1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2009 Red Hat, Inc.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
38 #include <asm/pgalloc.h>
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.
49 unsigned long transparent_hugepage_flags __read_mostly =
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
60 static struct shrinker deferred_split_shrinker;
62 static atomic_t huge_zero_refcount;
63 struct page *huge_zero_page __read_mostly;
64 unsigned long huge_zero_pfn __read_mostly = ~0UL;
66 static inline bool file_thp_enabled(struct vm_area_struct *vma)
68 return transhuge_vma_enabled(vma, vma->vm_flags) && vma->vm_file &&
69 !inode_is_open_for_write(vma->vm_file->f_inode) &&
70 (vma->vm_flags & VM_EXEC);
73 bool transparent_hugepage_active(struct vm_area_struct *vma)
75 /* The addr is used to check if the vma size fits */
76 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
78 if (!transhuge_vma_suitable(vma, addr))
80 if (vma_is_anonymous(vma))
81 return __transparent_hugepage_enabled(vma);
82 if (vma_is_shmem(vma))
83 return shmem_huge_enabled(vma);
84 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
85 return file_thp_enabled(vma);
90 static struct page *get_huge_zero_page(void)
92 struct page *zero_page;
94 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
95 return READ_ONCE(huge_zero_page);
97 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
100 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
103 count_vm_event(THP_ZERO_PAGE_ALLOC);
105 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
107 __free_pages(zero_page, compound_order(zero_page));
110 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
112 /* We take additional reference here. It will be put back by shrinker */
113 atomic_set(&huge_zero_refcount, 2);
115 return READ_ONCE(huge_zero_page);
118 static void put_huge_zero_page(void)
121 * Counter should never go to zero here. Only shrinker can put
124 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
127 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
129 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
130 return READ_ONCE(huge_zero_page);
132 if (!get_huge_zero_page())
135 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
136 put_huge_zero_page();
138 return READ_ONCE(huge_zero_page);
141 void mm_put_huge_zero_page(struct mm_struct *mm)
143 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
144 put_huge_zero_page();
147 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
148 struct shrink_control *sc)
150 /* we can free zero page only if last reference remains */
151 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
154 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
155 struct shrink_control *sc)
157 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
158 struct page *zero_page = xchg(&huge_zero_page, NULL);
159 BUG_ON(zero_page == NULL);
160 WRITE_ONCE(huge_zero_pfn, ~0UL);
161 __free_pages(zero_page, compound_order(zero_page));
168 static struct shrinker huge_zero_page_shrinker = {
169 .count_objects = shrink_huge_zero_page_count,
170 .scan_objects = shrink_huge_zero_page_scan,
171 .seeks = DEFAULT_SEEKS,
175 static ssize_t enabled_show(struct kobject *kobj,
176 struct kobj_attribute *attr, char *buf)
178 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
179 return sprintf(buf, "[always] madvise never\n");
180 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
181 return sprintf(buf, "always [madvise] never\n");
183 return sprintf(buf, "always madvise [never]\n");
186 static ssize_t enabled_store(struct kobject *kobj,
187 struct kobj_attribute *attr,
188 const char *buf, size_t count)
192 if (sysfs_streq(buf, "always")) {
193 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
194 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
195 } else if (sysfs_streq(buf, "madvise")) {
196 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
197 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
198 } else if (sysfs_streq(buf, "never")) {
199 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
200 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
205 int err = start_stop_khugepaged();
211 static struct kobj_attribute enabled_attr =
212 __ATTR(enabled, 0644, enabled_show, enabled_store);
214 ssize_t single_hugepage_flag_show(struct kobject *kobj,
215 struct kobj_attribute *attr, char *buf,
216 enum transparent_hugepage_flag flag)
218 return sprintf(buf, "%d\n",
219 !!test_bit(flag, &transparent_hugepage_flags));
222 ssize_t single_hugepage_flag_store(struct kobject *kobj,
223 struct kobj_attribute *attr,
224 const char *buf, size_t count,
225 enum transparent_hugepage_flag flag)
230 ret = kstrtoul(buf, 10, &value);
237 set_bit(flag, &transparent_hugepage_flags);
239 clear_bit(flag, &transparent_hugepage_flags);
244 static ssize_t defrag_show(struct kobject *kobj,
245 struct kobj_attribute *attr, char *buf)
247 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
248 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
249 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
250 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
251 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
252 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
253 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
254 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
255 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
258 static ssize_t defrag_store(struct kobject *kobj,
259 struct kobj_attribute *attr,
260 const char *buf, size_t count)
262 if (sysfs_streq(buf, "always")) {
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
266 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
267 } else if (sysfs_streq(buf, "defer+madvise")) {
268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
269 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
270 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
271 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
272 } else if (sysfs_streq(buf, "defer")) {
273 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
275 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
276 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
277 } else if (sysfs_streq(buf, "madvise")) {
278 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
279 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
280 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
281 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
282 } else if (sysfs_streq(buf, "never")) {
283 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
284 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
285 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
286 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
292 static struct kobj_attribute defrag_attr =
293 __ATTR(defrag, 0644, defrag_show, defrag_store);
295 static ssize_t use_zero_page_show(struct kobject *kobj,
296 struct kobj_attribute *attr, char *buf)
298 return single_hugepage_flag_show(kobj, attr, buf,
299 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
301 static ssize_t use_zero_page_store(struct kobject *kobj,
302 struct kobj_attribute *attr, const char *buf, size_t count)
304 return single_hugepage_flag_store(kobj, attr, buf, count,
305 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
307 static struct kobj_attribute use_zero_page_attr =
308 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
310 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf)
313 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
315 static struct kobj_attribute hpage_pmd_size_attr =
316 __ATTR_RO(hpage_pmd_size);
318 static struct attribute *hugepage_attr[] = {
321 &use_zero_page_attr.attr,
322 &hpage_pmd_size_attr.attr,
324 &shmem_enabled_attr.attr,
329 static const struct attribute_group hugepage_attr_group = {
330 .attrs = hugepage_attr,
333 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
337 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
338 if (unlikely(!*hugepage_kobj)) {
339 pr_err("failed to create transparent hugepage kobject\n");
343 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
345 pr_err("failed to register transparent hugepage group\n");
349 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
351 pr_err("failed to register transparent hugepage group\n");
352 goto remove_hp_group;
358 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
360 kobject_put(*hugepage_kobj);
364 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
366 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
367 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
368 kobject_put(hugepage_kobj);
371 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
376 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
379 #endif /* CONFIG_SYSFS */
381 static int __init hugepage_init(void)
384 struct kobject *hugepage_kobj;
386 if (!has_transparent_hugepage()) {
388 * Hardware doesn't support hugepages, hence disable
391 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
396 * hugepages can't be allocated by the buddy allocator
398 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
400 * we use page->mapping and page->index in second tail page
401 * as list_head: assuming THP order >= 2
403 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
405 err = hugepage_init_sysfs(&hugepage_kobj);
409 err = khugepaged_init();
413 err = register_shrinker(&huge_zero_page_shrinker);
415 goto err_hzp_shrinker;
416 err = register_shrinker(&deferred_split_shrinker);
418 goto err_split_shrinker;
421 * By default disable transparent hugepages on smaller systems,
422 * where the extra memory used could hurt more than TLB overhead
423 * is likely to save. The admin can still enable it through /sys.
425 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
426 transparent_hugepage_flags = 0;
430 err = start_stop_khugepaged();
436 unregister_shrinker(&deferred_split_shrinker);
438 unregister_shrinker(&huge_zero_page_shrinker);
440 khugepaged_destroy();
442 hugepage_exit_sysfs(hugepage_kobj);
446 subsys_initcall(hugepage_init);
448 static int __init setup_transparent_hugepage(char *str)
453 if (!strcmp(str, "always")) {
454 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
455 &transparent_hugepage_flags);
456 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
457 &transparent_hugepage_flags);
459 } else if (!strcmp(str, "madvise")) {
460 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
461 &transparent_hugepage_flags);
462 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
463 &transparent_hugepage_flags);
465 } else if (!strcmp(str, "never")) {
466 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
467 &transparent_hugepage_flags);
468 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
469 &transparent_hugepage_flags);
474 pr_warn("transparent_hugepage= cannot parse, ignored\n");
477 __setup("transparent_hugepage=", setup_transparent_hugepage);
479 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
481 if (likely(vma->vm_flags & VM_WRITE))
482 pmd = pmd_mkwrite(pmd);
487 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
489 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
490 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
493 return &memcg->deferred_split_queue;
495 return &pgdat->deferred_split_queue;
498 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
500 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
502 return &pgdat->deferred_split_queue;
506 void prep_transhuge_page(struct page *page)
509 * we use page->mapping and page->indexlru in second tail page
510 * as list_head: assuming THP order >= 2
513 INIT_LIST_HEAD(page_deferred_list(page));
514 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
517 bool is_transparent_hugepage(struct page *page)
519 if (!PageCompound(page))
522 page = compound_head(page);
523 return is_huge_zero_page(page) ||
524 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
526 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
528 static unsigned long __thp_get_unmapped_area(struct file *filp,
529 unsigned long addr, unsigned long len,
530 loff_t off, unsigned long flags, unsigned long size)
532 loff_t off_end = off + len;
533 loff_t off_align = round_up(off, size);
534 unsigned long len_pad, ret;
536 if (off_end <= off_align || (off_end - off_align) < size)
539 len_pad = len + size;
540 if (len_pad < len || (off + len_pad) < off)
543 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
544 off >> PAGE_SHIFT, flags);
547 * The failure might be due to length padding. The caller will retry
548 * without the padding.
