Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v6.9.4
   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;
  48const int page_cluster_max = 31;
  49
  50/* Protecting only lru_rotate.fbatch which requires disabling interrupts */
  51struct lru_rotate {
  52	local_lock_t lock;
  53	struct folio_batch fbatch;
  54};
  55static DEFINE_PER_CPU(struct lru_rotate, lru_rotate) = {
  56	.lock = INIT_LOCAL_LOCK(lock),
  57};
  58
  59/*
  60 * The following folio batches are grouped together because they are protected
  61 * by disabling preemption (and interrupts remain enabled).
  62 */
  63struct cpu_fbatches {
 
 
 
 
  64	local_lock_t lock;
  65	struct folio_batch lru_add;
  66	struct folio_batch lru_deactivate_file;
  67	struct folio_batch lru_deactivate;
  68	struct folio_batch lru_lazyfree;
  69#ifdef CONFIG_SMP
  70	struct folio_batch activate;
  71#endif
 
 
 
  72};
 
  73static DEFINE_PER_CPU(struct cpu_fbatches, cpu_fbatches) = {
  74	.lock = INIT_LOCAL_LOCK(lock),
 
  75};
  76
  77static void __page_cache_release(struct folio *folio, struct lruvec **lruvecp,
  78		unsigned long *flagsp)
  79{
  80	if (folio_test_lru(folio)) {
  81		folio_lruvec_relock_irqsave(folio, lruvecp, flagsp);
  82		lruvec_del_folio(*lruvecp, folio);
  83		__folio_clear_lru_flags(folio);
  84	}
  85
  86	/*
  87	 * In rare cases, when truncation or holepunching raced with
  88	 * munlock after VM_LOCKED was cleared, Mlocked may still be
  89	 * found set here.  This does not indicate a problem, unless
  90	 * "unevictable_pgs_cleared" appears worryingly large.
  91	 */
  92	if (unlikely(folio_test_mlocked(folio))) {
  93		long nr_pages = folio_nr_pages(folio);
  94
  95		__folio_clear_mlocked(folio);
  96		zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
  97		count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
  98	}
  99}
 100
 101/*
 102 * This path almost never happens for VM activity - pages are normally freed
 103 * in batches.  But it gets used by networking - and for compound pages.
 104 */
 105static void page_cache_release(struct folio *folio)
 106{
 107	struct lruvec *lruvec = NULL;
 108	unsigned long flags;
 109
 110	__page_cache_release(folio, &lruvec, &flags);
 111	if (lruvec)
 112		unlock_page_lruvec_irqrestore(lruvec, flags);
 113}
 114
 115static void __folio_put_small(struct folio *folio)
 116{
 117	page_cache_release(folio);
 118	mem_cgroup_uncharge(folio);
 119	free_unref_page(&folio->page, 0);
 120}
 121
 122static void __folio_put_large(struct folio *folio)
 123{
 124	/*
 125	 * __page_cache_release() is supposed to be called for thp, not for
 126	 * hugetlb. This is because hugetlb page does never have PageLRU set
 127	 * (it's never listed to any LRU lists) and no memcg routines should
 128	 * be called for hugetlb (it has a separate hugetlb_cgroup.)
 129	 */
 130	if (!folio_test_hugetlb(folio))
 131		page_cache_release(folio);
 132	destroy_large_folio(folio);
 133}
 134
 135void __folio_put(struct folio *folio)
 136{
 137	if (unlikely(folio_is_zone_device(folio)))
 138		free_zone_device_page(&folio->page);
 139	else if (unlikely(folio_test_large(folio)))
 140		__folio_put_large(folio);
 141	else
 142		__folio_put_small(folio);
 143}
 144EXPORT_SYMBOL(__folio_put);
 145
 146/**
 147 * put_pages_list() - release a list of pages
 148 * @pages: list of pages threaded on page->lru
 149 *
 150 * Release a list of pages which are strung together on page.lru.
 151 */
 152void put_pages_list(struct list_head *pages)
 153{
 154	struct folio_batch fbatch;
 155	struct folio *folio, *next;
 156
 157	folio_batch_init(&fbatch);
 158	list_for_each_entry_safe(folio, next, pages, lru) {
 159		if (!folio_put_testzero(folio))
 160			continue;
 161		if (folio_test_large(folio)) {
 162			__folio_put_large(folio);
 163			continue;
 164		}
 165		/* LRU flag must be clear because it's passed using the lru */
 166		if (folio_batch_add(&fbatch, folio) > 0)
 167			continue;
 168		free_unref_folios(&fbatch);
 169	}
 170
 171	if (fbatch.nr)
 172		free_unref_folios(&fbatch);
 173	INIT_LIST_HEAD(pages);
 
 174}
 175EXPORT_SYMBOL(put_pages_list);
 176
 177typedef void (*move_fn_t)(struct lruvec *lruvec, struct folio *folio);
 178
 179static void lru_add_fn(struct lruvec *lruvec, struct folio *folio)
 180{
 181	int was_unevictable = folio_test_clear_unevictable(folio);
 182	long nr_pages = folio_nr_pages(folio);
 183
 184	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 185
 186	/*
 187	 * Is an smp_mb__after_atomic() still required here, before
 188	 * folio_evictable() tests the mlocked flag, to rule out the possibility
 189	 * of stranding an evictable folio on an unevictable LRU?  I think
 190	 * not, because __munlock_folio() only clears the mlocked flag
 191	 * while the LRU lock is held.
 192	 *
 193	 * (That is not true of __page_cache_release(), and not necessarily
 194	 * true of folios_put(): but those only clear the mlocked flag after
 195	 * folio_put_testzero() has excluded any other users of the folio.)
 196	 */
 197	if (folio_evictable(folio)) {
 198		if (was_unevictable)
 199			__count_vm_events(UNEVICTABLE_PGRESCUED, nr_pages);
 200	} else {
 201		folio_clear_active(folio);
 202		folio_set_unevictable(folio);
 203		/*
 204		 * folio->mlock_count = !!folio_test_mlocked(folio)?
 205		 * But that leaves __mlock_folio() in doubt whether another
 206		 * actor has already counted the mlock or not.  Err on the
 207		 * safe side, underestimate, let page reclaim fix it, rather
 208		 * than leaving a page on the unevictable LRU indefinitely.
 209		 */
 210		folio->mlock_count = 0;
 211		if (!was_unevictable)
 212			__count_vm_events(UNEVICTABLE_PGCULLED, nr_pages);
 213	}
 214
 215	lruvec_add_folio(lruvec, folio);
 216	trace_mm_lru_insertion(folio);
 217}
 218
 219static void folio_batch_move_lru(struct folio_batch *fbatch, move_fn_t move_fn)
 220{
 221	int i;
 222	struct lruvec *lruvec = NULL;
 223	unsigned long flags = 0;
 224
 225	for (i = 0; i < folio_batch_count(fbatch); i++) {
 226		struct folio *folio = fbatch->folios[i];
 227
 228		/* block memcg migration while the folio moves between lru */
 229		if (move_fn != lru_add_fn && !folio_test_clear_lru(folio))
 230			continue;
 231
 232		folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
 233		move_fn(lruvec, folio);
 234
 235		folio_set_lru(folio);
 236	}
 237
 238	if (lruvec)
 239		unlock_page_lruvec_irqrestore(lruvec, flags);
 240	folios_put(fbatch);
 241}
 242
 243static void folio_batch_add_and_move(struct folio_batch *fbatch,
 244		struct folio *folio, move_fn_t move_fn)
 
 245{
 246	if (folio_batch_add(fbatch, folio) && !folio_test_large(folio) &&
 247	    !lru_cache_disabled())
 
 248		return;
 249	folio_batch_move_lru(fbatch, move_fn);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 250}
 251
 252static void lru_move_tail_fn(struct lruvec *lruvec, struct folio *folio)
 
 
 
 
 
 
 
 
 
 253{
 254	if (!folio_test_unevictable(folio)) {
 255		lruvec_del_folio(lruvec, folio);
 256		folio_clear_active(folio);
 257		lruvec_add_folio_tail(lruvec, folio);
 258		__count_vm_events(PGROTATED, folio_nr_pages(folio));
 259	}
 
