Linux Audio

Check our new training course

Linux kernel drivers training

May 6-19, 2025
Register
Loading...
v4.17
 
   1/*
   2 *  Copyright (C) 2009  Red Hat, Inc.
   3 *
   4 *  This work is licensed under the terms of the GNU GPL, version 2. See
   5 *  the COPYING file in the top-level directory.
   6 */
   7
   8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10#include <linux/mm.h>
  11#include <linux/sched.h>
  12#include <linux/sched/coredump.h>
  13#include <linux/sched/numa_balancing.h>
  14#include <linux/highmem.h>
  15#include <linux/hugetlb.h>
  16#include <linux/mmu_notifier.h>
  17#include <linux/rmap.h>
  18#include <linux/swap.h>
  19#include <linux/shrinker.h>
  20#include <linux/mm_inline.h>
  21#include <linux/swapops.h>
  22#include <linux/dax.h>
  23#include <linux/khugepaged.h>
  24#include <linux/freezer.h>
  25#include <linux/pfn_t.h>
  26#include <linux/mman.h>
  27#include <linux/memremap.h>
  28#include <linux/pagemap.h>
  29#include <linux/debugfs.h>
  30#include <linux/migrate.h>
  31#include <linux/hashtable.h>
  32#include <linux/userfaultfd_k.h>
  33#include <linux/page_idle.h>
  34#include <linux/shmem_fs.h>
  35#include <linux/oom.h>
 
 
  36
  37#include <asm/tlb.h>
  38#include <asm/pgalloc.h>
  39#include "internal.h"
  40
  41/*
  42 * By default, transparent hugepage support is disabled in order to avoid
  43 * risking an increased memory footprint for applications that are not
  44 * guaranteed to benefit from it. When transparent hugepage support is
  45 * enabled, it is for all mappings, and khugepaged scans all mappings.
  46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47 * for all hugepage allocations.
  48 */
  49unsigned long transparent_hugepage_flags __read_mostly =
  50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52#endif
  53#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55#endif
  56	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60static struct shrinker deferred_split_shrinker;
  61
  62static atomic_t huge_zero_refcount;
  63struct page *huge_zero_page __read_mostly;
  64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  65static struct page *get_huge_zero_page(void)
  66{
  67	struct page *zero_page;
  68retry:
  69	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  70		return READ_ONCE(huge_zero_page);
  71
  72	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  73			HPAGE_PMD_ORDER);
  74	if (!zero_page) {
  75		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  76		return NULL;
  77	}
  78	count_vm_event(THP_ZERO_PAGE_ALLOC);
  79	preempt_disable();
  80	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  81		preempt_enable();
  82		__free_pages(zero_page, compound_order(zero_page));
  83		goto retry;
  84	}
  85
  86	/* We take additional reference here. It will be put back by shrinker */
  87	atomic_set(&huge_zero_refcount, 2);
  88	preempt_enable();
  89	return READ_ONCE(huge_zero_page);
  90}
  91
  92static void put_huge_zero_page(void)
  93{
  94	/*
  95	 * Counter should never go to zero here. Only shrinker can put
  96	 * last reference.
  97	 */
  98	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
  99}
 100
 101struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 102{
 103	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 104		return READ_ONCE(huge_zero_page);
 105
 106	if (!get_huge_zero_page())
 107		return NULL;
 108
 109	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 110		put_huge_zero_page();
 111
 112	return READ_ONCE(huge_zero_page);
 113}
 114
 115void mm_put_huge_zero_page(struct mm_struct *mm)
 116{
 117	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 118		put_huge_zero_page();
 119}
 120
 121static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 122					struct shrink_control *sc)
 123{
 124	/* we can free zero page only if last reference remains */
 125	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 126}
 127
 128static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 129				       struct shrink_control *sc)
 130{
 131	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 132		struct page *zero_page = xchg(&huge_zero_page, NULL);
 133		BUG_ON(zero_page == NULL);
 134		__free_pages(zero_page, compound_order(zero_page));
 135		return HPAGE_PMD_NR;
 136	}
 137
 138	return 0;
 139}
 140
 141static struct shrinker huge_zero_page_shrinker = {
 142	.count_objects = shrink_huge_zero_page_count,
 143	.scan_objects = shrink_huge_zero_page_scan,
 144	.seeks = DEFAULT_SEEKS,
 145};
 146
 147#ifdef CONFIG_SYSFS
 148static ssize_t enabled_show(struct kobject *kobj,
 149			    struct kobj_attribute *attr, char *buf)
 150{
 151	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 152		return sprintf(buf, "[always] madvise never\n");
 153	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 154		return sprintf(buf, "always [madvise] never\n");
 155	else
 156		return sprintf(buf, "always madvise [never]\n");
 157}
 158
 159static ssize_t enabled_store(struct kobject *kobj,
 160			     struct kobj_attribute *attr,
 161			     const char *buf, size_t count)
 162{
 163	ssize_t ret = count;
 164
 165	if (!memcmp("always", buf,
 166		    min(sizeof("always")-1, count))) {
 167		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 168		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 169	} else if (!memcmp("madvise", buf,
 170			   min(sizeof("madvise")-1, count))) {
 171		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 172		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 173	} else if (!memcmp("never", buf,
 174			   min(sizeof("never")-1, count))) {
 175		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 176		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 177	} else
 178		ret = -EINVAL;
 179
 180	if (ret > 0) {
 181		int err = start_stop_khugepaged();
 182		if (err)
 183			ret = err;
 184	}
 185	return ret;
 186}
 187static struct kobj_attribute enabled_attr =
 188	__ATTR(enabled, 0644, enabled_show, enabled_store);
 189
 190ssize_t single_hugepage_flag_show(struct kobject *kobj,
 191				struct kobj_attribute *attr, char *buf,
 192				enum transparent_hugepage_flag flag)
 193{
 194	return sprintf(buf, "%d\n",
 195		       !!test_bit(flag, &transparent_hugepage_flags));
 196}
 197
 198ssize_t single_hugepage_flag_store(struct kobject *kobj,
 199				 struct kobj_attribute *attr,
 200				 const char *buf, size_t count,
 201				 enum transparent_hugepage_flag flag)
 202{
 203	unsigned long value;
 204	int ret;
 205
 206	ret = kstrtoul(buf, 10, &value);
 207	if (ret < 0)
 208		return ret;
 209	if (value > 1)
 210		return -EINVAL;
 211
 212	if (value)
 213		set_bit(flag, &transparent_hugepage_flags);
 214	else
 215		clear_bit(flag, &transparent_hugepage_flags);
 216
 217	return count;
 218}
 219
 220static ssize_t defrag_show(struct kobject *kobj,
 221			   struct kobj_attribute *attr, char *buf)
 222{
 223	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 224		return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 225	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 226		return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 227	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 228		return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 229	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 230		return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 231	return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 232}
 233
 234static ssize_t defrag_store(struct kobject *kobj,
 235			    struct kobj_attribute *attr,
 236			    const char *buf, size_t count)
 237{
 238	if (!memcmp("always", buf,
 239		    min(sizeof("always")-1, count))) {
 240		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 241		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 242		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 243		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 244	} else if (!memcmp("defer+madvise", buf,
 245		    min(sizeof("defer+madvise")-1, count))) {
 246		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 247		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 248		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 249		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 250	} else if (!memcmp("defer", buf,
 251		    min(sizeof("defer")-1, count))) {
 252		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 253		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 254		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 255		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 256	} else if (!memcmp("madvise", buf,
 257			   min(sizeof("madvise")-1, count))) {
 258		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 259		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 260		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 261		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 262	} else if (!memcmp("never", buf,
 263			   min(sizeof("never")-1, count))) {
 264		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 265		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 266		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 267		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 268	} else
 269		return -EINVAL;
 270
 271	return count;
 272}
 273static struct kobj_attribute defrag_attr =
 274	__ATTR(defrag, 0644, defrag_show, defrag_store);
 275
 276static ssize_t use_zero_page_show(struct kobject *kobj,
 277		struct kobj_attribute *attr, char *buf)
 278{
 279	return single_hugepage_flag_show(kobj, attr, buf,
 280				TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 281}
 282static ssize_t use_zero_page_store(struct kobject *kobj,
 283		struct kobj_attribute *attr, const char *buf, size_t count)
 284{
 285	return single_hugepage_flag_store(kobj, attr, buf, count,
 286				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 287}
 288static struct kobj_attribute use_zero_page_attr =
 289	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 290
 291static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 292		struct kobj_attribute *attr, char *buf)
 293{
 294	return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 295}
 296static struct kobj_attribute hpage_pmd_size_attr =
 297	__ATTR_RO(hpage_pmd_size);
 298
 299#ifdef CONFIG_DEBUG_VM
 300static ssize_t debug_cow_show(struct kobject *kobj,
 301				struct kobj_attribute *attr, char *buf)
 302{
 303	return single_hugepage_flag_show(kobj, attr, buf,
 304				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 305}
 306static ssize_t debug_cow_store(struct kobject *kobj,
 307			       struct kobj_attribute *attr,
 308			       const char *buf, size_t count)
 309{
 310	return single_hugepage_flag_store(kobj, attr, buf, count,
 311				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 312}
 313static struct kobj_attribute debug_cow_attr =
 314	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 315#endif /* CONFIG_DEBUG_VM */
 316
 317static struct attribute *hugepage_attr[] = {
 318	&enabled_attr.attr,
 319	&defrag_attr.attr,
 320	&use_zero_page_attr.attr,
 321	&hpage_pmd_size_attr.attr,
 322#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 323	&shmem_enabled_attr.attr,
 324#endif
 325#ifdef CONFIG_DEBUG_VM
 326	&debug_cow_attr.attr,
 327#endif
 328	NULL,
 329};
 330
 331static const struct attribute_group hugepage_attr_group = {
 332	.attrs = hugepage_attr,
 333};
 334
 335static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 336{
 337	int err;
 338
 339	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 340	if (unlikely(!*hugepage_kobj)) {
 341		pr_err("failed to create transparent hugepage kobject\n");
 342		return -ENOMEM;
 343	}
 344
 345	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 346	if (err) {
 347		pr_err("failed to register transparent hugepage group\n");
 348		goto delete_obj;
 349	}
 350
 351	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 352	if (err) {
 353		pr_err("failed to register transparent hugepage group\n");
 354		goto remove_hp_group;
 355	}
 356
 357	return 0;
 358
 359remove_hp_group:
 360	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 361delete_obj:
 362	kobject_put(*hugepage_kobj);
 363	return err;
 364}
 365
 366static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 367{
 368	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 369	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 370	kobject_put(hugepage_kobj);
 371}
 372#else
 373static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 374{
 375	return 0;
 376}
 377
 378static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 379{
 380}
 381#endif /* CONFIG_SYSFS */
 382
 383static int __init hugepage_init(void)
 384{
 385	int err;
 386	struct kobject *hugepage_kobj;
 387
 388	if (!has_transparent_hugepage()) {
 389		transparent_hugepage_flags = 0;
 390		return -EINVAL;
 391	}
 392
 393	/*
 394	 * hugepages can't be allocated by the buddy allocator
 395	 */
 396	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 397	/*
 398	 * we use page->mapping and page->index in second tail page
 399	 * as list_head: assuming THP order >= 2
 400	 */
 401	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 402
 403	err = hugepage_init_sysfs(&hugepage_kobj);
 404	if (err)
 405		goto err_sysfs;
 406
 407	err = khugepaged_init();
 408	if (err)
 409		goto err_slab;
 410
 411	err = register_shrinker(&huge_zero_page_shrinker);
 412	if (err)
 413		goto err_hzp_shrinker;
 414	err = register_shrinker(&deferred_split_shrinker);
 415	if (err)
 416		goto err_split_shrinker;
 417
 418	/*
 419	 * By default disable transparent hugepages on smaller systems,
 420	 * where the extra memory used could hurt more than TLB overhead
 421	 * is likely to save.  The admin can still enable it through /sys.
 422	 */
 423	if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
 424		transparent_hugepage_flags = 0;
 425		return 0;
 426	}
 427
 428	err = start_stop_khugepaged();
 429	if (err)
 430		goto err_khugepaged;
 431
 432	return 0;
 433err_khugepaged:
 434	unregister_shrinker(&deferred_split_shrinker);
 435err_split_shrinker:
 436	unregister_shrinker(&huge_zero_page_shrinker);
 437err_hzp_shrinker:
 438	khugepaged_destroy();
 439err_slab:
 440	hugepage_exit_sysfs(hugepage_kobj);
 441err_sysfs:
 442	return err;
 443}
 444subsys_initcall(hugepage_init);
 445
 446static int __init setup_transparent_hugepage(char *str)
 447{
 448	int ret = 0;
 449	if (!str)
 450		goto out;
 451	if (!strcmp(str, "always")) {
 452		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 453			&transparent_hugepage_flags);
 454		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 455			  &transparent_hugepage_flags);
 456		ret = 1;
 457	} else if (!strcmp(str, "madvise")) {
 458		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 459			  &transparent_hugepage_flags);
 460		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 461			&transparent_hugepage_flags);
 462		ret = 1;
 463	} else if (!strcmp(str, "never")) {
 464		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 465			  &transparent_hugepage_flags);
 466		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 467			  &transparent_hugepage_flags);
 468		ret = 1;
 469	}
 470out:
 471	if (!ret)
 472		pr_warn("transparent_hugepage= cannot parse, ignored\n");
 473	return ret;
 474}
 475__setup("transparent_hugepage=", setup_transparent_hugepage);
 476
 477pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 478{
 479	if (likely(vma->vm_flags & VM_WRITE))
 480		pmd = pmd_mkwrite(pmd);
 481	return pmd;
 482}
 483
 484static inline struct list_head *page_deferred_list(struct page *page)
 
 485{
 486	/*
 487	 * ->lru in the tail pages is occupied by compound_head.
 488	 * Let's use ->mapping + ->index in the second tail page as list_head.
 489	 */
 490	return (struct list_head *)&page[2].mapping;
 
 
 
 
 
 
 
 
 
 491}
 
 492
 493void prep_transhuge_page(struct page *page)
 494{
 495	/*
 496	 * we use page->mapping and page->indexlru in second tail page
 497	 * as list_head: assuming THP order >= 2
 498	 */
 499
 500	INIT_LIST_HEAD(page_deferred_list(page));
 501	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 502}
 503
 504unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 505		loff_t off, unsigned long flags, unsigned long size)
 506{
 507	unsigned long addr;
 508	loff_t off_end = off + len;
 509	loff_t off_align = round_up(off, size);
 510	unsigned long len_pad;
 511
 512	if (off_end <= off_align || (off_end - off_align) < size)
 513		return 0;
 514
 515	len_pad = len + size;
 516	if (len_pad < len || (off + len_pad) < off)
 517		return 0;
 518
 519	addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 520					      off >> PAGE_SHIFT, flags);
 521	if (IS_ERR_VALUE(addr))
 522		return 0;
 523
 524	addr += (off - addr) & (size - 1);
 525	return addr;
 526}
 527
 528unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 529		unsigned long len, unsigned long pgoff, unsigned long flags)
 530{
 531	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 532
 533	if (addr)
 534		goto out;
 535	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 536		goto out;
 537
 538	addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 539	if (addr)
 540		return addr;
 541
 542 out:
 543	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 544}
 545EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 546
 547static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
 548		gfp_t gfp)
 549{
 550	struct vm_area_struct *vma = vmf->vma;
 551	struct mem_cgroup *memcg;
 552	pgtable_t pgtable;
 553	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 554	int ret = 0;
 555
 556	VM_BUG_ON_PAGE(!PageCompound(page), page);
 557
 558	if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
 559		put_page(page);
 560		count_vm_event(THP_FAULT_FALLBACK);
 561		return VM_FAULT_FALLBACK;
 562	}
 563
 564	pgtable = pte_alloc_one(vma->vm_mm, haddr);
 565	if (unlikely(!pgtable)) {
 566		ret = VM_FAULT_OOM;
 567		goto release;
 568	}
 569
 570	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 571	/*
 572	 * The memory barrier inside __SetPageUptodate makes sure that
 573	 * clear_huge_page writes become visible before the set_pmd_at()
 574	 * write.
 575	 */
 576	__SetPageUptodate(page);
 577
 578	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 579	if (unlikely(!pmd_none(*vmf->pmd))) {
 580		goto unlock_release;
 581	} else {
 582		pmd_t entry;
 583
 584		ret = check_stable_address_space(vma->vm_mm);
 585		if (ret)
 586			goto unlock_release;
 587
 588		/* Deliver the page fault to userland */
 589		if (userfaultfd_missing(vma)) {
 590			int ret;
 591
 592			spin_unlock(vmf->ptl);
 593			mem_cgroup_cancel_charge(page, memcg, true);
 594			put_page(page);
 595			pte_free(vma->vm_mm, pgtable);
 596			ret = handle_userfault(vmf, VM_UFFD_MISSING);
 597			VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 598			return ret;
 599		}
 600
 601		entry = mk_huge_pmd(page, vma->vm_page_prot);
 602		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 603		page_add_new_anon_rmap(page, vma, haddr, true);
 604		mem_cgroup_commit_charge(page, memcg, false, true);
 605		lru_cache_add_active_or_unevictable(page, vma);
 606		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 607		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 608		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 609		mm_inc_nr_ptes(vma->vm_mm);
 610		spin_unlock(vmf->ptl);
 611		count_vm_event(THP_FAULT_ALLOC);
 
 612	}
 613
 614	return 0;
 615unlock_release:
 616	spin_unlock(vmf->ptl);
 617release:
 618	if (pgtable)
 619		pte_free(vma->vm_mm, pgtable);
 620	mem_cgroup_cancel_charge(page, memcg, true);
 621	put_page(page);
 622	return ret;
 623
 624}
 625
 626/*
 627 * always: directly stall for all thp allocations
 628 * defer: wake kswapd and fail if not immediately available
 629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 630 *		  fail if not immediately available
 631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 632 *	    available
 633 * never: never stall for any thp allocation
 634 */
 635static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 636{
 637	const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 638
 
 639	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 640		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 
 
 641	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 642		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 
 
 643	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 644		return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 645							     __GFP_KSWAPD_RECLAIM);
 
 
 
 646	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 647		return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
 648							     0);
 
 649	return GFP_TRANSHUGE_LIGHT;
 650}
 651
 652/* Caller must hold page table lock. */
 653static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 654		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 655		struct page *zero_page)
 656{
 657	pmd_t entry;
 658	if (!pmd_none(*pmd))
 659		return false;
 660	entry = mk_pmd(zero_page, vma->vm_page_prot);
 661	entry = pmd_mkhuge(entry);
 662	if (pgtable)
 663		pgtable_trans_huge_deposit(mm, pmd, pgtable);
 664	set_pmd_at(mm, haddr, pmd, entry);
 665	mm_inc_nr_ptes(mm);
 666	return true;
 667}
 668
 669int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 670{
 671	struct vm_area_struct *vma = vmf->vma;
 672	gfp_t gfp;
 673	struct page *page;
 674	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 675
 676	if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 677		return VM_FAULT_FALLBACK;
 678	if (unlikely(anon_vma_prepare(vma)))
 679		return VM_FAULT_OOM;
 680	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 681		return VM_FAULT_OOM;
 682	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 683			!mm_forbids_zeropage(vma->vm_mm) &&
 684			transparent_hugepage_use_zero_page()) {
 685		pgtable_t pgtable;
 686		struct page *zero_page;
 687		bool set;
 688		int ret;
 689		pgtable = pte_alloc_one(vma->vm_mm, haddr);
 690		if (unlikely(!pgtable))
 691			return VM_FAULT_OOM;
 692		zero_page = mm_get_huge_zero_page(vma->vm_mm);
 693		if (unlikely(!zero_page)) {
 694			pte_free(vma->vm_mm, pgtable);
 695			count_vm_event(THP_FAULT_FALLBACK);
 696			return VM_FAULT_FALLBACK;
 697		}
 698		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 699		ret = 0;
 700		set = false;
 701		if (pmd_none(*vmf->pmd)) {
 702			ret = check_stable_address_space(vma->vm_mm);
 703			if (ret) {
 704				spin_unlock(vmf->ptl);
 705			} else if (userfaultfd_missing(vma)) {
 706				spin_unlock(vmf->ptl);
 707				ret = handle_userfault(vmf, VM_UFFD_MISSING);
 708				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 709			} else {
 710				set_huge_zero_page(pgtable, vma->vm_mm, vma,
 711						   haddr, vmf->pmd, zero_page);
 712				spin_unlock(vmf->ptl);
 713				set = true;
 714			}
 715		} else
 716			spin_unlock(vmf->ptl);
 717		if (!set)
 718			pte_free(vma->vm_mm, pgtable);
 719		return ret;
 720	}
 721	gfp = alloc_hugepage_direct_gfpmask(vma);
 722	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 723	if (unlikely(!page)) {
 724		count_vm_event(THP_FAULT_FALLBACK);
 725		return VM_FAULT_FALLBACK;
 726	}
 727	prep_transhuge_page(page);
 728	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 729}
 730
 731static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 732		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 733		pgtable_t pgtable)
 734{
 735	struct mm_struct *mm = vma->vm_mm;
 736	pmd_t entry;
 737	spinlock_t *ptl;
 738
 739	ptl = pmd_lock(mm, pmd);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 740	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 741	if (pfn_t_devmap(pfn))
 742		entry = pmd_mkdevmap(entry);
 743	if (write) {
 744		entry = pmd_mkyoung(pmd_mkdirty(entry));
 745		entry = maybe_pmd_mkwrite(entry, vma);
 746	}
 747
 748	if (pgtable) {
 749		pgtable_trans_huge_deposit(mm, pmd, pgtable);
 750		mm_inc_nr_ptes(mm);
 
 751	}
 752
 753	set_pmd_at(mm, addr, pmd, entry);
 754	update_mmu_cache_pmd(vma, addr, pmd);
 
 
 755	spin_unlock(ptl);
 
 
 756}
 757
 758int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 759			pmd_t *pmd, pfn_t pfn, bool write)
 760{
 
 
 761	pgprot_t pgprot = vma->vm_page_prot;
 762	pgtable_t pgtable = NULL;
 
 763	/*
 764	 * If we had pmd_special, we could avoid all these restrictions,
 765	 * but we need to be consistent with PTEs and architectures that
 766	 * can't support a 'special' bit.
 767	 */
 768	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 
 769	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 770						(VM_PFNMAP|VM_MIXEDMAP));
 771	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 772	BUG_ON(!pfn_t_devmap(pfn));
 773
 774	if (addr < vma->vm_start || addr >= vma->vm_end)
 775		return VM_FAULT_SIGBUS;
 776
 777	if (arch_needs_pgtable_deposit()) {
 778		pgtable = pte_alloc_one(vma->vm_mm, addr);
 779		if (!pgtable)
 780			return VM_FAULT_OOM;
 781	}
 782
 783	track_pfn_insert(vma, &pgprot, pfn);
 784
 785	insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
 786	return VM_FAULT_NOPAGE;
 787}
 788EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 789
 790#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 791static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 792{
 793	if (likely(vma->vm_flags & VM_WRITE))
 794		pud = pud_mkwrite(pud);
 795	return pud;
 796}
 797
 798static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 799		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 800{
 801	struct mm_struct *mm = vma->vm_mm;
 802	pud_t entry;
 803	spinlock_t *ptl;
 804
 805	ptl = pud_lock(mm, pud);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 806	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 807	if (pfn_t_devmap(pfn))
 808		entry = pud_mkdevmap(entry);
 809	if (write) {
 810		entry = pud_mkyoung(pud_mkdirty(entry));
 811		entry = maybe_pud_mkwrite(entry, vma);
 812	}
 813	set_pud_at(mm, addr, pud, entry);
 814	update_mmu_cache_pud(vma, addr, pud);
 