550 if (IS_ERR_VALUE(ret))
554 * Do not try to align to THP boundary if allocation at the address
560 ret += (off - ret) & (size - 1);
564 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
565 unsigned long len, unsigned long pgoff, unsigned long flags)
568 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
570 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
573 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
577 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
579 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
581 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
582 struct page *page, gfp_t gfp)
584 struct vm_area_struct *vma = vmf->vma;
586 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
589 VM_BUG_ON_PAGE(!PageCompound(page), page);
591 if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
593 count_vm_event(THP_FAULT_FALLBACK);
594 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
595 return VM_FAULT_FALLBACK;
597 cgroup_throttle_swaprate(page, gfp);
599 pgtable = pte_alloc_one(vma->vm_mm);
600 if (unlikely(!pgtable)) {
605 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
607 * The memory barrier inside __SetPageUptodate makes sure that
608 * clear_huge_page writes become visible before the set_pmd_at()
611 __SetPageUptodate(page);
613 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
614 if (unlikely(!pmd_none(*vmf->pmd))) {
619 ret = check_stable_address_space(vma->vm_mm);
623 /* Deliver the page fault to userland */
624 if (userfaultfd_missing(vma)) {
627 spin_unlock(vmf->ptl);
629 pte_free(vma->vm_mm, pgtable);
630 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
631 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
635 entry = mk_huge_pmd(page, vma->vm_page_prot);
636 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
637 page_add_new_anon_rmap(page, vma, haddr, true);
638 lru_cache_add_inactive_or_unevictable(page, vma);
639 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
640 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
641 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
642 mm_inc_nr_ptes(vma->vm_mm);
643 spin_unlock(vmf->ptl);
644 count_vm_event(THP_FAULT_ALLOC);
645 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
650 spin_unlock(vmf->ptl);
653 pte_free(vma->vm_mm, pgtable);
660 * always: directly stall for all thp allocations
661 * defer: wake kswapd and fail if not immediately available
662 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
663 * fail if not immediately available
664 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
666 * never: never stall for any thp allocation
668 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
670 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
672 /* Always do synchronous compaction */
673 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
674 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
676 /* Kick kcompactd and fail quickly */
677 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
678 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
680 /* Synchronous compaction if madvised, otherwise kick kcompactd */
681 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
682 return GFP_TRANSHUGE_LIGHT |
683 (vma_madvised ? __GFP_DIRECT_RECLAIM :
684 __GFP_KSWAPD_RECLAIM);
686 /* Only do synchronous compaction if madvised */
687 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
688 return GFP_TRANSHUGE_LIGHT |
689 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
691 return GFP_TRANSHUGE_LIGHT;
694 /* Caller must hold page table lock. */
695 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
696 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
697 struct page *zero_page)
702 entry = mk_pmd(zero_page, vma->vm_page_prot);
703 entry = pmd_mkhuge(entry);
705 pgtable_trans_huge_deposit(mm, pmd, pgtable);
706 set_pmd_at(mm, haddr, pmd, entry);
711 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
713 struct vm_area_struct *vma = vmf->vma;
716 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
718 if (!transhuge_vma_suitable(vma, haddr))
719 return VM_FAULT_FALLBACK;
720 if (unlikely(anon_vma_prepare(vma)))
722 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
724 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
725 !mm_forbids_zeropage(vma->vm_mm) &&
726 transparent_hugepage_use_zero_page()) {
728 struct page *zero_page;
730 pgtable = pte_alloc_one(vma->vm_mm);
731 if (unlikely(!pgtable))
733 zero_page = mm_get_huge_zero_page(vma->vm_mm);
734 if (unlikely(!zero_page)) {
735 pte_free(vma->vm_mm, pgtable);
736 count_vm_event(THP_FAULT_FALLBACK);
737 return VM_FAULT_FALLBACK;
739 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
741 if (pmd_none(*vmf->pmd)) {
742 ret = check_stable_address_space(vma->vm_mm);
744 spin_unlock(vmf->ptl);
745 pte_free(vma->vm_mm, pgtable);
746 } else if (userfaultfd_missing(vma)) {
747 spin_unlock(vmf->ptl);
748 pte_free(vma->vm_mm, pgtable);
749 ret = handle_userfault(vmf, VM_UFFD_MISSING);
750 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
752 set_huge_zero_page(pgtable, vma->vm_mm, vma,
753 haddr, vmf->pmd, zero_page);
754 spin_unlock(vmf->ptl);
757 spin_unlock(vmf->ptl);
758 pte_free(vma->vm_mm, pgtable);
762 gfp = alloc_hugepage_direct_gfpmask(vma);
763 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
764 if (unlikely(!page)) {
765 count_vm_event(THP_FAULT_FALLBACK);
766 return VM_FAULT_FALLBACK;
768 prep_transhuge_page(page);
769 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
772 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
773 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
776 struct mm_struct *mm = vma->vm_mm;
780 ptl = pmd_lock(mm, pmd);
781 if (!pmd_none(*pmd)) {
783 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
784 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
787 entry = pmd_mkyoung(*pmd);
788 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
789 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
790 update_mmu_cache_pmd(vma, addr, pmd);
796 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
797 if (pfn_t_devmap(pfn))
798 entry = pmd_mkdevmap(entry);
800 entry = pmd_mkyoung(pmd_mkdirty(entry));
801 entry = maybe_pmd_mkwrite(entry, vma);
805 pgtable_trans_huge_deposit(mm, pmd, pgtable);
810 set_pmd_at(mm, addr, pmd, entry);
811 update_mmu_cache_pmd(vma, addr, pmd);
816 pte_free(mm, pgtable);
820 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
821 * @vmf: Structure describing the fault
822 * @pfn: pfn to insert
823 * @pgprot: page protection to use
824 * @write: whether it's a write fault
826 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
827 * also consult the vmf_insert_mixed_prot() documentation when
828 * @pgprot != @vmf->vma->vm_page_prot.
830 * Return: vm_fault_t value.
832 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
833 pgprot_t pgprot, bool write)
835 unsigned long addr = vmf->address & PMD_MASK;
836 struct vm_area_struct *vma = vmf->vma;
837 pgtable_t pgtable = NULL;
840 * If we had pmd_special, we could avoid all these restrictions,
841 * but we need to be consistent with PTEs and architectures that
842 * can't support a 'special' bit.
844 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
846 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
847 (VM_PFNMAP|VM_MIXEDMAP));
848 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
850 if (addr < vma->vm_start || addr >= vma->vm_end)
851 return VM_FAULT_SIGBUS;
853 if (arch_needs_pgtable_deposit()) {
854 pgtable = pte_alloc_one(vma->vm_mm);
859 track_pfn_insert(vma, &pgprot, pfn);
861 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
862 return VM_FAULT_NOPAGE;
864 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
866 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
867 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
869 if (likely(vma->vm_flags & VM_WRITE))
870 pud = pud_mkwrite(pud);
874 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
875 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
877 struct mm_struct *mm = vma->vm_mm;
881 ptl = pud_lock(mm, pud);
882 if (!pud_none(*pud)) {
884 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
885 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
888 entry = pud_mkyoung(*pud);
889 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
890 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
891 update_mmu_cache_pud(vma, addr, pud);
896 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
897 if (pfn_t_devmap(pfn))
898 entry = pud_mkdevmap(entry);
900 entry = pud_mkyoung(pud_mkdirty(entry));
901 entry = maybe_pud_mkwrite(entry, vma);
903 set_pud_at(mm, addr, pud, entry);
904 update_mmu_cache_pud(vma, addr, pud);
911 * vmf_insert_pfn_pud_prot - insert a pud size pfn
912 * @vmf: Structure describing the fault
913 * @pfn: pfn to insert
914 * @pgprot: page protection to use
915 * @write: whether it's a write fault
917 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
918 * also consult the vmf_insert_mixed_prot() documentation when
919 * @pgprot != @vmf->vma->vm_page_prot.