 260}
 261
 262/*
 263 * Writeback is about to end against a folio which has been marked for
 264 * immediate reclaim.  If it still appears to be reclaimable, move it
 265 * to the tail of the inactive list.
 266 *
 267 * folio_rotate_reclaimable() must disable IRQs, to prevent nasty races.
 268 */
 269void folio_rotate_reclaimable(struct folio *folio)
 270{
 271	if (!folio_test_locked(folio) && !folio_test_dirty(folio) &&
 272	    !folio_test_unevictable(folio) && folio_test_lru(folio)) {
 273		struct folio_batch *fbatch;
 274		unsigned long flags;
 275
 276		folio_get(folio);
 277		local_lock_irqsave(&lru_rotate.lock, flags);
 278		fbatch = this_cpu_ptr(&lru_rotate.fbatch);
 279		folio_batch_add_and_move(fbatch, folio, lru_move_tail_fn);
 280		local_unlock_irqrestore(&lru_rotate.lock, flags);
 281	}
 282}
 283
 284void lru_note_cost(struct lruvec *lruvec, bool file,
 285		   unsigned int nr_io, unsigned int nr_rotated)
 286{
 287	unsigned long cost;
 288
 289	/*
 290	 * Reflect the relative cost of incurring IO and spending CPU
 291	 * time on rotations. This doesn't attempt to make a precise
 292	 * comparison, it just says: if reloads are about comparable
 293	 * between the LRU lists, or rotations are overwhelmingly
 294	 * different between them, adjust scan balance for CPU work.
 295	 */
 296	cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
 297
 298	do {
 299		unsigned long lrusize;
 300
 301		/*
 302		 * Hold lruvec->lru_lock is safe here, since
 303		 * 1) The pinned lruvec in reclaim, or
 304		 * 2) From a pre-LRU page during refault (which also holds the
 305		 *    rcu lock, so would be safe even if the page was on the LRU
 306		 *    and could move simultaneously to a new lruvec).
 307		 */
 308		spin_lock_irq(&lruvec->lru_lock);
 309		/* Record cost event */
 310		if (file)
 311			lruvec->file_cost += cost;
 312		else
 313			lruvec->anon_cost += cost;
 314
 315		/*
 316		 * Decay previous events
 317		 *
 318		 * Because workloads change over time (and to avoid
 319		 * overflow) we keep these statistics as a floating
 320		 * average, which ends up weighing recent refaults
 321		 * more than old ones.
 322		 */
 323		lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
 324			  lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
 325			  lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
 326			  lruvec_page_state(lruvec, NR_ACTIVE_FILE);
 327
 328		if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
 329			lruvec->file_cost /= 2;
 330			lruvec->anon_cost /= 2;
 331		}
 332		spin_unlock_irq(&lruvec->lru_lock);
 333	} while ((lruvec = parent_lruvec(lruvec)));
 334}
 335
 336void lru_note_cost_refault(struct folio *folio)
 337{
 338	lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
 339		      folio_nr_pages(folio), 0);
 340}
 341
 342static void folio_activate_fn(struct lruvec *lruvec, struct folio *folio)
 343{
 344	if (!folio_test_active(folio) && !folio_test_unevictable(folio)) {
 345		long nr_pages = folio_nr_pages(folio);
 346
 347		lruvec_del_folio(lruvec, folio);
 348		folio_set_active(folio);
 349		lruvec_add_folio(lruvec, folio);
 350		trace_mm_lru_activate(folio);
 351
 352		__count_vm_events(PGACTIVATE, nr_pages);
 353		__count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE,
 354				     nr_pages);
 355	}
 
 
 
 
 356}
 357
 358#ifdef CONFIG_SMP
 359static void folio_activate_drain(int cpu)
 360{
 361	struct folio_batch *fbatch = &per_cpu(cpu_fbatches.activate, cpu);
 362
 363	if (folio_batch_count(fbatch))
 364		folio_batch_move_lru(fbatch, folio_activate_fn);
 365}
 366
 367void folio_activate(struct folio *folio)
 368{
 369	if (folio_test_lru(folio) && !folio_test_active(folio) &&
 370	    !folio_test_unevictable(folio)) {
 371		struct folio_batch *fbatch;
 372
 373		folio_get(folio);
 374		local_lock(&cpu_fbatches.lock);
 375		fbatch = this_cpu_ptr(&cpu_fbatches.activate);
 376		folio_batch_add_and_move(fbatch, folio, folio_activate_fn);
 377		local_unlock(&cpu_fbatches.lock);
 378	}
 379}
 380
 381#else
 382static inline void folio_activate_drain(int cpu)
 383{
 384}
 385
 386void folio_activate(struct folio *folio)
 387{
 388	struct lruvec *lruvec;
 389
 390	if (folio_test_clear_lru(folio)) {
 391		lruvec = folio_lruvec_lock_irq(folio);
 392		folio_activate_fn(lruvec, folio);
 393		unlock_page_lruvec_irq(lruvec);
 394		folio_set_lru(folio);
 395	}
 
 396}
 397#endif
 398
 399static void __lru_cache_activate_folio(struct folio *folio)
 400{
 401	struct folio_batch *fbatch;
 402	int i;
 403
 404	local_lock(&cpu_fbatches.lock);
 405	fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
 406
 407	/*
 408	 * Search backwards on the optimistic assumption that the folio being
 409	 * activated has just been added to this batch. Note that only
 410	 * the local batch is examined as a !LRU folio could be in the
 411	 * process of being released, reclaimed, migrated or on a remote
 412	 * batch that is currently being drained. Furthermore, marking
 413	 * a remote batch's folio active potentially hits a race where
 414	 * a folio is marked active just after it is added to the inactive
 415	 * list causing accounting errors and BUG_ON checks to trigger.
 416	 */
 417	for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
 418		struct folio *batch_folio = fbatch->folios[i];
 419
 420		if (batch_folio == folio) {
 421			folio_set_active(folio);
 422			break;
 423		}
 424	}
 425
 426	local_unlock(&cpu_fbatches.lock);
 427}
 428
 429#ifdef CONFIG_LRU_GEN
 430static void folio_inc_refs(struct folio *folio)
 431{
 432	unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
 433
 434	if (folio_test_unevictable(folio))
 435		return;
 436
 437	if (!folio_test_referenced(folio)) {
 438		folio_set_referenced(folio);
 439		return;
 440	}
 441
 442	if (!folio_test_workingset(folio)) {
 443		folio_set_workingset(folio);
 444		return;
 445	}
 446
 447	/* see the comment on MAX_NR_TIERS */
 448	do {
 449		new_flags = old_flags & LRU_REFS_MASK;
 450		if (new_flags == LRU_REFS_MASK)
 451			break;
 452
 453		new_flags += BIT(LRU_REFS_PGOFF);
 454		new_flags |= old_flags & ~LRU_REFS_MASK;
 455	} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
 456}
 457#else
 458static void folio_inc_refs(struct folio *folio)
 459{
 460}
 461#endif /* CONFIG_LRU_GEN */
 462
 463/*
 464 * Mark a page as having seen activity.
 
 465 *
 466 * inactive,unreferenced	->	inactive,referenced
 467 * inactive,referenced		->	active,unreferenced
 468 * active,unreferenced		->	active,referenced
 469 *
 470 * When a newly allocated page is not yet visible, so safe for non-atomic ops,
 471 * __SetPageReferenced(page) may be substituted for mark_page_accessed(page).
 
 
 
