swap.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/swap.c
   4 *
   5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6 */
   7
   8/*
   9 * This file contains the default values for the operation of the
  10 * Linux VM subsystem. Fine-tuning documentation can be found in
  11 * Documentation/admin-guide/sysctl/vm.rst.
  12 * Started 18.12.91
  13 * Swap aging added 23.2.95, Stephen Tweedie.
  14 * Buffermem limits added 12.3.98, Rik van Riel.
  15 */
  16
  17#include <linux/mm.h>
  18#include <linux/sched.h>
  19#include <linux/kernel_stat.h>
  20#include <linux/swap.h>
  21#include <linux/mman.h>
  22#include <linux/pagemap.h>
  23#include <linux/pagevec.h>
  24#include <linux/init.h>
  25#include <linux/export.h>
  26#include <linux/mm_inline.h>
  27#include <linux/percpu_counter.h>
  28#include <linux/memremap.h>
  29#include <linux/percpu.h>
  30#include <linux/cpu.h>
  31#include <linux/notifier.h>
  32#include <linux/backing-dev.h>
  33#include <linux/memcontrol.h>
  34#include <linux/gfp.h>
  35#include <linux/uio.h>
  36#include <linux/hugetlb.h>
  37#include <linux/page_idle.h>
  38#include <linux/local_lock.h>
  39#include <linux/buffer_head.h>
  40
  41#include "internal.h"
  42
  43#define CREATE_TRACE_POINTS
  44#include <trace/events/pagemap.h>
  45
  46/* How many pages do we try to swap or page in/out together? As a power of 2 */
  47int page_cluster;
  48static const int page_cluster_max = 31;
  49
  50struct cpu_fbatches {
  51	/*
  52	 * The following folio batches are grouped together because they are protected
  53	 * by disabling preemption (and interrupts remain enabled).
  54	 */
  55	local_lock_t lock;
  56	struct folio_batch lru_add;
  57	struct folio_batch lru_deactivate_file;
  58	struct folio_batch lru_deactivate;
  59	struct folio_batch lru_lazyfree;
  60#ifdef CONFIG_SMP
  61	struct folio_batch lru_activate;
  62#endif
  63	/* Protecting the following batches which require disabling interrupts */
  64	local_lock_t lock_irq;
  65	struct folio_batch lru_move_tail;
  66};
  67
  68static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
  69	.lock = INIT_LOCAL_LOCK(lock),
  70	.lock_irq = INIT_LOCAL_LOCK(lock_irq),
  71};
  72
  73static void __page_cache_release(struct folio *folio, struct lruvec **lruvecp,
  74		unsigned long *flagsp)
  75{
  76	if (folio_test_lru(folio)) {
  77		folio_lruvec_relock_irqsave(folio, lruvecp, flagsp);
  78		lruvec_del_folio(*lruvecp, folio);
  79		__folio_clear_lru_flags(folio);
  80	}
  81}
  82
  83/*
  84 * This path almost never happens for VM activity - pages are normally freed
  85 * in batches.  But it gets used by networking - and for compound pages.
  86 */
  87static void page_cache_release(struct folio *folio)
  88{
  89	struct lruvec *lruvec = NULL;
  90	unsigned long flags;
  91
  92	__page_cache_release(folio, &lruvec, &flags);
  93	if (lruvec)
  94		unlock_page_lruvec_irqrestore(lruvec, flags);
  95}
  96
  97void __folio_put(struct folio *folio)
  98{
  99	if (unlikely(folio_is_zone_device(folio))) {
 100		free_zone_device_folio(folio);
 101		return;
 102	}
 103
 104	if (folio_test_hugetlb(folio)) {
 105		free_huge_folio(folio);
 106		return;
 107	}
 108
 109	page_cache_release(folio);
 110	folio_unqueue_deferred_split(folio);
 111	mem_cgroup_uncharge(folio);
 112	free_frozen_pages(&folio->page, folio_order(folio));
 113}
 114EXPORT_SYMBOL(__folio_put);
 115
 116typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
 117
 118static void lru_add(struct lruvec *lruvec, struct folio *folio)
 119{
 120	int was_unevictable = folio_test_clear_unevictable(folio);
 121	long nr_pages = folio_nr_pages(folio);
 122
 123	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 124
 125	/*
 126	 * Is an smp_mb__after_atomic() still required here, before
 127	 * folio_evictable() tests the mlocked flag, to rule out the possibility
 128	 * of stranding an evictable folio on an unevictable LRU?  I think
 129	 * not, because __munlock_folio() only clears the mlocked flag
 130	 * while the LRU lock is held.
 131	 *
 132	 * (That is not true of __page_cache_release(), and not necessarily
 133	 * true of folios_put(): but those only clear the mlocked flag after
 134	 * folio_put_testzero() has excluded any other users of the folio.)
 135	 */
 136	if (folio_evictable(folio)) {
 137		if (was_unevictable)
 138			__count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
 139	} else {
 140		folio_clear_active(folio);
 141		folio_set_unevictable(folio);
 142		/*
 143		 * folio->mlock_count = !!folio_test_mlocked(folio)?