921 * Return: vm_fault_t value.
923 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
924 pgprot_t pgprot, bool write)
926 unsigned long addr = vmf->address & PUD_MASK;
927 struct vm_area_struct *vma = vmf->vma;
930 * If we had pud_special, we could avoid all these restrictions,
931 * but we need to be consistent with PTEs and architectures that
932 * can't support a 'special' bit.
934 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
936 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
937 (VM_PFNMAP|VM_MIXEDMAP));
938 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
940 if (addr < vma->vm_start || addr >= vma->vm_end)
941 return VM_FAULT_SIGBUS;
943 track_pfn_insert(vma, &pgprot, pfn);
945 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
946 return VM_FAULT_NOPAGE;
948 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
949 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
951 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
952 pmd_t *pmd, int flags)
956 _pmd = pmd_mkyoung(*pmd);
957 if (flags & FOLL_WRITE)
958 _pmd = pmd_mkdirty(_pmd);
959 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
960 pmd, _pmd, flags & FOLL_WRITE))
961 update_mmu_cache_pmd(vma, addr, pmd);
964 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
965 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
967 unsigned long pfn = pmd_pfn(*pmd);
968 struct mm_struct *mm = vma->vm_mm;
971 assert_spin_locked(pmd_lockptr(mm, pmd));
974 * When we COW a devmap PMD entry, we split it into PTEs, so we should
975 * not be in this function with `flags & FOLL_COW` set.
977 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
979 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
980 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
981 (FOLL_PIN | FOLL_GET)))
984 if (flags & FOLL_WRITE && !pmd_write(*pmd))
987 if (pmd_present(*pmd) && pmd_devmap(*pmd))
992 if (flags & FOLL_TOUCH)
993 touch_pmd(vma, addr, pmd, flags);
996 * device mapped pages can only be returned if the
997 * caller will manage the page reference count.
999 if (!(flags & (FOLL_GET | FOLL_PIN)))
1000 return ERR_PTR(-EEXIST);
1002 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1003 *pgmap = get_dev_pagemap(pfn, *pgmap);
1005 return ERR_PTR(-EFAULT);
1006 page = pfn_to_page(pfn);
1007 if (!try_grab_page(page, flags))
1008 page = ERR_PTR(-ENOMEM);
1013 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1014 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1015 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1017 spinlock_t *dst_ptl, *src_ptl;
1018 struct page *src_page;
1020 pgtable_t pgtable = NULL;
1023 /* Skip if can be re-fill on fault */
1024 if (!vma_is_anonymous(dst_vma))
1027 pgtable = pte_alloc_one(dst_mm);
1028 if (unlikely(!pgtable))
1031 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1032 src_ptl = pmd_lockptr(src_mm, src_pmd);
1033 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1038 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1039 if (unlikely(is_swap_pmd(pmd))) {
1040 swp_entry_t entry = pmd_to_swp_entry(pmd);
1042 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1043 if (is_write_migration_entry(entry)) {
1044 make_migration_entry_read(&entry);
1045 pmd = swp_entry_to_pmd(entry);
1046 if (pmd_swp_soft_dirty(*src_pmd))
1047 pmd = pmd_swp_mksoft_dirty(pmd);
1048 if (pmd_swp_uffd_wp(*src_pmd))
1049 pmd = pmd_swp_mkuffd_wp(pmd);
1050 set_pmd_at(src_mm, addr, src_pmd, pmd);
1052 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1053 mm_inc_nr_ptes(dst_mm);
1054 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1055 if (!userfaultfd_wp(dst_vma))
1056 pmd = pmd_swp_clear_uffd_wp(pmd);
1057 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1063 if (unlikely(!pmd_trans_huge(pmd))) {
1064 pte_free(dst_mm, pgtable);
1068 * When page table lock is held, the huge zero pmd should not be
1069 * under splitting since we don't split the page itself, only pmd to
1072 if (is_huge_zero_pmd(pmd)) {
1074 * get_huge_zero_page() will never allocate a new page here,
1075 * since we already have a zero page to copy. It just takes a
1078 mm_get_huge_zero_page(dst_mm);
1082 src_page = pmd_page(pmd);
1083 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1086 * If this page is a potentially pinned page, split and retry the fault
1087 * with smaller page size. Normally this should not happen because the
1088 * userspace should use MADV_DONTFORK upon pinned regions. This is a
1089 * best effort that the pinned pages won't be replaced by another
1090 * random page during the coming copy-on-write.
1092 if (unlikely(is_cow_mapping(src_vma->vm_flags) &&
1093 atomic_read(&src_mm->has_pinned) &&
1094 page_maybe_dma_pinned(src_page))) {
1095 pte_free(dst_mm, pgtable);
1096 spin_unlock(src_ptl);
1097 spin_unlock(dst_ptl);
1098 __split_huge_pmd(src_vma, src_pmd, addr, false, NULL);
1103 page_dup_rmap(src_page, true);
1104 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1106 mm_inc_nr_ptes(dst_mm);
1107 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1108 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1109 if (!userfaultfd_wp(dst_vma))
1110 pmd = pmd_clear_uffd_wp(pmd);
1111 pmd = pmd_mkold(pmd_wrprotect(pmd));
1112 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1116 spin_unlock(src_ptl);
1117 spin_unlock(dst_ptl);
1122 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1123 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1124 pud_t *pud, int flags)
1128 _pud = pud_mkyoung(*pud);
1129 if (flags & FOLL_WRITE)
1130 _pud = pud_mkdirty(_pud);
1131 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1132 pud, _pud, flags & FOLL_WRITE))
1133 update_mmu_cache_pud(vma, addr, pud);
1136 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1137 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1139 unsigned long pfn = pud_pfn(*pud);
1140 struct mm_struct *mm = vma->vm_mm;
1143 assert_spin_locked(pud_lockptr(mm, pud));
1145 if (flags & FOLL_WRITE && !pud_write(*pud))
1148 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1149 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1150 (FOLL_PIN | FOLL_GET)))
1153 if (pud_present(*pud) && pud_devmap(*pud))
1158 if (flags & FOLL_TOUCH)
1159 touch_pud(vma, addr, pud, flags);
1162 * device mapped pages can only be returned if the
1163 * caller will manage the page reference count.
1165 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1167 if (!(flags & (FOLL_GET | FOLL_PIN)))
1168 return ERR_PTR(-EEXIST);
1170 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1171 *pgmap = get_dev_pagemap(pfn, *pgmap);
1173 return ERR_PTR(-EFAULT);
1174 page = pfn_to_page(pfn);
1175 if (!try_grab_page(page, flags))
1176 page = ERR_PTR(-ENOMEM);
1181 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1182 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1183 struct vm_area_struct *vma)
1185 spinlock_t *dst_ptl, *src_ptl;
1189 dst_ptl = pud_lock(dst_mm, dst_pud);
1190 src_ptl = pud_lockptr(src_mm, src_pud);
1191 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1195 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1199 * When page table lock is held, the huge zero pud should not be
1200 * under splitting since we don't split the page itself, only pud to
1203 if (is_huge_zero_pud(pud)) {
1204 /* No huge zero pud yet */
1207 /* Please refer to comments in copy_huge_pmd() */
1208 if (unlikely(is_cow_mapping(vma->vm_flags) &&
1209 atomic_read(&src_mm->has_pinned) &&
1210 page_maybe_dma_pinned(pud_page(pud)))) {
1211 spin_unlock(src_ptl);
1212 spin_unlock(dst_ptl);
1213 __split_huge_pud(vma, src_pud, addr);
1217 pudp_set_wrprotect(src_mm, addr, src_pud);
1218 pud = pud_mkold(pud_wrprotect(pud));
1219 set_pud_at(dst_mm, addr, dst_pud, pud);
1223 spin_unlock(src_ptl);
1224 spin_unlock(dst_ptl);
1228 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1231 unsigned long haddr;
1232 bool write = vmf->flags & FAULT_FLAG_WRITE;
1234 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1235 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1238 entry = pud_mkyoung(orig_pud);
1240 entry = pud_mkdirty(entry);
1241 haddr = vmf->address & HPAGE_PUD_MASK;
1242 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1243 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1246 spin_unlock(vmf->ptl);
1248 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1250 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1253 unsigned long haddr;
1254 bool write = vmf->flags & FAULT_FLAG_WRITE;
1256 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1257 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1260 entry = pmd_mkyoung(orig_pmd);
1262 entry = pmd_mkdirty(entry);
1263 haddr = vmf->address & HPAGE_PMD_MASK;
1264 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1265 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1268 spin_unlock(vmf->ptl);
1271 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1273 struct vm_area_struct *vma = vmf->vma;
1275 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1277 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1278 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1280 if (is_huge_zero_pmd(orig_pmd))
1283 spin_lock(vmf->ptl);
1285 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1286 spin_unlock(vmf->ptl);
1290 page = pmd_page(orig_pmd);
1291 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1293 /* Lock page for reuse_swap_page() */
1294 if (!trylock_page(page)) {
1296 spin_unlock(vmf->ptl);
1298 spin_lock(vmf->ptl);
1299 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1300 spin_unlock(vmf->ptl);
1309 * We can only reuse the page if nobody else maps the huge page or it's
1312 if (reuse_swap_page(page, NULL)) {
1314 entry = pmd_mkyoung(orig_pmd);
1315 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1316 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1317 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1319 spin_unlock(vmf->ptl);
1320 return VM_FAULT_WRITE;
1324 spin_unlock(vmf->ptl);
1326 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1327 return VM_FAULT_FALLBACK;
1331 * FOLL_FORCE can write to even unwritable pmd's, but only
1332 * after we've gone through a COW cycle and they are dirty.