 
 472 */
 473void folio_mark_accessed(struct folio *folio)
 474{
 475	if (lru_gen_enabled()) {
 476		folio_inc_refs(folio);
 477		return;
 478	}
 479
 480	if (!folio_test_referenced(folio)) {
 481		folio_set_referenced(folio);
 482	} else if (folio_test_unevictable(folio)) {
 483		/*
 484		 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
 485		 * this list is never rotated or maintained, so marking an
 486		 * unevictable page accessed has no effect.
 487		 */
 488	} else if (!folio_test_active(folio)) {
 489		/*
 490		 * If the folio is on the LRU, queue it for activation via
 491		 * cpu_fbatches.activate. Otherwise, assume the folio is in a
 492		 * folio_batch, mark it active and it'll be moved to the active
 493		 * LRU on the next drain.
 494		 */
 495		if (folio_test_lru(folio))
 496			folio_activate(folio);
 497		else
 498			__lru_cache_activate_folio(folio);
 499		folio_clear_referenced(folio);
 500		workingset_activation(folio);
 501	}
 502	if (folio_test_idle(folio))
 503		folio_clear_idle(folio);
 504}
 505EXPORT_SYMBOL(folio_mark_accessed);
 506
 507/**
 508 * folio_add_lru - Add a folio to an LRU list.
 509 * @folio: The folio to be added to the LRU.
 510 *
 511 * Queue the folio for addition to the LRU. The decision on whether
 512 * to add the page to the [in]active [file|anon] list is deferred until the
 513 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
 514 * have the folio added to the active list using folio_mark_accessed().
 515 */
 516void folio_add_lru(struct folio *folio)
 517{
 518	struct folio_batch *fbatch;
 519
 520	VM_BUG_ON_FOLIO(folio_test_active(folio) &&
 521			folio_test_unevictable(folio), folio);
 522	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 523
 524	/* see the comment in lru_gen_add_folio() */
 525	if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
 526	    lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
 527		folio_set_active(folio);
 528
 529	folio_get(folio);
 530	local_lock(&cpu_fbatches.lock);
 531	fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
 532	folio_batch_add_and_move(fbatch, folio, lru_add_fn);
 533	local_unlock(&cpu_fbatches.lock);
 534}
 535EXPORT_SYMBOL(folio_add_lru);
 536
 537/**
 538 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
 539 * @folio: The folio to be added to the LRU.
 540 * @vma: VMA in which the folio is mapped.
 541 *
 542 * If the VMA is mlocked, @folio is added to the unevictable list.
 543 * Otherwise, it is treated the same way as folio_add_lru().
 544 */
 545void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
 546{
 547	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 548
 549	if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
 550		mlock_new_folio(folio);
 551	else
 552		folio_add_lru(folio);
 553}
 554
 555/*
 556 * If the folio cannot be invalidated, it is moved to the
 557 * inactive list to speed up its reclaim.  It is moved to the
 558 * head of the list, rather than the tail, to give the flusher
 559 * threads some time to write it out, as this is much more
 560 * effective than the single-page writeout from reclaim.
 561 *
 562 * If the folio isn't mapped and dirty/writeback, the folio
 563 * could be reclaimed asap using the reclaim flag.
 564 *
 565 * 1. active, mapped folio -> none
 566 * 2. active, dirty/writeback folio -> inactive, head, reclaim
 567 * 3. inactive, mapped folio -> none
 568 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
 569 * 5. inactive, clean -> inactive, tail
 570 * 6. Others -> none
 571 *
 572 * In 4, it moves to the head of the inactive list so the folio is
 573 * written out by flusher threads as this is much more efficient
 574 * than the single-page writeout from reclaim.
 575 */
 576static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio)
 577{
 578	bool active = folio_test_active(folio);
 579	long nr_pages = folio_nr_pages(folio);
 580
 581	if (folio_test_unevictable(folio))
 582		return;
 583
 584	/* Some processes are using the folio */
 585	if (folio_mapped(folio))
 586		return;
 587
 588	lruvec_del_folio(lruvec, folio);
 589	folio_clear_active(folio);
 590	folio_clear_referenced(folio);
 591
 592	if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
 593		/*
 594		 * Setting the reclaim flag could race with
 595		 * folio_end_writeback() and confuse readahead.  But the
 596		 * race window is _really_ small and  it's not a critical
 597		 * problem.
 598		 */
 599		lruvec_add_folio(lruvec, folio);
 600		folio_set_reclaim(folio);
 601	} else {
 602		/*
 603		 * The folio's writeback ended while it was in the batch.
 604		 * We move that folio to the tail of the inactive list.
 605		 */
 606		lruvec_add_folio_tail(lruvec, folio);
 607		__count_vm_events(PGROTATED, nr_pages);
 608	}
 609
 610	if (active) {
 611		__count_vm_events(PGDEACTIVATE, nr_pages);
 612		__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 613				     nr_pages);
 614	}
 615}
 616
 617static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio)
 618{
 619	if (!folio_test_unevictable(folio) && (folio_test_active(folio) || lru_gen_enabled())) {
 620		long nr_pages = folio_nr_pages(folio);
 621
 622		lruvec_del_folio(lruvec, folio);
 623		folio_clear_active(folio);
 624		folio_clear_referenced(folio);
 625		lruvec_add_folio(lruvec, folio);
 626
 627		__count_vm_events(PGDEACTIVATE, nr_pages);
 628		__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 629				     nr_pages);
 630	}
 
 
 
 631}
 632
 633static void lru_lazyfree_fn(struct lruvec *lruvec, struct folio *folio)
 634{
 635	if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
 636	    !folio_test_swapcache(folio) && !folio_test_unevictable(folio)) {
 637		long nr_pages = folio_nr_pages(folio);
 638
 639		lruvec_del_folio(lruvec, folio);
 640		folio_clear_active(folio);
 641		folio_clear_referenced(folio);
 642		/*
 643		 * Lazyfree folios are clean anonymous folios.  They have
 644		 * the swapbacked flag cleared, to distinguish them from normal
 645		 * anonymous folios
 646		 */
 647		folio_clear_swapbacked(folio);
 648		lruvec_add_folio(lruvec, folio);
 649
 650		__count_vm_events(PGLAZYFREE, nr_pages);
 651		__count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE,
 652				     nr_pages);
 653	}
 
 
 
 
 
 
 
 
 
