2 * Copyright (C) 2009 Red Hat, Inc.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/sched.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/page_owner.h>
36 #include <asm/pgalloc.h>
40 * By default transparent hugepage support is disabled in order that avoid
41 * to risk increase the memory footprint of applications without a guaranteed
42 * benefit. When transparent hugepage support is enabled, is for all mappings,
43 * and khugepaged scans all mappings.
44 * Defrag is invoked by khugepaged hugepage allocations and by page faults
45 * for all hugepage allocations.
47 unsigned long transparent_hugepage_flags __read_mostly =
48 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
49 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
52 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
54 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
55 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
56 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
58 static struct shrinker deferred_split_shrinker;
60 static atomic_t huge_zero_refcount;
61 struct page *huge_zero_page __read_mostly;
63 static struct page *get_huge_zero_page(void)
65 struct page *zero_page;
67 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
68 return READ_ONCE(huge_zero_page);
70 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
73 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
76 count_vm_event(THP_ZERO_PAGE_ALLOC);
78 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
80 __free_pages(zero_page, compound_order(zero_page));
84 /* We take additional reference here. It will be put back by shrinker */
85 atomic_set(&huge_zero_refcount, 2);
87 return READ_ONCE(huge_zero_page);
90 static void put_huge_zero_page(void)
93 * Counter should never go to zero here. Only shrinker can put
96 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
99 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
101 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
102 return READ_ONCE(huge_zero_page);
104 if (!get_huge_zero_page())
107 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
108 put_huge_zero_page();
110 return READ_ONCE(huge_zero_page);
113 void mm_put_huge_zero_page(struct mm_struct *mm)
115 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
116 put_huge_zero_page();
119 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
120 struct shrink_control *sc)
122 /* we can free zero page only if last reference remains */
123 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
126 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
127 struct shrink_control *sc)
129 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
130 struct page *zero_page = xchg(&huge_zero_page, NULL);
131 BUG_ON(zero_page == NULL);
132 __free_pages(zero_page, compound_order(zero_page));
139 static struct shrinker huge_zero_page_shrinker = {
140 .count_objects = shrink_huge_zero_page_count,
141 .scan_objects = shrink_huge_zero_page_scan,
142 .seeks = DEFAULT_SEEKS,
147 static ssize_t triple_flag_store(struct kobject *kobj,
148 struct kobj_attribute *attr,
149 const char *buf, size_t count,
150 enum transparent_hugepage_flag enabled,
151 enum transparent_hugepage_flag deferred,
152 enum transparent_hugepage_flag req_madv)
154 if (!memcmp("defer", buf,
155 min(sizeof("defer")-1, count))) {
156 if (enabled == deferred)
158 clear_bit(enabled, &transparent_hugepage_flags);
159 clear_bit(req_madv, &transparent_hugepage_flags);
160 set_bit(deferred, &transparent_hugepage_flags);
161 } else if (!memcmp("always", buf,
162 min(sizeof("always")-1, count))) {
163 clear_bit(deferred, &transparent_hugepage_flags);
164 clear_bit(req_madv, &transparent_hugepage_flags);
165 set_bit(enabled, &transparent_hugepage_flags);
166 } else if (!memcmp("madvise", buf,
167 min(sizeof("madvise")-1, count))) {
168 clear_bit(enabled, &transparent_hugepage_flags);
169 clear_bit(deferred, &transparent_hugepage_flags);
170 set_bit(req_madv, &transparent_hugepage_flags);
171 } else if (!memcmp("never", buf,
172 min(sizeof("never")-1, count))) {
173 clear_bit(enabled, &transparent_hugepage_flags);
174 clear_bit(req_madv, &transparent_hugepage_flags);
175 clear_bit(deferred, &transparent_hugepage_flags);
182 static ssize_t enabled_show(struct kobject *kobj,
183 struct kobj_attribute *attr, char *buf)
185 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
186 return sprintf(buf, "[always] madvise never\n");
187 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
188 return sprintf(buf, "always [madvise] never\n");
190 return sprintf(buf, "always madvise [never]\n");
193 static ssize_t enabled_store(struct kobject *kobj,
194 struct kobj_attribute *attr,
195 const char *buf, size_t count)
199 ret = triple_flag_store(kobj, attr, buf, count,
200 TRANSPARENT_HUGEPAGE_FLAG,
201 TRANSPARENT_HUGEPAGE_FLAG,
202 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
205 int err = start_stop_khugepaged();
212 static struct kobj_attribute enabled_attr =
213 __ATTR(enabled, 0644, enabled_show, enabled_store);
215 ssize_t single_hugepage_flag_show(struct kobject *kobj,
216 struct kobj_attribute *attr, char *buf,
217 enum transparent_hugepage_flag flag)
219 return sprintf(buf, "%d\n",
220 !!test_bit(flag, &transparent_hugepage_flags));
223 ssize_t single_hugepage_flag_store(struct kobject *kobj,
224 struct kobj_attribute *attr,
225 const char *buf, size_t count,
226 enum transparent_hugepage_flag flag)
231 ret = kstrtoul(buf, 10, &value);
238 set_bit(flag, &transparent_hugepage_flags);
240 clear_bit(flag, &transparent_hugepage_flags);
246 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
247 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
248 * memory just to allocate one more hugepage.
