brintos / linux-shallow public Read only

1// SPDX-License-Identifier: GPL-2.0-only2/*3 *  Copyright (C) 2009  Red Hat, Inc.4 */5 6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt7 8#include <linux/mm.h>9#include <linux/sched.h>10#include <linux/sched/mm.h>11#include <linux/sched/coredump.h>12#include <linux/sched/numa_balancing.h>13#include <linux/highmem.h>14#include <linux/hugetlb.h>15#include <linux/mmu_notifier.h>16#include <linux/rmap.h>17#include <linux/swap.h>18#include <linux/shrinker.h>19#include <linux/mm_inline.h>20#include <linux/swapops.h>21#include <linux/backing-dev.h>22#include <linux/dax.h>23#include <linux/mm_types.h>24#include <linux/khugepaged.h>25#include <linux/freezer.h>26#include <linux/pfn_t.h>27#include <linux/mman.h>28#include <linux/memremap.h>29#include <linux/pagemap.h>30#include <linux/debugfs.h>31#include <linux/migrate.h>32#include <linux/hashtable.h>33#include <linux/userfaultfd_k.h>34#include <linux/page_idle.h>35#include <linux/shmem_fs.h>36#include <linux/oom.h>37#include <linux/numa.h>38#include <linux/page_owner.h>39#include <linux/sched/sysctl.h>40#include <linux/memory-tiers.h>41#include <linux/compat.h>42#include <linux/pgalloc_tag.h>43#include <linux/pagewalk.h>44 45#include <asm/tlb.h>46#include <asm/pgalloc.h>47#include "internal.h"48#include "swap.h"49 50#define CREATE_TRACE_POINTS51#include <trace/events/thp.h>52 53/*54 * By default, transparent hugepage support is disabled in order to avoid55 * risking an increased memory footprint for applications that are not56 * guaranteed to benefit from it. When transparent hugepage support is57 * enabled, it is for all mappings, and khugepaged scans all mappings.58 * Defrag is invoked by khugepaged hugepage allocations and by page faults59 * for all hugepage allocations.60 */61unsigned long transparent_hugepage_flags __read_mostly =62#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS63	(1<<TRANSPARENT_HUGEPAGE_FLAG)|64#endif65#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE66	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|67#endif68	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|69	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|70	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);71 72static struct shrinker *deferred_split_shrinker;73static unsigned long deferred_split_count(struct shrinker *shrink,74					  struct shrink_control *sc);75static unsigned long deferred_split_scan(struct shrinker *shrink,76					 struct shrink_control *sc);77static bool split_underused_thp = true;78 79static atomic_t huge_zero_refcount;80struct folio *huge_zero_folio __read_mostly;81unsigned long huge_zero_pfn __read_mostly = ~0UL;82unsigned long huge_anon_orders_always __read_mostly;83unsigned long huge_anon_orders_madvise __read_mostly;84unsigned long huge_anon_orders_inherit __read_mostly;85static bool anon_orders_configured __initdata;86 87unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,88					 unsigned long vm_flags,89					 unsigned long tva_flags,90					 unsigned long orders)91{92	bool smaps = tva_flags & TVA_SMAPS;93	bool in_pf = tva_flags & TVA_IN_PF;94	bool enforce_sysfs = tva_flags & TVA_ENFORCE_SYSFS;95	unsigned long supported_orders;96 97	/* Check the intersection of requested and supported orders. */98	if (vma_is_anonymous(vma))99		supported_orders = THP_ORDERS_ALL_ANON;100	else if (vma_is_special_huge(vma))101		supported_orders = THP_ORDERS_ALL_SPECIAL;102	else103		supported_orders = THP_ORDERS_ALL_FILE_DEFAULT;104 105	orders &= supported_orders;106	if (!orders)107		return 0;108 109	if (!vma->vm_mm)		/* vdso */110		return 0;111 112	if (thp_disabled_by_hw() || vma_thp_disabled(vma, vm_flags))113		return 0;114 115	/* khugepaged doesn't collapse DAX vma, but page fault is fine. */116	if (vma_is_dax(vma))117		return in_pf ? orders : 0;118 119	/*120	 * khugepaged special VMA and hugetlb VMA.121	 * Must be checked after dax since some dax mappings may have122	 * VM_MIXEDMAP set.123	 */124	if (!in_pf && !smaps && (vm_flags & VM_NO_KHUGEPAGED))125		return 0;126 127	/*128	 * Check alignment for file vma and size for both file and anon vma by129	 * filtering out the unsuitable orders.130	 *131	 * Skip the check for page fault. Huge fault does the check in fault132	 * handlers.133	 */134	if (!in_pf) {135		int order = highest_order(orders);136		unsigned long addr;137 138		while (orders) {139			addr = vma->vm_end - (PAGE_SIZE << order);140			if (thp_vma_suitable_order(vma, addr, order))141				break;142			order = next_order(&orders, order);143		}144 145		if (!orders)146			return 0;147	}148 149	/*150	 * Enabled via shmem mount options or sysfs settings.151	 * Must be done before hugepage flags check since shmem has its152	 * own flags.153	 */154	if (!in_pf && shmem_file(vma->vm_file))155		return shmem_allowable_huge_orders(file_inode(vma->vm_file),156						   vma, vma->vm_pgoff, 0,157						   !enforce_sysfs);158 159	if (!vma_is_anonymous(vma)) {160		/*161		 * Enforce sysfs THP requirements as necessary. Anonymous vmas162		 * were already handled in thp_vma_allowable_orders().163		 */164		if (enforce_sysfs &&165		    (!hugepage_global_enabled() || (!(vm_flags & VM_HUGEPAGE) &&166						    !hugepage_global_always())))167			return 0;168 169		/*170		 * Trust that ->huge_fault() handlers know what they are doing171		 * in fault path.172		 */173		if (((in_pf || smaps)) && vma->vm_ops->huge_fault)174			return orders;175		/* Only regular file is valid in collapse path */176		if (((!in_pf || smaps)) && file_thp_enabled(vma))177			return orders;178		return 0;179	}180 181	if (vma_is_temporary_stack(vma))182		return 0;183 184	/*185	 * THPeligible bit of smaps should show 1 for proper VMAs even186	 * though anon_vma is not initialized yet.187	 *188	 * Allow page fault since anon_vma may be not initialized until189	 * the first page fault.190	 */191	if (!vma->anon_vma)192		return (smaps || in_pf) ? orders : 0;193 194	return orders;195}196 197static bool get_huge_zero_page(void)198{199	struct folio *zero_folio;200retry:201	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))202		return true;203 204	zero_folio = folio_alloc((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,205			HPAGE_PMD_ORDER);206	if (!zero_folio) {207		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);208		return false;209	}210	/* Ensure zero folio won't have large_rmappable flag set. */211	folio_clear_large_rmappable(zero_folio);212	preempt_disable();213	if (cmpxchg(&huge_zero_folio, NULL, zero_folio)) {214		preempt_enable();215		folio_put(zero_folio);216		goto retry;217	}218	WRITE_ONCE(huge_zero_pfn, folio_pfn(zero_folio));219 220	/* We take additional reference here. It will be put back by shrinker */221	atomic_set(&huge_zero_refcount, 2);222	preempt_enable();223	count_vm_event(THP_ZERO_PAGE_ALLOC);224	return true;225}226 227static void put_huge_zero_page(void)228{229	/*230	 * Counter should never go to zero here. Only shrinker can put231	 * last reference.232	 */233	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));234}235 236struct folio *mm_get_huge_zero_folio(struct mm_struct *mm)237{238	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))239		return READ_ONCE(huge_zero_folio);240 241	if (!get_huge_zero_page())242		return NULL;243 244	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))245		put_huge_zero_page();246 247	return READ_ONCE(huge_zero_folio);248}249 250void mm_put_huge_zero_folio(struct mm_struct *mm)251{252	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))253		put_huge_zero_page();254}255 256static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,257					struct shrink_control *sc)258{259	/* we can free zero page only if last reference remains */260	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;261}262 263static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,264				       struct shrink_control *sc)265{266	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {267		struct folio *zero_folio = xchg(&huge_zero_folio, NULL);268		BUG_ON(zero_folio == NULL);269		WRITE_ONCE(huge_zero_pfn, ~0UL);270		folio_put(zero_folio);271		return HPAGE_PMD_NR;272	}273 274	return 0;275}276 277static struct shrinker *huge_zero_page_shrinker;278 279#ifdef CONFIG_SYSFS280static ssize_t enabled_show(struct kobject *kobj,281			    struct kobj_attribute *attr, char *buf)282{283	const char *output;284 285	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))286		output = "[always] madvise never";287	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,288			  &transparent_hugepage_flags))289		output = "always [madvise] never";290	else291		output = "always madvise [never]";292 293	return sysfs_emit(buf, "%s\n", output);294}295 296static ssize_t enabled_store(struct kobject *kobj,297			     struct kobj_attribute *attr,298			     const char *buf, size_t count)299{300	ssize_t ret = count;301 302	if (sysfs_streq(buf, "always")) {303		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);304		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);305	} else if (sysfs_streq(buf, "madvise")) {306		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);307		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);308	} else if (sysfs_streq(buf, "never")) {309		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);310		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);311	} else312		ret = -EINVAL;313 314	if (ret > 0) {315		int err = start_stop_khugepaged();316		if (err)317			ret = err;318	}319	return ret;320}321 322static struct kobj_attribute enabled_attr = __ATTR_RW(enabled);323 324ssize_t single_hugepage_flag_show(struct kobject *kobj,325				  struct kobj_attribute *attr, char *buf,326				  enum transparent_hugepage_flag flag)327{328	return sysfs_emit(buf, "%d\n",329			  !!test_bit(flag, &transparent_hugepage_flags));330}331 332ssize_t single_hugepage_flag_store(struct kobject *kobj,333				 struct kobj_attribute *attr,334				 const char *buf, size_t count,335				 enum transparent_hugepage_flag flag)336{337	unsigned long value;338	int ret;339 340	ret = kstrtoul(buf, 10, &value);341	if (ret < 0)342		return ret;343	if (value > 1)344		return -EINVAL;345 346	if (value)347		set_bit(flag, &transparent_hugepage_flags);348	else349		clear_bit(flag, &transparent_hugepage_flags);350 351	return count;352}353 354static ssize_t defrag_show(struct kobject *kobj,355			   struct kobj_attribute *attr, char *buf)356{357	const char *output;358 359	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,360		     &transparent_hugepage_flags))361		output = "[always] defer defer+madvise madvise never";362	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,363			  &transparent_hugepage_flags))364		output = "always [defer] defer+madvise madvise never";365	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,366			  &transparent_hugepage_flags))367		output = "always defer [defer+madvise] madvise never";368	else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,369			  &transparent_hugepage_flags))370		output = "always defer defer+madvise [madvise] never";371	else372		output = "always defer defer+madvise madvise [never]";373 374	return sysfs_emit(buf, "%s\n", output);375}376 377static ssize_t defrag_store(struct kobject *kobj,378			    struct kobj_attribute *attr,379			    const char *buf, size_t count)380{381	if (sysfs_streq(buf, "always")) {382		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);383		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);384		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);385		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);386	} else if (sysfs_streq(buf, "defer+madvise")) {387		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);388		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);389		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);390		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);391	} else if (sysfs_streq(buf, "defer")) {392		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);393		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);394		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);395		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);396	} else if (sysfs_streq(buf, "madvise")) {397		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);398		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);399		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);400		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);401	} else if (sysfs_streq(buf, "never")) {402		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);403		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);404		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);405		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);406	} else407		return -EINVAL;408 409	return count;410}411static struct kobj_attribute defrag_attr = __ATTR_RW(defrag);412 413static ssize_t use_zero_page_show(struct kobject *kobj,414				  struct kobj_attribute *attr, char *buf)415{416	return single_hugepage_flag_show(kobj, attr, buf,417					 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);418}419static ssize_t use_zero_page_store(struct kobject *kobj,420		struct kobj_attribute *attr, const char *buf, size_t count)421{422	return single_hugepage_flag_store(kobj, attr, buf, count,423				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);424}425static struct kobj_attribute use_zero_page_attr = __ATTR_RW(use_zero_page);426 427static ssize_t hpage_pmd_size_show(struct kobject *kobj,428				   struct kobj_attribute *attr, char *buf)429{430	return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);431}432static struct kobj_attribute hpage_pmd_size_attr =433	__ATTR_RO(hpage_pmd_size);434 435static ssize_t split_underused_thp_show(struct kobject *kobj,436			    struct kobj_attribute *attr, char *buf)437{438	return sysfs_emit(buf, "%d\n", split_underused_thp);439}440 441static ssize_t split_underused_thp_store(struct kobject *kobj,442			     struct kobj_attribute *attr,443			     const char *buf, size_t count)444{445	int err = kstrtobool(buf, &split_underused_thp);446 447	if (err < 0)448		return err;449 450	return count;451}452 453static struct kobj_attribute split_underused_thp_attr = __ATTR(454	shrink_underused, 0644, split_underused_thp_show, split_underused_thp_store);455 456static struct attribute *hugepage_attr[] = {457	&enabled_attr.attr,458	&defrag_attr.attr,459	&use_zero_page_attr.attr,460	&hpage_pmd_size_attr.attr,461#ifdef CONFIG_SHMEM462	&shmem_enabled_attr.attr,463#endif464	&split_underused_thp_attr.attr,465	NULL,466};467 468static const struct attribute_group hugepage_attr_group = {469	.attrs = hugepage_attr,470};471 472static void hugepage_exit_sysfs(struct kobject *hugepage_kobj);473static void thpsize_release(struct kobject *kobj);474static DEFINE_SPINLOCK(huge_anon_orders_lock);475static LIST_HEAD(thpsize_list);476 477static ssize_t anon_enabled_show(struct kobject *kobj,478				 struct kobj_attribute *attr, char *buf)479{480	int order = to_thpsize(kobj)->order;481	const char *output;482 483	if (test_bit(order, &huge_anon_orders_always))484		output = "[always] inherit madvise never";485	else if (test_bit(order, &huge_anon_orders_inherit))486		output = "always [inherit] madvise never";487	else if (test_bit(order, &huge_anon_orders_madvise))488		output = "always inherit [madvise] never";489	else490		output = "always inherit madvise [never]";491 492	return sysfs_emit(buf, "%s\n", output);493}494 495static ssize_t anon_enabled_store(struct kobject *kobj,496				  struct kobj_attribute *attr,497				  const char *buf, size_t count)498{499	int order = to_thpsize(kobj)->order;500	ssize_t ret = count;501 502	if (sysfs_streq(buf, "always")) {503		spin_lock(&huge_anon_orders_lock);504		clear_bit(order, &huge_anon_orders_inherit);505		clear_bit(order, &huge_anon_orders_madvise);506		set_bit(order, &huge_anon_orders_always);507		spin_unlock(&huge_anon_orders_lock);508	} else if (sysfs_streq(buf, "inherit")) {509		spin_lock(&huge_anon_orders_lock);510		clear_bit(order, &huge_anon_orders_always);511		