 654}
 655
 656/*
 657 * Drain pages out of the cpu's folio_batch.
 658 * Either "cpu" is the current CPU, and preemption has already been
 659 * disabled; or "cpu" is being hot-unplugged, and is already dead.
 660 */
 661void lru_add_drain_cpu(int cpu)
 662{
 663	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
 664	struct folio_batch *fbatch = &fbatches->lru_add;
 665
 666	if (folio_batch_count(fbatch))
 667		folio_batch_move_lru(fbatch, lru_add_fn);
 668
 669	fbatch = &per_cpu(lru_rotate.fbatch, cpu);
 670	/* Disabling interrupts below acts as a compiler barrier. */
 671	if (data_race(folio_batch_count(fbatch))) {
 672		unsigned long flags;
 673
 674		/* No harm done if a racing interrupt already did this */
 675		local_lock_irqsave(&lru_rotate.lock, flags);
 676		folio_batch_move_lru(fbatch, lru_move_tail_fn);
 677		local_unlock_irqrestore(&lru_rotate.lock, flags);
 678	}
 679
 680	fbatch = &fbatches->lru_deactivate_file;
 681	if (folio_batch_count(fbatch))
 682		folio_batch_move_lru(fbatch, lru_deactivate_file_fn);
 683
 684	fbatch = &fbatches->lru_deactivate;
 685	if (folio_batch_count(fbatch))
 686		folio_batch_move_lru(fbatch, lru_deactivate_fn);
 687
 688	fbatch = &fbatches->lru_lazyfree;
 689	if (folio_batch_count(fbatch))
 690		folio_batch_move_lru(fbatch, lru_lazyfree_fn);
 691
 692	folio_activate_drain(cpu);
 693}
 694
 695/**
 696 * deactivate_file_folio() - Deactivate a file folio.
 697 * @folio: Folio to deactivate.
 698 *
 699 * This function hints to the VM that @folio is a good reclaim candidate,
 700 * for example if its invalidation fails due to the folio being dirty
 701 * or under writeback.
 702 *
 703 * Context: Caller holds a reference on the folio.
 704 */
 705void deactivate_file_folio(struct folio *folio)
 706{
 707	struct folio_batch *fbatch;
 708
 709	/* Deactivating an unevictable folio will not accelerate reclaim */
 710	if (folio_test_unevictable(folio))
 711		return;
 712
 713	folio_get(folio);
 714	local_lock(&cpu_fbatches.lock);
 715	fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate_file);
 716	folio_batch_add_and_move(fbatch, folio, lru_deactivate_file_fn);
 717	local_unlock(&cpu_fbatches.lock);
 718}
 719
 720/*
 721 * folio_deactivate - deactivate a folio
 722 * @folio: folio to deactivate
 723 *
 724 * folio_deactivate() moves @folio to the inactive list if @folio was on the
 725 * active list and was not unevictable. This is done to accelerate the
 726 * reclaim of @folio.
 727 */
 728void folio_deactivate(struct folio *folio)
 729{
 730	if (folio_test_lru(folio) && !folio_test_unevictable(folio) &&
 731	    (folio_test_active(folio) || lru_gen_enabled())) {
 732		struct folio_batch *fbatch;
 733
 734		folio_get(folio);
 735		local_lock(&cpu_fbatches.lock);
 736		fbatch = this_cpu_ptr(&cpu_fbatches.lru_deactivate);
 737		folio_batch_add_and_move(fbatch, folio, lru_deactivate_fn);
 738		local_unlock(&cpu_fbatches.lock);
 739	}
 740}
 741
 742/**
 743 * folio_mark_lazyfree - make an anon folio lazyfree
 744 * @folio: folio to deactivate
 745 *
 746 * folio_mark_lazyfree() moves @folio to the inactive file list.
 747 * This is done to accelerate the reclaim of @folio.
 748 */
 749void folio_mark_lazyfree(struct folio *folio)
 750{
 751	if (folio_test_lru(folio) && folio_test_anon(folio) &&
 752	    folio_test_swapbacked(folio) && !folio_test_swapcache(folio) &&
 753	    !folio_test_unevictable(folio)) {
 754		struct folio_batch *fbatch;
 755
 756		folio_get(folio);
 757		local_lock(&cpu_fbatches.lock);
 758		fbatch = this_cpu_ptr(&cpu_fbatches.lru_lazyfree);
 759		folio_batch_add_and_move(fbatch, folio, lru_lazyfree_fn);
 760		local_unlock(&cpu_fbatches.lock);
 761	}
 762}
 763
 764void lru_add_drain(void)
 765{
 766	local_lock(&cpu_fbatches.lock);
 767	lru_add_drain_cpu(smp_processor_id());
 768	local_unlock(&cpu_fbatches.lock);
 769	mlock_drain_local();
 770}
 771
 772/*
 773 * It's called from per-cpu workqueue context in SMP case so
 774 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
 775 * the same cpu. It shouldn't be a problem in !SMP case since
 776 * the core is only one and the locks will disable preemption.
 777 */
 778static void lru_add_and_bh_lrus_drain(void)
 779{
 780	local_lock(&cpu_fbatches.lock);
 781	lru_add_drain_cpu(smp_processor_id());
 782	local_unlock(&cpu_fbatches.lock);
 783	invalidate_bh_lrus_cpu();
 784	mlock_drain_local();
 785}
 786
 787void lru_add_drain_cpu_zone(struct zone *zone)
 788{
 789	local_lock(&cpu_fbatches.lock);
 790	lru_add_drain_cpu(smp_processor_id());
 791	drain_local_pages(zone);
 792	local_unlock(&cpu_fbatches.lock);
 793	mlock_drain_local();
 794}
 795
 796#ifdef CONFIG_SMP
 797
 798static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 799
 800static void lru_add_drain_per_cpu(struct work_struct *dummy)
 801{
 802	lru_add_and_bh_lrus_drain();
 803}
 804
 805static bool cpu_needs_drain(unsigned int cpu)
 806{
 807	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
 808
 809	/* Check these in order of likelihood that they're not zero */
 810	return folio_batch_count(&fbatches->lru_add) ||
 811		data_race(folio_batch_count(&per_cpu(lru_rotate.fbatch, cpu))) ||
 812		folio_batch_count(&fbatches->lru_deactivate_file) ||
 813		folio_batch_count(&fbatches->lru_deactivate) ||
 814		folio_batch_count(&fbatches->lru_lazyfree) ||
 815		folio_batch_count(&fbatches->activate) ||
 816		need_mlock_drain(cpu) ||
 817		has_bh_in_lru(cpu, NULL);
 818}
 819
 820/*
 821 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 822 * kworkers being shut down before our page_alloc_cpu_dead callback is
 823 * executed on the offlined cpu.
 824 * Calling this function with cpu hotplug locks held can actually lead
 825 * to obscure indirect dependencies via WQ context.
 826 */
 827static inline void __lru_add_drain_all(bool force_all_cpus)
 828{
 829	/*
 830	 * lru_drain_gen - Global pages generation number
 831	 *
 832	 * (A) Definition: global lru_drain_gen = x implies that all generations
 833	 *     0 < n <= x are already *scheduled* for draining.
 834	 *
 835	 * This is an optimization for the highly-contended use case where a
 836	 * user space workload keeps constantly generating a flow of pages for
 837	 * each CPU.
 838	 */
 839	static unsigned int lru_drain_gen;
 840	static struct cpumask has_work;
 841	static DEFINE_MUTEX(lock);
 842	unsigned cpu, this_gen;
 843
 844	/*
 845	 * Make sure nobody triggers this path before mm_percpu_wq is fully
 846	 * initialized.
 847	 */
 848	if (WARN_ON(!mm_percpu_wq))
 849		return;
 850
 851	/*
 852	 * Guarantee folio_batch counter stores visible by this CPU
 853	 * are visible to other CPUs before loading the current drain
 854	 * generation.
 855	 */
 856	smp_mb();
 857
 858	/*
 859	 * (B) Locally cache global LRU draining generation number
 860	 *
 861	 * The read barrier ensures that the counter is loaded before the mutex
 862	 * is taken. It pairs with smp_mb() inside the mutex critical section
 863	 * at (D).
 864	 */
 865	this_gen = smp_load_acquire(&lru_drain_gen);
 866
 867	mutex_lock(&lock);
 868
 869	/*
 870	 * (C) Exit the draining operation if a newer generation, from another
 871	 * lru_add_drain_all(), was already scheduled for draining. Check (A).
 872	 */
 873	if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
 874		goto done;
 875
 876	/*
 877	 * (D) Increment global generation number
 878	 *
 879	 * Pairs with smp_load_acquire() at (B), outside of the critical
 880	 * section. Use a full memory barrier to guarantee that the
 881	 * new global drain generation number is stored before loading
 882	 * folio_batch counters.
 883	 *
 884	 * This pairing must be done here, before the for_each_online_cpu loop
 885	 * below which drains the page vectors.
 886	 *
 887	 * Let x, y, and z represent some system CPU numbers, where x < y < z.
 888	 * Assume CPU #z is in the middle of the for_each_online_cpu loop
 889	 * below and has already reached CPU #y's per-cpu data. CPU #x comes
 890	 * along, adds some pages to its per-cpu vectors, then calls
 891	 * lru_add_drain_all().
 892	 *
 893	 * If the paired barrier is done at any later step, e.g. after the
 894	 * loop, CPU #x will just exit at (C) and miss flushing out all of its
 895	 * added pages.
 896	 */
 897	WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
 898	smp_mb();
 899
 900	cpumask_clear(&has_work);
 901	for_each_online_cpu(cpu) {
 902		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 903
 904		if (cpu_needs_drain(cpu)) {
 905			INIT_WORK(work, lru_add_drain_per_cpu);
 906			queue_work_on(cpu, mm_percpu_wq, work);
 907			__cpumask_set_cpu(cpu, &has_work);
 908		}
 909	}
 910
 911	for_each_cpu(cpu, &has_work)
 912		flush_work(&per_cpu(lru_add_drain_work, cpu));
 913
 914done:
 915	mutex_unlock(&lock);
 916}
 917
 918void lru_add_drain_all(void)
 919{
 920	__lru_add_drain_all(false);
 921}
 922#else
 923void lru_add_drain_all(void)
 924{
 925	lru_add_drain();
 926}
 927#endif /* CONFIG_SMP */
 928
 929atomic_t lru_disable_count = ATOMIC_INIT(0);
 930
 931/*
 932 * lru_cache_disable() needs to be called before we start compiling
 933 * a list of pages to be migrated using isolate_lru_page().
 934 * It drains pages on LRU cache and then disable on all cpus until
 935 * lru_cache_enable is called.
 936 *
 937 * Must be paired with a call to lru_cache_enable().
 938 */
 939void lru_cache_disable(void)
 940{
 941	atomic_inc(&lru_disable_count);
 942	/*
 943	 * Readers of lru_disable_count are protected by either disabling
 944	 * preemption or rcu_read_lock:
 945	 *
 946	 * preempt_disable, local_irq_disable  [bh_lru_lock()]
 947	 * rcu_read_lock		       [rt_spin_lock CONFIG_PREEMPT_RT]
 948	 * preempt_disable		       [local_lock !CONFIG_PREEMPT_RT]
 949	 *
 950	 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
 951	 * preempt_disable() regions of code. So any CPU which sees
 952	 * lru_disable_count = 0 will have exited the critical
 953	 * section when synchronize_rcu() returns.
 954	 */
 955	synchronize_rcu_expedited();
 956#ifdef CONFIG_SMP
 957	__lru_add_drain_all(true);
 958#else
 959	lru_add_and_bh_lrus_drain();
 960#endif
 961}
 962
 963/**
 964 * folios_put_refs - Reduce the reference count on a batch of folios.
 965 * @folios: The folios.
 966 * @refs: The number of refs to subtract from each folio.
 967 *
 968 * Like folio_put(), but for a batch of folios.  This is more efficient
 969 * than writing the loop yourself as it will optimise the locks which need
 970 * to be taken if the folios are freed.  The folios batch is returned
 971 * empty and ready to be reused for another batch; there is no need
 972 * to reinitialise it.  If @refs is NULL, we subtract one from each
 973 * folio refcount.
 974 *
 975 * Context: May be called in process or interrupt context, but not in NMI
 976 * context.  May be called while holding a spinlock.
 977 */
 978void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 979{
 980	int i, j;
 981	struct lruvec *lruvec = NULL;
 982	unsigned long flags = 0;
 983
 984	for (i = 0, j = 0; i < folios->nr; i++) {
 985		struct folio *folio = folios->folios[i];
 986		unsigned int nr_refs = refs ? refs[i] : 1;
 987
 988		if (is_huge_zero_page(&folio->page))
 989			continue;
 990
 991		if (folio_is_zone_device(folio)) {
 992			if (lruvec) {
 993				unlock_page_lruvec_irqrestore(lruvec, flags);
 994				lruvec = NULL;
 995			}
 996			if (put_devmap_managed_page_refs(&folio->page, nr_refs))
 997				continue;
 998			if (folio_ref_sub_and_test(folio, nr_refs))
 999				free_zone_device_page(&folio->page);
1000			continue;
1001		}
1002
1003		if (!folio_ref_sub_and_test(folio, nr_refs))
1004			continue;
1005
1006		/* hugetlb has its own memcg */
1007		if (folio_test_hugetlb(folio)) {
1008			if (lruvec) {
1009				unlock_page_lruvec_irqrestore(lruvec, flags);
1010				lruvec = NULL;
1011			}
1012			free_huge_folio(folio);
1013			continue;
1014		}
1015		if (folio_test_large(folio) &&
1016		    folio_test_large_rmappable(folio))
1017			folio_undo_large_rmappable(folio);
1018
1019		__page_cache_release(folio, &lruvec, &flags);
1020
1021		if (j != i)
1022			folios->folios[j] = folio;
1023		j++;
1024	}
1025	if (lruvec)
1026		unlock_page_lruvec_irqrestore(lruvec, flags);
1027	if (!j) {
1028		folio_batch_reinit(folios);
1029		return;
1030	}
1031
1032	folios->nr = j;
1033	mem_cgroup_uncharge_folios(folios);
1034	free_unref_folios(folios);
1035}
1036EXPORT_SYMBOL(folios_put_refs);
1037
1038/**
1039 * release_pages - batched put_page()
1040 * @arg: array of pages to release
1041 * @nr: number of pages
1042 *
1043 * Decrement the reference count on all the pages in @arg.  If it
1044 * fell to zero, remove the page from the LRU and free it.
1045 *
1046 * Note that the argument can be an array of pages, encoded pages,
1047 * or folio pointers. We ignore any encoded bits, and turn any of
1048 * them into just a folio that gets free'd.
1049 */
1050void release_pages(release_pages_arg arg, int nr)
1051{
1052	struct folio_batch fbatch;
1053	int refs[PAGEVEC_SIZE];
1054	struct encoded_page **encoded = arg.encoded_pages;
1055	int i;
1056
1057	folio_batch_init(&fbatch);
1058	for (i = 0; i < nr; i++) {
1059		/* Turn any of the argument types into a folio */
1060		struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
1061
1062		/* Is our next entry actually "nr_pages" -> "nr_refs" ? */
1063		refs[fbatch.nr] = 1;
1064		if (unlikely(encoded_page_flags(encoded[i]) &
1065			     ENCODED_PAGE_BIT_NR_PAGES_NEXT))
1066			refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
1067
1068		if (folio_batch_add(&fbatch, folio) > 0)
1069			continue;
1070		folios_put_refs(&fbatch, refs);
1071	}
1072
1073	if (fbatch.nr)
1074		folios_put_refs(&fbatch, refs);
1075}
1076EXPORT_SYMBOL(release_pages);
1077
1078/*
1079 * The folios which we're about to release may be in the deferred lru-addition
1080 * queues.  That would prevent them from really being freed right now.  That's
1081 * OK from a correctness point of view but is inefficient - those folios may be
1082 * cache-warm and we want to give them back to the page allocator ASAP.
1083 *
1084 * So __folio_batch_release() will drain those queues here.
1085 * folio_batch_move_lru() calls folios_put() directly to avoid
1086 * mutual recursion.
1087 */
1088void __folio_batch_release(struct folio_batch *fbatch)
1089{
1090	if (!fbatch->percpu_pvec_drained) {
1091		lru_add_drain();
1092		fbatch->percpu_pvec_drained = true;
1093	}
1094	folios_put(fbatch);
1095}
1096EXPORT_SYMBOL(__folio_batch_release);
1097
1098/**
1099 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1100 * @fbatch: The batch to prune
1101 *
1102 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1103 * entries.  This function prunes all the non-folio entries from @fbatch
1104 * without leaving holes, so that it can be passed on to folio-only batch
1105 * operations.
1106 */
1107void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1108{
1109	unsigned int i, j;
1110
1111	for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1112		struct folio *folio = fbatch->folios[i];
1113		if (!xa_is_value(folio))
1114			fbatch->folios[j++] = folio;
1115	}
1116	fbatch->nr = j;
1117}
1118
1119/*
1120 * Perform any setup for the swap system
1121 */
1122void __init swap_setup(void)
1123{
1124	unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1125
1126	/* Use a smaller cluster for small-memory machines */
1127	if (megs < 16)
1128		page_cluster = 2;
1129	else
1130		page_cluster = 3;
1131	/*
1132	 * Right now other parts of the system means that we
1133	 * _really_ don't want to cluster much more
1134	 */
1135}
v6.13.7
   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;
  48const 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_unref_page(&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		folio_lruvec_relock_irqsave(folio, &lruvec, &flags);
 168		move_fn(lruvec, folio);
 169
 170		folio_set_lru(folio);
 171	}
 172
 173	if (lruvec)
 174		unlock_page_lruvec_irqrestore(lruvec, flags);
 175	folios_put(fbatch);
 176}
 177
 178static void __folio_batch_add_and_move(struct folio_batch __percpu *fbatch,
 179		struct folio *folio, move_fn_t move_fn,
 180		bool on_lru, bool disable_irq)
 181{
 182	unsigned long flags;
 183
 184	if (on_lru && !folio_test_clear_lru(folio))
 185		return;
 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) || folio_test_large(folio) ||
 195	    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, on_lru)						\
 205	__folio_batch_add_and_move(								\
 206		&cpu_fbatches.op,								\
 207		folio,										\
 208		op,										\
 209		on_lru,										\
 210		offsetof(struct cpu_fbatches, op) >= 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))
 235		return;
 