1334 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1336 return pmd_write(pmd) ||
1337 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1340 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1345 struct mm_struct *mm = vma->vm_mm;
1346 struct page *page = NULL;
1348 assert_spin_locked(pmd_lockptr(mm, pmd));
1350 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1353 /* Avoid dumping huge zero page */
1354 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1355 return ERR_PTR(-EFAULT);
1357 /* Full NUMA hinting faults to serialise migration in fault paths */
1358 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1361 page = pmd_page(*pmd);
1362 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1364 if (!try_grab_page(page, flags))
1365 return ERR_PTR(-ENOMEM);
1367 if (flags & FOLL_TOUCH)
1368 touch_pmd(vma, addr, pmd, flags);
1370 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1372 * We don't mlock() pte-mapped THPs. This way we can avoid
1373 * leaking mlocked pages into non-VM_LOCKED VMAs.
1377 * In most cases the pmd is the only mapping of the page as we
1378 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1379 * writable private mappings in populate_vma_page_range().
1381 * The only scenario when we have the page shared here is if we
1382 * mlocking read-only mapping shared over fork(). We skip
1383 * mlocking such pages.
1387 * We can expect PageDoubleMap() to be stable under page lock:
1388 * for file pages we set it in page_add_file_rmap(), which
1389 * requires page to be locked.
1392 if (PageAnon(page) && compound_mapcount(page) != 1)
1394 if (PageDoubleMap(page) || !page->mapping)
1396 if (!trylock_page(page))
1398 if (page->mapping && !PageDoubleMap(page))
1399 mlock_vma_page(page);
1403 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1404 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1410 /* NUMA hinting page fault entry point for trans huge pmds */
1411 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1413 struct vm_area_struct *vma = vmf->vma;
1414 struct anon_vma *anon_vma = NULL;
1416 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1417 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1418 int target_nid, last_cpupid = -1;
1420 bool migrated = false;
1424 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1425 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1429 * If there are potential migrations, wait for completion and retry
1430 * without disrupting NUMA hinting information. Do not relock and
1431 * check_same as the page may no longer be mapped.
1433 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1434 page = pmd_page(*vmf->pmd);
1435 if (!get_page_unless_zero(page))
1437 spin_unlock(vmf->ptl);
1438 put_and_wait_on_page_locked(page);
1442 page = pmd_page(pmd);
1443 BUG_ON(is_huge_zero_page(page));
1444 page_nid = page_to_nid(page);
1445 last_cpupid = page_cpupid_last(page);
1446 count_vm_numa_event(NUMA_HINT_FAULTS);
1447 if (page_nid == this_nid) {
1448 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1449 flags |= TNF_FAULT_LOCAL;
1452 /* See similar comment in do_numa_page for explanation */
1453 if (!pmd_savedwrite(pmd))
1454 flags |= TNF_NO_GROUP;
1457 * Acquire the page lock to serialise THP migrations but avoid dropping
1458 * page_table_lock if at all possible
1460 page_locked = trylock_page(page);
1461 target_nid = mpol_misplaced(page, vma, haddr);
1462 if (target_nid == NUMA_NO_NODE) {
1463 /* If the page was locked, there are no parallel migrations */
1468 /* Migration could have started since the pmd_trans_migrating check */
1470 page_nid = NUMA_NO_NODE;
1471 if (!get_page_unless_zero(page))
1473 spin_unlock(vmf->ptl);
1474 put_and_wait_on_page_locked(page);
1479 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1480 * to serialises splits
1483 spin_unlock(vmf->ptl);
1484 anon_vma = page_lock_anon_vma_read(page);
1486 /* Confirm the PMD did not change while page_table_lock was released */
1487 spin_lock(vmf->ptl);
1488 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1491 page_nid = NUMA_NO_NODE;
1495 /* Bail if we fail to protect against THP splits for any reason */
1496 if (unlikely(!anon_vma)) {
1498 page_nid = NUMA_NO_NODE;
1503 * Since we took the NUMA fault, we must have observed the !accessible
1504 * bit. Make sure all other CPUs agree with that, to avoid them
1505 * modifying the page we're about to migrate.
1507 * Must be done under PTL such that we'll observe the relevant
1508 * inc_tlb_flush_pending().
1510 * We are not sure a pending tlb flush here is for a huge page
1511 * mapping or not. Hence use the tlb range variant
1513 if (mm_tlb_flush_pending(vma->vm_mm)) {
1514 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1516 * change_huge_pmd() released the pmd lock before
1517 * invalidating the secondary MMUs sharing the primary
1518 * MMU pagetables (with ->invalidate_range()). The
1519 * mmu_notifier_invalidate_range_end() (which
1520 * internally calls ->invalidate_range()) in
1521 * change_pmd_range() will run after us, so we can't
1522 * rely on it here and we need an explicit invalidate.
1524 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1525 haddr + HPAGE_PMD_SIZE);
1529 * Migrate the THP to the requested node, returns with page unlocked
1530 * and access rights restored.
1532 spin_unlock(vmf->ptl);
1534 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1535 vmf->pmd, pmd, vmf->address, page, target_nid);
1537 flags |= TNF_MIGRATED;
1538 page_nid = target_nid;
1540 flags |= TNF_MIGRATE_FAIL;
1544 BUG_ON(!PageLocked(page));
1545 was_writable = pmd_savedwrite(pmd);
1546 pmd = pmd_modify(pmd, vma->vm_page_prot);
1547 pmd = pmd_mkyoung(pmd);
1549 pmd = pmd_mkwrite(pmd);
1550 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1551 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1554 spin_unlock(vmf->ptl);
1558 page_unlock_anon_vma_read(anon_vma);
1560 if (page_nid != NUMA_NO_NODE)
1561 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1568 * Return true if we do MADV_FREE successfully on entire pmd page.
1569 * Otherwise, return false.
1571 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1572 pmd_t *pmd, unsigned long addr, unsigned long next)
1577 struct mm_struct *mm = tlb->mm;
1580 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1582 ptl = pmd_trans_huge_lock(pmd, vma);
1587 if (is_huge_zero_pmd(orig_pmd))
1590 if (unlikely(!pmd_present(orig_pmd))) {
1591 VM_BUG_ON(thp_migration_supported() &&
1592 !is_pmd_migration_entry(orig_pmd));
1596 page = pmd_page(orig_pmd);
1598 * If other processes are mapping this page, we couldn't discard
1599 * the page unless they all do MADV_FREE so let's skip the page.
1601 if (total_mapcount(page) != 1)
1604 if (!trylock_page(page))
1608 * If user want to discard part-pages of THP, split it so MADV_FREE
1609 * will deactivate only them.
