bit(PGDAT_DIRTY, &pgdat->flags);
2635
2636			/*
2637			 * If kswapd scans pages marked marked for immediate
2638			 * reclaim and under writeback (nr_immediate), it
2639			 * implies that pages are cycling through the LRU
2640			 * faster than they are written so also forcibly stall.
2641			 */
2642			if (sc->nr.immediate)
2643				congestion_wait(BLK_RW_ASYNC, HZ/10);
2644		}
2645
2646		/*
2647		 * Legacy memcg will stall in page writeback so avoid forcibly
2648		 * stalling in wait_iff_congested().
2649		 */
2650		if (!global_reclaim(sc) && sane_reclaim(sc) &&
2651		    sc->nr.dirty && sc->nr.dirty == sc->nr.congested)
2652			set_memcg_congestion(pgdat, root, true);
2653
2654		/*
2655		 * Stall direct reclaim for IO completions if underlying BDIs
2656		 * and node is congested. Allow kswapd to continue until it
2657		 * starts encountering unqueued dirty pages or cycling through
2658		 * the LRU too quickly.
2659		 */
2660		if (!sc->hibernation_mode && !current_is_kswapd() &&
2661		   current_may_throttle() && pgdat_memcg_congested(pgdat, root))
2662			wait_iff_congested(BLK_RW_ASYNC, HZ/10);
2663
2664	} while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
2665					 sc->nr_scanned - nr_scanned, sc));
2666
2667	/*
2668	 * Kswapd gives up on balancing particular nodes after too
2669	 * many failures to reclaim anything from them and goes to
2670	 * sleep. On reclaim progress, reset the failure counter. A
2671	 * successful direct reclaim run will revive a dormant kswapd.
2672	 */
2673	if (reclaimable)
2674		pgdat->kswapd_failures = 0;
2675
2676	return reclaimable;
2677}
2678
2679/*
2680 * Returns true if compaction should go ahead for a costly-order request, or
2681 * the allocation would already succeed without compaction. Return false if we
2682 * should reclaim first.
2683 */
2684static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2685{
2686	unsigned long watermark;
2687	enum compact_result suitable;
2688
2689	suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2690	if (suitable == COMPACT_SUCCESS)
2691		/* Allocation should succeed already. Don't reclaim. */
2692		return true;
2693	if (suitable == COMPACT_SKIPPED)
2694		/* Compaction cannot yet proceed. Do reclaim. */
2695		return false;
2696
2697	/*
2698	 * Compaction is already possible, but it takes time to run and there
2699	 * are potentially other callers using the pages just freed. So proceed
2700	 * with reclaim to make a buffer of free pages available to give
2701	 * compaction a reasonable chance of completing and allocating the page.
2702	 * Note that we won't actually reclaim the whole buffer in one attempt
2703	 * as the target watermark in should_continue_reclaim() is lower. But if
2704	 * we are already above the high+gap watermark, don't reclaim at all.
2705	 */
2706	watermark = high_wmark_pages(zone) + compact_gap(sc->order);
2707
2708	return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
2709}
2710
2711/*
2712 * This is the direct reclaim path, for page-allocating processes.  We only
2713 * try to reclaim pages from zones which will satisfy the caller's allocation
2714 * request.
2715 *
2716 * If a zone is deemed to be full of pinned pages then just give it a light
2717 * scan then give up on it.
2718 */
2719static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2720{
2721	struct zoneref *z;
2722	struct zone *zone;
2723	unsigned long nr_soft_reclaimed;
2724	unsigned long nr_soft_scanned;
2725	gfp_t orig_mask;
2726	pg_data_t *last_pgdat = NULL;
2727
2728	/*
2729	 * If the number of buffer_heads in the machine exceeds the maximum
2730	 * allowed level, force direct reclaim to scan the highmem zone as
2731	 * highmem pages could be pinning lowmem pages storing buffer_heads
2732	 */
2733	orig_mask = sc->gfp_mask;
2734	if (buffer_heads_over_limit) {
2735		sc->gfp_mask |= __GFP_HIGHMEM;
2736		sc->reclaim_idx = gfp_zone(sc->gfp_mask);
2737	}
2738
2739	for_each_zone_zonelist_nodemask(zone, z, zonelist,
2740					sc->reclaim_idx, sc->nodemask) {
2741		/*
2742		 * Take care memory controller reclaiming has small influence
2743		 * to global LRU.
2744		 */
2745		if (global_reclaim(sc)) {
2746			if (!cpuset_zone_allowed(zone,
2747						 GFP_KERNEL | __GFP_HARDWALL))
2748				continue;
2749
2750			/*
2751			 * If we already have plenty of memory free for
2752			 * compaction in this zone, don't free any more.
2753			 * Even though compaction is invoked for any
2754			 * non-zero order, only frequent costly order
2755			 * reclamation is disruptive enough to become a
2756			 * noticeable problem, like transparent huge
2757			 * page allocations.
2758			 */
2759			if (IS_ENABLED(CONFIG_COMPACTION) &&
2760			    sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2761			    compaction_ready(zone, sc)) {
2762				sc->compaction_ready = true;
2763				continue;
2764			}
2765
2766			/*
2767			 * Shrink each node in the zonelist once. If the
2768			 * zonelist is ordered by zone (not the default) then a
2769			 * node may be shrunk multiple times but in that case
2770			 * the user prefers lower zones being preserved.
2771			 */
2772			if (zone->zone_pgdat == last_pgdat)
2773				continue;
2774
2775			/*
2776			 * This steals pages from memory cgroups over softlimit
2777			 * and returns the number of reclaimed pages and
2778			 * scanned pages. This works for global memory pressure
2779			 * and balancing, not for a memcg's limit.
2780			 */
2781			nr_soft_scanned = 0;
2782			nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
2783						sc->order, sc->gfp_mask,
2784						&nr_soft_scanned);
2785			sc->nr_reclaimed += nr_soft_reclaimed;
2786			sc->nr_scanned += nr_soft_scanned;
2787			/* need some check for avoid more shrink_zone() */
2788		}
2789
2790		/* See comment about same check for global reclaim above */
2791		if (zone->zone_pgdat == last_pgdat)
2792			continue;
2793		last_pgdat = zone->zone_pgdat;
2794		shrink_node(zone->zone_pgdat, sc);
2795	}
2796
2797	/*
2798	 * Restore to original mask to avoid the impact on the caller if we
2799	 * promoted it to __GFP_HIGHMEM.
2800	 */
2801	sc->gfp_mask = orig_mask;
2802}
2803
2804static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat)
2805{
2806	struct mem_cgroup *memcg;
2807
2808	memcg = mem_cgroup_iter(root_memcg, NULL, NULL);
2809	do {
2810		unsigned long refaults;
2811		struct lruvec *lruvec;
2812
2813		if (memcg)
2814			refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE);
2815		else
2816			refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE);
2817
2818		lruvec = mem_cgroup_lruvec(pgdat, memcg);
 