 236
 237	folio_batch_add_and_move(folio, lru_move_tail, true);
 
 
 
 
 
 238}
 239
 240void lru_note_cost(struct lruvec *lruvec, bool file,
 241		   unsigned int nr_io, unsigned int nr_rotated)
 242{
 243	unsigned long cost;
 244
 245	/*
 246	 * Reflect the relative cost of incurring IO and spending CPU
 247	 * time on rotations. This doesn't attempt to make a precise
 248	 * comparison, it just says: if reloads are about comparable
 249	 * between the LRU lists, or rotations are overwhelmingly
 250	 * different between them, adjust scan balance for CPU work.
 251	 */
 252	cost = nr_io * SWAP_CLUSTER_MAX + nr_rotated;
 253
 254	do {
 255		unsigned long lrusize;
 256
 257		/*
 258		 * Hold lruvec->lru_lock is safe here, since
 259		 * 1) The pinned lruvec in reclaim, or
 260		 * 2) From a pre-LRU page during refault (which also holds the
 261		 *    rcu lock, so would be safe even if the page was on the LRU
 262		 *    and could move simultaneously to a new lruvec).
 263		 */
 264		spin_lock_irq(&lruvec->lru_lock);
 265		/* Record cost event */
 266		if (file)
 267			lruvec->file_cost += cost;
 268		else
 269			lruvec->anon_cost += cost;
 270
 271		/*
 272		 * Decay previous events
 273		 *
 274		 * Because workloads change over time (and to avoid
 275		 * overflow) we keep these statistics as a floating
 276		 * average, which ends up weighing recent refaults
 277		 * more than old ones.
 278		 */
 279		lrusize = lruvec_page_state(lruvec, NR_INACTIVE_ANON) +
 280			  lruvec_page_state(lruvec, NR_ACTIVE_ANON) +
 281			  lruvec_page_state(lruvec, NR_INACTIVE_FILE) +
 282			  lruvec_page_state(lruvec, NR_ACTIVE_FILE);
 283
 284		if (lruvec->file_cost + lruvec->anon_cost > lrusize / 4) {
 285			lruvec->file_cost /= 2;
 286			lruvec->anon_cost /= 2;
 287		}
 288		spin_unlock_irq(&lruvec->lru_lock);
 289	} while ((lruvec = parent_lruvec(lruvec)));
 290}
 291
 292void lru_note_cost_refault(struct folio *folio)
 293{
 294	lru_note_cost(folio_lruvec(folio), folio_is_file_lru(folio),
 295		      folio_nr_pages(folio), 0);
 296}
 297
 298static void lru_activate(struct lruvec *lruvec, struct folio *folio)
 299{
 300	long nr_pages = folio_nr_pages(folio);
 
 301
 302	if (folio_test_active(folio) || folio_test_unevictable(folio))
 303		return;
 
 
 304
 305
 306	lruvec_del_folio(lruvec, folio);
 307	folio_set_active(folio);
 308	lruvec_add_folio(lruvec, folio);
 309	trace_mm_lru_activate(folio);
 310
 311	__count_vm_events(PGACTIVATE, nr_pages);
 312	__count_memcg_events(lruvec_memcg(lruvec), PGACTIVATE, nr_pages);
 313}
 314
 315#ifdef CONFIG_SMP
 316static void folio_activate_drain(int cpu)
 317{
 318	struct folio_batch *fbatch = &per_cpu(cpu_fbatches.lru_activate, cpu);
 319
 320	if (folio_batch_count(fbatch))
 321		folio_batch_move_lru(fbatch, lru_activate);
 322}
 323
 324void folio_activate(struct folio *folio)
 325{
 326	if (folio_test_active(folio) || folio_test_unevictable(folio))
 327		return;
 
 328
 329	folio_batch_add_and_move(folio, lru_activate, true);
 
 
 
 
 
