khugepaged.c

   1// SPDX-License-Identifier: GPL-2.0
   2#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   3
   4#include <linux/mm.h>
   5#include <linux/sched.h>
   6#include <linux/sched/mm.h>
   7#include <linux/mmu_notifier.h>
   8#include <linux/rmap.h>
   9#include <linux/swap.h>
  10#include <linux/mm_inline.h>
  11#include <linux/kthread.h>
  12#include <linux/khugepaged.h>
  13#include <linux/freezer.h>
  14#include <linux/mman.h>
  15#include <linux/hashtable.h>
  16#include <linux/userfaultfd_k.h>
  17#include <linux/page_idle.h>
  18#include <linux/page_table_check.h>
  19#include <linux/rcupdate_wait.h>
  20#include <linux/leafops.h>
  21#include <linux/shmem_fs.h>
  22#include <linux/dax.h>
  23#include <linux/ksm.h>
  24#include <linux/pgalloc.h>
  25
  26#include <asm/tlb.h>
  27#include "internal.h"
  28#include "mm_slot.h"
  29
  30enum scan_result {
  31	SCAN_FAIL,
  32	SCAN_SUCCEED,
  33	SCAN_NO_PTE_TABLE,
  34	SCAN_PMD_MAPPED,
  35	SCAN_EXCEED_NONE_PTE,
  36	SCAN_EXCEED_SWAP_PTE,
  37	SCAN_EXCEED_SHARED_PTE,
  38	SCAN_PTE_NON_PRESENT,
  39	SCAN_PTE_UFFD_WP,
  40	SCAN_PTE_MAPPED_HUGEPAGE,
  41	SCAN_LACK_REFERENCED_PAGE,
  42	SCAN_PAGE_NULL,
  43	SCAN_SCAN_ABORT,
  44	SCAN_PAGE_COUNT,
  45	SCAN_PAGE_LRU,
  46	SCAN_PAGE_LOCK,
  47	SCAN_PAGE_ANON,
  48	SCAN_PAGE_COMPOUND,
  49	SCAN_ANY_PROCESS,
  50	SCAN_VMA_NULL,
  51	SCAN_VMA_CHECK,
  52	SCAN_ADDRESS_RANGE,
  53	SCAN_DEL_PAGE_LRU,
  54	SCAN_ALLOC_HUGE_PAGE_FAIL,
  55	SCAN_CGROUP_CHARGE_FAIL,
  56	SCAN_TRUNCATED,
  57	SCAN_PAGE_HAS_PRIVATE,
  58	SCAN_STORE_FAILED,
  59	SCAN_COPY_MC,
  60	SCAN_PAGE_FILLED,
  61};
  62
  63#define CREATE_TRACE_POINTS
  64#include <trace/events/huge_memory.h>
  65
  66static struct task_struct *khugepaged_thread __read_mostly;
  67static DEFINE_MUTEX(khugepaged_mutex);
  68
  69/* default scan 8*HPAGE_PMD_NR ptes (or vmas) every 10 second */
  70static unsigned int khugepaged_pages_to_scan __read_mostly;
  71static unsigned int khugepaged_pages_collapsed;
  72static unsigned int khugepaged_full_scans;
  73static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  74/* during fragmentation poll the hugepage allocator once every minute */
  75static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  76static unsigned long khugepaged_sleep_expire;
  77static DEFINE_SPINLOCK(khugepaged_mm_lock);
  78static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  79/*
  80 * default collapse hugepages if there is at least one pte mapped like
  81 * it would have happened if the vma was large enough during page
  82 * fault.
  83 *
  84 * Note that these are only respected if collapse was initiated by khugepaged.
  85 */
  86unsigned int khugepaged_max_ptes_none __read_mostly;
  87static unsigned int khugepaged_max_ptes_swap __read_mostly;
  88static unsigned int khugepaged_max_ptes_shared __read_mostly;
  89
  90#define MM_SLOTS_HASH_BITS 10
  91static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  92
  93static struct kmem_cache *mm_slot_cache __ro_after_init;
  94
  95struct collapse_control {
  96	bool is_khugepaged;
  97
  98	/* Num pages scanned per node */
  99	u32 node_load[MAX_NUMNODES];
 100
 101	/* nodemask for allocation fallback */
 102	nodemask_t alloc_nmask;
 103};
 104
 105/**
 106 * struct khugepaged_scan - cursor for scanning
 107 * @mm_head: the head of the mm list to scan
 108 * @mm_slot: the current mm_slot we are scanning
 109 * @address: the next address inside that to be scanned
 110 *
 111 * There is only the one khugepaged_scan instance of this cursor structure.
 112 */
 113struct khugepaged_scan {
 114	struct list_head mm_head;
 115	struct mm_slot *mm_slot;
 116	unsigned long address;
 117};
 118
 119static struct khugepaged_scan khugepaged_scan = {
 120	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 121};
 122
 123#ifdef CONFIG_SYSFS
 124static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 125					 struct kobj_attribute *attr,
 126					 char *buf)
 127{
 128	return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 129}
 130
 131static ssize_t __sleep_millisecs_store(const char *buf, size_t count,
 132				       unsigned int *millisecs)
 133{
 134	unsigned int msecs;
 135	int err;
 136
 137	err = kstrtouint(buf, 10, &msecs);
 138	if (err)
 139		return -EINVAL;
 140
 141	*millisecs = msecs;
 142	khugepaged_sleep_expire = 0;
 143	wake_up_interruptible(&khugepaged_wait);
 144
 145	return count;
 146}
 147
 148static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 149					  struct kobj_attribute *attr,
 150					  const char *buf, size_t count)
 151{
 152	return __sleep_millisecs_store(buf, count, &khugepaged_scan_sleep_millisecs);
 153}
 154static struct kobj_attribute scan_sleep_millisecs_attr =
 155	__ATTR_RW(scan_sleep_millisecs);
 156
 157static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 158					  struct kobj_attribute *attr,
 159					  char *buf)
 160{
 161	return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 162}
 163
 164static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 165					   struct kobj_attribute *attr,
 166					   const char *buf, size_t count)
 167{
 168	return __sleep_millisecs_store(buf, count, &khugepaged_alloc_sleep_millisecs);
 169}
 170static struct kobj_attribute alloc_sleep_millisecs_attr =
 171	__ATTR_RW(alloc_sleep_millisecs);
 172
 173static ssize_t pages_to_scan_show(struct kobject *kobj,
 174				  struct kobj_attribute *attr,
 175				  char *buf)
 176{
 177	return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);
 178}
 179static ssize_t pages_to_scan_store(struct kobject *kobj,
 180				   struct kobj_attribute *attr,
 181				   const char *buf, size_t count)
 182{
 183	unsigned int pages;
 184	int err;
 185
 186	err = kstrtouint(buf, 10, &pages);
 187	if (err || !pages)
 188		return -EINVAL;
 189
 190	khugepaged_pages_to_scan = pages;
 191
 192	return count;
 193}
 194static struct kobj_attribute pages_to_scan_attr =
 195	__ATTR_RW(pages_to_scan);
 196
 197static ssize_t pages_collapsed_show(struct kobject *kobj,
 198				    struct kobj_attribute *attr,
 199				    char *buf)
 200{
 201	return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);
 202}
 203static struct kobj_attribute pages_collapsed_attr =
 204	__ATTR_RO(pages_collapsed);
 205
 206static ssize_t full_scans_show(struct kobject *kobj,
 207			       struct kobj_attribute *attr,
 208			       char *buf)
 209{
 210	return sysfs_emit(buf, "%u\n", khugepaged_full_scans);
 211}
 212static struct kobj_attribute full_scans_attr =
 213	__ATTR_RO(full_scans);
 214
 215static ssize_t defrag_show(struct kobject *kobj,
 216			   struct kobj_attribute *attr, char *buf)
 217{
 218	return single_hugepage_flag_show(kobj, attr, buf,
 219					 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 220}
 221static ssize_t defrag_store(struct kobject *kobj,
 222			    struct kobj_attribute *attr,
 223			    const char *buf, size_t count)
 224{
 225	return single_hugepage_flag_store(kobj, attr, buf, count,
 226				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 227}
 228static struct kobj_attribute khugepaged_defrag_attr =
 229	__ATTR_RW(defrag);
 230
 231/*
 232 * max_ptes_none controls if khugepaged should collapse hugepages over
 233 * any unmapped ptes in turn potentially increasing the memory
 234 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 235 * reduce the available free memory in the system as it
 236 * runs. Increasing max_ptes_none will instead potentially reduce the
 237 * free memory in the system during the khugepaged scan.
 238 */
 239static ssize_t max_ptes_none_show(struct kobject *kobj,
 240				  struct kobj_attribute *attr,
 241				  char *buf)
 242{
 243	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);
 244}
 245static ssize_t max_ptes_none_store(struct kobject *kobj,
 246				   struct kobj_attribute *attr,
 247				   const char *buf, size_t count)
 248{
 249	int err;
 250	unsigned long max_ptes_none;
 251
 252	err = kstrtoul(buf, 10, &max_ptes_none);
 253	if (err || max_ptes_none > HPAGE_PMD_NR - 1)
 254		return -EINVAL;
 255
 256	khugepaged_max_ptes_none = max_ptes_none;
 257
 258	return count;
 259}
 260static struct kobj_attribute khugepaged_max_ptes_none_attr =
 261	__ATTR_RW(max_ptes_none);
 262
 263static ssize_t max_ptes_swap_show(struct kobject *kobj,
 264				  struct kobj_attribute *attr,
 265				  char *buf)
 266{
 267	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);
 268}
 269
 270static ssize_t max_ptes_swap_store(struct kobject *kobj,
 271				   struct kobj_attribute *attr,
 272				   const char *buf, size_t count)
 273{
 274	int err;
 275	unsigned long max_ptes_swap;
 276
 277	err  = kstrtoul(buf, 10, &max_ptes_swap);
 278	if (err || max_ptes_swap > HPAGE_PMD_NR - 1)
 279		return -EINVAL;
 280
 281	khugepaged_max_ptes_swap = max_ptes_swap;
 282
 283	return count;
 284}
 285
 286static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 287	__ATTR_RW(max_ptes_swap);
 288
 289static ssize_t max_ptes_shared_show(struct kobject *kobj,
 290				    struct kobj_attribute *attr,
 291				    char *buf)
 292{
 293	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);
 294}
 295
 296static ssize_t max_ptes_shared_store(struct kobject *kobj,
 297				     struct kobj_attribute *attr,
 298				     const char *buf, size_t count)
 299{
 300	int err;
 301	unsigned long max_ptes_shared;
 302
 303	err  = kstrtoul(buf, 10, &max_ptes_shared);
 304	if (err || max_ptes_shared > HPAGE_PMD_NR - 1)
 305		return -EINVAL;
 306
 307	khugepaged_max_ptes_shared = max_ptes_shared;
 308
 309	return count;
 310}
 311
 312static struct kobj_attribute khugepaged_max_ptes_shared_attr =
 313	__ATTR_RW(max_ptes_shared);
 314
 315static struct attribute *khugepaged_attr[] = {
 316	&khugepaged_defrag_attr.attr,
 317	&khugepaged_max_ptes_none_attr.attr,
 318	&khugepaged_max_ptes_swap_attr.attr,
 319	&khugepaged_max_ptes_shared_attr.attr,
 320	&pages_to_scan_attr.attr,
 321	&pages_collapsed_attr.attr,
 322	&full_scans_attr.attr,
 323	&scan_sleep_millisecs_attr.attr,
 324	&alloc_sleep_millisecs_attr.attr,
 325	NULL,
 326};
 327
 328struct attribute_group khugepaged_attr_group = {
 329	.attrs = khugepaged_attr,
 330	.name = "khugepaged",
 331};
 332#endif /* CONFIG_SYSFS */
 333
 334static bool pte_none_or_zero(pte_t pte)
 335{
 336	if (pte_none(pte))
 337		return true;
 338	return pte_present(pte) && is_zero_pfn(pte_pfn(pte));
 339}
 340
 341int hugepage_madvise(struct vm_area_struct *vma,
 342		     vm_flags_t *vm_flags, int advice)
 343{
 344	switch (advice) {
 345	case MADV_HUGEPAGE:
 346#ifdef CONFIG_S390
 347		/*
 348		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 349		 * can't handle this properly after s390_enable_sie, so we simply
 350		 * ignore the madvise to prevent qemu from causing a SIGSEGV.
 351		 */
 352		if (mm_has_pgste(vma->vm_mm))
 353			return 0;
 354#endif
 355		*vm_flags &= ~VM_NOHUGEPAGE;
 356		*vm_flags |= VM_HUGEPAGE;
 357		/*
 358		 * If the vma become good for khugepaged to scan,
 359		 * register it here without waiting a page fault that
 360		 * may not happen any time soon.