2819		lruvec->refaults = refaults;
2820	} while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL)));
2821}
2822
2823/*
2824 * This is the main entry point to direct page reclaim.
2825 *
2826 * If a full scan of the inactive list fails to free enough memory then we
2827 * are "out of memory" and something needs to be killed.
2828 *
2829 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2830 * high - the zone may be full of dirty or under-writeback pages, which this
2831 * caller can't do much about.  We kick the writeback threads and take explicit
2832 * naps in the hope that some of these pages can be written.  But if the
2833 * allocating task holds filesystem locks which prevent writeout this might not
2834 * work, and the allocation attempt will fail.
2835 *
2836 * returns:	0, if no pages reclaimed
2837 * 		else, the number of pages reclaimed
2838 */
2839static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2840					  struct scan_control *sc)
2841{
2842	int initial_priority = sc->priority;
2843	pg_data_t *last_pgdat;
2844	struct zoneref *z;
2845	struct zone *zone;
2846retry:
2847	delayacct_freepages_start();
2848
2849	if (global_reclaim(sc))
2850		__count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
2851
2852	do {
2853		vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2854				sc->priority);
2855		sc->nr_scanned = 0;
2856		shrink_zones(zonelist, sc);
2857
2858		if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2859			break;
2860
2861		if (sc->compaction_ready)
2862			break;
2863
2864		/*
2865		 * If we're getting trouble reclaiming, start doing
2866		 * writepage even in laptop mode.
2867		 */
2868		if (sc->priority < DEF_PRIORITY - 2)
2869			sc->may_writepage = 1;
2870	} while (--sc->priority >= 0);
2871
2872	last_pgdat = NULL;
2873	for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx,
2874					sc->nodemask) {
2875		if (zone->zone_pgdat == last_pgdat)
2876			continue;
2877		last_pgdat = zone->zone_pgdat;
2878		snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat);
2879		set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false);
2880	}
2881
2882	delayacct_freepages_end();
2883
2884	if (sc->nr_reclaimed)
2885		return sc->nr_reclaimed;
2886
2887	/* Aborted reclaim to try compaction? don't OOM, then */
2888	if (sc->compaction_ready)
2889		return 1;
2890
2891	/* Untapped cgroup reserves?  Don't OOM, retry. */
2892	if (sc->memcg_low_skipped) {
2893		sc->priority = initial_priority;
2894		sc->memcg_low_reclaim = 1;
2895		sc->memcg_low_skipped = 0;
2896		goto retry;
2897	}
2898
2899	return 0;
2900}
2901
2902static bool allow_direct_reclaim(pg_data_t *pgdat)
2903{
2904	struct zone *zone;
2905	unsigned long pfmemalloc_reserve = 0;
2906	unsigned long free_pages = 0;
2907	int i;
2908	bool wmark_ok;
2909
2910	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
2911		return true;
2912
2913	for (i = 0; i <= ZONE_NORMAL; i++) {
2914		zone = &pgdat->node_zones[i];
2915		if (!managed_zone(zone))
2916			continue;
2917
2918		if (!zone_reclaimable_pages(zone))
2919			continue;
2920
2921		pfmemalloc_reserve += min_wmark_pages(zone);
2922		free_pages += zone_page_state(zone, NR_FREE_PAGES);
2923	}
2924
2925	/* If there are no reserves (unexpected config) then do not throttle */
2926	if (!pfmemalloc_reserve)
2927		return true;
2928
2929	wmark_ok = free_pages > pfmemalloc_reserve / 2;
2930
2931	/* kswapd must be awake if processes are being throttled */
2932	if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2933		pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
2934						(enum zone_type)ZONE_NORMAL);
2935		wake_up_interruptible(&pgdat->kswapd_wait);
2936	}
2937
2938	return wmark_ok;
2939}
2940
2941/*
2942 * Throttle direct reclaimers if backing storage is backed by the network
2943 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2944 * depleted. kswapd will continue to make progress and wake the processes
2945 * when the low watermark is reached.
2946 *
2947 * Returns true if a fatal signal was delivered during throttling. If this
2948 * happens, the page allocator should not consider triggering the OOM killer.
2949 */
2950static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2951					nodemask_t *nodemask)
2952{
2953	struct zoneref *z;
2954	struct zone *zone;
2955	pg_data_t *pgdat = NULL;
2956
2957	/*
2958	 * Kernel threads should not be throttled as they may be indirectly
2959	 * responsible for cleaning pages necessary for reclaim to make forward
2960	 * progress. kjournald for example may enter direct reclaim while
2961	 * committing a transaction where throttling it could forcing other
2962	 * processes to block on log_wait_commit().
2963	 */
2964	if (current->flags & PF_KTHREAD)
2965		goto out;
2966
2967	/*
2968	 * If a fatal signal is pending, this process should not throttle.
2969	 * It should return quickly so it can exit and free its memory
2970	 */
2971	if (fatal_signal_pending(current))
2972		goto out;
2973
2974	/*
2975	 * Check if the pfmemalloc reserves are ok by finding the first node
2976	 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2977	 * GFP_KERNEL will be required for allocating network buffers when
2978	 * swapping over the network so ZONE_HIGHMEM is unusable.
2979	 *
2980	 * Throttling is based on the first usable node and throttled processes
2981	 * wait on a queue until kswapd makes progress and wakes them. There
2982	 * is an affinity then between processes waking up and where reclaim
2983	 * progress has been made assuming the process wakes on the same node.
2984	 * More importantly, processes running on remote nodes will not compete
2985	 * for remote pfmemalloc reserves and processes on different nodes
2986	 * should make reasonable progress.
2987	 */
2988	for_each_zone_zonelist_nodemask(zone, z, zonelist,
2989					gfp_zone(gfp_mask), nodemask) {
2990		if (zone_idx(zone) > ZONE_NORMAL)
2991			continue;
2992
2993		/* Throttle based on the first usable node */
2994		pgdat = zone->zone_pgdat;
2995		if (allow_direct_reclaim(pgdat))
2996			goto out;
2997		break;
2998	}
2999
3000	/* If no zone was usable by the allocation flags then do not throttle */
3001	if (!pgdat)
3002		goto out;
3003
3004	/* Account for the throttling */
3005	count_vm_event(PGSCAN_DIRECT_THROTTLE);
3006
3007	/*
3008	 * If the caller cannot enter the filesystem, it's possible that it
3009	 * is due to the caller holding an FS lock or performing a journal
3010	 * transaction in the case of a filesystem like ext[3|4]. In this case,
3011	 * it is not safe to block on pfmemalloc_wait as kswapd could be
3012	 * blocked waiting on the same lock. Instead, throttle for up to a
3013	 * second before continuing.
3014	 */
3015	if (!(gfp_mask & __GFP_FS)) {
3016		wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
3017			allow_direct_reclaim(pgdat), HZ);
3018
3019		goto check_pending;
3020	}
3021
3022	/* Throttle until kswapd wakes the process */
3023	wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
3024		allow_direct_reclaim(pgdat));
3025
3026check_pending:
3027	if (fatal_signal_pending(current))
3028		return true;
3029
3030out:
3031	return false;
3032}
3033
3034unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
3035				gfp_t gfp_mask, nodemask_t *nodemask)
3036{
3037	unsigned long nr_reclaimed;
3038	struct scan_control sc = {
3039		.nr_to_reclaim = SWAP_CLUSTER_MAX,
3040		.gfp_mask = current_gfp_context(gfp_mask),
3041		.reclaim_idx = gfp_zone(gfp_mask),
3042		.order = order,
3043		.nodemask = nodemask,
3044		.priority = DEF_PRIORITY,
3045		.may_writepage = !laptop_mode,
3046		.may_unmap = 1,
3047		.may_swap = 1,
3048	};
3049
3050	/*
 