 330}
 331
 332#else
 333static inline void folio_activate_drain(int cpu)
 334{
 335}
 336
 337void folio_activate(struct folio *folio)
 338{
 339	struct lruvec *lruvec;
 340
 341	if (!folio_test_clear_lru(folio))
 342		return;
 343
 344	lruvec = folio_lruvec_lock_irq(folio);
 345	lru_activate(lruvec, folio);
 346	unlock_page_lruvec_irq(lruvec);
 347	folio_set_lru(folio);
 348}
 349#endif
 350
 351static void __lru_cache_activate_folio(struct folio *folio)
 352{
 353	struct folio_batch *fbatch;
 354	int i;
 355
 356	local_lock(&cpu_fbatches.lock);
 357	fbatch = this_cpu_ptr(&cpu_fbatches.lru_add);
 358
 359	/*
 360	 * Search backwards on the optimistic assumption that the folio being
 361	 * activated has just been added to this batch. Note that only
 362	 * the local batch is examined as a !LRU folio could be in the
 363	 * process of being released, reclaimed, migrated or on a remote
 364	 * batch that is currently being drained. Furthermore, marking
 365	 * a remote batch's folio active potentially hits a race where
 366	 * a folio is marked active just after it is added to the inactive
 367	 * list causing accounting errors and BUG_ON checks to trigger.
 368	 */
 369	for (i = folio_batch_count(fbatch) - 1; i >= 0; i--) {
 370		struct folio *batch_folio = fbatch->folios[i];
 371
 372		if (batch_folio == folio) {
 373			folio_set_active(folio);
 374			break;
 375		}
 376	}
 377
 378	local_unlock(&cpu_fbatches.lock);
 379}
 380
 381#ifdef CONFIG_LRU_GEN
 382static void folio_inc_refs(struct folio *folio)
 383{
 384	unsigned long new_flags, old_flags = READ_ONCE(folio->flags);
 385
 386	if (folio_test_unevictable(folio))
 387		return;
 388
 389	if (!folio_test_referenced(folio)) {
 390		folio_set_referenced(folio);
 391		return;
 392	}
 393
 394	if (!folio_test_workingset(folio)) {
 395		folio_set_workingset(folio);
 396		return;
 397	}
 398
 399	/* see the comment on MAX_NR_TIERS */
 400	do {
 401		new_flags = old_flags & LRU_REFS_MASK;
 402		if (new_flags == LRU_REFS_MASK)
 403			break;
 404
 405		new_flags += BIT(LRU_REFS_PGOFF);
 406		new_flags |= old_flags & ~LRU_REFS_MASK;
 407	} while (!try_cmpxchg(&folio->flags, &old_flags, new_flags));
 408}
 409#else
 410static void folio_inc_refs(struct folio *folio)
 411{
 412}
 413#endif /* CONFIG_LRU_GEN */
 414
 415/**
 416 * folio_mark_accessed - Mark a folio as having seen activity.
 417 * @folio: The folio to mark.
 418 *
 419 * This function will perform one of the following transitions:
 
 
 420 *
 421 * * inactive,unreferenced	->	inactive,referenced
 422 * * inactive,referenced	->	active,unreferenced
 423 * * active,unreferenced	->	active,referenced
 424 *
 425 * When a newly allocated folio is not yet visible, so safe for non-atomic ops,
 426 * __folio_set_referenced() may be substituted for folio_mark_accessed().
 427 */
 428void folio_mark_accessed(struct folio *folio)
 429{
 430	if (lru_gen_enabled()) {
 431		folio_inc_refs(folio);
 432		return;
 433	}
 434
 435	if (!folio_test_referenced(folio)) {
 436		folio_set_referenced(folio);
 437	} else if (folio_test_unevictable(folio)) {
 438		/*
 439		 * Unevictable pages are on the "LRU_UNEVICTABLE" list. But,
 440		 * this list is never rotated or maintained, so marking an
 441		 * unevictable page accessed has no effect.
 442		 */
 443	} else if (!folio_test_active(folio)) {
 444		/*
 445		 * If the folio is on the LRU, queue it for activation via
 446		 * cpu_fbatches.lru_activate. Otherwise, assume the folio is in a
 447		 * folio_batch, mark it active and it'll be moved to the active
 448		 * LRU on the next drain.
 449		 */
 450		if (folio_test_lru(folio))
 451			folio_activate(folio);
 452		else
 453			__lru_cache_activate_folio(folio);
 454		folio_clear_referenced(folio);
 455		workingset_activation(folio);
 456	}
 457	if (folio_test_idle(folio))
 458		folio_clear_idle(folio);
 459}
 460EXPORT_SYMBOL(folio_mark_accessed);
 461
 462/**
 463 * folio_add_lru - Add a folio to an LRU list.
 464 * @folio: The folio to be added to the LRU.
 465 *
 466 * Queue the folio for addition to the LRU. The decision on whether
 467 * to add the page to the [in]active [file|anon] list is deferred until the
 468 * folio_batch is drained. This gives a chance for the caller of folio_add_lru()
 469 * have the folio added to the active list using folio_mark_accessed().
 470 */
 471void folio_add_lru(struct folio *folio)
 472{
 
 
 473	VM_BUG_ON_FOLIO(folio_test_active(folio) &&
 474			folio_test_unevictable(folio), folio);
 475	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 476
 477	/* see the comment in lru_gen_add_folio() */
 478	if (lru_gen_enabled() && !folio_test_unevictable(folio) &&
 479	    lru_gen_in_fault() && !(current->flags & PF_MEMALLOC))
 480		folio_set_active(folio);
 481
 482	folio_batch_add_and_move(folio, lru_add, false);
 
 
 
 
 483}
 484EXPORT_SYMBOL(folio_add_lru);
 485
 486/**
 487 * folio_add_lru_vma() - Add a folio to the appropate LRU list for this VMA.
 488 * @folio: The folio to be added to the LRU.
 489 * @vma: VMA in which the folio is mapped.
 490 *
 491 * If the VMA is mlocked, @folio is added to the unevictable list.
 492 * Otherwise, it is treated the same way as folio_add_lru().
 493 */
 494void folio_add_lru_vma(struct folio *folio, struct vm_area_struct *vma)
 495{
 496	VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 497
 498	if (unlikely((vma->vm_flags & (VM_LOCKED | VM_SPECIAL)) == VM_LOCKED))
 499		mlock_new_folio(folio);
 500	else
 501		folio_add_lru(folio);
 502}
 503
 504/*
 505 * If the folio cannot be invalidated, it is moved to the
 506 * inactive list to speed up its reclaim.  It is moved to the
 507 * head of the list, rather than the tail, to give the flusher
 508 * threads some time to write it out, as this is much more
 509 * effective than the single-page writeout from reclaim.
 510 *
 511 * If the folio isn't mapped and dirty/writeback, the folio
 512 * could be reclaimed asap using the reclaim flag.
 513 *
 514 * 1. active, mapped folio -> none
 515 * 2. active, dirty/writeback folio -> inactive, head, reclaim
 516 * 3. inactive, mapped folio -> none
 517 * 4. inactive, dirty/writeback folio -> inactive, head, reclaim
 518 * 5. inactive, clean -> inactive, tail
 519 * 6. Others -> none
 520 *
 521 * In 4, it moves to the head of the inactive list so the folio is
 522 * written out by flusher threads as this is much more efficient
 523 * than the single-page writeout from reclaim.
 524 */
 525static void lru_deactivate_file(struct lruvec *lruvec, struct folio *folio)
 526{
 527	bool active = folio_test_active(folio);
 528	long nr_pages = folio_nr_pages(folio);
 529
 530	if (folio_test_unevictable(folio))
 531		return;
 532
 533	/* Some processes are using the folio */
 534	if (folio_mapped(folio))
 535		return;
 536
 537	lruvec_del_folio(lruvec, folio);
 538	folio_clear_active(folio);
 539	folio_clear_referenced(folio);
 540
 541	if (folio_test_writeback(folio) || folio_test_dirty(folio)) {
 542		/*
 543		 * Setting the reclaim flag could race with
 544		 * folio_end_writeback() and confuse readahead.  But the
 545		 * race window is _really_ small and  it's not a critical
 546		 * problem.
 547		 */
 548		lruvec_add_folio(lruvec, folio);
 549		folio_set_reclaim(folio);
 550	} else {
 551		/*
 552		 * The folio's writeback ended while it was in the batch.
 553		 * We move that folio to the tail of the inactive list.
 554		 */
 555		lruvec_add_folio_tail(lruvec, folio);
 556		__count_vm_events(PGROTATED, nr_pages);
 557	}
 558
 559	if (active) {
 560		__count_vm_events(PGDEACTIVATE, nr_pages);
 561		__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE,
 562				     nr_pages);
 563	}
 564}
 565
 566static void lru_deactivate(struct lruvec *lruvec, struct folio *folio)
 567{
 568	long nr_pages = folio_nr_pages(folio);
 
 569
 570	if (folio_test_unevictable(folio) || !(folio_test_active(folio) || lru_gen_enabled()))
 571		return;
 
 
 572
 573	lruvec_del_folio(lruvec, folio);
 574	folio_clear_active(folio);
 575	folio_clear_referenced(folio);
 576	lruvec_add_folio(lruvec, folio);
 577
 578	__count_vm_events(PGDEACTIVATE, nr_pages);
 579	__count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_pages);
 580}
 581
 582static void lru_lazyfree(struct lruvec *lruvec, struct folio *folio)
 583{
 584	long nr_pages = folio_nr_pages(folio);
 
 
 585
 586	if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
 587	    folio_test_swapcache(folio) || folio_test_unevictable(folio))
 588		return;
 
 
 
 
 
 
 