clear_bit(order, &huge_anon_orders_madvise);512		set_bit(order, &huge_anon_orders_inherit);513		spin_unlock(&huge_anon_orders_lock);514	} else if (sysfs_streq(buf, "madvise")) {515		spin_lock(&huge_anon_orders_lock);516		clear_bit(order, &huge_anon_orders_always);517		clear_bit(order, &huge_anon_orders_inherit);518		set_bit(order, &huge_anon_orders_madvise);519		spin_unlock(&huge_anon_orders_lock);520	} else if (sysfs_streq(buf, "never")) {521		spin_lock(&huge_anon_orders_lock);522		clear_bit(order, &huge_anon_orders_always);523		clear_bit(order, &huge_anon_orders_inherit);524		clear_bit(order, &huge_anon_orders_madvise);525		spin_unlock(&huge_anon_orders_lock);526	} else527		ret = -EINVAL;528 529	if (ret > 0) {530		int err;531 532		err = start_stop_khugepaged();533		if (err)534			ret = err;535	}536	return ret;537}538 539static struct kobj_attribute anon_enabled_attr =540	__ATTR(enabled, 0644, anon_enabled_show, anon_enabled_store);541 542static struct attribute *anon_ctrl_attrs[] = {543	&anon_enabled_attr.attr,544	NULL,545};546 547static const struct attribute_group anon_ctrl_attr_grp = {548	.attrs = anon_ctrl_attrs,549};550 551static struct attribute *file_ctrl_attrs[] = {552#ifdef CONFIG_SHMEM553	&thpsize_shmem_enabled_attr.attr,554#endif555	NULL,556};557 558static const struct attribute_group file_ctrl_attr_grp = {559	.attrs = file_ctrl_attrs,560};561 562static struct attribute *any_ctrl_attrs[] = {563	NULL,564};565 566static const struct attribute_group any_ctrl_attr_grp = {567	.attrs = any_ctrl_attrs,568};569 570static const struct kobj_type thpsize_ktype = {571	.release = &thpsize_release,572	.sysfs_ops = &kobj_sysfs_ops,573};574 575DEFINE_PER_CPU(struct mthp_stat, mthp_stats) = {{{0}}};576 577static unsigned long sum_mthp_stat(int order, enum mthp_stat_item item)578{579	unsigned long sum = 0;580	int cpu;581 582	for_each_possible_cpu(cpu) {583		struct mthp_stat *this = &per_cpu(mthp_stats, cpu);584 585		sum += this->stats[order][item];586	}587 588	return sum;589}590 591#define DEFINE_MTHP_STAT_ATTR(_name, _index)				\592static ssize_t _name##_show(struct kobject *kobj,			\593			struct kobj_attribute *attr, char *buf)		\594{									\595	int order = to_thpsize(kobj)->order;				\596									\597	return sysfs_emit(buf, "%lu\n", sum_mthp_stat(order, _index));	\598}									\599static struct kobj_attribute _name##_attr = __ATTR_RO(_name)600 601DEFINE_MTHP_STAT_ATTR(anon_fault_alloc, MTHP_STAT_ANON_FAULT_ALLOC);602DEFINE_MTHP_STAT_ATTR(anon_fault_fallback, MTHP_STAT_ANON_FAULT_FALLBACK);603DEFINE_MTHP_STAT_ATTR(anon_fault_fallback_charge, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);604DEFINE_MTHP_STAT_ATTR(swpout, MTHP_STAT_SWPOUT);605DEFINE_MTHP_STAT_ATTR(swpout_fallback, MTHP_STAT_SWPOUT_FALLBACK);606#ifdef CONFIG_SHMEM607DEFINE_MTHP_STAT_ATTR(shmem_alloc, MTHP_STAT_SHMEM_ALLOC);608DEFINE_MTHP_STAT_ATTR(shmem_fallback, MTHP_STAT_SHMEM_FALLBACK);609DEFINE_MTHP_STAT_ATTR(shmem_fallback_charge, MTHP_STAT_SHMEM_FALLBACK_CHARGE);610#endif611DEFINE_MTHP_STAT_ATTR(split, MTHP_STAT_SPLIT);612DEFINE_MTHP_STAT_ATTR(split_failed, MTHP_STAT_SPLIT_FAILED);613DEFINE_MTHP_STAT_ATTR(split_deferred, MTHP_STAT_SPLIT_DEFERRED);614DEFINE_MTHP_STAT_ATTR(nr_anon, MTHP_STAT_NR_ANON);615DEFINE_MTHP_STAT_ATTR(nr_anon_partially_mapped, MTHP_STAT_NR_ANON_PARTIALLY_MAPPED);616 617static struct attribute *anon_stats_attrs[] = {618	&anon_fault_alloc_attr.attr,619	&anon_fault_fallback_attr.attr,620	&anon_fault_fallback_charge_attr.attr,621#ifndef CONFIG_SHMEM622	&swpout_attr.attr,623	&swpout_fallback_attr.attr,624#endif625	&split_deferred_attr.attr,626	&nr_anon_attr.attr,627	&nr_anon_partially_mapped_attr.attr,628	NULL,629};630 631static struct attribute_group anon_stats_attr_grp = {632	.name = "stats",633	.attrs = anon_stats_attrs,634};635 636static struct attribute *file_stats_attrs[] = {637#ifdef CONFIG_SHMEM638	&shmem_alloc_attr.attr,639	&shmem_fallback_attr.attr,640	&shmem_fallback_charge_attr.attr,641#endif642	NULL,643};644 645static struct attribute_group file_stats_attr_grp = {646	.name = "stats",647	.attrs = file_stats_attrs,648};649 650static struct attribute *any_stats_attrs[] = {651#ifdef CONFIG_SHMEM652	&swpout_attr.attr,653	&swpout_fallback_attr.attr,654#endif655	&split_attr.attr,656	&split_failed_attr.attr,657	NULL,658};659 660static struct attribute_group any_stats_attr_grp = {661	.name = "stats",662	.attrs = any_stats_attrs,663};664 665static int sysfs_add_group(struct kobject *kobj,666			   const struct attribute_group *grp)667{668	int ret = -ENOENT;669 670	/*671	 * If the group is named, try to merge first, assuming the subdirectory672	 * was already created. This avoids the warning emitted by673	 * sysfs_create_group() if the directory already exists.674	 */675	if (grp->name)676		ret = sysfs_merge_group(kobj, grp);677	if (ret)678		ret = sysfs_create_group(kobj, grp);679 680	return ret;681}682 683static struct thpsize *thpsize_create(int order, struct kobject *parent)684{685	unsigned long size = (PAGE_SIZE << order) / SZ_1K;686	struct thpsize *thpsize;687	int ret = -ENOMEM;688 689	thpsize = kzalloc(sizeof(*thpsize), GFP_KERNEL);690	if (!thpsize)691		goto err;692 693	thpsize->order = order;694 695	ret = kobject_init_and_add(&thpsize->kobj, &thpsize_ktype, parent,696				   "hugepages-%lukB", size);697	if (ret) {698		kfree(thpsize);699		goto err;700	}701 702 703	ret = sysfs_add_group(&thpsize->kobj, &any_ctrl_attr_grp);704	if (ret)705		goto err_put;706 707	ret = sysfs_add_group(&thpsize->kobj, &any_stats_attr_grp);708	if (ret)709		goto err_put;710 711	if (BIT(order) & THP_ORDERS_ALL_ANON) {712		ret = sysfs_add_group(&thpsize->kobj, &anon_ctrl_attr_grp);713		if (ret)714			goto err_put;715 716		ret = sysfs_add_group(&thpsize->kobj, &anon_stats_attr_grp);717		if (ret)718			goto err_put;719	}720 721	if (BIT(order) & THP_ORDERS_ALL_FILE_DEFAULT) {722		ret = sysfs_add_group(&thpsize->kobj, &file_ctrl_attr_grp);723		if (ret)724			goto err_put;725 726		ret = sysfs_add_group(&thpsize->kobj, &file_stats_attr_grp);727		if (ret)728			goto err_put;729	}730 731	return thpsize;732err_put:733	kobject_put(&thpsize->kobj);734err:735	return ERR_PTR(ret);736}737 738static void thpsize_release(struct kobject *kobj)739{740	kfree(to_thpsize(kobj));741}742 743static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)744{745	int err;746	struct thpsize *thpsize;747	unsigned long orders;748	int order;749 750	/*751	 * Default to setting PMD-sized THP to inherit the global setting and752	 * disable all other sizes. powerpc's PMD_ORDER isn't a compile-time753	 * constant so we have to do this here.754	 */755	if (!anon_orders_configured)756		huge_anon_orders_inherit = BIT(PMD_ORDER);757 758	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);759	if (unlikely(!*hugepage_kobj)) {760		pr_err("failed to create transparent hugepage kobject\n");761		return -ENOMEM;762	}763 764	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);765	if (err) {766		pr_err("failed to register transparent hugepage group\n");767		goto delete_obj;768	}769 770	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);771	if (err) {772		pr_err("failed to register transparent hugepage group\n");773		goto remove_hp_group;774	}775 776	orders = THP_ORDERS_ALL_ANON | THP_ORDERS_ALL_FILE_DEFAULT;777	order = highest_order(orders);778	while (orders) {779		thpsize = thpsize_create(order, *hugepage_kobj);780		if (IS_ERR(thpsize)) {781			pr_err("failed to create thpsize for order %d\n", order);782			err = PTR_ERR(thpsize);783			goto remove_all;784		}785		list_add(&thpsize->node, &thpsize_list);786		order = next_order(&orders, order);787	}788 789	return 0;790 791remove_all:792	hugepage_exit_sysfs(*hugepage_kobj);793	return err;794remove_hp_group:795	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);796delete_obj:797	kobject_put(*hugepage_kobj);798	return err;799}800 801static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)802{803	struct thpsize *thpsize, *tmp;804 805	list_for_each_entry_safe(thpsize, tmp, &thpsize_list, node) {806		list_del(&thpsize->node);807		kobject_put(&thpsize->kobj);808	}809 810	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);811	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);812	kobject_put(hugepage_kobj);813}814#else815static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)816{817	return 0;818}819 820static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)821{822}823#endif /* CONFIG_SYSFS */824 825static int __init thp_shrinker_init(void)826{827	huge_zero_page_shrinker = shrinker_alloc(0, "thp-zero");828	if (!huge_zero_page_shrinker)829		return -ENOMEM;830 831	deferred_split_shrinker = shrinker_alloc(SHRINKER_NUMA_AWARE |832						 SHRINKER_MEMCG_AWARE |833						 SHRINKER_NONSLAB,834						 "thp-deferred_split");835	if (!deferred_split_shrinker) {836		shrinker_free(huge_zero_page_shrinker);837		return -ENOMEM;838	}839 840	huge_zero_page_shrinker->count_objects = shrink_huge_zero_page_count;841	huge_zero_page_shrinker->scan_objects = shrink_huge_zero_page_scan;842	shrinker_register(huge_zero_page_shrinker);843 844	deferred_split_shrinker->count_objects = deferred_split_count;845	deferred_split_shrinker->scan_objects = deferred_split_scan;846	shrinker_register(deferred_split_shrinker);847 848	return 0;849}850 851static void __init thp_shrinker_exit(void)852{853	shrinker_free(huge_zero_page_shrinker);854	shrinker_free(deferred_split_shrinker);855}856 857static int __init hugepage_init(void)858{859	int err;860	struct kobject *hugepage_kobj;861 862	if (!has_transparent_hugepage()) {863		transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED;864		return -EINVAL;865	}866 867	/*868	 * hugepages can't be allocated by the buddy allocator869	 */870	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER > MAX_PAGE_ORDER);871 872	err = hugepage_init_sysfs(&hugepage_kobj);873	if (err)874		goto err_sysfs;875 876	err = khugepaged_init();877	if (err)878		goto err_slab;879 880	err = thp_shrinker_init();881	if (err)882		goto err_shrinker;883 884	/*885	 * By default disable transparent hugepages on smaller systems,886	 * where the extra memory used could hurt more than TLB overhead887	 * is likely to save.  The admin can still enable it through /sys.888	 */889	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {890		transparent_hugepage_flags = 0;891		return 0;892	}893 894	err = start_stop_khugepaged();895	if (err)896		goto err_khugepaged;897 898	return 0;899err_khugepaged:900	thp_shrinker_exit();901err_shrinker:902	khugepaged_destroy();903err_slab:904	hugepage_exit_sysfs(hugepage_kobj);905err_sysfs:906	return err;907}908subsys_initcall(hugepage_init);909 910static int __init setup_transparent_hugepage(char *str)911{912	int ret = 0;913	if (!str)914		goto out;915	if (!strcmp(str, "always")) {916		set_bit(TRANSPARENT_HUGEPAGE_FLAG,917			&transparent_hugepage_flags);918		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,919			  &transparent_hugepage_flags);920		ret = 1;921	} else if (!strcmp(str, "madvise")) {922		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,923			  &transparent_hugepage_flags);924		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,925			&transparent_hugepage_flags);926		ret = 1;927	} else if (!strcmp(str, "never")) {928		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,929			  &transparent_hugepage_flags);930		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,931			  &transparent_hugepage_flags);932		ret = 1;933	}934out:935	if (!ret)936		pr_warn("transparent_hugepage= cannot parse, ignored\n");937	return ret;938}939__setup("transparent_hugepage=", setup_transparent_hugepage);940 941static inline int get_order_from_str(const char *size_str)942{943	unsigned long size;944	char *endptr;945	int order;946 947	size = memparse(size_str, &endptr);948 949	if (!is_power_of_2(size))950		goto err;951	order = get_order(size);952	if (BIT(order) & ~THP_ORDERS_ALL_ANON)953		goto err;954 955	return order;956err:957	pr_err("invalid size %s in thp_anon boot parameter\n", size_str);958	return -EINVAL;959}960 961static char str_dup[PAGE_SIZE] __initdata;962static int __init setup_thp_anon(char *str)963{964	char *token, *range, *policy, *subtoken;965	unsigned long always, inherit, madvise;966	char *start_size, *end_size;967	int start, end, nr;968	char *p;969 970	if (!str || strlen(str) + 1 > PAGE_SIZE)971		goto err;972	strcpy(str_dup, str);973 974	always = huge_anon_orders_always;975	madvise = huge_anon_orders_madvise;976	inherit = huge_anon_orders_inherit;977	p = str_dup;978	while ((token = strsep(&p, ";")) != NULL) {979		range = strsep(&token, ":");980		policy = token;981 982		if (!policy)983			goto err;984 985		while ((subtoken = strsep(&range, ",")) != NULL) {986			if (strchr(subtoken, '-')) {987				start_size = strsep(&subtoken, "-");988				end_size = subtoken;989 990				start = get_order_from_str(start_size);991				end = get_order_from_str(end_size);992			} else {993				start = end = get_order_from_str(subtoken);994			}995 996			if (start < 0 || end < 0 || start > end)997				goto err;998 999			nr = end - start + 1;1000			if (!strcmp(policy, "always")) {1001				bitmap_set(&always, start, nr);1002				bitmap_clear(&inherit, start, nr);1003				bitmap_clear(&madvise, start, nr);1004			} else if (!strcmp(policy, "madvise")) {1005				bitmap_set(&madvise, start, nr);1006				bitmap_clear(&inherit, start, nr);1007				bitmap_clear(&always, start, nr);1008			} else if (!strcmp(policy, "inherit")) {1009				bitmap_set(&inherit, start, nr);1010				bitmap_clear(&madvise, start, nr);1011				bitmap_clear(&always, start, nr);1012			} else if (!strcmp(policy, "never")) {1013				bitmap_clear(&inherit, start, nr);1014				bitmap_clear(&madvise, start, nr);1015				bitmap_clear(&always, start, nr);1016			} else {1017				pr_err("invalid policy %s in thp_anon boot parameter\n", policy);1018				goto err;1019			}1020		}1021	}1022 1023	huge_anon_orders_always = always;1024	huge_anon_orders_madvise = madvise;1025	huge_anon_orders_inherit = inherit;1026	anon_orders_configured = true;1027	return 1;1028 1029err:1030	pr_warn("thp_anon=%s: error parsing string, ignoring setting\n", str);1031	return 0;1032}1033__setup("thp_anon=", setup_thp_anon);1034 1035pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)1036{1037	if (likely(vma->vm_flags & VM_WRITE))1038		pmd = pmd_mkwrite(pmd, vma);1039	return pmd;1040}1041 1042#ifdef CONFIG_MEMCG1043static inline1044struct deferred_split *get_deferred_split_queue(struct folio *folio)1045{1046	struct mem_cgroup *memcg = folio_memcg(folio);1047	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));1048 1049	if (memcg)1050		return &memcg->deferred_split_queue;1051	else1052		return &pgdat->deferred_split_queue;1053}1054#else1055static inline1056struct deferred_split *get_deferred_split_queue(struct folio *folio)1057{1058	struct pglist_data *pgdat = NODE_DATA(folio_nid(folio));1059 1060	return &pgdat->deferred_split_queue;1061}1062#endif1063 1064static inline bool is_transparent_hugepage(const struct folio *folio)1065{1066	if (!folio_test_large(folio))1067		return false;1068 1069	return is_huge_zero_folio(folio) ||1070		folio_test_large_rmappable(folio);1071}1072 1073static unsigned long __thp_get_unmapped_area(struct file *filp,1074		unsigned long addr, unsigned long len,1075		loff_t off, unsigned long flags, unsigned long size,1076		vm_flags_t vm_flags)1077{1078	loff_t off_end = off + len;1079	loff_t off_align = round_up(off, size);1080	unsigned long len_pad, ret, off_sub;1081 1082	if (!IS_ENABLED(CONFIG_64BIT) || in_compat_syscall())1083		return 0;1084 1085	if (off_end <= off_align || (off_end - off_align) < size)1086		return 0;1087 1088	len_pad = len + size;1089	if (len_pad < len || (off + len_pad) < off)1090		return 0;1091 1092	ret = mm_get_unmapped_area_vmflags(current->mm, filp, addr, len_pad,1093					   off >> PAGE_SHIFT, flags, vm_flags);1094 1095	/*1096	 * The failure might be due to length padding. The caller will retry1097	 * without the padding.1098	 */1099	if (IS_ERR_VALUE(ret))1100		return 0;1101 1102	/*1103	 * Do not try to align to THP boundary if