brintos

brintos / linux-shallow public Read only

0
0
Text · 72.3 KiB · b538c3d Raw
2815 lines · c
1// SPDX-License-Identifier: GPL-2.02#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt3 4#include <linux/mm.h>5#include <linux/sched.h>6#include <linux/sched/mm.h>7#include <linux/sched/coredump.h>8#include <linux/mmu_notifier.h>9#include <linux/rmap.h>10#include <linux/swap.h>11#include <linux/mm_inline.h>12#include <linux/kthread.h>13#include <linux/khugepaged.h>14#include <linux/freezer.h>15#include <linux/mman.h>16#include <linux/hashtable.h>17#include <linux/userfaultfd_k.h>18#include <linux/page_idle.h>19#include <linux/page_table_check.h>20#include <linux/rcupdate_wait.h>21#include <linux/swapops.h>22#include <linux/shmem_fs.h>23#include <linux/ksm.h>24 25#include <asm/tlb.h>26#include <asm/pgalloc.h>27#include "internal.h"28#include "mm_slot.h"29 30enum scan_result {31	SCAN_FAIL,32	SCAN_SUCCEED,33	SCAN_PMD_NULL,34	SCAN_PMD_NONE,35	SCAN_PMD_MAPPED,36	SCAN_EXCEED_NONE_PTE,37	SCAN_EXCEED_SWAP_PTE,38	SCAN_EXCEED_SHARED_PTE,39	SCAN_PTE_NON_PRESENT,40	SCAN_PTE_UFFD_WP,41	SCAN_PTE_MAPPED_HUGEPAGE,42	SCAN_PAGE_RO,43	SCAN_LACK_REFERENCED_PAGE,44	SCAN_PAGE_NULL,45	SCAN_SCAN_ABORT,46	SCAN_PAGE_COUNT,47	SCAN_PAGE_LRU,48	SCAN_PAGE_LOCK,49	SCAN_PAGE_ANON,50	SCAN_PAGE_COMPOUND,51	SCAN_ANY_PROCESS,52	SCAN_VMA_NULL,53	SCAN_VMA_CHECK,54	SCAN_ADDRESS_RANGE,55	SCAN_DEL_PAGE_LRU,56	SCAN_ALLOC_HUGE_PAGE_FAIL,57	SCAN_CGROUP_CHARGE_FAIL,58	SCAN_TRUNCATED,59	SCAN_PAGE_HAS_PRIVATE,60	SCAN_STORE_FAILED,61	SCAN_COPY_MC,62	SCAN_PAGE_FILLED,63};64 65#define CREATE_TRACE_POINTS66#include <trace/events/huge_memory.h>67 68static struct task_struct *khugepaged_thread __read_mostly;69static DEFINE_MUTEX(khugepaged_mutex);70 71/* default scan 8*512 pte (or vmas) every 30 second */72static unsigned int khugepaged_pages_to_scan __read_mostly;73static unsigned int khugepaged_pages_collapsed;74static unsigned int khugepaged_full_scans;75static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;76/* during fragmentation poll the hugepage allocator once every minute */77static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;78static unsigned long khugepaged_sleep_expire;79static DEFINE_SPINLOCK(khugepaged_mm_lock);80static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);81/*82 * default collapse hugepages if there is at least one pte mapped like83 * it would have happened if the vma was large enough during page84 * fault.85 *86 * Note that these are only respected if collapse was initiated by khugepaged.87 */88unsigned int khugepaged_max_ptes_none __read_mostly;89static unsigned int khugepaged_max_ptes_swap __read_mostly;90static unsigned int khugepaged_max_ptes_shared __read_mostly;91 92#define MM_SLOTS_HASH_BITS 1093static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);94 95static struct kmem_cache *mm_slot_cache __ro_after_init;96 97struct collapse_control {98	bool is_khugepaged;99 100	/* Num pages scanned per node */101	u32 node_load[MAX_NUMNODES];102 103	/* nodemask for allocation fallback */104	nodemask_t alloc_nmask;105};106 107/**108 * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned109 * @slot: hash lookup from mm to mm_slot110 */111struct khugepaged_mm_slot {112	struct mm_slot slot;113};114 115/**116 * struct khugepaged_scan - cursor for scanning117 * @mm_head: the head of the mm list to scan118 * @mm_slot: the current mm_slot we are scanning119 * @address: the next address inside that to be scanned120 *121 * There is only the one khugepaged_scan instance of this cursor structure.122 */123struct khugepaged_scan {124	struct list_head mm_head;125	struct khugepaged_mm_slot *mm_slot;126	unsigned long address;127};128 129static struct khugepaged_scan khugepaged_scan = {130	.mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),131};132 133#ifdef CONFIG_SYSFS134static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,135					 struct kobj_attribute *attr,136					 char *buf)137{138	return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs);139}140 141static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,142					  struct kobj_attribute *attr,143					  const char *buf, size_t count)144{145	unsigned int msecs;146	int err;147 148	err = kstrtouint(buf, 10, &msecs);149	if (err)150		return -EINVAL;151 152	khugepaged_scan_sleep_millisecs = msecs;153	khugepaged_sleep_expire = 0;154	wake_up_interruptible(&khugepaged_wait);155 156	return count;157}158static struct kobj_attribute scan_sleep_millisecs_attr =159	__ATTR_RW(scan_sleep_millisecs);160 161static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,162					  struct kobj_attribute *attr,163					  char *buf)164{165	return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs);166}167 168static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,169					   struct kobj_attribute *attr,170					   const char *buf, size_t count)171{172	unsigned int msecs;173	int err;174 175	err = kstrtouint(buf, 10, &msecs);176	if (err)177		return -EINVAL;178 179	khugepaged_alloc_sleep_millisecs = msecs;180	khugepaged_sleep_expire = 0;181	wake_up_interruptible(&khugepaged_wait);182 183	return count;184}185static struct kobj_attribute alloc_sleep_millisecs_attr =186	__ATTR_RW(alloc_sleep_millisecs);187 188static ssize_t pages_to_scan_show(struct kobject *kobj,189				  struct kobj_attribute *attr,190				  char *buf)191{192	return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan);193}194static ssize_t pages_to_scan_store(struct kobject *kobj,195				   struct kobj_attribute *attr,196				   const char *buf, size_t count)197{198	unsigned int pages;199	int err;200 201	err = kstrtouint(buf, 10, &pages);202	if (err || !pages)203		return -EINVAL;204 205	khugepaged_pages_to_scan = pages;206 207	return count;208}209static struct kobj_attribute pages_to_scan_attr =210	__ATTR_RW(pages_to_scan);211 212static ssize_t pages_collapsed_show(struct kobject *kobj,213				    struct kobj_attribute *attr,214				    char *buf)215{216	return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed);217}218static struct kobj_attribute pages_collapsed_attr =219	__ATTR_RO(pages_collapsed);220 221static ssize_t full_scans_show(struct kobject *kobj,222			       struct kobj_attribute *attr,223			       char *buf)224{225	return sysfs_emit(buf, "%u\n", khugepaged_full_scans);226}227static struct kobj_attribute full_scans_attr =228	__ATTR_RO(full_scans);229 230static ssize_t defrag_show(struct kobject *kobj,231			   struct kobj_attribute *attr, char *buf)232{233	return single_hugepage_flag_show(kobj, attr, buf,234					 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);235}236static ssize_t defrag_store(struct kobject *kobj,237			    struct kobj_attribute *attr,238			    const char *buf, size_t count)239{240	return single_hugepage_flag_store(kobj, attr, buf, count,241				 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);242}243static struct kobj_attribute khugepaged_defrag_attr =244	__ATTR_RW(defrag);245 246/*247 * max_ptes_none controls if khugepaged should collapse hugepages over248 * any unmapped ptes in turn potentially increasing the memory249 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not250 * reduce the available free memory in the system as it251 * runs. Increasing max_ptes_none will instead potentially reduce the252 * free memory in the system during the khugepaged scan.253 */254static ssize_t max_ptes_none_show(struct kobject *kobj,255				  struct kobj_attribute *attr,256				  char *buf)257{258	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none);259}260static ssize_t max_ptes_none_store(struct kobject *kobj,261				   struct kobj_attribute *attr,262				   const char *buf, size_t count)263{264	int err;265	unsigned long max_ptes_none;266 267	err = kstrtoul(buf, 10, &max_ptes_none);268	if (err || max_ptes_none > HPAGE_PMD_NR - 1)269		return -EINVAL;270 271	khugepaged_max_ptes_none = max_ptes_none;272 273	return count;274}275static struct kobj_attribute khugepaged_max_ptes_none_attr =276	__ATTR_RW(max_ptes_none);277 278static ssize_t max_ptes_swap_show(struct kobject *kobj,279				  struct kobj_attribute *attr,280				  char *buf)281{282	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap);283}284 285static ssize_t max_ptes_swap_store(struct kobject *kobj,286				   struct kobj_attribute *attr,287				   const char *buf, size_t count)288{289	int err;290	unsigned long max_ptes_swap;291 292	err  = kstrtoul(buf, 10, &max_ptes_swap);293	if (err || max_ptes_swap > HPAGE_PMD_NR - 1)294		return -EINVAL;295 296	khugepaged_max_ptes_swap = max_ptes_swap;297 298	return count;299}300 301static struct kobj_attribute khugepaged_max_ptes_swap_attr =302	__ATTR_RW(max_ptes_swap);303 304static ssize_t max_ptes_shared_show(struct kobject *kobj,305				    struct kobj_attribute *attr,306				    char *buf)307{308	return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared);309}310 311static ssize_t max_ptes_shared_store(struct kobject *kobj,312				     struct kobj_attribute *attr,313				     const char *buf, size_t count)314{315	int err;316	unsigned long max_ptes_shared;317 318	err  = kstrtoul(buf, 10, &max_ptes_shared);319	if (err || max_ptes_shared > HPAGE_PMD_NR - 1)320		return -EINVAL;321 322	khugepaged_max_ptes_shared = max_ptes_shared;323 324	return count;325}326 327static struct kobj_attribute khugepaged_max_ptes_shared_attr =328	__ATTR_RW(max_ptes_shared);329 330static struct attribute *khugepaged_attr[] = {331	&khugepaged_defrag_attr.attr,332	&khugepaged_max_ptes_none_attr.attr,333	&khugepaged_max_ptes_swap_attr.attr,334	&khugepaged_max_ptes_shared_attr.attr,335	&pages_to_scan_attr.attr,336	&pages_collapsed_attr.attr,337	&full_scans_attr.attr,338	&scan_sleep_millisecs_attr.attr,339	&alloc_sleep_millisecs_attr.attr,340	NULL,341};342 343struct attribute_group khugepaged_attr_group = {344	.attrs = khugepaged_attr,345	.name = "khugepaged",346};347#endif /* CONFIG_SYSFS */348 349int hugepage_madvise(struct vm_area_struct *vma,350		     unsigned long *vm_flags, int advice)351{352	switch (advice) {353	case MADV_HUGEPAGE:354#ifdef CONFIG_S390355		/*356		 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390357		 * can't handle this properly after s390_enable_sie, so we simply358		 * ignore the madvise to