 144		 * But that leaves __mlock_folio() in doubt whether another
 145		 * actor has already counted the mlock or not.  Err on the
 146		 * safe side, underestimate, let page reclaim fix it, rather
 147		 * than leaving a page on the unevictable LRU indefinitely.
 148		 */
 149		folio->mlock_count = 0;
 150		if (!was_unevictable)
 151			__count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
 152	}
 153
 154	lruvec_add_folio(lruvec, folio);
 155	trace_mm_lru_insertion(folio);
 156}
 157
 158static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
 159{
 160	int i;
 161	struct lruvec *lruvec = NULL;
 162	unsigned long flags = 0;
 163
 164	for (i = 0; i < folio_batch_count(fbatch); i++) {
 165		struct folio *folio = fbatch->folios[i];
 166
 167		/* block memcg migration while the folio moves between lru */
 168		if (move_fn != lru_add && !folio_test_clear_lru(folio))
 169			continue;
 170
 171		folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
 172		move_fn(lruvec, folio);
 173
 174		folio_set_lru(folio);
 175	}
 176
 177	if (lruvec)
 178		unlock_page_lruvec_irqrestore(lruvec, flags);
 179	folios_put(fbatch);
 180}
 181
 182static void __folio_batch_add_and_move(struct folio_batch __percpu *fbatch,
 183		struct folio *folio, move_fn_t move_fn, bool disable_irq)
 184{
 185	unsigned long flags;
 186
 187	folio_get(folio);
 188
 189	if (disable_irq)
 190		local_lock_irqsave(&cpu_fbatches.lock_irq, flags);
 191	else
 192		local_lock(&cpu_fbatches.lock);
 193
 194	if (!folio_batch_add(this_cpu_ptr(fbatch), folio) ||
 195			!folio_may_be_lru_cached(folio) || lru_cache_disabled())
 196		folio_batch_move_lru(this_cpu_ptr(fbatch), move_fn);
 197
 198	if (disable_irq)
 199		local_unlock_irqrestore(&cpu_fbatches.lock_irq, flags);
 200	else
 201		local_unlock(&cpu_fbatches.lock);
 202}
 203
 204#define folio_batch_add_and_move(folio, op)		\
 205	__folio_batch_add_and_move(			\
 206		&cpu_fbatches.op,			\
 207		folio,					\
 208		op,					\
 209		offsetof(struct cpu_fbatches, op) >=	\
 210		offsetof(struct cpu_fbatches, lock_irq)	\
 211	)
 212
 213static void lru_move_tail(struct lruvec *lruvec, struct folio *folio)
 214{
 215	if (folio_test_unevictable(folio))
 216		return;
 217
 218	lruvec_del_folio(lruvec, folio);
 219	folio_clear_active(folio);
 220	lruvec_add_folio_tail(lruvec, folio);
 221	__count_vm_events(PGROTATED, folio_nr_pages(folio));
 222}
 223
 224/*
 225 * Writeback is about to end against a folio which has been marked for
 226 * immediate reclaim.  If it still appears to be reclaimable, move it
 227 * to the tail of the inactive list.
 228 *
 229 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
 230 */
 231void folio_rotate_reclaimable(struct folio *folio)
 232{
 233	if (folio_test_locked(folio) || folio_test_dirty(folio) ||
 234	    folio_test_unevictable(folio) || !folio_test_lru(folio))
 235		return;
 236
 237	folio_batch_add_and_move(folio, lru_move_tail);
 238}
 239
 240void lru_note_cost_unlock_irq(struct lruvec *lruvec, bool file,
 241		unsigned int nr_io, unsigned int nr_rotated)
 242		__releases(lruvec->lru_lock)
 243{
 244	unsigned long cost;
 245
 246	/*
 247	 * Reflect the relative cost of incurring IO and spending CPU
 248	 * time on rotations. This doesn't attempt to make a precise
 249	 * comparison, it just says: if reloads are about comparable
 250	 * between the LRU lists, or rotations are overwhelmingly
 251	 * different between them, adjust scan balance for CPU work.
 252	 */
 253	cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
 254	if (!cost) {
 255		spin_unlock_irq(&lruvec->lru_lock);
 256		return;
 257	}
 258
 259	for (;;) {
 260		unsigned long lrusize;
 261
 262		/* Record cost event */
 263		if (file)
 264			lruvec->file_cost += cost;
 265		else
 266			lruvec->anon_cost += cost;
 267
 268		/*
 269		 * Decay previous events
 270		 *
 271		 * Because workloads change over time (and to avoid
 272		 * overflow) we keep these statistics as a floating
 273		 * average, which ends up weighing recent refaults
 274		 * more than old ones.