250 static ssize_t defrag_show(struct kobject *kobj,
251 struct kobj_attribute *attr, char *buf)
253 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
254 return sprintf(buf, "[always] defer madvise never\n");
255 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
256 return sprintf(buf, "always [defer] madvise never\n");
257 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
258 return sprintf(buf, "always defer [madvise] never\n");
260 return sprintf(buf, "always defer madvise [never]\n");
263 static ssize_t defrag_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
267 return triple_flag_store(kobj, attr, buf, count,
268 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
269 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
270 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
272 static struct kobj_attribute defrag_attr =
273 __ATTR(defrag, 0644, defrag_show, defrag_store);
275 static ssize_t use_zero_page_show(struct kobject *kobj,
276 struct kobj_attribute *attr, char *buf)
278 return single_hugepage_flag_show(kobj, attr, buf,
279 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
281 static ssize_t use_zero_page_store(struct kobject *kobj,
282 struct kobj_attribute *attr, const char *buf, size_t count)
284 return single_hugepage_flag_store(kobj, attr, buf, count,
285 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
287 static struct kobj_attribute use_zero_page_attr =
288 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
289 #ifdef CONFIG_DEBUG_VM
290 static ssize_t debug_cow_show(struct kobject *kobj,
291 struct kobj_attribute *attr, char *buf)
293 return single_hugepage_flag_show(kobj, attr, buf,
294 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
296 static ssize_t debug_cow_store(struct kobject *kobj,
297 struct kobj_attribute *attr,
298 const char *buf, size_t count)
300 return single_hugepage_flag_store(kobj, attr, buf, count,
301 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303 static struct kobj_attribute debug_cow_attr =
304 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
305 #endif /* CONFIG_DEBUG_VM */
307 static struct attribute *hugepage_attr[] = {
310 &use_zero_page_attr.attr,
311 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
312 &shmem_enabled_attr.attr,
314 #ifdef CONFIG_DEBUG_VM
315 &debug_cow_attr.attr,
320 static struct attribute_group hugepage_attr_group = {
321 .attrs = hugepage_attr,
324 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
328 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
329 if (unlikely(!*hugepage_kobj)) {
330 pr_err("failed to create transparent hugepage kobject\n");
334 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
336 pr_err("failed to register transparent hugepage group\n");
340 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
342 pr_err("failed to register transparent hugepage group\n");
343 goto remove_hp_group;
349 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
351 kobject_put(*hugepage_kobj);
355 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
357 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
358 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
359 kobject_put(hugepage_kobj);
362 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
367 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
370 #endif /* CONFIG_SYSFS */
372 static int __init hugepage_init(void)
375 struct kobject *hugepage_kobj;
377 if (!has_transparent_hugepage()) {
378 transparent_hugepage_flags = 0;
383 * hugepages can't be allocated by the buddy allocator
385 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
387 * we use page->mapping and page->index in second tail page
388 * as list_head: assuming THP order >= 2
390 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
392 err = hugepage_init_sysfs(&hugepage_kobj);
396 err = khugepaged_init();
400 err = register_shrinker(&huge_zero_page_shrinker);
402 goto err_hzp_shrinker;
403 err = register_shrinker(&deferred_split_shrinker);
405 goto err_split_shrinker;
408 * By default disable transparent hugepages on smaller systems,
409 * where the extra memory used could hurt more than TLB overhead
410 * is likely to save. The admin can still enable it through /sys.
412 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
413 transparent_hugepage_flags = 0;
417 err = start_stop_khugepaged();
423 unregister_shrinker(&deferred_split_shrinker);
425 unregister_shrinker(&huge_zero_page_shrinker);
427 khugepaged_destroy();
429 hugepage_exit_sysfs(hugepage_kobj);
433 subsys_initcall(hugepage_init);
435 static int __init setup_transparent_hugepage(char *str)
440 if (!strcmp(str, "always")) {
441 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
442 &transparent_hugepage_flags);
443 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
444 &transparent_hugepage_flags);
446 } else if (!strcmp(str, "madvise")) {
447 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
448 &transparent_hugepage_flags);
449 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
450 &transparent_hugepage_flags);
452 } else if (!strcmp(str, "never")) {
453 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
454 &transparent_hugepage_flags);
455 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
456 &transparent_hugepage_flags);
461 pr_warn("transparent_hugepage= cannot parse, ignored\n");
464 __setup("transparent_hugepage=", setup_transparent_hugepage);
466 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
468 if (likely(vma->vm_flags & VM_WRITE))
469 pmd = pmd_mkwrite(pmd);
473 static inline struct list_head *page_deferred_list(struct page *page)
476 * ->lru in the tail pages is occupied by compound_head.
477 * Let's use ->mapping + ->index in the second tail page as list_head.
479 return (struct list_head *)&page[2].mapping;
482 void prep_transhuge_page(struct page *page)
485 * we use page->mapping and page->indexlru in second tail page
486 * as list_head: assuming THP order >= 2
489 INIT_LIST_HEAD(page_deferred_list(page));
490 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
493 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
494 loff_t off, unsigned long flags, unsigned long size)
497 loff_t off_end = off + len;
498 loff_t off_align = round_up(off, size);
499 unsigned long len_pad;
501 if (off_end <= off_align || (off_end - off_align) < size)
504 len_pad = len + size;
505 if (len_pad < len || (off + len_pad) < off)
508 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
509 off >> PAGE_SHIFT, flags);
510 if (IS_ERR_VALUE(addr))
513 addr += (off - addr) & (size - 1);
517 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
518 unsigned long len, unsigned long pgoff, unsigned long flags)
520 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
524 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
527 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
532 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