prevent qemu from causing a SIGSEGV.359		 */360		if (mm_has_pgste(vma->vm_mm))361			return 0;362#endif363		*vm_flags &= ~VM_NOHUGEPAGE;364		*vm_flags |= VM_HUGEPAGE;365		/*366		 * If the vma become good for khugepaged to scan,367		 * register it here without waiting a page fault that368		 * may not happen any time soon.369		 */370		khugepaged_enter_vma(vma, *vm_flags);371		break;372	case MADV_NOHUGEPAGE:373		*vm_flags &= ~VM_HUGEPAGE;374		*vm_flags |= VM_NOHUGEPAGE;375		/*376		 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning377		 * this vma even if we leave the mm registered in khugepaged if378		 * it got registered before VM_NOHUGEPAGE was set.379		 */380		break;381	}382 383	return 0;384}385 386int __init khugepaged_init(void)387{388	mm_slot_cache = KMEM_CACHE(khugepaged_mm_slot, 0);389	if (!mm_slot_cache)390		return -ENOMEM;391 392	khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;393	khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;394	khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;395	khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2;396 397	return 0;398}399 400void __init khugepaged_destroy(void)401{402	kmem_cache_destroy(mm_slot_cache);403}404 405static inline int hpage_collapse_test_exit(struct mm_struct *mm)406{407	return atomic_read(&mm->mm_users) == 0;408}409 410static inline int hpage_collapse_test_exit_or_disable(struct mm_struct *mm)411{412	return hpage_collapse_test_exit(mm) ||413	       test_bit(MMF_DISABLE_THP, &mm->flags);414}415 416static bool hugepage_pmd_enabled(void)417{418	/*419	 * We cover both the anon and the file-backed case here; file-backed420	 * hugepages, when configured in, are determined by the global control.421	 * Anon pmd-sized hugepages are determined by the pmd-size control.422	 */423	if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&424	    hugepage_global_enabled())425		return true;426	if (test_bit(PMD_ORDER, &huge_anon_orders_always))427		return true;428	if (test_bit(PMD_ORDER, &huge_anon_orders_madvise))429		return true;430	if (test_bit(PMD_ORDER, &huge_anon_orders_inherit) &&431	    hugepage_global_enabled())432		return true;433	return false;434}435 436void __khugepaged_enter(struct mm_struct *mm)437{438	struct khugepaged_mm_slot *mm_slot;439	struct mm_slot *slot;440	int wakeup;441 442	/* __khugepaged_exit() must not run from under us */443	VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm);444	if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags)))445		return;446 447	mm_slot = mm_slot_alloc(mm_slot_cache);448	if (!mm_slot)449		return;450 451	slot = &mm_slot->slot;452 453	spin_lock(&khugepaged_mm_lock);454	mm_slot_insert(mm_slots_hash, mm, slot);455	/*456	 * Insert just behind the scanning cursor, to let the area settle457	 * down a little.458	 */459	wakeup = list_empty(&khugepaged_scan.mm_head);460	list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head);461	spin_unlock(&khugepaged_mm_lock);462 463	mmgrab(mm);464	if (wakeup)465		wake_up_interruptible(&khugepaged_wait);466}467 468void khugepaged_enter_vma(struct vm_area_struct *vma,469			  unsigned long vm_flags)470{471	if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) &&472	    hugepage_pmd_enabled()) {473		if (thp_vma_allowable_order(vma, vm_flags, TVA_ENFORCE_SYSFS,474					    PMD_ORDER))475			__khugepaged_enter(vma->vm_mm);476	}477}478 479void __khugepaged_exit(struct mm_struct *mm)480{481	struct khugepaged_mm_slot *mm_slot;482	struct mm_slot *slot;483	int free = 0;484 485	spin_lock(&khugepaged_mm_lock);486	slot = mm_slot_lookup(mm_slots_hash, mm);487	mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);488	if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {489		hash_del(&slot->hash);490		list_del(&slot->mm_node);491		free = 1;492	}493	spin_unlock(&khugepaged_mm_lock);494 495	if (free) {496		clear_bit(MMF_VM_HUGEPAGE, &mm->flags);497		mm_slot_free(mm_slot_cache, mm_slot);498		mmdrop(mm);499	} else if (mm_slot) {500		/*501		 * This is required to serialize against502		 * hpage_collapse_test_exit() (which is guaranteed to run503		 * under mmap sem read mode). Stop here (after we return all504		 * pagetables will be destroyed) until khugepaged has finished505		 * working on the pagetables under the mmap_lock.506		 */507		mmap_write_lock(mm);508		mmap_write_unlock(mm);509	}510}511 512static void release_pte_folio(struct folio *folio)513{514	node_stat_mod_folio(folio,515			NR_ISOLATED_ANON + folio_is_file_lru(folio),516			-folio_nr_pages(folio));517	folio_unlock(folio);518	folio_putback_lru(folio);519}520 521static void release_pte_pages(pte_t *pte, pte_t *_pte,522		struct list_head *compound_pagelist)523{524	struct folio *folio, *tmp;525 526	while (--_pte >= pte) {527		pte_t pteval = ptep_get(_pte);528		unsigned long pfn;529 530		if (pte_none(pteval))531			continue;532		pfn = pte_pfn(pteval);533		if (is_zero_pfn(pfn))534			continue;535		folio = pfn_folio(pfn);536		if (folio_test_large(folio))537			continue;538		release_pte_folio(folio);539	}540 541	list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) {542		list_del(&folio->lru);543		release_pte_folio(folio);544	}545}546 547static bool is_refcount_suitable(struct folio *folio)548{549	int expected_refcount = folio_mapcount(folio);550 551	if (!folio_test_anon(folio) || folio_test_swapcache(folio))552		expected_refcount += folio_nr_pages(folio);553 554	if (folio_test_private(folio))555		expected_refcount++;556 557	return folio_ref_count(folio) == expected_refcount;558}559 560static int __collapse_huge_page_isolate(struct vm_area_struct *vma,561					unsigned long address,562					pte_t *pte,563					struct collapse_control *cc,564					struct list_head *compound_pagelist)565{566	struct page *page = NULL;567	struct folio *folio = NULL;568	pte_t *_pte;569	int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0;570	bool writable = false;571 572	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;573	     _pte++, address += PAGE_SIZE) {574		pte_t pteval = ptep_get(_pte);575		if (pte_none(pteval) || (pte_present(pteval) &&576				is_zero_pfn(pte_pfn(pteval)))) {577			++none_or_zero;578			if (!userfaultfd_armed(vma) &&579			    (!cc->is_khugepaged ||580			     none_or_zero <= khugepaged_max_ptes_none)) {581				continue;582			} else {583				result = SCAN_EXCEED_NONE_PTE;584				count_vm_event(THP_SCAN_EXCEED_NONE_PTE);585				goto out;586			}587		}588		if (!pte_present(pteval)) {589			result = SCAN_PTE_NON_PRESENT;590			goto out;591		}592		if (pte_uffd_wp(pteval)) {593			result = SCAN_PTE_UFFD_WP;594			goto out;595		}596		page = vm_normal_page(vma, address, pteval);597		if (unlikely(!page) || unlikely(is_zone_device_page(page))) {598			result = SCAN_PAGE_NULL;599			goto out;600		}601 602		folio = page_folio(page);603		VM_BUG_ON_FOLIO(!folio_test_anon(folio), folio);604 605		/* See hpage_collapse_scan_pmd(). */606		if (folio_likely_mapped_shared(folio)) {607			++shared;608			if (cc->is_khugepaged &&609			    shared > khugepaged_max_ptes_shared) {610				result = SCAN_EXCEED_SHARED_PTE;611				count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);612				goto out;613			}614		}615 616		if (folio_test_large(folio)) {617			struct folio *f;618 619			/*620			 * Check if we have dealt with the compound page621			 * already622			 */623			list_for_each_entry(f, compound_pagelist, lru) {624				if (folio == f)625					goto next;626			}627		}628 629		/*630		 * We can do it before folio_isolate_lru because the631		 * folio can't be freed from under us. NOTE: PG_lock632		 * is needed to serialize against split_huge_page633		 * when invoked from the VM.634		 */635		if (!folio_trylock(folio)) {636			result = SCAN_PAGE_LOCK;637			goto out;638		}639 640		/*641		 * Check if the page has any GUP (or other external) pins.642		 *643		 * The page table that maps the page has been already unlinked644		 * from the page table tree and this process cannot get645		 * an additional pin on the page.646		 *647		 * New pins can come later if the page is shared across fork,648		 * but not from this process. The other process cannot write to649		 * the page, only trigger CoW.650		 */651		if (!is_refcount_suitable(folio)) {652			folio_unlock(folio);653			result = SCAN_PAGE_COUNT;654			goto out;655		}656 657		/*658		 * Isolate the page to avoid collapsing an hugepage659		 * currently in use by the VM.660		 */661		if (!folio_isolate_lru(folio)) {662			folio_unlock(folio);663			result = SCAN_DEL_PAGE_LRU;664			goto out;665		}666		node_stat_mod_folio(folio,667				NR_ISOLATED_ANON + folio_is_file_lru(folio),668				folio_nr_pages(folio));669		VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);670		VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);671 672		if (folio_test_large(folio))673			list_add_tail(&folio->lru, compound_pagelist);674next:675		/*676		 * If collapse was initiated by khugepaged, check that there is677		 * enough young pte to justify collapsing the page678		 */679		if (cc->is_khugepaged &&680		    (pte_young(pteval) || folio_test_young(folio) ||681		     folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,682								     address)))683			referenced++;684 685		if (pte_write(pteval))686			writable = true;687	}688 689	if (unlikely(!writable)) {690		result = SCAN_PAGE_RO;691	} else if (unlikely(cc->is_khugepaged && !referenced)) {692		result = SCAN_LACK_REFERENCED_PAGE;693	} else {694		result = SCAN_SUCCEED;695		trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,696						    referenced, writable, result);697		return result;698	}699out:700	release_pte_pages(pte, _pte, compound_pagelist);701	trace_mm_collapse_huge_page_isolate(&folio->page, none_or_zero,702					    referenced, writable, result);703	return result;704}705 706static void __collapse_huge_page_copy_succeeded(pte_t *pte,707						struct vm_area_struct *vma,708						unsigned long address,709						spinlock_t *ptl,710						struct list_head *compound_pagelist)711{712	struct folio *src, *tmp;713	pte_t *_pte;714	pte_t pteval;715 716	for (_pte = pte; _pte < pte + HPAGE_PMD_NR;717	     _pte++, address += PAGE_SIZE) {718		pteval = ptep_get(_pte);719		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {720			add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);721			if (is_zero_pfn(pte_pfn(pteval))) {722				/*723				 * ptl mostly unnecessary.724				 */725				spin_lock(ptl);726				ptep_clear(vma->vm_mm, address, _pte);727				spin_unlock(ptl);728				ksm_might_unmap_zero_page(vma->vm_mm, pteval);729			}730		} else {731			struct page *src_page = pte_page(pteval);732 733			src = page_folio(src_page);734			if (!folio_test_large(src))735				release_pte_folio(src);736			/*737			 * ptl mostly unnecessary, but preempt has to738			 * be disabled to update the per-cpu stats739			 * inside folio_remove_rmap_pte().740			 */741			spin_lock(ptl);742			ptep_clear(vma->vm_mm, address, _pte);743			folio_remove_rmap_pte(src, src_page, vma);744			spin_unlock(ptl);745			free_page_and_swap_cache(src_page);746		}747	}748 749	list_for_each_entry_safe(src, tmp, compound_pagelist, lru) {750		list_del(&src->lru);751		node_stat_sub_folio(src, NR_ISOLATED_ANON +752				folio_is_file_lru(src));753		folio_unlock(src);754		free_swap_cache(src);755		folio_putback_lru(src);756	}757}758 759static void __collapse_huge_page_copy_failed(pte_t *pte,760					     pmd_t *pmd,761					     pmd_t orig_pmd,762					     struct vm_area_struct *vma,763					     struct list_head *compound_pagelist)764{765	spinlock_t *pmd_ptl;766 767	/*768	 * Re-establish the PMD to point to the original page table769	 * entry. Restoring PMD needs to be done prior to releasing770	 * pages. Since pages are still isolated and locked here,771	 * acquiring anon_vma_lock_write is unnecessary.772	 */773	pmd_ptl = pmd_lock(vma->vm_mm, pmd);774	pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd));775	spin_unlock(pmd_ptl);776	/*777	 * Release both raw and compound pages isolated778	 * in __collapse_huge_page_isolate.779	 */780	release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist);781}782 783/*784 * __collapse_huge_page_copy - attempts to copy memory contents from raw785 * pages to a hugepage. Cleans up the raw pages if copying succeeds;786 * otherwise restores the original page table and releases isolated raw pages.787 * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC.788 *789 * @pte: starting of the PTEs to copy from790 * @folio: the new hugepage to copy contents to791 * @pmd: pointer to the new hugepage's PMD792 * @orig_pmd: the original raw pages' PMD793 * @vma: the original raw pages' virtual memory area794 * @address: starting address to copy795 * @ptl: lock on raw pages' PTEs796 * @compound_pagelist: list that stores compound pages797 */798static int __collapse_huge_page_copy(pte_t *pte, struct folio *folio,799		pmd_t *pmd, pmd_t orig_pmd, struct vm_area_struct *vma,800		unsigned long address, spinlock_t *ptl,801		struct list_head *compound_pagelist)802{803	unsigned int i;804	int result = SCAN_SUCCEED;805 806	/*807	 * Copying pages' contents is subject to memory poison at any iteration.808	 */809	for (i = 0; i < HPAGE_PMD_NR; i++) {810		pte_t pteval = ptep_get(pte + i);811		struct page *page = folio_page(folio, i);812		unsigned long src_addr = address + i * PAGE_SIZE;813		struct page *src_page;814 815		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {816			clear_user_highpage(page, src_addr);817			continue;818		}819		src_page = pte_page(pteval);820		if (copy_mc_user_highpage(page, src_page, src_addr, vma) > 0) {821			result = SCAN_COPY_MC;822			break;823		}824	}825 826	if (likely(result == SCAN_SUCCEED))827		__collapse_huge_page_copy_succeeded(pte, vma, address, ptl,828						    compound_pagelist);829	else830		__collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma,831						 compound_pagelist);832 833	return result;834}835 836static void khugepaged_alloc_sleep(void)837{838	DEFINE_WAIT(wait);839 840	add_wait_queue(&khugepaged_wait, &wait);841	__set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);842	schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));843	remove_wait_queue(&khugepaged_wait, &wait);844}845 846struct collapse_control khugepaged_collapse_control = {847	.is_khugepaged = true,848};849 850static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc)851{852	int i;853 854	/*855	 * If node_reclaim_mode is disabled, then no extra effort is made to856	 * allocate memory locally.857	 */858	if (!node_reclaim_enabled())859		return false;860 861	/* If there is a count for this node already, it must be acceptable */862	if (cc->node_load[nid])863		return false;864 865	for (i = 0; i < MAX_NUMNODES; i++) {866		if (!cc->node_load[i])867			continue;868		if (node_distance(nid, i) > node_reclaim_distance)869			return true;870	}871	return false;872}873 874#define khugepaged_defrag()					\875	(transparent_hugepage_flags &				\876	 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG))877 878/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */879static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)880{881	return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;882}883 884#ifdef CONFIG_NUMA885static int hpage_collapse_find_target_node(struct collapse_control *cc)886{887	int nid, target_node = 0, max_value = 0;888 889	/* find first node with max normal pages hit */890	for (nid = 0; nid < MAX_NUMNODES; nid++)891		if (cc->node_load[nid] > max_value) {892			max_value = cc->node_load[nid];893			target_node = nid;894		}895 896	for_each_online_node(nid) {897		if (max_value == cc->node_load[nid])898			node_set(nid, cc->alloc_nmask);899	}900 901	return target_node;902}903#else904static int hpage_collapse_find_target_node(struct collapse_control *cc)905{906	return 0;907}908#endif909 910/*911 * If mmap_lock temporarily dropped, revalidate vma912 * before taking mmap_lock.913 * Returns enum scan_result value.914 */915 916static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,917				   bool expect_anon,918				   struct vm_area_struct **vmap,919				   struct collapse_control *cc)920{921	struct vm_area_struct *vma;922	unsigned long tva_flags = cc->is_khugepaged ? TVA_ENFORCE_SYSFS : 0;923 924	if (unlikely(hpage_collapse_test_exit_or_disable(mm)))925		return SCAN_ANY_PROCESS;926 927	*vmap = vma = find_vma(mm, address);928	if (!vma)929		return SCAN_VMA_NULL;930 931	if (!thp_vma_suitable_order(vma, address, PMD_ORDER))932		return SCAN_ADDRESS_RANGE;933	if (!thp_vma_allowable_order(vma, vma->vm_flags, tva_flags, PMD_ORDER))934		return SCAN_VMA_CHECK;935	/*936	 * Anon VMA expected, the address may be unmapped then937	 * remapped to file after khugepaged reaquired the mmap_lock.938	 *939	 * thp_vma_allowable_order may return true for qualified file940	 * vmas.941	 */942	if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap)))943		return SCAN_PAGE_ANON;944	return SCAN_SUCCEED;945}946 947static int find_pmd_or_thp_or_none(struct mm_struct *mm,948				   unsigned long address,949				   pmd_t **pmd)950{951	pmd_t pmde;952 953	*pmd = mm_find_pmd(mm, address);954	if (!*pmd)955		return SCAN_PMD_NULL;956 957	pmde = pmdp_get_lockless(*pmd);958	if (pmd_none(pmde))959		return SCAN_PMD_NONE;960	if (!pmd_present(pmde))961		return SCAN_PMD_NULL;962	if (pmd_trans_huge(pmde))963		return SCAN_PMD_MAPPED;964	if (pmd_devmap(pmde))965		return SCAN_PMD_NULL;966	if (pmd_bad(pmde))967		return SCAN_PMD_NULL;968	return SCAN_SUCCEED;969}970 971static int check_pmd_still_valid(struct mm_struct *mm,972				 unsigned long address,973				 pmd_t *pmd)974{975	pmd_t *new_pmd;976	int result = find_pmd_or_thp_or_none(mm, address, &new_pmd);977 978	if (result != SCAN_SUCCEED)979		return result;980	if (new_pmd != pmd)981		return SCAN_FAIL;982	return SCAN_SUCCEED;983}984 985/*986 * Bring missing pages in from swap, to complete THP collapse.987 * Only done if hpage_collapse_scan_pmd believes it is worthwhile.988 *989 * Called and returns without pte mapped or spinlocks held.990 * Returns result: if not SCAN_SUCCEED, mmap_lock has been released.991 */992static int __collapse_huge_page_swapin(struct mm_struct *mm,993				       struct vm_area_struct *vma,994				       unsigned long haddr, pmd_t *pmd,995				       int referenced)996{997	int swapped_in = 0;998	vm_fault_t ret = 0;999	unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE);1000	int result;1001	pte_t *pte = NULL;1002	spinlock_t *ptl;1003 1004	for (address = haddr; address < end; address += PAGE_SIZE) {1005		struct vm_fault vmf = {1006			.vma = vma,1007			.address = address,1008			.pgoff = linear_page_index(vma, address),1009			.flags = FAULT_FLAG_ALLOW_RETRY,1010			.pmd = pmd,1011		};1012 1013		if (!pte++) {1014			pte = pte_offset_map_nolock(mm, pmd, address, &ptl);1015			if (!pte) {1016				mmap_read_unlock(mm);1017				result = SCAN_PMD_NULL;1018				goto out;1019			}1020		}1021 1022		vmf.orig_pte = ptep_get_lockless(pte);1023		if (!is_swap_pte(vmf.orig_pte))1024			continue;1025 1026		vmf.pte = pte;1027		vmf.ptl = ptl;1028		ret = do_swap_page(&vmf);1029		/* Which unmaps pte (after perhaps re-checking the entry) */1030		pte = NULL;1031 1032		/*1033		 * do_swap_page returns VM_FAULT_RETRY with released mmap_lock.1034		 * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because1035		 * we do not retry here and swap entry will remain in pagetable1036		 * resulting in later failure.1037		 */1038		if (ret & VM_FAULT_RETRY) {1039			/* Likely, but not guaranteed, that page lock failed */1040			result = SCAN_PAGE_LOCK;1041			goto out;1042		}1043		if (ret & VM_FAULT_ERROR) {1044			mmap_read_unlock(mm);1045			result = SCAN_FAIL;1046			goto out;1047		}1048		swapped_in++;1049	}1050 1051	if (pte)1052		pte_unmap(pte);1053 1054	/* Drain LRU cache to remove extra pin on the swapped in pages */1055	if (swapped_in)1056		lru_add_drain();1057 1058	result = SCAN_SUCCEED;1059out:1060	trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result);1061	return result;1062}1063 1064static int alloc_charge_folio(struct folio **foliop, struct mm_struct *mm,1065			      struct collapse_control *cc)1066{1067	gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() :1068		     GFP_TRANSHUGE);1069	int node = hpage_collapse_find_target_node(cc);1070	struct folio *folio;1071 1072	folio = __folio_alloc(gfp, HPAGE_PMD_ORDER, node, &cc->alloc_nmask);1073	if (!folio) {1074		*foliop = NULL;1075		count_vm_event(THP_COLLAPSE_ALLOC_FAILED);1076		return SCAN_ALLOC_HUGE_PAGE_FAIL;1077	}1078 1079	count_vm_event(THP_COLLAPSE_ALLOC);1080	if (unlikely(mem_cgroup_charge(folio, mm, gfp))) {1081		folio_put(folio);1082		*foliop = NULL;1083		return SCAN_CGROUP_CHARGE_FAIL;1084	}1085 1086	count_memcg_folio_events(folio, THP_COLLAPSE_ALLOC, 1);1087 1088	*foliop = folio;1089	return SCAN_SUCCEED;1090}1091 1092static int collapse_huge_page(struct mm_struct *mm, unsigned long address,1093			      int referenced, int unmapped,1094			      struct collapse_control *cc)1095{1096	LIST_HEAD(compound_pagelist);1097	pmd_t *pmd, _pmd;1098	pte_t *pte;1099	pgtable_t pgtable;1100	struct folio *folio;1101	spinlock_t *pmd_ptl, *pte_ptl;1102	int result = SCAN_FAIL;1103	struct vm_area_struct *vma;1104	struct mmu_notifier_range range;1105 1106	VM_BUG_ON(address & ~HPAGE_PMD_MASK);1107 1108	/*1109	 * Before allocating the hugepage, release the mmap_lock read lock.1110	 * The allocation can take potentially a long time if it involves1111	 * sync compaction, and we do not need to hold the mmap_lock during1112	 * that. We will recheck the vma after taking it again in write mode.1113	 */1114	mmap_read_unlock(mm);1115 1116	result = alloc_charge_folio(&folio, mm, cc);1117	if (result != SCAN_SUCCEED)1118		goto out_nolock;1119 1120	mmap_read_lock(mm);1121	result = hugepage_vma_revalidate(mm, address, true, &vma, cc);1122	if (result != SCAN_SUCCEED) {1123		mmap_read_unlock(mm);1124		goto out_nolock;1125	}1126 1127	result = find_pmd_or_thp_or_none(mm, address, &pmd);1128	if (result != SCAN_SUCCEED) {1129		mmap_read_unlock(mm);1130		goto out_nolock;1131	}1132 1133	if (unmapped) {1134		/*1135		 * __collapse_huge_page_swapin will return with mmap_lock1136		 * released when it fails. So we jump out_nolock directly in1137		 * that case.  Continuing to collapse causes inconsistency.1138		 */1139		result = __collapse_huge_page_swapin(mm, vma, address, pmd,1140						     referenced);1141		if (result != SCAN_SUCCEED)1142			goto out_nolock;1143	}1144 1145	mmap_read_unlock(mm);1146	/*1147	 * Prevent all access to pagetables with the exception of1148	 * gup_fast later handled by the ptep_clear_flush and the VM1149	 * handled by the anon_vma lock + PG_lock.1150	 *1151	 * UFFDIO_MOVE is prevented to race as well thanks to the1152	 * mmap_lock.1153	 */1154	mmap_write_lock(mm);1155	result = hugepage_vma_revalidate(mm, address, true, &vma, cc);1156	if (result != SCAN_SUCCEED)1157		goto out_up_write;1158	/* check if the pmd is still valid */1159	result = check_pmd_still_valid(mm, address, pmd);1160	if (result != SCAN_SUCCEED)1161		goto out_up_write;1162 1163	vma_start_write(vma);1164	anon_vma_lock_write(vma->anon_vma);1165 1166	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,1167				address + HPAGE_PMD_SIZE);1168	mmu_notifier_invalidate_range_start(&range);1169 1170	pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */1171	/*1172	 * This removes any huge TLB entry from the CPU so we won't allow1173	 * huge and small TLB entries for the same virtual address to1174	 * avoid the risk of CPU bugs in that area.1175	 *1176	 * Parallel GUP-fast is fine since GUP-fast will back off when1177	 * it detects PMD is changed.1178	 */1179	_pmd = pmdp_collapse_flush(vma, address, pmd);1180	spin_unlock(pmd_ptl);1181	mmu_notifier_invalidate_range_end(&range);1182	tlb_remove_table_sync_one();1183 1184	pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl);1185	if (pte) {1186		result = __collapse_huge_page_isolate(vma, address, pte, cc,1187						      &compound_pagelist);1188		spin_unlock(pte_ptl);1189	} else {1190		result = SCAN_PMD_NULL;1191	}1192 1193	if (unlikely(result != SCAN_SUCCEED)) {1194		if (pte)1195			pte_unmap(pte);1196		spin_lock(pmd_ptl);1197		BUG_ON(!pmd_none(*pmd));1198		/*1199		 * We can only use set_pmd_at when establishing1200		 * hugepmds and never for establishing regular pmds that1201		 * points to regular pagetables. Use pmd_populate for that1202		 */1203		pmd_populate(mm, pmd, pmd_pgtable(_pmd));1204		spin_unlock(pmd_ptl);1205		anon_vma_unlock_write(vma->anon_vma);1206		goto out_up_write;1207	}1208 1209	/*1210	 * All pages are isolated and locked so anon_vma rmap1211	 * can't run anymore.1212	 */1213	anon_vma_unlock_write(vma->anon_vma);1214 1215	result = __collapse_huge_page_copy(pte, folio, pmd, _pmd,1216					   vma, address, pte_ptl,1217					   &compound_pagelist);1218	pte_unmap(pte);1219	if (unlikely(result != SCAN_SUCCEED))1220		goto out_up_write;1221 1222	/*1223	 * The smp_wmb() inside __folio_mark_uptodate() ensures the1224	 * copy_huge_page writes become visible before the set_pmd_at()1225	 * write.1226	 */1227	__folio_mark_uptodate(folio);1228	pgtable = pmd_pgtable(_pmd);1229 1230	_pmd = mk_huge_pmd(&folio->page, vma->vm_page_prot);1231	_pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);1232 1233	spin_lock(pmd_ptl);1234	BUG_ON(!pmd_none(*pmd));1235	folio_add_new_anon_rmap(folio, vma, address, RMAP_EXCLUSIVE);1236	folio_add_lru_vma(folio, vma);1237	pgtable_trans_huge_deposit(mm, pmd, pgtable);1238	set_pmd_at(mm, address, pmd, _pmd);1239	update_mmu_cache_pmd(vma, address, pmd);1240	deferred_split_folio(folio, false);1241	spin_unlock(pmd_ptl);1242 1243	folio = NULL;1244 1245	result = SCAN_SUCCEED;1246out_up_write:1247	mmap_write_unlock(mm);1248out_nolock:1249	if (folio)1250		folio_put(folio);1251	trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result);1252	return result;1253}1254 1255static int hpage_collapse_scan_pmd(struct mm_struct *mm,1256				   struct vm_area_struct *vma,1257				   unsigned long address, bool *mmap_locked,1258				   struct collapse_control *cc)1259{1260	pmd_t *pmd;1261	pte_t *pte, *_pte;1262	int result = SCAN_FAIL, referenced = 0;1263	int none_or_zero = 0, shared = 0;1264	struct page *page = NULL;1265	struct folio *folio = NULL;1266	unsigned long _address;1267	spinlock_t *ptl;1268	int node = NUMA_NO_NODE, unmapped = 0;1269	bool writable = false;1270 1271	VM_BUG_ON(address & ~HPAGE_PMD_MASK);1272 1273	result = find_pmd_or_thp_or_none(mm, address, &pmd);1274	if (result != SCAN_SUCCEED)1275		goto out;1276 1277	memset(cc->node_load, 0, sizeof(cc->node_load));1278	nodes_clear(cc->alloc_nmask);1279	pte = pte_offset_map_lock(mm, pmd, address, &ptl);1280	if (!pte) {1281		result = SCAN_PMD_NULL;1282		goto out;1283	}1284 1285	for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR;1286	     _pte++, _address += PAGE_SIZE) {1287		pte_t pteval = ptep_get(_pte);1288		if (is_swap_pte(pteval)) {1289			++unmapped;1290			if (!cc->is_khugepaged ||1291			    unmapped <= khugepaged_max_ptes_swap) {1292				/*1293				 * Always be strict with uffd-wp1294				 * enabled swap entries.  Please see1295				 * comment below for pte_uffd_wp().1296				 */1297				if (pte_swp_uffd_wp_any(pteval)) {1298					result = SCAN_PTE_UFFD_WP;1299					goto out_unmap;1300				}1301				continue;1302			} else {1303				result = SCAN_EXCEED_SWAP_PTE;1304				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);1305				goto out_unmap;1306			}1307		}1308		if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {1309			++none_or_zero;1310			if (!userfaultfd_armed(vma) &&1311			    (!cc->is_khugepaged ||1312			     none_or_zero <= khugepaged_max_ptes_none)) {1313				continue;1314			} else {1315				result = SCAN_EXCEED_NONE_PTE;1316				count_vm_event(THP_SCAN_EXCEED_NONE_PTE);1317				goto out_unmap;1318			}1319		}1320		if (pte_uffd_wp(pteval)) {1321			/*1322			 * Don't collapse the page if any of the small1323			 * PTEs are armed with uffd write protection.1324			 * Here we can also mark the new huge pmd as1325			 * write protected if any of the small ones is1326			 * marked but that could bring unknown1327			 * userfault messages that falls outside of1328			 * the registered range.  So, just be simple.1329			 */1330			result = SCAN_PTE_UFFD_WP;1331			goto out_unmap;1332		}1333		if (pte_write(pteval))1334			writable = true;1335 1336		page = vm_normal_page(vma, _address, pteval);1337		if (unlikely(!page) || unlikely(is_zone_device_page(page))) {1338			result = SCAN_PAGE_NULL;1339			goto out_unmap;1340		}1341		folio = page_folio(page);1342 1343		if (!folio_test_anon(folio)) {1344			result = SCAN_PAGE_ANON;1345			goto out_unmap;1346		}1347 1348		/*1349		 * We treat a single page as shared if any part of the THP1350		 * is shared. "False negatives" from1351		 * folio_likely_mapped_shared() are not expected to matter1352		 * much in practice.1353		 */1354		if (folio_likely_mapped_shared(folio)) {1355			++shared;1356			if (cc->is_khugepaged &&1357			    shared > khugepaged_max_ptes_shared) {1358				result = SCAN_EXCEED_SHARED_PTE;1359				count_vm_event(THP_SCAN_EXCEED_SHARED_PTE);1360				goto out_unmap;1361			}1362		}1363 1364		/*1365		 * Record which node the original page is from and save this1366		 * information to cc->node_load[].1367		 * Khugepaged will allocate hugepage from the node has the max1368		 * hit record.1369		 */1370		node = folio_nid(folio);1371		if (hpage_collapse_scan_abort(node, cc)) {1372			result = SCAN_SCAN_ABORT;1373			goto out_unmap;1374		}1375		cc->node_load[node]++;1376		if (!folio_test_lru(folio)) {1377			result = SCAN_PAGE_LRU;1378			goto out_unmap;1379		}1380		if (folio_test_locked(folio)) {1381			result = SCAN_PAGE_LOCK;1382			goto out_unmap;1383		}1384 1385		/*1386		 * Check if the page has any GUP (or other external) pins.1387		 *1388		 * Here the check may be racy:1389		 * it may see folio_mapcount() > folio_ref_count().1390		 * But such case is ephemeral we could always retry collapse1391		 * later.  However it may report false positive if the page1392		 * has excessive GUP pins (i.e. 512).  