 275		 */
 276		lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
 277			  lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
 278			  lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
 279			  lruvec_page_state(lruvec, NR_ACTIVE_FILE);
 280
 281		if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
 282			lruvec->file_cost /= 2;
 283			lruvec->anon_cost /= 2;
 284		}
 285
 286		spin_unlock_irq(&lruvec->lru_lock);
 287		lruvec = parent_lruvec(lruvec);
 288		if (!lruvec)
 289			break;
 290		spin_lock_irq(&lruvec->lru_lock);
 291	}
 292}
 293
 294void lru_note_cost_refault(struct folio *folio)
 295{
 296	struct lruvec *lruvec;
 297
 298	lruvec = folio_lruvec_lock_irq(folio);
 299	lru_note_cost_unlock_irq(lruvec, folio_is_file_lru(folio),
 300				folio_nr_pages(folio), 0);
 301}
 302
 303static void lru_activate(struct lruvec *lruvec, struct folio *folio)
 304{
 305	long nr_pages = folio_nr_pages(folio);
 306
 307	if (folio_test_active(folio) || folio_test_unevictable(folio))
 308		return;
 309
 310
 311	lruvec_del_folio(lruvec, folio);
 312	folio_set_active(folio);
 313	lruvec_add_folio(lruvec, folio);
 314	trace_mm_lru_activate(folio);
 315
 316	__count_vm_events(PGACTIVATE, nr_pages);
 317	count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE, nr_pages);
 318}
 319
 320#ifdef CONFIG_SMP
 321static void folio_activate_drain(int cpu)
 322{
 323	struct folio_batch *fbatch = &per_cpu(cpu_fbatches.lru_activate, cpu);
 324
 325	if (folio_batch_count(fbatch))
 326		folio_batch_move_lru(fbatch, lru_activate);
 327}
 328
 329void folio_activate(struct folio *folio)
 330{
 331	if (folio_test_active(folio) || folio_test_unevictable(folio) ||
 332	    !folio_test_lru(folio))
 333		return;
 334
 335	folio_batch_add_and_move(folio, lru_activate);
 336}
 337
 338#else
 339static inline void folio_activate_drain(int cpu)
 340{
 341}
 342
 343void folio_activate(struct folio *folio)
 344{
 345	struct lruvec *lruvec;
 346
 347	if (!folio_test_clear_lru(folio))
 348		return;
 349
 350	lruvec = folio_lruvec_lock_irq(folio);
 351	lru_activate(lruvec, folio);
 352	unlock_page_lruvec_irq(lruvec);
 353	folio_set_lru(folio);
 354}
 355#endif
 356
 357static void __lru_cache_activate_folio(struct folio *folio)
 358{
 359	struct folio_batch *fbatch;
 360	int i;
 361
 362	local_lock(&cpu_fbatches.lock);
 363	fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
 364
 365	/*
 366	 * Search backwards on the optimistic assumption that the folio being
 367	 * activated has just been added to this batch. Note that only
 368	 * the local batch is examined as a !LRU folio could be in the
 369	 * process of being released, reclaimed, migrated or on a remote
 370	 * batch that is currently being drained. Furthermore, marking
 371	 * a remote batch's folio active potentially hits a race where
 372	 * a folio is marked active just after it is added to the inactive
 373	 * list causing accounting errors and BUG_ON checks to trigger.
 374	 */
 375	for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
 376		struct folio *batch_folio = fbatch->folios[i];
 377
 378		if (batch_folio == folio) {
 379			folio_set_active(folio);
 380			break;
 381		}
 382	}
 383
 384	local_unlock(&cpu_fbatches.lock);
 385}
 386
 387#ifdef CONFIG_LRU_GEN
 388
 389static void lru_gen_inc_refs(struct folio *folio)
 390{
 391	unsigned long new_flags, old_flags = READ_ONCE(folio->flags.f);
 392
 393	if (folio_test_unevictable(folio))
 394		return;
 395
 396	/* see the comment on LRU_REFS_FLAGS */
 397	if (!folio_test_referenced(folio)) {
 398		set_mask_bits(&folio->flags.f, LRU_REFS_MASK, BIT(PG_referenced));
 399		return;
 400	}
 401
 402	do {
 403		if ((old_flags & LRU_REFS_MASK) == LRU_REFS_MASK) {
 404			if (!folio_test_workingset(folio))
 405				folio_set_workingset(folio);
 406			return;
 407		}
 408
 409		new_flags = old_flags + BIT(LRU_REFS_PGOFF);
 410	} while (!try_cmpxchg(&folio->flags.f, &old_flags, new_flags));
 411}
 412
 413static bool lru_gen_clear_refs(struct folio *folio)
 414{
 415	struct lru_gen_folio *lrugen;
 416	int gen = folio_lru_gen(folio);
 417	int type = folio_is_file_lru(folio);
 418
 419	if (gen < 0)
 420		return true;
 421
 422	set_mask_bits(&folio->flags.f, LRU_REFS_FLAGS | BIT(PG_workingset), 0);
 423
 424	lrugen = &folio_lruvec(folio)->lrugen;
 425	/* whether can do without shuffling under the LRU lock */
 426	return gen == lru_gen_from_seq(READ_ONCE(lrugen->min_seq[type]));
 427}
 428
 429#else /* !CONFIG_LRU_GEN */
 430
 431static void lru_gen_inc_refs(struct folio *folio)
 432{
 433}
 434
 435static bool lru_gen_clear_refs(struct folio *folio)
 436{
 437	return false;
 438}
 439
 440#endif /* CONFIG_LRU_GEN */
 441
 442/**
 443 * folio_mark_accessed - Mark a folio as having seen activity.