 589
 590	lruvec_del_folio(lruvec, folio);
 591	folio_clear_active(folio);
 592	folio_clear_referenced(folio);
 593	/*
 594	 * Lazyfree folios are clean anonymous folios.  They have
 595	 * the swapbacked flag cleared, to distinguish them from normal
 596	 * anonymous folios
 597	 */
 598	folio_clear_swapbacked(folio);
 599	lruvec_add_folio(lruvec, folio);
 600
 601	__count_vm_events(PGLAZYFREE, nr_pages);
 602	__count_memcg_events(lruvec_memcg(lruvec), PGLAZYFREE, nr_pages);
 603}
 604
 605/*
 606 * Drain pages out of the cpu's folio_batch.
 607 * Either "cpu" is the current CPU, and preemption has already been
 608 * disabled; or "cpu" is being hot-unplugged, and is already dead.
 609 */
 610void lru_add_drain_cpu(int cpu)
 611{
 612	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
 613	struct folio_batch *fbatch = &fbatches->lru_add;
 614
 615	if (folio_batch_count(fbatch))
 616		folio_batch_move_lru(fbatch, lru_add);
 617
 618	fbatch = &fbatches->lru_move_tail;
 619	/* Disabling interrupts below acts as a compiler barrier. */
 620	if (data_race(folio_batch_count(fbatch))) {
 621		unsigned long flags;
 622
 623		/* No harm done if a racing interrupt already did this */
 624		local_lock_irqsave(&cpu_fbatches.lock_irq, flags);
 625		folio_batch_move_lru(fbatch, lru_move_tail);
 626		local_unlock_irqrestore(&cpu_fbatches.lock_irq, flags);
 627	}
 628
 629	fbatch = &fbatches->lru_deactivate_file;
 630	if (folio_batch_count(fbatch))
 631		folio_batch_move_lru(fbatch, lru_deactivate_file);
 632
 633	fbatch = &fbatches->lru_deactivate;
 634	if (folio_batch_count(fbatch))
 635		folio_batch_move_lru(fbatch, lru_deactivate);
 636
 637	fbatch = &fbatches->lru_lazyfree;
 638	if (folio_batch_count(fbatch))
 639		folio_batch_move_lru(fbatch, lru_lazyfree);
 640
 641	folio_activate_drain(cpu);
 642}
 643
 644/**
 645 * deactivate_file_folio() - Deactivate a file folio.
 646 * @folio: Folio to deactivate.
 647 *
 648 * This function hints to the VM that @folio is a good reclaim candidate,
 649 * for example if its invalidation fails due to the folio being dirty
 650 * or under writeback.
 651 *
 652 * Context: Caller holds a reference on the folio.
 653 */
 654void deactivate_file_folio(struct folio *folio)
 655{
 
 
 656	/* Deactivating an unevictable folio will not accelerate reclaim */
 657	if (folio_test_unevictable(folio))
 658		return;
 659
 660	folio_batch_add_and_move(folio, lru_deactivate_file, true);
 
 
 
 
 661}
 662
 663/*
 664 * folio_deactivate - deactivate a folio
 665 * @folio: folio to deactivate
 666 *
 667 * folio_deactivate() moves @folio to the inactive list if @folio was on the
 668 * active list and was not unevictable. This is done to accelerate the
 669 * reclaim of @folio.
 670 */
 671void folio_deactivate(struct folio *folio)
 672{
 673	if (folio_test_unevictable(folio) || !(folio_test_active(folio) || lru_gen_enabled()))
 674		return;
 
 675
 676	folio_batch_add_and_move(folio, lru_deactivate, true);
 
 
 
 
 
 677}
 678
 679/**
 680 * folio_mark_lazyfree - make an anon folio lazyfree
 681 * @folio: folio to deactivate
 682 *
 683 * folio_mark_lazyfree() moves @folio to the inactive file list.
 684 * This is done to accelerate the reclaim of @folio.
 685 */
 686void folio_mark_lazyfree(struct folio *folio)
 687{
 688	if (!folio_test_anon(folio) || !folio_test_swapbacked(folio) ||
 689	    folio_test_swapcache(folio) || folio_test_unevictable(folio))
 690		return;
 
 691
 692	folio_batch_add_and_move(folio, lru_lazyfree, true);
 
 
 
 
 
 693}
 694
 695void lru_add_drain(void)
 696{
 697	local_lock(&cpu_fbatches.lock);
 698	lru_add_drain_cpu(smp_processor_id());
 699	local_unlock(&cpu_fbatches.lock);
 700	mlock_drain_local();
 701}
 702
 703/*
 704 * It's called from per-cpu workqueue context in SMP case so
 705 * lru_add_drain_cpu and invalidate_bh_lrus_cpu should run on
 706 * the same cpu. It shouldn't be a problem in !SMP case since
 707 * the core is only one and the locks will disable preemption.
 708 */
 709static void lru_add_and_bh_lrus_drain(void)
 710{
 711	local_lock(&cpu_fbatches.lock);
 712	lru_add_drain_cpu(smp_processor_id());
 713	local_unlock(&cpu_fbatches.lock);
 714	invalidate_bh_lrus_cpu();
 715	mlock_drain_local();
 716}
 717
 718void lru_add_drain_cpu_zone(struct zone *zone)
 719{
 720	local_lock(&cpu_fbatches.lock);
 721	lru_add_drain_cpu(smp_processor_id());
 722	drain_local_pages(zone);
 723	local_unlock(&cpu_fbatches.lock);
 724	mlock_drain_local();
 725}
 726
 727#ifdef CONFIG_SMP
 728
 729static DEFINE_PER_CPU(struct work_struct, lru_add_drain_work);
 730
 731static void lru_add_drain_per_cpu(struct work_struct *dummy)
 732{
 733	lru_add_and_bh_lrus_drain();
 734}
 735
 736static bool cpu_needs_drain(unsigned int cpu)
 737{
 738	struct cpu_fbatches *fbatches = &per_cpu(cpu_fbatches, cpu);
 739
 740	/* Check these in order of likelihood that they're not zero */
 741	return folio_batch_count(&fbatches->lru_add) ||
 742		folio_batch_count(&fbatches->lru_move_tail) ||
 743		folio_batch_count(&fbatches->lru_deactivate_file) ||
 744		folio_batch_count(&fbatches->lru_deactivate) ||
 745		folio_batch_count(&fbatches->lru_lazyfree) ||
 746		folio_batch_count(&fbatches->lru_activate) ||
 747		need_mlock_drain(cpu) ||
 748		has_bh_in_lru(cpu, NULL);
 749}
 750
 751/*
 752 * Doesn't need any cpu hotplug locking because we do rely on per-cpu
 753 * kworkers being shut down before our page_alloc_cpu_dead callback is
 754 * executed on the offlined cpu.
 755 * Calling this function with cpu hotplug locks held can actually lead
 756 * to obscure indirect dependencies via WQ context.
 757 */
 758static inline void __lru_add_drain_all(bool force_all_cpus)
 759{
 760	/*
 761	 * lru_drain_gen - Global pages generation number
 762	 *
 763	 * (A) Definition: global lru_drain_gen = x implies that all generations
 764	 *     0 < n <= x are already *scheduled* for draining.
 765	 *
 766	 * This is an optimization for the highly-contended use case where a
 767	 * user space workload keeps constantly generating a flow of pages for
 768	 * each CPU.
 769	 */
 770	static unsigned int lru_drain_gen;
 771	static struct cpumask has_work;
 772	static DEFINE_MUTEX(lock);
 773	unsigned cpu, this_gen;
 774
 775	/*
 776	 * Make sure nobody triggers this path before mm_percpu_wq is fully
 777	 * initialized.
 778	 */
 779	if (WARN_ON(!mm_percpu_wq))
 780		return;
 781
 782	/*
 783	 * Guarantee folio_batch counter stores visible by this CPU
 784	 * are visible to other CPUs before loading the current drain
 785	 * generation.
 786	 */
 787	smp_mb();
 788
 789	/*
 790	 * (B) Locally cache global LRU draining generation number
 791	 *
 792	 * The read barrier ensures that the counter is loaded before the mutex
 793	 * is taken. It pairs with smp_mb() inside the mutex critical section
 794	 * at (D).
 795	 */
 796	this_gen = smp_load_acquire(&lru_drain_gen);
 797
 798	mutex_lock(&lock);
 799
 800	/*
 801	 * (C) Exit the draining operation if a newer generation, from another
 802	 * lru_add_drain_all(), was already scheduled for draining. Check (A).
 803	 */
 804	if (unlikely(this_gen != lru_drain_gen && !force_all_cpus))
 805		goto done;
 806
 807	/*
 808	 * (D) Increment global generation number
 809	 *
 810	 * Pairs with smp_load_acquire() at (B), outside of the critical
 811	 * section. Use a full memory barrier to guarantee that the
 812	 * new global drain generation number is stored before loading
 813	 * folio_batch counters.
 814	 *
 815	 * This pairing must be done here, before the for_each_online_cpu loop
 816	 * below which drains the page vectors.
 817	 *
 818	 * Let x, y, and z represent some system CPU numbers, where x < y < z.
 819	 * Assume CPU #z is in the middle of the for_each_online_cpu loop
 820	 * below and has already reached CPU #y's per-cpu data. CPU #x comes
 821	 * along, adds some pages to its per-cpu vectors, then calls
 822	 * lru_add_drain_all().
 823	 *
 824	 * If the paired barrier is done at any later step, e.g. after the
 825	 * loop, CPU #x will just exit at (C) and miss flushing out all of its
 826	 * added pages.
 827	 */
 828	WRITE_ONCE(lru_drain_gen, lru_drain_gen + 1);
 829	smp_mb();
 830
 831	cpumask_clear(&has_work);
 832	for_each_online_cpu(cpu) {
 833		struct work_struct *work = &per_cpu(lru_add_drain_work, cpu);
 834
 835		if (cpu_needs_drain(cpu)) {
 836			INIT_WORK(work, lru_add_drain_per_cpu);
 837			queue_work_on(cpu, mm_percpu_wq, work);
 838			__cpumask_set_cpu(cpu, &has_work);
 839		}
 840	}
 841
 842	for_each_cpu(cpu, &has_work)
 843		flush_work(&per_cpu(lru_add_drain_work, cpu));
 844
 845done:
 846	mutex_unlock(&lock);
 847}
 848
 849void lru_add_drain_all(void)
 850{
 851	__lru_add_drain_all(false);
 852}
 853#else
 854void lru_add_drain_all(void)
 855{
 856	lru_add_drain();
 857}
 858#endif /* CONFIG_SMP */
 859
 860atomic_t lru_disable_count = ATOMIC_INIT(0);
 861
 862/*
 863 * lru_cache_disable() needs to be called before we start compiling
 864 * a list of folios to be migrated using folio_isolate_lru().
 865 * It drains folios on LRU cache and then disable on all cpus until
 866 * lru_cache_enable is called.
 867 *
 868 * Must be paired with a call to lru_cache_enable().
 869 */
 870void lru_cache_disable(void)
 871{
 872	atomic_inc(&lru_disable_count);
 873	/*
 874	 * Readers of lru_disable_count are protected by either disabling
 875	 * preemption or rcu_read_lock:
 876	 *
 877	 * preempt_disable, local_irq_disable  [bh_lru_lock()]
 878	 * rcu_read_lock		       [rt_spin_lock CONFIG_PREEMPT_RT]
 879	 * preempt_disable		       [local_lock !CONFIG_PREEMPT_RT]
 880	 *
 881	 * Since v5.1 kernel, synchronize_rcu() is guaranteed to wait on
 882	 * preempt_disable() regions of code. So any CPU which sees
 883	 * lru_disable_count = 0 will have exited the critical
 884	 * section when synchronize_rcu() returns.
 885	 */
 886	synchronize_rcu_expedited();
 887#ifdef CONFIG_SMP
 888	__lru_add_drain_all(true);
 889#else
 890	lru_add_and_bh_lrus_drain();
 891#endif
 892}
 893
 894/**
 895 * folios_put_refs - Reduce the reference count on a batch of folios.
 896 * @folios: The folios.
 897 * @refs: The number of refs to subtract from each folio.
 898 *
 899 * Like folio_put(), but for a batch of folios.  This is more efficient
 900 * than writing the loop yourself as it will optimise the locks which need
 901 * to be taken if the folios are freed.  The folios batch is returned
 902 * empty and ready to be reused for another batch; there is no need
 903 * to reinitialise it.  If @refs is NULL, we subtract one from each
 904 * folio refcount.
 905 *
 906 * Context: May be called in process or interrupt context, but not in NMI
 907 * context.  May be called while holding a spinlock.
 908 */
 909void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 910{
 911	int i, j;
 912	struct lruvec *lruvec = NULL;
 913	unsigned long flags = 0;
 914
 915	for (i = 0, j = 0; i < folios->nr; i++) {
 916		struct folio *folio = folios->folios[i];
 917		unsigned int nr_refs = refs ? refs[i] : 1;
 918
 919		if (is_huge_zero_folio(folio))
 920			continue;
 921
 922		if (folio_is_zone_device(folio)) {
 923			if (lruvec) {
 924				unlock_page_lruvec_irqrestore(lruvec, flags);
 925				lruvec = NULL;
 926			}
 927			if (put_devmap_managed_folio_refs(folio, nr_refs))
 928				continue;
 929			if (folio_ref_sub_and_test(folio, nr_refs))
 930				free_zone_device_folio(folio);
 931			continue;
 932		}
 933
 934		if (!folio_ref_sub_and_test(folio, nr_refs))
 935			continue;
 936
 937		/* hugetlb has its own memcg */
 938		if (folio_test_hugetlb(folio)) {
 939			if (lruvec) {
 940				unlock_page_lruvec_irqrestore(lruvec, flags);
 941				lruvec = NULL;
 942			}
 943			free_huge_folio(folio);
 944			continue;
 945		}
 946		folio_unqueue_deferred_split(folio);
 