allocation at the address1104	 * hint succeeds.1105	 */1106	if (ret == addr)1107		return addr;1108 1109	off_sub = (off - ret) & (size - 1);1110 1111	if (test_bit(MMF_TOPDOWN, &current->mm->flags) && !off_sub)1112		return ret + size;1113 1114	ret += off_sub;1115	return ret;1116}1117 1118unsigned long thp_get_unmapped_area_vmflags(struct file *filp, unsigned long addr,1119		unsigned long len, unsigned long pgoff, unsigned long flags,1120		vm_flags_t vm_flags)1121{1122	unsigned long ret;1123	loff_t off = (loff_t)pgoff << PAGE_SHIFT;1124 1125	ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE, vm_flags);1126	if (ret)1127		return ret;1128 1129	return mm_get_unmapped_area_vmflags(current->mm, filp, addr, len, pgoff, flags,1130					    vm_flags);1131}1132 1133unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,1134		unsigned long len, unsigned long pgoff, unsigned long flags)1135{1136	return thp_get_unmapped_area_vmflags(filp, addr, len, pgoff, flags, 0);1137}1138EXPORT_SYMBOL_GPL(thp_get_unmapped_area);1139 1140static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,1141			struct page *page, gfp_t gfp)1142{1143	struct vm_area_struct *vma = vmf->vma;1144	struct folio *folio = page_folio(page);1145	pgtable_t pgtable;1146	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;1147	vm_fault_t ret = 0;1148 1149	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);1150 1151	if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {1152		folio_put(folio);1153		count_vm_event(THP_FAULT_FALLBACK);1154		count_vm_event(THP_FAULT_FALLBACK_CHARGE);1155		count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK);1156		count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);1157		return VM_FAULT_FALLBACK;1158	}1159	folio_throttle_swaprate(folio, gfp);1160 1161	pgtable = pte_alloc_one(vma->vm_mm);1162	if (unlikely(!pgtable)) {1163		ret = VM_FAULT_OOM;1164		goto release;1165	}1166 1167	folio_zero_user(folio, vmf->address);1168	/*1169	 * The memory barrier inside __folio_mark_uptodate makes sure that1170	 * folio_zero_user writes become visible before the set_pmd_at()1171	 * write.1172	 */1173	__folio_mark_uptodate(folio);1174 1175	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);1176	if (unlikely(!pmd_none(*vmf->pmd))) {1177		goto unlock_release;1178	} else {1179		pmd_t entry;1180 1181		ret = check_stable_address_space(vma->vm_mm);1182		if (ret)1183			goto unlock_release;1184 1185		/* Deliver the page fault to userland */1186		if (userfaultfd_missing(vma)) {1187			spin_unlock(vmf->ptl);1188			folio_put(folio);1189			pte_free(vma->vm_mm, pgtable);1190			ret = handle_userfault(vmf, VM_UFFD_MISSING);1191			VM_BUG_ON(ret & VM_FAULT_FALLBACK);1192			return ret;1193		}1194 1195		entry = mk_huge_pmd(page, vma->vm_page_prot);1196		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);1197		folio_add_new_anon_rmap(folio, vma, haddr, RMAP_EXCLUSIVE);1198		folio_add_lru_vma(folio, vma);1199		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);1200		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);1201		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);1202		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);1203		mm_inc_nr_ptes(vma->vm_mm);1204		deferred_split_folio(folio, false);1205		spin_unlock(vmf->ptl);1206		count_vm_event(THP_FAULT_ALLOC);1207		count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC);1208		count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);1209	}1210 1211	return 0;1212unlock_release:1213	spin_unlock(vmf->ptl);1214release:1215	if (pgtable)1216		pte_free(vma->vm_mm, pgtable);1217	folio_put(folio);1218	return ret;1219 1220}1221 1222/*1223 * always: directly stall for all thp allocations1224 * defer: wake kswapd and fail if not immediately available1225 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise1226 *		  fail if not immediately available1227 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately1228 *	    available1229 * never: never stall for any thp allocation1230 */1231gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)1232{1233	const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);1234 1235	/* Always do synchronous compaction */1236	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))1237		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);1238 1239	/* Kick kcompactd and fail quickly */1240	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))1241		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;1242 1243	/* Synchronous compaction if madvised, otherwise kick kcompactd */1244	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))1245		return GFP_TRANSHUGE_LIGHT |1246			(vma_madvised ? __GFP_DIRECT_RECLAIM :1247					__GFP_KSWAPD_RECLAIM);1248 1249	/* Only do synchronous compaction if madvised */1250	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))1251		return GFP_TRANSHUGE_LIGHT |1252		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);1253 1254	return GFP_TRANSHUGE_LIGHT;1255}1256 1257/* Caller must hold page table lock. */1258static void set_huge_zero_folio(pgtable_t pgtable, struct mm_struct *mm,1259		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,1260		struct folio *zero_folio)1261{1262	pmd_t entry;1263	if (!pmd_none(*pmd))1264		return;1265	entry = mk_pmd(&zero_folio->page, vma->vm_page_prot);1266	entry = pmd_mkhuge(entry);1267	pgtable_trans_huge_deposit(mm, pmd, pgtable);1268	set_pmd_at(mm, haddr, pmd, entry);1269	mm_inc_nr_ptes(mm);1270}1271 1272vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)1273{1274	struct vm_area_struct *vma = vmf->vma;1275	gfp_t gfp;1276	struct folio *folio;1277	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;1278	vm_fault_t ret;1279 1280	if (!thp_vma_suitable_order(vma, haddr, PMD_ORDER))1281		return VM_FAULT_FALLBACK;1282	ret = vmf_anon_prepare(vmf);1283	if (ret)1284		return ret;1285	khugepaged_enter_vma(vma, vma->vm_flags);1286 1287	if (!(vmf->flags & FAULT_FLAG_WRITE) &&1288			!mm_forbids_zeropage(vma->vm_mm) &&1289			transparent_hugepage_use_zero_page()) {1290		pgtable_t pgtable;1291		struct folio *zero_folio;1292		vm_fault_t ret;1293 1294		pgtable = pte_alloc_one(vma->vm_mm);1295		if (unlikely(!pgtable))1296			return VM_FAULT_OOM;1297		zero_folio = mm_get_huge_zero_folio(vma->vm_mm);1298		if (unlikely(!zero_folio)) {1299			pte_free(vma->vm_mm, pgtable);1300			count_vm_event(THP_FAULT_FALLBACK);1301			return VM_FAULT_FALLBACK;1302		}1303		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);1304		ret = 0;1305		if (pmd_none(*vmf->pmd)) {1306			ret = check_stable_address_space(vma->vm_mm);1307			if (ret) {1308				spin_unlock(vmf->ptl);1309				pte_free(vma->vm_mm, pgtable);1310			} else if (userfaultfd_missing(vma)) {1311				spin_unlock(vmf->ptl);1312				pte_free(vma->vm_mm, pgtable);1313				ret = handle_userfault(vmf, VM_UFFD_MISSING);1314				VM_BUG_ON(ret & VM_FAULT_FALLBACK);1315			} else {1316				set_huge_zero_folio(pgtable, vma->vm_mm, vma,1317						   haddr, vmf->pmd, zero_folio);1318				update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);1319				spin_unlock(vmf->ptl);1320			}1321		} else {1322			spin_unlock(vmf->ptl);1323			pte_free(vma->vm_mm, pgtable);1324		}1325		return ret;1326	}1327	gfp = vma_thp_gfp_mask(vma);1328	folio = vma_alloc_folio(gfp, HPAGE_PMD_ORDER, vma, haddr, true);1329	if (unlikely(!folio)) {1330		count_vm_event(THP_FAULT_FALLBACK);1331		count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK);1332		return VM_FAULT_FALLBACK;1333	}1334	return __do_huge_pmd_anonymous_page(vmf, &folio->page, gfp);1335}1336 1337static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,1338		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,1339		pgtable_t pgtable)1340{1341	struct mm_struct *mm = vma->vm_mm;1342	pmd_t entry;1343	spinlock_t *ptl;1344 1345	ptl = pmd_lock(mm, pmd);1346	if (!pmd_none(*pmd)) {1347		if (write) {1348			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {1349				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));1350				goto out_unlock;1351			}1352			entry = pmd_mkyoung(*pmd);1353			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);1354			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))1355				update_mmu_cache_pmd(vma, addr, pmd);1356		}1357 1358		goto out_unlock;1359	}1360 1361	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));1362	if (pfn_t_devmap(pfn))1363		entry = pmd_mkdevmap(entry);1364	else1365		entry = pmd_mkspecial(entry);1366	if (write) {1367		entry = pmd_mkyoung(pmd_mkdirty(entry));1368		entry = maybe_pmd_mkwrite(entry, vma);1369	}1370 1371	if (pgtable) {1372		pgtable_trans_huge_deposit(mm, pmd, pgtable);1373		mm_inc_nr_ptes(mm);1374		pgtable = NULL;1375	}1376 1377	set_pmd_at(mm, addr, pmd, entry);1378	update_mmu_cache_pmd(vma, addr, pmd);1379 1380out_unlock:1381	spin_unlock(ptl);1382	if (pgtable)1383		pte_free(mm, pgtable);1384}1385 1386/**1387 * vmf_insert_pfn_pmd - insert a pmd size pfn1388 * @vmf: Structure describing the fault1389 * @pfn: pfn to insert1390 * @write: whether it's a write fault1391 *1392 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.1393 *1394 * Return: vm_fault_t value.1395 */1396vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)1397{1398	unsigned long addr = vmf->address & PMD_MASK;1399	struct vm_area_struct *vma = vmf->vma;1400	pgprot_t pgprot = vma->vm_page_prot;1401	pgtable_t pgtable = NULL;1402 1403	/*1404	 * If we had pmd_special, we could avoid all these restrictions,1405	 * but we need to be consistent with PTEs and architectures that1406	 * can't support a 'special' bit.1407	 */1408	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&1409			!pfn_t_devmap(pfn));1410	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==1411						(VM_PFNMAP|VM_MIXEDMAP));1412	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));1413 1414	if (addr < vma->vm_start || addr >= vma->vm_end)1415		return VM_FAULT_SIGBUS;1416 1417	if (arch_needs_pgtable_deposit()) {1418		pgtable = pte_alloc_one(vma->vm_mm);1419		if (!pgtable)1420			return VM_FAULT_OOM;1421	}1422 1423	track_pfn_insert(vma, &pgprot, pfn);1424 1425	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);1426	return VM_FAULT_NOPAGE;1427}1428EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);1429 1430#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD1431static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)1432{1433	if (likely(vma->vm_flags & VM_WRITE))1434		pud = pud_mkwrite(pud);1435	return pud;1436}1437 1438static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,1439		pud_t *pud, pfn_t pfn, bool write)1440{1441	struct mm_struct *mm = vma->vm_mm;1442	pgprot_t prot = vma->vm_page_prot;1443	pud_t entry;1444	spinlock_t *ptl;1445 1446	ptl = pud_lock(mm, pud);1447	if (!pud_none(*pud)) {1448		if (write) {1449			if (WARN_ON_ONCE(pud_pfn(*pud) != pfn_t_to_pfn(pfn)))1450				goto out_unlock;1451			entry = pud_mkyoung(*pud);1452			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);1453			if (pudp_set_access_flags(vma, addr, pud, entry, 1))1454				update_mmu_cache_pud(vma, addr, pud);1455		}1456		goto out_unlock;1457	}1458 1459	entry = pud_mkhuge(pfn_t_pud(pfn, prot));1460	if (pfn_t_devmap(pfn))1461		entry = pud_mkdevmap(entry);1462	else1463		entry = pud_mkspecial(entry);1464	if (write) {1465		entry = pud_mkyoung(pud_mkdirty(entry));1466		entry = maybe_pud_mkwrite(entry, vma);1467	}1468	set_pud_at(mm, addr, pud, entry);1469	update_mmu_cache_pud(vma, addr, pud);1470 1471out_unlock:1472	spin_unlock(ptl);1473}1474 1475/**1476 * vmf_insert_pfn_pud - insert a pud size pfn1477 * @vmf: Structure describing the fault1478 * @pfn: pfn to insert1479 * @write: whether it's a write fault1480 *1481 * Insert a pud size pfn. See vmf_insert_pfn() for additional info.1482 *1483 * Return: vm_fault_t value.1484 */1485vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)1486{1487	unsigned long addr = vmf->address & PUD_MASK;1488	struct vm_area_struct *vma = vmf->vma;1489	pgprot_t pgprot = vma->vm_page_prot;1490 1491	/*1492	 * If we had pud_special, we could avoid all these restrictions,1493	 * but we need to be consistent with PTEs and architectures that1494	 * can't support a 'special' bit.1495	 */1496	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&1497			!pfn_t_devmap(pfn));1498	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==1499						(VM_PFNMAP|VM_MIXEDMAP));1500	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));1501 1502	if (addr < vma->vm_start || addr >= vma->vm_end)1503		return VM_FAULT_SIGBUS;1504 1505	track_pfn_insert(vma, &pgprot, pfn);1506 1507	insert_pfn_pud(vma, addr, vmf->pud, pfn, write);1508	return VM_FAULT_NOPAGE;1509}1510EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);1511#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */1512 1513void touch_pmd(struct vm_area_struct *vma, unsigned long addr,1514	       pmd_t *pmd, bool write)1515{1516	pmd_t _pmd;1517 1518	_pmd = pmd_mkyoung(*pmd);1519	if (write)1520		_pmd = pmd_mkdirty(_pmd);1521	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,1522				  pmd, _pmd, write))1523		update_mmu_cache_pmd(vma, addr, pmd);1524}1525 1526struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,1527		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)1528{1529	unsigned long pfn = pmd_pfn(*pmd);1530	struct mm_struct *mm = vma->vm_mm;1531	struct page *page;1532	int ret;1533 1534	assert_spin_locked(pmd_lockptr(mm, pmd));1535 1536	if (flags & FOLL_WRITE && !pmd_write(*pmd))1537		return NULL;1538 1539	if (pmd_present(*pmd) && pmd_devmap(*pmd))1540		/* pass */;1541	else1542		return NULL;1543 1544	if (flags & FOLL_TOUCH)1545		touch_pmd(vma, addr, pmd, flags & FOLL_WRITE);1546 1547	/*1548	 * device mapped pages can only be returned if the1549	 * caller will manage the page reference count.1550	 */1551	if (!(flags & (FOLL_GET | FOLL_PIN)))1552		return ERR_PTR(-EEXIST);1553 1554	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;1555	*pgmap = get_dev_pagemap(pfn, *pgmap);1556	if (!