 361		 */
 362		khugepaged_enter_vma(vma, *vm_flags);
 363		break;
 364	case MADV_NOHUGEPAGE:
 365		*vm_flags &= ~VM_HUGEPAGE;
 366		*vm_flags |= VM_NOHUGEPAGE;
 367		/*
 368		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 369		 * this vma even if we leave the mm registered in khugepaged if
 370		 * it got registered before VM_NOHUGEPAGE was set.
 371		 */
 372		break;
 373	}
 374
 375	return 0;
 376}
 377
 378int __init khugepaged_init(void)
 379{
 380	mm_slot_cache = KMEM_CACHE(mm_slot, 0);
 381	if (!mm_slot_cache)
 382		return -ENOMEM;
 383
 384	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 385	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 386	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 387	khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;
 388
 389	return 0;
 390}
 391
 392void __init khugepaged_destroy(void)
 393{
 394	kmem_cache_destroy(mm_slot_cache);
 395}
 396
 397static inline int hpage_collapse_test_exit(struct mm_struct *mm)
 398{
 399	return atomic_read(&mm->mm_users) == 0;
 400}
 401
 402static inline int hpage_collapse_test_exit_or_disable(struct mm_struct *mm)
 403{
 404	return hpage_collapse_test_exit(mm) ||
 405		mm_flags_test(MMF_DISABLE_THP_COMPLETELY, mm);
 406}
 407
 408static bool hugepage_pmd_enabled(void)
 409{
 410	/*
 411	 * We cover the anon, shmem and the file-backed case here; file-backed
 412	 * hugepages, when configured in, are determined by the global control.
 413	 * Anon pmd-sized hugepages are determined by the pmd-size control.
 414	 * Shmem pmd-sized hugepages are also determined by its pmd-size control,
 415	 * except when the global shmem_huge is set to SHMEM_HUGE_DENY.
 416	 */
 417	if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
 418	    hugepage_global_enabled())
 419		return true;
 420	if (test_bit(PMD_ORDER, &huge_anon_orders_always))
 421		return true;
 422	if (test_bit(PMD_ORDER, &huge_anon_orders_madvise))
 423		return true;
 424	if (test_bit(PMD_ORDER, &huge_anon_orders_inherit) &&
 425	    hugepage_global_enabled())
 426		return true;
 427	if (IS_ENABLED(CONFIG_SHMEM) && shmem_hpage_pmd_enabled())
 428		return true;
 429	return false;
 430}
 431
 432void __khugepaged_enter(struct mm_struct *mm)
 433{
 434	struct mm_slot *slot;
 435	int wakeup;
 436
 437	/* __khugepaged_exit() must not run from under us */
 438	VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);
 439	if (unlikely(mm_flags_test_and_set(MMF_VM_HUGEPAGE, mm)))
 440		return;
 441
 442	slot = mm_slot_alloc(mm_slot_cache);
 443	if (!slot)
 444		return;
 445
 446	spin_lock(&khugepaged_mm_lock);
 447	mm_slot_insert(mm_slots_hash, mm, slot);
 448	/*
 449	 * Insert just behind the scanning cursor, to let the area settle
 450	 * down a little.
 451	 */
 452	wakeup = list_empty(&khugepaged_scan.mm_head);
 453	list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);
 454	spin_unlock(&khugepaged_mm_lock);
 455
 456	mmgrab(mm);
 457	if (wakeup)
 458		wake_up_interruptible(&khugepaged_wait);
 459}
 460
 461void khugepaged_enter_vma(struct vm_area_struct *vma,
 462			  vm_flags_t vm_flags)
 463{
 464	if (!mm_flags_test(MMF_VM_HUGEPAGE, vma->vm_mm) &&
 465	    hugepage_pmd_enabled()) {
 466		if (thp_vma_allowable_order(vma, vm_flags, TVA_KHUGEPAGED, PMD_ORDER))
 467			__khugepaged_enter(vma->vm_mm);
 468	}
 469}
 470
 471void __khugepaged_exit(struct mm_struct *mm)
 472{
 473	struct mm_slot *slot;
 474	int free = 0;
 475
 476	spin_lock(&khugepaged_mm_lock);
 477	slot = mm_slot_lookup(mm_slots_hash, mm);
 478	if (slot && khugepaged_scan.mm_slot != slot) {
 479		hash_del(&slot->hash);
 480		list_del(&slot->mm_node);
 481		free = 1;
 482	}
 483	spin_unlock(&khugepaged_mm_lock);
 484
 485	if (free) {
 486		mm_flags_clear(MMF_VM_HUGEPAGE, mm);
 487		mm_slot_free(mm_slot_cache, slot);
 488		mmdrop(mm);
 489	} else if (slot) {
 490		/*
 491		 * This is required to serialize against
 492		 * hpage_collapse_test_exit() (which is guaranteed to run
 493		 * under mmap sem read mode). Stop here (after we return all
 494		 * pagetables will be destroyed) until khugepaged has finished
 495		 * working on the pagetables under the mmap_lock.
 496		 */
 497		mmap_write_lock(mm);
 498		mmap_write_unlock(mm);
 499	}
 500}
 501
 502static void release_pte_folio(struct folio *folio)
 503{
 504	node_stat_mod_folio(folio,
 505			NR_ISOLATED_ANON + folio_is_file_lru(folio),
 506			-folio_nr_pages(folio));
 507	folio_unlock(folio);
 508	folio_putback_lru(folio);
 509}
 510
 511static void release_pte_pages(pte_t *pte, pte_t *_pte,
 512		struct list_head *compound_pagelist)
 513{
 514	struct folio *folio, *tmp;
 515
 516	while (--_pte >= pte) {
 517		pte_t pteval = ptep_get(_pte);
 518		unsigned long pfn;
 519
 520		if (pte_none(pteval))
 521			continue;
 522		VM_WARN_ON_ONCE(!pte_present(pteval));
 523		pfn = pte_pfn(pteval);
 524		if (is_zero_pfn(pfn))
 525			continue;
 526		folio = pfn_folio(pfn);
 527		if (folio_test_large(folio))
 528			continue;
 529		release_pte_folio(folio);
 530	}
 531
 532	list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) {
 533		list_del(&folio->lru);
 534		release_pte_folio(folio);
 535	}
 536}
 537
 538static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 539					unsigned long start_addr,
 540					pte_t *pte,
 541					struct collapse_control *cc,
 542					struct list_head *compound_pagelist)
 543{
 544	struct page *page = NULL;
 545	struct folio *folio = NULL;
 546	unsigned long addr = start_addr;
 547	pte_t *_pte;
 548	int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;
 549
 550	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
 551	     _pte++, addr += PAGE_SIZE) {
 552		pte_t pteval = ptep_get(_pte);
 553		if (pte_none_or_zero(pteval)) {
 554			++none_or_zero;
 555			if (!userfaultfd_armed(vma) &&
 556			    (!cc->is_khugepaged ||
 557			     none_or_zero <= khugepaged_max_ptes_none)) {
 558				continue;
 559			} else {
 560				result = SCAN_EXCEED_NONE_PTE;
 561				count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
 562				goto out;
 563			}
 564		}
 565		if (!pte_present(pteval)) {
 566			result = SCAN_PTE_NON_PRESENT;
 567			goto out;
 568		}
 569		if (pte_uffd_wp(pteval)) {
 570			result = SCAN_PTE_UFFD_WP;
 571			goto out;
 572		}
 573		page = vm_normal_page(vma, addr, pteval);
 574		if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
 575			result = SCAN_PAGE_NULL;
 576			goto out;
 577		}
 578
 579		folio = page_folio(page);
 580		VM_BUG_ON_FOLIO(!folio_test_anon(folio), folio);
 581
 582		/* See hpage_collapse_scan_pmd(). */
 583		if (folio_maybe_mapped_shared(folio)) {
 584			++shared;
 585			if (cc->is_khugepaged &&
 586			    shared > khugepaged_max_ptes_shared) {
 587				result = SCAN_EXCEED_SHARED_PTE;
 588				count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
 589				goto out;
 590			}
 591		}
 592
 593		if (folio_test_large(folio)) {
 594			struct folio *f;
 595
 596			/*
 597			 * Check if we have dealt with the compound page
 598			 * already
 599			 */
 600			list_for_each_entry(f, compound_pagelist, lru) {
 601				if (folio == f)
 602					goto next;
 603			}
 604		}
 605
 606		/*
 607		 * We can do it before folio_isolate_lru because the
 608		 * folio can't be freed from under us. NOTE: PG_lock
 609		 * is needed to serialize against split_huge_page
 610		 * when invoked from the VM.
 611		 */
 612		if (!folio_trylock(folio)) {
 613			result = SCAN_PAGE_LOCK;
 614			goto out;
 615		}
 616
 617		/*
 618		 * Check if the page has any GUP (or other external) pins.
 619		 *
 620		 * The page table that maps the page has been already unlinked
 621		 * from the page table tree and this process cannot get
 622		 * an additional pin on the page.
 623		 *
 624		 * New pins can come later if the page is shared across fork,
 625		 * but not from this process. The other process cannot write to
 626		 * the page, only trigger CoW.
 627		 */
 628		if (folio_expected_ref_count(folio) != folio_ref_count(folio)) {
 629			folio_unlock(folio);
 630			result = SCAN_PAGE_COUNT;
 631			goto out;
 632		}
 633
 634		/*
 635		 * Isolate the page to avoid collapsing an hugepage
 636		 * currently in use by the VM.
 637		 */
 638		if (!folio_isolate_lru(folio)) {
 639			folio_unlock(folio);
 640			result = SCAN_DEL_PAGE_LRU;
 641			goto out;
 642		}
 643		node_stat_mod_folio(folio,
 644				NR_ISOLATED_ANON + folio_is_file_lru(folio),
 645				folio_nr_pages(folio));
 646		VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 647		VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
 648
 649		if (folio_test_large(folio))
 650			list_add_tail(&folio->lru, compound_pagelist);
 651next:
 652		/*
 653		 * If collapse was initiated by khugepaged, check that there is
 654		 * enough young pte to justify collapsing the page
 655		 */
 656		if (cc->is_khugepaged &&
 657		    (pte_young(pteval) || folio_test_young(folio) ||
 658		     folio_test_referenced(folio) ||
 659		     mmu_notifier_test_young(vma->vm_mm, addr)))
 660			referenced++;
 661	}
 662
 663	if (unlikely(cc->is_khugepaged && !referenced)) {
 664		result = SCAN_LACK_REFERENCED_PAGE;
 665	} else {
 666		result = SCAN_SUCCEED;
 667		trace_mm_collapse_huge_page_isolate(folio, none_or_zero,
 668						    referenced, result);
 669		return result;
 670	}
 671out:
 672	release_pte_pages(pte, _pte, compound_pagelist);
 673	trace_mm_collapse_huge_page_isolate(folio, none_or_zero,
 674					    referenced, result);
 675	return result;
 676}
 677
 678static void __collapse_huge_page_copy_succeeded(pte_t *pte,
 679						struct vm_area_struct *vma,
 680						unsigned long address,
 681						spinlock_t *ptl,
 682						struct list_head *compound_pagelist)
 683{
 684	unsigned long end = address + HPAGE_PMD_SIZE;
 685	struct folio *src, *tmp;
 686	pte_t pteval;
 687	pte_t *_pte;
 688	unsigned int nr_ptes;
 689
 690	for (_pte = pte; _pte < pte + HPAGE_PMD_NR; _pte += nr_ptes,
 691	     address += nr_ptes * PAGE_SIZE) {
 692		nr_ptes = 1;
 693		pteval = ptep_get(_pte);
 694		if (pte_none_or_zero(pteval)) {
 695			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 696			if (pte_none(pteval))
 697				continue;
 698			/*
 699			 * ptl mostly unnecessary.
 700			 */
 701			spin_lock(ptl);
 702			ptep_clear(vma->vm_mm, address, _pte);
 703			spin_unlock(ptl);
 704			ksm_might_unmap_zero_page(vma->vm_mm, pteval);
 705		} else {
 706			struct page *src_page = pte_page(pteval);
 707
 708			src = page_folio(src_page);
 709
 710			if (folio_test_large(src)) {
 711				unsigned int max_nr_ptes = (end - address) >> PAGE_SHIFT;
 712
 713				nr_ptes = folio_pte_batch(src, _pte, pteval, max_nr_ptes);
 714			} else {
 715				release_pte_folio(src);
 716			}
 717
 718			/*
 719			 * ptl mostly unnecessary, but preempt has to
 720			 * be disabled to update the per-cpu stats
 721			 * inside folio_remove_rmap_pte().