534 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
536 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
539 struct vm_area_struct *vma = fe->vma;
540 struct mem_cgroup *memcg;
542 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
544 VM_BUG_ON_PAGE(!PageCompound(page), page);
546 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg,
549 count_vm_event(THP_FAULT_FALLBACK);
550 return VM_FAULT_FALLBACK;
553 pgtable = pte_alloc_one(vma->vm_mm, haddr);
554 if (unlikely(!pgtable)) {
555 mem_cgroup_cancel_charge(page, memcg, true);
560 clear_huge_page(page, haddr, HPAGE_PMD_NR);
562 * The memory barrier inside __SetPageUptodate makes sure that
563 * clear_huge_page writes become visible before the set_pmd_at()
566 __SetPageUptodate(page);
568 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
569 if (unlikely(!pmd_none(*fe->pmd))) {
570 spin_unlock(fe->ptl);
571 mem_cgroup_cancel_charge(page, memcg, true);
573 pte_free(vma->vm_mm, pgtable);
577 /* Deliver the page fault to userland */
578 if (userfaultfd_missing(vma)) {
581 spin_unlock(fe->ptl);
582 mem_cgroup_cancel_charge(page, memcg, true);
584 pte_free(vma->vm_mm, pgtable);
585 ret = handle_userfault(fe, VM_UFFD_MISSING);
586 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
590 entry = mk_huge_pmd(page, vma->vm_page_prot);
591 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
592 page_add_new_anon_rmap(page, vma, haddr, true);
593 mem_cgroup_commit_charge(page, memcg, false, true);
594 lru_cache_add_active_or_unevictable(page, vma);
595 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
596 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
597 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
598 atomic_long_inc(&vma->vm_mm->nr_ptes);
599 spin_unlock(fe->ptl);
600 count_vm_event(THP_FAULT_ALLOC);
607 * If THP defrag is set to always then directly reclaim/compact as necessary
608 * If set to defer then do only background reclaim/compact and defer to khugepaged
609 * If set to madvise and the VMA is flagged then directly reclaim/compact
610 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
612 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
614 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
616 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
617 &transparent_hugepage_flags) && vma_madvised)
618 return GFP_TRANSHUGE;
619 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
620 &transparent_hugepage_flags))
621 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
622 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
623 &transparent_hugepage_flags))
624 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
626 return GFP_TRANSHUGE_LIGHT;
629 /* Caller must hold page table lock. */
630 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
631 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
632 struct page *zero_page)
637 entry = mk_pmd(zero_page, vma->vm_page_prot);
638 entry = pmd_mkhuge(entry);
640 pgtable_trans_huge_deposit(mm, pmd, pgtable);
641 set_pmd_at(mm, haddr, pmd, entry);
642 atomic_long_inc(&mm->nr_ptes);
646 int do_huge_pmd_anonymous_page(struct fault_env *fe)
648 struct vm_area_struct *vma = fe->vma;
651 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
653 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
654 return VM_FAULT_FALLBACK;
655 if (unlikely(anon_vma_prepare(vma)))
657 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
659 if (!(fe->flags & FAULT_FLAG_WRITE) &&
660 !mm_forbids_zeropage(vma->vm_mm) &&
661 transparent_hugepage_use_zero_page()) {
663 struct page *zero_page;
665 pgtable = pte_alloc_one(vma->vm_mm, haddr);
666 if (unlikely(!pgtable))
668 zero_page = mm_get_huge_zero_page(vma->vm_mm);
669 if (unlikely(!zero_page)) {
670 pte_free(vma->vm_mm, pgtable);
671 count_vm_event(THP_FAULT_FALLBACK);
672 return VM_FAULT_FALLBACK;
674 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
676 if (pmd_none(*fe->pmd)) {
677 if (userfaultfd_missing(vma)) {
678 spin_unlock(fe->ptl);
679 pte_free(vma->vm_mm, pgtable);
680 ret = handle_userfault(fe, VM_UFFD_MISSING);
681 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
683 set_huge_zero_page(pgtable, vma->vm_mm, vma,
684 haddr, fe->pmd, zero_page);
685 spin_unlock(fe->ptl);
688 spin_unlock(fe->ptl);
689 pte_free(vma->vm_mm, pgtable);
693 gfp = alloc_hugepage_direct_gfpmask(vma);
694 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
695 if (unlikely(!page)) {
696 count_vm_event(THP_FAULT_FALLBACK);
697 return VM_FAULT_FALLBACK;
699 prep_transhuge_page(page);
700 return __do_huge_pmd_anonymous_page(fe, page, gfp);
703 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
704 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
706 struct mm_struct *mm = vma->vm_mm;
710 ptl = pmd_lock(mm, pmd);
711 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
712 if (pfn_t_devmap(pfn))
713 entry = pmd_mkdevmap(entry);
715 entry = pmd_mkyoung(pmd_mkdirty(entry));
716 entry = maybe_pmd_mkwrite(entry, vma);
718 set_pmd_at(mm, addr, pmd, entry);
719 update_mmu_cache_pmd(vma, addr, pmd);
723 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
724 pmd_t *pmd, pfn_t pfn, bool write)
726 pgprot_t pgprot = vma->vm_page_prot;
728 * If we had pmd_special, we could avoid all these restrictions,
729 * but we need to be consistent with PTEs and architectures that
730 * can't support a 'special' bit.
732 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
733 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
734 (VM_PFNMAP|VM_MIXEDMAP));
735 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
736 BUG_ON(!pfn_t_devmap(pfn));
738 if (addr < vma->vm_start || addr >= vma->vm_end)
739 return VM_FAULT_SIGBUS;
740 if (track_pfn_insert(vma, &pgprot, pfn))
741 return VM_FAULT_SIGBUS;
742 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
743 return VM_FAULT_NOPAGE;
745 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
747 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
748 pmd_t *pmd, int flags)
752 _pmd = pmd_mkyoung(*pmd);
753 if (flags & FOLL_WRITE)
754 _pmd = pmd_mkdirty(_pmd);
755 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
756 pmd, _pmd, flags & FOLL_WRITE))
757 update_mmu_cache_pmd(vma, addr, pmd);
760 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
761 pmd_t *pmd, int flags)
763 unsigned long pfn = pmd_pfn(*pmd);
764 struct mm_struct *mm = vma->vm_mm;
765 struct dev_pagemap *pgmap;
768 assert_spin_locked(pmd_lockptr(mm, pmd));
771 * When we COW a devmap PMD entry, we split it into PTEs, so we should
772 * not be in this function with `flags & FOLL_COW` set.
774 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
776 if (flags & FOLL_WRITE && !pmd_write(*pmd))
779 if (pmd_present(*pmd) && pmd_devmap(*pmd))
784 if (flags & FOLL_TOUCH)
785 touch_pmd(vma, addr, pmd, flags);
788 * device mapped pages can only be returned if the
789 * caller will manage the page reference count.