 444 * @folio: The folio to mark.
 445 *
 446 * This function will perform one of the following transitions:
 447 *
 448 * * inactive,unreferenced	->	inactive,referenced
 449 * * inactive,referenced	->	active,unreferenced
 450 * * active,unreferenced	->	active,referenced
 451 *
 452 * When a newly allocated folio is not yet visible, so safe for non-atomic ops,
 453 * __folio_set_referenced() may be substituted for folio_mark_accessed().
 454 */
 455void folio_mark_accessed(struct folio *folio)
 456{
 457	if (folio_test_dropbehind(folio))
 458		return;
 459	if (lru_gen_enabled()) {
 460		lru_gen_inc_refs(folio);
 461		return;
 462	}
 463
 464	if (!folio_test_referenced(folio)) {
 465		folio_set_referenced(folio);
 466	} else if (folio_test_unevictable(folio)) {
 467		/*
 468		 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
 469		 * this list is never rotated or maintained, so marking an
 470		 * unevictable page accessed has no effect.
 471		 */
 472	} else if (!folio_test_active(folio)) {
 473		/*
 474		 * If the folio is on the LRU, queue it for activation via
 475		 * cpu_fbatches.lru_activate. Otherwise, assume the folio is in a
 476		 * folio_batch, mark it active and it'll be moved to the active
 477		 * LRU on the next drain.
 478		 */
 479		if (folio_test_lru(folio))
 480			folio_activate(folio);
 481		else
 482			__lru_cache_activate_folio(folio);
 483		folio_clear_referenced(folio);
 484		workingset_activation(folio);
 485	}
 486	if (folio_test_idle(folio))
 487		folio_clear_idle(folio);
 488}
 489EXPORT_SYMBOL(folio_mark_accessed);
 490
 491/**
 492 * folio_add_lru - Add a folio to an LRU list.
 493 * @folio: The folio to be added to the LRU.
 494 *
 495 * Queue the folio for addition to the LRU. The decision on whether
 496 * to add the page to the [in]active [file|anon] list is deferred until the
 497 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
 498 * have the folio added to the active list using folio_mark_accessed().
 499 */
 500void folio_add_lru(struct folio *folio)
 501{
 502	VM_BUG_ON_FOLIO(folio_test_active(folio) &&
 503			folio_test_unevictable(folio), folio);
 504	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 505
 506	/* see the comment in lru_gen_folio_seq() */
 507	if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
 508	    lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
 509		folio_set_active(folio);
 510
 511	folio_batch_add_and_move(folio, lru_add);
 512}
 513EXPORT_SYMBOL(folio_add_lru);
 514
 515/**
 516 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
 517 * @folio: The folio to be added to the LRU.
 518 * @vma: VMA in which the folio is mapped.
 519 *
 520 * If the VMA is mlocked, @folio is added to the unevictable list.
 521 * Otherwise, it is treated the same way as folio_add_lru().
 522 */
 523void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
 524{
 525	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 526
 527	if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
 528		mlock_new_folio(folio);
 529	else
 530		folio_add_lru(folio);
 531}
 532
 533/*
 534 * If the folio cannot be invalidated, it is moved to the
 535 * inactive list to speed up its reclaim.  It is moved to the
 536 * head of the list, rather than the tail, to give the flusher
 537 * threads some time to write it out, as this is much more
 538 * effective than the single-page writeout from reclaim.
 539 *
 540 * If the folio isn't mapped and dirty/writeback, the folio
 541 * could be reclaimed asap using the reclaim flag.
 542 *
 543 * 1. active, mapped folio -> none
 544 * 2. active, dirty/writeback folio -> inactive, head, reclaim
 545 * 3. inactive, mapped folio -> none
 546 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
 547 * 5. inactive, clean -> inactive, tail
 548 * 6. Others -> none
 549 *
 550 * In 4, it moves to the head of the inactive list so the folio is
 551 * written out by flusher threads as this is much more efficient
 552 * than the single-page writeout from reclaim.
 553 */
 554static void lru_deactivate_file(struct lruvec *lruvec, struct folio *folio)
 555{
 556	bool active = folio_test_active(folio) || lru_gen_enabled();
 557	long nr_pages = folio_nr_pages(folio);
 558
 559	if (folio_test_unevictable(folio))
 560		return;
 561
 562	/* Some processes are using the folio */
 563	if (folio_mapped(folio))
 564		return;
 565
 566	lruvec_del_folio(lruvec, folio);
 567	folio_clear_active(folio);
 568	folio_clear_referenced(folio);
 569
 570	if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
 571		/*
 572		 * Setting the reclaim flag could race with
 573		 * folio_end_writeback() and confuse readahead.  But the
 574		 * race window is _really_ small and  it's not a critical
 575		 * problem.