*pgmap)1557		return ERR_PTR(-EFAULT);1558	page = pfn_to_page(pfn);1559	ret = try_grab_folio(page_folio(page), 1, flags);1560	if (ret)1561		page = ERR_PTR(ret);1562 1563	return page;1564}1565 1566int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,1567		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,1568		  struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)1569{1570	spinlock_t *dst_ptl, *src_ptl;1571	struct page *src_page;1572	struct folio *src_folio;1573	pmd_t pmd;1574	pgtable_t pgtable = NULL;1575	int ret = -ENOMEM;1576 1577	pmd = pmdp_get_lockless(src_pmd);1578	if (unlikely(pmd_present(pmd) && pmd_special(pmd))) {1579		dst_ptl = pmd_lock(dst_mm, dst_pmd);1580		src_ptl = pmd_lockptr(src_mm, src_pmd);1581		spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);1582		/*1583		 * No need to recheck the pmd, it can't change with write1584		 * mmap lock held here.1585		 *1586		 * Meanwhile, making sure it's not a CoW VMA with writable1587		 * mapping, otherwise it means either the anon page wrongly1588		 * applied special bit, or we made the PRIVATE mapping be1589		 * able to wrongly write to the backend MMIO.1590		 */1591		VM_WARN_ON_ONCE(is_cow_mapping(src_vma->vm_flags) && pmd_write(pmd));1592		goto set_pmd;1593	}1594 1595	/* Skip if can be re-fill on fault */1596	if (!vma_is_anonymous(dst_vma))1597		return 0;1598 1599	pgtable = pte_alloc_one(dst_mm);1600	if (unlikely(!pgtable))1601		goto out;1602 1603	dst_ptl = pmd_lock(dst_mm, dst_pmd);1604	src_ptl = pmd_lockptr(src_mm, src_pmd);1605	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);1606 1607	ret = -EAGAIN;1608	pmd = *src_pmd;1609 1610#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION1611	if (unlikely(is_swap_pmd(pmd))) {1612		swp_entry_t entry = pmd_to_swp_entry(pmd);1613 1614		VM_BUG_ON(!is_pmd_migration_entry(pmd));1615		if (!is_readable_migration_entry(entry)) {1616			entry = make_readable_migration_entry(1617							swp_offset(entry));1618			pmd = swp_entry_to_pmd(entry);1619			if (pmd_swp_soft_dirty(*src_pmd))1620				pmd = pmd_swp_mksoft_dirty(pmd);1621			if (pmd_swp_uffd_wp(*src_pmd))1622				pmd = pmd_swp_mkuffd_wp(pmd);1623			set_pmd_at(src_mm, addr, src_pmd, pmd);1624		}1625		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);1626		mm_inc_nr_ptes(dst_mm);1627		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);1628		if (!userfaultfd_wp(dst_vma))1629			pmd = pmd_swp_clear_uffd_wp(pmd);1630		set_pmd_at(dst_mm, addr, dst_pmd, pmd);1631		ret = 0;1632		goto out_unlock;1633	}1634#endif1635 1636	if (unlikely(!pmd_trans_huge(pmd))) {1637		pte_free(dst_mm, pgtable);1638		goto out_unlock;1639	}1640	/*1641	 * When page table lock is held, the huge zero pmd should not be1642	 * under splitting since we don't split the page itself, only pmd to1643	 * a page table.1644	 */1645	if (is_huge_zero_pmd(pmd)) {1646		/*1647		 * mm_get_huge_zero_folio() will never allocate a new1648		 * folio here, since we already have a zero page to1649		 * copy. It just takes a reference.1650		 */1651		mm_get_huge_zero_folio(dst_mm);1652		goto out_zero_page;1653	}1654 1655	src_page = pmd_page(pmd);1656	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);1657	src_folio = page_folio(src_page);1658 1659	folio_get(src_folio);1660	if (unlikely(folio_try_dup_anon_rmap_pmd(src_folio, src_page, src_vma))) {1661		/* Page maybe pinned: split and retry the fault on PTEs. */1662		folio_put(src_folio);1663		pte_free(dst_mm, pgtable);1664		spin_unlock(src_ptl);1665		spin_unlock(dst_ptl);1666		__split_huge_pmd(src_vma, src_pmd, addr, false, NULL);1667		return -EAGAIN;1668	}1669	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);1670out_zero_page:1671	mm_inc_nr_ptes(dst_mm);1672	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);1673	pmdp_set_wrprotect(src_mm, addr, src_pmd);1674	if (!userfaultfd_wp(dst_vma))1675		pmd = pmd_clear_uffd_wp(pmd);1676	pmd = pmd_wrprotect(pmd);1677set_pmd:1678	pmd = pmd_mkold(pmd);1679	set_pmd_at(dst_mm, addr, dst_pmd, pmd);1680 1681	ret = 0;1682out_unlock:1683	spin_unlock(src_ptl);1684	spin_unlock(dst_ptl);1685out:1686	return ret;1687}1688 1689#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD1690void touch_pud(struct vm_area_struct *vma, unsigned long addr,1691	       pud_t *pud, bool write)1692{1693	pud_t _pud;1694 1695	_pud = pud_mkyoung(*pud);1696	if (write)1697		_pud = pud_mkdirty(_pud);1698	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,1699				  pud, _pud, write))1700		update_mmu_cache_pud(vma, addr, pud);1701}1702 1703int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,1704		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,1705		  struct vm_area_struct *vma)1706{1707	spinlock_t *dst_ptl, *src_ptl;1708	pud_t pud;1709	int ret;1710 1711	dst_ptl = pud_lock(dst_mm, dst_pud);1712	src_ptl = pud_lockptr(src_mm, src_pud);1713	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);1714 1715	ret = -EAGAIN;1716	pud = *src_pud;1717	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))1718		goto out_unlock;1719 1720	/*1721	 * TODO: once we support anonymous pages, use1722	 * folio_try_dup_anon_rmap_*() and split if duplicating fails.1723	 */1724	if (is_cow_mapping(vma->vm_flags) && pud_write(pud)) {1725		pudp_set_wrprotect(src_mm, addr, src_pud);1726		pud = pud_wrprotect(pud);1727	}1728	pud = pud_mkold(pud);1729	set_pud_at(dst_mm, addr, dst_pud, pud);1730 1731	ret = 0;1732out_unlock:1733	spin_unlock(src_ptl);1734	spin_unlock(dst_ptl);1735	return ret;1736}1737 1738void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)1739{1740	bool write = vmf->flags & FAULT_FLAG_WRITE;1741 1742	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);1743	if (unlikely(!pud_same(*vmf->pud, orig_pud)))1744		goto unlock;1745 1746	touch_pud(vmf->vma, vmf->address, vmf->pud, write);1747unlock:1748	spin_unlock(vmf->ptl);1749}1750#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */1751 1752void huge_pmd_set_accessed(struct vm_fault *vmf)1753{1754	bool write = vmf->flags & FAULT_FLAG_WRITE;1755 1756	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);1757	if (unlikely(!pmd_same(*vmf->pmd, vmf->orig_pmd)))1758		goto unlock;1759 1760	touch_pmd(vmf->vma, vmf->address, vmf->pmd, write);1761 1762unlock:1763	spin_unlock(vmf->ptl);1764}1765 1766vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf)1767{1768	const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE;1769	struct vm_area_struct *vma = vmf->vma;1770	struct folio *folio;1771	struct page *page;1772	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;1773	pmd_t orig_pmd = vmf->orig_pmd;1774 1775	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);1776	VM_BUG_ON_VMA(!vma->anon_vma, vma);1777 1778	if (is_huge_zero_pmd(orig_pmd))1779		goto fallback;1780 1781	spin_lock(vmf->ptl);1782 1783	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {1784		spin_unlock(vmf->ptl);1785		return 0;1786	}1787 1788	page = pmd_page(orig_pmd);1789	folio = page_folio(page);1790	VM_BUG_ON_PAGE(!PageHead(page), page);1791 1792	/* Early check when only holding the PT lock. */1793	if (PageAnonExclusive(page))1794		goto reuse;1795 1796	if (!folio_trylock(folio)) {1797		folio_get(folio);1798		spin_unlock(vmf->ptl);1799		folio_lock(folio);1800		spin_lock(vmf->ptl);1801		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {1802			spin_unlock(vmf->ptl);1803			folio_unlock(folio);1804			folio_put(folio);1805			return 0;1806		}1807		folio_put(folio);1808	}1809 1810	/* Recheck after temporarily dropping the PT lock. */1811	if (PageAnonExclusive(page)) {1812		folio_unlock(folio);1813		goto reuse;1814	}1815 1816	/*1817	 * See do_wp_page(): we can only reuse the folio exclusively if1818	 * there are no additional references. Note that we always drain1819	 * the LRU cache immediately after adding a THP.1820	 */1821	if (folio_ref_count(folio) >1822			1 + folio_test_swapcache(folio) * folio_nr_pages(folio))1823		goto unlock_fallback;1824	if (folio_test_swapcache(folio))1825		folio_free_swap(folio);1826	if (folio_ref_count(folio) == 1) {1827		pmd_t entry;1828 1829		folio_move_anon_rmap(folio, vma);1830		SetPageAnonExclusive(page);1831		folio_unlock(folio);1832reuse:1833		if (unlikely(unshare)) {1834			spin_unlock(vmf->ptl);1835			return 0;1836		}1837		entry = pmd_mkyoung(orig_pmd);1838		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);1839		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))1840			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);1841		spin_unlock(vmf->ptl);1842		return 0;1843	}1844 1845unlock_fallback:1846	folio_unlock(folio);1847	spin_unlock(vmf->ptl);1848fallback:1849	__split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);1850	return VM_FAULT_FALLBACK;1851}1852 1853static inline bool can_change_pmd_writable(struct vm_area_struct *vma,1854					   unsigned long addr, pmd_t pmd)1855{1856	struct page *page;1857 1858	if (WARN_ON_ONCE(!(vma->vm_flags & VM_WRITE)))1859		return false;1860 1861	/* Don't touch entries that are not even readable (NUMA hinting). */1862	if (pmd_protnone(pmd))1863		return false;1864 1865	/* Do we need write faults for softdirty tracking? */1866	if (pmd_needs_soft_dirty_wp(vma, pmd))1867		return false;1868 1869	/* Do we need write faults for uffd-wp tracking? */1870	if (userfaultfd_huge_pmd_wp(vma, pmd))1871		return false;1872 1873	if (!(vma->vm_flags & VM_SHARED)) {1874		/* See can_change_pte_writable(). */1875		page = vm_normal_page_pmd(vma, addr, pmd);1876		return page && PageAnon(page) && PageAnonExclusive(page);1877	}1878 1879	/* See can_change_pte_writable(). */1880	return pmd_dirty(pmd);1881}1882 1883/* NUMA hinting page fault entry point for trans huge pmds */1884vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)1885{1886	struct vm_area_struct *vma = vmf->vma;1887	struct folio *folio;1888	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;1889	int nid = NUMA_NO_NODE;1890	int target_nid, last_cpupid;1891	pmd_t pmd, old_pmd;1892	bool writable = false;1893	int flags = 0;1894 1895	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);1896	old_pmd = pmdp_get(vmf->pmd);1897 1898	if (unlikely(!pmd_same(old_pmd, vmf->orig_pmd))) {1899		spin_unlock(vmf->ptl);1900		return 0;1901	}1902 1903	pmd = pmd_modify(old_pmd, vma->vm_page_prot);1904 1905	/*1906	 * Detect now whether the PMD could be writable; this information1907	 * is only valid while holding the PT lock.1908	 */1909	writable = pmd_write(pmd);1910	if (!writable && vma_wants_manual_pte_write_upgrade(vma) &&1911	    can_change_pmd_writable(vma, vmf->address, pmd))1912		writable = true;1913 1914	folio = vm_normal_folio_pmd(vma, haddr, pmd);1915	if (!folio)1916		goto out_map;1917 1918	nid = folio_nid(folio);1919 1920	target_nid = numa_migrate_check(folio, vmf, haddr, &flags, writable,1921					&last_cpupid);1922	if (target_nid == NUMA_NO_NODE)1923		goto out_map;1924	if (migrate_misplaced_folio_prepare(folio, vma, target_nid)) {1925		flags |= TNF_MIGRATE_FAIL;1926		goto out_map;1927	}1928	/* The folio is isolated and isolation code holds a folio reference. */1929	spin_unlock(vmf->ptl);1930	writable = false;1931 1932	if (!migrate_misplaced_folio(folio, vma, target_nid)) {1933		flags |= TNF_MIGRATED;1934		nid = target_nid;1935		task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);1936		return 0;1937	}1938 1939	flags |= TNF_MIGRATE_FAIL;1940	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);1941	if (unlikely(!pmd_same(pmdp_get(vmf->pmd), vmf->orig_pmd))) {1942		spin_unlock(vmf->ptl);1943		return 0;1944	}1945out_map:1946	/* Restore the PMD */1947	pmd = pmd_modify(pmdp_get(vmf->pmd), vma->vm_page_prot);1948	pmd = pmd_mkyoung(pmd);1949	if (writable)1950		pmd = pmd_mkwrite(pmd, vma);1951	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);1952	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);1953	spin_unlock(vmf->ptl);1954 1955	if (nid != NUMA_NO_NODE)1956		task_numa_fault(last_cpupid, nid, HPAGE_PMD_NR, flags);1957	return 0;1958}1959 1960/*1961 * Return true if we do MADV_FREE successfully on entire pmd page.1962 * Otherwise, return false.1963 */1964bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,1965		pmd_t *pmd, unsigned long addr, unsigned long next)1966{1967	spinlock_t *ptl;1968	pmd_t orig_pmd;1969	struct folio *folio;1970	struct mm_struct *mm = tlb->mm;1971	bool ret = false;1972 1973	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);1974 1975	ptl = pmd_trans_huge_lock(pmd, vma);1976	if (!ptl)1977		goto out_unlocked;1978 1979	orig_pmd = *pmd;1980	if (is_huge_zero_pmd(orig_pmd))1981		goto out;1982 1983	if (unlikely(!pmd_present(orig_pmd))) {1984		VM_BUG_ON(thp_migration_supported() &&1985				  !is_pmd_migration_entry(orig_pmd));1986		goto out;1987	}1988 1989	folio = pmd_folio(orig_pmd);1990	/*1991	 * If other processes are mapping this folio, we couldn't discard1992	 * the folio unless they all do MADV_FREE so let's skip the folio.1993	 */1994	if (folio_likely_mapped_shared(folio))1995		goto out;1996 1997	if (!folio_trylock(folio))1998		goto out;1999 2000	/*2001	 * If user want to discard part-pages of THP, split it so MADV_FREE2002	 * will deactivate only them.2003	 */2004	if (next - addr != HPAGE_PMD_SIZE) {2005		folio_get(folio);2006		spin_unlock(ptl);2007		split_folio(folio);2008		folio_unlock(folio);2009		folio_put(folio);2010		goto out_unlocked;2011	}2012 2013	if (folio_test_dirty(folio))2014		folio_clear_dirty(folio);2015	folio_unlock(folio);2016 2017	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {2018		pmdp_invalidate(vma, addr, pmd);2019		orig_pmd = pmd_mkold(orig_pmd);2020		orig_pmd = pmd_mkclean(orig_pmd);2021 2022		set_pmd_at(mm, addr, pmd, orig_pmd);2023		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);2024	}2025 2026	folio_mark_lazyfree(folio);2027	ret = true;2028out:2029	spin_unlock(ptl);2030out_unlocked:2031	return ret;2032}2033 2034static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)2035{2036	pgtable_t pgtable;2037 2038	pgtable = pgtable_trans_huge_withdraw(mm, pmd);2039	pte_free(mm, pgtable);2040	mm_dec_nr_ptes(mm);2041}2042 2043int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,2044		 pmd_t *pmd, unsigned long addr)2045{2046	pmd_t orig_pmd;2047	spinlock_t *ptl;2048 2049	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);2050 2051	ptl = __pmd_trans_huge_lock(pmd, vma);2052	if (!ptl)2053		return 0;2054	/*2055	 * For architectures like ppc64 we look at deposited pgtable2056	 * when calling pmdp_huge_get_and_clear. So do the2057	 * pgtable_trans_huge_withdraw after finishing pmdp related2058	 * operations.2059	 */2060	orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,2061						tlb->fullmm);2062	arch_check_zapped_pmd(vma, orig_pmd);2063	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);2064	if (vma_is_special_huge(vma)) {2065		if (arch_needs_pgtable_deposit())2066			zap_deposited_table(tlb->mm, pmd);2067		spin_unlock(ptl);2068	} else if (is_huge_zero_pmd(orig_pmd)) {2069		zap_deposited_table(tlb->mm, pmd);2070		spin_unlock(ptl);2071	} else {2072		struct folio *folio = NULL;2073		int flush_needed = 1;2074 2075		if (pmd_present(orig_pmd)) {2076			struct page *page = pmd_page(orig_pmd);2077 2078			folio = page_folio(page);2079			folio_remove_rmap_pmd(folio, page, vma);2080			WARN_ON_ONCE(folio_mapcount(folio) < 0);2081			VM_BUG_ON_PAGE(!PageHead(page), page);2082		} else if (thp_migration_supported()) {2083			swp_entry_t entry;2084 2085			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));2086			entry = pmd_to_swp_entry(orig_pmd);2087			folio = pfn_swap_entry_folio(entry);2088			flush_needed = 0;2089		} else2090			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");2091 2092		if (folio_test_anon(folio)) {2093			zap_deposited_table(tlb->mm, pmd);2094			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);2095		} else {2096			if (arch_needs_pgtable_deposit())2097				zap_deposited_table(tlb->mm, pmd);2098			add_mm_counter(tlb->mm, mm_counter_file(folio),2099				       -HPAGE_PMD_NR);2100		}2101 2102		spin_unlock(ptl);2103		if (flush_needed)2104			tlb_remove_page_size(tlb, &folio->page, HPAGE_PMD_SIZE);2105	}2106	return 1;2107}2108 2109#ifndef pmd_move_must_withdraw2110static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,2111					 spinlock_t *old_pmd_ptl,2112					 struct vm_area_struct *vma)2113{2114	/*2115	 * With split pmd lock we also need to move preallocated2116	 * PTE page table if new_pmd is on different PMD page table.2117	 *2118	 * We also don't deposit and withdraw tables for file pages.2119	 */2120	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);2121}2122#endif2123 2124static pmd_t move_soft_dirty_pmd(pmd_t pmd)2125{2126#ifdef CONFIG_MEM_SOFT_DIRTY2127	if (unlikely(is_pmd_migration_entry(pmd)))2128		pmd = pmd_swp_mksoft_dirty(pmd);2129	else if (pmd_present(pmd))2130		pmd = pmd_mksoft_dirty(pmd);2131#endif2132	return pmd;2133}2134 2135bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,2136		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)2137{2138	spinlock_t *old_ptl, *new_ptl;2139	pmd_t pmd;2140	struct mm_struct *mm = vma->vm_mm;2141	bool force_flush = false;2142 2143	/*2144	 * The destination pmd shouldn't be established, free_pgtables()2145	 * should have released it; but move_page_tables() might have already2146	 * inserted a page table, if racing against shmem/file collapse.2147	 */2148	if (!pmd_none(*new_pmd)) {2149		VM_BUG_ON(pmd_trans_huge(*new_pmd));2150		return false;2151	}2152 2153	/*2154	 * We don't have to worry about the ordering of src and dst2155	 * ptlocks because exclusive mmap_lock prevents deadlock.2156	 */2157	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);2158	if (old_ptl) {2159		new_ptl = pmd_lockptr(mm, new_pmd);2160		if (new_ptl != old_ptl)2161			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);2162		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);2163		if (pmd_present(pmd))2164			force_flush = true;2165		VM_BUG_ON(!pmd_none(*new_pmd));2166 2167		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {2168			pgtable_t pgtable;2169			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);2170			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);2171		}2172		pmd = move_soft_dirty_pmd(pmd);2173		set_pmd_at(mm, new_addr, new_pmd, pmd);2174		if (force_flush)2175			flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE);2176		if (new_ptl != old_ptl)2177			spin_unlock(new_ptl);2178		spin_unlock(old_ptl);2179		return true;2180	}2181	return false;2182}2183 2184/*2185 * Returns2186 *  - 0 if PMD could not be locked2187 *  - 1 if PMD was locked but protections unchanged and TLB flush unnecessary2188 *      or if prot_numa but THP migration is not supported2189 *  - HPAGE_PMD_NR if protections changed and TLB flush necessary2190 */2191int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,2192		    pmd_t *pmd, unsigned long addr, pgprot_t newprot,2193		    unsigned long cp_flags)2194{2195	struct mm_struct *mm = vma->vm_mm;2196	spinlock_t *ptl;2197	pmd_t oldpmd, entry;2198	bool prot_numa = cp_flags & MM_CP_PROT_NUMA;2199	bool uffd_wp = cp_flags & MM_CP_UFFD_WP;2200	bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;2201	int ret = 1;2202 2203	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);2204 2205	if (prot_numa && !thp_migration_supported())2206		return 1;2207 2208	ptl = __pmd_trans_huge_lock(pmd, vma);2209	if (!ptl)2210		return 0;2211 2212#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION2213	if (is_swap_pmd(*pmd)) {2214		swp_entry_t entry = pmd_to_swp_entry(*pmd);2215		struct folio *folio = pfn_swap_entry_folio(entry);2216		pmd_t newpmd;2217 2218		VM_BUG_ON(!is_pmd_migration_entry(*pmd));2219		if (is_writable_migration_entry(entry)) {2220			/*2221			 * A protection check is difficult so2222			 * just be safe and disable write2223			 */2224			if (folio_test_anon(folio))2225				entry = make_readable_exclusive_migration_entry(swp_offset(entry));2226			else2227				entry = make_readable_migration_entry(swp_offset(entry));2228			newpmd = swp_entry_to_pmd(entry);2229			if (pmd_swp_soft_dirty(*pmd))2230				newpmd = pmd_swp_mksoft_dirty(newpmd);2231		} else {2232			newpmd = *pmd;2233		}2234 2235		if (uffd_wp)2236			newpmd = pmd_swp_mkuffd_wp(newpmd);2237		else if (uffd_wp_resolve)2238			newpmd = pmd_swp_clear_uffd_wp(newpmd);2239		if (!pmd_same(*pmd, newpmd))2240			set_pmd_at(mm, addr, pmd, newpmd);2241		goto unlock;2242	}2243#endif2244 2245	if (prot_numa) {2246		struct folio *folio;2247		bool toptier;2248		/*2249		 * Avoid trapping faults against the zero page. The read-only2250		 * data is likely to be read-cached on the local CPU and2251		 * local/remote hits to the zero page are not interesting.2252		 */2253		if (is_huge_zero_pmd(*pmd))2254			goto unlock;2255 2256		if (pmd_protnone(*pmd))2257			goto unlock;2258 2259		folio = pmd_folio(*pmd);2260		toptier = node_is_toptier(folio_nid(folio));2261		/*2262		 * Skip scanning top tier node if normal numa2263		 * balancing is disabled2264		 */2265		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&2266		    toptier)2267			goto unlock;2268 2269		if (folio_use_access_time(folio))2270			folio_xchg_access_time(folio,2271					       jiffies_to_msecs(jiffies));2272	}2273	/*2274	 * In case prot_numa, we are under mmap_read_lock(mm). It's critical2275	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED2276	 * which is also under mmap_read_lock(mm):2277	 *2278	 *	CPU0:				CPU1:2279	 *				change_huge_pmd(prot_numa=1)2280	 *				 pmdp_huge_get_and_clear_notify()2281	 * madvise_dontneed()2282	 *  zap_pmd_range()2283	 *   pmd_trans_huge(*pmd) == 0 (without ptl)2284	 *   // skip the pmd2285	 *				 set_pmd_at();2286	 *				 // pmd is re-established2287	 *2288	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it2289	 * which may break userspace.2290	 *2291	 * pmdp_invalidate_ad() is required to make sure we don't miss2292	 * dirty/young flags set by hardware.2293	 */2294	oldpmd = pmdp_invalidate_ad(vma, addr, pmd);2295 2296	entry = pmd_modify(oldpmd, newprot);2297	if (uffd_wp)2298		entry = pmd_mkuffd_wp(entry);2299	else if (uffd_wp_resolve)2300		/*2301		 * Leave the write bit to be handled by PF interrupt2302		 * handler, then things like COW could be properly2303		 * handled.2304		 */2305		entry = pmd_clear_uffd_wp(entry);2306 2307	/* See change_pte_range(). */2308	if ((cp_flags & MM_CP_TRY_CHANGE_WRITABLE) && !pmd_write(entry) &&2309	    can_change_pmd_writable(vma, addr, entry))2310		entry = pmd_mkwrite(entry, vma);2311 2312	ret = HPAGE_PMD_NR;2313	set_pmd_at(mm, addr, pmd, entry);2314 2315	if (huge_pmd_needs_flush(oldpmd, entry))2316		tlb_flush_pmd_range(tlb, addr, HPAGE_PMD_SIZE);2317unlock:2318	spin_unlock(ptl);2319	return ret;2320}2321 2322/*2323 * Returns:2324 *2325 * - 0: if pud leaf changed from under us2326 * - 1: if pud can be skipped2327 * - HPAGE_PUD_NR: if pud was successfully processed2328 */2329#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD2330int change_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,2331		    pud_t *pudp, unsigned long addr, pgprot_t newprot,2332		    unsigned long cp_flags)2333{2334	struct mm_struct *mm = vma->vm_mm;2335	pud_t oldpud, entry;2336	spinlock_t *ptl;2337 2338	tlb_change_page_size(tlb, HPAGE_PUD_SIZE);2339 2340	/* NUMA balancing doesn't apply to dax */2341	if (cp_flags & MM_CP_PROT_NUMA)2342		return 1;2343 2344	/*2345	 * Huge entries on userfault-wp only works with anonymous, while we2346	 * don't have anonymous PUDs yet.2347	 */2348	if (WARN_ON_ONCE(cp_flags & MM_CP_UFFD_WP_ALL))2349		return 1;2350 2351	ptl = __pud_trans_huge_lock(pudp, vma);2352	if (!ptl)2353		return 0;2354 2355	/*2356	 * Can't clear PUD or it can race with concurrent zapping.  See2357	 * change_huge_pmd().2358	 */2359	oldpud = pudp_invalidate(vma, addr, pudp);2360	entry = pud_modify(oldpud, newprot);2361	set_pud_at(mm, addr, pudp, entry);2362	tlb_flush_pud_range(tlb, addr, HPAGE_PUD_SIZE);2363 2364	spin_unlock(ptl);2365	return HPAGE_PUD_NR;2366}2367#endif2368 2369#ifdef CONFIG_USERFAULTFD2370/*2371 * The PT lock for src_pmd and dst_vma/src_vma (for reading) are locked by2372 * the caller, but it must return after releasing the page_table_lock.2373 * Just move the page from src_pmd to dst_pmd if possible.2374 * Return zero if succeeded in moving the page, -EAGAIN if it needs to be2375 * repeated by the caller, or other errors in case of failure.2376 */2377int move_pages_huge_pmd(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd, pmd_t dst_pmdval,2378			struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,2379			unsigned long dst_addr, unsigned long src_addr)2380{2381	pmd_t _dst_pmd, src_pmdval;2382	struct page *src_page;2383	struct folio *src_folio;2384	struct anon_vma *src_anon_vma;2385	spinlock_t *src_ptl, *dst_ptl;2386	pgtable_t src_pgtable;2387	struct mmu_notifier_range range;2388	int err = 0;2389 2390	src_pmdval = *src_pmd;2391	src_ptl = pmd_lockptr(mm, src_pmd);2392 2393	lockdep_assert_held(src_ptl);2394	vma_assert_locked(src_vma);2395	vma_assert_locked(dst_vma);2396 2397	/* Sanity checks before the operation */2398	if (WARN_ON_ONCE(!pmd_none(dst_pmdval)) || WARN_ON_ONCE(src_addr & ~HPAGE_PMD_MASK) ||2399	    WARN_ON_ONCE(dst_addr & ~HPAGE_PMD_MASK)) {2400		spin_unlock(src_ptl);2401		return -EINVAL;2402	}2403 2404	if (!pmd_trans_huge(src_pmdval)) {2405		spin_unlock(src_ptl);2406		if (is_pmd_migration_entry(src_pmdval)) {2407			pmd_migration_entry_wait(mm, &src_pmdval);2408			return -EAGAIN;2409		}2410		return -ENOENT;2411	}2412 2413	src_page = pmd_page(src_pmdval);2414 2415	if (!is_huge_zero_pmd(src_pmdval)) {2416		if (unlikely(!PageAnonExclusive(src_page))) {2417			spin_unlock(src_ptl);2418			return -EBUSY;2419		}2420 2421		src_folio = page_folio(src_page);2422		folio_get(src_folio);2423	} else2424		src_folio = NULL;2425 2426	spin_unlock(src_ptl);2427 2428	flush_cache_range(src_vma, src_addr, src_addr + HPAGE_PMD_SIZE);2429	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, src_addr,2430				src_addr + HPAGE_PMD_SIZE);2431	mmu_notifier_invalidate_range_start(&range);2432 2433	if (src_folio) {2434		folio_lock(src_folio);2435 2436		/*2437		 * split_huge_page walks the anon_vma chain without the page2438		 * lock. Serialize against it with the anon_vma lock, the page2439		 * lock is not enough.2440		 */2441		src_anon_vma = folio_get_anon_vma(src_folio);2442		if (!src_anon_vma) {2443			err = -EAGAIN;2444			goto unlock_folio;2445		}2446		anon_vma_lock_write(src_anon_vma);2447	} else2448		src_anon_vma = NULL;2449 2450	dst_ptl = pmd_lockptr(mm, dst_pmd);2451	double_pt_lock(src_ptl, dst_ptl);2452	if (unlikely(!pmd_same(*src_pmd, src_pmdval) ||2453		     !pmd_same(*dst_pmd, dst_pmdval))) {2454		err = -EAGAIN;2455		goto unlock_ptls;2456	}2457	if (src_folio) {2458		if (folio_maybe_dma_pinned(src_folio) ||2459		    !PageAnonExclusive(&src_folio->page)) {2460			err = -EBUSY;2461			goto unlock_ptls;2462		}2463 2464		if (WARN_ON_ONCE(!folio_test_head(src_folio)) ||2465		    WARN_ON_ONCE(!folio_test_anon(src_folio))) {2466			err = -EBUSY;2467			goto unlock_ptls;2468		}2469 2470		src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);2471		/* Folio got pinned from under us. Put it back and fail the move. */2472		if (folio_maybe_dma_pinned(src_folio)) {2473			set_pmd_at(mm, src_addr, src_pmd, src_pmdval);2474			err = -EBUSY;2475			goto unlock_ptls;2476		}2477 2478		folio_move_anon_rmap(src_folio, dst_vma);2479		src_folio->index = linear_page_index(dst_vma, dst_addr);2480 2481		_dst_pmd = mk_huge_pmd(&src_folio->page, dst_vma->vm_page_prot);2482		/* Follow mremap() behavior and treat the entry dirty after the move */2483		_dst_pmd = pmd_mkwrite(pmd_mkdirty(_dst_pmd), dst_vma);2484	} else {2485		src_pmdval = pmdp_huge_clear_flush(src_vma, src_addr, src_pmd);2486		_dst_pmd = mk_huge_pmd(src_page, dst_vma->vm_page_prot);2487	}2488	set_pmd_at(mm, dst_addr, dst_pmd, _dst_pmd);2489 2490	src_pgtable = pgtable_trans_huge_withdraw(mm, src_pmd);2491	pgtable_trans_huge_deposit(mm, dst_pmd, src_pgtable);2492unlock_ptls:2493	double_pt_unlock(src_ptl, dst_ptl);2494	if (src_anon_vma) {2495		anon_vma_unlock_write(src_anon_vma);2496		put_anon_vma(src_anon_vma);2497	}2498unlock_folio:2499	/* unblock rmap walks */2500	if (src_folio)2501		folio_unlock(src_folio);2502	mmu_notifier_invalidate_range_end(&range);2503	if (src_folio)2504		folio_put(src_folio);2505	return err;2506}2507#endif /* CONFIG_USERFAULTFD */2508 2509/*2510 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.2511 *2512 * Note that if it returns page table lock pointer, this routine returns without2513 * unlocking page table lock. So callers must unlock it.2514 */2515spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)2516{2517	spinlock_t *ptl;2518	ptl = pmd_lock(vma->vm_mm, pmd);2519	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||2520			pmd_devmap(*pmd)))2521		return ptl;2522	spin_unlock(ptl);2523	return NULL;2524}2525 2526/*2527 * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.2528 *2529 * Note that if it returns page table lock pointer, this routine returns without2530 * unlocking page table lock. So callers must unlock it.2531 */2532spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)2533{2534	spinlock_t *ptl;2535 2536	ptl = pud_lock(vma->vm_mm, pud);2537	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))2538		return ptl;2539	spin_unlock(ptl);2540	return NULL;2541}2542 2543#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD2544int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,2545		 pud_t *pud, unsigned long addr)2546{2547	spinlock_t *ptl;2548	pud_t orig_pud;2549 2550	ptl = __pud_trans_huge_lock(pud, vma);2551	if (!ptl)2552		return 0;2553 2554	orig_pud = pudp_huge_get_and_clear_full(vma, addr, pud, tlb->fullmm);2555	arch_check_zapped_pud(vma, orig_pud);2556	tlb_remove_pud_tlb_entry(tlb, pud, addr);2557	if (vma_is_special_huge(vma)) {2558		spin_unlock(ptl);2559		/* No zero page support yet */2560	} else {2561		/* No support for anonymous PUD pages yet */2562		BUG();2563	}2564	return 1;2565}2566 2567static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,2568		unsigned long haddr)2569{2570	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);2571	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);2572	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);2573	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));2574 2575	count_vm_event(THP_SPLIT_PUD);2576 2577	pudp_huge_clear_flush(vma, haddr, pud);2578}2579 2580void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,2581		unsigned long address)2582{2583	spinlock_t *ptl;2584	struct mmu_notifier_range range;2585 2586	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,2587				address & HPAGE_PUD_MASK,2588				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);2589	mmu_notifier_invalidate_range_start(&range);2590	ptl = pud_lock(vma->vm_mm, pud);2591	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))2592		goto out;2593	__split_huge_pud_locked(vma, pud, range.start);2594 2595out:2596	spin_unlock(ptl);2597	mmu_notifier_invalidate_range_end(&range);2598}2599#else2600void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,2601		unsigned long address)2602{2603}2604#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */2605 2606static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,2607		unsigned long haddr, pmd_t *pmd)2608{2609	struct mm_struct *mm = vma->vm_mm;2610	pgtable_t pgtable;2611	pmd_t _pmd, old_pmd;2612	unsigned long addr;2613	pte_t *pte;2614	int i;2615 2616	/*2617	 * Leave pmd empty until pte is filled note that it is fine to delay2618	 * notification until mmu_notifier_invalidate_range_end() as we are2619	 * replacing a zero pmd write protected page with a zero pte write2620	 * protected page.2621	 *2622	 * See Documentation/mm/mmu_notifier.rst2623	 */2624	old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);2625 2626	pgtable = pgtable_trans_huge_withdraw(mm, pmd);2627	pmd_populate(mm, &_pmd, pgtable);2628 2629	pte = pte_offset_map(&_pmd, haddr);2630	VM_BUG_ON(!pte);2631	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {2632		pte_t entry;2633 2634		entry = pfn_pte(my_zero_pfn(addr), vma->vm_page_prot);2635		entry = pte_mkspecial(entry);2636		if (pmd_uffd_wp(old_pmd))2637			entry = pte_mkuffd_wp(entry);2638		VM_BUG_ON(!pte_none(ptep_get(pte)));2639		set_pte_at(mm, addr, pte, entry);2640		pte++;2641	}2642	pte_unmap(pte - 1);2643	smp_wmb(); /* make pte visible before pmd */2644	pmd_populate(mm, pmd, pgtable);2645}2646 2647static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,2648		unsigned long haddr, bool freeze)2649{2650	struct mm_struct *mm = vma->vm_mm;2651	struct folio *folio;2652	struct page *page;2653	pgtable_t pgtable;2654	pmd_t old_pmd, _pmd;2655	bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;2656	bool anon_exclusive = false, dirty = false;2657	unsigned long addr;2658	pte_t *pte;2659	int i;2660 2661	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);2662	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);2663	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);2664	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)2665				&& !pmd_devmap(*pmd));2666 2667	count_vm_event(THP_SPLIT_PMD);2668 2669	if (!vma_is_anonymous(vma)) {2670		old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);2671		/*2672		 * We are going to unmap this huge page. So2673		 * just go ahead and zap it2674		 */2675		if (arch_needs_pgtable_deposit())2676			zap_deposited_table(mm, pmd);2677		if (vma_is_special_huge(vma))2678			return;2679		if (unlikely(is_pmd_migration_entry(old_pmd))) {2680			swp_entry_t entry;2681 2682			entry = pmd_to_swp_entry(old_pmd);2683			folio = pfn_swap_entry_folio(entry);2684		} else {2685			page = pmd_page(old_pmd);2686			folio = page_folio(page);2687			if (!folio_test_dirty(folio) && pmd_dirty(old_pmd))2688				folio_mark_dirty(folio);2689			if (!folio_test_referenced(folio) && pmd_young(old_pmd))2690				folio_set_referenced(folio);2691			folio_remove_rmap_pmd(folio, page, vma);2692			folio_put(folio);2693		}2694		add_mm_counter(mm, mm_counter_file(folio), -HPAGE_PMD_NR);2695		return;2696	}2697 2698	if (is_huge_zero_pmd(*pmd)) {2699		/*2700		 * FIXME: Do we want to invalidate secondary mmu by calling2701		 * mmu_notifier_arch_invalidate_secondary_tlbs() see comments below2702		 * inside __split_huge_pmd() ?2703		 *2704		 * We are going from a zero huge page write protected to zero2705		 * small page also write protected so it does not seems useful2706		 * to invalidate secondary mmu at this time.2707		 */2708		return __split_huge_zero_page_pmd(vma, haddr, pmd);2709	}2710 2711	pmd_migration = is_pmd_migration_entry(*pmd);2712	if (unlikely(pmd_migration)) {2713		swp_entry_t entry;2714 2715		old_pmd = *pmd;2716		entry = pmd_to_swp_entry(old_pmd);2717		page = pfn_swap_entry_to_page(entry);2718		write = is_writable_migration_entry(entry);2719		if (PageAnon(page))2720			anon_exclusive = is_readable_exclusive_migration_entry(entry);2721		young = is_migration_entry_young(entry);2722		dirty = is_migration_entry_dirty(entry);2723		soft_dirty = pmd_swp_soft_dirty(old_pmd);2724		uffd_wp = pmd_swp_uffd_wp(old_pmd);2725	} else {2726		/*2727		 * Up to this point the pmd is present and huge and userland has2728		 * the whole access to the hugepage during the split (which2729		 * happens in place). If we overwrite the pmd with the not-huge2730		 * version pointing to the pte here (which of course we could if2731		 * all CPUs were bug free), userland could trigger a small page2732		 * size TLB miss on the small sized TLB while the hugepage TLB2733		 * entry is still established in the huge TLB. Some CPU doesn't2734		 * like that. See2735		 * http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum2736		 * 383 on page 105. Intel should be safe but is also warns that2737		 * it's only safe if the permission and cache attributes of the2738		 * two entries loaded in the two TLB is identical (which should2739		 * be the case here). But it is generally safer to never allow2740		 * small and huge TLB entries for the same virtual address to be2741		 * loaded simultaneously. So instead of doing "pmd_populate();2742		 * flush_pmd_tlb_range();" we first mark the current pmd2743		 * notpresent (atomically because here the pmd_trans_huge must2744		 * remain set at all times on the pmd until the split is2745		 * complete for this pmd), then we flush the SMP TLB and finally2746		 * we write the non-huge version of the pmd entry with2747		 * pmd_populate.2748		 */2749		old_pmd = pmdp_invalidate(vma, haddr, pmd);2750		page = pmd_page(old_pmd);2751		folio = page_folio(page);2752		if (pmd_dirty(old_pmd)) {2753			dirty = true;2754			folio_set_dirty(folio);2755		}2756		write = pmd_write(old_pmd);2757		young = pmd_young(old_pmd);2758		soft_dirty = pmd_soft_dirty(old_pmd);2759		uffd_wp = pmd_uffd_wp(old_pmd);2760 2761		VM_WARN_ON_FOLIO(!folio_ref_count(folio), folio);2762		VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);2763 2764		/*2765		 * Without "freeze", we'll simply split the PMD, propagating the2766		 * PageAnonExclusive() flag for each PTE by setting it for2767		 * each subpage -- no need to (temporarily) clear.2768		 *2769		 * With "freeze" we want to replace mapped pages by2770		 * migration entries right away. This is only possible if we2771		 * managed to clear PageAnonExclusive() -- see2772		 * set_pmd_migration_entry().2773		 *2774		 * In case we cannot clear PageAnonExclusive(), split the PMD2775		 * only and let try_to_migrate_one() fail later.2776		 *2777		 * See folio_try_share_anon_rmap_pmd(): invalidate PMD first.2778		 */2779		anon_exclusive = PageAnonExclusive(page);2780		if (freeze && anon_exclusive &&2781		    folio_try_share_anon_rmap_pmd(folio, page))2782			freeze = false;2783		if (!freeze) {2784			rmap_t rmap_flags = RMAP_NONE;2785 2786			folio_ref_add(folio, HPAGE_PMD_NR - 1);2787			if (anon_exclusive)2788				rmap_flags |= RMAP_EXCLUSIVE;2789			folio_add_anon_rmap_ptes(folio, page, HPAGE_PMD_NR,2790						 vma, haddr, rmap_flags);2791		}2792	}2793 2794	/*2795	 * Withdraw the table only after we mark the pmd entry invalid.2796	 * This's critical for some architectures (Power).2797	 */2798	pgtable = pgtable_trans_huge_withdraw(mm, pmd);2799	pmd_populate(mm, &_pmd, pgtable);2800 2801	pte = pte_offset_map(&_pmd, haddr);2802	VM_BUG_ON(!pte);2803 2804	/*2805	 * Note that NUMA hinting access restrictions are not transferred to2806	 * avoid any possibility of altering permissions across VMAs.2807	 */2808	if (freeze || pmd_migration) {2809		for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {2810			pte_t entry;2811			swp_entry_t swp_entry;2812 2813			if (write)2814				swp_entry = make_writable_migration_entry(2815							page_to_pfn(page + i));2816			else if (anon_exclusive)2817				swp_entry = make_readable_exclusive_migration_entry(2818							page_to_pfn(page + i));2819			else2820				swp_entry = make_readable_migration_entry(2821							page_to_pfn(page + i));2822			if (young)2823				swp_entry = make_migration_entry_young(swp_entry);2824			if (dirty)2825				swp_entry = make_migration_entry_dirty(swp_entry);2826			entry = swp_entry_to_pte(swp_entry);2827			if (soft_dirty)2828				entry = pte_swp_mksoft_dirty(entry);2829			if (uffd_wp)2830				entry = pte_swp_mkuffd_wp(entry);2831 2832			VM_WARN_ON(!pte_none(ptep_get(pte + i)));2833			set_pte_at(mm, addr, pte + i, entry);2834		}2835	} else {2836		pte_t entry;2837 2838		entry = mk_pte(page, READ_ONCE(vma->vm_page_prot));2839		if (write)2840			entry = pte_mkwrite(entry, vma);2841		if (!young)2842			entry = pte_mkold(entry);2843		/* NOTE: this may set soft-dirty too on some archs */2844		if (dirty)2845			entry = pte_mkdirty(entry);2846		if (soft_dirty)2847			entry = pte_mksoft_dirty(entry);2848		if (uffd_wp)2849			entry = pte_mkuffd_wp(entry);2850 2851		for (i = 0; i < HPAGE_PMD_NR; i++)2852			VM_WARN_ON(!pte_none(ptep_get(pte + i)));2853 2854		set_ptes(mm, haddr, pte, entry, HPAGE_PMD_NR);2855	}2856	pte_unmap(pte);2857 2858	if (!pmd_migration)2859		folio_remove_rmap_pmd(folio, page, vma);2860	if (freeze)2861		put_page(page);2862 2863	smp_wmb(); /* make pte visible before pmd */2864	pmd_populate(mm, pmd, pgtable);2865}2866 2867void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,2868			   pmd_t *pmd, bool freeze, struct folio *folio)2869{2870	VM_WARN_ON_ONCE(folio && !folio_test_pmd_mappable(folio));2871	VM_WARN_ON_ONCE(!IS_ALIGNED(address, HPAGE_PMD_SIZE));2872	VM_WARN_ON_ONCE(folio && !folio_test_locked(folio));2873	VM_BUG_ON(freeze && !folio);2874 2875	/*2876	 * When the caller requests to set up a migration entry, we2877	 * require a folio to check the PMD against. Otherwise, there2878	 * is a risk of replacing the wrong folio.2879	 */2880	if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) ||2881	    is_pmd_migration_entry(*pmd)) {2882		if (folio && folio != pmd_folio(*pmd))2883			return;2884		__split_huge_pmd_locked(vma, pmd, address, freeze);2885	}2886}2887 2888void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,2889		unsigned long address, bool freeze, struct folio *folio)2890{2891	spinlock_t *ptl;2892	struct mmu_notifier_range range;2893 2894	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,2895				address & HPAGE_PMD_MASK,2896				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);2897	mmu_notifier_invalidate_range_start(&range);2898	ptl = pmd_lock(vma->vm_mm, pmd);2899	split_huge_pmd_locked(vma, range.start, pmd, freeze, folio);2900	spin_unlock(ptl);2901	mmu_notifier_invalidate_range_end(&range);2902}2903 2904void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,2905		bool freeze, struct folio *folio)2906{2907	pmd_t *pmd = mm_find_pmd(vma->vm_mm, address);2908 2909	if (!pmd)2910		return;2911 2912	__split_huge_pmd(vma, pmd, address, freeze, folio);2913}2914 2915static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)2916{2917	/*2918	 * If the new address isn't hpage aligned and it could previously2919	 * contain an hugepage: check if we need to split an huge pmd.2920	 */2921	if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&2922	    range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),2923			 ALIGN(address, HPAGE_PMD_SIZE)))2924		split_huge_pmd_address(vma, address, false, NULL);2925}2926 2927void vma_adjust_trans_huge(struct vm_area_struct *vma,2928			     unsigned long start,2929			     unsigned long end,2930			     long adjust_next)2931{2932	/* Check if we need to split start first. */2933	split_huge_pmd_if_needed(vma, start);2934 2935	/* Check if we need to split end next. */2936	split_huge_pmd_if_needed(vma, end);2937 2938	/*2939	 * If we're also updating the next vma vm_start,2940	 * check if we need to split it.2941	 */2942	if (adjust_next > 0) {2943		struct vm_area_struct *next = find_vma(vma->vm_mm, vma->vm_end);2944		unsigned long nstart = next->vm_start;2945		nstart += adjust_next;2946		split_huge_pmd_if_needed(next, nstart);2947	}2948}2949 2950static void unmap_folio(struct folio *folio)2951{2952	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SYNC |2953		TTU_BATCH_FLUSH;2954 2955	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);2956 2957	if (folio_test_pmd_mappable(folio))2958		ttu_flags |= TTU_SPLIT_HUGE_PMD;2959 2960	/*2961	 * Anon pages need migration entries to preserve them, but file2962	 * pages can simply be left unmapped, then faulted back on demand.2963	 * If that is ever changed (perhaps for mlock), update remap_page().2964	 */2965	if (folio_test_anon(folio))2966		try_to_migrate(folio, ttu_flags);2967	else2968		try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);2969 2970	try_to_unmap_flush();2971}2972 2973static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma,2974					    unsigned long addr, pmd_t *pmdp,2975					    struct folio *folio)2976{2977	struct mm_struct *mm = vma->vm_mm;2978	int ref_count, map_count;2979	pmd_t orig_pmd = *pmdp;2980 2981	if (folio_test_dirty(folio) || pmd_dirty(orig_pmd))2982		return false;2983 2984	orig_pmd = pmdp_huge_clear_flush(vma, addr, pmdp);2985 2986	/*2987	 * Syncing against concurrent GUP-fast:2988	 * - clear PMD; barrier; read refcount2989	 * - inc refcount; barrier; read PMD2990	 */2991	smp_mb();2992 2993	ref_count = folio_ref_count(folio);2994	map_count = folio_mapcount(folio);2995 2996	/*2997	 * Order reads for folio refcount and dirty flag2998	 * (see comments in __remove_mapping()).2999	 */3000	smp_rmb();3001 3002	/*3003	 * If the folio or its PMD is redirtied at this point, or if there3004	 * are unexpected references, we will give up to discard this folio3005	 * and remap it.3006	 *3007	 * The only folio refs must be one from isolation plus the rmap(s).3008	 */3009	if (folio_test_dirty(folio) || pmd_dirty(orig_pmd) ||3010	    ref_count != map_count + 1) {3011		set_pmd_at(mm, addr, pmdp, orig_pmd);3012		return false;3013	}3014 3015	folio_remove_rmap_pmd(folio, pmd_page(orig_pmd), vma);3016	zap_deposited_table(mm, pmdp);3017	add_mm_counter(mm, MM_ANONPAGES, -HPAGE_PMD_NR);3018	if (vma->vm_flags & VM_LOCKED)3019		mlock_drain_local();3020	folio_put(folio);3021 3022	return true;3023}3024 3025bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,3026			   pmd_t *pmdp, struct folio *folio)3027{3028	VM_WARN_ON_FOLIO(!folio_test_pmd_mappable(folio), folio);3029	VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);3030	VM_WARN_ON_ONCE(!IS_ALIGNED(addr, HPAGE_PMD_SIZE));3031 3032	if (folio_test_anon(folio) && !folio_test_swapbacked(folio))3033		return __discard_anon_folio_pmd_locked(vma, addr, pmdp, folio);3034 3035	return false;3036}3037 3038static void remap_page(struct folio *folio, unsigned long nr, int flags)3039{3040	int i = 0;3041 3042	/* If unmap_folio() uses try_to_migrate() on file, remove this check */3043	if (!folio_test_anon(folio))3044		return;3045	for (;;) {3046		remove_migration_ptes(folio, folio, RMP_LOCKED | flags);3047		i += folio_nr_pages(folio);3048		if (i >= nr)3049			break;3050		folio = folio_next(folio);3051	}3052}3053 3054static void lru_add_page_tail(struct folio *folio, struct page *tail,3055		struct lruvec *lruvec, struct list_head *list)3056{3057	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);3058	VM_BUG_ON_FOLIO(PageLRU(tail), folio);3059	lockdep_assert_held(&lruvec->lru_lock);3060 3061	if (list) {3062		/* page reclaim is reclaiming a huge page */3063		VM_WARN_ON(folio_test_lru(folio));3064		get_page(tail);3065		list_add_tail(&tail->lru, list);3066	} else {3067		/* head is still on lru (and we have it frozen) */3068		VM_WARN_ON(!folio_test_lru(folio));3069		if (folio_test_unevictable(folio))3070			tail->mlock_count = 0;3071		else3072			list_add_tail(&tail->lru, &folio->lru);3073		SetPageLRU(tail);3074	}3075}3076 3077static void __split_huge_page_tail(struct folio *folio, int tail,3078		struct lruvec *lruvec, struct list_head *list,3079		unsigned int new_order)3080{3081	struct page *head = &folio->page;3082	struct page *page_tail = head + tail;3083	/*3084	 * Careful: new_folio is not a "real" folio before we cleared PageTail.3085	 * Don't pass it around before clear_compound_head().3086	 */3087	struct folio *new_folio = (struct folio *)page_tail;3088 3089	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);3090 3091	/*3092	 * Clone page flags before unfreezing refcount.3093	 *3094	 * After successful get_page_unless_zero() might follow flags change,3095	 * for example lock_page() which set PG_waiters.3096	 *3097	 * Note that for mapped sub-pages of an anonymous THP,3098	 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in3099	 * the migration entry instead from where remap_page() will restore it.3100	 * We can still have PG_anon_exclusive set on effectively unmapped and3101	 * unreferenced sub-pages of an anonymous THP: we can simply drop3102	 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.3103	 */3104	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;3105	page_tail->flags |= (head->flags &3106			((1L << PG_referenced) |3107			 (1L << PG_swapbacked) |3108			 (1L << PG_swapcache) |3109			 (1L << PG_mlocked) |3110			 (1L << PG_uptodate) |3111			 (1L << PG_active) |3112			 (1L << PG_workingset) |3113			 (1L << PG_locked) |3114			 (1L << PG_unevictable) |3115#ifdef CONFIG_ARCH_USES_PG_ARCH_23116			 (1L << PG_arch_2) |3117#endif3118#ifdef CONFIG_ARCH_USES_PG_ARCH_33119			 (1L << PG_arch_3) |3120#endif3121			 (1L << PG_dirty) |3122			 LRU_GEN_MASK | LRU_REFS_MASK));3123 3124	/* ->mapping in first and second tail page is replaced by other uses */3125	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,3126			page_tail);3127	page_tail->mapping = head->mapping;3128	page_tail->index = head->index + tail;3129 3130	/*3131	 * page->private should not be set in tail pages. Fix up and warn once3132	 * if private is unexpectedly set.3133	 */3134	if (unlikely(page_tail->private)) {3135		VM_WARN_ON_ONCE_PAGE(true, page_tail);3136		page_tail->private = 0;3137	}3138	if (folio_test_swapcache(folio))3139		new_folio->swap.val = folio->swap.val + tail;3140 3141	/* Page flags must be visible before we make the page non-compound. */3142	smp_wmb();3143 3144	/*3145	 * Clear PageTail before unfreezing page refcount.3146	 *3147	 * After successful get_page_unless_zero() might follow put_page()3148	 * which needs correct compound_head().3149	 */3150	clear_compound_head(page_tail);3151	if (new_order) {3152		prep_compound_page(page_tail, new_order);3153		folio_set_large_rmappable(new_folio);3154	}3155 3156	/* Finally unfreeze refcount. Additional reference from page cache. */3157	page_ref_unfreeze(page_tail,3158		1 + ((!folio_test_anon(folio) || folio_test_swapcache(folio)) ?3159			     folio_nr_pages(new_folio) : 0));3160 3161	if (folio_test_young(folio))3162		folio_set_young(new_folio);3163	if (folio_test_idle(folio))3164		folio_set_idle(new_folio);3165 3166	folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));3167 3168	/*3169	 * always add to the tail because some iterators expect new3170	 * pages to show after the currently processed elements - e.g.3171	 * migrate_pages3172	 */3173	lru_add_page_tail(folio, page_tail, lruvec, list);3174}3175 3176static void __split_huge_page(struct page *page, struct list_head *list,3177		pgoff_t end, unsigned int new_order)3178{3179	struct folio *folio = page_folio(page);3180	struct page *head = &folio->page;3181	struct lruvec *lruvec;3182	struct address_space *swap_cache = NULL;3183	unsigned long offset = 