1611 if (next - addr != HPAGE_PMD_SIZE) {
1614 split_huge_page(page);
1620 if (PageDirty(page))
1621 ClearPageDirty(page);
1624 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1625 pmdp_invalidate(vma, addr, pmd);
1626 orig_pmd = pmd_mkold(orig_pmd);
1627 orig_pmd = pmd_mkclean(orig_pmd);
1629 set_pmd_at(mm, addr, pmd, orig_pmd);
1630 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1633 mark_page_lazyfree(page);
1641 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1645 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1646 pte_free(mm, pgtable);
1650 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1651 pmd_t *pmd, unsigned long addr)
1656 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1658 ptl = __pmd_trans_huge_lock(pmd, vma);
1662 * For architectures like ppc64 we look at deposited pgtable
1663 * when calling pmdp_huge_get_and_clear. So do the
1664 * pgtable_trans_huge_withdraw after finishing pmdp related
1667 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1669 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1670 if (vma_is_special_huge(vma)) {
1671 if (arch_needs_pgtable_deposit())
1672 zap_deposited_table(tlb->mm, pmd);
1674 if (is_huge_zero_pmd(orig_pmd))
1675 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1676 } else if (is_huge_zero_pmd(orig_pmd)) {
1677 zap_deposited_table(tlb->mm, pmd);
1679 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1681 struct page *page = NULL;
1682 int flush_needed = 1;
1684 if (pmd_present(orig_pmd)) {
1685 page = pmd_page(orig_pmd);
1686 page_remove_rmap(page, true);
1687 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1688 VM_BUG_ON_PAGE(!PageHead(page), page);
1689 } else if (thp_migration_supported()) {
1692 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1693 entry = pmd_to_swp_entry(orig_pmd);
1694 page = migration_entry_to_page(entry);
1697 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1699 if (PageAnon(page)) {
1700 zap_deposited_table(tlb->mm, pmd);
1701 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1703 if (arch_needs_pgtable_deposit())
1704 zap_deposited_table(tlb->mm, pmd);
1705 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1710 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1715 #ifndef pmd_move_must_withdraw
1716 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1717 spinlock_t *old_pmd_ptl,
1718 struct vm_area_struct *vma)
1721 * With split pmd lock we also need to move preallocated
1722 * PTE page table if new_pmd is on different PMD page table.
1724 * We also don't deposit and withdraw tables for file pages.
1726 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1730 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1732 #ifdef CONFIG_MEM_SOFT_DIRTY
1733 if (unlikely(is_pmd_migration_entry(pmd)))
1734 pmd = pmd_swp_mksoft_dirty(pmd);
1735 else if (pmd_present(pmd))
1736 pmd = pmd_mksoft_dirty(pmd);
1741 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1742 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1744 spinlock_t *old_ptl, *new_ptl;
1746 struct mm_struct *mm = vma->vm_mm;
1747 bool force_flush = false;
1750 * The destination pmd shouldn't be established, free_pgtables()
1751 * should have release it.
1753 if (WARN_ON(!pmd_none(*new_pmd))) {
1754 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1759 * We don't have to worry about the ordering of src and dst
1760 * ptlocks because exclusive mmap_lock prevents deadlock.
1762 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1764 new_ptl = pmd_lockptr(mm, new_pmd);
1765 if (new_ptl != old_ptl)
1766 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1767 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1768 if (pmd_present(pmd))
1770 VM_BUG_ON(!pmd_none(*new_pmd));
1772 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1774 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1775 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1777 pmd = move_soft_dirty_pmd(pmd);
1778 set_pmd_at(mm, new_addr, new_pmd, pmd);
1780 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1781 if (new_ptl != old_ptl)
1782 spin_unlock(new_ptl);
1783 spin_unlock(old_ptl);
1791 * - 0 if PMD could not be locked
1792 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1793 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1795 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1796 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1798 struct mm_struct *mm = vma->vm_mm;
1801 bool preserve_write;
1803 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1804 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1805 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1807 ptl = __pmd_trans_huge_lock(pmd, vma);
1811 preserve_write = prot_numa && pmd_write(*pmd);
1814 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1815 if (is_swap_pmd(*pmd)) {
1816 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1818 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1819 if (is_write_migration_entry(entry)) {
1822 * A protection check is difficult so
1823 * just be safe and disable write
1825 make_migration_entry_read(&entry);
1826 newpmd = swp_entry_to_pmd(entry);
1827 if (pmd_swp_soft_dirty(*pmd))
1828 newpmd = pmd_swp_mksoft_dirty(newpmd);
1829 if (pmd_swp_uffd_wp(*pmd))
1830 newpmd = pmd_swp_mkuffd_wp(newpmd);
1831 set_pmd_at(mm, addr, pmd, newpmd);
1838 * Avoid trapping faults against the zero page. The read-only
1839 * data is likely to be read-cached on the local CPU and
1840 * local/remote hits to the zero page are not interesting.
1842 if (prot_numa && is_huge_zero_pmd(*pmd))
1845 if (prot_numa && pmd_protnone(*pmd))
1849 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1850 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1851 * which is also under mmap_read_lock(mm):
1854 * change_huge_pmd(prot_numa=1)
1855 * pmdp_huge_get_and_clear_notify()
1856 * madvise_dontneed()
1858 * pmd_trans_huge(*pmd) == 0 (without ptl)
1861 * // pmd is re-established
1863 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1864 * which may break userspace.
1866 * pmdp_invalidate() is required to make sure we don't miss
1867 * dirty/young flags set by hardware.
1869 entry = pmdp_invalidate(vma, addr, pmd);
1871 entry = pmd_modify(entry, newprot);
1873 entry = pmd_mk_savedwrite(entry);
1875 entry = pmd_wrprotect(entry);
1876 entry = pmd_mkuffd_wp(entry);
1877 } else if (uffd_wp_resolve) {
1879 * Leave the write bit to be handled by PF interrupt
1880 * handler, then things like COW could be properly
1883 entry = pmd_clear_uffd_wp(entry);
1886 set_pmd_at(mm, addr, pmd, entry);
1887 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1894 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1896 * Note that if it returns page table lock pointer, this routine returns without
1897 * unlocking page table lock. So callers must unlock it.
1899 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1902 ptl = pmd_lock(vma->vm_mm, pmd);
1903 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1911 * Returns true if a given pud maps a thp, false otherwise.
1913 * Note that if it returns true, this routine returns without unlocking page
1914 * table lock. So callers must unlock it.
1916 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1920 ptl = pud_lock(vma->vm_mm, pud);
1921 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1927 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1928 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1929 pud_t *pud, unsigned long addr)
1933 ptl = __pud_trans_huge_lock(pud, vma);
1937 * For architectures like ppc64 we look at deposited pgtable
1938 * when calling pudp_huge_get_and_clear. So do the
1939 * pgtable_trans_huge_withdraw after finishing pudp related
1942 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1943 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1944 if (vma_is_special_huge(vma)) {
1946 /* No zero page support yet */
1948 /* No support for anonymous PUD pages yet */
1954 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1955 unsigned long haddr)
1957 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1958 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1959 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1960 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1962 count_vm_event(THP_SPLIT_PUD);
1964 pudp_huge_clear_flush_notify(vma, haddr, pud);
1967 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1968 unsigned long address)
1971 struct mmu_notifier_range range;
1973 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1974 address & HPAGE_PUD_MASK,
1975 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1976 mmu_notifier_invalidate_range_start(&range);
1977 ptl = pud_lock(vma->vm_mm, pud);
1978 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1980 __split_huge_pud_locked(vma, pud, range.start);
1985 * No need to double call mmu_notifier->invalidate_range() callback as
1986 * the above pudp_huge_clear_flush_notify() did already call it.