 576		 */
 577		lruvec_add_folio(lruvec, folio);
 578		folio_set_reclaim(folio);
 579	} else {
 580		/*
 581		 * The folio's writeback ended while it was in the batch.
 582		 * We move that folio to the tail of the inactive list.
 583		 */
 584		lruvec_add_folio_tail(lruvec, folio);
 585		__count_vm_events(PGROTATED, nr_pages);
 586	}
 587
 588	if (active) {
 589		__count_vm_events(PGDEACTIVATE, nr_pages);
 590		count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 591				     nr_pages);
 592	}
 593}
 594
 595static void lru_deactivate(struct lruvec *lruvec, struct folio *folio)
 596{
 597	long nr_pages = folio_nr_pages(folio);
 598
 599	if (folio_test_unevictable(folio) || !(folio_test_active(folio) || lru_gen_enabled()))
 600		return;
 601
 602	lruvec_del_folio(lruvec, folio);
 603	folio_clear_active(folio);
 604	folio_clear_referenced(folio);
 605	lruvec_add_folio(lruvec, folio);
 606
 607	__count_vm_events(PGDEACTIVATE, nr_pages);
 608	count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_pages);
 609}
 610
 611static void lru_lazyfree(struct lruvec *lruvec, struct folio *folio)
 612{
 613	long nr_pages = folio_nr_pages(folio);
 614
 615	if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
 616	    folio_test_swapcache(folio) || folio_test_unevictable(folio))
 617		return;
 618
 619	lruvec_del_folio(lruvec, folio);
 620	folio_clear_active(folio);
 621	if (lru_gen_enabled())
 622		lru_gen_clear_refs(folio);
 623	else
 624		folio_clear_referenced(folio);
 625	/*
 626	 * Lazyfree folios are clean anonymous folios.  They have
 627	 * the swapbacked flag cleared, to distinguish them from normal
 628	 * anonymous folios
 629	 */
 630	folio_clear_swapbacked(folio);
 631	lruvec_add_folio(lruvec, folio);
 632
 633	__count_vm_events(PGLAZYFREE, nr_pages);
 634	count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE, nr_pages);
 635}
 636
 637/*
 638 * Drain pages out of the cpu's folio_batch.
 639 * Either "cpu" is the current CPU, and preemption has already been
 640 * disabled; or "cpu" is being hot-unplugged, and is already dead.
 641 */
 642void lru_add_drain_cpu(int cpu)
 643{
 644	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
 645	struct folio_batch *fbatch = &fbatches->lru_add;
 646
 647	if (folio_batch_count(fbatch))
 648		folio_batch_move_lru(fbatch, lru_add);
 649
 650	fbatch = &fbatches->lru_move_tail;
 651	/* Disabling interrupts below acts as a compiler barrier. */
 652	if (data_race(folio_batch_count(fbatch))) {
 653		unsigned long flags;
 654
 655		/* No harm done if a racing interrupt already did this */
 656		local_lock_irqsave(&cpu_fbatches.lock_irq, flags);
 657		folio_batch_move_lru(fbatch, lru_move_tail);
 658		local_unlock_irqrestore(&cpu_fbatches.lock_irq, flags);
 659	}
 660
 661	fbatch = &fbatches->lru_deactivate_file;
 662	if (folio_batch_count(fbatch))
 663		folio_batch_move_lru(fbatch, lru_deactivate_file);
 664
 665	fbatch = &fbatches->lru_deactivate;
 666	if (folio_batch_count(fbatch))
 667		folio_batch_move_lru(fbatch, lru_deactivate);
 668
 669	fbatch = &fbatches->lru_lazyfree;
 670	if (folio_batch_count(fbatch))
 671		folio_batch_move_lru(fbatch, lru_lazyfree);
 672
 673	folio_activate_drain(cpu);
 674}
 675
 676/**
 677 * deactivate_file_folio() - Deactivate a file folio.
 678 * @folio: Folio to deactivate.
 679 *
 680 * This function hints to the VM that @folio is a good reclaim candidate,
 681 * for example if its invalidation fails due to the folio being dirty
 682 * or under writeback.
 683 *
 684 * Context: Caller holds a reference on the folio.
 685 */
 686void deactivate_file_folio(struct folio *folio)
 687{
 688	/* Deactivating an unevictable folio will not accelerate reclaim */
 689	if (folio_test_unevictable(folio) || !folio_test_lru(folio))
 690		return;
 691
 692	if (lru_gen_enabled() && lru_gen_clear_refs(folio))
 693		return;
 694
 695	folio_batch_add_and_move(folio, lru_deactivate_file);
 696}
 697
 698/*
 699 * folio_deactivate - deactivate a folio
 700 * @folio: folio to deactivate
 701 *
 702 * folio_deactivate() moves @folio to the inactive list if @folio was on the
 703 * active list and was not unevictable. This is done to accelerate the
 704 * reclaim of @folio.