 
 
 
 
 
 
 
3051	 * Do not enter reclaim if fatal signal was delivered while throttled.
3052	 * 1 is returned so that the page allocator does not OOM kill at this
3053	 * point.
3054	 */
3055	if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
3056		return 1;
3057
3058	trace_mm_vmscan_direct_reclaim_begin(order,
3059				sc.may_writepage,
3060				sc.gfp_mask,
3061				sc.reclaim_idx);
3062
3063	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3064
3065	trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
 
3066
3067	return nr_reclaimed;
3068}
3069
3070#ifdef CONFIG_MEMCG
3071
 
3072unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
3073						gfp_t gfp_mask, bool noswap,
3074						pg_data_t *pgdat,
3075						unsigned long *nr_scanned)
3076{
3077	struct scan_control sc = {
3078		.nr_to_reclaim = SWAP_CLUSTER_MAX,
3079		.target_mem_cgroup = memcg,
3080		.may_writepage = !laptop_mode,
3081		.may_unmap = 1,
3082		.reclaim_idx = MAX_NR_ZONES - 1,
3083		.may_swap = !noswap,
3084	};
3085	unsigned long lru_pages;
3086
 
 
3087	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3088			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
3089
3090	trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
3091						      sc.may_writepage,
3092						      sc.gfp_mask,
3093						      sc.reclaim_idx);
3094
3095	/*
3096	 * NOTE: Although we can get the priority field, using it
3097	 * here is not a good idea, since it limits the pages we can scan.
3098	 * if we don't reclaim here, the shrink_node from balance_pgdat
3099	 * will pick up pages from other mem cgroup's as well. We hack
3100	 * the priority and make it zero.
3101	 */
3102	shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
3103
3104	trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3105
3106	*nr_scanned = sc.nr_scanned;
 
3107	return sc.nr_reclaimed;
3108}
3109
3110unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
3111					   unsigned long nr_pages,
3112					   gfp_t gfp_mask,
3113					   bool may_swap)
3114{
3115	struct zonelist *zonelist;
3116	unsigned long nr_reclaimed;
 
3117	int nid;
3118	unsigned int noreclaim_flag;
3119	struct scan_control sc = {
3120		.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3121		.gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) |
3122				(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
3123		.reclaim_idx = MAX_NR_ZONES - 1,
3124		.target_mem_cgroup = memcg,
3125		.priority = DEF_PRIORITY,
3126		.may_writepage = !laptop_mode,
3127		.may_unmap = 1,
3128		.may_swap = may_swap,
3129	};
3130
 
3131	/*
3132	 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3133	 * take care of from where we get pages. So the node where we start the
3134	 * scan does not need to be the current node.
3135	 */
3136	nid = mem_cgroup_select_victim_node(memcg);
3137
3138	zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
3139
3140	trace_mm_vmscan_memcg_reclaim_begin(0,
3141					    sc.may_writepage,
3142					    sc.gfp_mask,
3143					    sc.reclaim_idx);
3144
 
3145	noreclaim_flag = memalloc_noreclaim_save();
 