 722			 */
 723			spin_lock(ptl);
 724			clear_ptes(vma->vm_mm, address, _pte, nr_ptes);
 725			folio_remove_rmap_ptes(src, src_page, nr_ptes, vma);
 726			spin_unlock(ptl);
 727			free_swap_cache(src);
 728			folio_put_refs(src, nr_ptes);
 729		}
 730	}
 731
 732	list_for_each_entry_safe(src, tmp, compound_pagelist, lru) {
 733		list_del(&src->lru);
 734		node_stat_sub_folio(src, NR_ISOLATED_ANON +
 735				folio_is_file_lru(src));
 736		folio_unlock(src);
 737		free_swap_cache(src);
 738		folio_putback_lru(src);
 739	}
 740}
 741
 742static void __collapse_huge_page_copy_failed(pte_t *pte,
 743					     pmd_t *pmd,
 744					     pmd_t orig_pmd,
 745					     struct vm_area_struct *vma,
 746					     struct list_head *compound_pagelist)
 747{
 748	spinlock_t *pmd_ptl;
 749
 750	/*
 751	 * Re-establish the PMD to point to the original page table
 752	 * entry. Restoring PMD needs to be done prior to releasing
 753	 * pages. Since pages are still isolated and locked here,
 754	 * acquiring anon_vma_lock_write is unnecessary.
 755	 */
 756	pmd_ptl = pmd_lock(vma->vm_mm, pmd);
 757	pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd));
 758	spin_unlock(pmd_ptl);
 759	/*
 760	 * Release both raw and compound pages isolated
 761	 * in __collapse_huge_page_isolate.
 762	 */
 763	release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist);
 764}
 765
 766/*
 767 * __collapse_huge_page_copy - attempts to copy memory contents from raw
 768 * pages to a hugepage. Cleans up the raw pages if copying succeeds;
 769 * otherwise restores the original page table and releases isolated raw pages.
 770 * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC.
 771 *
 772 * @pte: starting of the PTEs to copy from
 773 * @folio: the new hugepage to copy contents to
 774 * @pmd: pointer to the new hugepage's PMD
 775 * @orig_pmd: the original raw pages' PMD
 776 * @vma: the original raw pages' virtual memory area
 777 * @address: starting address to copy
 778 * @ptl: lock on raw pages' PTEs
 779 * @compound_pagelist: list that stores compound pages
 780 */
 781static int __collapse_huge_page_copy(pte_t *pte, struct folio *folio,
 782		pmd_t *pmd, pmd_t orig_pmd, struct vm_area_struct *vma,
 783		unsigned long address, spinlock_t *ptl,
 784		struct list_head *compound_pagelist)
 785{
 786	unsigned int i;
 787	int result = SCAN_SUCCEED;
 788
 789	/*
 790	 * Copying pages' contents is subject to memory poison at any iteration.
 791	 */
 792	for (i = 0; i < HPAGE_PMD_NR; i++) {
 793		pte_t pteval = ptep_get(pte + i);
 794		struct page *page = folio_page(folio, i);
 795		unsigned long src_addr = address + i * PAGE_SIZE;
 796		struct page *src_page;
 797
 798		if (pte_none_or_zero(pteval)) {
 799			clear_user_highpage(page, src_addr);
 800			continue;
 801		}
 802		src_page = pte_page(pteval);
 803		if (copy_mc_user_highpage(page, src_page, src_addr, vma) > 0) {
 804			result = SCAN_COPY_MC;
 805			break;
 806		}
 807	}
 808
 809	if (likely(result == SCAN_SUCCEED))
 810		__collapse_huge_page_copy_succeeded(pte, vma, address, ptl,
 811						    compound_pagelist);
 812	else
 813		__collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma,
 814						 compound_pagelist);
 815
 816	return result;
 817}
 818
 819static void khugepaged_alloc_sleep(void)
 820{
 821	DEFINE_WAIT(wait);
 822
 823	add_wait_queue(&khugepaged_wait, &wait);
 824	__set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 825	schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 826	remove_wait_queue(&khugepaged_wait, &wait);
 827}
 828
 829struct collapse_control khugepaged_collapse_control = {
 830	.is_khugepaged = true,
 831};
 832
 833static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)
 834{
 835	int i;
 836
 837	/*
 838	 * If node_reclaim_mode is disabled, then no extra effort is made to
 839	 * allocate memory locally.
 840	 */
 841	if (!node_reclaim_enabled())
 842		return false;
 843
 844	/* If there is a count for this node already, it must be acceptable */
 845	if (cc->node_load[nid])
 846		return false;
 847
 848	for (i = 0; i < MAX_NUMNODES; i++) {
 849		if (!cc->node_load[i])
 850			continue;
 851		if (node_distance(nid, i) > node_reclaim_distance)
 852			return true;
 853	}
 854	return false;
 855}
 856
 857#define khugepaged_defrag()					\
 858	(transparent_hugepage_flags &				\
 859	 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))
 860
 861/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 862static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 863{
 864	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 865}
 866
 867#ifdef CONFIG_NUMA
 868static int hpage_collapse_find_target_node(struct collapse_control *cc)
 869{
 870	int nid, target_node = 0, max_value = 0;
 871
 872	/* find first node with max normal pages hit */
 873	for (nid = 0; nid < MAX_NUMNODES; nid++)
 874		if (cc->node_load[nid] > max_value) {
 875			max_value = cc->node_load[nid];
 876			target_node = nid;
 877		}
 878
 879	for_each_online_node(nid) {
 880		if (max_value == cc->node_load[nid])
 881			node_set(nid, cc->alloc_nmask);
 882	}
 883
 884	return target_node;
 885}
 886#else
 887static int hpage_collapse_find_target_node(struct collapse_control *cc)
 888{
 889	return 0;
 890}
 891#endif
 892
 893/*
 894 * If mmap_lock temporarily dropped, revalidate vma
 895 * before taking mmap_lock.
 896 * Returns enum scan_result value.
 897 */
 898
 899static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 900				   bool expect_anon,
 901				   struct vm_area_struct **vmap,
 902				   struct collapse_control *cc)
 903{
 904	struct vm_area_struct *vma;
 905	enum tva_type type = cc->is_khugepaged ? TVA_KHUGEPAGED :
 906				 TVA_FORCED_COLLAPSE;
 907
 908	if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
 909		return SCAN_ANY_PROCESS;
 910
 911	*vmap = vma = find_vma(mm, address);
 912	if (!vma)
 913		return SCAN_VMA_NULL;
 914
 915	if (!thp_vma_suitable_order(vma, address, PMD_ORDER))
 916		return SCAN_ADDRESS_RANGE;
 917	if (!thp_vma_allowable_order(vma, vma->vm_flags, type, PMD_ORDER))
 918		return SCAN_VMA_CHECK;
 919	/*
 920	 * Anon VMA expected, the address may be unmapped then
 921	 * remapped to file after khugepaged reaquired the mmap_lock.
 922	 *
 923	 * thp_vma_allowable_order may return true for qualified file
 924	 * vmas.
 925	 */
 926	if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))
 927		return SCAN_PAGE_ANON;
 928	return SCAN_SUCCEED;
 929}
 930
 931static inline int check_pmd_state(pmd_t *pmd)
 932{
 933	pmd_t pmde = pmdp_get_lockless(pmd);
 934
 935	if (pmd_none(pmde))
 936		return SCAN_NO_PTE_TABLE;
 937
 938	/*
 939	 * The folio may be under migration when khugepaged is trying to
 940	 * collapse it. Migration success or failure will eventually end
 941	 * up with a present PMD mapping a folio again.
 942	 */
 943	if (pmd_is_migration_entry(pmde))
 944		return SCAN_PMD_MAPPED;
 945	if (!pmd_present(pmde))
 946		return SCAN_NO_PTE_TABLE;
 947	if (pmd_trans_huge(pmde))
 948		return SCAN_PMD_MAPPED;
 949	if (pmd_bad(pmde))
 950		return SCAN_NO_PTE_TABLE;
 951	return SCAN_SUCCEED;
 952}
 953
 954static int find_pmd_or_thp_or_none(struct mm_struct *mm,
 955				   unsigned long address,
 956				   pmd_t **pmd)
 957{
 958	*pmd = mm_find_pmd(mm, address);
 959	if (!*pmd)
 960		return SCAN_NO_PTE_TABLE;
 961
 962	return check_pmd_state(*pmd);
 963}
 964
 965static int check_pmd_still_valid(struct mm_struct *mm,
 966				 unsigned long address,
 967				 pmd_t *pmd)
 968{
 969	pmd_t *new_pmd;
 970	int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);
 971
 972	if (result != SCAN_SUCCEED)
 973		return result;
 974	if (new_pmd != pmd)
 975		return SCAN_FAIL;
 976	return SCAN_SUCCEED;
 977}
 978
 979/*
 980 * Bring missing pages in from swap, to complete THP collapse.
 981 * Only done if hpage_collapse_scan_pmd believes it is worthwhile.
 982 *
 983 * Called and returns without pte mapped or spinlocks held.
 984 * Returns result: if not SCAN_SUCCEED, mmap_lock has been released.
 985 */
 986static int __collapse_huge_page_swapin(struct mm_struct *mm,
 987				       struct vm_area_struct *vma,
 988				       unsigned long start_addr, pmd_t *pmd,
 989				       int referenced)
 990{
 991	int swapped_in = 0;
 992	vm_fault_t ret = 0;
 993	unsigned long addr, end = start_addr + (HPAGE_PMD_NR * PAGE_SIZE);
 994	int result;
 995	pte_t *pte = NULL;
 996	spinlock_t *ptl;
 997
 998	for (addr = start_addr; addr < end; addr += PAGE_SIZE) {
 999		struct vm_fault vmf = {
1000			.vma = vma,
1001			.address = addr,
1002			.pgoff = linear_page_index(vma, addr),
1003			.flags = FAULT_FLAG_ALLOW_RETRY,
1004			.pmd = pmd,
1005		};
1006
1007		if (!pte++) {
1008			/*
1009			 * Here the ptl is only used to check pte_same() in
1010			 * do_swap_page(), so readonly version is enough.
1011			 */
1012			pte = pte_offset_map_ro_nolock(mm, pmd, addr, &ptl);
1013			if (!pte) {
1014				mmap_read_unlock(mm);
1015				result = SCAN_NO_PTE_TABLE;
1016				goto out;
1017			}
1018		}
1019
1020		vmf.orig_pte = ptep_get_lockless(pte);
1021		if (pte_none(vmf.orig_pte) ||
1022		    pte_present(vmf.orig_pte))
1023			continue;
1024
1025		vmf.pte = pte;
1026		vmf.ptl = ptl;
1027		ret = do_swap_page(&vmf);
1028		/* Which unmaps pte (after perhaps re-checking the entry) */
1029		pte = NULL;
1030
1031		/*
1032		 * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.
1033		 * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because
1034		 * we do not retry here and swap entry will remain in pagetable
1035		 * resulting in later failure.
1036		 */
1037		if (ret & VM_FAULT_RETRY) {
1038			/* Likely, but not guaranteed, that page lock failed */
1039			result = SCAN_PAGE_LOCK;
1040			goto out;
1041		}
1042		if (ret & VM_FAULT_ERROR) {
1043			mmap_read_unlock(mm);
1044			result = SCAN_FAIL;
1045			goto out;
1046		}
1047		swapped_in++;
1048	}
1049
1050	if (pte)
1051		pte_unmap(pte);
1052
1053	/* Drain LRU cache to remove extra pin on the swapped in pages */
1054	if (swapped_in)
1055		lru_add_drain();
1056
1057	result = SCAN_SUCCEED;
1058out:
1059	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result);
1060	return result;
1061}
1062
1063static int alloc_charge_folio(struct folio **foliop, struct mm_struct *mm,
1064			      struct collapse_control *cc)
1065{
1066	gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :
1067		     GFP_TRANSHUGE);
1068	int node = hpage_collapse_find_target_node(cc);
1069	struct folio *folio;
1070
1071	folio = __folio_alloc(gfp, HPAGE_PMD_ORDER, node, &cc->alloc_nmask);
1072	if (!folio) {
1073		*foliop = NULL;
1074		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
1075		return SCAN_ALLOC_HUGE_PAGE_FAIL;
1076	}
1077
1078	count_vm_event(THP_COLLAPSE_ALLOC);
1079	if (unlikely(mem_cgroup_charge(folio, mm, gfp))) {
1080		folio_put(folio);
1081		*foliop = NULL;
1082		return SCAN_CGROUP_CHARGE_FAIL;
1083	}
1084
1085	count_memcg_folio_events(folio, THP_COLLAPSE_ALLOC, 1);
1086
1087	*foliop = folio;
1088	return SCAN_SUCCEED;
1089}
1090
1091static int collapse_huge_page(struct mm_struct *mm, unsigned long address,
1092			      int referenced, int unmapped,
1093			      struct collapse_control *cc)
1094{
1095	LIST_HEAD(compound_pagelist);
1096	pmd_t *pmd, _pmd;
1097	pte_t *pte;
1098	pgtable_t pgtable;
1099	struct folio *folio;
1100	spinlock_t *pmd_ptl, *pte_ptl;
1101	int result = SCAN_FAIL;
1102	struct vm_area_struct *vma;
1103	struct mmu_notifier_range range;
1104
1105	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1106
1107	/*
1108	 * Before allocating the hugepage, release the mmap_lock read lock.