791 if (!(flags & FOLL_GET))
792 return ERR_PTR(-EEXIST);
794 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
795 pgmap = get_dev_pagemap(pfn, NULL);
797 return ERR_PTR(-EFAULT);
798 page = pfn_to_page(pfn);
800 put_dev_pagemap(pgmap);
805 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
806 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
807 struct vm_area_struct *vma)
809 spinlock_t *dst_ptl, *src_ptl;
810 struct page *src_page;
812 pgtable_t pgtable = NULL;
815 /* Skip if can be re-fill on fault */
816 if (!vma_is_anonymous(vma))
819 pgtable = pte_alloc_one(dst_mm, addr);
820 if (unlikely(!pgtable))
823 dst_ptl = pmd_lock(dst_mm, dst_pmd);
824 src_ptl = pmd_lockptr(src_mm, src_pmd);
825 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
829 if (unlikely(!pmd_trans_huge(pmd))) {
830 pte_free(dst_mm, pgtable);
834 * When page table lock is held, the huge zero pmd should not be
835 * under splitting since we don't split the page itself, only pmd to
838 if (is_huge_zero_pmd(pmd)) {
839 struct page *zero_page;
841 * get_huge_zero_page() will never allocate a new page here,
842 * since we already have a zero page to copy. It just takes a
845 zero_page = mm_get_huge_zero_page(dst_mm);
846 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
852 src_page = pmd_page(pmd);
853 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
855 page_dup_rmap(src_page, true);
856 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
857 atomic_long_inc(&dst_mm->nr_ptes);
858 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
860 pmdp_set_wrprotect(src_mm, addr, src_pmd);
861 pmd = pmd_mkold(pmd_wrprotect(pmd));
862 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
866 spin_unlock(src_ptl);
867 spin_unlock(dst_ptl);
872 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
876 bool write = fe->flags & FAULT_FLAG_WRITE;
878 fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
879 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
882 entry = pmd_mkyoung(orig_pmd);
884 entry = pmd_mkdirty(entry);
885 haddr = fe->address & HPAGE_PMD_MASK;
886 if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, write))
887 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
890 spin_unlock(fe->ptl);
893 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
896 struct vm_area_struct *vma = fe->vma;
897 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
898 struct mem_cgroup *memcg;
903 unsigned long mmun_start; /* For mmu_notifiers */
904 unsigned long mmun_end; /* For mmu_notifiers */
906 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
908 if (unlikely(!pages)) {
913 for (i = 0; i < HPAGE_PMD_NR; i++) {
914 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
915 __GFP_OTHER_NODE, vma,
916 fe->address, page_to_nid(page));
917 if (unlikely(!pages[i] ||
918 mem_cgroup_try_charge(pages[i], vma->vm_mm,
919 GFP_KERNEL, &memcg, false))) {
923 memcg = (void *)page_private(pages[i]);
924 set_page_private(pages[i], 0);
925 mem_cgroup_cancel_charge(pages[i], memcg,
933 set_page_private(pages[i], (unsigned long)memcg);
936 for (i = 0; i < HPAGE_PMD_NR; i++) {
937 copy_user_highpage(pages[i], page + i,
938 haddr + PAGE_SIZE * i, vma);
939 __SetPageUptodate(pages[i]);
944 mmun_end = haddr + HPAGE_PMD_SIZE;
945 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
947 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
948 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
950 VM_BUG_ON_PAGE(!PageHead(page), page);
952 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
953 /* leave pmd empty until pte is filled */
955 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
956 pmd_populate(vma->vm_mm, &_pmd, pgtable);
958 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
960 entry = mk_pte(pages[i], vma->vm_page_prot);
961 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
962 memcg = (void *)page_private(pages[i]);
963 set_page_private(pages[i], 0);
964 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
965 mem_cgroup_commit_charge(pages[i], memcg, false, false);
966 lru_cache_add_active_or_unevictable(pages[i], vma);
967 fe->pte = pte_offset_map(&_pmd, haddr);
968 VM_BUG_ON(!pte_none(*fe->pte));
969 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
974 smp_wmb(); /* make pte visible before pmd */
975 pmd_populate(vma->vm_mm, fe->pmd, pgtable);
976 page_remove_rmap(page, true);
977 spin_unlock(fe->ptl);
979 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
981 ret |= VM_FAULT_WRITE;
988 spin_unlock(fe->ptl);
989 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
990 for (i = 0; i < HPAGE_PMD_NR; i++) {
991 memcg = (void *)page_private(pages[i]);
992 set_page_private(pages[i], 0);
993 mem_cgroup_cancel_charge(pages[i], memcg, false);
1000 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
1002 struct vm_area_struct *vma = fe->vma;
1003 struct page *page = NULL, *new_page;
1004 struct mem_cgroup *memcg;
1005 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1006 unsigned long mmun_start; /* For mmu_notifiers */
1007 unsigned long mmun_end; /* For mmu_notifiers */
1008 gfp_t huge_gfp; /* for allocation and charge */
1011 fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1012 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1013 if (is_huge_zero_pmd(orig_pmd))
1016 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1019 page = pmd_page(orig_pmd);
1020 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1022 * We can only reuse the page if nobody else maps the huge page or it's
1025 if (!trylock_page(page)) {
1027 spin_unlock(fe->ptl);
1030 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1038 if (page_trans_huge_mapcount(page, NULL) == 1) {
1040 entry = pmd_mkyoung(orig_pmd);
1041 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1042 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1))
1043 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1044 ret |= VM_FAULT_WRITE;
1050 spin_unlock(fe->ptl);
1052 if (transparent_hugepage_enabled(vma) &&
1053 !transparent_hugepage_debug_cow()) {
1054 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1055 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1059 if (likely(new_page)) {
1060 prep_transhuge_page(new_page);
1063 split_huge_pmd(vma, fe->pmd, fe->address);
1064 ret |= VM_FAULT_FALLBACK;
1066 ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1067 if (ret & VM_FAULT_OOM) {
1068 split_huge_pmd(vma, fe->pmd, fe->address);
1069 ret |= VM_FAULT_FALLBACK;
1073 count_vm_event(THP_FAULT_FALLBACK);
1077 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1078 huge_gfp | __GFP_NORETRY, &memcg, true))) {
1080 split_huge_pmd(vma, fe->pmd, fe->address);
1083 ret |= VM_FAULT_FALLBACK;
1084 count_vm_event(THP_FAULT_FALLBACK);
1088 count_vm_event(THP_FAULT_ALLOC);
1091 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1093 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1094 __SetPageUptodate(new_page);
1097 mmun_end = haddr + HPAGE_PMD_SIZE;
1098 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1103 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1104 spin_unlock(fe->ptl);
1105 mem_cgroup_cancel_charge(new_page, memcg, true);
1110 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1111 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1112 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1113 page_add_new_anon_rmap(new_page, vma, haddr, true);
1114 mem_cgroup_commit_charge(new_page, memcg, false, true);
1115 lru_cache_add_active_or_unevictable(new_page, vma);
1116 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1117 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1119 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1121 VM_BUG_ON_PAGE(!PageHead(page), page);
1122 page_remove_rmap(page, true);
1125 ret |= VM_FAULT_WRITE;
1127 spin_unlock(fe->ptl);
1129 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1133 spin_unlock(fe->ptl);
1138 * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
1139 * but only after we've gone through a COW cycle and they are dirty.
1141 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1143 return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
1146 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1151 struct mm_struct *mm = vma->vm_mm;
1152 struct page *page = NULL;
1154 assert_spin_locked(pmd_lockptr(mm, pmd));
1156 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1159 /* Avoid dumping huge zero page */
1160 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1161 return ERR_PTR(-EFAULT);
1163 /* Full NUMA hinting faults to serialise migration in fault paths */
1164 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1167 page = pmd_page(*pmd);
1168 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1169 if (flags & FOLL_TOUCH)
1170 touch_pmd(vma, addr, pmd, flags);
1171 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1173 * We don't mlock() pte-mapped THPs. This way we can avoid
1174 * leaking mlocked pages into non-VM_LOCKED VMAs.