 
 815	spin_unlock(ptl);
 816}
 817
 818int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 819			pud_t *pud, pfn_t pfn, bool write)
 820{
 
 
 821	pgprot_t pgprot = vma->vm_page_prot;
 
 822	/*
 823	 * If we had pud_special, we could avoid all these restrictions,
 824	 * but we need to be consistent with PTEs and architectures that
 825	 * can't support a 'special' bit.
 826	 */
 827	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
 
 828	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 829						(VM_PFNMAP|VM_MIXEDMAP));
 830	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 831	BUG_ON(!pfn_t_devmap(pfn));
 832
 833	if (addr < vma->vm_start || addr >= vma->vm_end)
 834		return VM_FAULT_SIGBUS;
 835
 836	track_pfn_insert(vma, &pgprot, pfn);
 837
 838	insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
 839	return VM_FAULT_NOPAGE;
 840}
 841EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 842#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 843
 844static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 845		pmd_t *pmd, int flags)
 846{
 847	pmd_t _pmd;
 848
 849	_pmd = pmd_mkyoung(*pmd);
 850	if (flags & FOLL_WRITE)
 851		_pmd = pmd_mkdirty(_pmd);
 852	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 853				pmd, _pmd, flags & FOLL_WRITE))
 854		update_mmu_cache_pmd(vma, addr, pmd);
 855}
 856
 857struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 858		pmd_t *pmd, int flags)
 859{
 860	unsigned long pfn = pmd_pfn(*pmd);
 861	struct mm_struct *mm = vma->vm_mm;
 862	struct dev_pagemap *pgmap;
 863	struct page *page;
 864
 865	assert_spin_locked(pmd_lockptr(mm, pmd));
 866
 867	/*
 868	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
 869	 * not be in this function with `flags & FOLL_COW` set.
 870	 */
 871	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 872
 873	if (flags & FOLL_WRITE && !pmd_write(*pmd))
 874		return NULL;
 875
 876	if (pmd_present(*pmd) && pmd_devmap(*pmd))
 877		/* pass */;
 878	else
 879		return NULL;
 880
 881	if (flags & FOLL_TOUCH)
 882		touch_pmd(vma, addr, pmd, flags);
 883
 884	/*
 885	 * device mapped pages can only be returned if the
 886	 * caller will manage the page reference count.
 887	 */
 888	if (!(flags & FOLL_GET))
 889		return ERR_PTR(-EEXIST);
 890
 891	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 892	pgmap = get_dev_pagemap(pfn, NULL);
 893	if (!pgmap)
 894		return ERR_PTR(-EFAULT);
 895	page = pfn_to_page(pfn);
 896	get_page(page);
 897	put_dev_pagemap(pgmap);
 898
 899	return page;
 900}
 901
 902int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 903		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 904		  struct vm_area_struct *vma)
 905{
 906	spinlock_t *dst_ptl, *src_ptl;
 907	struct page *src_page;
 908	pmd_t pmd;
 909	pgtable_t pgtable = NULL;
 910	int ret = -ENOMEM;
 911
 912	/* Skip if can be re-fill on fault */
 913	if (!vma_is_anonymous(vma))
 914		return 0;
 915
 916	pgtable = pte_alloc_one(dst_mm, addr);
 917	if (unlikely(!pgtable))
 918		goto out;
 919
 920	dst_ptl = pmd_lock(dst_mm, dst_pmd);
 921	src_ptl = pmd_lockptr(src_mm, src_pmd);
 922	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 923
 924	ret = -EAGAIN;
 925	pmd = *src_pmd;
 926
 927#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 928	if (unlikely(is_swap_pmd(pmd))) {
 929		swp_entry_t entry = pmd_to_swp_entry(pmd);
 930
 931		VM_BUG_ON(!is_pmd_migration_entry(pmd));
 932		if (is_write_migration_entry(entry)) {
 933			make_migration_entry_read(&entry);
 934			pmd = swp_entry_to_pmd(entry);
 935			if (pmd_swp_soft_dirty(*src_pmd))
 936				pmd = pmd_swp_mksoft_dirty(pmd);
 937			set_pmd_at(src_mm, addr, src_pmd, pmd);
 938		}
 939		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 940		mm_inc_nr_ptes(dst_mm);
 941		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 942		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 943		ret = 0;
 944		goto out_unlock;
 945	}
 946#endif
 947
 948	if (unlikely(!pmd_trans_huge(pmd))) {
 949		pte_free(dst_mm, pgtable);
 950		goto out_unlock;
 951	}
 952	/*
 953	 * When page table lock is held, the huge zero pmd should not be
 954	 * under splitting since we don't split the page itself, only pmd to
 955	 * a page table.
 956	 */
 957	if (is_huge_zero_pmd(pmd)) {
 958		struct page *zero_page;
 959		/*
 960		 * get_huge_zero_page() will never allocate a new page here,
 961		 * since we already have a zero page to copy. It just takes a
 962		 * reference.
 963		 */
 964		zero_page = mm_get_huge_zero_page(dst_mm);
 965		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
 966				zero_page);
 967		ret = 0;
 968		goto out_unlock;
 969	}
 970
 971	src_page = pmd_page(pmd);
 972	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
 973	get_page(src_page);
 974	page_dup_rmap(src_page, true);
 975	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 976	mm_inc_nr_ptes(dst_mm);
 977	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
 978
 979	pmdp_set_wrprotect(src_mm, addr, src_pmd);
 980	pmd = pmd_mkold(pmd_wrprotect(pmd));
 981	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 982
 983	ret = 0;
 984out_unlock:
 985	spin_unlock(src_ptl);
 986	spin_unlock(dst_ptl);
 987out:
 988	return ret;
 989}
 990
 991#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 992static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
 993		pud_t *pud, int flags)
 994{
 995	pud_t _pud;
 996
 997	_pud = pud_mkyoung(*pud);
 998	if (flags & FOLL_WRITE)
 999		_pud = pud_mkdirty(_pud);
1000	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1001				pud, _pud, flags & FOLL_WRITE))
1002		update_mmu_cache_pud(vma, addr, pud);
1003}
1004
1005struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1006		pud_t *pud, int flags)
1007{
1008	unsigned long pfn = pud_pfn(*pud);
1009	struct mm_struct *mm = vma->vm_mm;
1010	struct dev_pagemap *pgmap;
1011	struct page *page;
1012
1013	assert_spin_locked(pud_lockptr(mm, pud));
1014
1015	if (flags & FOLL_WRITE && !pud_write(*pud))
1016		return NULL;
1017
1018	if (pud_present(*pud) && pud_devmap(*pud))
1019		/* pass */;
1020	else
1021		return NULL;
1022
1023	if (flags & FOLL_TOUCH)
1024		touch_pud(vma, addr, pud, flags);
1025
1026	/*
1027	 * device mapped pages can only be returned if the
1028	 * caller will manage the page reference count.
1029	 */
1030	if (!(flags & FOLL_GET))
1031		return ERR_PTR(-EEXIST);
1032
1033	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1034	pgmap = get_dev_pagemap(pfn, NULL);
1035	if (!pgmap)
1036		return ERR_PTR(-EFAULT);
1037	page = pfn_to_page(pfn);
1038	get_page(page);
1039	put_dev_pagemap(pgmap);
1040
1041	return page;
1042}
1043
1044int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1046		  struct vm_area_struct *vma)
1047{
1048	spinlock_t *dst_ptl, *src_ptl;
1049	pud_t pud;
1050	int ret;
1051
1052	dst_ptl = pud_lock(dst_mm, dst_pud);
1053	src_ptl = pud_lockptr(src_mm, src_pud);
1054	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1055
1056	ret = -EAGAIN;
1057	pud = *src_pud;
1058	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1059		goto out_unlock;
1060
1061	/*
1062	 * When page table lock is held, the huge zero pud should not be
1063	 * under splitting since we don't split the page itself, only pud to
1064	 * a page table.
1065	 */
1066	if (is_huge_zero_pud(pud)) {
1067		/* No huge zero pud yet */
1068	}
1069
1070	pudp_set_wrprotect(src_mm, addr, src_pud);
1071	pud = pud_mkold(pud_wrprotect(pud));
1072	set_pud_at(dst_mm, addr, dst_pud, pud);
1073
1074	ret = 0;
1075out_unlock:
1076	spin_unlock(src_ptl);
1077	spin_unlock(dst_ptl);
1078	return ret;
1079}
1080
1081void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1082{
1083	pud_t entry;
1084	unsigned long haddr;
1085	bool write = vmf->flags & FAULT_FLAG_WRITE;
1086
1087	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1088	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1089		goto unlock;
1090
1091	entry = pud_mkyoung(orig_pud);
1092	if (write)
1093		entry = pud_mkdirty(entry);
1094	haddr = vmf->address & HPAGE_PUD_MASK;
1095	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1096		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1097
1098unlock:
1099	spin_unlock(vmf->ptl);
1100}
1101#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1102
1103void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1104{
1105	pmd_t entry;
1106	unsigned long haddr;
1107	bool write = vmf->flags & FAULT_FLAG_WRITE;
1108
1109	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1110	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1111		goto unlock;
1112
1113	entry = pmd_mkyoung(orig_pmd);
1114	if (write)
1115		entry = pmd_mkdirty(entry);
1116	haddr = vmf->address & HPAGE_PMD_MASK;
1117	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1118		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1119
1120unlock:
1121	spin_unlock(vmf->ptl);
1122}
1123
1124static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1125		struct page *page)
1126{
1127	struct vm_area_struct *vma = vmf->vma;
1128	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1129	struct mem_cgroup *memcg;
1130	pgtable_t pgtable;
1131	pmd_t _pmd;
1132	int ret = 0, i;
 
1133	struct page **pages;
1134	unsigned long mmun_start;	/* For mmu_notifiers */
1135	unsigned long mmun_end;		/* For mmu_notifiers */
1136
1137	pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1138			GFP_KERNEL);
1139	if (unlikely(!pages)) {
1140		ret |= VM_FAULT_OOM;
1141		goto out;
1142	}
1143
1144	for (i = 0; i < HPAGE_PMD_NR; i++) {
1145		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1146					       vmf->address, page_to_nid(page));
1147		if (unlikely(!pages[i] ||
1148			     mem_cgroup_try_charge(pages[i], vma->vm_mm,
1149				     GFP_KERNEL, &memcg, false))) {
1150			if (pages[i])
1151				put_page(pages[i]);
1152			while (--i >= 0) {
1153				memcg = (void *)page_private(pages[i]);
1154				set_page_private(pages[i], 0);
1155				mem_cgroup_cancel_charge(pages[i], memcg,
1156						false);
1157				put_page(pages[i]);
1158			}
1159			kfree(pages);
1160			ret |= VM_FAULT_OOM;
1161			goto out;
1162		}
1163		set_page_private(pages[i], (unsigned long)memcg);
1164	}
1165
1166	for (i = 0; i < HPAGE_PMD_NR; i++) {
1167		copy_user_highpage(pages[i], page + i,
1168				   haddr + PAGE_SIZE * i, vma);
1169		__SetPageUptodate(pages[i]);
1170		cond_resched();
1171	}
1172
1173	mmun_start = haddr;
1174	mmun_end   = haddr + HPAGE_PMD_SIZE;
1175	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1176
1177	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1178	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1179		goto out_free_pages;
1180	VM_BUG_ON_PAGE(!PageHead(page), page);
1181
1182	/*
1183	 * Leave pmd empty until pte is filled note we must notify here as
1184	 * concurrent CPU thread might write to new page before the call to
1185	 * mmu_notifier_invalidate_range_end() happens which can lead to a
1186	 * device seeing memory write in different order than CPU.
1187	 *
1188	 * See Documentation/vm/mmu_notifier.txt
1189	 */
1190	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1191
1192	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1193	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1194
1195	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1196		pte_t entry;
1197		entry = mk_pte(pages[i], vma->vm_page_prot);
1198		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1199		memcg = (void *)page_private(pages[i]);
1200		set_page_private(pages[i], 0);
1201		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1202		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1203		lru_cache_add_active_or_unevictable(pages[i], vma);
1204		vmf->pte = pte_offset_map(&_pmd, haddr);
1205		VM_BUG_ON(!pte_none(*vmf->pte));
1206		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1207		pte_unmap(vmf->pte);
1208	}
1209	kfree(pages);
1210
1211	smp_wmb(); /* make pte visible before pmd */
1212	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1213	page_remove_rmap(page, true);
1214	spin_unlock(vmf->ptl);
1215
1216	/*
1217	 * No need to double call mmu_notifier->invalidate_range() callback as
1218	 * the above pmdp_huge_clear_flush_notify() did already call it.
1219	 */
1220	mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1221						mmun_end);
1222
1223	ret |= VM_FAULT_WRITE;
1224	put_page(page);
1225
1226out:
1227	return ret;
1228
1229out_free_pages:
1230	spin_unlock(vmf->ptl);
1231	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1232	for (i = 0; i < HPAGE_PMD_NR; i++) {
1233		memcg = (void *)page_private(pages[i]);
1234		set_page_private(pages[i], 0);
1235		mem_cgroup_cancel_charge(pages[i], memcg, false);
1236		put_page(pages[i]);
1237	}
1238	kfree(pages);
1239	goto out;
1240}
1241
1242int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1243{
1244	struct vm_area_struct *vma = vmf->vma;
1245	struct page *page = NULL, *new_page;
1246	struct mem_cgroup *memcg;
1247	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1248	unsigned long mmun_start;	/* For mmu_notifiers */
1249	unsigned long mmun_end;		/* For mmu_notifiers */
1250	gfp_t huge_gfp;			/* for allocation and charge */
1251	int ret = 0;
1252
1253	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1254	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1255	if (is_huge_zero_pmd(orig_pmd))
1256		goto alloc;
1257	spin_lock(vmf->ptl);
1258	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1259		goto out_unlock;
1260
1261	page = pmd_page(orig_pmd);
1262	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1263	/*
1264	 * We can only reuse the page if nobody else maps the huge page or it's
1265	 * part.
1266	 */
1267	if (!trylock_page(page)) {
1268		get_page(page);
1269		spin_unlock(vmf->ptl);
1270		lock_page(page);
1271		spin_lock(vmf->ptl);
1272		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1273			unlock_page(page);
1274			put_page(page);
1275			goto out_unlock;
1276		}
1277		put_page(page);
1278	}
1279	if (reuse_swap_page(page, NULL)) {
1280		pmd_t entry;
1281		entry = pmd_mkyoung(orig_pmd);
1282		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1283		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1284			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1285		ret |= VM_FAULT_WRITE;
1286		unlock_page(page);
1287		goto out_unlock;
1288	}
1289	unlock_page(page);
1290	get_page(page);
1291	spin_unlock(vmf->ptl);
1292alloc:
1293	if (transparent_hugepage_enabled(vma) &&
1294	    !transparent_hugepage_debug_cow()) {
1295		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1296		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1297	} else
1298		new_page = NULL;
1299
1300	if (likely(new_page)) {
1301		prep_transhuge_page(new_page);
1302	} else {
1303		if (!page) {
1304			split_huge_pmd(vma, vmf->pmd, vmf->address);
1305			ret |= VM_FAULT_FALLBACK;
1306		} else {
1307			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1308			if (ret & VM_FAULT_OOM) {
1309				split_huge_pmd(vma, vmf->pmd, vmf->address);
1310				ret |= VM_FAULT_FALLBACK;
1311			}
1312			put_page(page);
1313		}
1314		count_vm_event(THP_FAULT_FALLBACK);
1315		goto out;
1316	}
1317
1318	if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1319					huge_gfp, &memcg, true))) {
1320		put_page(new_page);
1321		split_huge_pmd(vma, vmf->pmd, vmf->address);
1322		if (page)
1323			put_page(page);
1324		ret |= VM_FAULT_FALLBACK;
1325		count_vm_event(THP_FAULT_FALLBACK);
1326		goto out;
1327	}
1328
1329	count_vm_event(THP_FAULT_ALLOC);
 
1330
1331	if (!page)
1332		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1333	else
1334		copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
 
1335	__SetPageUptodate(new_page);
1336
1337	mmun_start = haddr;
1338	mmun_end   = haddr + HPAGE_PMD_SIZE;
1339	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1340
1341	spin_lock(vmf->ptl);
1342	if (page)
1343		put_page(page);
1344	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345		spin_unlock(vmf->ptl);
1346		mem_cgroup_cancel_charge(new_page, memcg, true);
1347		put_page(new_page);
1348		goto out_mn;
1349	} else {
1350		pmd_t entry;
1351		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1352		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1353		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1354		page_add_new_anon_rmap(new_page, vma, haddr, true);
1355		mem_cgroup_commit_charge(new_page, memcg, false, true);
1356		lru_cache_add_active_or_unevictable(new_page, vma);
1357		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1358		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1359		if (!page) {
1360			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1361		} else {
1362			VM_BUG_ON_PAGE(!PageHead(page), page);
1363			page_remove_rmap(page, true);
1364			put_page(page);
1365		}
1366		ret |= VM_FAULT_WRITE;
1367	}
1368	spin_unlock(vmf->ptl);
1369out_mn:
1370	/*
1371	 * No need to double call mmu_notifier->invalidate_range() callback as
1372	 * the above pmdp_huge_clear_flush_notify() did already call it.
1373	 */
1374	mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1375					       mmun_end);
1376out:
1377	return ret;
1378out_unlock:
1379	spin_unlock(vmf->ptl);
1380	return ret;
1381}
1382
1383/*
1384 * FOLL_FORCE can write to even unwritable pmd's, but only
1385 * after we've gone through a COW cycle and they are dirty.
1386 */
1387static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1388{
1389	return pmd_write(pmd) ||
1390	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1391}
1392
1393struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1394				   unsigned long addr,
1395				   pmd_t *pmd,
1396				   unsigned int flags)
1397{
1398	struct mm_struct *mm = vma->vm_mm;
1399	struct page *page = NULL;
1400
1401	assert_spin_locked(pmd_lockptr(mm, pmd));
1402
1403	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1404		goto out;
1405
1406	/* Avoid dumping huge zero page */
1407	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1408		return ERR_PTR(-EFAULT);
1409
1410	/* Full NUMA hinting faults to serialise migration in fault paths */
1411	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1412		goto out;
1413
1414	page = pmd_page(*pmd);
1415	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1416	if (flags & FOLL_TOUCH)
1417		touch_pmd(vma, addr, pmd, flags);
1418	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1419		/*
1420		 * We don't mlock() pte-mapped THPs. This way we can avoid
1421		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1422		 *
1423		 * For anon THP:
1424		 *
1425		 * In most cases the pmd is the only mapping of the page as we
1426		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1427		 * writable private mappings in populate_vma_page_range().
1428		 *
1429		 * The only scenario when we have the page shared here is if we
1430		 * mlocking read-only mapping shared over fork(). We skip
1431		 * mlocking such pages.
1432		 *
1433		 * For file THP:
1434		 *
1435		 * We can expect PageDoubleMap() to be stable under page lock:
1436		 * for file pages we set it in page_add_file_rmap(), which
1437		 * requires page to be locked.
1438		 */
1439
1440		if (PageAnon(page) && compound_mapcount(page) != 1)
1441			goto skip_mlock;
1442		if (PageDoubleMap(page) || !page->mapping)
1443			goto skip_mlock;
1444		if (!trylock_page(page))
1445			goto skip_mlock;
1446		lru_add_drain();
1447		if (page->mapping && !PageDoubleMap(page))
1448			mlock_vma_page(page);
1449		unlock_page(page);
1450	}
1451skip_mlock:
1452	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1453	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1454	if (flags & FOLL_GET)
1455		get_page(page);
1456
1457out:
1458	return page;
1459}
1460
1461/* NUMA hinting page fault entry point for trans huge pmds */
1462int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1463{
1464	struct vm_area_struct *vma = vmf->vma;
1465	struct anon_vma *anon_vma = NULL;
1466	struct page *page;
1467	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1468	int page_nid = -1, this_nid = numa_node_id();
1469	int target_nid, last_cpupid = -1;
1470	bool page_locked;
1471	bool migrated = false;
1472	bool was_writable;
1473	int flags = 0;
1474
1475	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1476	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1477		goto out_unlock;
1478
1479	/*
1480	 * If there are potential migrations, wait for completion and retry
1481	 * without disrupting NUMA hinting information. Do not relock and
1482	 * check_same as the page may no longer be mapped.
1483	 */
1484	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1485		page = pmd_page(*vmf->pmd);
1486		if (!get_page_unless_zero(page))
1487			goto out_unlock;
1488		spin_unlock(vmf->ptl);
1489		wait_on_page_locked(page);
1490		put_page(page);
1491		goto out;
1492	}
1493
1494	page = pmd_page(pmd);
1495	BUG_ON(is_huge_zero_page(page));
1496	page_nid = page_to_nid(page);
1497	last_cpupid = page_cpupid_last(page);
1498	count_vm_numa_event(NUMA_HINT_FAULTS);
1499	if (page_nid == this_nid) {
1500		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1501		flags |= TNF_FAULT_LOCAL;
1502	}
1503
1504	/* See similar comment in do_numa_page for explanation */
1505	if (!pmd_savedwrite(pmd))
1506		flags |= TNF_NO_GROUP;
1507
1508	/*
1509	 * Acquire the page lock to serialise THP migrations but avoid dropping
1510	 * page_table_lock if at all possible
1511	 */
1512	page_locked = trylock_page(page);
1513	target_nid = mpol_misplaced(page, vma, haddr);
1514	if (target_nid == -1) {
1515		/* If the page was locked, there are no parallel migrations */
1516		if (page_locked)
1517			goto clear_pmdnuma;
1518	}
1519
1520	/* Migration could have started since the pmd_trans_migrating check */
1521	if (!page_locked) {
1522		page_nid = -1;
1523		if (!get_page_unless_zero(page))
1524			goto out_unlock;
1525		spin_unlock(vmf->ptl);
1526		wait_on_page_locked(page);
1527		put_page(page);
1528		goto out;
1529	}
1530
1531	/*
1532	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1533	 * to serialises splits
1534	 */
1535	get_page(page);
1536	spin_unlock(vmf->ptl);
1537	anon_vma = page_lock_anon_vma_read(page);
1538
1539	/* Confirm the PMD did not change while page_table_lock was released */
1540	spin_lock(vmf->ptl);
1541	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1542		unlock_page(page);
1543		put_page(page);
1544		page_nid = -1;
1545		goto out_unlock;
1546	}
1547
1548	/* Bail if we fail to protect against THP splits for any reason */
1549	if (unlikely(!anon_vma)) {
1550		put_page(page);
1551		page_nid = -1;
1552		goto clear_pmdnuma;
1553	}
1554
1555	/*
1556	 * Since we took the NUMA fault, we must have observed the !accessible
1557	 * bit. Make sure all other CPUs agree with that, to avoid them
1558	 * modifying the page we're about to migrate.
1559	 *
1560	 * Must be done under PTL such that we'll observe the relevant
1561	 * inc_tlb_flush_pending().
1562	 *
1563	 * We are not sure a pending tlb flush here is for a huge page
1564	 * mapping or not. Hence use the tlb range variant
1565	 */
1566	if (mm_tlb_flush_pending(vma->vm_mm))
1567		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
 
 
 
 
 
 
 
 
 
 
 