1109	 * The allocation can take potentially a long time if it involves
1110	 * sync compaction, and we do not need to hold the mmap_lock during
1111	 * that. We will recheck the vma after taking it again in write mode.
1112	 */
1113	mmap_read_unlock(mm);
1114
1115	result = alloc_charge_folio(&folio, mm, cc);
1116	if (result != SCAN_SUCCEED)
1117		goto out_nolock;
1118
1119	mmap_read_lock(mm);
1120	result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1121	if (result != SCAN_SUCCEED) {
1122		mmap_read_unlock(mm);
1123		goto out_nolock;
1124	}
1125
1126	result = find_pmd_or_thp_or_none(mm, address, &pmd);
1127	if (result != SCAN_SUCCEED) {
1128		mmap_read_unlock(mm);
1129		goto out_nolock;
1130	}
1131
1132	if (unmapped) {
1133		/*
1134		 * __collapse_huge_page_swapin will return with mmap_lock
1135		 * released when it fails. So we jump out_nolock directly in
1136		 * that case.  Continuing to collapse causes inconsistency.
1137		 */
1138		result = __collapse_huge_page_swapin(mm, vma, address, pmd,
1139						     referenced);
1140		if (result != SCAN_SUCCEED)
1141			goto out_nolock;
1142	}
1143
1144	mmap_read_unlock(mm);
1145	/*
1146	 * Prevent all access to pagetables with the exception of
1147	 * gup_fast later handled by the ptep_clear_flush and the VM
1148	 * handled by the anon_vma lock + PG_lock.
1149	 *
1150	 * UFFDIO_MOVE is prevented to race as well thanks to the
1151	 * mmap_lock.
1152	 */
1153	mmap_write_lock(mm);
1154	result = hugepage_vma_revalidate(mm, address, true, &vma, cc);
1155	if (result != SCAN_SUCCEED)
1156		goto out_up_write;
1157	/* check if the pmd is still valid */
1158	vma_start_write(vma);
1159	result = check_pmd_still_valid(mm, address, pmd);
1160	if (result != SCAN_SUCCEED)
1161		goto out_up_write;
1162
1163	anon_vma_lock_write(vma->anon_vma);
1164
1165	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
1166				address + HPAGE_PMD_SIZE);
1167	mmu_notifier_invalidate_range_start(&range);
1168
1169	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1170	/*
1171	 * This removes any huge TLB entry from the CPU so we won't allow
1172	 * huge and small TLB entries for the same virtual address to
1173	 * avoid the risk of CPU bugs in that area.
1174	 *
1175	 * Parallel GUP-fast is fine since GUP-fast will back off when
1176	 * it detects PMD is changed.
1177	 */
1178	_pmd = pmdp_collapse_flush(vma, address, pmd);
1179	spin_unlock(pmd_ptl);
1180	mmu_notifier_invalidate_range_end(&range);
1181	tlb_remove_table_sync_one();
1182
1183	pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl);
1184	if (pte) {
1185		result = __collapse_huge_page_isolate(vma, address, pte, cc,
1186						      &compound_pagelist);
1187		spin_unlock(pte_ptl);
1188	} else {
1189		result = SCAN_NO_PTE_TABLE;
1190	}
1191
1192	if (unlikely(result != SCAN_SUCCEED)) {
1193		if (pte)
1194			pte_unmap(pte);
1195		spin_lock(pmd_ptl);
1196		BUG_ON(!pmd_none(*pmd));
1197		/*
1198		 * We can only use set_pmd_at when establishing
1199		 * hugepmds and never for establishing regular pmds that
1200		 * points to regular pagetables. Use pmd_populate for that
1201		 */
1202		pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1203		spin_unlock(pmd_ptl);
1204		anon_vma_unlock_write(vma->anon_vma);
1205		goto out_up_write;
1206	}
1207
1208	/*
1209	 * All pages are isolated and locked so anon_vma rmap
1210	 * can't run anymore.
1211	 */
1212	anon_vma_unlock_write(vma->anon_vma);
1213
1214	result = __collapse_huge_page_copy(pte, folio, pmd, _pmd,
1215					   vma, address, pte_ptl,
1216					   &compound_pagelist);
1217	pte_unmap(pte);
1218	if (unlikely(result != SCAN_SUCCEED))
1219		goto out_up_write;
1220
1221	/*
1222	 * The smp_wmb() inside __folio_mark_uptodate() ensures the
1223	 * copy_huge_page writes become visible before the set_pmd_at()
1224	 * write.
1225	 */
1226	__folio_mark_uptodate(folio);
1227	pgtable = pmd_pgtable(_pmd);
1228
1229	spin_lock(pmd_ptl);
1230	BUG_ON(!pmd_none(*pmd));
1231	pgtable_trans_huge_deposit(mm, pmd, pgtable);
1232	map_anon_folio_pmd_nopf(folio, pmd, vma, address);
1233	spin_unlock(pmd_ptl);
1234
1235	folio = NULL;
1236
1237	result = SCAN_SUCCEED;
1238out_up_write:
1239	mmap_write_unlock(mm);
1240out_nolock:
1241	if (folio)
1242		folio_put(folio);
1243	trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);
1244	return result;
1245}
1246
1247static int hpage_collapse_scan_pmd(struct mm_struct *mm,
1248				   struct vm_area_struct *vma,
1249				   unsigned long start_addr, bool *mmap_locked,
1250				   struct collapse_control *cc)
1251{
1252	pmd_t *pmd;
1253	pte_t *pte, *_pte;
1254	int result = SCAN_FAIL, referenced = 0;
1255	int none_or_zero = 0, shared = 0;
1256	struct page *page = NULL;
1257	struct folio *folio = NULL;
1258	unsigned long addr;
1259	spinlock_t *ptl;
1260	int node = NUMA_NO_NODE, unmapped = 0;
1261
1262	VM_BUG_ON(start_addr & ~HPAGE_PMD_MASK);
1263
1264	result = find_pmd_or_thp_or_none(mm, start_addr, &pmd);
1265	if (result != SCAN_SUCCEED)
1266		goto out;
1267
1268	memset(cc->node_load, 0, sizeof(cc->node_load));
1269	nodes_clear(cc->alloc_nmask);
1270	pte = pte_offset_map_lock(mm, pmd, start_addr, &ptl);
1271	if (!pte) {
1272		result = SCAN_NO_PTE_TABLE;
1273		goto out;
1274	}
1275
1276	for (addr = start_addr, _pte = pte; _pte < pte + HPAGE_PMD_NR;
1277	     _pte++, addr += PAGE_SIZE) {
1278		pte_t pteval = ptep_get(_pte);
1279		if (pte_none_or_zero(pteval)) {
1280			++none_or_zero;
1281			if (!userfaultfd_armed(vma) &&
1282			    (!cc->is_khugepaged ||
1283			     none_or_zero <= khugepaged_max_ptes_none)) {
1284				continue;
1285			} else {
1286				result = SCAN_EXCEED_NONE_PTE;
1287				count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
1288				goto out_unmap;
1289			}
1290		}
1291		if (!pte_present(pteval)) {
1292			++unmapped;
1293			if (!cc->is_khugepaged ||
1294			    unmapped <= khugepaged_max_ptes_swap) {
1295				/*
1296				 * Always be strict with uffd-wp
1297				 * enabled swap entries.  Please see
1298				 * comment below for pte_uffd_wp().
1299				 */
1300				if (pte_swp_uffd_wp_any(pteval)) {
1301					result = SCAN_PTE_UFFD_WP;
1302					goto out_unmap;
1303				}
1304				continue;
1305			} else {
1306				result = SCAN_EXCEED_SWAP_PTE;
1307				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
1308				goto out_unmap;
1309			}
1310		}
1311		if (pte_uffd_wp(pteval)) {
1312			/*
1313			 * Don't collapse the page if any of the small
1314			 * PTEs are armed with uffd write protection.
1315			 * Here we can also mark the new huge pmd as
1316			 * write protected if any of the small ones is
1317			 * marked but that could bring unknown
1318			 * userfault messages that falls outside of
1319			 * the registered range.  So, just be simple.
1320			 */
1321			result = SCAN_PTE_UFFD_WP;
1322			goto out_unmap;
1323		}
1324
1325		page = vm_normal_page(vma, addr, pteval);
1326		if (unlikely(!page) || unlikely(is_zone_device_page(page))) {
1327			result = SCAN_PAGE_NULL;
1328			goto out_unmap;
1329		}
1330		folio = page_folio(page);
1331
1332		if (!folio_test_anon(folio)) {
1333			result = SCAN_PAGE_ANON;
1334			goto out_unmap;
1335		}
1336
1337		/*
1338		 * We treat a single page as shared if any part of the THP
1339		 * is shared.
1340		 */
1341		if (folio_maybe_mapped_shared(folio)) {
1342			++shared;
1343			if (cc->is_khugepaged &&
1344			    shared > khugepaged_max_ptes_shared) {
1345				result = SCAN_EXCEED_SHARED_PTE;
1346				count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);
1347				goto out_unmap;
1348			}
1349		}
1350
1351		/*
1352		 * Record which node the original page is from and save this
1353		 * information to cc->node_load[].
1354		 * Khugepaged will allocate hugepage from the node has the max
1355		 * hit record.
1356		 */
1357		node = folio_nid(folio);
1358		if (hpage_collapse_scan_abort(node, cc)) {
1359			result = SCAN_SCAN_ABORT;
1360			goto out_unmap;
1361		}
1362		cc->node_load[node]++;
1363		if (!folio_test_lru(folio)) {
1364			result = SCAN_PAGE_LRU;
1365			goto out_unmap;
1366		}
1367		if (folio_test_locked(folio)) {
1368			result = SCAN_PAGE_LOCK;
1369			goto out_unmap;
1370		}
1371
1372		/*
1373		 * Check if the page has any GUP (or other external) pins.
1374		 *
1375		 * Here the check may be racy:
1376		 * it may see folio_mapcount() > folio_ref_count().
1377		 * But such case is ephemeral we could always retry collapse
1378		 * later.  However it may report false positive if the page
1379		 * has excessive GUP pins (i.e. 512).  Anyway the same check
1380		 * will be done again later the risk seems low.
1381		 */
1382		if (folio_expected_ref_count(folio) != folio_ref_count(folio)) {
1383			result = SCAN_PAGE_COUNT;
1384			goto out_unmap;
1385		}
1386
1387		/*
1388		 * If collapse was initiated by khugepaged, check that there is
1389		 * enough young pte to justify collapsing the page
1390		 */
1391		if (cc->is_khugepaged &&
1392		    (pte_young(pteval) || folio_test_young(folio) ||
1393		     folio_test_referenced(folio) ||
1394		     mmu_notifier_test_young(vma->vm_mm, addr)))
1395			referenced++;
1396	}
1397	if (cc->is_khugepaged &&
1398		   (!referenced ||
1399		    (unmapped && referenced < HPAGE_PMD_NR / 2))) {
1400		result = SCAN_LACK_REFERENCED_PAGE;
1401	} else {
1402		result = SCAN_SUCCEED;
1403	}
1404out_unmap:
1405	pte_unmap_unlock(pte, ptl);
1406	if (result == SCAN_SUCCEED) {
1407		result = collapse_huge_page(mm, start_addr, referenced,
1408					    unmapped, cc);
1409		/* collapse_huge_page will return with the mmap_lock released */
1410		*mmap_locked = false;
1411	}
1412out:
1413	trace_mm_khugepaged_scan_pmd(mm, folio, referenced,
1414				     none_or_zero, result, unmapped);
1415	return result;
1416}
1417
1418static void collect_mm_slot(struct mm_slot *slot)
1419{
1420	struct mm_struct *mm = slot->mm;
1421
1422	lockdep_assert_held(&khugepaged_mm_lock);
1423
1424	if (hpage_collapse_test_exit(mm)) {
1425		/* free mm_slot */
1426		hash_del(&slot->hash);
1427		list_del(&slot->mm_node);
1428
1429		/*
1430		 * Not strictly needed because the mm exited already.