3146	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
 
3147	memalloc_noreclaim_restore(noreclaim_flag);
 
3148
3149	trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
 
3150
3151	return nr_reclaimed;
3152}
3153#endif
3154
3155static void age_active_anon(struct pglist_data *pgdat,
3156				struct scan_control *sc)
3157{
3158	struct mem_cgroup *memcg;
3159
3160	if (!total_swap_pages)
3161		return;
3162
3163	memcg = mem_cgroup_iter(NULL, NULL, NULL);
3164	do {
3165		struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
3166
3167		if (inactive_list_is_low(lruvec, false, memcg, sc, true))
3168			shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
3169					   sc, LRU_ACTIVE_ANON);
3170
3171		memcg = mem_cgroup_iter(NULL, memcg, NULL);
3172	} while (memcg);
3173}
3174
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3175/*
3176 * Returns true if there is an eligible zone balanced for the request order
3177 * and classzone_idx
3178 */
3179static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
3180{
3181	int i;
3182	unsigned long mark = -1;
3183	struct zone *zone;
3184
 
 
 
 
3185	for (i = 0; i <= classzone_idx; i++) {
3186		zone = pgdat->node_zones + i;
3187
3188		if (!managed_zone(zone))
3189			continue;
3190
3191		mark = high_wmark_pages(zone);
3192		if (zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3193			return true;
3194	}
3195
3196	/*
3197	 * If a node has no populated zone within classzone_idx, it does not
3198	 * need balancing by definition. This can happen if a zone-restricted
3199	 * allocation tries to wake a remote kswapd.
3200	 */
3201	if (mark == -1)
3202		return true;
3203
3204	return false;
3205}
3206
3207/* Clear pgdat state for congested, dirty or under writeback. */
3208static void clear_pgdat_congested(pg_data_t *pgdat)
3209{
3210	clear_bit(PGDAT_CONGESTED, &pgdat->flags);
3211	clear_bit(PGDAT_DIRTY, &pgdat->flags);
3212	clear_bit(PGDAT_WRITEBACK, &pgdat->flags);
3213}
3214
3215/*
3216 * Prepare kswapd for sleeping. This verifies that there are no processes
3217 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3218 *
3219 * Returns true if kswapd is ready to sleep
3220 */
3221static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3222{
3223	/*
3224	 * The throttled processes are normally woken up in balance_pgdat() as
3225	 * soon as allow_direct_reclaim() is true. But there is a potential
3226	 * race between when kswapd checks the watermarks and a process gets
3227	 * throttled. There is also a potential race if processes get
3228	 * throttled, kswapd wakes, a large process exits thereby balancing the
3229	 * zones, which causes kswapd to exit balance_pgdat() before reaching
3230	 * the wake up checks. If kswapd is going to sleep, no process should
3231	 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3232	 * the wake up is premature, processes will wake kswapd and get
3233	 * throttled again. The difference from wake ups in balance_pgdat() is
3234	 * that here we are under prepare_to_wait().
3235	 */
3236	if (waitqueue_active(&pgdat->pfmemalloc_wait))
3237		wake_up_all(&pgdat->pfmemalloc_wait);
3238
3239	/* Hopeless node, leave it to direct reclaim */
3240	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
3241		return true;
3242
3243	if (pgdat_balanced(pgdat, order, classzone_idx)) {
3244		clear_pgdat_congested(pgdat);
3245		return true;
3246	}
3247
3248	return false;
3249}
3250
3251/*
3252 * kswapd shrinks a node of pages that are at or below the highest usable
3253 * zone that is currently unbalanced.
3254 *
3255 * Returns true if kswapd scanned at least the requested number of pages to
3256 * reclaim or if the lack of progress was due to pages under writeback.
3257 * This is used to determine if the scanning priority needs to be raised.
3258 */
3259static bool kswapd_shrink_node(pg_data_t *pgdat,
3260			       struct scan_control *sc)
3261{
3262	struct zone *zone;
3263	int z;
3264
3265	/* Reclaim a number of pages proportional to the number of zones */
3266	sc->nr_to_reclaim = 0;
3267	for (z = 0; z <= sc->reclaim_idx; z++) {
3268		zone = pgdat->node_zones + z;
3269		if (!managed_zone(zone))
3270			continue;
3271
3272		sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3273	}
3274
3275	/*
3276	 * Historically care was taken to put equal pressure on all zones but
3277	 * now pressure is applied based on node LRU order.
3278	 */
3279	shrink_node(pgdat, sc);
3280
3281	/*
3282	 * Fragmentation may mean that the system cannot be rebalanced for
3283	 * high-order allocations. If twice the allocation size has been
3284	 * reclaimed then recheck watermarks only at order-0 to prevent
3285	 * excessive reclaim. Assume that a process requested a high-order
3286	 * can direct reclaim/compact.
3287	 */
3288	if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
3289		sc->order = 0;
3290
3291	return sc->nr_scanned >= sc->nr_to_reclaim;
3292}
3293
3294/*
3295 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3296 * that are eligible for use by the caller until at least one zone is
3297 * balanced.
3298 *
3299 * Returns the order kswapd finished reclaiming at.
3300 *
3301 * kswapd scans the zones in the highmem->normal->dma direction.  It skips
3302 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3303 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3304 * or lower is eligible for reclaim until at least one usable zone is
3305 * balanced.
3306 */
3307static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
3308{
3309	int i;
3310	unsigned long nr_soft_reclaimed;
3311	unsigned long nr_soft_scanned;
 
 
 
 
3312	struct zone *zone;
3313	struct scan_control sc = {
3314		.gfp_mask = GFP_KERNEL,
3315		.order = order,
3316		.priority = DEF_PRIORITY,
3317		.may_writepage = !laptop_mode,
3318		.may_unmap = 1,
3319		.may_swap = 1,
3320	};
 
 
 
 
 
3321	count_vm_event(PAGEOUTRUN);
3322
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3323	do {
3324		unsigned long nr_reclaimed = sc.nr_reclaimed;
3325		bool raise_priority = true;
 