 
1568
1569	/*
1570	 * Migrate the THP to the requested node, returns with page unlocked
1571	 * and access rights restored.
1572	 */
1573	spin_unlock(vmf->ptl);
1574
1575	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1576				vmf->pmd, pmd, vmf->address, page, target_nid);
1577	if (migrated) {
1578		flags |= TNF_MIGRATED;
1579		page_nid = target_nid;
1580	} else
1581		flags |= TNF_MIGRATE_FAIL;
1582
1583	goto out;
1584clear_pmdnuma:
1585	BUG_ON(!PageLocked(page));
1586	was_writable = pmd_savedwrite(pmd);
1587	pmd = pmd_modify(pmd, vma->vm_page_prot);
1588	pmd = pmd_mkyoung(pmd);
1589	if (was_writable)
1590		pmd = pmd_mkwrite(pmd);
1591	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1592	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1593	unlock_page(page);
1594out_unlock:
1595	spin_unlock(vmf->ptl);
1596
1597out:
1598	if (anon_vma)
1599		page_unlock_anon_vma_read(anon_vma);
1600
1601	if (page_nid != -1)
1602		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1603				flags);
1604
1605	return 0;
1606}
1607
1608/*
1609 * Return true if we do MADV_FREE successfully on entire pmd page.
1610 * Otherwise, return false.
1611 */
1612bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1613		pmd_t *pmd, unsigned long addr, unsigned long next)
1614{
1615	spinlock_t *ptl;
1616	pmd_t orig_pmd;
1617	struct page *page;
1618	struct mm_struct *mm = tlb->mm;
1619	bool ret = false;
1620
1621	tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1622
1623	ptl = pmd_trans_huge_lock(pmd, vma);
1624	if (!ptl)
1625		goto out_unlocked;
1626
1627	orig_pmd = *pmd;
1628	if (is_huge_zero_pmd(orig_pmd))
1629		goto out;
1630
1631	if (unlikely(!pmd_present(orig_pmd))) {
1632		VM_BUG_ON(thp_migration_supported() &&
1633				  !is_pmd_migration_entry(orig_pmd));
1634		goto out;
1635	}
1636
1637	page = pmd_page(orig_pmd);
1638	/*
1639	 * If other processes are mapping this page, we couldn't discard
1640	 * the page unless they all do MADV_FREE so let's skip the page.
1641	 */
1642	if (page_mapcount(page) != 1)
1643		goto out;
1644
1645	if (!trylock_page(page))
1646		goto out;
1647
1648	/*
1649	 * If user want to discard part-pages of THP, split it so MADV_FREE
1650	 * will deactivate only them.
1651	 */
1652	if (next - addr != HPAGE_PMD_SIZE) {
1653		get_page(page);
1654		spin_unlock(ptl);
1655		split_huge_page(page);
1656		unlock_page(page);
1657		put_page(page);
1658		goto out_unlocked;
1659	}
1660
1661	if (PageDirty(page))
1662		ClearPageDirty(page);
1663	unlock_page(page);
1664
1665	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1666		pmdp_invalidate(vma, addr, pmd);
1667		orig_pmd = pmd_mkold(orig_pmd);
1668		orig_pmd = pmd_mkclean(orig_pmd);
1669
1670		set_pmd_at(mm, addr, pmd, orig_pmd);
1671		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1672	}
1673
1674	mark_page_lazyfree(page);
1675	ret = true;
1676out:
1677	spin_unlock(ptl);
1678out_unlocked:
1679	return ret;
1680}
1681
1682static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1683{
1684	pgtable_t pgtable;
1685
1686	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1687	pte_free(mm, pgtable);
1688	mm_dec_nr_ptes(mm);
1689}
1690
1691int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1692		 pmd_t *pmd, unsigned long addr)
1693{
1694	pmd_t orig_pmd;
1695	spinlock_t *ptl;
1696
1697	tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1698
1699	ptl = __pmd_trans_huge_lock(pmd, vma);
1700	if (!ptl)
1701		return 0;
1702	/*
1703	 * For architectures like ppc64 we look at deposited pgtable
1704	 * when calling pmdp_huge_get_and_clear. So do the
1705	 * pgtable_trans_huge_withdraw after finishing pmdp related
1706	 * operations.
1707	 */
1708	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1709			tlb->fullmm);
1710	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1711	if (vma_is_dax(vma)) {
1712		if (arch_needs_pgtable_deposit())
1713			zap_deposited_table(tlb->mm, pmd);
1714		spin_unlock(ptl);
1715		if (is_huge_zero_pmd(orig_pmd))
1716			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1717	} else if (is_huge_zero_pmd(orig_pmd)) {
1718		zap_deposited_table(tlb->mm, pmd);
1719		spin_unlock(ptl);
1720		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1721	} else {
1722		struct page *page = NULL;
1723		int flush_needed = 1;
1724
1725		if (pmd_present(orig_pmd)) {
1726			page = pmd_page(orig_pmd);
1727			page_remove_rmap(page, true);
1728			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1729			VM_BUG_ON_PAGE(!PageHead(page), page);
1730		} else if (thp_migration_supported()) {
1731			swp_entry_t entry;
1732
1733			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1734			entry = pmd_to_swp_entry(orig_pmd);
1735			page = pfn_to_page(swp_offset(entry));
1736			flush_needed = 0;
1737		} else
1738			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1739
1740		if (PageAnon(page)) {
1741			zap_deposited_table(tlb->mm, pmd);
1742			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1743		} else {
1744			if (arch_needs_pgtable_deposit())
1745				zap_deposited_table(tlb->mm, pmd);
1746			add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1747		}
1748
1749		spin_unlock(ptl);
1750		if (flush_needed)
1751			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1752	}
1753	return 1;
1754}
1755
1756#ifndef pmd_move_must_withdraw
1757static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1758					 spinlock_t *old_pmd_ptl,
1759					 struct vm_area_struct *vma)
1760{
1761	/*
1762	 * With split pmd lock we also need to move preallocated
1763	 * PTE page table if new_pmd is on different PMD page table.
1764	 *
1765	 * We also don't deposit and withdraw tables for file pages.
1766	 */
1767	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1768}
1769#endif
1770
1771static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1772{
1773#ifdef CONFIG_MEM_SOFT_DIRTY
1774	if (unlikely(is_pmd_migration_entry(pmd)))
1775		pmd = pmd_swp_mksoft_dirty(pmd);
1776	else if (pmd_present(pmd))
1777		pmd = pmd_mksoft_dirty(pmd);
1778#endif
1779	return pmd;
1780}
1781
1782bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1783		  unsigned long new_addr, unsigned long old_end,
1784		  pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1785{
1786	spinlock_t *old_ptl, *new_ptl;
1787	pmd_t pmd;
1788	struct mm_struct *mm = vma->vm_mm;
1789	bool force_flush = false;
1790
1791	if ((old_addr & ~HPAGE_PMD_MASK) ||
1792	    (new_addr & ~HPAGE_PMD_MASK) ||
1793	    old_end - old_addr < HPAGE_PMD_SIZE)
1794		return false;
1795
1796	/*
1797	 * The destination pmd shouldn't be established, free_pgtables()
1798	 * should have release it.
1799	 */
1800	if (WARN_ON(!pmd_none(*new_pmd))) {
1801		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1802		return false;
1803	}
1804
1805	/*
1806	 * We don't have to worry about the ordering of src and dst
1807	 * ptlocks because exclusive mmap_sem prevents deadlock.
1808	 */
1809	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1810	if (old_ptl) {
1811		new_ptl = pmd_lockptr(mm, new_pmd);
1812		if (new_ptl != old_ptl)
1813			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1814		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1815		if (pmd_present(pmd) && pmd_dirty(pmd))
1816			force_flush = true;
1817		VM_BUG_ON(!pmd_none(*new_pmd));
1818
1819		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1820			pgtable_t pgtable;
1821			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1822			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1823		}
1824		pmd = move_soft_dirty_pmd(pmd);
1825		set_pmd_at(mm, new_addr, new_pmd, pmd);
1826		if (new_ptl != old_ptl)
1827			spin_unlock(new_ptl);
1828		if (force_flush)
1829			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1830		else
1831			*need_flush = true;
1832		spin_unlock(old_ptl);
1833		return true;
1834	}
1835	return false;
1836}
1837
1838/*
1839 * Returns
1840 *  - 0 if PMD could not be locked
1841 *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1842 *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1843 */
1844int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1845		unsigned long addr, pgprot_t newprot, int prot_numa)
1846{
1847	struct mm_struct *mm = vma->vm_mm;
1848	spinlock_t *ptl;
1849	pmd_t entry;
1850	bool preserve_write;
1851	int ret;
1852
1853	ptl = __pmd_trans_huge_lock(pmd, vma);
1854	if (!ptl)
1855		return 0;
1856
1857	preserve_write = prot_numa && pmd_write(*pmd);
1858	ret = 1;
1859
1860#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1861	if (is_swap_pmd(*pmd)) {
1862		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1863
1864		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1865		if (is_write_migration_entry(entry)) {
1866			pmd_t newpmd;
1867			/*
1868			 * A protection check is difficult so
1869			 * just be safe and disable write
1870			 */
1871			make_migration_entry_read(&entry);
1872			newpmd = swp_entry_to_pmd(entry);
1873			if (pmd_swp_soft_dirty(*pmd))
1874				newpmd = pmd_swp_mksoft_dirty(newpmd);
1875			set_pmd_at(mm, addr, pmd, newpmd);
1876		}
1877		goto unlock;
1878	}
1879#endif
1880
1881	/*
1882	 * Avoid trapping faults against the zero page. The read-only
1883	 * data is likely to be read-cached on the local CPU and
1884	 * local/remote hits to the zero page are not interesting.
1885	 */
1886	if (prot_numa && is_huge_zero_pmd(*pmd))
1887		goto unlock;
1888
1889	if (prot_numa && pmd_protnone(*pmd))
1890		goto unlock;
1891
1892	/*
1893	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1894	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1895	 * which is also under down_read(mmap_sem):
1896	 *
1897	 *	CPU0:				CPU1:
1898	 *				change_huge_pmd(prot_numa=1)
1899	 *				 pmdp_huge_get_and_clear_notify()
1900	 * madvise_dontneed()
1901	 *  zap_pmd_range()
1902	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1903	 *   // skip the pmd
1904	 *				 set_pmd_at();
1905	 *				 // pmd is re-established
1906	 *
1907	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1908	 * which may break userspace.
1909	 *
1910	 * pmdp_invalidate() is required to make sure we don't miss
1911	 * dirty/young flags set by hardware.
1912	 */
1913	entry = pmdp_invalidate(vma, addr, pmd);
1914
1915	entry = pmd_modify(entry, newprot);
1916	if (preserve_write)
1917		entry = pmd_mk_savedwrite(entry);
1918	ret = HPAGE_PMD_NR;
1919	set_pmd_at(mm, addr, pmd, entry);
1920	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1921unlock:
1922	spin_unlock(ptl);
1923	return ret;
1924}
1925
1926/*
1927 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1928 *
1929 * Note that if it returns page table lock pointer, this routine returns without
1930 * unlocking page table lock. So callers must unlock it.
1931 */
1932spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1933{
1934	spinlock_t *ptl;
1935	ptl = pmd_lock(vma->vm_mm, pmd);
1936	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1937			pmd_devmap(*pmd)))
1938		return ptl;
1939	spin_unlock(ptl);
1940	return NULL;
1941}
1942
1943/*
1944 * Returns true if a given pud maps a thp, false otherwise.
1945 *
1946 * Note that if it returns true, this routine returns without unlocking page
1947 * table lock. So callers must unlock it.
1948 */
1949spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1950{
1951	spinlock_t *ptl;
1952
1953	ptl = pud_lock(vma->vm_mm, pud);
1954	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1955		return ptl;
1956	spin_unlock(ptl);
1957	return NULL;
1958}
1959
1960#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1961int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1962		 pud_t *pud, unsigned long addr)
1963{
1964	pud_t orig_pud;
1965	spinlock_t *ptl;
1966
1967	ptl = __pud_trans_huge_lock(pud, vma);
1968	if (!ptl)
1969		return 0;
1970	/*
1971	 * For architectures like ppc64 we look at deposited pgtable
1972	 * when calling pudp_huge_get_and_clear. So do the
1973	 * pgtable_trans_huge_withdraw after finishing pudp related
1974	 * operations.
1975	 */
1976	orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1977			tlb->fullmm);
1978	tlb_remove_pud_tlb_entry(tlb, pud, addr);
1979	if (vma_is_dax(vma)) {
1980		spin_unlock(ptl);
1981		/* No zero page support yet */
1982	} else {
1983		/* No support for anonymous PUD pages yet */
1984		BUG();
1985	}
1986	return 1;
1987}
1988
1989static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1990		unsigned long haddr)
1991{
1992	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1993	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1994	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1995	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1996
1997	count_vm_event(THP_SPLIT_PUD);
1998
1999	pudp_huge_clear_flush_notify(vma, haddr, pud);
2000}
2001
2002void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2003		unsigned long address)
2004{
2005	spinlock_t *ptl;
2006	struct mm_struct *mm = vma->vm_mm;
2007	unsigned long haddr = address & HPAGE_PUD_MASK;
2008
2009	mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2010	ptl = pud_lock(mm, pud);
 
 
 
2011	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2012		goto out;
2013	__split_huge_pud_locked(vma, pud, haddr);
2014
2015out:
2016	spin_unlock(ptl);
2017	/*
2018	 * No need to double call mmu_notifier->invalidate_range() callback as
2019	 * the above pudp_huge_clear_flush_notify() did already call it.
2020	 */
2021	mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2022					       HPAGE_PUD_SIZE);
2023}
2024#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2025
2026static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2027		unsigned long haddr, pmd_t *pmd)
2028{
2029	struct mm_struct *mm = vma->vm_mm;
2030	pgtable_t pgtable;
2031	pmd_t _pmd;
2032	int i;
2033
2034	/*
2035	 * Leave pmd empty until pte is filled note that it is fine to delay
2036	 * notification until mmu_notifier_invalidate_range_end() as we are
2037	 * replacing a zero pmd write protected page with a zero pte write
2038	 * protected page.
2039	 *
2040	 * See Documentation/vm/mmu_notifier.txt
2041	 */
2042	pmdp_huge_clear_flush(vma, haddr, pmd);
2043
2044	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2045	pmd_populate(mm, &_pmd, pgtable);
2046
2047	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2048		pte_t *pte, entry;
2049		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2050		entry = pte_mkspecial(entry);
2051		pte = pte_offset_map(&_pmd, haddr);
2052		VM_BUG_ON(!pte_none(*pte));
2053		set_pte_at(mm, haddr, pte, entry);
2054		pte_unmap(pte);
2055	}
2056	smp_wmb(); /* make pte visible before pmd */
2057	pmd_populate(mm, pmd, pgtable);
2058}
2059
2060static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2061		unsigned long haddr, bool freeze)
2062{
2063	struct mm_struct *mm = vma->vm_mm;
2064	struct page *page;
2065	pgtable_t pgtable;
2066	pmd_t old_pmd, _pmd;
2067	bool young, write, soft_dirty, pmd_migration = false;
2068	unsigned long addr;
2069	int i;
2070
2071	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2072	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2073	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2074	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2075				&& !pmd_devmap(*pmd));
2076
2077	count_vm_event(THP_SPLIT_PMD);
2078
2079	if (!vma_is_anonymous(vma)) {
2080		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2081		/*
2082		 * We are going to unmap this huge page. So
2083		 * just go ahead and zap it
2084		 */
2085		if (arch_needs_pgtable_deposit())
2086			zap_deposited_table(mm, pmd);
2087		if (vma_is_dax(vma))
2088			return;
2089		page = pmd_page(_pmd);
 
 
2090		if (!PageReferenced(page) && pmd_young(_pmd))
2091			SetPageReferenced(page);
2092		page_remove_rmap(page, true);
2093		put_page(page);
2094		add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2095		return;
2096	} else if (is_huge_zero_pmd(*pmd)) {
2097		/*
2098		 * FIXME: Do we want to invalidate secondary mmu by calling
2099		 * mmu_notifier_invalidate_range() see comments below inside
2100		 * __split_huge_pmd() ?
2101		 *
2102		 * We are going from a zero huge page write protected to zero
2103		 * small page also write protected so it does not seems useful
2104		 * to invalidate secondary mmu at this time.
2105		 */
2106		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2107	}
2108
2109	/*
2110	 * Up to this point the pmd is present and huge and userland has the
2111	 * whole access to the hugepage during the split (which happens in
2112	 * place). If we overwrite the pmd with the not-huge version pointing
2113	 * to the pte here (which of course we could if all CPUs were bug
2114	 * free), userland could trigger a small page size TLB miss on the
2115	 * small sized TLB while the hugepage TLB entry is still established in
2116	 * the huge TLB. Some CPU doesn't like that.
2117	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2118	 * 383 on page 93. Intel should be safe but is also warns that it's
2119	 * only safe if the permission and cache attributes of the two entries
2120	 * loaded in the two TLB is identical (which should be the case here).
2121	 * But it is generally safer to never allow small and huge TLB entries
2122	 * for the same virtual address to be loaded simultaneously. So instead
2123	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2124	 * current pmd notpresent (atomically because here the pmd_trans_huge
2125	 * must remain set at all times on the pmd until the split is complete
2126	 * for this pmd), then we flush the SMP TLB and finally we write the
2127	 * non-huge version of the pmd entry with pmd_populate.
2128	 */
2129	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2130
2131#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2132	pmd_migration = is_pmd_migration_entry(old_pmd);
2133	if (pmd_migration) {
2134		swp_entry_t entry;
2135
2136		entry = pmd_to_swp_entry(old_pmd);
2137		page = pfn_to_page(swp_offset(entry));
2138	} else
2139#endif
 
 
2140		page = pmd_page(old_pmd);
 
 
 
 
 
 
2141	VM_BUG_ON_PAGE(!page_count(page), page);
2142	page_ref_add(page, HPAGE_PMD_NR - 1);
2143	if (pmd_dirty(old_pmd))
2144		SetPageDirty(page);
2145	write = pmd_write(old_pmd);
2146	young = pmd_young(old_pmd);
2147	soft_dirty = pmd_soft_dirty(old_pmd);
2148
2149	/*
2150	 * Withdraw the table only after we mark the pmd entry invalid.
2151	 * This's critical for some architectures (Power).
2152	 */
2153	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2154	pmd_populate(mm, &_pmd, pgtable);
2155
2156	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2157		pte_t entry, *pte;
2158		/*
2159		 * Note that NUMA hinting access restrictions are not
2160		 * transferred to avoid any possibility of altering
2161		 * permissions across VMAs.
2162		 */
2163		if (freeze || pmd_migration) {
2164			swp_entry_t swp_entry;
2165			swp_entry = make_migration_entry(page + i, write);
2166			entry = swp_entry_to_pte(swp_entry);
2167			if (soft_dirty)
2168				entry = pte_swp_mksoft_dirty(entry);
2169		} else {
2170			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2171			entry = maybe_mkwrite(entry, vma);
2172			if (!write)
2173				entry = pte_wrprotect(entry);
2174			if (!young)
2175				entry = pte_mkold(entry);
2176			if (soft_dirty)
2177				entry = pte_mksoft_dirty(entry);
2178		}
2179		pte = pte_offset_map(&_pmd, addr);
2180		BUG_ON(!pte_none(*pte));
2181		set_pte_at(mm, addr, pte, entry);
2182		atomic_inc(&page[i]._mapcount);
2183		pte_unmap(pte);
2184	}
2185
2186	/*
2187	 * Set PG_double_map before dropping compound_mapcount to avoid
2188	 * false-negative page_mapped().
2189	 */
2190	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2191		for (i = 0; i < HPAGE_PMD_NR; i++)
2192			atomic_inc(&page[i]._mapcount);
2193	}
2194
2195	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2196		/* Last compound_mapcount is gone. */
2197		__dec_node_page_state(page, NR_ANON_THPS);
2198		if (TestClearPageDoubleMap(page)) {
2199			/* No need in mapcount reference anymore */
2200			for (i = 0; i < HPAGE_PMD_NR; i++)
2201				atomic_dec(&page[i]._mapcount);
2202		}
2203	}
2204
2205	smp_wmb(); /* make pte visible before pmd */
2206	pmd_populate(mm, pmd, pgtable);
2207
2208	if (freeze) {
2209		for (i = 0; i < HPAGE_PMD_NR; i++) {
2210			page_remove_rmap(page + i, false);
2211			put_page(page + i);
2212		}
2213	}
2214}
2215
2216void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2217		unsigned long address, bool freeze, struct page *page)
2218{
2219	spinlock_t *ptl;
2220	struct mm_struct *mm = vma->vm_mm;
2221	unsigned long haddr = address & HPAGE_PMD_MASK;
2222
2223	mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2224	ptl = pmd_lock(mm, pmd);
 
 
 
2225
2226	/*
2227	 * If caller asks to setup a migration entries, we need a page to check
2228	 * pmd against. Otherwise we can end up replacing wrong page.
2229	 */
2230	VM_BUG_ON(freeze && !page);
2231	if (page && page != pmd_page(*pmd))
2232	        goto out;
2233
2234	if (pmd_trans_huge(*pmd)) {
2235		page = pmd_page(*pmd);
2236		if (PageMlocked(page))
2237			clear_page_mlock(page);
2238	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2239		goto out;
2240	__split_huge_pmd_locked(vma, pmd, haddr, freeze);
2241out:
2242	spin_unlock(ptl);
2243	/*
2244	 * No need to double call mmu_notifier->invalidate_range() callback.
2245	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2246	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2247	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2248	 *    fault will trigger a flush_notify before pointing to a new page
2249	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2250	 *    page in the meantime)
2251	 *  3) Split a huge pmd into pte pointing to the same page. No need
2252	 *     to invalidate secondary tlb entry they are all still valid.
2253	 *     any further changes to individual pte will notify. So no need
2254	 *     to call mmu_notifier->invalidate_range()
2255	 */
2256	mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2257					       HPAGE_PMD_SIZE);
2258}
2259
2260void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2261		bool freeze, struct page *page)
2262{
2263	pgd_t *pgd;
2264	p4d_t *p4d;
2265	pud_t *pud;
2266	pmd_t *pmd;
2267
2268	pgd = pgd_offset(vma->vm_mm, address);
2269	if (!pgd_present(*pgd))
2270		return;
2271
2272	p4d = p4d_offset(pgd, address);
2273	if (!p4d_present(*p4d))
2274		return;
2275
2276	pud = pud_offset(p4d, address);
2277	if (!pud_present(*pud))
2278		return;
2279
2280	pmd = pmd_offset(pud, address);
2281
2282	__split_huge_pmd(vma, pmd, address, freeze, page);
2283}
2284
2285void vma_adjust_trans_huge(struct vm_area_struct *vma,
2286			     unsigned long start,
2287			     unsigned long end,
2288			     long adjust_next)
2289{
2290	/*
2291	 * If the new start address isn't hpage aligned and it could
2292	 * previously contain an hugepage: check if we need to split
2293	 * an huge pmd.
2294	 */
2295	if (start & ~HPAGE_PMD_MASK &&
2296	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2297	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2298		split_huge_pmd_address(vma, start, false, NULL);
2299
2300	/*
2301	 * If the new end address isn't hpage aligned and it could
2302	 * previously contain an hugepage: check if we need to split
2303	 * an huge pmd.
2304	 */
2305	if (end & ~HPAGE_PMD_MASK &&
2306	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2307	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2308		split_huge_pmd_address(vma, end, false, NULL);
2309
2310	/*
2311	 * If we're also updating the vma->vm_next->vm_start, if the new
2312	 * vm_next->vm_start isn't page aligned and it could previously
2313	 * contain an hugepage: check if we need to split an huge pmd.
2314	 */
2315	if (adjust_next > 0) {
2316		struct vm_area_struct *next = vma->vm_next;
2317		unsigned long nstart = next->vm_start;
2318		nstart += adjust_next << PAGE_SHIFT;
2319		if (nstart & ~HPAGE_PMD_MASK &&
2320		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2321		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2322			split_huge_pmd_address(next, nstart, false, NULL);
2323	}
2324}
2325
2326static void freeze_page(struct page *page)
2327{
2328	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2329		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2330	bool unmap_success;
2331
2332	VM_BUG_ON_PAGE(!PageHead(page), page);
2333
2334	if (PageAnon(page))
2335		ttu_flags |= TTU_SPLIT_FREEZE;
2336
2337	unmap_success = try_to_unmap(page, ttu_flags);
2338	VM_BUG_ON_PAGE(!unmap_success, page);
2339}
2340
2341static void unfreeze_page(struct page *page)
2342{
2343	int i;
2344	if (PageTransHuge(page)) {
2345		remove_migration_ptes(page, page, true);
2346	} else {
2347		for (i = 0; i < HPAGE_PMD_NR; i++)
2348			remove_migration_ptes(page + i, page + i, true);
2349	}
2350}
2351
2352static void __split_huge_page_tail(struct page *head, int tail,
2353		struct lruvec *lruvec, struct list_head *list)
2354{
2355	struct page *page_tail = head + tail;
2356
2357	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2358
2359	/*
2360	 * Clone page flags before unfreezing refcount.
2361	 *
2362	 * After successful get_page_unless_zero() might follow flags change,
2363	 * for exmaple lock_page() which set PG_waiters.
2364	 */
2365	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2366	page_tail->flags |= (head->flags &
2367			((1L << PG_referenced) |
2368			 (1L << PG_swapbacked) |
2369			 (1L << PG_swapcache) |
2370			 (1L << PG_mlocked) |
2371			 (1L << PG_uptodate) |
2372			 (1L << PG_active) |
 