1178 * In most cases the pmd is the only mapping of the page as we
1179 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1180 * writable private mappings in populate_vma_page_range().
1182 * The only scenario when we have the page shared here is if we
1183 * mlocking read-only mapping shared over fork(). We skip
1184 * mlocking such pages.
1188 * We can expect PageDoubleMap() to be stable under page lock:
1189 * for file pages we set it in page_add_file_rmap(), which
1190 * requires page to be locked.
1193 if (PageAnon(page) && compound_mapcount(page) != 1)
1195 if (PageDoubleMap(page) || !page->mapping)
1197 if (!trylock_page(page))
1200 if (page->mapping && !PageDoubleMap(page))
1201 mlock_vma_page(page);
1205 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1206 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1207 if (flags & FOLL_GET)
1214 /* NUMA hinting page fault entry point for trans huge pmds */
1215 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1217 struct vm_area_struct *vma = fe->vma;
1218 struct anon_vma *anon_vma = NULL;
1220 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1221 int page_nid = -1, this_nid = numa_node_id();
1222 int target_nid, last_cpupid = -1;
1224 bool migrated = false;
1228 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1229 if (unlikely(!pmd_same(pmd, *fe->pmd)))
1233 * If there are potential migrations, wait for completion and retry
1234 * without disrupting NUMA hinting information. Do not relock and
1235 * check_same as the page may no longer be mapped.
1237 if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1238 page = pmd_page(*fe->pmd);
1239 if (!get_page_unless_zero(page))
1241 spin_unlock(fe->ptl);
1242 wait_on_page_locked(page);
1247 page = pmd_page(pmd);
1248 BUG_ON(is_huge_zero_page(page));
1249 page_nid = page_to_nid(page);
1250 last_cpupid = page_cpupid_last(page);
1251 count_vm_numa_event(NUMA_HINT_FAULTS);
1252 if (page_nid == this_nid) {
1253 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1254 flags |= TNF_FAULT_LOCAL;
1257 /* See similar comment in do_numa_page for explanation */
1258 if (!pmd_write(pmd))
1259 flags |= TNF_NO_GROUP;
1262 * Acquire the page lock to serialise THP migrations but avoid dropping
1263 * page_table_lock if at all possible
1265 page_locked = trylock_page(page);
1266 target_nid = mpol_misplaced(page, vma, haddr);
1267 if (target_nid == -1) {
1268 /* If the page was locked, there are no parallel migrations */
1273 /* Migration could have started since the pmd_trans_migrating check */
1276 if (!get_page_unless_zero(page))
1278 spin_unlock(fe->ptl);
1279 wait_on_page_locked(page);
1285 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1286 * to serialises splits
1289 spin_unlock(fe->ptl);
1290 anon_vma = page_lock_anon_vma_read(page);
1292 /* Confirm the PMD did not change while page_table_lock was released */
1294 if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1301 /* Bail if we fail to protect against THP splits for any reason */
1302 if (unlikely(!anon_vma)) {
1309 * Migrate the THP to the requested node, returns with page unlocked
1310 * and access rights restored.
1312 spin_unlock(fe->ptl);
1313 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1314 fe->pmd, pmd, fe->address, page, target_nid);
1316 flags |= TNF_MIGRATED;
1317 page_nid = target_nid;
1319 flags |= TNF_MIGRATE_FAIL;
1323 BUG_ON(!PageLocked(page));
1324 was_writable = pmd_write(pmd);
1325 pmd = pmd_modify(pmd, vma->vm_page_prot);
1326 pmd = pmd_mkyoung(pmd);
1328 pmd = pmd_mkwrite(pmd);
1329 set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1330 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1333 spin_unlock(fe->ptl);
1337 page_unlock_anon_vma_read(anon_vma);
1340 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1346 * Return true if we do MADV_FREE successfully on entire pmd page.
1347 * Otherwise, return false.
1349 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1350 pmd_t *pmd, unsigned long addr, unsigned long next)
1355 struct mm_struct *mm = tlb->mm;
1358 ptl = pmd_trans_huge_lock(pmd, vma);
1363 if (is_huge_zero_pmd(orig_pmd))
1366 page = pmd_page(orig_pmd);
1368 * If other processes are mapping this page, we couldn't discard
1369 * the page unless they all do MADV_FREE so let's skip the page.
1371 if (total_mapcount(page) != 1)
1374 if (!trylock_page(page))
1378 * If user want to discard part-pages of THP, split it so MADV_FREE
1379 * will deactivate only them.
1381 if (next - addr != HPAGE_PMD_SIZE) {
1384 split_huge_page(page);
1390 if (PageDirty(page))
1391 ClearPageDirty(page);
1394 if (PageActive(page))
1395 deactivate_page(page);
1397 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1398 pmdp_invalidate(vma, addr, pmd);
1399 orig_pmd = pmd_mkold(orig_pmd);
1400 orig_pmd = pmd_mkclean(orig_pmd);
1402 set_pmd_at(mm, addr, pmd, orig_pmd);
1403 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1412 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1413 pmd_t *pmd, unsigned long addr)
1418 ptl = __pmd_trans_huge_lock(pmd, vma);
1422 * For architectures like ppc64 we look at deposited pgtable
1423 * when calling pmdp_huge_get_and_clear. So do the
1424 * pgtable_trans_huge_withdraw after finishing pmdp related
1427 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1429 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1430 if (vma_is_dax(vma)) {
1432 if (is_huge_zero_pmd(orig_pmd))
1433 tlb_remove_page(tlb, pmd_page(orig_pmd));
1434 } else if (is_huge_zero_pmd(orig_pmd)) {
1435 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1436 atomic_long_dec(&tlb->mm->nr_ptes);
1438 tlb_remove_page(tlb, pmd_page(orig_pmd));
1440 struct page *page = pmd_page(orig_pmd);
1441 page_remove_rmap(page, true);
1442 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1443 VM_BUG_ON_PAGE(!PageHead(page), page);
1444 if (PageAnon(page)) {
1446 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1447 pte_free(tlb->mm, pgtable);
1448 atomic_long_dec(&tlb->mm->nr_ptes);
1449 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1451 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1454 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1459 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1460 unsigned long new_addr, unsigned long old_end,
1461 pmd_t *old_pmd, pmd_t *new_pmd)
1463 spinlock_t *old_ptl, *new_ptl;
1465 struct mm_struct *mm = vma->vm_mm;
1466 bool force_flush = false;
1468 if ((old_addr & ~HPAGE_PMD_MASK) ||
1469 (new_addr & ~HPAGE_PMD_MASK) ||
1470 old_end - old_addr < HPAGE_PMD_SIZE)
1474 * The destination pmd shouldn't be established, free_pgtables()
1475 * should have release it.