1431		 *
1432		 * mm_flags_clear(MMF_VM_HUGEPAGE, mm);
1433		 */
1434
1435		/* khugepaged_mm_lock actually not necessary for the below */
1436		mm_slot_free(mm_slot_cache, slot);
1437		mmdrop(mm);
1438	}
1439}
1440
1441/* folio must be locked, and mmap_lock must be held */
1442static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
1443			pmd_t *pmdp, struct folio *folio, struct page *page)
1444{
1445	struct mm_struct *mm = vma->vm_mm;
1446	struct vm_fault vmf = {
1447		.vma = vma,
1448		.address = addr,
1449		.flags = 0,
1450	};
1451	pgd_t *pgdp;
1452	p4d_t *p4dp;
1453	pud_t *pudp;
1454
1455	mmap_assert_locked(vma->vm_mm);
1456
1457	if (!pmdp) {
1458		pgdp = pgd_offset(mm, addr);
1459		p4dp = p4d_alloc(mm, pgdp, addr);
1460		if (!p4dp)
1461			return SCAN_FAIL;
1462		pudp = pud_alloc(mm, p4dp, addr);
1463		if (!pudp)
1464			return SCAN_FAIL;
1465		pmdp = pmd_alloc(mm, pudp, addr);
1466		if (!pmdp)
1467			return SCAN_FAIL;
1468	}
1469
1470	vmf.pmd = pmdp;
1471	if (do_set_pmd(&vmf, folio, page))
1472		return SCAN_FAIL;
1473
1474	folio_get(folio);
1475	return SCAN_SUCCEED;
1476}
1477
1478/**
1479 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1480 * address haddr.
1481 *
1482 * @mm: process address space where collapse happens
1483 * @addr: THP collapse address
1484 * @install_pmd: If a huge PMD should be installed
1485 *
1486 * This function checks whether all the PTEs in the PMD are pointing to the
1487 * right THP. If so, retract the page table so the THP can refault in with
1488 * as pmd-mapped. Possibly install a huge PMD mapping the THP.
1489 */
1490int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
1491			    bool install_pmd)
1492{
1493	int nr_mapped_ptes = 0, result = SCAN_FAIL;
1494	unsigned int nr_batch_ptes;
1495	struct mmu_notifier_range range;
1496	bool notified = false;
1497	unsigned long haddr = addr & HPAGE_PMD_MASK;
1498	unsigned long end = haddr + HPAGE_PMD_SIZE;
1499	struct vm_area_struct *vma = vma_lookup(mm, haddr);
1500	struct folio *folio;
1501	pte_t *start_pte, *pte;
1502	pmd_t *pmd, pgt_pmd;
1503	spinlock_t *pml = NULL, *ptl;
1504	int i;
1505
1506	mmap_assert_locked(mm);
1507
1508	/* First check VMA found, in case page tables are being torn down */
1509	if (!vma || !vma->vm_file ||
1510	    !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
1511		return SCAN_VMA_CHECK;
1512
1513	/* Fast check before locking page if already PMD-mapped */
1514	result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1515	if (result == SCAN_PMD_MAPPED)
1516		return result;
1517
1518	/*
1519	 * If we are here, we've succeeded in replacing all the native pages
1520	 * in the page cache with a single hugepage. If a mm were to fault-in
1521	 * this memory (mapped by a suitably aligned VMA), we'd get the hugepage
1522	 * and map it by a PMD, regardless of sysfs THP settings. As such, let's
1523	 * analogously elide sysfs THP settings here and force collapse.
1524	 */
1525	if (!thp_vma_allowable_order(vma, vma->vm_flags, TVA_FORCED_COLLAPSE, PMD_ORDER))
1526		return SCAN_VMA_CHECK;
1527
1528	/* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */
1529	if (userfaultfd_wp(vma))
1530		return SCAN_PTE_UFFD_WP;
1531
1532	folio = filemap_lock_folio(vma->vm_file->f_mapping,
1533			       linear_page_index(vma, haddr));
1534	if (IS_ERR(folio))
1535		return SCAN_PAGE_NULL;
1536
1537	if (folio_order(folio) != HPAGE_PMD_ORDER) {
1538		result = SCAN_PAGE_COMPOUND;
1539		goto drop_folio;
1540	}
1541
1542	result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
1543	switch (result) {
1544	case SCAN_SUCCEED:
1545		break;
1546	case SCAN_NO_PTE_TABLE:
1547		/*
1548		 * All pte entries have been removed and pmd cleared.
1549		 * Skip all the pte checks and just update the pmd mapping.
1550		 */
1551		goto maybe_install_pmd;
1552	default:
1553		goto drop_folio;
1554	}
1555
1556	result = SCAN_FAIL;
1557	start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1558	if (!start_pte)		/* mmap_lock + page lock should prevent this */
1559		goto drop_folio;
1560
1561	/* step 1: check all mapped PTEs are to the right huge page */
1562	for (i = 0, addr = haddr, pte = start_pte;
1563	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1564		struct page *page;
1565		pte_t ptent = ptep_get(pte);
1566
1567		/* empty pte, skip */
1568		if (pte_none(ptent))
1569			continue;
1570
1571		/* page swapped out, abort */
1572		if (!pte_present(ptent)) {
1573			result = SCAN_PTE_NON_PRESENT;
1574			goto abort;
1575		}
1576
1577		page = vm_normal_page(vma, addr, ptent);
1578		if (WARN_ON_ONCE(page && is_zone_device_page(page)))
1579			page = NULL;
1580		/*
1581		 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1582		 * page table, but the new page will not be a subpage of hpage.
1583		 */
1584		if (folio_page(folio, i) != page)
1585			goto abort;
1586	}
1587
1588	pte_unmap_unlock(start_pte, ptl);
1589	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1590				haddr, haddr + HPAGE_PMD_SIZE);
1591	mmu_notifier_invalidate_range_start(&range);
1592	notified = true;
1593
1594	/*
1595	 * pmd_lock covers a wider range than ptl, and (if split from mm's
1596	 * page_table_lock) ptl nests inside pml. The less time we hold pml,
1597	 * the better; but userfaultfd's mfill_atomic_pte() on a private VMA
1598	 * inserts a valid as-if-COWed PTE without even looking up page cache.
1599	 * So page lock of folio does not protect from it, so we must not drop
1600	 * ptl before pgt_pmd is removed, so uffd private needs pml taken now.
1601	 */
1602	if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED))
1603		pml = pmd_lock(mm, pmd);
1604
1605	start_pte = pte_offset_map_rw_nolock(mm, pmd, haddr, &pgt_pmd, &ptl);
1606	if (!start_pte)		/* mmap_lock + page lock should prevent this */
1607		goto abort;
1608	if (!pml)
1609		spin_lock(ptl);
1610	else if (ptl != pml)
1611		spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1612
1613	if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
1614		goto abort;
1615
1616	/* step 2: clear page table and adjust rmap */
1617	for (i = 0, addr = haddr, pte = start_pte; i < HPAGE_PMD_NR;
1618	     i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
1619	     pte += nr_batch_ptes) {
1620		unsigned int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
1621		struct page *page;
1622		pte_t ptent = ptep_get(pte);
1623
1624		nr_batch_ptes = 1;
1625
1626		if (pte_none(ptent))
1627			continue;
1628		/*
1629		 * We dropped ptl after the first scan, to do the mmu_notifier:
1630		 * page lock stops more PTEs of the folio being faulted in, but
1631		 * does not stop write faults COWing anon copies from existing
1632		 * PTEs; and does not stop those being swapped out or migrated.
1633		 */
1634		if (!pte_present(ptent)) {
1635			result = SCAN_PTE_NON_PRESENT;
1636			goto abort;
1637		}
1638		page = vm_normal_page(vma, addr, ptent);
1639
1640		if (folio_page(folio, i) != page)
1641			goto abort;
1642
1643		nr_batch_ptes = folio_pte_batch(folio, pte, ptent, max_nr_batch_ptes);
1644
1645		/*
1646		 * Must clear entry, or a racing truncate may re-remove it.
1647		 * TLB flush can be left until pmdp_collapse_flush() does it.
1648		 * PTE dirty? Shmem page is already dirty; file is read-only.
1649		 */
1650		clear_ptes(mm, addr, pte, nr_batch_ptes);
1651		folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
1652		nr_mapped_ptes += nr_batch_ptes;
1653	}
1654
1655	if (!pml)
1656		spin_unlock(ptl);
1657
1658	/* step 3: set proper refcount and mm_counters. */
1659	if (nr_mapped_ptes) {
1660		folio_ref_sub(folio, nr_mapped_ptes);
1661		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
1662	}
1663
1664	/* step 4: remove empty page table */
1665	if (!pml) {
1666		pml = pmd_lock(mm, pmd);
1667		if (ptl != pml) {
1668			spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1669			if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd)))) {
1670				flush_tlb_mm(mm);
1671				goto unlock;
1672			}
1673		}
1674	}
1675	pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd);
1676	pmdp_get_lockless_sync();
1677	pte_unmap_unlock(start_pte, ptl);
1678	if (ptl != pml)
1679		spin_unlock(pml);
1680
1681	mmu_notifier_invalidate_range_end(&range);
1682
1683	mm_dec_nr_ptes(mm);
1684	page_table_check_pte_clear_range(mm, haddr, pgt_pmd);
1685	pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1686
1687maybe_install_pmd:
1688	/* step 5: install pmd entry */
1689	result = install_pmd
1690			? set_huge_pmd(vma, haddr, pmd, folio, &folio->page)
1691			: SCAN_SUCCEED;
1692	goto drop_folio;
1693abort:
1694	if (nr_mapped_ptes) {
1695		flush_tlb_mm(mm);
1696		folio_ref_sub(folio, nr_mapped_ptes);
1697		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
1698	}
1699unlock:
1700	if (start_pte)
1701		pte_unmap_unlock(start_pte, ptl);
1702	if (pml && pml != ptl)
1703		spin_unlock(pml);
1704	if (notified)
1705		mmu_notifier_invalidate_range_end(&range);
1706drop_folio:
1707	folio_unlock(folio);
1708	folio_put(folio);
1709	return result;
1710}
1711
1712/* Can we retract page tables for this file-backed VMA? */
1713static bool file_backed_vma_is_retractable(struct vm_area_struct *vma)
1714{
1715	/*
1716	 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1717	 * got written to. These VMAs are likely not worth removing
1718	 * page tables from, as PMD-mapping is likely to be split later.
1719	 */
1720	if (READ_ONCE(vma->anon_vma))
1721		return false;
1722
1723	/*
1724	 * When a vma is registered with uffd-wp, we cannot recycle
1725	 * the page table because there may be pte markers installed.
1726	 * Other vmas can still have the same file mapped hugely, but
1727	 * skip this one: it will always be mapped in small page size
1728	 * for uffd-wp registered ranges.
1729	 */
1730	if (userfaultfd_wp(vma))
1731		return false;
1732
1733	/*
1734	 * If the VMA contains guard regions then we can't collapse it.
1735	 *
1736	 * This is set atomically on guard marker installation under mmap/VMA
1737	 * read lock, and here we may not hold any VMA or mmap lock at all.
1738	 *
1739	 * This is therefore serialised on the PTE page table lock, which is
1740	 * obtained on guard region installation after the flag is set, so this
1741	 * check being performed under this lock excludes races.
1742	 */
1743	if (vma_flag_test_atomic(vma, VMA_MAYBE_GUARD_BIT))
1744		return false;
1745
1746	return true;
1747}
1748
1749static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1750{
1751	struct vm_area_struct *vma;
1752
1753	i_mmap_lock_read(mapping);
1754	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1755		struct mmu_notifier_range range;
1756		struct mm_struct *mm;
1757		unsigned long addr;
1758		pmd_t *pmd, pgt_pmd;
1759		spinlock_t *pml;
1760		spinlock_t *ptl;
1761		bool success = false;
1762
1763		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1764		if (addr & ~HPAGE_PMD_MASK ||
1765		    vma->vm_end < addr + HPAGE_PMD_SIZE)
1766			continue;
1767
1768		mm = vma->vm_mm;
1769		if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED)
1770			continue;
1771
1772		if (hpage_collapse_test_exit(mm))
1773			continue;
1774
1775		if (!file_backed_vma_is_retractable(vma))
1776			continue;
1777
1778		/* PTEs were notified when unmapped; but now for the PMD? */
1779		mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
1780					addr, addr + HPAGE_PMD_SIZE);
1781		mmu_notifier_invalidate_range_start(&range);
1782
1783		pml = pmd_lock(mm, pmd);
1784		/*
1785		 * The lock of new_folio is still held, we will be blocked in
1786		 * the page fault path, which prevents the pte entries from
1787		 * being set again. So even though the old empty PTE page may be
1788		 * concurrently freed and a new PTE page is filled into the pmd
1789		 * entry, it is still empty and can be removed.