2373			 (1L << PG_locked) |
2374			 (1L << PG_unevictable) |
2375			 (1L << PG_dirty)));
2376
 
 
 
 
 
 
2377	/* Page flags must be visible before we make the page non-compound. */
2378	smp_wmb();
2379
2380	/*
2381	 * Clear PageTail before unfreezing page refcount.
2382	 *
2383	 * After successful get_page_unless_zero() might follow put_page()
2384	 * which needs correct compound_head().
2385	 */
2386	clear_compound_head(page_tail);
2387
2388	/* Finally unfreeze refcount. Additional reference from page cache. */
2389	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2390					  PageSwapCache(head)));
2391
2392	if (page_is_young(head))
2393		set_page_young(page_tail);
2394	if (page_is_idle(head))
2395		set_page_idle(page_tail);
2396
2397	/* ->mapping in first tail page is compound_mapcount */
2398	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2399			page_tail);
2400	page_tail->mapping = head->mapping;
2401
2402	page_tail->index = head->index + tail;
2403	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2404
2405	/*
2406	 * always add to the tail because some iterators expect new
2407	 * pages to show after the currently processed elements - e.g.
2408	 * migrate_pages
2409	 */
2410	lru_add_page_tail(head, page_tail, lruvec, list);
2411}
2412
2413static void __split_huge_page(struct page *page, struct list_head *list,
2414		unsigned long flags)
2415{
2416	struct page *head = compound_head(page);
2417	struct zone *zone = page_zone(head);
2418	struct lruvec *lruvec;
2419	pgoff_t end = -1;
 
2420	int i;
2421
2422	lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2423
2424	/* complete memcg works before add pages to LRU */
2425	mem_cgroup_split_huge_fixup(head);
2426
2427	if (!PageAnon(page))
2428		end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
 
 
 
 
 
2429
2430	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2431		__split_huge_page_tail(head, i, lruvec, list);
2432		/* Some pages can be beyond i_size: drop them from page cache */
2433		if (head[i].index >= end) {
2434			ClearPageDirty(head + i);
2435			__delete_from_page_cache(head + i, NULL);
2436			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2437				shmem_uncharge(head->mapping->host, 1);
2438			put_page(head + i);
 
 
 
 
 
 
2439		}
2440	}
2441
2442	ClearPageCompound(head);
 
 
 
2443	/* See comment in __split_huge_page_tail() */
2444	if (PageAnon(head)) {
2445		/* Additional pin to radix tree of swap cache */
2446		if (PageSwapCache(head))
2447			page_ref_add(head, 2);
2448		else
 
2449			page_ref_inc(head);
 
2450	} else {
2451		/* Additional pin to radix tree */
2452		page_ref_add(head, 2);
2453		xa_unlock(&head->mapping->i_pages);
2454	}
2455
2456	spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2457
2458	unfreeze_page(head);
2459
2460	for (i = 0; i < HPAGE_PMD_NR; i++) {
2461		struct page *subpage = head + i;
2462		if (subpage == page)
2463			continue;
2464		unlock_page(subpage);
2465
2466		/*
2467		 * Subpages may be freed if there wasn't any mapping
2468		 * like if add_to_swap() is running on a lru page that
2469		 * had its mapping zapped. And freeing these pages
2470		 * requires taking the lru_lock so we do the put_page
2471		 * of the tail pages after the split is complete.
2472		 */
2473		put_page(subpage);
2474	}
2475}
2476
2477int total_mapcount(struct page *page)
2478{
2479	int i, compound, ret;
2480
2481	VM_BUG_ON_PAGE(PageTail(page), page);
2482
2483	if (likely(!PageCompound(page)))
2484		return atomic_read(&page->_mapcount) + 1;
2485
2486	compound = compound_mapcount(page);
2487	if (PageHuge(page))
2488		return compound;
2489	ret = compound;
2490	for (i = 0; i < HPAGE_PMD_NR; i++)
2491		ret += atomic_read(&page[i]._mapcount) + 1;
2492	/* File pages has compound_mapcount included in _mapcount */
2493	if (!PageAnon(page))
2494		return ret - compound * HPAGE_PMD_NR;
2495	if (PageDoubleMap(page))
2496		ret -= HPAGE_PMD_NR;
2497	return ret;
2498}
2499
2500/*
2501 * This calculates accurately how many mappings a transparent hugepage
2502 * has (unlike page_mapcount() which isn't fully accurate). This full
2503 * accuracy is primarily needed to know if copy-on-write faults can
2504 * reuse the page and change the mapping to read-write instead of
2505 * copying them. At the same time this returns the total_mapcount too.
2506 *
2507 * The function returns the highest mapcount any one of the subpages
2508 * has. If the return value is one, even if different processes are
2509 * mapping different subpages of the transparent hugepage, they can
2510 * all reuse it, because each process is reusing a different subpage.
2511 *
2512 * The total_mapcount is instead counting all virtual mappings of the
2513 * subpages. If the total_mapcount is equal to "one", it tells the
2514 * caller all mappings belong to the same "mm" and in turn the
2515 * anon_vma of the transparent hugepage can become the vma->anon_vma
2516 * local one as no other process may be mapping any of the subpages.
2517 *
2518 * It would be more accurate to replace page_mapcount() with
2519 * page_trans_huge_mapcount(), however we only use
2520 * page_trans_huge_mapcount() in the copy-on-write faults where we
2521 * need full accuracy to avoid breaking page pinning, because
2522 * page_trans_huge_mapcount() is slower than page_mapcount().
2523 */
2524int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2525{
2526	int i, ret, _total_mapcount, mapcount;
2527
2528	/* hugetlbfs shouldn't call it */
2529	VM_BUG_ON_PAGE(PageHuge(page), page);
2530
2531	if (likely(!PageTransCompound(page))) {
2532		mapcount = atomic_read(&page->_mapcount) + 1;
2533		if (total_mapcount)
2534			*total_mapcount = mapcount;
2535		return mapcount;
2536	}
2537
2538	page = compound_head(page);
2539
2540	_total_mapcount = ret = 0;
2541	for (i = 0; i < HPAGE_PMD_NR; i++) {
2542		mapcount = atomic_read(&page[i]._mapcount) + 1;
2543		ret = max(ret, mapcount);
2544		_total_mapcount += mapcount;
2545	}
2546	if (PageDoubleMap(page)) {
2547		ret -= 1;
2548		_total_mapcount -= HPAGE_PMD_NR;
2549	}
2550	mapcount = compound_mapcount(page);
2551	ret += mapcount;
2552	_total_mapcount += mapcount;
2553	if (total_mapcount)
2554		*total_mapcount = _total_mapcount;
2555	return ret;
2556}
2557
2558/* Racy check whether the huge page can be split */
2559bool can_split_huge_page(struct page *page, int *pextra_pins)
2560{
2561	int extra_pins;
2562
2563	/* Additional pins from radix tree */
2564	if (PageAnon(page))
2565		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2566	else
2567		extra_pins = HPAGE_PMD_NR;
2568	if (pextra_pins)
2569		*pextra_pins = extra_pins;
2570	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2571}
2572
2573/*
2574 * This function splits huge page into normal pages. @page can point to any
2575 * subpage of huge page to split. Split doesn't change the position of @page.
2576 *
2577 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2578 * The huge page must be locked.
2579 *
2580 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2581 *
2582 * Both head page and tail pages will inherit mapping, flags, and so on from
2583 * the hugepage.
2584 *
2585 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2586 * they are not mapped.
2587 *
2588 * Returns 0 if the hugepage is split successfully.
2589 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2590 * us.
2591 */
2592int split_huge_page_to_list(struct page *page, struct list_head *list)
2593{
2594	struct page *head = compound_head(page);
2595	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
 
2596	struct anon_vma *anon_vma = NULL;
2597	struct address_space *mapping = NULL;
2598	int count, mapcount, extra_pins, ret;
2599	bool mlocked;
2600	unsigned long flags;
 
2601
2602	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2603	VM_BUG_ON_PAGE(!PageLocked(page), page);
2604	VM_BUG_ON_PAGE(!PageCompound(page), page);
2605
2606	if (PageWriteback(page))
2607		return -EBUSY;
2608
2609	if (PageAnon(head)) {
2610		/*
2611		 * The caller does not necessarily hold an mmap_sem that would
2612		 * prevent the anon_vma disappearing so we first we take a
2613		 * reference to it and then lock the anon_vma for write. This
2614		 * is similar to page_lock_anon_vma_read except the write lock
2615		 * is taken to serialise against parallel split or collapse
2616		 * operations.
2617		 */
2618		anon_vma = page_get_anon_vma(head);
2619		if (!anon_vma) {
2620			ret = -EBUSY;
2621			goto out;
2622		}
 
2623		mapping = NULL;
2624		anon_vma_lock_write(anon_vma);
2625	} else {
2626		mapping = head->mapping;
2627
2628		/* Truncated ? */
2629		if (!mapping) {
2630			ret = -EBUSY;
2631			goto out;
2632		}
2633
2634		anon_vma = NULL;
2635		i_mmap_lock_read(mapping);
 
 
 
 
 
 
 
 
 
2636	}
2637
2638	/*
2639	 * Racy check if we can split the page, before freeze_page() will
2640	 * split PMDs
2641	 */
2642	if (!can_split_huge_page(head, &extra_pins)) {
2643		ret = -EBUSY;
2644		goto out_unlock;
2645	}
2646
2647	mlocked = PageMlocked(page);
2648	freeze_page(head);
2649	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2650
2651	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2652	if (mlocked)
2653		lru_add_drain();
2654
2655	/* prevent PageLRU to go away from under us, and freeze lru stats */
2656	spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2657
2658	if (mapping) {
2659		void **pslot;
2660
2661		xa_lock(&mapping->i_pages);
2662		pslot = radix_tree_lookup_slot(&mapping->i_pages,
2663				page_index(head));
2664		/*
2665		 * Check if the head page is present in radix tree.
2666		 * We assume all tail are present too, if head is there.
2667		 */
2668		if (radix_tree_deref_slot_protected(pslot,
2669					&mapping->i_pages.xa_lock) != head)
2670			goto fail;
2671	}
2672
2673	/* Prevent deferred_split_scan() touching ->_refcount */
2674	spin_lock(&pgdata->split_queue_lock);
2675	count = page_count(head);
2676	mapcount = total_mapcount(head);
2677	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2678		if (!list_empty(page_deferred_list(head))) {
2679			pgdata->split_queue_len--;
2680			list_del(page_deferred_list(head));
2681		}
2682		if (mapping)
2683			__dec_node_page_state(page, NR_SHMEM_THPS);
2684		spin_unlock(&pgdata->split_queue_lock);
2685		__split_huge_page(page, list, flags);
 
 
 
 
 
2686		if (PageSwapCache(head)) {
2687			swp_entry_t entry = { .val = page_private(head) };
2688
2689			ret = split_swap_cluster(entry);
2690		} else
2691			ret = 0;
2692	} else {
2693		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2694			pr_alert("total_mapcount: %u, page_count(): %u\n",
2695					mapcount, count);
2696			if (PageTail(page))
2697				dump_page(head, NULL);
2698			dump_page(page, "total_mapcount(head) > 0");
2699			BUG();
2700		}
2701		spin_unlock(&pgdata->split_queue_lock);
2702fail:		if (mapping)
2703			xa_unlock(&mapping->i_pages);
2704		spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2705		unfreeze_page(head);
2706		ret = -EBUSY;
2707	}
2708
2709out_unlock:
2710	if (anon_vma) {
2711		anon_vma_unlock_write(anon_vma);
2712		put_anon_vma(anon_vma);
2713	}
2714	if (mapping)
2715		i_mmap_unlock_read(mapping);
2716out:
2717	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2718	return ret;
2719}
2720
2721void free_transhuge_page(struct page *page)
2722{
2723	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2724	unsigned long flags;
2725
2726	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2727	if (!list_empty(page_deferred_list(page))) {
2728		pgdata->split_queue_len--;
2729		list_del(page_deferred_list(page));
2730	}
2731	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2732	free_compound_page(page);
2733}
2734
2735void deferred_split_huge_page(struct page *page)
2736{
2737	struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
 
 
 
2738	unsigned long flags;
2739
2740	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2741
2742	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
2743	if (list_empty(page_deferred_list(page))) {
2744		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2745		list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2746		pgdata->split_queue_len++;
 
 
 
 
 
2747	}
2748	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2749}
2750
2751static unsigned long deferred_split_count(struct shrinker *shrink,
2752		struct shrink_control *sc)
2753{
2754	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2755	return READ_ONCE(pgdata->split_queue_len);
 
 
 
 
 
 
2756}
2757
2758static unsigned long deferred_split_scan(struct shrinker *shrink,
2759		struct shrink_control *sc)
2760{
2761	struct pglist_data *pgdata = NODE_DATA(sc->nid);
 
2762	unsigned long flags;
2763	LIST_HEAD(list), *pos, *next;
2764	struct page *page;
2765	int split = 0;
2766
2767	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
 
 
 
 
 
2768	/* Take pin on all head pages to avoid freeing them under us */
2769	list_for_each_safe(pos, next, &pgdata->split_queue) {
2770		page = list_entry((void *)pos, struct page, mapping);
2771		page = compound_head(page);
2772		if (get_page_unless_zero(page)) {
2773			list_move(page_deferred_list(page), &list);
2774		} else {
2775			/* We lost race with put_compound_page() */
2776			list_del_init(page_deferred_list(page));
2777			pgdata->split_queue_len--;
2778		}
2779		if (!--sc->nr_to_scan)
2780			break;
2781	}
2782	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2783
2784	list_for_each_safe(pos, next, &list) {
2785		page = list_entry((void *)pos, struct page, mapping);
2786		if (!trylock_page(page))
2787			goto next;
2788		/* split_huge_page() removes page from list on success */
2789		if (!split_huge_page(page))
2790			split++;
2791		unlock_page(page);
2792next:
2793		put_page(page);
2794	}
2795
2796	spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2797	list_splice_tail(&list, &pgdata->split_queue);
2798	spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2799
2800	/*
2801	 * Stop shrinker if we didn't split any page, but the queue is empty.
2802	 * This can happen if pages were freed under us.
2803	 */
2804	if (!split && list_empty(&pgdata->split_queue))
2805		return SHRINK_STOP;
2806	return split;
2807}
2808
2809static struct shrinker deferred_split_shrinker = {
2810	.count_objects = deferred_split_count,
2811	.scan_objects = deferred_split_scan,
2812	.seeks = DEFAULT_SEEKS,
2813	.flags = SHRINKER_NUMA_AWARE,
 
2814};
2815
2816#ifdef CONFIG_DEBUG_FS
2817static int split_huge_pages_set(void *data, u64 val)
2818{
2819	struct zone *zone;
2820	struct page *page;
2821	unsigned long pfn, max_zone_pfn;
2822	unsigned long total = 0, split = 0;
2823
2824	if (val != 1)
2825		return -EINVAL;
2826
2827	for_each_populated_zone(zone) {
2828		max_zone_pfn = zone_end_pfn(zone);
2829		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2830			if (!pfn_valid(pfn))
2831				continue;
2832
2833			page = pfn_to_page(pfn);
2834			if (!get_page_unless_zero(page))
2835				continue;
2836
2837			if (zone != page_zone(page))
2838				goto next;
2839
2840			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2841				goto next;
2842
2843			total++;
2844			lock_page(page);
2845			if (!split_huge_page(page))
2846				split++;
2847			unlock_page(page);
2848next:
2849			put_page(page);
2850		}
2851	}
2852
2853	pr_info("%lu of %lu THP split\n", split, total);
2854
2855	return 0;
2856}
2857DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2858		"%llu\n");
2859
2860static int __init split_huge_pages_debugfs(void)
2861{
2862	void *ret;
2863
2864	ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2865			&split_huge_pages_fops);
2866	if (!ret)
2867		pr_warn("Failed to create split_huge_pages in debugfs");
2868	return 0;
2869}
2870late_initcall(split_huge_pages_debugfs);
2871#endif
2872
2873#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2874void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2875		struct page *page)
2876{
2877	struct vm_area_struct *vma = pvmw->vma;
2878	struct mm_struct *mm = vma->vm_mm;
2879	unsigned long address = pvmw->address;
2880	pmd_t pmdval;
2881	swp_entry_t entry;
2882	pmd_t pmdswp;
2883
2884	if (!(pvmw->pmd && !pvmw->pte))
2885		return;
2886
2887	mmu_notifier_invalidate_range_start(mm, address,
2888			address + HPAGE_PMD_SIZE);
2889
2890	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2891	pmdval = *pvmw->pmd;
2892	pmdp_invalidate(vma, address, pvmw->pmd);
2893	if (pmd_dirty(pmdval))
2894		set_page_dirty(page);
2895	entry = make_migration_entry(page, pmd_write(pmdval));
2896	pmdswp = swp_entry_to_pmd(entry);
2897	if (pmd_soft_dirty(pmdval))
2898		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2899	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2900	page_remove_rmap(page, true);
2901	put_page(page);
2902
2903	mmu_notifier_invalidate_range_end(mm, address,
2904			address + HPAGE_PMD_SIZE);
2905}
2906
2907void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2908{
2909	struct vm_area_struct *vma = pvmw->vma;
2910	struct mm_struct *mm = vma->vm_mm;
2911	unsigned long address = pvmw->address;
2912	unsigned long mmun_start = address & HPAGE_PMD_MASK;
2913	pmd_t pmde;
2914	swp_entry_t entry;
2915
2916	if (!(pvmw->pmd && !pvmw->pte))
2917		return;
2918
2919	entry = pmd_to_swp_entry(*pvmw->pmd);
2920	get_page(new);
2921	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2922	if (pmd_swp_soft_dirty(*pvmw->pmd))
2923		pmde = pmd_mksoft_dirty(pmde);
2924	if (is_write_migration_entry(entry))
2925		pmde = maybe_pmd_mkwrite(pmde, vma);
2926
2927	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2928	if (PageAnon(new))
2929		page_add_anon_rmap(new, vma, mmun_start, true);
2930	else
2931		page_add_file_rmap(new, true);
2932	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2933	if (vma->vm_flags & VM_LOCKED)
2934		mlock_vma_page(new);
2935	update_mmu_cache_pmd(vma, address, pvmw->pmd);
2936}
2937#endif
v5.4
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  Copyright (C) 2009  Red Hat, Inc.
 