Anyway the same check1393		 * will be done again later the risk seems low.1394		 */1395		if (!is_refcount_suitable(folio)) {1396			result = SCAN_PAGE_COUNT;1397			goto out_unmap;1398		}1399 1400		/*1401		 * If collapse was initiated by khugepaged, check that there is1402		 * enough young pte to justify collapsing the page1403		 */1404		if (cc->is_khugepaged &&1405		    (pte_young(pteval) || folio_test_young(folio) ||1406		     folio_test_referenced(folio) || mmu_notifier_test_young(vma->vm_mm,1407								     address)))1408			referenced++;1409	}1410	if (!writable) {1411		result = SCAN_PAGE_RO;1412	} else if (cc->is_khugepaged &&1413		   (!referenced ||1414		    (unmapped && referenced < HPAGE_PMD_NR / 2))) {1415		result = SCAN_LACK_REFERENCED_PAGE;1416	} else {1417		result = SCAN_SUCCEED;1418	}1419out_unmap:1420	pte_unmap_unlock(pte, ptl);1421	if (result == SCAN_SUCCEED) {1422		result = collapse_huge_page(mm, address, referenced,1423					    unmapped, cc);1424		/* collapse_huge_page will return with the mmap_lock released */1425		*mmap_locked = false;1426	}1427out:1428	trace_mm_khugepaged_scan_pmd(mm, &folio->page, writable, referenced,1429				     none_or_zero, result, unmapped);1430	return result;1431}1432 1433static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot)1434{1435	struct mm_slot *slot = &mm_slot->slot;1436	struct mm_struct *mm = slot->mm;1437 1438	lockdep_assert_held(&khugepaged_mm_lock);1439 1440	if (hpage_collapse_test_exit(mm)) {1441		/* free mm_slot */1442		hash_del(&slot->hash);1443		list_del(&slot->mm_node);1444 1445		/*1446		 * Not strictly needed because the mm exited already.1447		 *1448		 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);1449		 */1450 1451		/* khugepaged_mm_lock actually not necessary for the below */1452		mm_slot_free(mm_slot_cache, mm_slot);1453		mmdrop(mm);1454	}1455}1456 1457#ifdef CONFIG_SHMEM1458/* hpage must be locked, and mmap_lock must be held */1459static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,1460			pmd_t *pmdp, struct page *hpage)1461{1462	struct vm_fault vmf = {1463		.vma = vma,1464		.address = addr,1465		.flags = 0,1466		.pmd = pmdp,1467	};1468 1469	VM_BUG_ON(!PageTransHuge(hpage));1470	mmap_assert_locked(vma->vm_mm);1471 1472	if (do_set_pmd(&vmf, hpage))1473		return SCAN_FAIL;1474 1475	get_page(hpage);1476	return SCAN_SUCCEED;1477}1478 1479/**1480 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at1481 * address haddr.1482 *1483 * @mm: process address space where collapse happens1484 * @addr: THP collapse address1485 * @install_pmd: If a huge PMD should be installed1486 *1487 * This function checks whether all the PTEs in the PMD are pointing to the1488 * right THP. If so, retract the page table so the THP can refault in with1489 * as pmd-mapped. Possibly install a huge PMD mapping the THP.1490 */1491int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,1492			    bool install_pmd)1493{1494	struct mmu_notifier_range range;1495	bool notified = false;1496	unsigned long haddr = addr & HPAGE_PMD_MASK;1497	struct vm_area_struct *vma = vma_lookup(mm, haddr);1498	struct folio *folio;1499	pte_t *start_pte, *pte;1500	pmd_t *pmd, pgt_pmd;1501	spinlock_t *pml = NULL, *ptl;1502	int nr_ptes = 0, result = SCAN_FAIL;1503	int i;1504 1505	mmap_assert_locked(mm);1506 1507	/* First check VMA found, in case page tables are being torn down */1508	if (!vma || !vma->vm_file ||1509	    !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))1510		return SCAN_VMA_CHECK;1511 1512	/* Fast check before locking page if already PMD-mapped */1513	result = find_pmd_or_thp_or_none(mm, haddr, &pmd);1514	if (result == SCAN_PMD_MAPPED)1515		return result;1516 1517	/*1518	 * If we are here, we've succeeded in replacing all the native pages1519	 * in the page cache with a single hugepage. If a mm were to fault-in1520	 * this memory (mapped by a suitably aligned VMA), we'd get the hugepage1521	 * and map it by a PMD, regardless of sysfs THP settings. As such, let's1522	 * analogously elide sysfs THP settings here.1523	 */1524	if (!thp_vma_allowable_order(vma, vma->vm_flags, 0, PMD_ORDER))1525		return SCAN_VMA_CHECK;1526 1527	/* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */1528	if (userfaultfd_wp(vma))1529		return SCAN_PTE_UFFD_WP;1530 1531	folio = filemap_lock_folio(vma->vm_file->f_mapping,1532			       linear_page_index(vma, haddr));1533	if (IS_ERR(folio))1534		return SCAN_PAGE_NULL;1535 1536	if (folio_order(folio) != HPAGE_PMD_ORDER) {1537		result = SCAN_PAGE_COMPOUND;1538		goto drop_folio;1539	}1540 1541	result = find_pmd_or_thp_or_none(mm, haddr, &pmd);1542	switch (result) {1543	case SCAN_SUCCEED:1544		break;1545	case SCAN_PMD_NONE:1546		/*1547		 * All pte entries have been removed and pmd cleared.1548		 * Skip all the pte checks and just update the pmd mapping.1549		 */1550		goto maybe_install_pmd;1551	default:1552		goto drop_folio;1553	}1554 1555	result = SCAN_FAIL;1556	start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);1557	if (!start_pte)		/* mmap_lock + page lock should prevent this */1558		goto drop_folio;1559 1560	/* step 1: check all mapped PTEs are to the right huge page */1561	for (i = 0, addr = haddr, pte = start_pte;1562	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {1563		struct page *page;1564		pte_t ptent = ptep_get(pte);1565 1566		/* empty pte, skip */1567		if (pte_none(ptent))1568			continue;1569 1570		/* page swapped out, abort */1571		if (!pte_present(ptent)) {1572			result = SCAN_PTE_NON_PRESENT;1573			goto abort;1574		}1575 1576		page = vm_normal_page(vma, addr, ptent);1577		if (WARN_ON_ONCE(page && is_zone_device_page(page)))1578			page = NULL;1579		/*1580		 * Note that uprobe, debugger, or MAP_PRIVATE may change the1581		 * page table, but the new page will not be a subpage of hpage.1582		 */1583		if (folio_page(folio, i) != page)1584			goto abort;1585	}1586 1587	pte_unmap_unlock(start_pte, ptl);1588	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,1589				haddr, haddr + HPAGE_PMD_SIZE);1590	mmu_notifier_invalidate_range_start(&range);1591	notified = true;1592 1593	/*1594	 * pmd_lock covers a wider range than ptl, and (if split from mm's1595	 * page_table_lock) ptl nests inside pml. The less time we hold pml,1596	 * the better; but userfaultfd's mfill_atomic_pte() on a private VMA1597	 * inserts a valid as-if-COWed PTE without even looking up page cache.1598	 * So page lock of folio does not protect from it, so we must not drop1599	 * ptl before pgt_pmd is removed, so uffd private needs pml taken now.1600	 */1601	if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED))1602		pml = pmd_lock(mm, pmd);1603 1604	start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl);1605	if (!start_pte)		/* mmap_lock + page lock should prevent this */1606		goto abort;1607	if (!pml)1608		spin_lock(ptl);1609	else if (ptl != pml)1610		spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);1611 1612	/* step 2: clear page table and adjust rmap */1613	for (i = 0, addr = haddr, pte = start_pte;1614	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {1615		struct page *page;1616		pte_t ptent = ptep_get(pte);1617 1618		if (pte_none(ptent))1619			continue;1620		/*1621		 * We dropped ptl after the first scan, to do the mmu_notifier:1622		 * page lock stops more PTEs of the folio being faulted in, but1623		 * does not stop write faults COWing anon copies from existing1624		 * PTEs; and does not stop those being swapped out or migrated.1625		 */1626		if (!pte_present(ptent)) {1627			result = SCAN_PTE_NON_PRESENT;1628			goto abort;1629		}1630		page = vm_normal_page(vma, addr, ptent);1631		if (folio_page(folio, i) != page)1632			goto abort;1633 1634		/*1635		 * Must clear entry, or a racing truncate may re-remove it.1636		 * TLB flush can be left until pmdp_collapse_flush() does it.1637		 * PTE dirty? Shmem page is already dirty; file is read-only.1638		 */1639		ptep_clear(mm, addr, pte);1640		folio_remove_rmap_pte(folio, page, vma);1641		nr_ptes++;1642	}1643 1644	pte_unmap(start_pte);1645	if (!pml)1646		spin_unlock(ptl);1647 1648	/* step 3: set proper refcount and mm_counters. */1649	if (nr_ptes) {1650		folio_ref_sub(folio, nr_ptes);1651		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);1652	}1653 1654	/* step 4: remove empty page table */1655	if (!pml) {1656		pml = pmd_lock(mm, pmd);1657		if (ptl != pml)1658			spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);1659	}1660	pgt_pmd = pmdp_collapse_flush(vma, haddr, pmd);1661	pmdp_get_lockless_sync();1662	if (ptl != pml)1663		spin_unlock(ptl);1664	spin_unlock(pml);1665 1666	mmu_notifier_invalidate_range_end(&range);1667 1668	mm_dec_nr_ptes(mm);1669	page_table_check_pte_clear_range(mm, haddr, pgt_pmd);1670	pte_free_defer(mm, pmd_pgtable(pgt_pmd));1671 1672maybe_install_pmd:1673	/* step 5: install pmd entry */1674	result = install_pmd1675			? set_huge_pmd(vma, haddr, pmd, &folio->page)1676			: SCAN_SUCCEED;1677	goto drop_folio;1678abort:1679	if (nr_ptes) {1680		flush_tlb_mm(mm);1681		folio_ref_sub(folio, nr_ptes);1682		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);1683	}1684	if (start_pte)1685		pte_unmap_unlock(start_pte, ptl);1686	if (pml && pml != ptl)1687		spin_unlock(pml);1688	if (notified)1689		mmu_notifier_invalidate_range_end(&range);1690drop_folio:1691	folio_unlock(folio);1692	folio_put(folio);1693	return result;1694}1695 1696static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)1697{1698	struct vm_area_struct *vma;1699 1700	i_mmap_lock_read(mapping);1701	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {1702		struct mmu_notifier_range range;1703		struct mm_struct *mm;1704		unsigned long addr;1705		pmd_t *pmd, pgt_pmd;1706		spinlock_t *pml;1707		spinlock_t *ptl;1708		bool skipped_uffd = false;1709 1710		/*1711		 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that1712		 * got written to. These VMAs are likely not worth removing1713		 * page tables from, as PMD-mapping is likely to be split later.1714		 */1715		if (READ_ONCE(vma->anon_vma))1716			continue;1717 1718		addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);1719		if (addr & ~HPAGE_PMD_MASK ||1720		    vma->vm_end < addr + HPAGE_PMD_SIZE)1721			continue;1722 1723		mm = vma->vm_mm;1724		if (find_pmd_or_thp_or_none(mm, addr, &pmd) != SCAN_SUCCEED)1725			continue;1726 1727		if (hpage_collapse_test_exit(mm))1728			continue;1729		/*1730		 * When a vma is registered with uffd-wp, we cannot recycle1731		 * the page table because there may be pte markers installed.1732		 * Other vmas can still have the same file mapped hugely, but1733		 * skip this one: it will always be mapped in small page size1734		 * for uffd-wp registered ranges.1735		 */1736		if (userfaultfd_wp(vma))1737			continue;1738 1739		/* PTEs were notified when unmapped; but now for the PMD? */1740		mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,1741					addr, addr + HPAGE_PMD_SIZE);1742		mmu_notifier_invalidate_range_start(&range);1743 1744		pml = pmd_lock(mm, pmd);1745		ptl = pte_lockptr(mm, pmd);1746		if (ptl != pml)1747			spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);1748 1749		/*1750		 * Huge page lock is still held, so normally the page table1751		 * must remain empty; and we have already skipped anon_vma1752		 * and userfaultfd_wp() vmas.  