1790		 *
1791		 * So here we only need to recheck if the state of pmd entry
1792		 * still meets our requirements, rather than checking pmd_same()
1793		 * like elsewhere.
1794		 */
1795		if (check_pmd_state(pmd) != SCAN_SUCCEED)
1796			goto drop_pml;
1797		ptl = pte_lockptr(mm, pmd);
1798		if (ptl != pml)
1799			spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
1800
1801		/*
1802		 * Huge page lock is still held, so normally the page table must
1803		 * remain empty; and we have already skipped anon_vma and
1804		 * userfaultfd_wp() vmas.  But since the mmap_lock is not held,
1805		 * it is still possible for a racing userfaultfd_ioctl() or
1806		 * madvise() to have inserted ptes or markers.  Now that we hold
1807		 * ptlock, repeating the retractable checks protects us from
1808		 * races against the prior checks.
1809		 */
1810		if (likely(file_backed_vma_is_retractable(vma))) {
1811			pgt_pmd = pmdp_collapse_flush(vma, addr, pmd);
1812			pmdp_get_lockless_sync();
1813			success = true;
1814		}
1815
1816		if (ptl != pml)
1817			spin_unlock(ptl);
1818drop_pml:
1819		spin_unlock(pml);
1820
1821		mmu_notifier_invalidate_range_end(&range);
1822
1823		if (success) {
1824			mm_dec_nr_ptes(mm);
1825			page_table_check_pte_clear_range(mm, addr, pgt_pmd);
1826			pte_free_defer(mm, pmd_pgtable(pgt_pmd));
1827		}
1828	}
1829	i_mmap_unlock_read(mapping);
1830}
1831
1832/**
1833 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1834 *
1835 * @mm: process address space where collapse happens
1836 * @addr: virtual collapse start address
1837 * @file: file that collapse on
1838 * @start: collapse start address
1839 * @cc: collapse context and scratchpad
1840 *
1841 * Basic scheme is simple, details are more complex:
1842 *  - allocate and lock a new huge page;
1843 *  - scan page cache, locking old pages
1844 *    + swap/gup in pages if necessary;
1845 *  - copy data to new page
1846 *  - handle shmem holes
1847 *    + re-validate that holes weren't filled by someone else
1848 *    + check for userfaultfd
1849 *  - finalize updates to the page cache;
1850 *  - if replacing succeeds:
1851 *    + unlock huge page;
1852 *    + free old pages;
1853 *  - if replacing failed;
1854 *    + unlock old pages
1855 *    + unlock and free huge page;
1856 */
1857static int collapse_file(struct mm_struct *mm, unsigned long addr,
1858			 struct file *file, pgoff_t start,
1859			 struct collapse_control *cc)
1860{
1861	struct address_space *mapping = file->f_mapping;
1862	struct page *dst;
1863	struct folio *folio, *tmp, *new_folio;
1864	pgoff_t index = 0, end = start + HPAGE_PMD_NR;
1865	LIST_HEAD(pagelist);
1866	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1867	int nr_none = 0, result = SCAN_SUCCEED;
1868	bool is_shmem = shmem_file(file);
1869
1870	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1871	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1872
1873	result = alloc_charge_folio(&new_folio, mm, cc);
1874	if (result != SCAN_SUCCEED)
1875		goto out;
1876
1877	mapping_set_update(&xas, mapping);
1878
1879	__folio_set_locked(new_folio);
1880	if (is_shmem)
1881		__folio_set_swapbacked(new_folio);
1882	new_folio->index = start;
1883	new_folio->mapping = mapping;
1884
1885	/*
1886	 * Ensure we have slots for all the pages in the range.  This is
1887	 * almost certainly a no-op because most of the pages must be present
1888	 */
1889	do {
1890		xas_lock_irq(&xas);
1891		xas_create_range(&xas);
1892		if (!xas_error(&xas))
1893			break;
1894		xas_unlock_irq(&xas);
1895		if (!xas_nomem(&xas, GFP_KERNEL)) {
1896			result = SCAN_FAIL;
1897			goto rollback;
1898		}
1899	} while (1);
1900
1901	for (index = start; index < end;) {
1902		xas_set(&xas, index);
1903		folio = xas_load(&xas);
1904
1905		VM_BUG_ON(index != xas.xa_index);
1906		if (is_shmem) {
1907			if (!folio) {
1908				/*
1909				 * Stop if extent has been truncated or
1910				 * hole-punched, and is now completely
1911				 * empty.
1912				 */
1913				if (index == start) {
1914					if (!xas_next_entry(&xas, end - 1)) {
1915						result = SCAN_TRUNCATED;
1916						goto xa_locked;
1917					}
1918				}
1919				nr_none++;
1920				index++;
1921				continue;
1922			}
1923
1924			if (xa_is_value(folio) || !folio_test_uptodate(folio)) {
1925				xas_unlock_irq(&xas);
1926				/* swap in or instantiate fallocated page */
1927				if (shmem_get_folio(mapping->host, index, 0,
1928						&folio, SGP_NOALLOC)) {
1929					result = SCAN_FAIL;
1930					goto xa_unlocked;
1931				}
1932				/* drain lru cache to help folio_isolate_lru() */
1933				lru_add_drain();
1934			} else if (folio_trylock(folio)) {
1935				folio_get(folio);
1936				xas_unlock_irq(&xas);
1937			} else {
1938				result = SCAN_PAGE_LOCK;
1939				goto xa_locked;
1940			}
1941		} else {	/* !is_shmem */
1942			if (!folio || xa_is_value(folio)) {
1943				xas_unlock_irq(&xas);
1944				page_cache_sync_readahead(mapping, &file->f_ra,
1945							  file, index,
1946							  end - index);
1947				/* drain lru cache to help folio_isolate_lru() */
1948				lru_add_drain();
1949				folio = filemap_lock_folio(mapping, index);
1950				if (IS_ERR(folio)) {
1951					result = SCAN_FAIL;
1952					goto xa_unlocked;
1953				}
1954			} else if (folio_test_dirty(folio)) {
1955				/*
1956				 * khugepaged only works on read-only fd,
1957				 * so this page is dirty because it hasn't
1958				 * been flushed since first write. There
1959				 * won't be new dirty pages.
1960				 *
1961				 * Trigger async flush here and hope the
1962				 * writeback is done when khugepaged
1963				 * revisits this page.
1964				 *
1965				 * This is a one-off situation. We are not
1966				 * forcing writeback in loop.
1967				 */
1968				xas_unlock_irq(&xas);
1969				filemap_flush(mapping);
1970				result = SCAN_FAIL;
1971				goto xa_unlocked;
1972			} else if (folio_test_writeback(folio)) {
1973				xas_unlock_irq(&xas);
1974				result = SCAN_FAIL;
1975				goto xa_unlocked;
1976			} else if (folio_trylock(folio)) {
1977				folio_get(folio);
1978				xas_unlock_irq(&xas);
1979			} else {
1980				result = SCAN_PAGE_LOCK;
1981				goto xa_locked;
1982			}
1983		}
1984
1985		/*
1986		 * The folio must be locked, so we can drop the i_pages lock
1987		 * without racing with truncate.
1988		 */
1989		VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1990
1991		/* make sure the folio is up to date */
1992		if (unlikely(!folio_test_uptodate(folio))) {
1993			result = SCAN_FAIL;
1994			goto out_unlock;
1995		}
1996
1997		/*
1998		 * If file was truncated then extended, or hole-punched, before
1999		 * we locked the first folio, then a THP might be there already.
2000		 * This will be discovered on the first iteration.
2001		 */
2002		if (folio_order(folio) == HPAGE_PMD_ORDER &&
2003		    folio->index == start) {
2004			/* Maybe PMD-mapped */
2005			result = SCAN_PTE_MAPPED_HUGEPAGE;
2006			goto out_unlock;
2007		}
2008
2009		if (folio_mapping(folio) != mapping) {
2010			result = SCAN_TRUNCATED;
2011			goto out_unlock;
2012		}
2013
2014		if (!is_shmem && (folio_test_dirty(folio) ||
2015				  folio_test_writeback(folio))) {
2016			/*
2017			 * khugepaged only works on read-only fd, so this
2018			 * folio is dirty because it hasn't been flushed
2019			 * since first write.
2020			 */
2021			result = SCAN_FAIL;
2022			goto out_unlock;
2023		}
2024
2025		if (!folio_isolate_lru(folio)) {
2026			result = SCAN_DEL_PAGE_LRU;
2027			goto out_unlock;
2028		}
2029
2030		if (!filemap_release_folio(folio, GFP_KERNEL)) {
2031			result = SCAN_PAGE_HAS_PRIVATE;
2032			folio_putback_lru(folio);
2033			goto out_unlock;
2034		}
2035
2036		if (folio_mapped(folio))
2037			try_to_unmap(folio,
2038					TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);
2039
2040		xas_lock_irq(&xas);
2041
2042		VM_BUG_ON_FOLIO(folio != xa_load(xas.xa, index), folio);
2043
2044		/*
2045		 * We control 2 + nr_pages references to the folio:
2046		 *  - we hold a pin on it;
2047		 *  - nr_pages reference from page cache;
2048		 *  - one from lru_isolate_folio;
2049		 * If those are the only references, then any new usage
2050		 * of the folio will have to fetch it from the page
2051		 * cache. That requires locking the folio to handle
2052		 * truncate, so any new usage will be blocked until we
2053		 * unlock folio after collapse/during rollback.
2054		 */
2055		if (folio_ref_count(folio) != 2 + folio_nr_pages(folio)) {
2056			result = SCAN_PAGE_COUNT;
2057			xas_unlock_irq(&xas);
2058			folio_putback_lru(folio);
2059			goto out_unlock;
2060		}
2061
2062		/*
2063		 * Accumulate the folios that are being collapsed.
2064		 */
2065		list_add_tail(&folio->lru, &pagelist);
2066		index += folio_nr_pages(folio);
2067		continue;
2068out_unlock:
2069		folio_unlock(folio);
2070		folio_put(folio);
2071		goto xa_unlocked;
2072	}
2073
2074	if (!is_shmem) {
2075		filemap_nr_thps_inc(mapping);
2076		/*
2077		 * Paired with the fence in do_dentry_open() -> get_write_access()
2078		 * to ensure i_writecount is up to date and the update to nr_thps
2079		 * is visible. Ensures the page cache will be truncated if the
2080		 * file is opened writable.
2081		 */
2082		smp_mb();
2083		if (inode_is_open_for_write(mapping->host)) {
2084			result = SCAN_FAIL;
2085			filemap_nr_thps_dec(mapping);
2086		}
2087	}
2088
2089xa_locked:
2090	xas_unlock_irq(&xas);
2091xa_unlocked:
2092
2093	/*
2094	 * If collapse is successful, flush must be done now before copying.
2095	 * If collapse is unsuccessful, does flush actually need to be done?
2096	 * Do it anyway, to clear the state.
2097	 */
2098	try_to_unmap_flush();
2099
2100	if (result == SCAN_SUCCEED && nr_none &&
2101	    !shmem_charge(mapping->host, nr_none))
2102		result = SCAN_FAIL;
2103	if (result != SCAN_SUCCEED) {
2104		nr_none = 0;
2105		goto rollback;
2106	}
2107
2108	/*
2109	 * The old folios are locked, so they won't change anymore.