 
 
   4 */
   5
   6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7
   8#include <linux/mm.h>
   9#include <linux/sched.h>
  10#include <linux/sched/coredump.h>
  11#include <linux/sched/numa_balancing.h>
  12#include <linux/highmem.h>
  13#include <linux/hugetlb.h>
  14#include <linux/mmu_notifier.h>
  15#include <linux/rmap.h>
  16#include <linux/swap.h>
  17#include <linux/shrinker.h>
  18#include <linux/mm_inline.h>
  19#include <linux/swapops.h>
  20#include <linux/dax.h>
  21#include <linux/khugepaged.h>
  22#include <linux/freezer.h>
  23#include <linux/pfn_t.h>
  24#include <linux/mman.h>
  25#include <linux/memremap.h>
  26#include <linux/pagemap.h>
  27#include <linux/debugfs.h>
  28#include <linux/migrate.h>
  29#include <linux/hashtable.h>
  30#include <linux/userfaultfd_k.h>
  31#include <linux/page_idle.h>
  32#include <linux/shmem_fs.h>
  33#include <linux/oom.h>
  34#include <linux/numa.h>
  35#include <linux/page_owner.h>
  36
  37#include <asm/tlb.h>
  38#include <asm/pgalloc.h>
  39#include "internal.h"
  40
  41/*
  42 * By default, transparent hugepage support is disabled in order to avoid
  43 * risking an increased memory footprint for applications that are not
  44 * guaranteed to benefit from it. When transparent hugepage support is
  45 * enabled, it is for all mappings, and khugepaged scans all mappings.
  46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
  47 * for all hugepage allocations.
  48 */
  49unsigned long transparent_hugepage_flags __read_mostly =
  50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  51	(1<<TRANSPARENT_HUGEPAGE_FLAG)|
  52#endif
  53#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  54	(1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  55#endif
  56	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
  57	(1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
  58	(1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
  59
  60static struct shrinker deferred_split_shrinker;
  61
  62static atomic_t huge_zero_refcount;
  63struct page *huge_zero_page __read_mostly;
  64
  65bool transparent_hugepage_enabled(struct vm_area_struct *vma)
  66{
  67	/* The addr is used to check if the vma size fits */
  68	unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
  69
  70	if (!transhuge_vma_suitable(vma, addr))
  71		return false;
  72	if (vma_is_anonymous(vma))
  73		return __transparent_hugepage_enabled(vma);
  74	if (vma_is_shmem(vma))
  75		return shmem_huge_enabled(vma);
  76
  77	return false;
  78}
  79
  80static struct page *get_huge_zero_page(void)
  81{
  82	struct page *zero_page;
  83retry:
  84	if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
  85		return READ_ONCE(huge_zero_page);
  86
  87	zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
  88			HPAGE_PMD_ORDER);
  89	if (!zero_page) {
  90		count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
  91		return NULL;
  92	}
  93	count_vm_event(THP_ZERO_PAGE_ALLOC);
  94	preempt_disable();
  95	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
  96		preempt_enable();
  97		__free_pages(zero_page, compound_order(zero_page));
  98		goto retry;
  99	}
 100
 101	/* We take additional reference here. It will be put back by shrinker */
 102	atomic_set(&huge_zero_refcount, 2);
 103	preempt_enable();
 104	return READ_ONCE(huge_zero_page);
 105}
 106
 107static void put_huge_zero_page(void)
 108{
 109	/*
 110	 * Counter should never go to zero here. Only shrinker can put
 111	 * last reference.
 112	 */
 113	BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
 114}
 115
 116struct page *mm_get_huge_zero_page(struct mm_struct *mm)
 117{
 118	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 119		return READ_ONCE(huge_zero_page);
 120
 121	if (!get_huge_zero_page())
 122		return NULL;
 123
 124	if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 125		put_huge_zero_page();
 126
 127	return READ_ONCE(huge_zero_page);
 128}
 129
 130void mm_put_huge_zero_page(struct mm_struct *mm)
 131{
 132	if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
 133		put_huge_zero_page();
 134}
 135
 136static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
 137					struct shrink_control *sc)
 138{
 139	/* we can free zero page only if last reference remains */
 140	return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
 141}
 142
 143static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 144				       struct shrink_control *sc)
 145{
 146	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 147		struct page *zero_page = xchg(&huge_zero_page, NULL);
 148		BUG_ON(zero_page == NULL);
 149		__free_pages(zero_page, compound_order(zero_page));
 150		return HPAGE_PMD_NR;
 151	}
 152
 153	return 0;
 154}
 155
 156static struct shrinker huge_zero_page_shrinker = {
 157	.count_objects = shrink_huge_zero_page_count,
 158	.scan_objects = shrink_huge_zero_page_scan,
 159	.seeks = DEFAULT_SEEKS,
 160};
 161
 162#ifdef CONFIG_SYSFS
 163static ssize_t enabled_show(struct kobject *kobj,
 164			    struct kobj_attribute *attr, char *buf)
 165{
 166	if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
 167		return sprintf(buf, "[always] madvise never\n");
 168	else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
 169		return sprintf(buf, "always [madvise] never\n");
 170	else
 171		return sprintf(buf, "always madvise [never]\n");
 172}
 173
 174static ssize_t enabled_store(struct kobject *kobj,
 175			     struct kobj_attribute *attr,
 176			     const char *buf, size_t count)
 177{
 178	ssize_t ret = count;
 179
 180	if (!memcmp("always", buf,
 181		    min(sizeof("always")-1, count))) {
 182		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 183		set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 184	} else if (!memcmp("madvise", buf,
 185			   min(sizeof("madvise")-1, count))) {
 186		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 187		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 188	} else if (!memcmp("never", buf,
 189			   min(sizeof("never")-1, count))) {
 190		clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
 191		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
 192	} else
 193		ret = -EINVAL;
 194
 195	if (ret > 0) {
 196		int err = start_stop_khugepaged();
 197		if (err)
 198			ret = err;
 199	}
 200	return ret;
 201}
 202static struct kobj_attribute enabled_attr =
 203	__ATTR(enabled, 0644, enabled_show, enabled_store);
 204
 205ssize_t single_hugepage_flag_show(struct kobject *kobj,
 206				struct kobj_attribute *attr, char *buf,
 207				enum transparent_hugepage_flag flag)
 208{
 209	return sprintf(buf, "%d\n",
 210		       !!test_bit(flag, &transparent_hugepage_flags));
 211}
 212
 213ssize_t single_hugepage_flag_store(struct kobject *kobj,
 214				 struct kobj_attribute *attr,
 215				 const char *buf, size_t count,
 216				 enum transparent_hugepage_flag flag)
 217{
 218	unsigned long value;
 219	int ret;
 220
 221	ret = kstrtoul(buf, 10, &value);
 222	if (ret < 0)
 223		return ret;
 224	if (value > 1)
 225		return -EINVAL;
 226
 227	if (value)
 228		set_bit(flag, &transparent_hugepage_flags);
 229	else
 230		clear_bit(flag, &transparent_hugepage_flags);
 231
 232	return count;
 233}
 234
 235static ssize_t defrag_show(struct kobject *kobj,
 236			   struct kobj_attribute *attr, char *buf)
 237{
 238	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 239		return sprintf(buf, "[always] defer defer+madvise madvise never\n");
 240	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 241		return sprintf(buf, "always [defer] defer+madvise madvise never\n");
 242	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 243		return sprintf(buf, "always defer [defer+madvise] madvise never\n");
 244	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 245		return sprintf(buf, "always defer defer+madvise [madvise] never\n");
 246	return sprintf(buf, "always defer defer+madvise madvise [never]\n");
 247}
 248
 249static ssize_t defrag_store(struct kobject *kobj,
 250			    struct kobj_attribute *attr,
 251			    const char *buf, size_t count)
 252{
 253	if (!memcmp("always", buf,
 254		    min(sizeof("always")-1, count))) {
 255		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 256		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 257		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 258		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 259	} else if (!memcmp("defer+madvise", buf,
 260		    min(sizeof("defer+madvise")-1, count))) {
 261		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 262		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 263		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 264		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 265	} else if (!memcmp("defer", buf,
 266		    min(sizeof("defer")-1, count))) {
 267		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 268		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 269		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 270		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 271	} else if (!memcmp("madvise", buf,
 272			   min(sizeof("madvise")-1, count))) {
 273		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 274		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 275		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 276		set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 277	} else if (!memcmp("never", buf,
 278			   min(sizeof("never")-1, count))) {
 279		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
 280		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
 281		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
 282		clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
 283	} else
 284		return -EINVAL;
 285
 286	return count;
 287}
 288static struct kobj_attribute defrag_attr =
 289	__ATTR(defrag, 0644, defrag_show, defrag_store);
 290
 291static ssize_t use_zero_page_show(struct kobject *kobj,
 292		struct kobj_attribute *attr, char *buf)
 293{
 294	return single_hugepage_flag_show(kobj, attr, buf,
 295				TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 296}
 297static ssize_t use_zero_page_store(struct kobject *kobj,
 298		struct kobj_attribute *attr, const char *buf, size_t count)
 299{
 300	return single_hugepage_flag_store(kobj, attr, buf, count,
 301				 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
 302}
 303static struct kobj_attribute use_zero_page_attr =
 304	__ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
 305
 306static ssize_t hpage_pmd_size_show(struct kobject *kobj,
 307		struct kobj_attribute *attr, char *buf)
 308{
 309	return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
 310}
 311static struct kobj_attribute hpage_pmd_size_attr =
 312	__ATTR_RO(hpage_pmd_size);
 313
 314#ifdef CONFIG_DEBUG_VM
 315static ssize_t debug_cow_show(struct kobject *kobj,
 316				struct kobj_attribute *attr, char *buf)
 317{
 318	return single_hugepage_flag_show(kobj, attr, buf,
 319				TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 320}
 321static ssize_t debug_cow_store(struct kobject *kobj,
 322			       struct kobj_attribute *attr,
 323			       const char *buf, size_t count)
 324{
 325	return single_hugepage_flag_store(kobj, attr, buf, count,
 326				 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 327}
 328static struct kobj_attribute debug_cow_attr =
 329	__ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 330#endif /* CONFIG_DEBUG_VM */
 331
 332static struct attribute *hugepage_attr[] = {
 333	&enabled_attr.attr,
 334	&defrag_attr.attr,
 335	&use_zero_page_attr.attr,
 336	&hpage_pmd_size_attr.attr,
 337#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
 338	&shmem_enabled_attr.attr,
 339#endif
 340#ifdef CONFIG_DEBUG_VM
 341	&debug_cow_attr.attr,
 342#endif
 343	NULL,
 344};
 345
 346static const struct attribute_group hugepage_attr_group = {
 347	.attrs = hugepage_attr,
 348};
 349
 350static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 351{
 352	int err;
 353
 354	*hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 355	if (unlikely(!*hugepage_kobj)) {
 356		pr_err("failed to create transparent hugepage kobject\n");
 357		return -ENOMEM;
 358	}
 359
 360	err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 361	if (err) {
 362		pr_err("failed to register transparent hugepage group\n");
 363		goto delete_obj;
 364	}
 365
 366	err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 367	if (err) {
 368		pr_err("failed to register transparent hugepage group\n");
 369		goto remove_hp_group;
 370	}
 371
 372	return 0;
 373
 374remove_hp_group:
 375	sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 376delete_obj:
 377	kobject_put(*hugepage_kobj);
 378	return err;
 379}
 380
 381static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 382{
 383	sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 384	sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 385	kobject_put(hugepage_kobj);
 386}
 387#else
 388static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 389{
 390	return 0;
 391}
 392
 393static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 394{
 395}
 396#endif /* CONFIG_SYSFS */
 397
 398static int __init hugepage_init(void)
 399{
 400	int err;
 401	struct kobject *hugepage_kobj;
 402
 403	if (!has_transparent_hugepage()) {
 404		transparent_hugepage_flags = 0;
 405		return -EINVAL;
 406	}
 407
 408	/*
 409	 * hugepages can't be allocated by the buddy allocator
 410	 */
 411	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
 412	/*
 413	 * we use page->mapping and page->index in second tail page
 414	 * as list_head: assuming THP order >= 2
 415	 */
 416	MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
 417
 418	err = hugepage_init_sysfs(&hugepage_kobj);
 419	if (err)
 420		goto err_sysfs;
 421
 422	err = khugepaged_init();
 423	if (err)
 424		goto err_slab;
 425
 426	err = register_shrinker(&huge_zero_page_shrinker);
 427	if (err)
 428		goto err_hzp_shrinker;
 429	err = register_shrinker(&deferred_split_shrinker);
 430	if (err)
 431		goto err_split_shrinker;
 432
 433	/*
 434	 * By default disable transparent hugepages on smaller systems,
 435	 * where the extra memory used could hurt more than TLB overhead
 436	 * is likely to save.  The admin can still enable it through /sys.
 437	 */
 438	if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
 439		transparent_hugepage_flags = 0;
 440		return 0;
 441	}
 442
 443	err = start_stop_khugepaged();
 444	if (err)
 445		goto err_khugepaged;
 446
 447	return 0;
 448err_khugepaged:
 449	unregister_shrinker(&deferred_split_shrinker);
 450err_split_shrinker:
 451	unregister_shrinker(&huge_zero_page_shrinker);
 452err_hzp_shrinker:
 453	khugepaged_destroy();
 454err_slab:
 455	hugepage_exit_sysfs(hugepage_kobj);
 456err_sysfs:
 457	return err;
 458}
 459subsys_initcall(hugepage_init);
 460
 461static int __init setup_transparent_hugepage(char *str)
 462{
 463	int ret = 0;
 464	if (!str)
 465		goto out;
 466	if (!strcmp(str, "always")) {
 467		set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 468			&transparent_hugepage_flags);
 469		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 470			  &transparent_hugepage_flags);
 471		ret = 1;
 472	} else if (!strcmp(str, "madvise")) {
 473		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 474			  &transparent_hugepage_flags);
 475		set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 476			&transparent_hugepage_flags);
 477		ret = 1;
 478	} else if (!strcmp(str, "never")) {
 479		clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 480			  &transparent_hugepage_flags);
 481		clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 482			  &transparent_hugepage_flags);
 483		ret = 1;
 484	}
 485out:
 486	if (!ret)
 487		pr_warn("transparent_hugepage= cannot parse, ignored\n");
 488	return ret;
 489}
 490__setup("transparent_hugepage=", setup_transparent_hugepage);
 491
 492pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 493{
 494	if (likely(vma->vm_flags & VM_WRITE))
 495		pmd = pmd_mkwrite(pmd);
 496	return pmd;
 497}
 498
 499#ifdef CONFIG_MEMCG
 500static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 501{
 502	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
 503	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 504
 505	if (memcg)
 506		return &memcg->deferred_split_queue;
 507	else
 508		return &pgdat->deferred_split_queue;
 509}
 510#else
 511static inline struct deferred_split *get_deferred_split_queue(struct page *page)
 512{
 513	struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
 514
 515	return &pgdat->deferred_split_queue;
 516}
 517#endif
 518
 519void prep_transhuge_page(struct page *page)
 520{
 521	/*
 522	 * we use page->mapping and page->indexlru in second tail page
 523	 * as list_head: assuming THP order >= 2
 524	 */
 525
 526	INIT_LIST_HEAD(page_deferred_list(page));
 527	set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
 528}
 529
 530static unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
 531		loff_t off, unsigned long flags, unsigned long size)
 532{
 533	unsigned long addr;
 534	loff_t off_end = off + len;
 535	loff_t off_align = round_up(off, size);
 536	unsigned long len_pad;
 537
 538	if (off_end <= off_align || (off_end - off_align) < size)
 539		return 0;
 540
 541	len_pad = len + size;
 542	if (len_pad < len || (off + len_pad) < off)
 543		return 0;
 544
 545	addr = current->mm->get_unmapped_area(filp, 0, len_pad,
 546					      off >> PAGE_SHIFT, flags);
 547	if (IS_ERR_VALUE(addr))
 548		return 0;
 549
 550	addr += (off - addr) & (size - 1);
 551	return addr;
 552}
 553
 554unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 555		unsigned long len, unsigned long pgoff, unsigned long flags)
 556{
 557	loff_t off = (loff_t)pgoff << PAGE_SHIFT;
 558
 559	if (addr)
 560		goto out;
 561	if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
 562		goto out;
 563
 564	addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
 565	if (addr)
 566		return addr;
 567
 568 out:
 569	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 570}
 571EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
 572
 573static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
 574			struct page *page, gfp_t gfp)
 575{
 576	struct vm_area_struct *vma = vmf->vma;
 577	struct mem_cgroup *memcg;
 578	pgtable_t pgtable;
 579	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 580	vm_fault_t ret = 0;
 581
 582	VM_BUG_ON_PAGE(!PageCompound(page), page);
 583
 584	if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
 585		put_page(page);
 586		count_vm_event(THP_FAULT_FALLBACK);
 587		return VM_FAULT_FALLBACK;
 588	}
 589
 590	pgtable = pte_alloc_one(vma->vm_mm);
 591	if (unlikely(!pgtable)) {
 592		ret = VM_FAULT_OOM;
 593		goto release;
 594	}
 595
 596	clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
 597	/*
 598	 * The memory barrier inside __SetPageUptodate makes sure that
 599	 * clear_huge_page writes become visible before the set_pmd_at()
 600	 * write.
 601	 */
 602	__SetPageUptodate(page);
 603
 604	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 605	if (unlikely(!pmd_none(*vmf->pmd))) {
 606		goto unlock_release;
 607	} else {
 608		pmd_t entry;
 609
 610		ret = check_stable_address_space(vma->vm_mm);
 611		if (ret)
 612			goto unlock_release;
 613
 614		/* Deliver the page fault to userland */
 615		if (userfaultfd_missing(vma)) {
 616			vm_fault_t ret2;
 617
 618			spin_unlock(vmf->ptl);
 619			mem_cgroup_cancel_charge(page, memcg, true);
 620			put_page(page);
 621			pte_free(vma->vm_mm, pgtable);
 622			ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
 623			VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
 624			return ret2;
 625		}
 626
 627		entry = mk_huge_pmd(page, vma->vm_page_prot);
 628		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 629		page_add_new_anon_rmap(page, vma, haddr, true);
 630		mem_cgroup_commit_charge(page, memcg, false, true);
 631		lru_cache_add_active_or_unevictable(page, vma);
 632		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
 633		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
 634		add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 635		mm_inc_nr_ptes(vma->vm_mm);
 636		spin_unlock(vmf->ptl);
 637		count_vm_event(THP_FAULT_ALLOC);
 638		count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
 639	}
 640
 641	return 0;
 642unlock_release:
 643	spin_unlock(vmf->ptl);
 644release:
 645	if (pgtable)
 646		pte_free(vma->vm_mm, pgtable);
 647	mem_cgroup_cancel_charge(page, memcg, true);
 648	put_page(page);
 649	return ret;
 650
 651}
 652
 653/*
 654 * always: directly stall for all thp allocations
 655 * defer: wake kswapd and fail if not immediately available
 656 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
 657 *		  fail if not immediately available
 658 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
 659 *	    available
 660 * never: never stall for any thp allocation
 661 */
 662static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
 663{
 664	const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
 665
 666	/* Always do synchronous compaction */
 667	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
 668		return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
 669
 670	/* Kick kcompactd and fail quickly */
 671	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
 672		return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
 673
 674	/* Synchronous compaction if madvised, otherwise kick kcompactd */
 675	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
 676		return GFP_TRANSHUGE_LIGHT |
 677			(vma_madvised ? __GFP_DIRECT_RECLAIM :
 678					__GFP_KSWAPD_RECLAIM);
 679
 680	/* Only do synchronous compaction if madvised */
 681	if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
 682		return GFP_TRANSHUGE_LIGHT |
 683		       (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
 684
 685	return GFP_TRANSHUGE_LIGHT;
 686}
 687
 688/* Caller must hold page table lock. */
 689static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 690		struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
 691		struct page *zero_page)
 692{
 693	pmd_t entry;
 694	if (!pmd_none(*pmd))
 695		return false;
 696	entry = mk_pmd(zero_page, vma->vm_page_prot);
 697	entry = pmd_mkhuge(entry);
 698	if (pgtable)
 699		pgtable_trans_huge_deposit(mm, pmd, pgtable);
 700	set_pmd_at(mm, haddr, pmd, entry);
 701	mm_inc_nr_ptes(mm);
 702	return true;
 703}
 704
 705vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
 706{
 707	struct vm_area_struct *vma = vmf->vma;
 708	gfp_t gfp;
 709	struct page *page;
 710	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
 711
 712	if (!transhuge_vma_suitable(vma, haddr))
 713		return VM_FAULT_FALLBACK;
 714	if (unlikely(anon_vma_prepare(vma)))
 715		return VM_FAULT_OOM;
 716	if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
 717		return VM_FAULT_OOM;
 718	if (!(vmf->flags & FAULT_FLAG_WRITE) &&
 719			!mm_forbids_zeropage(vma->vm_mm) &&
 720			transparent_hugepage_use_zero_page()) {
 721		pgtable_t pgtable;
 722		struct page *zero_page;
 723		bool set;
 724		vm_fault_t ret;
 725		pgtable = pte_alloc_one(vma->vm_mm);
 726		if (unlikely(!pgtable))
 727			return VM_FAULT_OOM;
 728		zero_page = mm_get_huge_zero_page(vma->vm_mm);
 729		if (unlikely(!zero_page)) {
 730			pte_free(vma->vm_mm, pgtable);
 731			count_vm_event(THP_FAULT_FALLBACK);
 732			return VM_FAULT_FALLBACK;
 733		}
 734		vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
 735		ret = 0;
 736		set = false;
 737		if (pmd_none(*vmf->pmd)) {
 738			ret = check_stable_address_space(vma->vm_mm);
 739			if (ret) {
 740				spin_unlock(vmf->ptl);
 741			} else if (userfaultfd_missing(vma)) {
 742				spin_unlock(vmf->ptl);
 743				ret = handle_userfault(vmf, VM_UFFD_MISSING);
 744				VM_BUG_ON(ret & VM_FAULT_FALLBACK);
 745			} else {
 746				set_huge_zero_page(pgtable, vma->vm_mm, vma,
 747						   haddr, vmf->pmd, zero_page);
 748				spin_unlock(vmf->ptl);
 749				set = true;
 750			}
 751		} else
 752			spin_unlock(vmf->ptl);
 753		if (!set)
 754			pte_free(vma->vm_mm, pgtable);
 755		return ret;
 756	}
 757	gfp = alloc_hugepage_direct_gfpmask(vma);
 758	page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
 759	if (unlikely(!page)) {
 760		count_vm_event(THP_FAULT_FALLBACK);
 761		return VM_FAULT_FALLBACK;
 762	}
 763	prep_transhuge_page(page);
 764	return __do_huge_pmd_anonymous_page(vmf, page, gfp);
 765}
 766
 767static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
 768		pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
 769		pgtable_t pgtable)
 770{
 771	struct mm_struct *mm = vma->vm_mm;
 772	pmd_t entry;
 773	spinlock_t *ptl;
 774
 775	ptl = pmd_lock(mm, pmd);
 776	if (!pmd_none(*pmd)) {
 777		if (write) {
 778			if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
 779				WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
 780				goto out_unlock;
 781			}
 782			entry = pmd_mkyoung(*pmd);
 783			entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 784			if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
 785				update_mmu_cache_pmd(vma, addr, pmd);
 786		}
 787
 788		goto out_unlock;
 789	}
 790
 791	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
 792	if (pfn_t_devmap(pfn))
 793		entry = pmd_mkdevmap(entry);
 794	if (write) {
 795		entry = pmd_mkyoung(pmd_mkdirty(entry));
 796		entry = maybe_pmd_mkwrite(entry, vma);
 797	}
 798
 799	if (pgtable) {
 800		pgtable_trans_huge_deposit(mm, pmd, pgtable);
 801		mm_inc_nr_ptes(mm);
 802		pgtable = NULL;
 803	}
 804
 805	set_pmd_at(mm, addr, pmd, entry);
 806	update_mmu_cache_pmd(vma, addr, pmd);
 807
 808out_unlock:
 809	spin_unlock(ptl);
 810	if (pgtable)
 811		pte_free(mm, pgtable);
 812}
 813
 814vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
 
 815{
 816	unsigned long addr = vmf->address & PMD_MASK;
 817	struct vm_area_struct *vma = vmf->vma;
 818	pgprot_t pgprot = vma->vm_page_prot;
 819	pgtable_t pgtable = NULL;
 820
 821	/*
 822	 * If we had pmd_special, we could avoid all these restrictions,
 823	 * but we need to be consistent with PTEs and architectures that
 824	 * can't support a 'special' bit.
 825	 */
 826	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 827			!pfn_t_devmap(pfn));
 828	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 829						(VM_PFNMAP|VM_MIXEDMAP));
 830	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 
 831
 832	if (addr < vma->vm_start || addr >= vma->vm_end)
 833		return VM_FAULT_SIGBUS;
 834
 835	if (arch_needs_pgtable_deposit()) {
 836		pgtable = pte_alloc_one(vma->vm_mm);
 837		if (!pgtable)
 838			return VM_FAULT_OOM;
 839	}
 840
 841	track_pfn_insert(vma, &pgprot, pfn);
 842
 843	insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
 844	return VM_FAULT_NOPAGE;
 845}
 846EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
 847
 848#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 849static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
 850{
 851	if (likely(vma->vm_flags & VM_WRITE))
 852		pud = pud_mkwrite(pud);
 853	return pud;
 854}
 855
 856static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
 857		pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
 858{
 859	struct mm_struct *mm = vma->vm_mm;
 860	pud_t entry;
 861	spinlock_t *ptl;
 862
 863	ptl = pud_lock(mm, pud);
 864	if (!pud_none(*pud)) {
 865		if (write) {
 866			if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
 867				WARN_ON_ONCE(!is_huge_zero_pud(*pud));
 868				goto out_unlock;
 869			}
 870			entry = pud_mkyoung(*pud);
 871			entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
 872			if (pudp_set_access_flags(vma, addr, pud, entry, 1))
 873				update_mmu_cache_pud(vma, addr, pud);
 874		}
 875		goto out_unlock;
 876	}
 877
 878	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
 879	if (pfn_t_devmap(pfn))
 880		entry = pud_mkdevmap(entry);
 881	if (write) {
 882		entry = pud_mkyoung(pud_mkdirty(entry));
 883		entry = maybe_pud_mkwrite(entry, vma);
 884	}
 885	set_pud_at(mm, addr, pud, entry);
 886	update_mmu_cache_pud(vma, addr, pud);
 887
 888out_unlock:
 889	spin_unlock(ptl);
 890}
 891
 892vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
 
 893{
 894	unsigned long addr = vmf->address & PUD_MASK;
 895	struct vm_area_struct *vma = vmf->vma;
 896	pgprot_t pgprot = vma->vm_page_prot;
 897
 898	/*
 899	 * If we had pud_special, we could avoid all these restrictions,
 900	 * but we need to be consistent with PTEs and architectures that
 901	 * can't support a 'special' bit.
 902	 */
 903	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
 904			!pfn_t_devmap(pfn));
 905	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
 906						(VM_PFNMAP|VM_MIXEDMAP));
 907	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
 