But since the mmap_lock is not1753		 * held, it is still possible for a racing userfaultfd_ioctl()1754		 * to have inserted ptes or markers.  Now that we hold ptlock,1755		 * repeating the anon_vma check protects from one category,1756		 * and repeating the userfaultfd_wp() check from another.1757		 */1758		if (unlikely(vma->anon_vma || userfaultfd_wp(vma))) {1759			skipped_uffd = true;1760		} else {1761			pgt_pmd = pmdp_collapse_flush(vma, addr, pmd);1762			pmdp_get_lockless_sync();1763		}1764 1765		if (ptl != pml)1766			spin_unlock(ptl);1767		spin_unlock(pml);1768 1769		mmu_notifier_invalidate_range_end(&range);1770 1771		if (!skipped_uffd) {1772			mm_dec_nr_ptes(mm);1773			page_table_check_pte_clear_range(mm, addr, pgt_pmd);1774			pte_free_defer(mm, pmd_pgtable(pgt_pmd));1775		}1776	}1777	i_mmap_unlock_read(mapping);1778}1779 1780/**1781 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.1782 *1783 * @mm: process address space where collapse happens1784 * @addr: virtual collapse start address1785 * @file: file that collapse on1786 * @start: collapse start address1787 * @cc: collapse context and scratchpad1788 *1789 * Basic scheme is simple, details are more complex:1790 *  - allocate and lock a new huge page;1791 *  - scan page cache, locking old pages1792 *    + swap/gup in pages if necessary;1793 *  - copy data to new page1794 *  - handle shmem holes1795 *    + re-validate that holes weren't filled by someone else1796 *    + check for userfaultfd1797 *  - finalize updates to the page cache;1798 *  - if replacing succeeds:1799 *    + unlock huge page;1800 *    + free old pages;1801 *  - if replacing failed;1802 *    + unlock old pages1803 *    + unlock and free huge page;1804 */1805static int collapse_file(struct mm_struct *mm, unsigned long addr,1806			 struct file *file, pgoff_t start,1807			 struct collapse_control *cc)1808{1809	struct address_space *mapping = file->f_mapping;1810	struct page *dst;1811	struct folio *folio, *tmp, *new_folio;1812	pgoff_t index = 0, end = start + HPAGE_PMD_NR;1813	LIST_HEAD(pagelist);1814	XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);1815	int nr_none = 0, result = SCAN_SUCCEED;1816	bool is_shmem = shmem_file(file);1817 1818	VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);1819	VM_BUG_ON(start & (HPAGE_PMD_NR - 1));1820 1821	result = alloc_charge_folio(&new_folio, mm, cc);1822	if (result != SCAN_SUCCEED)1823		goto out;1824 1825	__folio_set_locked(new_folio);1826	if (is_shmem)1827		__folio_set_swapbacked(new_folio);1828	new_folio->index = start;1829	new_folio->mapping = mapping;1830 1831	/*1832	 * Ensure we have slots for all the pages in the range.  This is1833	 * almost certainly a no-op because most of the pages must be present1834	 */1835	do {1836		xas_lock_irq(&xas);1837		xas_create_range(&xas);1838		if (!xas_error(&xas))1839			break;1840		xas_unlock_irq(&xas);1841		if (!xas_nomem(&xas, GFP_KERNEL)) {1842			result = SCAN_FAIL;1843			goto rollback;1844		}1845	} while (1);1846 1847	for (index = start; index < end;) {1848		xas_set(&xas, index);1849		folio = xas_load(&xas);1850 1851		VM_BUG_ON(index != xas.xa_index);1852		if (is_shmem) {1853			if (!folio) {1854				/*1855				 * Stop if extent has been truncated or1856				 * hole-punched, and is now completely1857				 * empty.1858				 */1859				if (index == start) {1860					if (!xas_next_entry(&xas, end - 1)) {1861						result = SCAN_TRUNCATED;1862						goto xa_locked;1863					}1864				}1865				nr_none++;1866				index++;1867				continue;1868			}1869 1870			if (xa_is_value(folio) || !folio_test_uptodate(folio)) {1871				xas_unlock_irq(&xas);1872				/* swap in or instantiate fallocated page */1873				if (shmem_get_folio(mapping->host, index, 0,1874						&folio, SGP_NOALLOC)) {1875					result = SCAN_FAIL;1876					goto xa_unlocked;1877				}1878				/* drain lru cache to help folio_isolate_lru() */1879				lru_add_drain();1880			} else if (folio_trylock(folio)) {1881				folio_get(folio);1882				xas_unlock_irq(&xas);1883			} else {1884				result = SCAN_PAGE_LOCK;1885				goto xa_locked;1886			}1887		} else {	/* !is_shmem */1888			if (!folio || xa_is_value(folio)) {1889				xas_unlock_irq(&xas);1890				page_cache_sync_readahead(mapping, &file->f_ra,1891							  file, index,1892							  end - index);1893				/* drain lru cache to help folio_isolate_lru() */1894				lru_add_drain();1895				folio = filemap_lock_folio(mapping, index);1896				if (IS_ERR(folio)) {1897					result = SCAN_FAIL;1898					goto xa_unlocked;1899				}1900			} else if (folio_test_dirty(folio)) {1901				/*1902				 * khugepaged only works on read-only fd,1903				 * so this page is dirty because it hasn't1904				 * been flushed since first write. There1905				 * won't be new dirty pages.1906				 *1907				 * Trigger async flush here and hope the1908				 * writeback is done when khugepaged1909				 * revisits this page.1910				 *1911				 * This is a one-off situation. We are not1912				 * forcing writeback in loop.1913				 */1914				xas_unlock_irq(&xas);1915				filemap_flush(mapping);1916				result = SCAN_FAIL;1917				goto xa_unlocked;1918			} else if (folio_test_writeback(folio)) {1919				xas_unlock_irq(&xas);1920				result = SCAN_FAIL;1921				goto xa_unlocked;1922			} else if (folio_trylock(folio)) {1923				folio_get(folio);1924				xas_unlock_irq(&xas);1925			} else {1926				result = SCAN_PAGE_LOCK;1927				goto xa_locked;1928			}1929		}1930 1931		/*1932		 * The folio must be locked, so we can drop the i_pages lock1933		 * without racing with truncate.1934		 */1935		VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);1936 1937		/* make sure the folio is up to date */1938		if (unlikely(!folio_test_uptodate(folio))) {1939			result = SCAN_FAIL;1940			goto out_unlock;1941		}1942 1943		/*1944		 * If file was truncated then extended, or hole-punched, before1945		 * we locked the first folio, then a THP might be there already.1946		 * This will be discovered on the first iteration.1947		 */1948		if (folio_order(folio) == HPAGE_PMD_ORDER &&1949		    folio->index == start) {1950			/* Maybe PMD-mapped */1951			result = SCAN_PTE_MAPPED_HUGEPAGE;1952			goto out_unlock;1953		}1954 1955		if (folio_mapping(folio) != mapping) {1956			result = SCAN_TRUNCATED;1957			goto out_unlock;1958		}1959 1960		if (!is_shmem && (folio_test_dirty(folio) ||1961				  folio_test_writeback(folio))) {1962			/*1963			 * khugepaged only works on read-only fd, so this1964			 * folio is dirty because it hasn't been flushed1965			 * since first write.1966			 */1967			result = SCAN_FAIL;1968			goto out_unlock;1969		}1970 1971		if (!folio_isolate_lru(folio)) {1972			result = SCAN_DEL_PAGE_LRU;1973			goto out_unlock;1974		}1975 1976		if (!filemap_release_folio(folio, GFP_KERNEL)) {1977			result = SCAN_PAGE_HAS_PRIVATE;1978			folio_putback_lru(folio);1979			goto out_unlock;1980		}1981 1982		if (folio_mapped(folio))1983			try_to_unmap(folio,1984					TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH);1985 1986		xas_lock_irq(&xas);1987 1988		VM_BUG_ON_FOLIO(folio != xa_load(xas.xa, index), folio);1989 1990		/*1991		 * We control 2 + nr_pages references to the folio:1992		 *  - we hold a pin on it;1993		 *  - nr_pages reference from page cache;1994		 *  - one from lru_isolate_folio;1995		 * If those are the only references, then any new usage1996		 * of the folio will have to fetch it from the page1997		 * cache. That requires locking the folio to handle1998		 * truncate, so any new usage will be blocked until we1999		 * unlock folio after collapse/during rollback.2000		 */2001		if (folio_ref_count(folio) != 2 + folio_nr_pages(folio)) {2002			result = SCAN_PAGE_COUNT;2003			xas_unlock_irq(&xas);2004			folio_putback_lru(folio);2005			goto out_unlock;2006		}2007 2008		/*2009		 * Accumulate the folios that are being collapsed.2010		 */2011		list_add_tail(&folio->lru, &pagelist);2012		index += folio_nr_pages(folio);2013		continue;2014out_unlock:2015		folio_unlock(folio);2016		folio_put(folio);2017		goto xa_unlocked;2018	}2019 2020	if (!is_shmem) {2021		filemap_nr_thps_inc(mapping);2022		/*2023		 * Paired with the fence in do_dentry_open() -> get_write_access()2024		 * to ensure i_writecount is up to date and the update to nr_thps2025		 * is visible. Ensures the page cache will be truncated if the2026		 * file is opened writable.2027		 */2028		smp_mb();2029		if (inode_is_open_for_write(mapping->host)) {2030			result = SCAN_FAIL;2031			filemap_nr_thps_dec(mapping);2032		}2033	}2034 2035xa_locked:2036	xas_unlock_irq(&xas);2037xa_unlocked:2038 2039	/*2040	 * If collapse is successful, flush must be done now before copying.2041	 * If collapse is unsuccessful, does flush actually need to be done?2042	 * Do it anyway, to clear the state.2043	 */2044	try_to_unmap_flush();2045 2046	if (result == SCAN_SUCCEED && nr_none &&2047	    !shmem_charge(mapping->host, nr_none))2048		result = SCAN_FAIL;2049	if (result != SCAN_SUCCEED) {2050		nr_none = 0;2051		goto rollback;2052	}2053 2054	/*2055	 * The old folios are locked, so they won't change anymore.2056	 */2057	index = start;2058	dst = folio_page(new_folio, 0);2059	list_for_each_entry(folio, &pagelist, lru) {2060		int i, nr_pages = folio_nr_pages(folio);2061 2062		while (index < folio->index) {2063			clear_highpage(dst);2064			index++;2065			dst++;2066		}2067 2068		for (i = 0; i < nr_pages; i++) {2069			if (copy_mc_highpage(dst, folio_page(folio, i)) > 0) {2070				result = SCAN_COPY_MC;2071				goto rollback;2072			}2073			index++;2074			dst++;2075		}2076	}2077	while (index < end) {2078		clear_highpage(dst);2079		index++;2080		dst++;2081	}2082 2083	if (nr_none) {2084		struct vm_area_struct *vma;2085		int nr_none_check = 0;2086 2087		i_mmap_lock_read(mapping);2088		xas_lock_irq(&xas);2089 2090		xas_set(&xas, start);2091		for (index = start; index < end; index++) {2092			if (!xas_next(&xas)) {2093				xas_store(&xas, XA_RETRY_ENTRY);2094				if (xas_error(&xas)) {2095					