1477 if (WARN_ON(!pmd_none(*new_pmd))) {
1478 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1483 * We don't have to worry about the ordering of src and dst
1484 * ptlocks because exclusive mmap_sem prevents deadlock.
1486 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1488 new_ptl = pmd_lockptr(mm, new_pmd);
1489 if (new_ptl != old_ptl)
1490 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1491 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1492 if (pmd_present(pmd))
1494 VM_BUG_ON(!pmd_none(*new_pmd));
1496 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1497 vma_is_anonymous(vma)) {
1499 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1500 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1502 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1504 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1505 if (new_ptl != old_ptl)
1506 spin_unlock(new_ptl);
1507 spin_unlock(old_ptl);
1515 * - 0 if PMD could not be locked
1516 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1517 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1519 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1520 unsigned long addr, pgprot_t newprot, int prot_numa)
1522 struct mm_struct *mm = vma->vm_mm;
1525 bool preserve_write;
1528 ptl = __pmd_trans_huge_lock(pmd, vma);
1532 preserve_write = prot_numa && pmd_write(*pmd);
1536 * Avoid trapping faults against the zero page. The read-only
1537 * data is likely to be read-cached on the local CPU and
1538 * local/remote hits to the zero page are not interesting.
1540 if (prot_numa && is_huge_zero_pmd(*pmd))
1543 if (prot_numa && pmd_protnone(*pmd))
1547 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1548 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1549 * which is also under down_read(mmap_sem):
1552 * change_huge_pmd(prot_numa=1)
1553 * pmdp_huge_get_and_clear_notify()
1554 * madvise_dontneed()
1556 * pmd_trans_huge(*pmd) == 0 (without ptl)
1559 * // pmd is re-established
1561 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1562 * which may break userspace.
1564 * pmdp_invalidate() is required to make sure we don't miss
1565 * dirty/young flags set by hardware.
1568 pmdp_invalidate(vma, addr, pmd);
1571 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1574 if (pmd_dirty(*pmd))
1575 entry = pmd_mkdirty(entry);
1576 if (pmd_young(*pmd))
1577 entry = pmd_mkyoung(entry);
1579 entry = pmd_modify(entry, newprot);
1581 entry = pmd_mkwrite(entry);
1583 set_pmd_at(mm, addr, pmd, entry);
1584 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1591 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1593 * Note that if it returns page table lock pointer, this routine returns without
1594 * unlocking page table lock. So callers must unlock it.
1596 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1599 ptl = pmd_lock(vma->vm_mm, pmd);
1600 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1606 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1607 unsigned long haddr, pmd_t *pmd)
1609 struct mm_struct *mm = vma->vm_mm;
1614 /* leave pmd empty until pte is filled */
1615 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1617 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1618 pmd_populate(mm, &_pmd, pgtable);
1620 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1622 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1623 entry = pte_mkspecial(entry);
1624 pte = pte_offset_map(&_pmd, haddr);
1625 VM_BUG_ON(!pte_none(*pte));
1626 set_pte_at(mm, haddr, pte, entry);
1629 smp_wmb(); /* make pte visible before pmd */
1630 pmd_populate(mm, pmd, pgtable);
1633 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1634 unsigned long haddr, bool freeze)
1636 struct mm_struct *mm = vma->vm_mm;
1640 bool young, write, dirty, soft_dirty;
1644 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1645 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1646 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1647 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1649 count_vm_event(THP_SPLIT_PMD);
1651 if (!vma_is_anonymous(vma)) {
1652 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1653 if (vma_is_dax(vma))
1655 page = pmd_page(_pmd);
1656 if (!PageDirty(page) && pmd_dirty(_pmd))
1657 set_page_dirty(page);
1658 if (!PageReferenced(page) && pmd_young(_pmd))
1659 SetPageReferenced(page);
1660 page_remove_rmap(page, true);
1662 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1664 } else if (is_huge_zero_pmd(*pmd)) {
1665 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1668 page = pmd_page(*pmd);
1669 VM_BUG_ON_PAGE(!page_count(page), page);
1670 page_ref_add(page, HPAGE_PMD_NR - 1);
1671 write = pmd_write(*pmd);
1672 young = pmd_young(*pmd);
1673 dirty = pmd_dirty(*pmd);
1674 soft_dirty = pmd_soft_dirty(*pmd);
1676 pmdp_huge_split_prepare(vma, haddr, pmd);
1677 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1678 pmd_populate(mm, &_pmd, pgtable);
1680 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1683 * Note that NUMA hinting access restrictions are not
1684 * transferred to avoid any possibility of altering
1685 * permissions across VMAs.
1688 swp_entry_t swp_entry;
1689 swp_entry = make_migration_entry(page + i, write);
1690 entry = swp_entry_to_pte(swp_entry);
1692 entry = pte_swp_mksoft_dirty(entry);
1694 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1695 entry = maybe_mkwrite(entry, vma);
1697 entry = pte_wrprotect(entry);
1699 entry = pte_mkold(entry);
1701 entry = pte_mksoft_dirty(entry);
1704 SetPageDirty(page + i);
1705 pte = pte_offset_map(&_pmd, addr);
1706 BUG_ON(!pte_none(*pte));
1707 set_pte_at(mm, addr, pte, entry);
1708 atomic_inc(&page[i]._mapcount);
1713 * Set PG_double_map before dropping compound_mapcount to avoid
1714 * false-negative page_mapped().
1716 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1717 for (i = 0; i < HPAGE_PMD_NR; i++)
1718 atomic_inc(&page[i]._mapcount);
1721 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1722 /* Last compound_mapcount is gone. */
1723 __dec_node_page_state(page, NR_ANON_THPS);
1724 if (TestClearPageDoubleMap(page)) {
1725 /* No need in mapcount reference anymore */
1726 for (i = 0; i < HPAGE_PMD_NR; i++)
1727 atomic_dec(&page[i]._mapcount);
1731 smp_wmb(); /* make pte visible before pmd */
1733 * Up to this point the pmd is present and huge and userland has the
1734 * whole access to the hugepage during the split (which happens in
1735 * place). If we overwrite the pmd with the not-huge version pointing
1736 * to the pte here (which of course we could if all CPUs were bug
1737 * free), userland could trigger a small page size TLB miss on the
1738 * small sized TLB while the hugepage TLB entry is still established in
1739 * the huge TLB. Some CPU doesn't like that.