 908
 909	if (addr < vma->vm_start || addr >= vma->vm_end)
 910		return VM_FAULT_SIGBUS;
 911
 912	track_pfn_insert(vma, &pgprot, pfn);
 913
 914	insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
 915	return VM_FAULT_NOPAGE;
 916}
 917EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 918#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 919
 920static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
 921		pmd_t *pmd, int flags)
 922{
 923	pmd_t _pmd;
 924
 925	_pmd = pmd_mkyoung(*pmd);
 926	if (flags & FOLL_WRITE)
 927		_pmd = pmd_mkdirty(_pmd);
 928	if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
 929				pmd, _pmd, flags & FOLL_WRITE))
 930		update_mmu_cache_pmd(vma, addr, pmd);
 931}
 932
 933struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 934		pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 935{
 936	unsigned long pfn = pmd_pfn(*pmd);
 937	struct mm_struct *mm = vma->vm_mm;
 
 938	struct page *page;
 939
 940	assert_spin_locked(pmd_lockptr(mm, pmd));
 941
 942	/*
 943	 * When we COW a devmap PMD entry, we split it into PTEs, so we should
 944	 * not be in this function with `flags & FOLL_COW` set.
 945	 */
 946	WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
 947
 948	if (flags & FOLL_WRITE && !pmd_write(*pmd))
 949		return NULL;
 950
 951	if (pmd_present(*pmd) && pmd_devmap(*pmd))
 952		/* pass */;
 953	else
 954		return NULL;
 955
 956	if (flags & FOLL_TOUCH)
 957		touch_pmd(vma, addr, pmd, flags);
 958
 959	/*
 960	 * device mapped pages can only be returned if the
 961	 * caller will manage the page reference count.
 962	 */
 963	if (!(flags & FOLL_GET))
 964		return ERR_PTR(-EEXIST);
 965
 966	pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
 967	*pgmap = get_dev_pagemap(pfn, *pgmap);
 968	if (!*pgmap)
 969		return ERR_PTR(-EFAULT);
 970	page = pfn_to_page(pfn);
 971	get_page(page);
 
 972
 973	return page;
 974}
 975
 976int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 977		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 978		  struct vm_area_struct *vma)
 979{
 980	spinlock_t *dst_ptl, *src_ptl;
 981	struct page *src_page;
 982	pmd_t pmd;
 983	pgtable_t pgtable = NULL;
 984	int ret = -ENOMEM;
 985
 986	/* Skip if can be re-fill on fault */
 987	if (!vma_is_anonymous(vma))
 988		return 0;
 989
 990	pgtable = pte_alloc_one(dst_mm);
 991	if (unlikely(!pgtable))
 992		goto out;
 993
 994	dst_ptl = pmd_lock(dst_mm, dst_pmd);
 995	src_ptl = pmd_lockptr(src_mm, src_pmd);
 996	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 997
 998	ret = -EAGAIN;
 999	pmd = *src_pmd;
1000
1001#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1002	if (unlikely(is_swap_pmd(pmd))) {
1003		swp_entry_t entry = pmd_to_swp_entry(pmd);
1004
1005		VM_BUG_ON(!is_pmd_migration_entry(pmd));
1006		if (is_write_migration_entry(entry)) {
1007			make_migration_entry_read(&entry);
1008			pmd = swp_entry_to_pmd(entry);
1009			if (pmd_swp_soft_dirty(*src_pmd))
1010				pmd = pmd_swp_mksoft_dirty(pmd);
1011			set_pmd_at(src_mm, addr, src_pmd, pmd);
1012		}
1013		add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1014		mm_inc_nr_ptes(dst_mm);
1015		pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1016		set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1017		ret = 0;
1018		goto out_unlock;
1019	}
1020#endif
1021
1022	if (unlikely(!pmd_trans_huge(pmd))) {
1023		pte_free(dst_mm, pgtable);
1024		goto out_unlock;
1025	}
1026	/*
1027	 * When page table lock is held, the huge zero pmd should not be
1028	 * under splitting since we don't split the page itself, only pmd to
1029	 * a page table.
1030	 */
1031	if (is_huge_zero_pmd(pmd)) {
1032		struct page *zero_page;
1033		/*
1034		 * get_huge_zero_page() will never allocate a new page here,
1035		 * since we already have a zero page to copy. It just takes a
1036		 * reference.
1037		 */
1038		zero_page = mm_get_huge_zero_page(dst_mm);
1039		set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1040				zero_page);
1041		ret = 0;
1042		goto out_unlock;
1043	}
1044
1045	src_page = pmd_page(pmd);
1046	VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1047	get_page(src_page);
1048	page_dup_rmap(src_page, true);
1049	add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1050	mm_inc_nr_ptes(dst_mm);
1051	pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1052
1053	pmdp_set_wrprotect(src_mm, addr, src_pmd);
1054	pmd = pmd_mkold(pmd_wrprotect(pmd));
1055	set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1056
1057	ret = 0;
1058out_unlock:
1059	spin_unlock(src_ptl);
1060	spin_unlock(dst_ptl);
1061out:
1062	return ret;
1063}
1064
1065#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1066static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1067		pud_t *pud, int flags)
1068{
1069	pud_t _pud;
1070
1071	_pud = pud_mkyoung(*pud);
1072	if (flags & FOLL_WRITE)
1073		_pud = pud_mkdirty(_pud);
1074	if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1075				pud, _pud, flags & FOLL_WRITE))
1076		update_mmu_cache_pud(vma, addr, pud);
1077}
1078
1079struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1080		pud_t *pud, int flags, struct dev_pagemap **pgmap)
1081{
1082	unsigned long pfn = pud_pfn(*pud);
1083	struct mm_struct *mm = vma->vm_mm;
 
1084	struct page *page;
1085
1086	assert_spin_locked(pud_lockptr(mm, pud));
1087
1088	if (flags & FOLL_WRITE && !pud_write(*pud))
1089		return NULL;
1090
1091	if (pud_present(*pud) && pud_devmap(*pud))
1092		/* pass */;
1093	else
1094		return NULL;
1095
1096	if (flags & FOLL_TOUCH)
1097		touch_pud(vma, addr, pud, flags);
1098
1099	/*
1100	 * device mapped pages can only be returned if the
1101	 * caller will manage the page reference count.
1102	 */
1103	if (!(flags & FOLL_GET))
1104		return ERR_PTR(-EEXIST);
1105
1106	pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1107	*pgmap = get_dev_pagemap(pfn, *pgmap);
1108	if (!*pgmap)
1109		return ERR_PTR(-EFAULT);
1110	page = pfn_to_page(pfn);
1111	get_page(page);
 
1112
1113	return page;
1114}
1115
1116int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1117		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1118		  struct vm_area_struct *vma)
1119{
1120	spinlock_t *dst_ptl, *src_ptl;
1121	pud_t pud;
1122	int ret;
1123
1124	dst_ptl = pud_lock(dst_mm, dst_pud);
1125	src_ptl = pud_lockptr(src_mm, src_pud);
1126	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1127
1128	ret = -EAGAIN;
1129	pud = *src_pud;
1130	if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1131		goto out_unlock;
1132
1133	/*
1134	 * When page table lock is held, the huge zero pud should not be
1135	 * under splitting since we don't split the page itself, only pud to
1136	 * a page table.
1137	 */
1138	if (is_huge_zero_pud(pud)) {
1139		/* No huge zero pud yet */
1140	}
1141
1142	pudp_set_wrprotect(src_mm, addr, src_pud);
1143	pud = pud_mkold(pud_wrprotect(pud));
1144	set_pud_at(dst_mm, addr, dst_pud, pud);
1145
1146	ret = 0;
1147out_unlock:
1148	spin_unlock(src_ptl);
1149	spin_unlock(dst_ptl);
1150	return ret;
1151}
1152
1153void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1154{
1155	pud_t entry;
1156	unsigned long haddr;
1157	bool write = vmf->flags & FAULT_FLAG_WRITE;
1158
1159	vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1160	if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1161		goto unlock;
1162
1163	entry = pud_mkyoung(orig_pud);
1164	if (write)
1165		entry = pud_mkdirty(entry);
1166	haddr = vmf->address & HPAGE_PUD_MASK;
1167	if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1168		update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1169
1170unlock:
1171	spin_unlock(vmf->ptl);
1172}
1173#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1174
1175void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1176{
1177	pmd_t entry;
1178	unsigned long haddr;
1179	bool write = vmf->flags & FAULT_FLAG_WRITE;
1180
1181	vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1182	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1183		goto unlock;
1184
1185	entry = pmd_mkyoung(orig_pmd);
1186	if (write)
1187		entry = pmd_mkdirty(entry);
1188	haddr = vmf->address & HPAGE_PMD_MASK;
1189	if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1190		update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1191
1192unlock:
1193	spin_unlock(vmf->ptl);
1194}
1195
1196static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1197			pmd_t orig_pmd, struct page *page)
1198{
1199	struct vm_area_struct *vma = vmf->vma;
1200	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1201	struct mem_cgroup *memcg;
1202	pgtable_t pgtable;
1203	pmd_t _pmd;
1204	int i;
1205	vm_fault_t ret = 0;
1206	struct page **pages;
1207	struct mmu_notifier_range range;
 
1208
1209	pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1210			      GFP_KERNEL);
1211	if (unlikely(!pages)) {
1212		ret |= VM_FAULT_OOM;
1213		goto out;
1214	}
1215
1216	for (i = 0; i < HPAGE_PMD_NR; i++) {
1217		pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1218					       vmf->address, page_to_nid(page));
1219		if (unlikely(!pages[i] ||
1220			     mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1221				     GFP_KERNEL, &memcg, false))) {
1222			if (pages[i])
1223				put_page(pages[i]);
1224			while (--i >= 0) {
1225				memcg = (void *)page_private(pages[i]);
1226				set_page_private(pages[i], 0);
1227				mem_cgroup_cancel_charge(pages[i], memcg,
1228						false);
1229				put_page(pages[i]);
1230			}
1231			kfree(pages);
1232			ret |= VM_FAULT_OOM;
1233			goto out;
1234		}
1235		set_page_private(pages[i], (unsigned long)memcg);
1236	}
1237
1238	for (i = 0; i < HPAGE_PMD_NR; i++) {
1239		copy_user_highpage(pages[i], page + i,
1240				   haddr + PAGE_SIZE * i, vma);
1241		__SetPageUptodate(pages[i]);
1242		cond_resched();
1243	}
1244
1245	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1246				haddr, haddr + HPAGE_PMD_SIZE);
1247	mmu_notifier_invalidate_range_start(&range);
1248
1249	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1250	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1251		goto out_free_pages;
1252	VM_BUG_ON_PAGE(!PageHead(page), page);
1253
1254	/*
1255	 * Leave pmd empty until pte is filled note we must notify here as
1256	 * concurrent CPU thread might write to new page before the call to
1257	 * mmu_notifier_invalidate_range_end() happens which can lead to a
1258	 * device seeing memory write in different order than CPU.
1259	 *
1260	 * See Documentation/vm/mmu_notifier.rst
1261	 */
1262	pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1263
1264	pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1265	pmd_populate(vma->vm_mm, &_pmd, pgtable);
1266
1267	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1268		pte_t entry;
1269		entry = mk_pte(pages[i], vma->vm_page_prot);
1270		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1271		memcg = (void *)page_private(pages[i]);
1272		set_page_private(pages[i], 0);
1273		page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1274		mem_cgroup_commit_charge(pages[i], memcg, false, false);
1275		lru_cache_add_active_or_unevictable(pages[i], vma);
1276		vmf->pte = pte_offset_map(&_pmd, haddr);
1277		VM_BUG_ON(!pte_none(*vmf->pte));
1278		set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1279		pte_unmap(vmf->pte);
1280	}
1281	kfree(pages);
1282
1283	smp_wmb(); /* make pte visible before pmd */
1284	pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1285	page_remove_rmap(page, true);
1286	spin_unlock(vmf->ptl);
1287
1288	/*
1289	 * No need to double call mmu_notifier->invalidate_range() callback as
1290	 * the above pmdp_huge_clear_flush_notify() did already call it.
1291	 */
1292	mmu_notifier_invalidate_range_only_end(&range);
 
1293
1294	ret |= VM_FAULT_WRITE;
1295	put_page(page);
1296
1297out:
1298	return ret;
1299
1300out_free_pages:
1301	spin_unlock(vmf->ptl);
1302	mmu_notifier_invalidate_range_end(&range);
1303	for (i = 0; i < HPAGE_PMD_NR; i++) {
1304		memcg = (void *)page_private(pages[i]);
1305		set_page_private(pages[i], 0);
1306		mem_cgroup_cancel_charge(pages[i], memcg, false);
1307		put_page(pages[i]);
1308	}
1309	kfree(pages);
1310	goto out;
1311}
1312
1313vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1314{
1315	struct vm_area_struct *vma = vmf->vma;
1316	struct page *page = NULL, *new_page;
1317	struct mem_cgroup *memcg;
1318	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1319	struct mmu_notifier_range range;
 
1320	gfp_t huge_gfp;			/* for allocation and charge */
1321	vm_fault_t ret = 0;
1322
1323	vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1324	VM_BUG_ON_VMA(!vma->anon_vma, vma);
1325	if (is_huge_zero_pmd(orig_pmd))
1326		goto alloc;
1327	spin_lock(vmf->ptl);
1328	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1329		goto out_unlock;
1330
1331	page = pmd_page(orig_pmd);
1332	VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1333	/*
1334	 * We can only reuse the page if nobody else maps the huge page or it's
1335	 * part.
1336	 */
1337	if (!trylock_page(page)) {
1338		get_page(page);
1339		spin_unlock(vmf->ptl);
1340		lock_page(page);
1341		spin_lock(vmf->ptl);
1342		if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1343			unlock_page(page);
1344			put_page(page);
1345			goto out_unlock;
1346		}
1347		put_page(page);
1348	}
1349	if (reuse_swap_page(page, NULL)) {
1350		pmd_t entry;
1351		entry = pmd_mkyoung(orig_pmd);
1352		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1353		if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry,  1))
1354			update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1355		ret |= VM_FAULT_WRITE;
1356		unlock_page(page);
1357		goto out_unlock;
1358	}
1359	unlock_page(page);
1360	get_page(page);
1361	spin_unlock(vmf->ptl);
1362alloc:
1363	if (__transparent_hugepage_enabled(vma) &&
1364	    !transparent_hugepage_debug_cow()) {
1365		huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1366		new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1367	} else
1368		new_page = NULL;
1369
1370	if (likely(new_page)) {
1371		prep_transhuge_page(new_page);
1372	} else {
1373		if (!page) {
1374			split_huge_pmd(vma, vmf->pmd, vmf->address);
1375			ret |= VM_FAULT_FALLBACK;
1376		} else {
1377			ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1378			if (ret & VM_FAULT_OOM) {
1379				split_huge_pmd(vma, vmf->pmd, vmf->address);
1380				ret |= VM_FAULT_FALLBACK;
1381			}
1382			put_page(page);
1383		}
1384		count_vm_event(THP_FAULT_FALLBACK);
1385		goto out;
1386	}
1387
1388	if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1389					huge_gfp, &memcg, true))) {
1390		put_page(new_page);
1391		split_huge_pmd(vma, vmf->pmd, vmf->address);
1392		if (page)
1393			put_page(page);
1394		ret |= VM_FAULT_FALLBACK;
1395		count_vm_event(THP_FAULT_FALLBACK);
1396		goto out;
1397	}
1398
1399	count_vm_event(THP_FAULT_ALLOC);
1400	count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1401
1402	if (!page)
1403		clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1404	else
1405		copy_user_huge_page(new_page, page, vmf->address,
1406				    vma, HPAGE_PMD_NR);
1407	__SetPageUptodate(new_page);
1408
1409	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1410				haddr, haddr + HPAGE_PMD_SIZE);
1411	mmu_notifier_invalidate_range_start(&range);
1412
1413	spin_lock(vmf->ptl);
1414	if (page)
1415		put_page(page);
1416	if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1417		spin_unlock(vmf->ptl);
1418		mem_cgroup_cancel_charge(new_page, memcg, true);
1419		put_page(new_page);
1420		goto out_mn;
1421	} else {
1422		pmd_t entry;
1423		entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1424		entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1425		pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1426		page_add_new_anon_rmap(new_page, vma, haddr, true);
1427		mem_cgroup_commit_charge(new_page, memcg, false, true);
1428		lru_cache_add_active_or_unevictable(new_page, vma);
1429		set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1430		update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1431		if (!page) {
1432			add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1433		} else {
1434			VM_BUG_ON_PAGE(!PageHead(page), page);
1435			page_remove_rmap(page, true);
1436			put_page(page);
1437		}
1438		ret |= VM_FAULT_WRITE;
1439	}
1440	spin_unlock(vmf->ptl);
1441out_mn:
1442	/*
1443	 * No need to double call mmu_notifier->invalidate_range() callback as
1444	 * the above pmdp_huge_clear_flush_notify() did already call it.
1445	 */
1446	mmu_notifier_invalidate_range_only_end(&range);
 
1447out:
1448	return ret;
1449out_unlock:
1450	spin_unlock(vmf->ptl);
1451	return ret;
1452}
1453
1454/*
1455 * FOLL_FORCE can write to even unwritable pmd's, but only
1456 * after we've gone through a COW cycle and they are dirty.
1457 */
1458static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1459{
1460	return pmd_write(pmd) ||
1461	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1462}
1463
1464struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1465				   unsigned long addr,
1466				   pmd_t *pmd,
1467				   unsigned int flags)
1468{
1469	struct mm_struct *mm = vma->vm_mm;
1470	struct page *page = NULL;
1471
1472	assert_spin_locked(pmd_lockptr(mm, pmd));
1473
1474	if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1475		goto out;
1476
1477	/* Avoid dumping huge zero page */
1478	if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1479		return ERR_PTR(-EFAULT);
1480
1481	/* Full NUMA hinting faults to serialise migration in fault paths */
1482	if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1483		goto out;
1484
1485	page = pmd_page(*pmd);
1486	VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1487	if (flags & FOLL_TOUCH)
1488		touch_pmd(vma, addr, pmd, flags);
1489	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1490		/*
1491		 * We don't mlock() pte-mapped THPs. This way we can avoid
1492		 * leaking mlocked pages into non-VM_LOCKED VMAs.
1493		 *
1494		 * For anon THP:
1495		 *
1496		 * In most cases the pmd is the only mapping of the page as we
1497		 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1498		 * writable private mappings in populate_vma_page_range().
1499		 *
1500		 * The only scenario when we have the page shared here is if we
1501		 * mlocking read-only mapping shared over fork(). We skip
1502		 * mlocking such pages.
1503		 *
1504		 * For file THP:
1505		 *
1506		 * We can expect PageDoubleMap() to be stable under page lock:
1507		 * for file pages we set it in page_add_file_rmap(), which
1508		 * requires page to be locked.
1509		 */
1510
1511		if (PageAnon(page) && compound_mapcount(page) != 1)
1512			goto skip_mlock;
1513		if (PageDoubleMap(page) || !page->mapping)
1514			goto skip_mlock;
1515		if (!trylock_page(page))
1516			goto skip_mlock;
1517		lru_add_drain();
1518		if (page->mapping && !PageDoubleMap(page))
1519			mlock_vma_page(page);
1520		unlock_page(page);
1521	}
1522skip_mlock:
1523	page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1524	VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1525	if (flags & FOLL_GET)
1526		get_page(page);
1527
1528out:
1529	return page;
1530}
1531
1532/* NUMA hinting page fault entry point for trans huge pmds */
1533vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1534{
1535	struct vm_area_struct *vma = vmf->vma;
1536	struct anon_vma *anon_vma = NULL;
1537	struct page *page;
1538	unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1539	int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1540	int target_nid, last_cpupid = -1;
1541	bool page_locked;
1542	bool migrated = false;
1543	bool was_writable;
1544	int flags = 0;
1545
1546	vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1547	if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1548		goto out_unlock;
1549
1550	/*
1551	 * If there are potential migrations, wait for completion and retry
1552	 * without disrupting NUMA hinting information. Do not relock and
1553	 * check_same as the page may no longer be mapped.
1554	 */
1555	if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1556		page = pmd_page(*vmf->pmd);
1557		if (!get_page_unless_zero(page))
1558			goto out_unlock;
1559		spin_unlock(vmf->ptl);
1560		put_and_wait_on_page_locked(page);
 
1561		goto out;
1562	}
1563
1564	page = pmd_page(pmd);
1565	BUG_ON(is_huge_zero_page(page));
1566	page_nid = page_to_nid(page);
1567	last_cpupid = page_cpupid_last(page);
1568	count_vm_numa_event(NUMA_HINT_FAULTS);
1569	if (page_nid == this_nid) {
1570		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1571		flags |= TNF_FAULT_LOCAL;
1572	}
1573
1574	/* See similar comment in do_numa_page for explanation */
1575	if (!pmd_savedwrite(pmd))
1576		flags |= TNF_NO_GROUP;
1577
1578	/*
1579	 * Acquire the page lock to serialise THP migrations but avoid dropping
1580	 * page_table_lock if at all possible
1581	 */
1582	page_locked = trylock_page(page);
1583	target_nid = mpol_misplaced(page, vma, haddr);
1584	if (target_nid == NUMA_NO_NODE) {
1585		/* If the page was locked, there are no parallel migrations */
1586		if (page_locked)
1587			goto clear_pmdnuma;
1588	}
1589
1590	/* Migration could have started since the pmd_trans_migrating check */
1591	if (!page_locked) {
1592		page_nid = NUMA_NO_NODE;
1593		if (!get_page_unless_zero(page))
1594			goto out_unlock;
1595		spin_unlock(vmf->ptl);
1596		put_and_wait_on_page_locked(page);
 