result = SCAN_STORE_FAILED;2096					goto immap_locked;2097				}2098				nr_none_check++;2099			}2100		}2101 2102		if (nr_none != nr_none_check) {2103			result = SCAN_PAGE_FILLED;2104			goto immap_locked;2105		}2106 2107		/*2108		 * If userspace observed a missing page in a VMA with2109		 * a MODE_MISSING userfaultfd, then it might expect a2110		 * UFFD_EVENT_PAGEFAULT for that page. If so, we need to2111		 * roll back to avoid suppressing such an event. Since2112		 * wp/minor userfaultfds don't give userspace any2113		 * guarantees that the kernel doesn't fill a missing2114		 * page with a zero page, so they don't matter here.2115		 *2116		 * Any userfaultfds registered after this point will2117		 * not be able to observe any missing pages due to the2118		 * previously inserted retry entries.2119		 */2120		vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) {2121			if (userfaultfd_missing(vma)) {2122				result = SCAN_EXCEED_NONE_PTE;2123				goto immap_locked;2124			}2125		}2126 2127immap_locked:2128		i_mmap_unlock_read(mapping);2129		if (result != SCAN_SUCCEED) {2130			xas_set(&xas, start);2131			for (index = start; index < end; index++) {2132				if (xas_next(&xas) == XA_RETRY_ENTRY)2133					xas_store(&xas, NULL);2134			}2135 2136			xas_unlock_irq(&xas);2137			goto rollback;2138		}2139	} else {2140		xas_lock_irq(&xas);2141	}2142 2143	if (is_shmem)2144		__lruvec_stat_mod_folio(new_folio, NR_SHMEM_THPS, HPAGE_PMD_NR);2145	else2146		__lruvec_stat_mod_folio(new_folio, NR_FILE_THPS, HPAGE_PMD_NR);2147 2148	if (nr_none) {2149		__lruvec_stat_mod_folio(new_folio, NR_FILE_PAGES, nr_none);2150		/* nr_none is always 0 for non-shmem. */2151		__lruvec_stat_mod_folio(new_folio, NR_SHMEM, nr_none);2152	}2153 2154	/*2155	 * Mark new_folio as uptodate before inserting it into the2156	 * page cache so that it isn't mistaken for an fallocated but2157	 * unwritten page.2158	 */2159	folio_mark_uptodate(new_folio);2160	folio_ref_add(new_folio, HPAGE_PMD_NR - 1);2161 2162	if (is_shmem)2163		folio_mark_dirty(new_folio);2164	folio_add_lru(new_folio);2165 2166	/* Join all the small entries into a single multi-index entry. */2167	xas_set_order(&xas, start, HPAGE_PMD_ORDER);2168	xas_store(&xas, new_folio);2169	WARN_ON_ONCE(xas_error(&xas));2170	xas_unlock_irq(&xas);2171 2172	/*2173	 * Remove pte page tables, so we can re-fault the page as huge.2174	 * If MADV_COLLAPSE, adjust result to call collapse_pte_mapped_thp().2175	 */2176	retract_page_tables(mapping, start);2177	if (cc && !cc->is_khugepaged)2178		result = SCAN_PTE_MAPPED_HUGEPAGE;2179	folio_unlock(new_folio);2180 2181	/*2182	 * The collapse has succeeded, so free the old folios.2183	 */2184	list_for_each_entry_safe(folio, tmp, &pagelist, lru) {2185		list_del(&folio->lru);2186		folio->mapping = NULL;2187		folio_clear_active(folio);2188		folio_clear_unevictable(folio);2189		folio_unlock(folio);2190		folio_put_refs(folio, 2 + folio_nr_pages(folio));2191	}2192 2193	goto out;2194 2195rollback:2196	/* Something went wrong: roll back page cache changes */2197	if (nr_none) {2198		xas_lock_irq(&xas);2199		mapping->nrpages -= nr_none;2200		xas_unlock_irq(&xas);2201		shmem_uncharge(mapping->host, nr_none);2202	}2203 2204	list_for_each_entry_safe(folio, tmp, &pagelist, lru) {2205		list_del(&folio->lru);2206		folio_unlock(folio);2207		folio_putback_lru(folio);2208		folio_put(folio);2209	}2210	/*2211	 * Undo the updates of filemap_nr_thps_inc for non-SHMEM2212	 * file only. This undo is not needed unless failure is2213	 * due to SCAN_COPY_MC.2214	 */2215	if (!is_shmem && result == SCAN_COPY_MC) {2216		filemap_nr_thps_dec(mapping);2217		/*2218		 * Paired with the fence in do_dentry_open() -> get_write_access()2219		 * to ensure the update to nr_thps is visible.2220		 */2221		smp_mb();2222	}2223 2224	new_folio->mapping = NULL;2225 2226	folio_unlock(new_folio);2227	folio_put(new_folio);2228out:2229	VM_BUG_ON(!list_empty(&pagelist));2230	trace_mm_khugepaged_collapse_file(mm, new_folio, index, addr, is_shmem, file, HPAGE_PMD_NR, result);2231	return result;2232}2233 2234static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,2235				    struct file *file, pgoff_t start,2236				    struct collapse_control *cc)2237{2238	struct folio *folio = NULL;2239	struct address_space *mapping = file->f_mapping;2240	XA_STATE(xas, &mapping->i_pages, start);2241	int present, swap;2242	int node = NUMA_NO_NODE;2243	int result = SCAN_SUCCEED;2244 2245	present = 0;2246	swap = 0;2247	memset(cc->node_load, 0, sizeof(cc->node_load));2248	nodes_clear(cc->alloc_nmask);2249	rcu_read_lock();2250	xas_for_each(&xas, folio, start + HPAGE_PMD_NR - 1) {2251		if (xas_retry(&xas, folio))2252			continue;2253 2254		if (xa_is_value(folio)) {2255			swap += 1 << xas_get_order(&xas);2256			if (cc->is_khugepaged &&2257			    swap > khugepaged_max_ptes_swap) {2258				result = SCAN_EXCEED_SWAP_PTE;2259				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);2260				break;2261			}2262			continue;2263		}2264 2265		if (folio_order(folio) == HPAGE_PMD_ORDER &&2266		    folio->index == start) {2267			/* Maybe PMD-mapped */2268			result = SCAN_PTE_MAPPED_HUGEPAGE;2269			/*2270			 * For SCAN_PTE_MAPPED_HUGEPAGE, further processing2271			 * by the caller won't touch the page cache, and so2272			 * it's safe to skip LRU and refcount checks before2273			 * returning.2274			 */2275			break;2276		}2277 2278		node = folio_nid(folio);2279		if (hpage_collapse_scan_abort(node, cc)) {2280			result = SCAN_SCAN_ABORT;2281			break;2282		}2283		cc->node_load[node]++;2284 2285		if (!folio_test_lru(folio)) {2286			result = SCAN_PAGE_LRU;2287			break;2288		}2289 2290		if (!is_refcount_suitable(folio)) {2291			result = SCAN_PAGE_COUNT;2292			break;2293		}2294 2295		/*2296		 * We probably should check if the folio is referenced2297		 * here, but nobody would transfer pte_young() to2298		 * folio_test_referenced() for us.  And rmap walk here2299		 * is just too costly...2300		 */2301 2302		present += folio_nr_pages(folio);2303 2304		if (need_resched()) {2305			xas_pause(&xas);2306			cond_resched_rcu();2307		}2308	}2309	rcu_read_unlock();2310 2311	if (result == SCAN_SUCCEED) {2312		if (cc->is_khugepaged &&2313		    present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {2314			result = SCAN_EXCEED_NONE_PTE;2315			count_vm_event(THP_SCAN_EXCEED_NONE_PTE);2316		} else {2317			result = collapse_file(mm, addr, file, start, cc);2318		}2319	}2320 2321	trace_mm_khugepaged_scan_file(mm, folio, file, present, swap, result);2322	return result;2323}2324#else2325static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr,2326				    struct file *file, pgoff_t start,2327				    struct collapse_control *cc)2328{2329	BUILD_BUG();2330}2331#endif2332 2333static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result,2334					    struct collapse_control *cc)2335	__releases(&khugepaged_mm_lock)2336	__acquires(&khugepaged_mm_lock)2337{2338	struct vma_iterator vmi;2339	struct khugepaged_mm_slot *mm_slot;2340	struct mm_slot *slot;2341	struct mm_struct *mm;2342	struct vm_area_struct *vma;2343	int progress = 0;2344 2345	VM_BUG_ON(!pages);2346	lockdep_assert_held(&khugepaged_mm_lock);2347	*result = SCAN_FAIL;2348 2349	if (khugepaged_scan.mm_slot) {2350		mm_slot = khugepaged_scan.mm_slot;2351		slot = &mm_slot->slot;2352	} else {2353		slot = list_entry(khugepaged_scan.mm_head.next,2354				     struct mm_slot, mm_node);2355		mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot);2356		khugepaged_scan.address = 0;2357		khugepaged_scan.mm_slot = mm_slot;2358	}2359	spin_unlock(&khugepaged_mm_lock);2360 2361	mm = slot->mm;2362	/*2363	 * Don't wait for semaphore (to avoid long wait times).  Just move to2364	 * the next mm on the list.2365	 */2366	vma = NULL;2367	if (unlikely(!mmap_read_trylock(mm)))2368		goto breakouterloop_mmap_lock;2369 2370	progress++;2371	if (unlikely(hpage_collapse_test_exit_or_disable(mm)))2372		goto breakouterloop;2373 2374	vma_iter_init(&vmi, mm, khugepaged_scan.address);2375	for_each_vma(vmi, vma) {2376		unsigned long hstart, hend;2377 2378		cond_resched();2379		if (unlikely(hpage_collapse_test_exit_or_disable(mm))) {2380			progress++;2381			break;2382		}2383		if (!thp_vma_allowable_order(vma, vma->vm_flags,2384					TVA_ENFORCE_SYSFS, PMD_ORDER)) {2385skip:2386			progress++;2387			continue;2388		}2389		hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE);2390		hend = round_down(vma->vm_end, HPAGE_PMD_SIZE);2391		if (khugepaged_scan.address > hend)2392			goto skip;2393		if (khugepaged_scan.address < hstart)2394			khugepaged_scan.address = hstart;2395		VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);2396 2397		while (khugepaged_scan.address < hend) {2398			bool mmap_locked = true;2399 2400			cond_resched();2401			if (unlikely(hpage_collapse_test_exit_or_disable(mm)))2402				goto breakouterloop;2403 2404			VM_BUG_ON(khugepaged_scan.address < hstart ||2405				  khugepaged_scan.address + HPAGE_PMD_SIZE >2406				  hend);2407			if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {2408				struct file *file = get_file(vma->vm_file);2409				pgoff_t pgoff = linear_page_index(vma,2410						khugepaged_scan.address);2411 2412				mmap_read_unlock(mm);2413				mmap_locked = false;2414				*result = hpage_collapse_scan_file(mm,2415					khugepaged_scan.address, file, pgoff, cc);2416				fput(file);2417				if (*result == SCAN_PTE_MAPPED_HUGEPAGE) {2418					mmap_read_lock(mm);2419					if (hpage_collapse_test_exit_or_disable(mm))2420						goto breakouterloop;2421					*result = collapse_pte_mapped_thp(mm,2422						khugepaged_scan.address, false);2423					if (*result == SCAN_PMD_MAPPED)2424						*result = SCAN_SUCCEED;2425					mmap_read_unlock(mm);2426				}2427			} else {2428				*result = hpage_collapse_scan_pmd(mm, vma,2429					khugepaged_scan.address, &mmap_locked, cc);2430			}2431 2432			if (*result == SCAN_SUCCEED)2433				++khugepaged_pages_collapsed;2434 2435			/* move to next address */2436			khugepaged_scan.address += HPAGE_PMD_SIZE;2437			progress += HPAGE_PMD_NR;2438			if (!mmap_locked)2439				/*2440				 * We released mmap_lock so break loop.  