1988 mmu_notifier_invalidate_range_only_end(&range);
1990 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1992 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1993 unsigned long haddr, pmd_t *pmd)
1995 struct mm_struct *mm = vma->vm_mm;
1997 pmd_t _pmd, old_pmd;
2001 * Leave pmd empty until pte is filled note that it is fine to delay
2002 * notification until mmu_notifier_invalidate_range_end() as we are
2003 * replacing a zero pmd write protected page with a zero pte write
2006 * See Documentation/vm/mmu_notifier.rst
2008 old_pmd = pmdp_huge_clear_flush(vma, haddr, pmd);
2010 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2011 pmd_populate(mm, &_pmd, pgtable);
2013 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2015 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2016 entry = pte_mkspecial(entry);
2017 if (pmd_uffd_wp(old_pmd))
2018 entry = pte_mkuffd_wp(entry);
2019 pte = pte_offset_map(&_pmd, haddr);
2020 VM_BUG_ON(!pte_none(*pte));
2021 set_pte_at(mm, haddr, pte, entry);
2024 smp_wmb(); /* make pte visible before pmd */
2025 pmd_populate(mm, pmd, pgtable);
2028 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2029 unsigned long haddr, bool freeze)
2031 struct mm_struct *mm = vma->vm_mm;
2034 pmd_t old_pmd, _pmd;
2035 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2039 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2040 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2041 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2042 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2043 && !pmd_devmap(*pmd));
2045 count_vm_event(THP_SPLIT_PMD);
2047 if (!vma_is_anonymous(vma)) {
2048 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2050 * We are going to unmap this huge page. So
2051 * just go ahead and zap it
2053 if (arch_needs_pgtable_deposit())
2054 zap_deposited_table(mm, pmd);
2055 if (vma_is_special_huge(vma))
2057 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2060 entry = pmd_to_swp_entry(old_pmd);
2061 page = migration_entry_to_page(entry);
2063 page = pmd_page(old_pmd);
2064 if (!PageDirty(page) && pmd_dirty(old_pmd))
2065 set_page_dirty(page);
2066 if (!PageReferenced(page) && pmd_young(old_pmd))
2067 SetPageReferenced(page);
2068 page_remove_rmap(page, true);
2071 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2075 if (is_huge_zero_pmd(*pmd)) {
2077 * FIXME: Do we want to invalidate secondary mmu by calling
2078 * mmu_notifier_invalidate_range() see comments below inside
2079 * __split_huge_pmd() ?
2081 * We are going from a zero huge page write protected to zero
2082 * small page also write protected so it does not seems useful
2083 * to invalidate secondary mmu at this time.
2085 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2089 * Up to this point the pmd is present and huge and userland has the
2090 * whole access to the hugepage during the split (which happens in
2091 * place). If we overwrite the pmd with the not-huge version pointing
2092 * to the pte here (which of course we could if all CPUs were bug
2093 * free), userland could trigger a small page size TLB miss on the
2094 * small sized TLB while the hugepage TLB entry is still established in
2095 * the huge TLB. Some CPU doesn't like that.
2096 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2097 * 383 on page 105. Intel should be safe but is also warns that it's
2098 * only safe if the permission and cache attributes of the two entries
2099 * loaded in the two TLB is identical (which should be the case here).
2100 * But it is generally safer to never allow small and huge TLB entries
2101 * for the same virtual address to be loaded simultaneously. So instead
2102 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2103 * current pmd notpresent (atomically because here the pmd_trans_huge
2104 * must remain set at all times on the pmd until the split is complete
2105 * for this pmd), then we flush the SMP TLB and finally we write the
2106 * non-huge version of the pmd entry with pmd_populate.
2108 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2110 pmd_migration = is_pmd_migration_entry(old_pmd);
2111 if (unlikely(pmd_migration)) {
2114 entry = pmd_to_swp_entry(old_pmd);
2115 page = migration_entry_to_page(entry);
2116 write = is_write_migration_entry(entry);
2118 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2119 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2121 page = pmd_page(old_pmd);
2122 if (pmd_dirty(old_pmd))
2124 write = pmd_write(old_pmd);
2125 young = pmd_young(old_pmd);
2126 soft_dirty = pmd_soft_dirty(old_pmd);
2127 uffd_wp = pmd_uffd_wp(old_pmd);
2129 VM_BUG_ON_PAGE(!page_count(page), page);
2130 page_ref_add(page, HPAGE_PMD_NR - 1);
2133 * Withdraw the table only after we mark the pmd entry invalid.
2134 * This's critical for some architectures (Power).
2136 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2137 pmd_populate(mm, &_pmd, pgtable);
2139 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2142 * Note that NUMA hinting access restrictions are not
2143 * transferred to avoid any possibility of altering
2144 * permissions across VMAs.
2146 if (freeze || pmd_migration) {
2147 swp_entry_t swp_entry;
2148 swp_entry = make_migration_entry(page + i, write);
2149 entry = swp_entry_to_pte(swp_entry);
2151 entry = pte_swp_mksoft_dirty(entry);
2153 entry = pte_swp_mkuffd_wp(entry);
2155 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2156 entry = maybe_mkwrite(entry, vma);
2158 entry = pte_wrprotect(entry);
2160 entry = pte_mkold(entry);
2162 entry = pte_mksoft_dirty(entry);
2164 entry = pte_mkuffd_wp(entry);
2166 pte = pte_offset_map(&_pmd, addr);
2167 BUG_ON(!pte_none(*pte));
2168 set_pte_at(mm, addr, pte, entry);
2170 atomic_inc(&page[i]._mapcount);
2174 if (!pmd_migration) {
2176 * Set PG_double_map before dropping compound_mapcount to avoid
2177 * false-negative page_mapped().
2179 if (compound_mapcount(page) > 1 &&
2180 !TestSetPageDoubleMap(page)) {
2181 for (i = 0; i < HPAGE_PMD_NR; i++)
2182 atomic_inc(&page[i]._mapcount);
2185 lock_page_memcg(page);
2186 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2187 /* Last compound_mapcount is gone. */
2188 __dec_lruvec_page_state(page, NR_ANON_THPS);
2189 if (TestClearPageDoubleMap(page)) {
2190 /* No need in mapcount reference anymore */
2191 for (i = 0; i < HPAGE_PMD_NR; i++)
2192 atomic_dec(&page[i]._mapcount);
2195 unlock_page_memcg(page);
2198 smp_wmb(); /* make pte visible before pmd */
2199 pmd_populate(mm, pmd, pgtable);
2202 for (i = 0; i < HPAGE_PMD_NR; i++) {
2203 page_remove_rmap(page + i, false);
2209 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2210 unsigned long address, bool freeze, struct page *page)
2213 struct mmu_notifier_range range;
2214 bool do_unlock_page = false;
2217 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2218 address & HPAGE_PMD_MASK,
2219 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2220 mmu_notifier_invalidate_range_start(&range);
2221 ptl = pmd_lock(vma->vm_mm, pmd);
2224 * If caller asks to setup a migration entries, we need a page to check
2225 * pmd against. Otherwise we can end up replacing wrong page.
2227 VM_BUG_ON(freeze && !page);
2229 VM_WARN_ON_ONCE(!PageLocked(page));
2230 if (page != pmd_page(*pmd))
2235 if (pmd_trans_huge(*pmd)) {
2237 page = pmd_page(*pmd);
2239 * An anonymous page must be locked, to ensure that a
2240 * concurrent reuse_swap_page() sees stable mapcount;
2241 * but reuse_swap_page() is not used on shmem or file,
2242 * and page lock must not be taken when zap_pmd_range()
2243 * calls __split_huge_pmd() while i_mmap_lock is held.
2245 if (PageAnon(page)) {
2246 if (unlikely(!trylock_page(page))) {
2252 if (unlikely(!pmd_same(*pmd, _pmd))) {
2260 do_unlock_page = true;
2263 if (PageMlocked(page))
2264 clear_page_mlock(page);
2265 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2267 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2273 * No need to double call mmu_notifier->invalidate_range() callback.
2274 * They are 3 cases to consider inside __split_huge_pmd_locked():
2275 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2276 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2277 * fault will trigger a flush_notify before pointing to a new page
2278 * (it is fine if the secondary mmu keeps pointing to the old zero
2279 * page in the meantime)
2280 * 3) Split a huge pmd into pte pointing to the same page. No need
2281 * to invalidate secondary tlb entry they are all still valid.
2282 * any further changes to individual pte will notify. So no need
2283 * to call mmu_notifier->invalidate_range()
2285 mmu_notifier_invalidate_range_only_end(&range);
2288 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2289 bool freeze, struct page *page)
2296 pgd = pgd_offset(vma->vm_mm, address);
2297 if (!pgd_present(*pgd))
2300 p4d = p4d_offset(pgd, address);
2301 if (!p4d_present(*p4d))
2304 pud = pud_offset(p4d, address);
2305 if (!pud_present(*pud))
2308 pmd = pmd_offset(pud, address);
2310 __split_huge_pmd(vma, pmd, address, freeze, page);
2313 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2314 unsigned long start,
2319 * If the new start address isn't hpage aligned and it could
2320 * previously contain an hugepage: check if we need to split
2323 if (start & ~HPAGE_PMD_MASK &&
2324 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2325 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2326 split_huge_pmd_address(vma, start, false, NULL);
2329 * If the new end address isn't hpage aligned and it could
2330 * previously contain an hugepage: check if we need to split
2333 if (end & ~HPAGE_PMD_MASK &&
2334 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2335 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2336 split_huge_pmd_address(vma, end, false, NULL);
2339 * If we're also updating the vma->vm_next->vm_start, if the new
2340 * vm_next->vm_start isn't hpage aligned and it could previously
2341 * contain an hugepage: check if we need to split an huge pmd.