1597		goto out;
1598	}
1599
1600	/*
1601	 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1602	 * to serialises splits
1603	 */
1604	get_page(page);
1605	spin_unlock(vmf->ptl);
1606	anon_vma = page_lock_anon_vma_read(page);
1607
1608	/* Confirm the PMD did not change while page_table_lock was released */
1609	spin_lock(vmf->ptl);
1610	if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1611		unlock_page(page);
1612		put_page(page);
1613		page_nid = NUMA_NO_NODE;
1614		goto out_unlock;
1615	}
1616
1617	/* Bail if we fail to protect against THP splits for any reason */
1618	if (unlikely(!anon_vma)) {
1619		put_page(page);
1620		page_nid = NUMA_NO_NODE;
1621		goto clear_pmdnuma;
1622	}
1623
1624	/*
1625	 * Since we took the NUMA fault, we must have observed the !accessible
1626	 * bit. Make sure all other CPUs agree with that, to avoid them
1627	 * modifying the page we're about to migrate.
1628	 *
1629	 * Must be done under PTL such that we'll observe the relevant
1630	 * inc_tlb_flush_pending().
1631	 *
1632	 * We are not sure a pending tlb flush here is for a huge page
1633	 * mapping or not. Hence use the tlb range variant
1634	 */
1635	if (mm_tlb_flush_pending(vma->vm_mm)) {
1636		flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1637		/*
1638		 * change_huge_pmd() released the pmd lock before
1639		 * invalidating the secondary MMUs sharing the primary
1640		 * MMU pagetables (with ->invalidate_range()). The
1641		 * mmu_notifier_invalidate_range_end() (which
1642		 * internally calls ->invalidate_range()) in
1643		 * change_pmd_range() will run after us, so we can't
1644		 * rely on it here and we need an explicit invalidate.
1645		 */
1646		mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1647					      haddr + HPAGE_PMD_SIZE);
1648	}
1649
1650	/*
1651	 * Migrate the THP to the requested node, returns with page unlocked
1652	 * and access rights restored.
1653	 */
1654	spin_unlock(vmf->ptl);
1655
1656	migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1657				vmf->pmd, pmd, vmf->address, page, target_nid);
1658	if (migrated) {
1659		flags |= TNF_MIGRATED;
1660		page_nid = target_nid;
1661	} else
1662		flags |= TNF_MIGRATE_FAIL;
1663
1664	goto out;
1665clear_pmdnuma:
1666	BUG_ON(!PageLocked(page));
1667	was_writable = pmd_savedwrite(pmd);
1668	pmd = pmd_modify(pmd, vma->vm_page_prot);
1669	pmd = pmd_mkyoung(pmd);
1670	if (was_writable)
1671		pmd = pmd_mkwrite(pmd);
1672	set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1673	update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1674	unlock_page(page);
1675out_unlock:
1676	spin_unlock(vmf->ptl);
1677
1678out:
1679	if (anon_vma)
1680		page_unlock_anon_vma_read(anon_vma);
1681
1682	if (page_nid != NUMA_NO_NODE)
1683		task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1684				flags);
1685
1686	return 0;
1687}
1688
1689/*
1690 * Return true if we do MADV_FREE successfully on entire pmd page.
1691 * Otherwise, return false.
1692 */
1693bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1694		pmd_t *pmd, unsigned long addr, unsigned long next)
1695{
1696	spinlock_t *ptl;
1697	pmd_t orig_pmd;
1698	struct page *page;
1699	struct mm_struct *mm = tlb->mm;
1700	bool ret = false;
1701
1702	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1703
1704	ptl = pmd_trans_huge_lock(pmd, vma);
1705	if (!ptl)
1706		goto out_unlocked;
1707
1708	orig_pmd = *pmd;
1709	if (is_huge_zero_pmd(orig_pmd))
1710		goto out;
1711
1712	if (unlikely(!pmd_present(orig_pmd))) {
1713		VM_BUG_ON(thp_migration_supported() &&
1714				  !is_pmd_migration_entry(orig_pmd));
1715		goto out;
1716	}
1717
1718	page = pmd_page(orig_pmd);
1719	/*
1720	 * If other processes are mapping this page, we couldn't discard
1721	 * the page unless they all do MADV_FREE so let's skip the page.
1722	 */
1723	if (page_mapcount(page) != 1)
1724		goto out;
1725
1726	if (!trylock_page(page))
1727		goto out;
1728
1729	/*
1730	 * If user want to discard part-pages of THP, split it so MADV_FREE
1731	 * will deactivate only them.
1732	 */
1733	if (next - addr != HPAGE_PMD_SIZE) {
1734		get_page(page);
1735		spin_unlock(ptl);
1736		split_huge_page(page);
1737		unlock_page(page);
1738		put_page(page);
1739		goto out_unlocked;
1740	}
1741
1742	if (PageDirty(page))
1743		ClearPageDirty(page);
1744	unlock_page(page);
1745
1746	if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1747		pmdp_invalidate(vma, addr, pmd);
1748		orig_pmd = pmd_mkold(orig_pmd);
1749		orig_pmd = pmd_mkclean(orig_pmd);
1750
1751		set_pmd_at(mm, addr, pmd, orig_pmd);
1752		tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1753	}
1754
1755	mark_page_lazyfree(page);
1756	ret = true;
1757out:
1758	spin_unlock(ptl);
1759out_unlocked:
1760	return ret;
1761}
1762
1763static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1764{
1765	pgtable_t pgtable;
1766
1767	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1768	pte_free(mm, pgtable);
1769	mm_dec_nr_ptes(mm);
1770}
1771
1772int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1773		 pmd_t *pmd, unsigned long addr)
1774{
1775	pmd_t orig_pmd;
1776	spinlock_t *ptl;
1777
1778	tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1779
1780	ptl = __pmd_trans_huge_lock(pmd, vma);
1781	if (!ptl)
1782		return 0;
1783	/*
1784	 * For architectures like ppc64 we look at deposited pgtable
1785	 * when calling pmdp_huge_get_and_clear. So do the
1786	 * pgtable_trans_huge_withdraw after finishing pmdp related
1787	 * operations.
1788	 */
1789	orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1790			tlb->fullmm);
1791	tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1792	if (vma_is_dax(vma)) {
1793		if (arch_needs_pgtable_deposit())
1794			zap_deposited_table(tlb->mm, pmd);
1795		spin_unlock(ptl);
1796		if (is_huge_zero_pmd(orig_pmd))
1797			tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1798	} else if (is_huge_zero_pmd(orig_pmd)) {
1799		zap_deposited_table(tlb->mm, pmd);
1800		spin_unlock(ptl);
1801		tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1802	} else {
1803		struct page *page = NULL;
1804		int flush_needed = 1;
1805
1806		if (pmd_present(orig_pmd)) {
1807			page = pmd_page(orig_pmd);
1808			page_remove_rmap(page, true);
1809			VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1810			VM_BUG_ON_PAGE(!PageHead(page), page);
1811		} else if (thp_migration_supported()) {
1812			swp_entry_t entry;
1813
1814			VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1815			entry = pmd_to_swp_entry(orig_pmd);
1816			page = pfn_to_page(swp_offset(entry));
1817			flush_needed = 0;
1818		} else
1819			WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1820
1821		if (PageAnon(page)) {
1822			zap_deposited_table(tlb->mm, pmd);
1823			add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1824		} else {
1825			if (arch_needs_pgtable_deposit())
1826				zap_deposited_table(tlb->mm, pmd);
1827			add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1828		}
1829
1830		spin_unlock(ptl);
1831		if (flush_needed)
1832			tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1833	}
1834	return 1;
1835}
1836
1837#ifndef pmd_move_must_withdraw
1838static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1839					 spinlock_t *old_pmd_ptl,
1840					 struct vm_area_struct *vma)
1841{
1842	/*
1843	 * With split pmd lock we also need to move preallocated
1844	 * PTE page table if new_pmd is on different PMD page table.
1845	 *
1846	 * We also don't deposit and withdraw tables for file pages.
1847	 */
1848	return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1849}
1850#endif
1851
1852static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1853{
1854#ifdef CONFIG_MEM_SOFT_DIRTY
1855	if (unlikely(is_pmd_migration_entry(pmd)))
1856		pmd = pmd_swp_mksoft_dirty(pmd);
1857	else if (pmd_present(pmd))
1858		pmd = pmd_mksoft_dirty(pmd);
1859#endif
1860	return pmd;
1861}
1862
1863bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1864		  unsigned long new_addr, unsigned long old_end,
1865		  pmd_t *old_pmd, pmd_t *new_pmd)
1866{
1867	spinlock_t *old_ptl, *new_ptl;
1868	pmd_t pmd;
1869	struct mm_struct *mm = vma->vm_mm;
1870	bool force_flush = false;
1871
1872	if ((old_addr & ~HPAGE_PMD_MASK) ||
1873	    (new_addr & ~HPAGE_PMD_MASK) ||
1874	    old_end - old_addr < HPAGE_PMD_SIZE)
1875		return false;
1876
1877	/*
1878	 * The destination pmd shouldn't be established, free_pgtables()
1879	 * should have release it.
1880	 */
1881	if (WARN_ON(!pmd_none(*new_pmd))) {
1882		VM_BUG_ON(pmd_trans_huge(*new_pmd));
1883		return false;
1884	}
1885
1886	/*
1887	 * We don't have to worry about the ordering of src and dst
1888	 * ptlocks because exclusive mmap_sem prevents deadlock.
1889	 */
1890	old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1891	if (old_ptl) {
1892		new_ptl = pmd_lockptr(mm, new_pmd);
1893		if (new_ptl != old_ptl)
1894			spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1895		pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1896		if (pmd_present(pmd))
1897			force_flush = true;
1898		VM_BUG_ON(!pmd_none(*new_pmd));
1899
1900		if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1901			pgtable_t pgtable;
1902			pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1903			pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1904		}
1905		pmd = move_soft_dirty_pmd(pmd);
1906		set_pmd_at(mm, new_addr, new_pmd, pmd);
 
 
1907		if (force_flush)
1908			flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1909		if (new_ptl != old_ptl)
1910			spin_unlock(new_ptl);
1911		spin_unlock(old_ptl);
1912		return true;
1913	}
1914	return false;
1915}
1916
1917/*
1918 * Returns
1919 *  - 0 if PMD could not be locked
1920 *  - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1921 *  - HPAGE_PMD_NR is protections changed and TLB flush necessary
1922 */
1923int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1924		unsigned long addr, pgprot_t newprot, int prot_numa)
1925{
1926	struct mm_struct *mm = vma->vm_mm;
1927	spinlock_t *ptl;
1928	pmd_t entry;
1929	bool preserve_write;
1930	int ret;
1931
1932	ptl = __pmd_trans_huge_lock(pmd, vma);
1933	if (!ptl)
1934		return 0;
1935
1936	preserve_write = prot_numa && pmd_write(*pmd);
1937	ret = 1;
1938
1939#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1940	if (is_swap_pmd(*pmd)) {
1941		swp_entry_t entry = pmd_to_swp_entry(*pmd);
1942
1943		VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1944		if (is_write_migration_entry(entry)) {
1945			pmd_t newpmd;
1946			/*
1947			 * A protection check is difficult so
1948			 * just be safe and disable write
1949			 */
1950			make_migration_entry_read(&entry);
1951			newpmd = swp_entry_to_pmd(entry);
1952			if (pmd_swp_soft_dirty(*pmd))
1953				newpmd = pmd_swp_mksoft_dirty(newpmd);
1954			set_pmd_at(mm, addr, pmd, newpmd);
1955		}
1956		goto unlock;
1957	}
1958#endif
1959
1960	/*
1961	 * Avoid trapping faults against the zero page. The read-only
1962	 * data is likely to be read-cached on the local CPU and
1963	 * local/remote hits to the zero page are not interesting.
1964	 */
1965	if (prot_numa && is_huge_zero_pmd(*pmd))
1966		goto unlock;
1967
1968	if (prot_numa && pmd_protnone(*pmd))
1969		goto unlock;
1970
1971	/*
1972	 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1973	 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1974	 * which is also under down_read(mmap_sem):
1975	 *
1976	 *	CPU0:				CPU1:
1977	 *				change_huge_pmd(prot_numa=1)
1978	 *				 pmdp_huge_get_and_clear_notify()
1979	 * madvise_dontneed()
1980	 *  zap_pmd_range()
1981	 *   pmd_trans_huge(*pmd) == 0 (without ptl)
1982	 *   // skip the pmd
1983	 *				 set_pmd_at();
1984	 *				 // pmd is re-established
1985	 *
1986	 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1987	 * which may break userspace.
1988	 *
1989	 * pmdp_invalidate() is required to make sure we don't miss
1990	 * dirty/young flags set by hardware.
1991	 */
1992	entry = pmdp_invalidate(vma, addr, pmd);
1993
1994	entry = pmd_modify(entry, newprot);
1995	if (preserve_write)
1996		entry = pmd_mk_savedwrite(entry);
1997	ret = HPAGE_PMD_NR;
1998	set_pmd_at(mm, addr, pmd, entry);
1999	BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
2000unlock:
2001	spin_unlock(ptl);
2002	return ret;
2003}
2004
2005/*
2006 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
2007 *
2008 * Note that if it returns page table lock pointer, this routine returns without
2009 * unlocking page table lock. So callers must unlock it.
2010 */
2011spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2012{
2013	spinlock_t *ptl;
2014	ptl = pmd_lock(vma->vm_mm, pmd);
2015	if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2016			pmd_devmap(*pmd)))
2017		return ptl;
2018	spin_unlock(ptl);
2019	return NULL;
2020}
2021
2022/*
2023 * Returns true if a given pud maps a thp, false otherwise.
2024 *
2025 * Note that if it returns true, this routine returns without unlocking page
2026 * table lock. So callers must unlock it.
2027 */
2028spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2029{
2030	spinlock_t *ptl;
2031
2032	ptl = pud_lock(vma->vm_mm, pud);
2033	if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2034		return ptl;
2035	spin_unlock(ptl);
2036	return NULL;
2037}
2038
2039#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2040int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2041		 pud_t *pud, unsigned long addr)
2042{
 
2043	spinlock_t *ptl;
2044
2045	ptl = __pud_trans_huge_lock(pud, vma);
2046	if (!ptl)
2047		return 0;
2048	/*
2049	 * For architectures like ppc64 we look at deposited pgtable
2050	 * when calling pudp_huge_get_and_clear. So do the
2051	 * pgtable_trans_huge_withdraw after finishing pudp related
2052	 * operations.
2053	 */
2054	pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
 
2055	tlb_remove_pud_tlb_entry(tlb, pud, addr);
2056	if (vma_is_dax(vma)) {
2057		spin_unlock(ptl);
2058		/* No zero page support yet */
2059	} else {
2060		/* No support for anonymous PUD pages yet */
2061		BUG();
2062	}
2063	return 1;
2064}
2065
2066static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2067		unsigned long haddr)
2068{
2069	VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2070	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2071	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2072	VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2073
2074	count_vm_event(THP_SPLIT_PUD);
2075
2076	pudp_huge_clear_flush_notify(vma, haddr, pud);
2077}
2078
2079void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2080		unsigned long address)
2081{
2082	spinlock_t *ptl;
2083	struct mmu_notifier_range range;
 
2084
2085	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2086				address & HPAGE_PUD_MASK,
2087				(address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2088	mmu_notifier_invalidate_range_start(&range);
2089	ptl = pud_lock(vma->vm_mm, pud);
2090	if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2091		goto out;
2092	__split_huge_pud_locked(vma, pud, range.start);
2093
2094out:
2095	spin_unlock(ptl);
2096	/*
2097	 * No need to double call mmu_notifier->invalidate_range() callback as
2098	 * the above pudp_huge_clear_flush_notify() did already call it.
2099	 */
2100	mmu_notifier_invalidate_range_only_end(&range);
 
2101}
2102#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2103
2104static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2105		unsigned long haddr, pmd_t *pmd)
2106{
2107	struct mm_struct *mm = vma->vm_mm;
2108	pgtable_t pgtable;
2109	pmd_t _pmd;
2110	int i;
2111
2112	/*
2113	 * Leave pmd empty until pte is filled note that it is fine to delay
2114	 * notification until mmu_notifier_invalidate_range_end() as we are
2115	 * replacing a zero pmd write protected page with a zero pte write
2116	 * protected page.
2117	 *
2118	 * See Documentation/vm/mmu_notifier.rst
2119	 */
2120	pmdp_huge_clear_flush(vma, haddr, pmd);
2121
2122	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2123	pmd_populate(mm, &_pmd, pgtable);
2124
2125	for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2126		pte_t *pte, entry;
2127		entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2128		entry = pte_mkspecial(entry);
2129		pte = pte_offset_map(&_pmd, haddr);
2130		VM_BUG_ON(!pte_none(*pte));
2131		set_pte_at(mm, haddr, pte, entry);
2132		pte_unmap(pte);
2133	}
2134	smp_wmb(); /* make pte visible before pmd */
2135	pmd_populate(mm, pmd, pgtable);
2136}
2137
2138static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2139		unsigned long haddr, bool freeze)
2140{
2141	struct mm_struct *mm = vma->vm_mm;
2142	struct page *page;
2143	pgtable_t pgtable;
2144	pmd_t old_pmd, _pmd;
2145	bool young, write, soft_dirty, pmd_migration = false;
2146	unsigned long addr;
2147	int i;
2148
2149	VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2150	VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2151	VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2152	VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2153				&& !pmd_devmap(*pmd));
2154
2155	count_vm_event(THP_SPLIT_PMD);
2156
2157	if (!vma_is_anonymous(vma)) {
2158		_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2159		/*
2160		 * We are going to unmap this huge page. So
2161		 * just go ahead and zap it
2162		 */
2163		if (arch_needs_pgtable_deposit())
2164			zap_deposited_table(mm, pmd);
2165		if (vma_is_dax(vma))
2166			return;
2167		page = pmd_page(_pmd);
2168		if (!PageDirty(page) && pmd_dirty(_pmd))
2169			set_page_dirty(page);
2170		if (!PageReferenced(page) && pmd_young(_pmd))
2171			SetPageReferenced(page);
2172		page_remove_rmap(page, true);
2173		put_page(page);
2174		add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2175		return;
2176	} else if (is_huge_zero_pmd(*pmd)) {
2177		/*
2178		 * FIXME: Do we want to invalidate secondary mmu by calling
2179		 * mmu_notifier_invalidate_range() see comments below inside
2180		 * __split_huge_pmd() ?
2181		 *
2182		 * We are going from a zero huge page write protected to zero
2183		 * small page also write protected so it does not seems useful
2184		 * to invalidate secondary mmu at this time.
2185		 */
2186		return __split_huge_zero_page_pmd(vma, haddr, pmd);
2187	}
2188
2189	/*
2190	 * Up to this point the pmd is present and huge and userland has the
2191	 * whole access to the hugepage during the split (which happens in
2192	 * place). If we overwrite the pmd with the not-huge version pointing
2193	 * to the pte here (which of course we could if all CPUs were bug
2194	 * free), userland could trigger a small page size TLB miss on the
2195	 * small sized TLB while the hugepage TLB entry is still established in
2196	 * the huge TLB. Some CPU doesn't like that.
2197	 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2198	 * 383 on page 93. Intel should be safe but is also warns that it's
2199	 * only safe if the permission and cache attributes of the two entries
2200	 * loaded in the two TLB is identical (which should be the case here).
2201	 * But it is generally safer to never allow small and huge TLB entries
2202	 * for the same virtual address to be loaded simultaneously. So instead
2203	 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2204	 * current pmd notpresent (atomically because here the pmd_trans_huge
2205	 * must remain set at all times on the pmd until the split is complete
2206	 * for this pmd), then we flush the SMP TLB and finally we write the
2207	 * non-huge version of the pmd entry with pmd_populate.
2208	 */
2209	old_pmd = pmdp_invalidate(vma, haddr, pmd);
2210
 
2211	pmd_migration = is_pmd_migration_entry(old_pmd);
2212	if (unlikely(pmd_migration)) {
2213		swp_entry_t entry;
2214
2215		entry = pmd_to_swp_entry(old_pmd);
2216		page = pfn_to_page(swp_offset(entry));
2217		write = is_write_migration_entry(entry);
2218		young = false;
2219		soft_dirty = pmd_swp_soft_dirty(old_pmd);
2220	} else {
2221		page = pmd_page(old_pmd);
2222		if (pmd_dirty(old_pmd))
2223			SetPageDirty(page);
2224		write = pmd_write(old_pmd);
2225		young = pmd_young(old_pmd);
2226		soft_dirty = pmd_soft_dirty(old_pmd);
2227	}
2228	VM_BUG_ON_PAGE(!page_count(page), page);
2229	page_ref_add(page, HPAGE_PMD_NR - 1);
 
 
 
 
 
2230
2231	/*
2232	 * Withdraw the table only after we mark the pmd entry invalid.
2233	 * This's critical for some architectures (Power).
2234	 */
2235	pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2236	pmd_populate(mm, &_pmd, pgtable);
2237
2238	for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2239		pte_t entry, *pte;
2240		/*
2241		 * Note that NUMA hinting access restrictions are not
2242		 * transferred to avoid any possibility of altering
2243		 * permissions across VMAs.
2244		 */
2245		if (freeze || pmd_migration) {
2246			swp_entry_t swp_entry;
2247			swp_entry = make_migration_entry(page + i, write);
2248			entry = swp_entry_to_pte(swp_entry);
2249			if (soft_dirty)
2250				entry = pte_swp_mksoft_dirty(entry);
2251		} else {
2252			entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2253			entry = maybe_mkwrite(entry, vma);
2254			if (!write)
2255				entry = pte_wrprotect(entry);
2256			if (!young)
2257				entry = pte_mkold(entry);
2258			if (soft_dirty)
2259				entry = pte_mksoft_dirty(entry);
2260		}
2261		pte = pte_offset_map(&_pmd, addr);
2262		BUG_ON(!pte_none(*pte));
2263		set_pte_at(mm, addr, pte, entry);
2264		atomic_inc(&page[i]._mapcount);
2265		pte_unmap(pte);
2266	}
2267
2268	/*
2269	 * Set PG_double_map before dropping compound_mapcount to avoid
2270	 * false-negative page_mapped().
2271	 */
2272	if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2273		for (i = 0; i < HPAGE_PMD_NR; i++)
2274			atomic_inc(&page[i]._mapcount);
2275	}
2276
2277	if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2278		/* Last compound_mapcount is gone. */
2279		__dec_node_page_state(page, NR_ANON_THPS);
2280		if (TestClearPageDoubleMap(page)) {
2281			/* No need in mapcount reference anymore */
2282			for (i = 0; i < HPAGE_PMD_NR; i++)
2283				atomic_dec(&page[i]._mapcount);
2284		}
2285	}
2286
2287	smp_wmb(); /* make pte visible before pmd */
2288	pmd_populate(mm, pmd, pgtable);
2289
2290	if (freeze) {
2291		for (i = 0; i < HPAGE_PMD_NR; i++) {
2292			page_remove_rmap(page + i, false);
2293			put_page(page + i);
2294		}
2295	}
2296}
2297
2298void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2299		unsigned long address, bool freeze, struct page *page)
2300{
2301	spinlock_t *ptl;
2302	struct mmu_notifier_range range;
 
2303
2304	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2305				address & HPAGE_PMD_MASK,
2306				(address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2307	mmu_notifier_invalidate_range_start(&range);
2308	ptl = pmd_lock(vma->vm_mm, pmd);
2309
2310	/*
2311	 * If caller asks to setup a migration entries, we need a page to check
2312	 * pmd against. Otherwise we can end up replacing wrong page.
2313	 */
2314	VM_BUG_ON(freeze && !page);
2315	if (page && page != pmd_page(*pmd))
2316	        goto out;
2317
2318	if (pmd_trans_huge(*pmd)) {
2319		page = pmd_page(*pmd);
2320		if (PageMlocked(page))
2321			clear_page_mlock(page);
2322	} else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2323		goto out;
2324	__split_huge_pmd_locked(vma, pmd, range.start, freeze);
2325out:
2326	spin_unlock(ptl);
2327	/*
2328	 * No need to double call mmu_notifier->invalidate_range() callback.
2329	 * They are 3 cases to consider inside __split_huge_pmd_locked():
2330	 *  1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2331	 *  2) __split_huge_zero_page_pmd() read only zero page and any write
2332	 *    fault will trigger a flush_notify before pointing to a new page
2333	 *    (it is fine if the secondary mmu keeps pointing to the old zero
2334	 *    page in the meantime)
2335	 *  3) Split a huge pmd into pte pointing to the same page. No need
2336	 *     to invalidate secondary tlb entry they are all still valid.
2337	 *     any further changes to individual pte will notify. So no need
2338	 *     to call mmu_notifier->invalidate_range()
2339	 */
2340	mmu_notifier_invalidate_range_only_end(&range);
 