 705 */
 706void folio_deactivate(struct folio *folio)
 707{
 708	if (folio_test_unevictable(folio) || !folio_test_lru(folio))
 709		return;
 710
 711	if (lru_gen_enabled() ? lru_gen_clear_refs(folio) : !folio_test_active(folio))
 712		return;
 713
 714	folio_batch_add_and_move(folio, lru_deactivate);
 715}
 716
 717/**
 718 * folio_mark_lazyfree - make an anon folio lazyfree
 719 * @folio: folio to deactivate
 720 *
 721 * folio_mark_lazyfree() moves @folio to the inactive file list.
 722 * This is done to accelerate the reclaim of @folio.
 723 */
 724void folio_mark_lazyfree(struct folio *folio)
 725{
 726	if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
 727	    !folio_test_lru(folio) ||
 728	    folio_test_swapcache(folio) || folio_test_unevictable(folio))
 729		return;
 730
 731	folio_batch_add_and_move(folio, lru_lazyfree);
 732}
 733
 734void lru_add_drain(void)
 735{
 736	local_lock(&cpu_fbatches.lock);
 737	lru_add_drain_cpu(smp_processor_id());
 738	local_unlock(&cpu_fbatches.lock);
 739	mlock_drain_local();
 740}
 741
 742/*
 743 * It's called from per-cpu workqueue context in SMP case so
 744 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
 745 * the same cpu. It shouldn't be a problem in !SMP case since
 746 * the core is only one and the locks will disable preemption.
 747 */
 748static void lru_add_and_bh_lrus_drain(void)
 749{
 750	local_lock(&cpu_fbatches.lock);
 751	lru_add_drain_cpu(smp_processor_id());
 752	local_unlock(&cpu_fbatches.lock);
 753	invalidate_bh_lrus_cpu();
 754	mlock_drain_local();
 755}
 756
 757void lru_add_drain_cpu_zone(struct zone *zone)
 758{
 759	local_lock(&cpu_fbatches.lock);
 760	lru_add_drain_cpu(smp_processor_id());
 761	drain_local_pages(zone);
 762	local_unlock(&cpu_fbatches.lock);
 763	mlock_drain_local();
 764}
 765
 766#ifdef CONFIG_SMP
 767
 768static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 769
 770static void lru_add_drain_per_cpu(struct work_struct *dummy)
 771{
 772	lru_add_and_bh_lrus_drain();
 773}
 774
 775static bool cpu_needs_drain(unsigned int cpu)
 776{
 777	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
 778
 779	/* Check these in order of likelihood that they're not zero */
 780	return folio_batch_count(&fbatches->lru_add) ||
 781		folio_batch_count(&fbatches->lru_move_tail) ||
 782		folio_batch_count(&fbatches->lru_deactivate_file) ||
 783		folio_batch_count(&fbatches->lru_deactivate) ||
 784		folio_batch_count(&fbatches->lru_lazyfree) ||
 785		folio_batch_count(&fbatches->lru_activate) ||
 786		need_mlock_drain(cpu) ||
 787		has_bh_in_lru(cpu, NULL);
 788}
 789
 790/*
 791 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 792 * kworkers being shut down before our page_alloc_cpu_dead callback is
 793 * executed on the offlined cpu.
 794 * Calling this function with cpu hotplug locks held can actually lead
 795 * to obscure indirect dependencies via WQ context.
 796 */
 797static inline void __lru_add_drain_all(bool force_all_cpus)
 798{
 799	/*
 800	 * lru_drain_gen - Global pages generation number
 801	 *
 802	 * (A) Definition: global lru_drain_gen = x implies that all generations
 803	 *     0 < n <= x are already *scheduled* for draining.
 804	 *
 805	 * This is an optimization for the highly-contended use case where a
 806	 * user space workload keeps constantly generating a flow of pages for
 807	 * each CPU.
 808	 */
 809	static unsigned int lru_drain_gen;
 810	static struct cpumask has_work;
 811	static DEFINE_MUTEX(lock);
 812	unsigned cpu, this_gen;
 813
 814	/*
 815	 * Make sure nobody triggers this path before mm_percpu_wq is fully
 816	 * initialized.
 817	 */
 818	if (WARN_ON(!mm_percpu_wq))
 819		return;
 820
 821	/*
 822	 * Guarantee folio_batch counter stores visible by this CPU
 823	 * are visible to other CPUs before loading the current drain
 824	 * generation.
 825	 */
 826	smp_mb();
 827
 828	/*
 829	 * (B) Locally cache global LRU draining generation number
 830	 *
 831	 * The read barrier ensures that the counter is loaded before the mutex
 832	 * is taken. It pairs with smp_mb() inside the mutex critical section
 833	 * at (D).