2343 if (adjust_next > 0) {
2344 struct vm_area_struct *next = vma->vm_next;
2345 unsigned long nstart = next->vm_start;
2346 nstart += adjust_next;
2347 if (nstart & ~HPAGE_PMD_MASK &&
2348 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2349 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2350 split_huge_pmd_address(next, nstart, false, NULL);
2354 static void unmap_page(struct page *page)
2356 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_SYNC |
2357 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2359 VM_BUG_ON_PAGE(!PageHead(page), page);
2362 ttu_flags |= TTU_SPLIT_FREEZE;
2364 try_to_unmap(page, ttu_flags);
2366 VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2369 static void remap_page(struct page *page, unsigned int nr)
2372 if (PageTransHuge(page)) {
2373 remove_migration_ptes(page, page, true);
2375 for (i = 0; i < nr; i++)
2376 remove_migration_ptes(page + i, page + i, true);
2380 static void __split_huge_page_tail(struct page *head, int tail,
2381 struct lruvec *lruvec, struct list_head *list)
2383 struct page *page_tail = head + tail;
2385 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2388 * Clone page flags before unfreezing refcount.
2390 * After successful get_page_unless_zero() might follow flags change,
2391 * for exmaple lock_page() which set PG_waiters.
2393 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2394 page_tail->flags |= (head->flags &
2395 ((1L << PG_referenced) |
2396 (1L << PG_swapbacked) |
2397 (1L << PG_swapcache) |
2398 (1L << PG_mlocked) |
2399 (1L << PG_uptodate) |
2401 (1L << PG_workingset) |
2403 (1L << PG_unevictable) |
2409 /* ->mapping in first tail page is compound_mapcount */
2410 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2412 page_tail->mapping = head->mapping;
2413 page_tail->index = head->index + tail;
2415 /* Page flags must be visible before we make the page non-compound. */
2419 * Clear PageTail before unfreezing page refcount.
2421 * After successful get_page_unless_zero() might follow put_page()
2422 * which needs correct compound_head().
2424 clear_compound_head(page_tail);
2426 /* Finally unfreeze refcount. Additional reference from page cache. */
2427 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2428 PageSwapCache(head)));
2430 if (page_is_young(head))
2431 set_page_young(page_tail);
2432 if (page_is_idle(head))
2433 set_page_idle(page_tail);
2435 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2438 * always add to the tail because some iterators expect new
2439 * pages to show after the currently processed elements - e.g.
2442 lru_add_page_tail(head, page_tail, lruvec, list);
2445 static void __split_huge_page(struct page *page, struct list_head *list,
2446 pgoff_t end, unsigned long flags)
2448 struct page *head = compound_head(page);
2449 pg_data_t *pgdat = page_pgdat(head);
2450 struct lruvec *lruvec;
2451 struct address_space *swap_cache = NULL;
2452 unsigned long offset = 0;
2453 unsigned int nr = thp_nr_pages(head);
2456 lruvec = mem_cgroup_page_lruvec(head, pgdat);
2458 /* complete memcg works before add pages to LRU */
2459 split_page_memcg(head, nr);
2461 if (PageAnon(head) && PageSwapCache(head)) {
2462 swp_entry_t entry = { .val = page_private(head) };
2464 offset = swp_offset(entry);
2465 swap_cache = swap_address_space(entry);
2466 xa_lock(&swap_cache->i_pages);
2469 for (i = nr - 1; i >= 1; i--) {
2470 __split_huge_page_tail(head, i, lruvec, list);
2471 /* Some pages can be beyond i_size: drop them from page cache */
2472 if (head[i].index >= end) {
2473 ClearPageDirty(head + i);
2474 __delete_from_page_cache(head + i, NULL);
2475 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2476 shmem_uncharge(head->mapping->host, 1);
2478 } else if (!PageAnon(page)) {
2479 __xa_store(&head->mapping->i_pages, head[i].index,
2481 } else if (swap_cache) {
2482 __xa_store(&swap_cache->i_pages, offset + i,
2487 ClearPageCompound(head);
2489 split_page_owner(head, nr);
2491 /* See comment in __split_huge_page_tail() */
2492 if (PageAnon(head)) {
2493 /* Additional pin to swap cache */
2494 if (PageSwapCache(head)) {
2495 page_ref_add(head, 2);
2496 xa_unlock(&swap_cache->i_pages);
2501 /* Additional pin to page cache */
2502 page_ref_add(head, 2);
2503 xa_unlock(&head->mapping->i_pages);
2506 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2508 remap_page(head, nr);
2510 if (PageSwapCache(head)) {
2511 swp_entry_t entry = { .val = page_private(head) };
2513 split_swap_cluster(entry);
2516 for (i = 0; i < nr; i++) {
2517 struct page *subpage = head + i;
2518 if (subpage == page)
2520 unlock_page(subpage);
2523 * Subpages may be freed if there wasn't any mapping
2524 * like if add_to_swap() is running on a lru page that
2525 * had its mapping zapped. And freeing these pages
2526 * requires taking the lru_lock so we do the put_page
2527 * of the tail pages after the split is complete.
2533 int total_mapcount(struct page *page)
2535 int i, compound, nr, ret;
2537 VM_BUG_ON_PAGE(PageTail(page), page);
2539 if (likely(!PageCompound(page)))
2540 return atomic_read(&page->_mapcount) + 1;
2542 compound = compound_mapcount(page);
2543 nr = compound_nr(page);
2547 for (i = 0; i < nr; i++)
2548 ret += atomic_read(&page[i]._mapcount) + 1;
2549 /* File pages has compound_mapcount included in _mapcount */
2550 if (!PageAnon(page))
2551 return ret - compound * nr;
2552 if (PageDoubleMap(page))
2558 * This calculates accurately how many mappings a transparent hugepage
2559 * has (unlike page_mapcount() which isn't fully accurate). This full
2560 * accuracy is primarily needed to know if copy-on-write faults can
2561 * reuse the page and change the mapping to read-write instead of
2562 * copying them. At the same time this returns the total_mapcount too.
2564 * The function returns the highest mapcount any one of the subpages
2565 * has. If the return value is one, even if different processes are
2566 * mapping different subpages of the transparent hugepage, they can
2567 * all reuse it, because each process is reusing a different subpage.
2569 * The total_mapcount is instead counting all virtual mappings of the
2570 * subpages. If the total_mapcount is equal to "one", it tells the
2571 * caller all mappings belong to the same "mm" and in turn the
2572 * anon_vma of the transparent hugepage can become the vma->anon_vma
2573 * local one as no other process may be mapping any of the subpages.
2575 * It would be more accurate to replace page_mapcount() with
2576 * page_trans_huge_mapcount(), however we only use
2577 * page_trans_huge_mapcount() in the copy-on-write faults where we
2578 * need full accuracy to avoid breaking page pinning, because
2579 * page_trans_huge_mapcount() is slower than page_mapcount().
2581 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2583 int i, ret, _total_mapcount, mapcount;
2585 /* hugetlbfs shouldn't call it */
2586 VM_BUG_ON_PAGE(PageHuge(page), page);
2588 if (likely(!PageTransCompound(page))) {
2589 mapcount = atomic_read(&page->_mapcount) + 1;
2591 *total_mapcount = mapcount;
2595 page = compound_head(page);
2597 _total_mapcount = ret = 0;
2598 for (i = 0; i < thp_nr_pages(page); i++) {
2599 mapcount = atomic_read(&page[i]._mapcount) + 1;
2600 ret = max(ret, mapcount);
2601 _total_mapcount += mapcount;
2603 if (PageDoubleMap(page)) {
2605 _total_mapcount -= thp_nr_pages(page);
2607 mapcount = compound_mapcount(page);
2609 _total_mapcount += mapcount;
2611 *total_mapcount = _total_mapcount;
2615 /* Racy check whether the huge page can be split */
2616 bool can_split_huge_page(struct page *page, int *pextra_pins)
2620 /* Additional pins from page cache */
2622 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2624 extra_pins = thp_nr_pages(page);
2626 *pextra_pins = extra_pins;
2627 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2631 * This function splits huge page into normal pages. @page can point to any
2632 * subpage of huge page to split. Split doesn't change the position of @page.
2634 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2635 * The huge page must be locked.