Note2441				 * that we drop mmap_lock before all hugepage2442				 * allocations, so if allocation fails, we are2443				 * guaranteed to break here and report the2444				 * correct result back to caller.2445				 */2446				goto breakouterloop_mmap_lock;2447			if (progress >= pages)2448				goto breakouterloop;2449		}2450	}2451breakouterloop:2452	mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */2453breakouterloop_mmap_lock:2454 2455	spin_lock(&khugepaged_mm_lock);2456	VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);2457	/*2458	 * Release the current mm_slot if this mm is about to die, or2459	 * if we scanned all vmas of this mm.2460	 */2461	if (hpage_collapse_test_exit(mm) || !vma) {2462		/*2463		 * Make sure that if mm_users is reaching zero while2464		 * khugepaged runs here, khugepaged_exit will find2465		 * mm_slot not pointing to the exiting mm.2466		 */2467		if (slot->mm_node.next != &khugepaged_scan.mm_head) {2468			slot = list_entry(slot->mm_node.next,2469					  struct mm_slot, mm_node);2470			khugepaged_scan.mm_slot =2471				mm_slot_entry(slot, struct khugepaged_mm_slot, slot);2472			khugepaged_scan.address = 0;2473		} else {2474			khugepaged_scan.mm_slot = NULL;2475			khugepaged_full_scans++;2476		}2477 2478		collect_mm_slot(mm_slot);2479	}2480 2481	return progress;2482}2483 2484static int khugepaged_has_work(void)2485{2486	return !list_empty(&khugepaged_scan.mm_head) && hugepage_pmd_enabled();2487}2488 2489static int khugepaged_wait_event(void)2490{2491	return !list_empty(&khugepaged_scan.mm_head) ||2492		kthread_should_stop();2493}2494 2495static void khugepaged_do_scan(struct collapse_control *cc)2496{2497	unsigned int progress = 0, pass_through_head = 0;2498	unsigned int pages = READ_ONCE(khugepaged_pages_to_scan);2499	bool wait = true;2500	int result = SCAN_SUCCEED;2501 2502	lru_add_drain_all();2503 2504	while (true) {2505		cond_resched();2506 2507		if (unlikely(kthread_should_stop()))2508			break;2509 2510		spin_lock(&khugepaged_mm_lock);2511		if (!khugepaged_scan.mm_slot)2512			pass_through_head++;2513		if (khugepaged_has_work() &&2514		    pass_through_head < 2)2515			progress += khugepaged_scan_mm_slot(pages - progress,2516							    &result, cc);2517		else2518			progress = pages;2519		spin_unlock(&khugepaged_mm_lock);2520 2521		if (progress >= pages)2522			break;2523 2524		if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) {2525			/*2526			 * If fail to allocate the first time, try to sleep for2527			 * a while.  When hit again, cancel the scan.2528			 */2529			if (!wait)2530				break;2531			wait = false;2532			khugepaged_alloc_sleep();2533		}2534	}2535}2536 2537static bool khugepaged_should_wakeup(void)2538{2539	return kthread_should_stop() ||2540	       time_after_eq(jiffies, khugepaged_sleep_expire);2541}2542 2543static void khugepaged_wait_work(void)2544{2545	if (khugepaged_has_work()) {2546		const unsigned long scan_sleep_jiffies =2547			msecs_to_jiffies(khugepaged_scan_sleep_millisecs);2548 2549		if (!scan_sleep_jiffies)2550			return;2551 2552		khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;2553		wait_event_freezable_timeout(khugepaged_wait,2554					     khugepaged_should_wakeup(),2555					     scan_sleep_jiffies);2556		return;2557	}2558 2559	if (hugepage_pmd_enabled())2560		wait_event_freezable(khugepaged_wait, khugepaged_wait_event());2561}2562 2563static int khugepaged(void *none)2564{2565	struct khugepaged_mm_slot *mm_slot;2566 2567	set_freezable();2568	set_user_nice(current, MAX_NICE);2569 2570	while (!kthread_should_stop()) {2571		khugepaged_do_scan(&khugepaged_collapse_control);2572		khugepaged_wait_work();2573	}2574 2575	spin_lock(&khugepaged_mm_lock);2576	mm_slot = khugepaged_scan.mm_slot;2577	khugepaged_scan.mm_slot = NULL;2578	if (mm_slot)2579		collect_mm_slot(mm_slot);2580	spin_unlock(&khugepaged_mm_lock);2581	return 0;2582}2583 2584static void set_recommended_min_free_kbytes(void)2585{2586	struct zone *zone;2587	int nr_zones = 0;2588	unsigned long recommended_min;2589 2590	if (!hugepage_pmd_enabled()) {2591		calculate_min_free_kbytes();2592		goto update_wmarks;2593	}2594 2595	for_each_populated_zone(zone) {2596		/*2597		 * We don't need to worry about fragmentation of2598		 * ZONE_MOVABLE since it only has movable pages.2599		 */2600		if (zone_idx(zone) > gfp_zone(GFP_USER))2601			continue;2602 2603		nr_zones++;2604	}2605 2606	/* Ensure 2 pageblocks are free to assist fragmentation avoidance */2607	recommended_min = pageblock_nr_pages * nr_zones * 2;2608 2609	/*2610	 * Make sure that on average at least two pageblocks are almost free2611	 * of another type, one for a migratetype to fall back to and a2612	 * second to avoid subsequent fallbacks of other types There are 32613	 * MIGRATE_TYPES we care about.2614	 */2615	recommended_min += pageblock_nr_pages * nr_zones *2616			   MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;2617 2618	/* don't ever allow to reserve more than 5% of the lowmem */2619	recommended_min = min(recommended_min,2620			      (unsigned long) nr_free_buffer_pages() / 20);2621	recommended_min <<= (PAGE_SHIFT-10);2622 2623	if (recommended_min > min_free_kbytes) {2624		if (user_min_free_kbytes >= 0)2625			pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",2626				min_free_kbytes, recommended_min);2627 2628		min_free_kbytes = recommended_min;2629	}2630 2631update_wmarks:2632	setup_per_zone_wmarks();2633}2634 2635int start_stop_khugepaged(void)2636{2637	int err = 0;2638 2639	mutex_lock(&khugepaged_mutex);2640	if (hugepage_pmd_enabled()) {2641		if (!khugepaged_thread)2642			khugepaged_thread = kthread_run(khugepaged, NULL,2643							"khugepaged");2644		if (IS_ERR(khugepaged_thread)) {2645			pr_err("khugepaged: kthread_run(khugepaged) failed\n");2646			err = PTR_ERR(khugepaged_thread);2647			khugepaged_thread = NULL;2648			goto fail;2649		}2650 2651		if (!list_empty(&khugepaged_scan.mm_head))2652			wake_up_interruptible(&khugepaged_wait);2653	} else if (khugepaged_thread) {2654		kthread_stop(khugepaged_thread);2655		khugepaged_thread = NULL;2656	}2657	set_recommended_min_free_kbytes();2658fail:2659	mutex_unlock(&khugepaged_mutex);2660	return err;2661}2662 2663void khugepaged_min_free_kbytes_update(void)2664{2665	mutex_lock(&khugepaged_mutex);2666	if (hugepage_pmd_enabled() && khugepaged_thread)2667		set_recommended_min_free_kbytes();2668	mutex_unlock(&khugepaged_mutex);2669}2670 2671bool current_is_khugepaged(void)2672{2673	return kthread_func(current) == khugepaged;2674}2675 2676static int madvise_collapse_errno(enum scan_result r)2677{2678	/*2679	 * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide2680	 * actionable feedback to caller, so they may take an appropriate2681	 * fallback measure depending on the nature of the failure.2682	 */2683	switch (r) {2684	case SCAN_ALLOC_HUGE_PAGE_FAIL:2685		return -ENOMEM;2686	case SCAN_CGROUP_CHARGE_FAIL:2687	case SCAN_EXCEED_NONE_PTE:2688		return -EBUSY;2689	/* Resource temporary unavailable - trying again might succeed */2690	case SCAN_PAGE_COUNT:2691	case SCAN_PAGE_LOCK:2692	case SCAN_PAGE_LRU:2693	case SCAN_DEL_PAGE_LRU:2694	case SCAN_PAGE_FILLED:2695		return -EAGAIN;2696	/*2697	 * Other: Trying again likely not to succeed / error intrinsic to2698	 * specified memory range. khugepaged likely won't be able to collapse2699	 * either.2700	 */2701	default:2702		return -EINVAL;2703	}2704}2705 2706int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev,2707		     unsigned long start, unsigned long end)2708{2709	struct collapse_control *cc;2710	struct mm_struct *mm = vma->vm_mm;2711	unsigned long hstart, hend, addr;2712	int thps = 0, last_fail = SCAN_FAIL;2713	bool mmap_locked = true;2714 2715	BUG_ON(vma->vm_start > start);2716	BUG_ON(vma->vm_end < end);2717 2718	*prev = vma;2719 2720	if (!thp_vma_allowable_order(vma, vma->vm_flags, 0, PMD_ORDER))2721		return -EINVAL;2722 2723	cc = kmalloc(sizeof(*cc), GFP_KERNEL);2724	if (!cc)2725		return -ENOMEM;2726	cc->is_khugepaged = false;2727 2728	mmgrab(mm);2729	lru_add_drain_all();2730 2731	hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;2732	hend = end & HPAGE_PMD_MASK;2733 2734	for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) {2735		int result = SCAN_FAIL;2736 2737		if (!mmap_locked) {2738			cond_resched();2739			mmap_read_lock(mm);2740			mmap_locked = true;2741			result = hugepage_vma_revalidate(mm, addr, false, &vma,2742							 cc);2743			if (result  != SCAN_SUCCEED) {2744				last_fail = result;2745				goto out_nolock;2746			}2747 2748			hend = min(hend, vma->vm_end & HPAGE_PMD_MASK);2749		}2750		mmap_assert_locked(mm);2751		memset(cc->node_load, 0, sizeof(cc->node_load));2752		nodes_clear(cc->alloc_nmask);2753		if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {2754			struct file *file = get_file(vma->vm_file);2755			pgoff_t pgoff = linear_page_index(vma, addr);2756 2757			mmap_read_unlock(mm);2758			mmap_locked = false;2759			result = hpage_collapse_scan_file(mm, addr, file, pgoff,2760							  cc);2761			fput(file);2762		} else {2763			result = hpage_collapse_scan_pmd(mm, vma, addr,2764							 &mmap_locked, cc);2765		}2766		if (!mmap_locked)2767			*prev = NULL;  /* Tell caller we dropped mmap_lock */2768 2769handle_result:2770		switch (result) {2771		case SCAN_SUCCEED:2772		case SCAN_PMD_MAPPED:2773			++thps;2774			break;2775		case SCAN_PTE_MAPPED_HUGEPAGE:2776			BUG_ON(mmap_locked);2777			BUG_ON(*prev);2778			mmap_read_lock(mm);2779			result = collapse_pte_mapped_thp(mm, addr, true);2780			mmap_read_unlock(mm);2781			goto handle_result;2782		/* Whitelisted set of results where continuing OK */2783		case SCAN_PMD_NULL:2784		case SCAN_PTE_NON_PRESENT:2785		case SCAN_PTE_UFFD_WP:2786		case SCAN_PAGE_RO:2787		case SCAN_LACK_REFERENCED_PAGE:2788		case SCAN_PAGE_NULL:2789		case SCAN_PAGE_COUNT:2790		case SCAN_PAGE_LOCK:2791		case SCAN_PAGE_COMPOUND:2792		case SCAN_PAGE_LRU:2793		case SCAN_DEL_PAGE_LRU:2794			last_fail = result;2795			break;2796		default:2797			last_fail = result;2798			/* Other error, exit */2799			goto out_maybelock;2800		}2801	}2802 2803out_maybelock:2804	/* Caller expects us to hold mmap_lock on return */2805	if (!mmap_locked)2806		mmap_read_lock(mm);2807out_nolock:2808	mmap_assert_locked(mm);2809	mmdrop(mm);2810	kfree(cc);2811 2812	return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 02813			: madvise_collapse_errno(last_fail);2814}2815