2110	 */
2111	index = start;
2112	dst = folio_page(new_folio, 0);
2113	list_for_each_entry(folio, &pagelist, lru) {
2114		int i, nr_pages = folio_nr_pages(folio);
2115
2116		while (index < folio->index) {
2117			clear_highpage(dst);
2118			index++;
2119			dst++;
2120		}
2121
2122		for (i = 0; i < nr_pages; i++) {
2123			if (copy_mc_highpage(dst, folio_page(folio, i)) > 0) {
2124				result = SCAN_COPY_MC;
2125				goto rollback;
2126			}
2127			index++;
2128			dst++;
2129		}
2130	}
2131	while (index < end) {
2132		clear_highpage(dst);
2133		index++;
2134		dst++;
2135	}
2136
2137	if (nr_none) {
2138		struct vm_area_struct *vma;
2139		int nr_none_check = 0;
2140
2141		i_mmap_lock_read(mapping);
2142		xas_lock_irq(&xas);
2143
2144		xas_set(&xas, start);
2145		for (index = start; index < end; index++) {
2146			if (!xas_next(&xas)) {
2147				xas_store(&xas, XA_RETRY_ENTRY);
2148				if (xas_error(&xas)) {
2149					result = SCAN_STORE_FAILED;
2150					goto immap_locked;
2151				}
2152				nr_none_check++;
2153			}
2154		}
2155
2156		if (nr_none != nr_none_check) {
2157			result = SCAN_PAGE_FILLED;
2158			goto immap_locked;
2159		}
2160
2161		/*
2162		 * If userspace observed a missing page in a VMA with
2163		 * a MODE_MISSING userfaultfd, then it might expect a
2164		 * UFFD_EVENT_PAGEFAULT for that page. If so, we need to
2165		 * roll back to avoid suppressing such an event. Since
2166		 * wp/minor userfaultfds don't give userspace any
2167		 * guarantees that the kernel doesn't fill a missing
2168		 * page with a zero page, so they don't matter here.
2169		 *
2170		 * Any userfaultfds registered after this point will
2171		 * not be able to observe any missing pages due to the
2172		 * previously inserted retry entries.
2173		 */
2174		vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) {
2175			if (userfaultfd_missing(vma)) {
2176				result = SCAN_EXCEED_NONE_PTE;
2177				goto immap_locked;
2178			}
2179		}
2180
2181immap_locked:
2182		i_mmap_unlock_read(mapping);
2183		if (result != SCAN_SUCCEED) {
2184			xas_set(&xas, start);
2185			for (index = start; index < end; index++) {
2186				if (xas_next(&xas) == XA_RETRY_ENTRY)
2187					xas_store(&xas, NULL);
2188			}
2189
2190			xas_unlock_irq(&xas);
2191			goto rollback;
2192		}
2193	} else {
2194		xas_lock_irq(&xas);
2195	}
2196
2197	if (is_shmem)
2198		lruvec_stat_mod_folio(new_folio, NR_SHMEM_THPS, HPAGE_PMD_NR);
2199	else
2200		lruvec_stat_mod_folio(new_folio, NR_FILE_THPS, HPAGE_PMD_NR);
2201
2202	if (nr_none) {
2203		lruvec_stat_mod_folio(new_folio, NR_FILE_PAGES, nr_none);
2204		/* nr_none is always 0 for non-shmem. */
2205		lruvec_stat_mod_folio(new_folio, NR_SHMEM, nr_none);
2206	}
2207
2208	/*
2209	 * Mark new_folio as uptodate before inserting it into the
2210	 * page cache so that it isn't mistaken for an fallocated but
2211	 * unwritten page.
2212	 */
2213	folio_mark_uptodate(new_folio);
2214	folio_ref_add(new_folio, HPAGE_PMD_NR - 1);
2215
2216	if (is_shmem)
2217		folio_mark_dirty(new_folio);
2218	folio_add_lru(new_folio);
2219
2220	/* Join all the small entries into a single multi-index entry. */
2221	xas_set_order(&xas, start, HPAGE_PMD_ORDER);
2222	xas_store(&xas, new_folio);
2223	WARN_ON_ONCE(xas_error(&xas));
2224	xas_unlock_irq(&xas);
2225
2226	/*
2227	 * Remove pte page tables, so we can re-fault the page as huge.
2228	 * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp().
2229	 */
2230	retract_page_tables(mapping, start);
2231	if (cc && !cc->is_khugepaged)
2232		result = SCAN_PTE_MAPPED_HUGEPAGE;
2233	folio_unlock(new_folio);
2234
2235	/*
2236	 * The collapse has succeeded, so free the old folios.
2237	 */
2238	list_for_each_entry_safe(folio, tmp, &pagelist, lru) {
2239		list_del(&folio->lru);
2240		folio->mapping = NULL;
2241		folio_clear_active(folio);
2242		folio_clear_unevictable(folio);
2243		folio_unlock(folio);
2244		folio_put_refs(folio, 2 + folio_nr_pages(folio));
2245	}
2246
2247	goto out;
2248
2249rollback:
2250	/* Something went wrong: roll back page cache changes */
2251	if (nr_none) {
2252		xas_lock_irq(&xas);
2253		mapping->nrpages -= nr_none;
2254		xas_unlock_irq(&xas);
2255		shmem_uncharge(mapping->host, nr_none);
2256	}
2257
2258	list_for_each_entry_safe(folio, tmp, &pagelist, lru) {
2259		list_del(&folio->lru);
2260		folio_unlock(folio);
2261		folio_putback_lru(folio);
2262		folio_put(folio);
2263	}
2264	/*
2265	 * Undo the updates of filemap_nr_thps_inc for non-SHMEM
2266	 * file only. This undo is not needed unless failure is
2267	 * due to SCAN_COPY_MC.
2268	 */
2269	if (!is_shmem && result == SCAN_COPY_MC) {
2270		filemap_nr_thps_dec(mapping);
2271		/*
2272		 * Paired with the fence in do_dentry_open() -> get_write_access()
2273		 * to ensure the update to nr_thps is visible.
2274		 */
2275		smp_mb();
2276	}
2277
2278	new_folio->mapping = NULL;
2279
2280	folio_unlock(new_folio);
2281	folio_put(new_folio);
2282out:
2283	VM_BUG_ON(!list_empty(&pagelist));
2284	trace_mm_khugepaged_collapse_file(mm, new_folio, index, addr, is_shmem, file, HPAGE_PMD_NR, result);
2285	return result;
2286}
2287
2288static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,
2289				    struct file *file, pgoff_t start,
2290				    struct collapse_control *cc)
2291{
2292	struct folio *folio = NULL;
2293	struct address_space *mapping = file->f_mapping;
2294	XA_STATE(xas, &mapping->i_pages, start);
2295	int present, swap;
2296	int node = NUMA_NO_NODE;
2297	int result = SCAN_SUCCEED;
2298
2299	present = 0;
2300	swap = 0;
2301	memset(cc->node_load, 0, sizeof(cc->node_load));
2302	nodes_clear(cc->alloc_nmask);
2303	rcu_read_lock();
2304	xas_for_each(&xas, folio, start + HPAGE_PMD_NR - 1) {
2305		if (xas_retry(&xas, folio))
2306			continue;
2307
2308		if (xa_is_value(folio)) {
2309			swap += 1 << xas_get_order(&xas);
2310			if (cc->is_khugepaged &&
2311			    swap > khugepaged_max_ptes_swap) {
2312				result = SCAN_EXCEED_SWAP_PTE;
2313				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
2314				break;
2315			}
2316			continue;
2317		}
2318
2319		if (!folio_try_get(folio)) {
2320			xas_reset(&xas);
2321			continue;
2322		}
2323
2324		if (unlikely(folio != xas_reload(&xas))) {
2325			folio_put(folio);
2326			xas_reset(&xas);
2327			continue;
2328		}
2329
2330		if (folio_order(folio) == HPAGE_PMD_ORDER &&
2331		    folio->index == start) {
2332			/* Maybe PMD-mapped */
2333			result = SCAN_PTE_MAPPED_HUGEPAGE;
2334			/*
2335			 * For SCAN_PTE_MAPPED_HUGEPAGE, further processing
2336			 * by the caller won't touch the page cache, and so
2337			 * it's safe to skip LRU and refcount checks before
2338			 * returning.
2339			 */
2340			folio_put(folio);
2341			break;
2342		}
2343
2344		node = folio_nid(folio);
2345		if (hpage_collapse_scan_abort(node, cc)) {
2346			result = SCAN_SCAN_ABORT;
2347			folio_put(folio);
2348			break;
2349		}
2350		cc->node_load[node]++;
2351
2352		if (!folio_test_lru(folio)) {
2353			result = SCAN_PAGE_LRU;
2354			folio_put(folio);
2355			break;
2356		}
2357
2358		if (folio_expected_ref_count(folio) + 1 != folio_ref_count(folio)) {
2359			result = SCAN_PAGE_COUNT;
2360			folio_put(folio);
2361			break;
2362		}
2363
2364		/*
2365		 * We probably should check if the folio is referenced
2366		 * here, but nobody would transfer pte_young() to
2367		 * folio_test_referenced() for us.  And rmap walk here
2368		 * is just too costly...
2369		 */
2370
2371		present += folio_nr_pages(folio);
2372		folio_put(folio);
2373
2374		if (need_resched()) {
2375			xas_pause(&xas);
2376			cond_resched_rcu();
2377		}
2378	}
2379	rcu_read_unlock();
2380
2381	if (result == SCAN_SUCCEED) {
2382		if (cc->is_khugepaged &&
2383		    present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
2384			result = SCAN_EXCEED_NONE_PTE;
2385			count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
2386		} else {
2387			result = collapse_file(mm, addr, file, start, cc);
2388		}
2389	}
2390
2391	trace_mm_khugepaged_scan_file(mm, folio, file, present, swap, result);
2392	return result;
2393}
2394
2395static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,
2396					    struct collapse_control *cc)
2397	__releases(&khugepaged_mm_lock)
2398	__acquires(&khugepaged_mm_lock)
2399{
2400	struct vma_iterator vmi;
2401	struct mm_slot *slot;
2402	struct mm_struct *mm;
2403	struct vm_area_struct *vma;
2404	int progress = 0;
2405
2406	VM_BUG_ON(!pages);
2407	lockdep_assert_held(&khugepaged_mm_lock);
2408	*result = SCAN_FAIL;
2409
2410	if (khugepaged_scan.mm_slot) {
2411		slot = khugepaged_scan.mm_slot;
2412	} else {
2413		slot = list_first_entry(&khugepaged_scan.mm_head,
2414				     struct mm_slot, mm_node);
2415		khugepaged_scan.address = 0;
2416		khugepaged_scan.mm_slot = slot;
2417	}
2418	spin_unlock(&khugepaged_mm_lock);
2419
2420	mm = slot->mm;
2421	/*
2422	 * Don't wait for semaphore (to avoid long wait times).  Just move to
2423	 * the next mm on the list.
2424	 */
2425	vma = NULL;
2426	if (unlikely(!mmap_read_trylock(mm)))
2427		goto breakouterloop_mmap_lock;
2428
2429	progress++;
2430	if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
2431		goto breakouterloop;
2432
2433	vma_iter_init(&vmi, mm, khugepaged_scan.address);
2434	for_each_vma(vmi, vma) {
2435		unsigned long hstart, hend;
2436
2437		cond_resched();
2438		if (unlikely(hpage_collapse_test_exit_or_disable(mm))) {
2439			progress++;
2440			break;
2441		}
2442		if (!thp_vma_allowable_order(vma, vma->vm_flags, TVA_KHUGEPAGED, PMD_ORDER)) {
2443skip:
2444			progress++;
2445			continue;
2446		}
2447		hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);
2448		hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);
2449		if (khugepaged_scan.address > hend)
2450			goto skip;
2451		if (khugepaged_scan.address < hstart)
2452			khugepaged_scan.address = hstart;
2453		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2454
2455		while (khugepaged_scan.address < hend) {
2456			bool mmap_locked = true;
2457
2458			cond_resched();
2459			if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
2460				goto breakouterloop;
2461
2462			VM_BUG_ON(khugepaged_scan.address < hstart ||
2463				  khugepaged_scan.address + HPAGE_PMD_SIZE >
2464				  hend);
2465			if (!vma_is_anonymous(vma)) {
2466				struct file *file = get_file(vma->vm_file);
2467				pgoff_t pgoff = linear_page_index(vma,
2468						khugepaged_scan.address);
2469
2470				mmap_read_unlock(mm);
2471				mmap_locked = false;
2472				*result = hpage_collapse_scan_file(mm,
2473					khugepaged_scan.address, file, pgoff, cc);
2474				fput(file);
2475				if (*result == SCAN_PTE_MAPPED_HUGEPAGE) {
2476					mmap_read_lock(mm);
2477					if (hpage_collapse_test_exit_or_disable(mm))
2478						goto breakouterloop;
2479					*result = collapse_pte_mapped_thp(mm,
2480						khugepaged_scan.address, false);
2481					if (*result == SCAN_PMD_MAPPED)
2482						*result = SCAN_SUCCEED;
2483					mmap_read_unlock(mm);
2484				}
2485			} else {
2486				*result = hpage_collapse_scan_pmd(mm, vma,
2487					khugepaged_scan.address, &mmap_locked, cc);
2488			}
2489
2490			if (*result == SCAN_SUCCEED)
2491				++khugepaged_pages_collapsed;
2492
2493			/* move to next address */
2494			khugepaged_scan.address += HPAGE_PMD_SIZE;
2495			progress += HPAGE_PMD_NR;
2496			if (!mmap_locked)
2497				/*
2498				 * We released mmap_lock so break loop.  Note
2499				 * that we drop mmap_lock before all hugepage
2500				 * allocations, so if allocation fails, we are
2501				 * guaranteed to break here and report the
2502				 * correct result back to caller.