2341}
2342
2343void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2344		bool freeze, struct page *page)
2345{
2346	pgd_t *pgd;
2347	p4d_t *p4d;
2348	pud_t *pud;
2349	pmd_t *pmd;
2350
2351	pgd = pgd_offset(vma->vm_mm, address);
2352	if (!pgd_present(*pgd))
2353		return;
2354
2355	p4d = p4d_offset(pgd, address);
2356	if (!p4d_present(*p4d))
2357		return;
2358
2359	pud = pud_offset(p4d, address);
2360	if (!pud_present(*pud))
2361		return;
2362
2363	pmd = pmd_offset(pud, address);
2364
2365	__split_huge_pmd(vma, pmd, address, freeze, page);
2366}
2367
2368void vma_adjust_trans_huge(struct vm_area_struct *vma,
2369			     unsigned long start,
2370			     unsigned long end,
2371			     long adjust_next)
2372{
2373	/*
2374	 * If the new start address isn't hpage aligned and it could
2375	 * previously contain an hugepage: check if we need to split
2376	 * an huge pmd.
2377	 */
2378	if (start & ~HPAGE_PMD_MASK &&
2379	    (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2380	    (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2381		split_huge_pmd_address(vma, start, false, NULL);
2382
2383	/*
2384	 * If the new end address isn't hpage aligned and it could
2385	 * previously contain an hugepage: check if we need to split
2386	 * an huge pmd.
2387	 */
2388	if (end & ~HPAGE_PMD_MASK &&
2389	    (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2390	    (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2391		split_huge_pmd_address(vma, end, false, NULL);
2392
2393	/*
2394	 * If we're also updating the vma->vm_next->vm_start, if the new
2395	 * vm_next->vm_start isn't page aligned and it could previously
2396	 * contain an hugepage: check if we need to split an huge pmd.
2397	 */
2398	if (adjust_next > 0) {
2399		struct vm_area_struct *next = vma->vm_next;
2400		unsigned long nstart = next->vm_start;
2401		nstart += adjust_next << PAGE_SHIFT;
2402		if (nstart & ~HPAGE_PMD_MASK &&
2403		    (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2404		    (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2405			split_huge_pmd_address(next, nstart, false, NULL);
2406	}
2407}
2408
2409static void unmap_page(struct page *page)
2410{
2411	enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2412		TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2413	bool unmap_success;
2414
2415	VM_BUG_ON_PAGE(!PageHead(page), page);
2416
2417	if (PageAnon(page))
2418		ttu_flags |= TTU_SPLIT_FREEZE;
2419
2420	unmap_success = try_to_unmap(page, ttu_flags);
2421	VM_BUG_ON_PAGE(!unmap_success, page);
2422}
2423
2424static void remap_page(struct page *page)
2425{
2426	int i;
2427	if (PageTransHuge(page)) {
2428		remove_migration_ptes(page, page, true);
2429	} else {
2430		for (i = 0; i < HPAGE_PMD_NR; i++)
2431			remove_migration_ptes(page + i, page + i, true);
2432	}
2433}
2434
2435static void __split_huge_page_tail(struct page *head, int tail,
2436		struct lruvec *lruvec, struct list_head *list)
2437{
2438	struct page *page_tail = head + tail;
2439
2440	VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2441
2442	/*
2443	 * Clone page flags before unfreezing refcount.
2444	 *
2445	 * After successful get_page_unless_zero() might follow flags change,
2446	 * for exmaple lock_page() which set PG_waiters.
2447	 */
2448	page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2449	page_tail->flags |= (head->flags &
2450			((1L << PG_referenced) |
2451			 (1L << PG_swapbacked) |
2452			 (1L << PG_swapcache) |
2453			 (1L << PG_mlocked) |
2454			 (1L << PG_uptodate) |
2455			 (1L << PG_active) |
2456			 (1L << PG_workingset) |
2457			 (1L << PG_locked) |
2458			 (1L << PG_unevictable) |
2459			 (1L << PG_dirty)));
2460
2461	/* ->mapping in first tail page is compound_mapcount */
2462	VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2463			page_tail);
2464	page_tail->mapping = head->mapping;
2465	page_tail->index = head->index + tail;
2466
2467	/* Page flags must be visible before we make the page non-compound. */
2468	smp_wmb();
2469
2470	/*
2471	 * Clear PageTail before unfreezing page refcount.
2472	 *
2473	 * After successful get_page_unless_zero() might follow put_page()
2474	 * which needs correct compound_head().
2475	 */
2476	clear_compound_head(page_tail);
2477
2478	/* Finally unfreeze refcount. Additional reference from page cache. */
2479	page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2480					  PageSwapCache(head)));
2481
2482	if (page_is_young(head))
2483		set_page_young(page_tail);
2484	if (page_is_idle(head))
2485		set_page_idle(page_tail);
2486
 
 
 
 
 
 
2487	page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2488
2489	/*
2490	 * always add to the tail because some iterators expect new
2491	 * pages to show after the currently processed elements - e.g.
2492	 * migrate_pages
2493	 */
2494	lru_add_page_tail(head, page_tail, lruvec, list);
2495}
2496
2497static void __split_huge_page(struct page *page, struct list_head *list,
2498		pgoff_t end, unsigned long flags)
2499{
2500	struct page *head = compound_head(page);
2501	pg_data_t *pgdat = page_pgdat(head);
2502	struct lruvec *lruvec;
2503	struct address_space *swap_cache = NULL;
2504	unsigned long offset = 0;
2505	int i;
2506
2507	lruvec = mem_cgroup_page_lruvec(head, pgdat);
2508
2509	/* complete memcg works before add pages to LRU */
2510	mem_cgroup_split_huge_fixup(head);
2511
2512	if (PageAnon(head) && PageSwapCache(head)) {
2513		swp_entry_t entry = { .val = page_private(head) };
2514
2515		offset = swp_offset(entry);
2516		swap_cache = swap_address_space(entry);
2517		xa_lock(&swap_cache->i_pages);
2518	}
2519
2520	for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2521		__split_huge_page_tail(head, i, lruvec, list);
2522		/* Some pages can be beyond i_size: drop them from page cache */
2523		if (head[i].index >= end) {
2524			ClearPageDirty(head + i);
2525			__delete_from_page_cache(head + i, NULL);
2526			if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2527				shmem_uncharge(head->mapping->host, 1);
2528			put_page(head + i);
2529		} else if (!PageAnon(page)) {
2530			__xa_store(&head->mapping->i_pages, head[i].index,
2531					head + i, 0);
2532		} else if (swap_cache) {
2533			__xa_store(&swap_cache->i_pages, offset + i,
2534					head + i, 0);
2535		}
2536	}
2537
2538	ClearPageCompound(head);
2539
2540	split_page_owner(head, HPAGE_PMD_ORDER);
2541
2542	/* See comment in __split_huge_page_tail() */
2543	if (PageAnon(head)) {
2544		/* Additional pin to swap cache */
2545		if (PageSwapCache(head)) {
2546			page_ref_add(head, 2);
2547			xa_unlock(&swap_cache->i_pages);
2548		} else {
2549			page_ref_inc(head);
2550		}
2551	} else {
2552		/* Additional pin to page cache */
2553		page_ref_add(head, 2);
2554		xa_unlock(&head->mapping->i_pages);
2555	}
2556
2557	spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2558
2559	remap_page(head);
2560
2561	for (i = 0; i < HPAGE_PMD_NR; i++) {
2562		struct page *subpage = head + i;
2563		if (subpage == page)
2564			continue;
2565		unlock_page(subpage);
2566
2567		/*
2568		 * Subpages may be freed if there wasn't any mapping
2569		 * like if add_to_swap() is running on a lru page that
2570		 * had its mapping zapped. And freeing these pages
2571		 * requires taking the lru_lock so we do the put_page
2572		 * of the tail pages after the split is complete.
2573		 */
2574		put_page(subpage);
2575	}
2576}
2577
2578int total_mapcount(struct page *page)
2579{
2580	int i, compound, ret;
2581
2582	VM_BUG_ON_PAGE(PageTail(page), page);
2583
2584	if (likely(!PageCompound(page)))
2585		return atomic_read(&page->_mapcount) + 1;
2586
2587	compound = compound_mapcount(page);
2588	if (PageHuge(page))
2589		return compound;
2590	ret = compound;
2591	for (i = 0; i < HPAGE_PMD_NR; i++)
2592		ret += atomic_read(&page[i]._mapcount) + 1;
2593	/* File pages has compound_mapcount included in _mapcount */
2594	if (!PageAnon(page))
2595		return ret - compound * HPAGE_PMD_NR;
2596	if (PageDoubleMap(page))
2597		ret -= HPAGE_PMD_NR;
2598	return ret;
2599}
2600
2601/*
2602 * This calculates accurately how many mappings a transparent hugepage
2603 * has (unlike page_mapcount() which isn't fully accurate). This full
2604 * accuracy is primarily needed to know if copy-on-write faults can
2605 * reuse the page and change the mapping to read-write instead of
2606 * copying them. At the same time this returns the total_mapcount too.
2607 *
2608 * The function returns the highest mapcount any one of the subpages
2609 * has. If the return value is one, even if different processes are
2610 * mapping different subpages of the transparent hugepage, they can
2611 * all reuse it, because each process is reusing a different subpage.
2612 *
2613 * The total_mapcount is instead counting all virtual mappings of the
2614 * subpages. If the total_mapcount is equal to "one", it tells the
2615 * caller all mappings belong to the same "mm" and in turn the
2616 * anon_vma of the transparent hugepage can become the vma->anon_vma
2617 * local one as no other process may be mapping any of the subpages.
2618 *
2619 * It would be more accurate to replace page_mapcount() with
2620 * page_trans_huge_mapcount(), however we only use
2621 * page_trans_huge_mapcount() in the copy-on-write faults where we
2622 * need full accuracy to avoid breaking page pinning, because
2623 * page_trans_huge_mapcount() is slower than page_mapcount().
2624 */
2625int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2626{
2627	int i, ret, _total_mapcount, mapcount;
2628
2629	/* hugetlbfs shouldn't call it */
2630	VM_BUG_ON_PAGE(PageHuge(page), page);
2631
2632	if (likely(!PageTransCompound(page))) {
2633		mapcount = atomic_read(&page->_mapcount) + 1;
2634		if (total_mapcount)
2635			*total_mapcount = mapcount;
2636		return mapcount;
2637	}
2638
2639	page = compound_head(page);
2640
2641	_total_mapcount = ret = 0;
2642	for (i = 0; i < HPAGE_PMD_NR; i++) {
2643		mapcount = atomic_read(&page[i]._mapcount) + 1;
2644		ret = max(ret, mapcount);
2645		_total_mapcount += mapcount;
2646	}
2647	if (PageDoubleMap(page)) {
2648		ret -= 1;
2649		_total_mapcount -= HPAGE_PMD_NR;
2650	}
2651	mapcount = compound_mapcount(page);
2652	ret += mapcount;
2653	_total_mapcount += mapcount;
2654	if (total_mapcount)
2655		*total_mapcount = _total_mapcount;
2656	return ret;
2657}
2658
2659/* Racy check whether the huge page can be split */
2660bool can_split_huge_page(struct page *page, int *pextra_pins)
2661{
2662	int extra_pins;
2663
2664	/* Additional pins from page cache */
2665	if (PageAnon(page))
2666		extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2667	else
2668		extra_pins = HPAGE_PMD_NR;
2669	if (pextra_pins)
2670		*pextra_pins = extra_pins;
2671	return total_mapcount(page) == page_count(page) - extra_pins - 1;
2672}
2673
2674/*
2675 * This function splits huge page into normal pages. @page can point to any
2676 * subpage of huge page to split. Split doesn't change the position of @page.
2677 *
2678 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2679 * The huge page must be locked.
2680 *
2681 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2682 *
2683 * Both head page and tail pages will inherit mapping, flags, and so on from
2684 * the hugepage.
2685 *
2686 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2687 * they are not mapped.
2688 *
2689 * Returns 0 if the hugepage is split successfully.
2690 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2691 * us.
2692 */
2693int split_huge_page_to_list(struct page *page, struct list_head *list)
2694{
2695	struct page *head = compound_head(page);
2696	struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2697	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2698	struct anon_vma *anon_vma = NULL;
2699	struct address_space *mapping = NULL;
2700	int count, mapcount, extra_pins, ret;
2701	bool mlocked;
2702	unsigned long flags;
2703	pgoff_t end;
2704
2705	VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2706	VM_BUG_ON_PAGE(!PageLocked(page), page);
2707	VM_BUG_ON_PAGE(!PageCompound(page), page);
2708
2709	if (PageWriteback(page))
2710		return -EBUSY;
2711
2712	if (PageAnon(head)) {
2713		/*
2714		 * The caller does not necessarily hold an mmap_sem that would
2715		 * prevent the anon_vma disappearing so we first we take a
2716		 * reference to it and then lock the anon_vma for write. This
2717		 * is similar to page_lock_anon_vma_read except the write lock
2718		 * is taken to serialise against parallel split or collapse
2719		 * operations.
2720		 */
2721		anon_vma = page_get_anon_vma(head);
2722		if (!anon_vma) {
2723			ret = -EBUSY;
2724			goto out;
2725		}
2726		end = -1;
2727		mapping = NULL;
2728		anon_vma_lock_write(anon_vma);
2729	} else {
2730		mapping = head->mapping;
2731
2732		/* Truncated ? */
2733		if (!mapping) {
2734			ret = -EBUSY;
2735			goto out;
2736		}
2737
2738		anon_vma = NULL;
2739		i_mmap_lock_read(mapping);
2740
2741		/*
2742		 *__split_huge_page() may need to trim off pages beyond EOF:
2743		 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2744		 * which cannot be nested inside the page tree lock. So note
2745		 * end now: i_size itself may be changed at any moment, but
2746		 * head page lock is good enough to serialize the trimming.
2747		 */
2748		end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2749	}
2750
2751	/*
2752	 * Racy check if we can split the page, before unmap_page() will
2753	 * split PMDs
2754	 */
2755	if (!can_split_huge_page(head, &extra_pins)) {
2756		ret = -EBUSY;
2757		goto out_unlock;
2758	}
2759
2760	mlocked = PageMlocked(page);
2761	unmap_page(head);
2762	VM_BUG_ON_PAGE(compound_mapcount(head), head);
2763
2764	/* Make sure the page is not on per-CPU pagevec as it takes pin */
2765	if (mlocked)
2766		lru_add_drain();
2767
2768	/* prevent PageLRU to go away from under us, and freeze lru stats */
2769	spin_lock_irqsave(&pgdata->lru_lock, flags);
2770
2771	if (mapping) {
2772		XA_STATE(xas, &mapping->i_pages, page_index(head));
2773
 
 
 
2774		/*
2775		 * Check if the head page is present in page cache.
2776		 * We assume all tail are present too, if head is there.
2777		 */
2778		xa_lock(&mapping->i_pages);
2779		if (xas_load(&xas) != head)
2780			goto fail;
2781	}
2782
2783	/* Prevent deferred_split_scan() touching ->_refcount */
2784	spin_lock(&ds_queue->split_queue_lock);
2785	count = page_count(head);
2786	mapcount = total_mapcount(head);
2787	if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2788		if (!list_empty(page_deferred_list(head))) {
2789			ds_queue->split_queue_len--;
2790			list_del(page_deferred_list(head));
2791		}
2792		if (mapping) {
2793			if (PageSwapBacked(page))
2794				__dec_node_page_state(page, NR_SHMEM_THPS);
2795			else
2796				__dec_node_page_state(page, NR_FILE_THPS);
2797		}
2798
2799		spin_unlock(&ds_queue->split_queue_lock);
2800		__split_huge_page(page, list, end, flags);
2801		if (PageSwapCache(head)) {
2802			swp_entry_t entry = { .val = page_private(head) };
2803
2804			ret = split_swap_cluster(entry);
2805		} else
2806			ret = 0;
2807	} else {
2808		if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2809			pr_alert("total_mapcount: %u, page_count(): %u\n",
2810					mapcount, count);
2811			if (PageTail(page))
2812				dump_page(head, NULL);
2813			dump_page(page, "total_mapcount(head) > 0");
2814			BUG();
2815		}
2816		spin_unlock(&ds_queue->split_queue_lock);
2817fail:		if (mapping)
2818			xa_unlock(&mapping->i_pages);
2819		spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2820		remap_page(head);
2821		ret = -EBUSY;
2822	}
2823
2824out_unlock:
2825	if (anon_vma) {
2826		anon_vma_unlock_write(anon_vma);
2827		put_anon_vma(anon_vma);
2828	}
2829	if (mapping)
2830		i_mmap_unlock_read(mapping);
2831out:
2832	count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2833	return ret;
2834}
2835
2836void free_transhuge_page(struct page *page)
2837{
2838	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2839	unsigned long flags;
2840
2841	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2842	if (!list_empty(page_deferred_list(page))) {
2843		ds_queue->split_queue_len--;
2844		list_del(page_deferred_list(page));
2845	}
2846	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2847	free_compound_page(page);
2848}
2849
2850void deferred_split_huge_page(struct page *page)
2851{
2852	struct deferred_split *ds_queue = get_deferred_split_queue(page);
2853#ifdef CONFIG_MEMCG
2854	struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2855#endif
2856	unsigned long flags;
2857
2858	VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2859
2860	/*
2861	 * The try_to_unmap() in page reclaim path might reach here too,
2862	 * this may cause a race condition to corrupt deferred split queue.
2863	 * And, if page reclaim is already handling the same page, it is
2864	 * unnecessary to handle it again in shrinker.
2865	 *
2866	 * Check PageSwapCache to determine if the page is being
2867	 * handled by page reclaim since THP swap would add the page into
2868	 * swap cache before calling try_to_unmap().
2869	 */
2870	if (PageSwapCache(page))
2871		return;
2872
2873	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2874	if (list_empty(page_deferred_list(page))) {
2875		count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2876		list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2877		ds_queue->split_queue_len++;
2878#ifdef CONFIG_MEMCG
2879		if (memcg)
2880			memcg_set_shrinker_bit(memcg, page_to_nid(page),
2881					       deferred_split_shrinker.id);
2882#endif
2883	}
2884	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2885}
2886
2887static unsigned long deferred_split_count(struct shrinker *shrink,
2888		struct shrink_control *sc)
2889{
2890	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2891	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2892
2893#ifdef CONFIG_MEMCG
2894	if (sc->memcg)
2895		ds_queue = &sc->memcg->deferred_split_queue;
2896#endif
2897	return READ_ONCE(ds_queue->split_queue_len);
2898}
2899
2900static unsigned long deferred_split_scan(struct shrinker *shrink,
2901		struct shrink_control *sc)
2902{
2903	struct pglist_data *pgdata = NODE_DATA(sc->nid);
2904	struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2905	unsigned long flags;
2906	LIST_HEAD(list), *pos, *next;
2907	struct page *page;
2908	int split = 0;
2909
2910#ifdef CONFIG_MEMCG
2911	if (sc->memcg)
2912		ds_queue = &sc->memcg->deferred_split_queue;
2913#endif
2914
2915	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2916	/* Take pin on all head pages to avoid freeing them under us */
2917	list_for_each_safe(pos, next, &ds_queue->split_queue) {
2918		page = list_entry((void *)pos, struct page, mapping);
2919		page = compound_head(page);
2920		if (get_page_unless_zero(page)) {
2921			list_move(page_deferred_list(page), &list);
2922		} else {
2923			/* We lost race with put_compound_page() */
2924			list_del_init(page_deferred_list(page));
2925			ds_queue->split_queue_len--;
2926		}
2927		if (!--sc->nr_to_scan)
2928			break;
2929	}
2930	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2931
2932	list_for_each_safe(pos, next, &list) {
2933		page = list_entry((void *)pos, struct page, mapping);
2934		if (!trylock_page(page))
2935			goto next;
2936		/* split_huge_page() removes page from list on success */
2937		if (!split_huge_page(page))
2938			split++;
2939		unlock_page(page);
2940next:
2941		put_page(page);
2942	}
2943
2944	spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2945	list_splice_tail(&list, &ds_queue->split_queue);
2946	spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2947
2948	/*
2949	 * Stop shrinker if we didn't split any page, but the queue is empty.
2950	 * This can happen if pages were freed under us.
2951	 */
2952	if (!split && list_empty(&ds_queue->split_queue))
2953		return SHRINK_STOP;
2954	return split;
2955}
2956
2957static struct shrinker deferred_split_shrinker = {
2958	.count_objects = deferred_split_count,
2959	.scan_objects = deferred_split_scan,
2960	.seeks = DEFAULT_SEEKS,
2961	.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2962		 SHRINKER_NONSLAB,
2963};
2964
2965#ifdef CONFIG_DEBUG_FS
2966static int split_huge_pages_set(void *data, u64 val)
2967{
2968	struct zone *zone;
2969	struct page *page;
2970	unsigned long pfn, max_zone_pfn;
2971	unsigned long total = 0, split = 0;
2972
2973	if (val != 1)
2974		return -EINVAL;
2975
2976	for_each_populated_zone(zone) {
2977		max_zone_pfn = zone_end_pfn(zone);
2978		for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2979			if (!pfn_valid(pfn))
2980				continue;
2981
2982			page = pfn_to_page(pfn);
2983			if (!get_page_unless_zero(page))
2984				continue;
2985
2986			if (zone != page_zone(page))
2987				goto next;
2988
2989			if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2990				goto next;
2991
2992			total++;
2993			lock_page(page);
2994			if (!split_huge_page(page))
2995				split++;
2996			unlock_page(page);
2997next:
2998			put_page(page);
2999		}
3000	}
3001
3002	pr_info("%lu of %lu THP split\n", split, total);
3003
3004	return 0;
3005}
3006DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3007		"%llu\n");
3008
3009static int __init split_huge_pages_debugfs(void)
3010{
3011	debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3012			    &split_huge_pages_fops);
 
 
 
 
3013	return 0;
3014}
3015late_initcall(split_huge_pages_debugfs);
3016#endif
3017
3018#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3019void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3020		struct page *page)
3021{
3022	struct vm_area_struct *vma = pvmw->vma;
3023	struct mm_struct *mm = vma->vm_mm;
3024	unsigned long address = pvmw->address;
3025	pmd_t pmdval;
3026	swp_entry_t entry;
3027	pmd_t pmdswp;
3028
3029	if (!(pvmw->pmd && !pvmw->pte))
3030		return;
3031
 
 
 
3032	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3033	pmdval = *pvmw->pmd;
3034	pmdp_invalidate(vma, address, pvmw->pmd);
3035	if (pmd_dirty(pmdval))
3036		set_page_dirty(page);
3037	entry = make_migration_entry(page, pmd_write(pmdval));
3038	pmdswp = swp_entry_to_pmd(entry);
3039	if (pmd_soft_dirty(pmdval))
3040		pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3041	set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3042	page_remove_rmap(page, true);
3043	put_page(page);
 
 
 
3044}
3045
3046void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3047{
3048	struct vm_area_struct *vma = pvmw->vma;
3049	struct mm_struct *mm = vma->vm_mm;
3050	unsigned long address = pvmw->address;
3051	unsigned long mmun_start = address & HPAGE_PMD_MASK;
3052	pmd_t pmde;
3053	swp_entry_t entry;
3054
3055	if (!(pvmw->pmd && !pvmw->pte))
3056		return;
3057
3058	entry = pmd_to_swp_entry(*pvmw->pmd);
3059	get_page(new);
3060	pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3061	if (pmd_swp_soft_dirty(*pvmw->pmd))
3062		pmde = pmd_mksoft_dirty(pmde);
3063	if (is_write_migration_entry(entry))
3064		pmde = maybe_pmd_mkwrite(pmde, vma);
3065
3066	flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3067	if (PageAnon(new))
3068		page_add_anon_rmap(new, vma, mmun_start, true);
3069	else
3070		page_add_file_rmap(new, true);
3071	set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3072	if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3073		mlock_vma_page(new);
3074	update_mmu_cache_pmd(vma, address, pvmw->pmd);
3075}
3076#endif