0;3184	int i, nr_dropped = 0;3185	unsigned int new_nr = 1 << new_order;3186	int order = folio_order(folio);3187	unsigned int nr = 1 << order;3188 3189	/* complete memcg works before add pages to LRU */3190	split_page_memcg(head, order, new_order);3191 3192	if (folio_test_anon(folio) && folio_test_swapcache(folio)) {3193		offset = swap_cache_index(folio->swap);3194		swap_cache = swap_address_space(folio->swap);3195		xa_lock(&swap_cache->i_pages);3196	}3197 3198	/* lock lru list/PageCompound, ref frozen by page_ref_freeze */3199	lruvec = folio_lruvec_lock(folio);3200 3201	ClearPageHasHWPoisoned(head);3202 3203	for (i = nr - new_nr; i >= new_nr; i -= new_nr) {3204		__split_huge_page_tail(folio, i, lruvec, list, new_order);3205		/* Some pages can be beyond EOF: drop them from page cache */3206		if (head[i].index >= end) {3207			struct folio *tail = page_folio(head + i);3208 3209			if (shmem_mapping(folio->mapping))3210				nr_dropped++;3211			else if (folio_test_clear_dirty(tail))3212				folio_account_cleaned(tail,3213					inode_to_wb(folio->mapping->host));3214			__filemap_remove_folio(tail, NULL);3215			folio_put(tail);3216		} else if (!PageAnon(page)) {3217			__xa_store(&folio->mapping->i_pages, head[i].index,3218					head + i, 0);3219		} else if (swap_cache) {3220			__xa_store(&swap_cache->i_pages, offset + i,3221					head + i, 0);3222		}3223	}3224 3225	if (!new_order)3226		ClearPageCompound(head);3227	else {3228		struct folio *new_folio = (struct folio *)head;3229 3230		folio_set_order(new_folio, new_order);3231	}3232	unlock_page_lruvec(lruvec);3233	/* Caller disabled irqs, so they are still disabled here */3234 3235	split_page_owner(head, order, new_order);3236	pgalloc_tag_split(folio, order, new_order);3237 3238	/* See comment in __split_huge_page_tail() */3239	if (folio_test_anon(folio)) {3240		/* Additional pin to swap cache */3241		if (folio_test_swapcache(folio)) {3242			folio_ref_add(folio, 1 + new_nr);3243			xa_unlock(&swap_cache->i_pages);3244		} else {3245			folio_ref_inc(folio);3246		}3247	} else {3248		/* Additional pin to page cache */3249		folio_ref_add(folio, 1 + new_nr);3250		xa_unlock(&folio->mapping->i_pages);3251	}3252	local_irq_enable();3253 3254	if (nr_dropped)3255		shmem_uncharge(folio->mapping->host, nr_dropped);3256	remap_page(folio, nr, PageAnon(head) ? RMP_USE_SHARED_ZEROPAGE : 0);3257 3258	/*3259	 * set page to its compound_head when split to non order-0 pages, so3260	 * we can skip unlocking it below, since PG_locked is transferred to3261	 * the compound_head of the page and the caller will unlock it.3262	 */3263	if (new_order)3264		page = compound_head(page);3265 3266	for (i = 0; i < nr; i += new_nr) {3267		struct page *subpage = head + i;3268		struct folio *new_folio = page_folio(subpage);3269		if (subpage == page)3270			continue;3271		folio_unlock(new_folio);3272 3273		/*3274		 * Subpages may be freed if there wasn't any mapping3275		 * like if add_to_swap() is running on a lru page that3276		 * had its mapping zapped. And freeing these pages3277		 * requires taking the lru_lock so we do the put_page3278		 * of the tail pages after the split is complete.3279		 */3280		free_page_and_swap_cache(subpage);3281	}3282}3283 3284/* Racy check whether the huge page can be split */3285bool can_split_folio(struct folio *folio, int caller_pins, int *pextra_pins)3286{3287	int extra_pins;3288 3289	/* Additional pins from page cache */3290	if (folio_test_anon(folio))3291		extra_pins = folio_test_swapcache(folio) ?3292				folio_nr_pages(folio) : 0;3293	else3294		extra_pins = folio_nr_pages(folio);3295	if (pextra_pins)3296		*pextra_pins = extra_pins;3297	return folio_mapcount(folio) == folio_ref_count(folio) - extra_pins -3298					caller_pins;3299}3300 3301/*3302 * This function splits a large folio into smaller folios of order @new_order.3303 * @page can point to any page of the large folio to split. The split operation3304 * does not change the position of @page.3305 *3306 * Prerequisites:3307 *3308 * 1) The caller must hold a reference on the @page's owning folio, also known3309 *    as the large folio.3310 *3311 * 2) The large folio must be locked.3312 *3313 * 3) The folio must not be pinned. Any unexpected folio references, including3314 *    GUP pins, will result in the folio not getting split; instead, the caller3315 *    will receive an -EAGAIN.3316 *3317 * 4) @new_order > 1, usually. Splitting to order-1 anonymous folios is not3318 *    supported for non-file-backed folios, because folio->_deferred_list, which3319 *    is used by partially mapped folios, is stored in subpage 2, but an order-13320 *    folio only has subpages 0 and 1. File-backed order-1 folios are supported,3321 *    since they do not use _deferred_list.3322 *3323 * After splitting, the caller's folio reference will be transferred to @page,3324 * resulting in a raised refcount of @page after this call. The other pages may3325 * be freed if they are not mapped.3326 *3327 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.3328 *3329 * Pages in @new_order will inherit the mapping, flags, and so on from the3330 * huge page.3331 *3332 * Returns 0 if the huge page was split successfully.3333 *3334 * Returns -EAGAIN if the folio has unexpected reference (e.g., GUP) or if3335 * the folio was concurrently removed from the page cache.3336 *3337 * Returns -EBUSY when trying to split the huge zeropage, if the folio is3338 * under writeback, if fs-specific folio metadata cannot currently be3339 * released, or if some unexpected race happened (e.g., anon VMA disappeared,3340 * truncation).3341 *3342 * Callers should ensure that the order respects the address space mapping3343 * min-order if one is set for non-anonymous folios.3344 *3345 * Returns -EINVAL when trying to split to an order that is incompatible3346 * with the folio. Splitting to order 0 is compatible with all folios.3347 */3348int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,3349				     unsigned int new_order)3350{3351	struct folio *folio = page_folio(page);3352	struct deferred_split *ds_queue = get_deferred_split_queue(folio);3353	/* reset xarray order to new order after split */3354	XA_STATE_ORDER(xas, &folio->mapping->i_pages, folio->index, new_order);3355	bool is_anon = folio_test_anon(folio);3356	struct address_space *mapping = NULL;3357	struct anon_vma *anon_vma = NULL;3358	int order = folio_order(folio);3359	int extra_pins, ret;3360	pgoff_t end;3361	bool is_hzp;3362 3363	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);3364	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);3365 3366	if (new_order >= folio_order(folio))3367		return -EINVAL;3368 3369	if (is_anon) {3370		/* order-1 is not supported for anonymous THP. */3371		if (new_order == 1) {3372			VM_WARN_ONCE(1, "Cannot split to order-1 folio");3373			return -EINVAL;3374		}3375	} else if (new_order) {3376		/* Split shmem folio to non-zero order not supported */3377		if (shmem_mapping(folio->mapping)) {3378			VM_WARN_ONCE(1,3379				"Cannot split shmem folio to non-0 order");3380			return -EINVAL;3381		}3382		/*3383		 * No split if the file system does not support large folio.3384		 * Note that we might still have THPs in such mappings due to3385		 * CONFIG_READ_ONLY_THP_FOR_FS. But in that case, the mapping3386		 * does not actually support large folios properly.3387		 */3388		if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&3389		    !mapping_large_folio_support(folio->mapping)) {3390			VM_WARN_ONCE(1,3391				"Cannot split file folio to non-0 order");3392			return -EINVAL;3393		}3394	}3395 3396	/* Only swapping a whole PMD-mapped folio is supported */3397	if (folio_test_swapcache(folio) && new_order)3398		return -EINVAL;3399 3400	is_hzp = is_huge_zero_folio(folio);3401	if (is_hzp) {3402		pr_warn_ratelimited("Called split_huge_page for huge zero page\n");3403		return -EBUSY;3404	}3405 3406	if (folio_test_writeback(folio))3407		return -EBUSY;3408 3409	if (is_anon) {3410		/*3411		 * The caller does not necessarily hold an mmap_lock that would3412		 * prevent the anon_vma disappearing so we first we take a3413		 * reference to it and then lock the anon_vma for write. This3414		 * is similar to folio_lock_anon_vma_read except the write lock3415		 * is taken to serialise against parallel split or collapse3416		 * operations.3417		 */3418		anon_vma = folio_get_anon_vma(folio);3419		if (!anon_vma) {3420			ret = -EBUSY;3421			goto out;3422		}3423		end = -1;3424		mapping = NULL;3425		anon_vma_lock_write(anon_vma);3426	} else {3427		unsigned int min_order;3428		gfp_t gfp;3429 3430		mapping = folio->mapping;3431 3432		/* Truncated ? */3433		if (!mapping) {3434			ret = -EBUSY;3435			goto out;3436		}3437 3438		min_order = mapping_min_folio_order(folio->mapping);3439		if (new_order < min_order) {3440			VM_WARN_ONCE(1, "Cannot split mapped folio below min-order: %u",3441				     min_order);3442			ret = -EINVAL;3443			goto out;3444		}3445 3446		gfp = current_gfp_context(mapping_gfp_mask(mapping) &3447							GFP_RECLAIM_MASK);3448 3449		if (!filemap_release_folio(folio, gfp)) {3450			ret = -EBUSY;3451			goto out;3452		}3453 3454		xas_split_alloc(&xas, folio, folio_order(folio), gfp);3455		if (xas_error(&xas)) {3456			ret = xas_error(&xas);3457			goto out;3458		}3459 3460		anon_vma = NULL;3461		i_mmap_lock_read(mapping);3462 3463		/*3464		 *__split_huge_page() may need to trim off pages beyond EOF:3465		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,3466		 * which cannot be nested inside the page tree lock. So note3467		 * end now: i_size itself may be changed at any moment, but3468		 * folio lock is good enough to serialize the trimming.3469		 */3470		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);3471		if (shmem_mapping(mapping))3472			end = shmem_fallocend(mapping->host, end);3473	}3474 3475	/*3476	 * Racy check if we can split the page, before unmap_folio() will3477	 * split PMDs3478	 */3479	if (!can_split_folio(folio, 1, &extra_pins)) {3480		ret = -EAGAIN;3481		goto out_unlock;3482	}3483 3484	unmap_folio(folio);3485 3486	/* block interrupt reentry in xa_lock and spinlock */3487	local_irq_disable();3488	if (mapping) {3489		/*3490		 * Check if the folio is present in page cache.3491		 * We assume all tail are present too, if folio is there.3492		 */3493		xas_lock(&xas);3494		xas_reset(&xas);3495		if (xas_load(&xas) != folio)3496			goto fail;3497	}3498 3499	/* Prevent deferred_split_scan() touching ->_refcount */3500	spin_lock(&ds_queue->split_queue_lock);3501	if (folio_ref_freeze(folio, 1 + extra_pins)) {3502		if (folio_order(folio) > 1 &&3503		    !list_empty(&folio->_deferred_list)) {3504			ds_queue->split_queue_len--;3505			if (folio_test_partially_mapped(folio)) {3506				__folio_clear_partially_mapped(folio);3507				mod_mthp_stat(folio_order(folio),3508					      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);3509			}3510			/*3511			 * Reinitialize page_deferred_list after removing the3512			 * page from the split_queue, otherwise a subsequent3513			 * split will see list corruption when checking the3514			 * page_deferred_list.3515			 */3516			list_del_init(&folio->_deferred_list);3517		}3518		spin_unlock(&ds_queue->split_queue_lock);3519		if (mapping) {3520			int nr = folio_nr_pages(folio);3521 3522			xas_split(&xas, folio, folio_order(folio));3523			if (folio_test_pmd_mappable(folio) &&3524			    new_order < HPAGE_PMD_ORDER) {3525				if (folio_test_swapbacked(folio)) {3526					__lruvec_stat_mod_folio(folio,3527							NR_SHMEM_THPS, -nr);3528				} else {3529					__lruvec_stat_mod_folio(folio,3530							NR_FILE_THPS, -nr);3531					filemap_nr_thps_dec(mapping);3532				}3533			}3534		}3535 3536		if (is_anon) {3537			mod_mthp_stat(order, MTHP_STAT_NR_ANON, -1);3538			mod_mthp_stat(new_order, MTHP_STAT_NR_ANON, 1 << (order - new_order));3539		}3540		__split_huge_page(page, list, end, new_order);3541		ret = 0;3542	} else {3543		spin_unlock(&ds_queue->split_queue_lock);3544fail:3545		if (mapping)3546			xas_unlock(&xas);3547		local_irq_enable();3548		remap_page(folio, folio_nr_pages(folio), 0);3549		ret = -EAGAIN;3550	}3551 3552out_unlock:3553	if (anon_vma) {3554		anon_vma_unlock_write(anon_vma);3555		put_anon_vma(anon_vma);3556	}3557	if (mapping)3558		i_mmap_unlock_read(mapping);3559out:3560	xas_destroy(&xas);3561	if (order == HPAGE_PMD_ORDER)3562		count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);3563	count_mthp_stat(order, !ret ? MTHP_STAT_SPLIT : MTHP_STAT_SPLIT_FAILED);3564	return ret;3565}3566 3567int min_order_for_split(struct folio *folio)3568{3569	if (folio_test_anon(folio))3570		return 0;3571 3572	if (!folio->mapping) {3573		if (folio_test_pmd_mappable(folio))3574			count_vm_event(THP_SPLIT_PAGE_FAILED);3575		return -EBUSY;3576	}3577 3578	return mapping_min_folio_order(folio->mapping);3579}3580 3581int split_folio_to_list(struct folio *folio, struct list_head *list)3582{3583	int ret = min_order_for_split(folio);3584 3585	if (ret < 0)3586		return ret;3587 3588	return split_huge_page_to_list_to_order(&folio->page, list, ret);3589}3590 3591/*3592 * __folio_unqueue_deferred_split() is not to be called directly:3593 * the folio_unqueue_deferred_split() inline wrapper in mm/internal.h3594 * limits its calls to those folios which may have a _deferred_list for3595 * queueing THP splits, and that list is (racily observed to be) non-empty.3596 *3597 * It is unsafe to call folio_unqueue_deferred_split() until folio refcount is3598 * zero: because even when split_queue_lock is held, a non-empty _deferred_list3599 * might be in use on deferred_split_scan()'s unlocked on-stack list.3600 *3601 * If memory cgroups are enabled, split_queue_lock is in the mem_cgroup: it is3602 * therefore important to unqueue deferred split before changing folio memcg.3603 */3604bool __folio_unqueue_deferred_split(struct folio *folio)3605{3606	struct deferred_split *ds_queue;3607	unsigned long flags;3608	bool unqueued = false;3609 3610	WARN_ON_ONCE(folio_ref_count(folio));3611	WARN_ON_ONCE(!mem_cgroup_disabled() && !folio_memcg(folio));3612 3613	ds_queue = get_deferred_split_queue(folio);3614	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);3615	if (!list_empty(&folio->_deferred_list)) {3616		ds_queue->split_queue_len--;3617		if (folio_test_partially_mapped(folio)) {3618			__folio_clear_partially_mapped(folio);3619			mod_mthp_stat(folio_order(folio),3620				      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);3621		}3622		list_del_init(&folio->_deferred_list);3623		unqueued = true;3624	}3625	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);3626 3627	return unqueued;	/* useful for debug warnings */3628}3629 3630/* partially_mapped=false won't clear PG_partially_mapped folio flag */3631void deferred_split_folio(struct folio *folio, bool partially_mapped)3632{3633	struct deferred_split *ds_queue = get_deferred_split_queue(folio);3634#ifdef CONFIG_MEMCG3635	struct mem_cgroup *memcg = folio_memcg(folio);3636#endif3637	unsigned long flags;3638 3639	/*3640	 * Order 1 folios have no space for a deferred list, but we also3641	 * won't waste much memory by not adding them to the deferred list.3642	 */3643	if (folio_order(folio) <= 1)3644		return;3645 3646	if (!partially_mapped && !split_underused_thp)3647		return;3648 3649	/*3650	 * Exclude swapcache: originally to avoid a corrupt deferred split3651	 * queue. Nowadays that is fully prevented by mem_cgroup_swapout();3652	 * but if page reclaim is already handling the same folio, it is3653	 * unnecessary to handle it again in the shrinker, so excluding3654	 * swapcache here may still be a useful optimization.3655	 */3656	if (folio_test_swapcache(folio))3657		return;3658 3659	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);3660	if (partially_mapped) {3661		if (!folio_test_partially_mapped(folio)) {3662			__folio_set_partially_mapped(folio);3663			if (folio_test_pmd_mappable(folio))3664				count_vm_event(THP_DEFERRED_SPLIT_PAGE);3665			count_mthp_stat(folio_order(folio), MTHP_STAT_SPLIT_DEFERRED);3666			mod_mthp_stat(folio_order(folio), MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, 