2637 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2639 * Both head page and tail pages will inherit mapping, flags, and so on from
2642 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2643 * they are not mapped.
2645 * Returns 0 if the hugepage is split successfully.
2646 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2649 int split_huge_page_to_list(struct page *page, struct list_head *list)
2651 struct page *head = compound_head(page);
2652 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2653 struct deferred_split *ds_queue = get_deferred_split_queue(head);
2654 struct anon_vma *anon_vma = NULL;
2655 struct address_space *mapping = NULL;
2656 int extra_pins, ret;
2657 unsigned long flags;
2660 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2661 VM_BUG_ON_PAGE(!PageLocked(head), head);
2662 VM_BUG_ON_PAGE(!PageCompound(head), head);
2664 if (PageWriteback(head))
2667 if (PageAnon(head)) {
2669 * The caller does not necessarily hold an mmap_lock that would
2670 * prevent the anon_vma disappearing so we first we take a
2671 * reference to it and then lock the anon_vma for write. This
2672 * is similar to page_lock_anon_vma_read except the write lock
2673 * is taken to serialise against parallel split or collapse
2676 anon_vma = page_get_anon_vma(head);
2683 anon_vma_lock_write(anon_vma);
2685 mapping = head->mapping;
2694 i_mmap_lock_read(mapping);
2697 *__split_huge_page() may need to trim off pages beyond EOF:
2698 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2699 * which cannot be nested inside the page tree lock. So note
2700 * end now: i_size itself may be changed at any moment, but
2701 * head page lock is good enough to serialize the trimming.
2703 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2707 * Racy check if we can split the page, before unmap_page() will
2710 if (!can_split_huge_page(head, &extra_pins)) {
2717 /* prevent PageLRU to go away from under us, and freeze lru stats */
2718 spin_lock_irqsave(&pgdata->lru_lock, flags);
2721 XA_STATE(xas, &mapping->i_pages, page_index(head));
2724 * Check if the head page is present in page cache.
2725 * We assume all tail are present too, if head is there.
2727 xa_lock(&mapping->i_pages);
2728 if (xas_load(&xas) != head)
2732 /* Prevent deferred_split_scan() touching ->_refcount */
2733 spin_lock(&ds_queue->split_queue_lock);
2734 if (page_ref_freeze(head, 1 + extra_pins)) {
2735 if (!list_empty(page_deferred_list(head))) {
2736 ds_queue->split_queue_len--;
2737 list_del(page_deferred_list(head));
2739 spin_unlock(&ds_queue->split_queue_lock);
2741 if (PageSwapBacked(head))
2742 __dec_node_page_state(head, NR_SHMEM_THPS);
2744 __dec_node_page_state(head, NR_FILE_THPS);
2747 __split_huge_page(page, list, end, flags);
2750 spin_unlock(&ds_queue->split_queue_lock);
2753 xa_unlock(&mapping->i_pages);
2754 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2755 remap_page(head, thp_nr_pages(head));
2761 anon_vma_unlock_write(anon_vma);
2762 put_anon_vma(anon_vma);
2765 i_mmap_unlock_read(mapping);
2767 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2771 void free_transhuge_page(struct page *page)
2773 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2774 unsigned long flags;
2776 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2777 if (!list_empty(page_deferred_list(page))) {
2778 ds_queue->split_queue_len--;
2779 list_del(page_deferred_list(page));
2781 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2782 free_compound_page(page);
2785 void deferred_split_huge_page(struct page *page)
2787 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2789 struct mem_cgroup *memcg = compound_head(page)->mem_cgroup;
2791 unsigned long flags;
2793 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2796 * The try_to_unmap() in page reclaim path might reach here too,
2797 * this may cause a race condition to corrupt deferred split queue.
2798 * And, if page reclaim is already handling the same page, it is
2799 * unnecessary to handle it again in shrinker.
2801 * Check PageSwapCache to determine if the page is being
2802 * handled by page reclaim since THP swap would add the page into
2803 * swap cache before calling try_to_unmap().
2805 if (PageSwapCache(page))
2808 if (!list_empty(page_deferred_list(page)))
2811 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2812 if (list_empty(page_deferred_list(page))) {
2813 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2814 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2815 ds_queue->split_queue_len++;
2818 memcg_set_shrinker_bit(memcg, page_to_nid(page),
2819 deferred_split_shrinker.id);
2822 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2825 static unsigned long deferred_split_count(struct shrinker *shrink,
2826 struct shrink_control *sc)
2828 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2829 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2833 ds_queue = &sc->memcg->deferred_split_queue;
2835 return READ_ONCE(ds_queue->split_queue_len);
2838 static unsigned long deferred_split_scan(struct shrinker *shrink,
2839 struct shrink_control *sc)
2841 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2842 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2843 unsigned long flags;
2844 LIST_HEAD(list), *pos, *next;
2850 ds_queue = &sc->memcg->deferred_split_queue;
2853 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2854 /* Take pin on all head pages to avoid freeing them under us */
2855 list_for_each_safe(pos, next, &ds_queue->split_queue) {
2856 page = list_entry((void *)pos, struct page, mapping);
2857 page = compound_head(page);
2858 if (get_page_unless_zero(page)) {
2859 list_move(page_deferred_list(page), &list);
2861 /* We lost race with put_compound_page() */
2862 list_del_init(page_deferred_list(page));
2863 ds_queue->split_queue_len--;
2865 if (!--sc->nr_to_scan)
2868 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2870 list_for_each_safe(pos, next, &list) {
2871 page = list_entry((void *)pos, struct page, mapping);
2872 if (!trylock_page(page))
2874 /* split_huge_page() removes page from list on success */
2875 if (!split_huge_page(page))
2882 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2883 list_splice_tail(&list, &ds_queue->split_queue);
2884 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2887 * Stop shrinker if we didn't split any page, but the queue is empty.
2888 * This can happen if pages were freed under us.
2890 if (!split && list_empty(&ds_queue->split_queue))
2895 static struct shrinker deferred_split_shrinker = {
2896 .count_objects = deferred_split_count,
2897 .scan_objects = deferred_split_scan,
2898 .seeks = DEFAULT_SEEKS,
2899 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2903 #ifdef CONFIG_DEBUG_FS
2904 static int split_huge_pages_set(void *data, u64 val)
2908 unsigned long pfn, max_zone_pfn;
2909 unsigned long total = 0, split = 0;
2914 for_each_populated_zone(zone) {
2915 max_zone_pfn = zone_end_pfn(zone);
2916 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2917 if (!pfn_valid(pfn))
2920 page = pfn_to_page(pfn);
2921 if (!get_page_unless_zero(page))
2924 if (zone != page_zone(page))
2927 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2932 if (!split_huge_page(page))
2940 pr_info("%lu of %lu THP split\n", split, total);
2944 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2947 static int __init split_huge_pages_debugfs(void)
2949 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2950 &split_huge_pages_fops);
2953 late_initcall(split_huge_pages_debugfs);
2956 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2957 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2960 struct vm_area_struct *vma = pvmw->vma;
2961 struct mm_struct *mm = vma->vm_mm;
2962 unsigned long address = pvmw->address;
2967 if (!(pvmw->pmd && !pvmw->pte))
2970 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2971 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
2972 if (pmd_dirty(pmdval))
2973 set_page_dirty(page);
2974 entry = make_migration_entry(page, pmd_write(pmdval));
2975 pmdswp = swp_entry_to_pmd(entry);
2976 if (pmd_soft_dirty(pmdval))
2977 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2978 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2979 page_remove_rmap(page, true);
2983 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2985 struct vm_area_struct *vma = pvmw->vma;
2986 struct mm_struct *mm = vma->vm_mm;
2987 unsigned long address = pvmw->address;
2988 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2992 if (!(pvmw->pmd && !pvmw->pte))
2995 entry = pmd_to_swp_entry(*pvmw->pmd);
2997 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2998 if (pmd_swp_soft_dirty(*pvmw->pmd))
2999 pmde = pmd_mksoft_dirty(pmde);
3000 if (is_write_migration_entry(entry))
3001 pmde = maybe_pmd_mkwrite(pmde, vma);
3002 if (pmd_swp_uffd_wp(*pvmw->pmd))
3003 pmde = pmd_wrprotect(pmd_mkuffd_wp(pmde));
3005 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3007 page_add_anon_rmap(new, vma, mmun_start, true);
3009 page_add_file_rmap(new, true);
3010 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3011 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3012 mlock_vma_page(new);
3013 update_mmu_cache_pmd(vma, address, pvmw->pmd);