 834	 */
 835	this_gen = smp_load_acquire(&lru_drain_gen);
 836
 837	/* It helps everyone if we do our own local drain immediately. */
 838	lru_add_drain();
 839
 840	mutex_lock(&lock);
 841
 842	/*
 843	 * (C) Exit the draining operation if a newer generation, from another
 844	 * lru_add_drain_all(), was already scheduled for draining. Check (A).
 845	 */
 846	if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
 847		goto done;
 848
 849	/*
 850	 * (D) Increment global generation number
 851	 *
 852	 * Pairs with smp_load_acquire() at (B), outside of the critical
 853	 * section. Use a full memory barrier to guarantee that the
 854	 * new global drain generation number is stored before loading
 855	 * folio_batch counters.
 856	 *
 857	 * This pairing must be done here, before the for_each_online_cpu loop
 858	 * below which drains the page vectors.
 859	 *
 860	 * Let x, y, and z represent some system CPU numbers, where x < y < z.
 861	 * Assume CPU #z is in the middle of the for_each_online_cpu loop
 862	 * below and has already reached CPU #y's per-cpu data. CPU #x comes
 863	 * along, adds some pages to its per-cpu vectors, then calls
 864	 * lru_add_drain_all().
 865	 *
 866	 * If the paired barrier is done at any later step, e.g. after the
 867	 * loop, CPU #x will just exit at (C) and miss flushing out all of its
 868	 * added pages.
 869	 */
 870	WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
 871	smp_mb();
 872
 873	cpumask_clear(&has_work);
 874	for_each_online_cpu(cpu) {
 875		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 876
 877		if (cpu_needs_drain(cpu)) {
 878			INIT_WORK(work, lru_add_drain_per_cpu);
 879			queue_work_on(cpu, mm_percpu_wq, work);
 880			__cpumask_set_cpu(cpu, &has_work);
 881		}
 882	}
 883
 884	for_each_cpu(cpu, &has_work)
 885		flush_work(&per_cpu(lru_add_drain_work, cpu));
 886
 887done:
 888	mutex_unlock(&lock);
 889}
 890
 891void lru_add_drain_all(void)
 892{
 893	__lru_add_drain_all(false);
 894}
 895#else
 896void lru_add_drain_all(void)
 897{
 898	lru_add_drain();
 899}
 900#endif /* CONFIG_SMP */
 901
 902atomic_t lru_disable_count = ATOMIC_INIT(0);
 903
 904/*
 905 * lru_cache_disable() needs to be called before we start compiling
 906 * a list of folios to be migrated using folio_isolate_lru().
 907 * It drains folios on LRU cache and then disable on all cpus until
 908 * lru_cache_enable is called.
 909 *
 910 * Must be paired with a call to lru_cache_enable().
 911 */
 912void lru_cache_disable(void)
 913{
 914	atomic_inc(&lru_disable_count);
 915	/*
 916	 * Readers of lru_disable_count are protected by either disabling
 917	 * preemption or rcu_read_lock:
 918	 *
 919	 * preempt_disable, local_irq_disable  [bh_lru_lock()]
 920	 * rcu_read_lock		       [rt_spin_lock CONFIG_PREEMPT_RT]
 921	 * preempt_disable		       [local_lock !CONFIG_PREEMPT_RT]
 922	 *
 923	 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
 924	 * preempt_disable() regions of code. So any CPU which sees
 925	 * lru_disable_count = 0 will have exited the critical
 926	 * section when synchronize_rcu() returns.
 927	 */
 928	synchronize_rcu_expedited();
 929#ifdef CONFIG_SMP
 930	__lru_add_drain_all(true);
 931#else
 932	lru_add_and_bh_lrus_drain();
 933#endif
 934}
 935
 936/**
 937 * folios_put_refs - Reduce the reference count on a batch of folios.
 938 * @folios: The folios.
 939 * @refs: The number of refs to subtract from each folio.
 940 *
 941 * Like folio_put(), but for a batch of folios.  This is more efficient
 942 * than writing the loop yourself as it will optimise the locks which need
 943 * to be taken if the folios are freed.  The folios batch is returned
 944 * empty and ready to be reused for another batch; there is no need
 945 * to reinitialise it.  If @refs is NULL, we subtract one from each
 946 * folio refcount.
 947 *
 948 * Context: May be called in process or interrupt context, but not in NMI
 949 * context.  May be called while holding a spinlock.