 
3326
3327		sc.reclaim_idx = classzone_idx;
3328
3329		/*
3330		 * If the number of buffer_heads exceeds the maximum allowed
3331		 * then consider reclaiming from all zones. This has a dual
3332		 * purpose -- on 64-bit systems it is expected that
3333		 * buffer_heads are stripped during active rotation. On 32-bit
3334		 * systems, highmem pages can pin lowmem memory and shrinking
3335		 * buffers can relieve lowmem pressure. Reclaim may still not
3336		 * go ahead if all eligible zones for the original allocation
3337		 * request are balanced to avoid excessive reclaim from kswapd.
3338		 */
3339		if (buffer_heads_over_limit) {
3340			for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3341				zone = pgdat->node_zones + i;
3342				if (!managed_zone(zone))
3343					continue;
3344
3345				sc.reclaim_idx = i;
3346				break;
3347			}
3348		}
3349
3350		/*
3351		 * Only reclaim if there are no eligible zones. Note that
3352		 * sc.reclaim_idx is not used as buffer_heads_over_limit may
3353		 * have adjusted it.
 
 
 
 
 
 
 
 
 
 
 
 
 
3354		 */
3355		if (pgdat_balanced(pgdat, sc.order, classzone_idx))
3356			goto out;
3357
 
 
 
 
 
 
 
 
 
 
 
 
 
3358		/*
3359		 * Do some background aging of the anon list, to give
3360		 * pages a chance to be referenced before reclaiming. All
3361		 * pages are rotated regardless of classzone as this is
3362		 * about consistent aging.
3363		 */
3364		age_active_anon(pgdat, &sc);
3365
3366		/*
3367		 * If we're getting trouble reclaiming, start doing writepage
3368		 * even in laptop mode.
3369		 */
3370		if (sc.priority < DEF_PRIORITY - 2)
3371			sc.may_writepage = 1;
3372
3373		/* Call soft limit reclaim before calling shrink_node. */
3374		sc.nr_scanned = 0;
3375		nr_soft_scanned = 0;
3376		nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
3377						sc.gfp_mask, &nr_soft_scanned);
3378		sc.nr_reclaimed += nr_soft_reclaimed;
3379
3380		/*
3381		 * There should be no need to raise the scanning priority if
3382		 * enough pages are already being scanned that that high
3383		 * watermark would be met at 100% efficiency.
3384		 */
3385		if (kswapd_shrink_node(pgdat, &sc))
3386			raise_priority = false;
3387
3388		/*
3389		 * If the low watermark is met there is no need for processes
3390		 * to be throttled on pfmemalloc_wait as they should not be
3391		 * able to safely make forward progress. Wake them
3392		 */
3393		if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3394				allow_direct_reclaim(pgdat))
3395			wake_up_all(&pgdat->pfmemalloc_wait);
3396
3397		/* Check if kswapd should be suspending */
3398		if (try_to_freeze() || kthread_should_stop())
 
 
 
3399			break;
3400
3401		/*
3402		 * Raise priority if scanning rate is too low or there was no
3403		 * progress in reclaiming pages
3404		 */
3405		nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
 
 
 
 
 
 
 
 
 
 
3406		if (raise_priority || !nr_reclaimed)
3407			sc.priority--;
3408	} while (sc.priority >= 1);
3409
3410	if (!sc.nr_reclaimed)
3411		pgdat->kswapd_failures++;
3412
3413out:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3414	snapshot_refaults(NULL, pgdat);
 
 
 