1740 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1741 * 383 on page 93. Intel should be safe but is also warns that it's
1742 * only safe if the permission and cache attributes of the two entries
1743 * loaded in the two TLB is identical (which should be the case here).
1744 * But it is generally safer to never allow small and huge TLB entries
1745 * for the same virtual address to be loaded simultaneously. So instead
1746 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1747 * current pmd notpresent (atomically because here the pmd_trans_huge
1748 * and pmd_trans_splitting must remain set at all times on the pmd
1749 * until the split is complete for this pmd), then we flush the SMP TLB
1750 * and finally we write the non-huge version of the pmd entry with
1753 pmdp_invalidate(vma, haddr, pmd);
1754 pmd_populate(mm, pmd, pgtable);
1757 for (i = 0; i < HPAGE_PMD_NR; i++) {
1758 page_remove_rmap(page + i, false);
1764 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1765 unsigned long address, bool freeze, struct page *page)
1768 struct mm_struct *mm = vma->vm_mm;
1769 unsigned long haddr = address & HPAGE_PMD_MASK;
1770 bool do_unlock_page = false;
1773 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1774 ptl = pmd_lock(mm, pmd);
1777 * If caller asks to setup a migration entries, we need a page to check
1778 * pmd against. Otherwise we can end up replacing wrong page.
1780 VM_BUG_ON(freeze && !page);
1782 VM_WARN_ON_ONCE(!PageLocked(page));
1783 if (page != pmd_page(*pmd))
1788 if (pmd_trans_huge(*pmd)) {
1790 page = pmd_page(*pmd);
1792 * An anonymous page must be locked, to ensure that a
1793 * concurrent reuse_swap_page() sees stable mapcount;
1794 * but reuse_swap_page() is not used on shmem or file,
1795 * and page lock must not be taken when zap_pmd_range()
1796 * calls __split_huge_pmd() while i_mmap_lock is held.
1798 if (PageAnon(page)) {
1799 if (unlikely(!trylock_page(page))) {
1805 if (unlikely(!pmd_same(*pmd, _pmd))) {
1813 do_unlock_page = true;
1816 if (PageMlocked(page))
1817 clear_page_mlock(page);
1818 } else if (!pmd_devmap(*pmd))
1820 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1825 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1828 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1829 bool freeze, struct page *page)
1835 pgd = pgd_offset(vma->vm_mm, address);
1836 if (!pgd_present(*pgd))
1839 pud = pud_offset(pgd, address);
1840 if (!pud_present(*pud))
1843 pmd = pmd_offset(pud, address);
1845 __split_huge_pmd(vma, pmd, address, freeze, page);
1848 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1849 unsigned long start,
1854 * If the new start address isn't hpage aligned and it could
1855 * previously contain an hugepage: check if we need to split
1858 if (start & ~HPAGE_PMD_MASK &&
1859 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1860 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1861 split_huge_pmd_address(vma, start, false, NULL);
1864 * If the new end address isn't hpage aligned and it could
1865 * previously contain an hugepage: check if we need to split
1868 if (end & ~HPAGE_PMD_MASK &&
1869 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1870 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1871 split_huge_pmd_address(vma, end, false, NULL);
1874 * If we're also updating the vma->vm_next->vm_start, if the new
1875 * vm_next->vm_start isn't page aligned and it could previously
1876 * contain an hugepage: check if we need to split an huge pmd.
1878 if (adjust_next > 0) {
1879 struct vm_area_struct *next = vma->vm_next;
1880 unsigned long nstart = next->vm_start;
1881 nstart += adjust_next << PAGE_SHIFT;
1882 if (nstart & ~HPAGE_PMD_MASK &&
1883 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1884 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1885 split_huge_pmd_address(next, nstart, false, NULL);
1889 static void unmap_page(struct page *page)
1891 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1895 VM_BUG_ON_PAGE(!PageHead(page), page);
1898 ttu_flags |= TTU_MIGRATION;
1900 /* We only need TTU_SPLIT_HUGE_PMD once */
1901 try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1902 for (i = 1; i < HPAGE_PMD_NR; i++) {
1903 /* Cut short if the page is unmapped */
1904 if (page_count(page) == 1)
1907 try_to_unmap(page + i, ttu_flags);
1910 VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
1913 static void remap_page(struct page *page)
1917 for (i = 0; i < HPAGE_PMD_NR; i++)
1918 remove_migration_ptes(page + i, page + i, true);
1921 static void __split_huge_page_tail(struct page *head, int tail,
1922 struct lruvec *lruvec, struct list_head *list)
1924 struct page *page_tail = head + tail;
1926 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1929 * Clone page flags before unfreezing refcount.
1931 * After successful get_page_unless_zero() might follow flags change,
1932 * for exmaple lock_page() which set PG_waiters.
1934 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1935 page_tail->flags |= (head->flags &
1936 ((1L << PG_referenced) |
1937 (1L << PG_swapbacked) |
1938 (1L << PG_mlocked) |
1939 (1L << PG_uptodate) |
1942 (1L << PG_unevictable) |
1945 /* ->mapping in first tail page is compound_mapcount */
1946 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1948 page_tail->mapping = head->mapping;
1949 page_tail->index = head->index + tail;
1951 /* Page flags must be visible before we make the page non-compound. */
1955 * Clear PageTail before unfreezing page refcount.
1957 * After successful get_page_unless_zero() might follow put_page()
1958 * which needs correct compound_head().