2503				 */
2504				goto breakouterloop_mmap_lock;
2505			if (progress >= pages)
2506				goto breakouterloop;
2507		}
2508	}
2509breakouterloop:
2510	mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */
2511breakouterloop_mmap_lock:
2512
2513	spin_lock(&khugepaged_mm_lock);
2514	VM_BUG_ON(khugepaged_scan.mm_slot != slot);
2515	/*
2516	 * Release the current mm_slot if this mm is about to die, or
2517	 * if we scanned all vmas of this mm.
2518	 */
2519	if (hpage_collapse_test_exit(mm) || !vma) {
2520		/*
2521		 * Make sure that if mm_users is reaching zero while
2522		 * khugepaged runs here, khugepaged_exit will find
2523		 * mm_slot not pointing to the exiting mm.
2524		 */
2525		if (!list_is_last(&slot->mm_node, &khugepaged_scan.mm_head)) {
2526			khugepaged_scan.mm_slot = list_next_entry(slot, mm_node);
2527			khugepaged_scan.address = 0;
2528		} else {
2529			khugepaged_scan.mm_slot = NULL;
2530			khugepaged_full_scans++;
2531		}
2532
2533		collect_mm_slot(slot);
2534	}
2535
2536	return progress;
2537}
2538
2539static int khugepaged_has_work(void)
2540{
2541	return !list_empty(&khugepaged_scan.mm_head) && hugepage_pmd_enabled();
2542}
2543
2544static int khugepaged_wait_event(void)
2545{
2546	return !list_empty(&khugepaged_scan.mm_head) ||
2547		kthread_should_stop();
2548}
2549
2550static void khugepaged_do_scan(struct collapse_control *cc)
2551{
2552	unsigned int progress = 0, pass_through_head = 0;
2553	unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);
2554	bool wait = true;
2555	int result = SCAN_SUCCEED;
2556
2557	lru_add_drain_all();
2558
2559	while (true) {
2560		cond_resched();
2561
2562		if (unlikely(kthread_should_stop()))
2563			break;
2564
2565		spin_lock(&khugepaged_mm_lock);
2566		if (!khugepaged_scan.mm_slot)
2567			pass_through_head++;
2568		if (khugepaged_has_work() &&
2569		    pass_through_head < 2)
2570			progress += khugepaged_scan_mm_slot(pages - progress,
2571							    &result, cc);
2572		else
2573			progress = pages;
2574		spin_unlock(&khugepaged_mm_lock);
2575
2576		if (progress >= pages)
2577			break;
2578
2579		if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {
2580			/*
2581			 * If fail to allocate the first time, try to sleep for
2582			 * a while.  When hit again, cancel the scan.
2583			 */
2584			if (!wait)
2585				break;
2586			wait = false;
2587			khugepaged_alloc_sleep();
2588		}
2589	}
2590}
2591
2592static bool khugepaged_should_wakeup(void)
2593{
2594	return kthread_should_stop() ||
2595	       time_after_eq(jiffies, khugepaged_sleep_expire);
2596}
2597
2598static void khugepaged_wait_work(void)
2599{
2600	if (khugepaged_has_work()) {
2601		const unsigned long scan_sleep_jiffies =
2602			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2603
2604		if (!scan_sleep_jiffies)
2605			return;
2606
2607		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2608		wait_event_freezable_timeout(khugepaged_wait,
2609					     khugepaged_should_wakeup(),
2610					     scan_sleep_jiffies);
2611		return;
2612	}
2613
2614	if (hugepage_pmd_enabled())
2615		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2616}
2617
2618static int khugepaged(void *none)
2619{
2620	struct mm_slot *slot;
2621
2622	set_freezable();
2623	set_user_nice(current, MAX_NICE);
2624
2625	while (!kthread_should_stop()) {
2626		khugepaged_do_scan(&khugepaged_collapse_control);
2627		khugepaged_wait_work();
2628	}
2629
2630	spin_lock(&khugepaged_mm_lock);
2631	slot = khugepaged_scan.mm_slot;
2632	khugepaged_scan.mm_slot = NULL;
2633	if (slot)
2634		collect_mm_slot(slot);
2635	spin_unlock(&khugepaged_mm_lock);
2636	return 0;
2637}
2638
2639static void set_recommended_min_free_kbytes(void)
2640{
2641	struct zone *zone;
2642	int nr_zones = 0;
2643	unsigned long recommended_min;
2644
2645	if (!hugepage_pmd_enabled()) {
2646		calculate_min_free_kbytes();
2647		goto update_wmarks;
2648	}
2649
2650	for_each_populated_zone(zone) {
2651		/*
2652		 * We don't need to worry about fragmentation of
2653		 * ZONE_MOVABLE since it only has movable pages.
2654		 */
2655		if (zone_idx(zone) > gfp_zone(GFP_USER))
2656			continue;
2657
2658		nr_zones++;
2659	}
2660
2661	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2662	recommended_min = pageblock_nr_pages * nr_zones * 2;
2663
2664	/*
2665	 * Make sure that on average at least two pageblocks are almost free
2666	 * of another type, one for a migratetype to fall back to and a
2667	 * second to avoid subsequent fallbacks of other types There are 3
2668	 * MIGRATE_TYPES we care about.
2669	 */
2670	recommended_min += pageblock_nr_pages * nr_zones *
2671			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2672
2673	/* don't ever allow to reserve more than 5% of the lowmem */
2674	recommended_min = min(recommended_min,
2675			      (unsigned long) nr_free_buffer_pages() / 20);
2676	recommended_min <<= (PAGE_SHIFT-10);
2677
2678	if (recommended_min > min_free_kbytes) {
2679		if (user_min_free_kbytes >= 0)
2680			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2681				min_free_kbytes, recommended_min);
2682
2683		min_free_kbytes = recommended_min;
2684	}
2685
2686update_wmarks:
2687	setup_per_zone_wmarks();
2688}
2689
2690int start_stop_khugepaged(void)
2691{
2692	int err = 0;
2693
2694	mutex_lock(&khugepaged_mutex);
2695	if (hugepage_pmd_enabled()) {
2696		if (!khugepaged_thread)
2697			khugepaged_thread = kthread_run(khugepaged, NULL,
2698							"khugepaged");
2699		if (IS_ERR(khugepaged_thread)) {
2700			pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2701			err = PTR_ERR(khugepaged_thread);
2702			khugepaged_thread = NULL;
2703			goto fail;
2704		}
2705
2706		if (!list_empty(&khugepaged_scan.mm_head))
2707			wake_up_interruptible(&khugepaged_wait);
2708	} else if (khugepaged_thread) {
2709		kthread_stop(khugepaged_thread);
2710		khugepaged_thread = NULL;
2711	}
2712	set_recommended_min_free_kbytes();
2713fail:
2714	mutex_unlock(&khugepaged_mutex);
2715	return err;
2716}
2717
2718void khugepaged_min_free_kbytes_update(void)
2719{
2720	mutex_lock(&khugepaged_mutex);
2721	if (hugepage_pmd_enabled() && khugepaged_thread)
2722		set_recommended_min_free_kbytes();
2723	mutex_unlock(&khugepaged_mutex);
2724}
2725
2726bool current_is_khugepaged(void)
2727{
2728	return kthread_func(current) == khugepaged;
2729}
2730
2731static int madvise_collapse_errno(enum scan_result r)
2732{
2733	/*
2734	 * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide
2735	 * actionable feedback to caller, so they may take an appropriate
2736	 * fallback measure depending on the nature of the failure.
2737	 */
2738	switch (r) {
2739	case SCAN_ALLOC_HUGE_PAGE_FAIL:
2740		return -ENOMEM;
2741	case SCAN_CGROUP_CHARGE_FAIL:
2742	case SCAN_EXCEED_NONE_PTE:
2743		return -EBUSY;
2744	/* Resource temporary unavailable - trying again might succeed */
2745	case SCAN_PAGE_COUNT:
2746	case SCAN_PAGE_LOCK:
2747	case SCAN_PAGE_LRU:
2748	case SCAN_DEL_PAGE_LRU:
2749	case SCAN_PAGE_FILLED:
2750		return -EAGAIN;
2751	/*
2752	 * Other: Trying again likely not to succeed / error intrinsic to
2753	 * specified memory range. khugepaged likely won't be able to collapse
2754	 * either.
2755	 */
2756	default:
2757		return -EINVAL;
2758	}
2759}
2760
2761int madvise_collapse(struct vm_area_struct *vma, unsigned long start,
2762		     unsigned long end, bool *lock_dropped)
2763{
2764	struct collapse_control *cc;
2765	struct mm_struct *mm = vma->vm_mm;
2766	unsigned long hstart, hend, addr;
2767	int thps = 0, last_fail = SCAN_FAIL;
2768	bool mmap_locked = true;
2769
2770	BUG_ON(vma->vm_start > start);
2771	BUG_ON(vma->vm_end < end);
2772
2773	if (!thp_vma_allowable_order(vma, vma->vm_flags, TVA_FORCED_COLLAPSE, PMD_ORDER))
2774		return -EINVAL;
2775
2776	cc = kmalloc(sizeof(*cc), GFP_KERNEL);
2777	if (!cc)
2778		return -ENOMEM;
2779	cc->is_khugepaged = false;
2780
2781	mmgrab(mm);
2782	lru_add_drain_all();
2783
2784	hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2785	hend = end & HPAGE_PMD_MASK;
2786
2787	for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {
2788		int result = SCAN_FAIL;
2789
2790		if (!mmap_locked) {
2791			cond_resched();
2792			mmap_read_lock(mm);
2793			mmap_locked = true;
2794			result = hugepage_vma_revalidate(mm, addr, false, &vma,
2795							 cc);
2796			if (result  != SCAN_SUCCEED) {
2797				last_fail = result;
2798				goto out_nolock;
2799			}
2800
2801			hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);
2802		}
2803		mmap_assert_locked(mm);
2804		if (!vma_is_anonymous(vma)) {
2805			struct file *file = get_file(vma->vm_file);
2806			pgoff_t pgoff = linear_page_index(vma, addr);
2807
2808			mmap_read_unlock(mm);
2809			mmap_locked = false;
2810			result = hpage_collapse_scan_file(mm, addr, file, pgoff,
2811							  cc);
2812			fput(file);
2813		} else {
2814			result = hpage_collapse_scan_pmd(mm, vma, addr,
2815							 &mmap_locked, cc);
2816		}
2817		if (!mmap_locked)
2818			*lock_dropped = true;
2819
2820handle_result:
2821		switch (result) {
2822		case SCAN_SUCCEED:
2823		case SCAN_PMD_MAPPED:
2824			++thps;
2825			break;
2826		case SCAN_PTE_MAPPED_HUGEPAGE:
2827			BUG_ON(mmap_locked);
2828			mmap_read_lock(mm);
2829			result = collapse_pte_mapped_thp(mm, addr, true);
2830			mmap_read_unlock(mm);
2831			goto handle_result;
2832		/* Whitelisted set of results where continuing OK */
2833		case SCAN_NO_PTE_TABLE:
2834		case SCAN_PTE_NON_PRESENT:
2835		case SCAN_PTE_UFFD_WP:
2836		case SCAN_LACK_REFERENCED_PAGE:
2837		case SCAN_PAGE_NULL:
2838		case SCAN_PAGE_COUNT:
2839		case SCAN_PAGE_LOCK:
2840		case SCAN_PAGE_COMPOUND:
2841		case SCAN_PAGE_LRU:
2842		case SCAN_DEL_PAGE_LRU:
2843			last_fail = result;
2844			break;
2845		default:
2846			last_fail = result;
2847			/* Other error, exit */
2848			goto out_maybelock;
2849		}
2850	}
2851
2852out_maybelock:
2853	/* Caller expects us to hold mmap_lock on return */
2854	if (!mmap_locked)
2855		mmap_read_lock(mm);
2856out_nolock:
2857	mmap_assert_locked(mm);
2858	mmdrop(mm);
2859	kfree(cc);
2860
2861	return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0
2862			: madvise_collapse_errno(last_fail);
2863}