 
3415	/*
3416	 * Return the order kswapd stopped reclaiming at as
3417	 * prepare_kswapd_sleep() takes it into account. If another caller
3418	 * entered the allocator slow path while kswapd was awake, order will
3419	 * remain at the higher level.
3420	 */
3421	return sc.order;
3422}
3423
3424/*
3425 * pgdat->kswapd_classzone_idx is the highest zone index that a recent
3426 * allocation request woke kswapd for. When kswapd has not woken recently,
3427 * the value is MAX_NR_ZONES which is not a valid index. This compares a
3428 * given classzone and returns it or the highest classzone index kswapd
3429 * was recently woke for.
3430 */
3431static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat,
3432					   enum zone_type classzone_idx)
3433{
3434	if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
3435		return classzone_idx;
3436
3437	return max(pgdat->kswapd_classzone_idx, classzone_idx);
3438}
3439
3440static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3441				unsigned int classzone_idx)
3442{
3443	long remaining = 0;
3444	DEFINE_WAIT(wait);
3445
3446	if (freezing(current) || kthread_should_stop())
3447		return;
3448
3449	prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3450
3451	/*
3452	 * Try to sleep for a short interval. Note that kcompactd will only be
3453	 * woken if it is possible to sleep for a short interval. This is
3454	 * deliberate on the assumption that if reclaim cannot keep an
3455	 * eligible zone balanced that it's also unlikely that compaction will
3456	 * succeed.
3457	 */
3458	if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3459		/*
3460		 * Compaction records what page blocks it recently failed to
3461		 * isolate pages from and skips them in the future scanning.
3462		 * When kswapd is going to sleep, it is reasonable to assume
3463		 * that pages and compaction may succeed so reset the cache.
3464		 */
3465		reset_isolation_suitable(pgdat);
3466
3467		/*
3468		 * We have freed the memory, now we should compact it to make
3469		 * allocation of the requested order possible.
3470		 */
3471		wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
3472
3473		remaining = schedule_timeout(HZ/10);
3474
3475		/*
3476		 * If woken prematurely then reset kswapd_classzone_idx and
3477		 * order. The values will either be from a wakeup request or
3478		 * the previous request that slept prematurely.
3479		 */
3480		if (remaining) {
3481			pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3482			pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3483		}
3484
3485		finish_wait(&pgdat->kswapd_wait, &wait);
3486		prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3487	}
3488
3489	/*
3490	 * After a short sleep, check if it was a premature sleep. If not, then
3491	 * go fully to sleep until explicitly woken up.
3492	 */
3493	if (!remaining &&
3494	    prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
3495		trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3496
3497		/*
3498		 * vmstat counters are not perfectly accurate and the estimated
3499		 * value for counters such as NR_FREE_PAGES can deviate from the
3500		 * true value by nr_online_cpus * threshold. To avoid the zone
3501		 * watermarks being breached while under pressure, we reduce the
3502		 * per-cpu vmstat threshold while kswapd is awake and restore
3503		 * them before going back to sleep.
3504		 */
3505		set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3506
3507		if (!kthread_should_stop())
3508			schedule();
3509
3510		set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3511	} else {
3512		if (remaining)
3513			count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3514		else
3515			count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3516	}
3517	finish_wait(&pgdat->kswapd_wait, &wait);
3518}
3519
3520/*
3521 * The background pageout daemon, started as a kernel thread
3522 * from the init process.
3523 *
3524 * This basically trickles out pages so that we have _some_
3525 * free memory available even if there is no other activity
3526 * that frees anything up. This is needed for things like routing
3527 * etc, where we otherwise might have all activity going on in
3528 * asynchronous contexts that cannot page things out.
3529 *
3530 * If there are applications that are active memory-allocators
3531 * (most normal use), this basically shouldn't matter.
3532 */
3533static int kswapd(void *p)
3534{
3535	unsigned int alloc_order, reclaim_order;
3536	unsigned int classzone_idx = MAX_NR_ZONES - 1;
3537	pg_data_t *pgdat = (pg_data_t*)p;
3538	struct task_struct *tsk = current;
3539
3540	struct reclaim_state reclaim_state = {
3541		.reclaimed_slab = 0,
3542	};
3543	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3544
3545	if (!cpumask_empty(cpumask))
3546		set_cpus_allowed_ptr(tsk, cpumask);
3547	current->reclaim_state = &reclaim_state;
3548
3549	/*
3550	 * Tell the memory management that we're a "memory allocator",
3551	 * and that if we need more memory we should get access to it
3552	 * regardless (see "__alloc_pages()"). "kswapd" should
3553	 * never get caught in the normal page freeing logic.
3554	 *
3555	 * (Kswapd normally doesn't need memory anyway, but sometimes
3556	 * you need a small amount of memory in order to be able to
3557	 * page out something else, and this flag essentially protects
3558	 * us from recursively trying to free more memory as we're
3559	 * trying to free the first piece of memory in the first place).
3560	 */
3561	tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3562	set_freezable();
3563
3564	pgdat->kswapd_order = 0;
3565	pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3566	for ( ; ; ) {
3567		bool ret;
3568
3569		alloc_order = reclaim_order = pgdat->kswapd_order;
3570		classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx);
3571
3572kswapd_try_sleep:
3573		kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3574					classzone_idx);
3575
3576		/* Read the new order and classzone_idx */
3577		alloc_order = reclaim_order = pgdat->kswapd_order;
3578		classzone_idx = kswapd_classzone_idx(pgdat, 0);
3579		pgdat->kswapd_order = 0;
3580		pgdat->kswapd_classzone_idx = MAX_NR_ZONES;
3581
3582		ret = try_to_freeze();
3583		if (kthread_should_stop())
3584			break;
3585
3586		/*
3587		 * We can speed up thawing tasks if we don't call balance_pgdat
3588		 * after returning from the refrigerator
3589		 */
3590		if (ret)
3591			continue;
3592
3593		/*
3594		 * Reclaim begins at the requested order but if a high-order
3595		 * reclaim fails then kswapd falls back to reclaiming for
3596		 * order-0. If that happens, kswapd will consider sleeping
3597		 * for the order it finished reclaiming at (reclaim_order)
3598		 * but kcompactd is woken to compact for the original
3599		 * request (alloc_order).
3600		 */
3601		trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3602						alloc_order);
3603		fs_reclaim_acquire(GFP_KERNEL);
3604		reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3605		fs_reclaim_release(GFP_KERNEL);
3606		if (reclaim_order < alloc_order)
3607			goto kswapd_try_sleep;
3608	}
3609
3610	tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
3611	current->reclaim_state = NULL;
3612
3613	return 0;
3614}
3615
3616/*
3617 * A zone is low on free memory or too fragmented for high-order memory.  If
3618 * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's
3619 * pgdat.  It will wake up kcompactd after reclaiming memory.  If kswapd reclaim
3620 * has failed or is not needed, still wake up kcompactd if only compaction is
3621 * needed.
3622 */
3623void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order,
3624		   enum zone_type classzone_idx)
3625{
3626	pg_data_t *pgdat;
3627
3628	if (!managed_zone(zone))
3629		return;
3630
3631	if (!cpuset_zone_allowed(zone, gfp_flags))
3632		return;
3633	pgdat = zone->zone_pgdat;
3634	pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat,
3635							   classzone_idx);
 
 
 
 
3636	pgdat->kswapd_order = max(pgdat->kswapd_order, order);
3637	if (!waitqueue_active(&pgdat->kswapd_wait))
3638		return;
3639
3640	/* Hopeless node, leave it to direct reclaim if possible */
3641	if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ||
3642	    pgdat_balanced(pgdat, order, classzone_idx)) {
 