1);3667 3668		}3669	} else {3670		/* partially mapped folios cannot become non-partially mapped */3671		VM_WARN_ON_FOLIO(folio_test_partially_mapped(folio), folio);3672	}3673	if (list_empty(&folio->_deferred_list)) {3674		list_add_tail(&folio->_deferred_list, &ds_queue->split_queue);3675		ds_queue->split_queue_len++;3676#ifdef CONFIG_MEMCG3677		if (memcg)3678			set_shrinker_bit(memcg, folio_nid(folio),3679					 deferred_split_shrinker->id);3680#endif3681	}3682	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);3683}3684 3685static unsigned long deferred_split_count(struct shrinker *shrink,3686		struct shrink_control *sc)3687{3688	struct pglist_data *pgdata = NODE_DATA(sc->nid);3689	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;3690 3691#ifdef CONFIG_MEMCG3692	if (sc->memcg)3693		ds_queue = &sc->memcg->deferred_split_queue;3694#endif3695	return READ_ONCE(ds_queue->split_queue_len);3696}3697 3698static bool thp_underused(struct folio *folio)3699{3700	int num_zero_pages = 0, num_filled_pages = 0;3701	void *kaddr;3702	int i;3703 3704	if (khugepaged_max_ptes_none == HPAGE_PMD_NR - 1)3705		return false;3706 3707	for (i = 0; i < folio_nr_pages(folio); i++) {3708		kaddr = kmap_local_folio(folio, i * PAGE_SIZE);3709		if (!memchr_inv(kaddr, 0, PAGE_SIZE)) {3710			num_zero_pages++;3711			if (num_zero_pages > khugepaged_max_ptes_none) {3712				kunmap_local(kaddr);3713				return true;3714			}3715		} else {3716			/*3717			 * Another path for early exit once the number3718			 * of non-zero filled pages exceeds threshold.3719			 */3720			num_filled_pages++;3721			if (num_filled_pages >= HPAGE_PMD_NR - khugepaged_max_ptes_none) {3722				kunmap_local(kaddr);3723				return false;3724			}3725		}3726		kunmap_local(kaddr);3727	}3728	return false;3729}3730 3731static unsigned long deferred_split_scan(struct shrinker *shrink,3732		struct shrink_control *sc)3733{3734	struct pglist_data *pgdata = NODE_DATA(sc->nid);3735	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;3736	unsigned long flags;3737	LIST_HEAD(list);3738	struct folio *folio, *next, *prev = NULL;3739	int split = 0, removed = 0;3740 3741#ifdef CONFIG_MEMCG3742	if (sc->memcg)3743		ds_queue = &sc->memcg->deferred_split_queue;3744#endif3745 3746	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);3747	/* Take pin on all head pages to avoid freeing them under us */3748	list_for_each_entry_safe(folio, next, &ds_queue->split_queue,3749							_deferred_list) {3750		if (folio_try_get(folio)) {3751			list_move(&folio->_deferred_list, &list);3752		} else {3753			/* We lost race with folio_put() */3754			if (folio_test_partially_mapped(folio)) {3755				__folio_clear_partially_mapped(folio);3756				mod_mthp_stat(folio_order(folio),3757					      MTHP_STAT_NR_ANON_PARTIALLY_MAPPED, -1);3758			}3759			list_del_init(&folio->_deferred_list);3760			ds_queue->split_queue_len--;3761		}3762		if (!--sc->nr_to_scan)3763			break;3764	}3765	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);3766 3767	list_for_each_entry_safe(folio, next, &list, _deferred_list) {3768		bool did_split = false;3769		bool underused = false;3770 3771		if (!folio_test_partially_mapped(folio)) {3772			underused = thp_underused(folio);3773			if (!underused)3774				goto next;3775		}3776		if (!folio_trylock(folio))3777			goto next;3778		if (!split_folio(folio)) {3779			did_split = true;3780			if (underused)3781				count_vm_event(THP_UNDERUSED_SPLIT_PAGE);3782			split++;3783		}3784		folio_unlock(folio);3785next:3786		/*3787		 * split_folio() removes folio from list on success.3788		 * Only add back to the queue if folio is partially mapped.3789		 * If thp_underused returns false, or if split_folio fails3790		 * in the case it was underused, then consider it used and3791		 * don't add it back to split_queue.3792		 */3793		if (did_split) {3794			; /* folio already removed from list */3795		} else if (!folio_test_partially_mapped(folio)) {3796			list_del_init(&folio->_deferred_list);3797			removed++;3798		} else {3799			/*3800			 * That unlocked list_del_init() above would be unsafe,3801			 * unless its folio is separated from any earlier folios3802			 * left on the list (which may be concurrently unqueued)3803			 * by one safe folio with refcount still raised.3804			 */3805			swap(folio, prev);3806		}3807		if (folio)3808			folio_put(folio);3809	}3810 3811	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);3812	list_splice_tail(&list, &ds_queue->split_queue);3813	ds_queue->split_queue_len -= removed;3814	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);3815 3816	if (prev)3817		folio_put(prev);3818 3819	/*3820	 * Stop shrinker if we didn't split any page, but the queue is empty.3821	 * This can happen if pages were freed under us.3822	 */3823	if (!split && list_empty(&ds_queue->split_queue))3824		return SHRINK_STOP;3825	return split;3826}3827 3828#ifdef CONFIG_DEBUG_FS3829static void split_huge_pages_all(void)3830{3831	struct zone *zone;3832	struct page *page;3833	struct folio *folio;3834	unsigned long pfn, max_zone_pfn;3835	unsigned long total = 0, split = 0;3836 3837	pr_debug("Split all THPs\n");3838	for_each_zone(zone) {3839		if (!managed_zone(zone))3840			continue;3841		max_zone_pfn = zone_end_pfn(zone);3842		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {3843			int nr_pages;3844 3845			page = pfn_to_online_page(pfn);3846			if (!page || PageTail(page))3847				continue;3848			folio = page_folio(page);3849			if (!folio_try_get(folio))3850				continue;3851 3852			if (unlikely(page_folio(page) != folio))3853				goto next;3854 3855			if (zone != folio_zone(folio))3856				goto next;3857 3858			if (!folio_test_large(folio)3859				|| folio_test_hugetlb(folio)3860				|| !folio_test_lru(folio))3861				goto next;3862 3863			total++;3864			folio_lock(folio);3865			nr_pages = folio_nr_pages(folio);3866			if (!split_folio(folio))3867				split++;3868			pfn += nr_pages - 1;3869			folio_unlock(folio);3870next:3871			folio_put(folio);3872			cond_resched();3873		}3874	}3875 3876	pr_debug("%lu of %lu THP split\n", split, total);3877}3878 3879static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)3880{3881	return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||3882		    is_vm_hugetlb_page(vma);3883}3884 3885static int split_huge_pages_pid(int pid, unsigned long vaddr_start,3886				unsigned long vaddr_end, unsigned int new_order)3887{3888	int ret = 0;3889	struct task_struct *task;3890	struct mm_struct *mm;3891	unsigned long total = 0, split = 0;3892	unsigned long addr;3893 3894	vaddr_start &= PAGE_MASK;3895	vaddr_end &= PAGE_MASK;3896 3897	task = find_get_task_by_vpid(pid);3898	if (!task) {3899		ret = -ESRCH;3900		goto out;3901	}3902 3903	/* Find the mm_struct */3904	mm = get_task_mm(task);3905	put_task_struct(task);3906 3907	if (!mm) {3908		ret = -EINVAL;3909		goto out;3910	}3911 3912	pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",3913		 pid, vaddr_start, vaddr_end);3914 3915	mmap_read_lock(mm);3916	/*3917	 * always increase addr by PAGE_SIZE, since we could have a PTE page3918	 * table filled with PTE-mapped THPs, each of which is distinct.3919	 */3920	for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {3921		struct vm_area_struct *vma = vma_lookup(mm, addr);3922		struct folio_walk fw;3923		struct folio *folio;3924		struct address_space *mapping;3925		unsigned int target_order = new_order;3926 3927		if (!vma)3928			break;3929 3930		/* skip special VMA and hugetlb VMA */3931		if (vma_not_suitable_for_thp_split(vma)) {3932			addr = vma->vm_end;3933			continue;3934		}3935 3936		folio = folio_walk_start(&fw, vma, addr, 0);3937		if (!folio)3938			continue;3939 3940		if (!is_transparent_hugepage(folio))3941			goto next;3942 3943		if (!folio_test_anon(folio)) {3944			mapping = folio->mapping;3945			target_order = max(new_order,3946					   mapping_min_folio_order(mapping));3947		}3948 3949		if (target_order >= folio_order(folio))3950			goto next;3951 3952		total++;3953		/*3954		 * For folios with private, split_huge_page_to_list_to_order()3955		 * will try to drop it before split and then check if the folio3956		 * can be split or not. So skip the check here.3957		 */3958		if (!folio_test_private(folio) &&3959		    !can_split_folio(folio, 0, NULL))3960			goto next;3961 3962		if (!folio_trylock(folio))3963			goto next;3964		folio_get(folio);3965		folio_walk_end(&fw, vma);3966 3967		if (!folio_test_anon(folio) && folio->mapping != mapping)3968			goto unlock;3969 3970		if (!split_folio_to_order(folio, target_order))3971			split++;3972 3973unlock:3974 3975		folio_unlock(folio);3976		folio_put(folio);3977 3978		cond_resched();3979		continue;3980next:3981		folio_walk_end(&fw, vma);3982		cond_resched();3983	}3984	mmap_read_unlock(mm);3985	mmput(mm);3986 3987	pr_debug("%lu of %lu THP split\n", split, total);3988 3989out:3990	return ret;3991}3992 3993static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,3994				pgoff_t off_end, unsigned int new_order)3995{3996	struct filename *file;3997	struct file *candidate;3998	struct address_space *mapping;3999	int ret = -EINVAL;4000	pgoff_t index;4001	int nr_pages = 1;4002	unsigned long total = 0, split = 0;4003	unsigned int min_order;4004	unsigned int target_order;4005 4006	file = getname_kernel(file_path);4007	if (IS_ERR(file))4008		return ret;4009 4010	candidate = file_open_name(file, O_RDONLY, 0);4011	if (IS_ERR(candidate))4012		goto out;4013 4014	pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",4015		 file_path, off_start, off_end);4016 4017	mapping = candidate->f_mapping;4018	min_order = mapping_min_folio_order(mapping);4019	target_order = max(new_order, min_order);4020 4021	for (index = off_start; index < off_end; index += nr_pages) {4022		struct folio *folio = filemap_get_folio(mapping, index);4023 4024		nr_pages = 1;4025		if (IS_ERR(folio))4026			continue;4027 4028		if (!folio_test_large(folio))4029			goto next;4030 4031		total++;4032		nr_pages = folio_nr_pages(folio);4033 4034		if (target_order >= folio_order(folio))4035			goto next;4036 4037		if (!folio_trylock(folio))4038			goto next;4039 4040		if (folio->mapping != mapping)4041			goto unlock;4042 4043		if (!split_folio_to_order(folio, target_order))4044			split++;4045 4046unlock:4047		folio_unlock(folio);4048next:4049		folio_put(folio);4050		cond_resched();4051	}4052 4053	filp_close(candidate, NULL);4054	ret = 0;4055 4056	pr_debug("%lu of %lu file-backed THP split\n", split, total);4057out:4058	putname(file);4059	return ret;4060}4061 4062#define MAX_INPUT_BUF_SZ 2554063 4064static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,4065				size_t count, loff_t *ppops)4066{4067	static DEFINE_MUTEX(split_debug_mutex);4068	ssize_t ret;4069	/*4070	 * hold pid, start_vaddr, end_vaddr, new_order or4071	 * file_path, off_start, off_end, new_order4072	 */4073	char input_buf[MAX_INPUT_BUF_SZ];4074	int pid;4075	unsigned long vaddr_start, vaddr_end;4076	unsigned int new_order = 0;4077 4078	ret = mutex_lock_interruptible(&split_debug_mutex);4079	if (ret)4080		return ret;4081 4082	ret = -EFAULT;4083 4084	memset(input_buf, 0, MAX_INPUT_BUF_SZ);4085	if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))4086		goto out;4087 4088	input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';4089 4090	if (input_buf[0] == '/') {4091		char *tok;4092		char *buf = input_buf;4093		char file_path[MAX_INPUT_BUF_SZ];4094		pgoff_t off_start = 0, off_end = 0;4095		size_t input_len = strlen(input_buf);4096 4097		tok = strsep(&buf, ",");4098		if (tok) {4099			strcpy(file_path, tok);4100		} else {4101			ret = -EINVAL;4102			goto out;4103		}4104 4105		ret = sscanf(buf, "0x%lx,0x%lx,%d", &off_start, &off_end, &new_order);4106		if (ret != 2 && ret != 3) {4107			ret = -EINVAL;4108			goto out;4109		}4110		ret = split_huge_pages_in_file(file_path, off_start, off_end, new_order);4111		if (!ret)4112			ret = input_len;4113 4114		goto out;4115	}4116 4117	ret = sscanf(input_buf, "%d,0x%lx,0x%lx,%d", &pid, &vaddr_start, &vaddr_end, &new_order);4118	if (ret == 1 && pid == 1) {4119		split_huge_pages_all();4120		ret = strlen(input_buf);4121		goto out;4122	} else if (ret != 3 && ret != 4) {4123		ret = -EINVAL;4124		goto out;4125	}4126 4127	ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end, new_order);4128	if (!ret)4129		ret = strlen(input_buf);4130out:4131	mutex_unlock(&split_debug_mutex);4132	return ret;4133 4134}4135 4136static const struct file_operations split_huge_pages_fops = {4137	.owner	 = THIS_MODULE,4138	.write	 = split_huge_pages_write,4139};4140 4141static int __init split_huge_pages_debugfs(void)4142{4143	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,4144			    &split_huge_pages_fops);4145	return 0;4146}4147late_initcall(split_huge_pages_debugfs);4148#endif4149 4150#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION4151int set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,4152		struct page *page)4153{4154	struct folio *folio = page_folio(page);4155	struct vm_area_struct *vma = pvmw->vma;4156	struct mm_struct *mm = vma->vm_mm;4157	unsigned long address = pvmw->address;4158	bool anon_exclusive;4159	pmd_t pmdval;4160	swp_entry_t entry;4161	pmd_t pmdswp;4162 4163	if (!(pvmw->pmd && !pvmw->pte))4164		return 0;4165 4166	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);4167	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);4168 4169	/* See folio_try_share_anon_rmap_pmd(): invalidate PMD first. */4170	anon_exclusive = folio_test_anon(folio) && PageAnonExclusive(page);4171	if (anon_exclusive && folio_try_share_anon_rmap_pmd(folio, page)) {4172		set_pmd_at(mm, address, pvmw->pmd, pmdval);4173		return -EBUSY;4174	}4175 4176	if (pmd_dirty(pmdval))4177		folio_mark_dirty(folio);4178	if (pmd_write(pmdval))4179		entry = make_writable_migration_entry(page_to_pfn(page));4180	else if (anon_exclusive)4181		entry = make_readable_exclusive_migration_entry(page_to_pfn(page));4182	else4183		entry = make_readable_migration_entry(page_to_pfn(page));4184	if (pmd_young(pmdval))4185		entry = make_migration_entry_young(entry);4186	if (pmd_dirty(pmdval))4187		entry = make_migration_entry_dirty(entry);4188	pmdswp = swp_entry_to_pmd(entry);4189	if (pmd_soft_dirty(pmdval))4190		pmdswp = pmd_swp_mksoft_dirty(pmdswp);4191	if (pmd_uffd_wp(pmdval))4192		pmdswp = pmd_swp_mkuffd_wp(pmdswp);4193	set_pmd_at(mm, address, pvmw->pmd, pmdswp);4194	folio_remove_rmap_pmd(folio, page, vma);4195	folio_put(folio);4196	trace_set_migration_pmd(address, pmd_val(pmdswp));4197 4198	return 0;4199}4200 4201void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)4202{4203	struct folio *folio = page_folio(new);4204	struct vm_area_struct *vma = pvmw->vma;4205	struct mm_struct *mm = vma->vm_mm;4206	unsigned long address = pvmw->address;4207	unsigned long haddr = address & HPAGE_PMD_MASK;4208	pmd_t pmde;4209	swp_entry_t entry;4210 4211	if (!(pvmw->pmd && !pvmw->pte))4212		return;4213 4214	entry = pmd_to_swp_entry(*pvmw->pmd);4215	folio_get(folio);4216	pmde = mk_huge_pmd(new, READ_ONCE(vma->vm_page_prot));4217	if (pmd_swp_soft_dirty(*pvmw->pmd))4218		pmde = pmd_mksoft_dirty(pmde);4219	if (is_writable_migration_entry(entry))4220		pmde = pmd_mkwrite(pmde, vma);4221	if (pmd_swp_uffd_wp(*pvmw->pmd))4222		pmde = pmd_mkuffd_wp(pmde);4223	if (!is_migration_entry_young(entry))4224		pmde = pmd_mkold(pmde);4225	/* NOTE: this may contain setting soft-dirty on some archs */4226	if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))4227		pmde = pmd_mkdirty(pmde);4228 4229	if (folio_test_anon(folio)) {4230		rmap_t rmap_flags = RMAP_NONE;4231 4232		if (!is_readable_migration_entry(entry))4233			rmap_flags |= RMAP_EXCLUSIVE;4234 4235		folio_add_anon_rmap_pmd(folio, new, vma, haddr, rmap_flags);4236	} else {4237		folio_add_file_rmap_pmd(folio, new, vma);4238	}4239	VM_BUG_ON(pmd_write(pmde) && folio_test_anon(folio) && !PageAnonExclusive(new));4240	set_pmd_at(mm, haddr, pvmw->pmd, pmde);4241 4242	/* No need to invalidate - it was non-present before */4243	update_mmu_cache_pmd(vma, address, pvmw->pmd);4244	trace_remove_migration_pmd(address, pmd_val(pmde));4245}4246#endif4247