 950 */
 951void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 952{
 953	int i, j;
 954	struct lruvec *lruvec = NULL;
 955	unsigned long flags = 0;
 956
 957	for (i = 0, j = 0; i < folios->nr; i++) {
 958		struct folio *folio = folios->folios[i];
 959		unsigned int nr_refs = refs ? refs[i] : 1;
 960
 961		if (is_huge_zero_folio(folio))
 962			continue;
 963
 964		if (folio_is_zone_device(folio)) {
 965			if (lruvec) {
 966				unlock_page_lruvec_irqrestore(lruvec, flags);
 967				lruvec = NULL;
 968			}
 969			if (folio_ref_sub_and_test(folio, nr_refs))
 970				free_zone_device_folio(folio);
 971			continue;
 972		}
 973
 974		if (!folio_ref_sub_and_test(folio, nr_refs))
 975			continue;
 976
 977		/* hugetlb has its own memcg */
 978		if (folio_test_hugetlb(folio)) {
 979			if (lruvec) {
 980				unlock_page_lruvec_irqrestore(lruvec, flags);
 981				lruvec = NULL;
 982			}
 983			free_huge_folio(folio);
 984			continue;
 985		}
 986		folio_unqueue_deferred_split(folio);
 987		__page_cache_release(folio, &lruvec, &flags);
 988
 989		if (j != i)
 990			folios->folios[j] = folio;
 991		j++;
 992	}
 993	if (lruvec)
 994		unlock_page_lruvec_irqrestore(lruvec, flags);
 995	if (!j) {
 996		folio_batch_reinit(folios);
 997		return;
 998	}
 999
1000	folios->nr = j;
1001	mem_cgroup_uncharge_folios(folios);
1002	free_unref_folios(folios);
1003}
1004EXPORT_SYMBOL(folios_put_refs);
1005
1006/**
1007 * release_pages - batched put_page()
1008 * @arg: array of pages to release
1009 * @nr: number of pages
1010 *
1011 * Decrement the reference count on all the pages in @arg.  If it
1012 * fell to zero, remove the page from the LRU and free it.
1013 *
1014 * Note that the argument can be an array of pages, encoded pages,
1015 * or folio pointers. We ignore any encoded bits, and turn any of
1016 * them into just a folio that gets free'd.
1017 */
1018void release_pages(release_pages_arg arg, int nr)
1019{
1020	struct folio_batch fbatch;
1021	int refs[PAGEVEC_SIZE];
1022	struct encoded_page **encoded = arg.encoded_pages;
1023	int i;
1024
1025	folio_batch_init(&fbatch);
1026	for (i = 0; i < nr; i++) {
1027		/* Turn any of the argument types into a folio */
1028		struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
1029
1030		/* Is our next entry actually "nr_pages" -> "nr_refs" ? */
1031		refs[fbatch.nr] = 1;
1032		if (unlikely(encoded_page_flags(encoded[i]) &
1033			     ENCODED_PAGE_BIT_NR_PAGES_NEXT))
1034			refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
1035
1036		if (folio_batch_add(&fbatch, folio) > 0)
1037			continue;
1038		folios_put_refs(&fbatch, refs);
1039	}
1040
1041	if (fbatch.nr)
1042		folios_put_refs(&fbatch, refs);
1043}
1044EXPORT_SYMBOL(release_pages);
1045
1046/*
1047 * The folios which we're about to release may be in the deferred lru-addition
1048 * queues.  That would prevent them from really being freed right now.  That's
1049 * OK from a correctness point of view but is inefficient - those folios may be
1050 * cache-warm and we want to give them back to the page allocator ASAP.
1051 *
1052 * So __folio_batch_release() will drain those queues here.
1053 * folio_batch_move_lru() calls folios_put() directly to avoid
1054 * mutual recursion.
1055 */
1056void __folio_batch_release(struct folio_batch *fbatch)
1057{
1058	if (!fbatch->percpu_pvec_drained) {
1059		lru_add_drain();
1060		fbatch->percpu_pvec_drained = true;
1061	}
1062	folios_put(fbatch);
1063}
1064EXPORT_SYMBOL(__folio_batch_release);
1065
1066/**
1067 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1068 * @fbatch: The batch to prune
1069 *
1070 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1071 * entries.  This function prunes all the non-folio entries from @fbatch
1072 * without leaving holes, so that it can be passed on to folio-only batch
1073 * operations.
1074 */
1075void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1076{
1077	unsigned int i, j;
1078
1079	for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1080		struct folio *folio = fbatch->folios[i];
1081		if (!xa_is_value(folio))
1082			fbatch->folios[j++] = folio;
1083	}
1084	fbatch->nr = j;
1085}
1086
1087static const struct ctl_table swap_sysctl_table[] = {
1088	{
1089		.procname	= "page-cluster",
1090		.data		= &page_cluster,
1091		.maxlen		= sizeof(int),
1092		.mode		= 0644,
1093		.proc_handler	= proc_dointvec_minmax,
1094		.extra1		= SYSCTL_ZERO,
1095		.extra2		= (void *)&page_cluster_max,
1096	}
1097};
1098
1099/*
1100 * Perform any setup for the swap system
1101 */
1102void __init swap_setup(void)
1103{
1104	unsigned long megs = PAGES_TO_MB(totalram_pages());
1105
1106	/* Use a smaller cluster for small-memory machines */
1107	if (megs < 16)
1108		page_cluster = 2;
1109	else
1110		page_cluster = 3;
1111	/*
1112	 * Right now other parts of the system means that we
1113	 * _really_ don't want to cluster much more
1114	 */
1115
1116	register_sysctl_init("vm", swap_sysctl_table);
1117}