3643		/*
3644		 * There may be plenty of free memory available, but it's too
3645		 * fragmented for high-order allocations.  Wake up kcompactd
3646		 * and rely on compaction_suitable() to determine if it's
3647		 * needed.  If it fails, it will defer subsequent attempts to
3648		 * ratelimit its work.
3649		 */
3650		if (!(gfp_flags & __GFP_DIRECT_RECLAIM))
3651			wakeup_kcompactd(pgdat, order, classzone_idx);
3652		return;
3653	}
3654
3655	trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order,
3656				      gfp_flags);
3657	wake_up_interruptible(&pgdat->kswapd_wait);
3658}
3659
3660#ifdef CONFIG_HIBERNATION
3661/*
3662 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3663 * freed pages.
3664 *
3665 * Rather than trying to age LRUs the aim is to preserve the overall
3666 * LRU order by reclaiming preferentially
3667 * inactive > active > active referenced > active mapped
3668 */
3669unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3670{
3671	struct reclaim_state reclaim_state;
3672	struct scan_control sc = {
3673		.nr_to_reclaim = nr_to_reclaim,
3674		.gfp_mask = GFP_HIGHUSER_MOVABLE,
3675		.reclaim_idx = MAX_NR_ZONES - 1,
3676		.priority = DEF_PRIORITY,
3677		.may_writepage = 1,
3678		.may_unmap = 1,
3679		.may_swap = 1,
3680		.hibernation_mode = 1,
3681	};
3682	struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3683	struct task_struct *p = current;
3684	unsigned long nr_reclaimed;
3685	unsigned int noreclaim_flag;
3686
 
3687	noreclaim_flag = memalloc_noreclaim_save();
3688	fs_reclaim_acquire(sc.gfp_mask);
3689	reclaim_state.reclaimed_slab = 0;
3690	p->reclaim_state = &reclaim_state;
3691
3692	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
3693
3694	p->reclaim_state = NULL;
 
3695	fs_reclaim_release(sc.gfp_mask);
3696	memalloc_noreclaim_restore(noreclaim_flag);
3697
3698	return nr_reclaimed;
3699}
3700#endif /* CONFIG_HIBERNATION */
3701
3702/* It's optimal to keep kswapds on the same CPUs as their memory, but
3703   not required for correctness.  So if the last cpu in a node goes
3704   away, we get changed to run anywhere: as the first one comes back,
3705   restore their cpu bindings. */
3706static int kswapd_cpu_online(unsigned int cpu)
3707{
3708	int nid;
3709
3710	for_each_node_state(nid, N_MEMORY) {
3711		pg_data_t *pgdat = NODE_DATA(nid);
3712		const struct cpumask *mask;
3713
3714		mask = cpumask_of_node(pgdat->node_id);
3715
3716		if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3717			/* One of our CPUs online: restore mask */
3718			set_cpus_allowed_ptr(pgdat->kswapd, mask);
3719	}
3720	return 0;
3721}
3722
3723/*
3724 * This kswapd start function will be called by init and node-hot-add.
3725 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3726 */
3727int kswapd_run(int nid)
3728{
3729	pg_data_t *pgdat = NODE_DATA(nid);
3730	int ret = 0;
3731
3732	if (pgdat->kswapd)
3733		return 0;
3734
3735	pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3736	if (IS_ERR(pgdat->kswapd)) {
3737		/* failure at boot is fatal */
3738		BUG_ON(system_state < SYSTEM_RUNNING);
3739		pr_err("Failed to start kswapd on node %d\n", nid);
3740		ret = PTR_ERR(pgdat->kswapd);
3741		pgdat->kswapd = NULL;
3742	}
3743	return ret;
3744}
3745
3746/*
3747 * Called by memory hotplug when all memory in a node is offlined.  Caller must
3748 * hold mem_hotplug_begin/end().
3749 */
3750void kswapd_stop(int nid)
3751{
3752	struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3753
3754	if (kswapd) {
3755		kthread_stop(kswapd);
3756		NODE_DATA(nid)->kswapd = NULL;
3757	}
3758}
3759
3760static int __init kswapd_init(void)
3761{
3762	int nid, ret;
3763
3764	swap_setup();
3765	for_each_node_state(nid, N_MEMORY)
3766 		kswapd_run(nid);
3767	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3768					"mm/vmscan:online", kswapd_cpu_online,
3769					NULL);
3770	WARN_ON(ret < 0);
3771	return 0;
3772}
3773
3774module_init(kswapd_init)
3775
3776#ifdef CONFIG_NUMA
3777/*
3778 * Node reclaim mode
3779 *
3780 * If non-zero call node_reclaim when the number of free pages falls below
3781 * the watermarks.
3782 */
3783int node_reclaim_mode __read_mostly;
3784
3785#define RECLAIM_OFF 0
3786#define RECLAIM_ZONE (1<<0)	/* Run shrink_inactive_list on the zone */
3787#define RECLAIM_WRITE (1<<1)	/* Writeout pages during reclaim */
3788#define RECLAIM_UNMAP (1<<2)	/* Unmap pages during reclaim */
3789
3790/*
3791 * Priority for NODE_RECLAIM. This determines the fraction of pages
3792 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3793 * a zone.
3794 */
3795#define NODE_RECLAIM_PRIORITY 4
3796
3797/*
3798 * Percentage of pages in a zone that must be unmapped for node_reclaim to
3799 * occur.
3800 */
3801int sysctl_min_unmapped_ratio = 1;
3802
3803/*
3804 * If the number of slab pages in a zone grows beyond this percentage then
3805 * slab reclaim needs to occur.
3806 */
3807int sysctl_min_slab_ratio = 5;
3808
3809static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
3810{
3811	unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3812	unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3813		node_page_state(pgdat, NR_ACTIVE_FILE);
3814
3815	/*
3816	 * It's possible for there to be more file mapped pages than
3817	 * accounted for by the pages on the file LRU lists because
3818	 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3819	 */
3820	return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3821}
3822
3823/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3824static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
3825{
3826	unsigned long nr_pagecache_reclaimable;
3827	unsigned long delta = 0;
3828
3829	/*
3830	 * If RECLAIM_UNMAP is set, then all file pages are considered
3831	 * potentially reclaimable. Otherwise, we have to worry about
3832	 * pages like swapcache and node_unmapped_file_pages() provides
3833	 * a better estimate
3834	 */
3835	if (node_reclaim_mode & RECLAIM_UNMAP)
3836		nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
3837	else
3838		nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
3839
3840	/* If we can't clean pages, remove dirty pages from consideration */
3841	if (!(node_reclaim_mode & RECLAIM_WRITE))
3842		delta += node_page_state(pgdat, NR_FILE_DIRTY);
3843
3844	/* Watch for any possible underflows due to delta */
3845	if (unlikely(delta > nr_pagecache_reclaimable))
3846		delta = nr_pagecache_reclaimable;
3847
3848	return nr_pagecache_reclaimable - delta;
3849}
3850
3851/*
3852 * Try to free up some pages from this node through reclaim.
3853 */
3854static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3855{
3856	/* Minimum pages needed in order to stay on node */
3857	const unsigned long nr_pages = 1 << order;
3858	struct task_struct *p = current;
3859	struct reclaim_state reclaim_state;
3860	unsigned int noreclaim_flag;
3861	struct scan_control sc = {
3862		.nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3863		.gfp_mask = current_gfp_context(gfp_mask),
3864		.order = order,
3865		.priority = NODE_RECLAIM_PRIORITY,
3866		.may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3867		.may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
3868		.may_swap = 1,
3869		.reclaim_idx = gfp_zone(gfp_mask),
3870	};
3871
 