 
 
 947		__page_cache_release(folio, &lruvec, &flags);
 948
 949		if (j != i)
 950			folios->folios[j] = folio;
 951		j++;
 952	}
 953	if (lruvec)
 954		unlock_page_lruvec_irqrestore(lruvec, flags);
 955	if (!j) {
 956		folio_batch_reinit(folios);
 957		return;
 958	}
 959
 960	folios->nr = j;
 961	mem_cgroup_uncharge_folios(folios);
 962	free_unref_folios(folios);
 963}
 964EXPORT_SYMBOL(folios_put_refs);
 965
 966/**
 967 * release_pages - batched put_page()
 968 * @arg: array of pages to release
 969 * @nr: number of pages
 970 *
 971 * Decrement the reference count on all the pages in @arg.  If it
 972 * fell to zero, remove the page from the LRU and free it.
 973 *
 974 * Note that the argument can be an array of pages, encoded pages,
 975 * or folio pointers. We ignore any encoded bits, and turn any of
 976 * them into just a folio that gets free'd.
 977 */
 978void release_pages(release_pages_arg arg, int nr)
 979{
 980	struct folio_batch fbatch;
 981	int refs[PAGEVEC_SIZE];
 982	struct encoded_page **encoded = arg.encoded_pages;
 983	int i;
 984
 985	folio_batch_init(&fbatch);
 986	for (i = 0; i < nr; i++) {
 987		/* Turn any of the argument types into a folio */
 988		struct folio *folio = page_folio(encoded_page_ptr(encoded[i]));
 989
 990		/* Is our next entry actually "nr_pages" -> "nr_refs" ? */
 991		refs[fbatch.nr] = 1;
 992		if (unlikely(encoded_page_flags(encoded[i]) &
 993			     ENCODED_PAGE_BIT_NR_PAGES_NEXT))
 994			refs[fbatch.nr] = encoded_nr_pages(encoded[++i]);
 995
 996		if (folio_batch_add(&fbatch, folio) > 0)
 997			continue;
 998		folios_put_refs(&fbatch, refs);
 999	}
1000
1001	if (fbatch.nr)
1002		folios_put_refs(&fbatch, refs);
1003}
1004EXPORT_SYMBOL(release_pages);
1005
1006/*
1007 * The folios which we're about to release may be in the deferred lru-addition
1008 * queues.  That would prevent them from really being freed right now.  That's
1009 * OK from a correctness point of view but is inefficient - those folios may be
1010 * cache-warm and we want to give them back to the page allocator ASAP.
1011 *
1012 * So __folio_batch_release() will drain those queues here.
1013 * folio_batch_move_lru() calls folios_put() directly to avoid
1014 * mutual recursion.
1015 */
1016void __folio_batch_release(struct folio_batch *fbatch)
1017{
1018	if (!fbatch->percpu_pvec_drained) {
1019		lru_add_drain();
1020		fbatch->percpu_pvec_drained = true;
1021	}
1022	folios_put(fbatch);
1023}
1024EXPORT_SYMBOL(__folio_batch_release);
1025
1026/**
1027 * folio_batch_remove_exceptionals() - Prune non-folios from a batch.
1028 * @fbatch: The batch to prune
1029 *
1030 * find_get_entries() fills a batch with both folios and shadow/swap/DAX
1031 * entries.  This function prunes all the non-folio entries from @fbatch
1032 * without leaving holes, so that it can be passed on to folio-only batch
1033 * operations.
1034 */
1035void folio_batch_remove_exceptionals(struct folio_batch *fbatch)
1036{
1037	unsigned int i, j;
1038
1039	for (i = 0, j = 0; i < folio_batch_count(fbatch); i++) {
1040		struct folio *folio = fbatch->folios[i];
1041		if (!xa_is_value(folio))
1042			fbatch->folios[j++] = folio;
1043	}
1044	fbatch->nr = j;
1045}
1046
1047/*
1048 * Perform any setup for the swap system
1049 */
1050void __init swap_setup(void)
1051{
1052	unsigned long megs = totalram_pages() >> (20 - PAGE_SHIFT);
1053
1054	/* Use a smaller cluster for small-memory machines */
1055	if (megs < 16)
1056		page_cluster = 2;
1057	else
1058		page_cluster = 3;
1059	/*
1060	 * Right now other parts of the system means that we
1061	 * _really_ don't want to cluster much more
1062	 */
1063}