 
 
3872	cond_resched();
 
3873	/*
3874	 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
3875	 * and we also need to be able to write out pages for RECLAIM_WRITE
3876	 * and RECLAIM_UNMAP.
3877	 */
3878	noreclaim_flag = memalloc_noreclaim_save();
3879	p->flags |= PF_SWAPWRITE;
3880	fs_reclaim_acquire(sc.gfp_mask);
3881	reclaim_state.reclaimed_slab = 0;
3882	p->reclaim_state = &reclaim_state;
3883
3884	if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
3885		/*
3886		 * Free memory by calling shrink node with increasing
3887		 * priorities until we have enough memory freed.
3888		 */
3889		do {
3890			shrink_node(pgdat, &sc);
3891		} while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3892	}
3893
3894	p->reclaim_state = NULL;
3895	fs_reclaim_release(gfp_mask);
3896	current->flags &= ~PF_SWAPWRITE;
3897	memalloc_noreclaim_restore(noreclaim_flag);
 
 
 
 
3898	return sc.nr_reclaimed >= nr_pages;
3899}
3900
3901int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
3902{
3903	int ret;
3904
3905	/*
3906	 * Node reclaim reclaims unmapped file backed pages and
3907	 * slab pages if we are over the defined limits.
3908	 *
3909	 * A small portion of unmapped file backed pages is needed for
3910	 * file I/O otherwise pages read by file I/O will be immediately
3911	 * thrown out if the node is overallocated. So we do not reclaim
3912	 * if less than a specified percentage of the node is used by
3913	 * unmapped file backed pages.
3914	 */
3915	if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3916	    node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3917		return NODE_RECLAIM_FULL;
3918
3919	/*
3920	 * Do not scan if the allocation should not be delayed.
3921	 */
3922	if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
3923		return NODE_RECLAIM_NOSCAN;
3924
3925	/*
3926	 * Only run node reclaim on the local node or on nodes that do not
3927	 * have associated processors. This will favor the local processor
3928	 * over remote processors and spread off node memory allocations
3929	 * as wide as possible.
3930	 */
3931	if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3932		return NODE_RECLAIM_NOSCAN;
3933
3934	if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3935		return NODE_RECLAIM_NOSCAN;
3936
3937	ret = __node_reclaim(pgdat, gfp_mask, order);
3938	clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
3939
3940	if (!ret)
3941		count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3942
3943	return ret;
3944}
3945#endif
3946
3947/*
3948 * page_evictable - test whether a page is evictable
3949 * @page: the page to test
3950 *
3951 * Test whether page is evictable--i.e., should be placed on active/inactive
3952 * lists vs unevictable list.
3953 *
3954 * Reasons page might not be evictable:
3955 * (1) page's mapping marked unevictable
3956 * (2) page is part of an mlocked VMA
3957 *
3958 */
3959int page_evictable(struct page *page)
3960{
3961	int ret;
3962
3963	/* Prevent address_space of inode and swap cache from being freed */
3964	rcu_read_lock();
3965	ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3966	rcu_read_unlock();
3967	return ret;
3968}
3969
3970#ifdef CONFIG_SHMEM
3971/**
3972 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3973 * @pages:	array of pages to check
3974 * @nr_pages:	number of pages to check
3975 *
3976 * Checks pages for evictability and moves them to the appropriate lru list.
3977 *
3978 * This function is only used for SysV IPC SHM_UNLOCK.
3979 */
3980void check_move_unevictable_pages(struct page **pages, int nr_pages)
3981{
3982	struct lruvec *lruvec;
3983	struct pglist_data *pgdat = NULL;
3984	int pgscanned = 0;
3985	int pgrescued = 0;
3986	int i;
3987
3988	for (i = 0; i < nr_pages; i++) {
3989		struct page *page = pages[i];
3990		struct pglist_data *pagepgdat = page_pgdat(page);
3991
3992		pgscanned++;
3993		if (pagepgdat != pgdat) {
3994			if (pgdat)
3995				spin_unlock_irq(&pgdat->lru_lock);
3996			pgdat = pagepgdat;
3997			spin_lock_irq(&pgdat->lru_lock);
3998		}
3999		lruvec = mem_cgroup_page_lruvec(page, pgdat);
4000
4001		if (!PageLRU(page) || !PageUnevictable(page))
4002			continue;
4003
4004		if (page_evictable(page)) {
4005			enum lru_list lru = page_lru_base_type(page);
4006
4007			VM_BUG_ON_PAGE(PageActive(page), page);
4008			ClearPageUnevictable(page);
4009			del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
4010			add_page_to_lru_list(page, lruvec, lru);
4011			pgrescued++;
4012		}
4013	}
4014
4015	if (pgdat) {
4016		__count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
4017		__count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
4018		spin_unlock_irq(&pgdat->lru_lock);
4019	}
4020}
4021#endif /* CONFIG_SHMEM */