1960 clear_compound_head(page_tail);
1962 /* Finally unfreeze refcount. Additional reference from page cache. */
1963 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
1964 PageSwapCache(head)));
1966 if (page_is_young(head))
1967 set_page_young(page_tail);
1968 if (page_is_idle(head))
1969 set_page_idle(page_tail);
1971 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1972 lru_add_page_tail(head, page_tail, lruvec, list);
1975 static void __split_huge_page(struct page *page, struct list_head *list,
1976 pgoff_t end, unsigned long flags)
1978 struct page *head = compound_head(page);
1979 struct zone *zone = page_zone(head);
1980 struct lruvec *lruvec;
1983 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1985 /* complete memcg works before add pages to LRU */
1986 mem_cgroup_split_huge_fixup(head);
1988 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1989 __split_huge_page_tail(head, i, lruvec, list);
1990 /* Some pages can be beyond i_size: drop them from page cache */
1991 if (head[i].index >= end) {
1992 __ClearPageDirty(head + i);
1993 __delete_from_page_cache(head + i, NULL);
1994 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1995 shmem_uncharge(head->mapping->host, 1);
2000 ClearPageCompound(head);
2002 split_page_owner(head, HPAGE_PMD_ORDER);
2004 /* See comment in __split_huge_page_tail() */
2005 if (PageAnon(head)) {
2008 /* Additional pin to radix tree */
2009 page_ref_add(head, 2);
2010 spin_unlock(&head->mapping->tree_lock);
2013 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2017 for (i = 0; i < HPAGE_PMD_NR; i++) {
2018 struct page *subpage = head + i;
2019 if (subpage == page)
2021 unlock_page(subpage);
2024 * Subpages may be freed if there wasn't any mapping
2025 * like if add_to_swap() is running on a lru page that
2026 * had its mapping zapped. And freeing these pages
2027 * requires taking the lru_lock so we do the put_page
2028 * of the tail pages after the split is complete.
2034 int total_mapcount(struct page *page)
2036 int i, compound, ret;
2038 VM_BUG_ON_PAGE(PageTail(page), page);
2040 if (likely(!PageCompound(page)))
2041 return atomic_read(&page->_mapcount) + 1;
2043 compound = compound_mapcount(page);
2047 for (i = 0; i < HPAGE_PMD_NR; i++)
2048 ret += atomic_read(&page[i]._mapcount) + 1;
2049 /* File pages has compound_mapcount included in _mapcount */
2050 if (!PageAnon(page))
2051 return ret - compound * HPAGE_PMD_NR;
2052 if (PageDoubleMap(page))
2053 ret -= HPAGE_PMD_NR;
2058 * This calculates accurately how many mappings a transparent hugepage
2059 * has (unlike page_mapcount() which isn't fully accurate). This full
2060 * accuracy is primarily needed to know if copy-on-write faults can
2061 * reuse the page and change the mapping to read-write instead of
2062 * copying them. At the same time this returns the total_mapcount too.
2064 * The function returns the highest mapcount any one of the subpages
2065 * has. If the return value is one, even if different processes are
2066 * mapping different subpages of the transparent hugepage, they can
2067 * all reuse it, because each process is reusing a different subpage.
2069 * The total_mapcount is instead counting all virtual mappings of the
2070 * subpages. If the total_mapcount is equal to "one", it tells the
2071 * caller all mappings belong to the same "mm" and in turn the
2072 * anon_vma of the transparent hugepage can become the vma->anon_vma
2073 * local one as no other process may be mapping any of the subpages.
2075 * It would be more accurate to replace page_mapcount() with
2076 * page_trans_huge_mapcount(), however we only use
2077 * page_trans_huge_mapcount() in the copy-on-write faults where we
2078 * need full accuracy to avoid breaking page pinning, because
2079 * page_trans_huge_mapcount() is slower than page_mapcount().
2081 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2083 int i, ret, _total_mapcount, mapcount;
2085 /* hugetlbfs shouldn't call it */
2086 VM_BUG_ON_PAGE(PageHuge(page), page);
2088 if (likely(!PageTransCompound(page))) {
2089 mapcount = atomic_read(&page->_mapcount) + 1;
2091 *total_mapcount = mapcount;
2095 page = compound_head(page);
2097 _total_mapcount = ret = 0;
2098 for (i = 0; i < HPAGE_PMD_NR; i++) {
2099 mapcount = atomic_read(&page[i]._mapcount) + 1;
2100 ret = max(ret, mapcount);
2101 _total_mapcount += mapcount;
2103 if (PageDoubleMap(page)) {
2105 _total_mapcount -= HPAGE_PMD_NR;
2107 mapcount = compound_mapcount(page);
2109 _total_mapcount += mapcount;
2111 *total_mapcount = _total_mapcount;
2116 * This function splits huge page into normal pages. @page can point to any
2117 * subpage of huge page to split. Split doesn't change the position of @page.
2119 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2120 * The huge page must be locked.
2122 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2124 * Both head page and tail pages will inherit mapping, flags, and so on from
2127 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2128 * they are not mapped.
2130 * Returns 0 if the hugepage is split successfully.
2131 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2134 int split_huge_page_to_list(struct page *page, struct list_head *list)
2136 struct page *head = compound_head(page);
2137 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2138 struct anon_vma *anon_vma = NULL;
2139 struct address_space *mapping = NULL;
2140 int extra_pins, ret;
2142 unsigned long flags;
2145 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2146 VM_BUG_ON_PAGE(!PageLocked(page), page);
2147 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2148 VM_BUG_ON_PAGE(!PageCompound(page), page);
2150 if (PageAnon(head)) {
2152 * The caller does not necessarily hold an mmap_sem that would
2153 * prevent the anon_vma disappearing so we first we take a
2154 * reference to it and then lock the anon_vma for write. This
2155 * is similar to page_lock_anon_vma_read except the write lock
2156 * is taken to serialise against parallel split or collapse
2159 anon_vma = page_get_anon_vma(head);
2167 anon_vma_lock_write(anon_vma);
2169 mapping = head->mapping;
2177 /* Addidional pins from radix tree */
2178 extra_pins = HPAGE_PMD_NR;
2180 i_mmap_lock_read(mapping);
2183 *__split_huge_page() may need to trim off pages beyond EOF:
2184 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2185 * which cannot be nested inside the page tree lock. So note
2186 * end now: i_size itself may be changed at any moment, but
2187 * head page lock is good enough to serialize the trimming.
2189 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2193 * Racy check if we can split the page, before unmap_page() will
2196 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2201 mlocked = PageMlocked(page);
2204 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2208 /* prevent PageLRU to go away from under us, and freeze lru stats */
2209 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2214 spin_lock(&mapping->tree_lock);
2215 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2218 * Check if the head page is present in radix tree.
2219 * We assume all tail are present too, if head is there.
2221 if (radix_tree_deref_slot_protected(pslot,
2222 &mapping->tree_lock) != head)
2226 /* Prevent deferred_split_scan() touching ->_refcount */
2227 spin_lock(&pgdata->split_queue_lock);
2228 if (page_ref_freeze(head, 1 + extra_pins)) {
2229 if (!list_empty(page_deferred_list(head))) {
2230 pgdata->split_queue_len--;
2231 list_del(page_deferred_list(head));
2234 __dec_node_page_state(page, NR_SHMEM_THPS);
2235 spin_unlock(&pgdata->split_queue_lock);
2236 __split_huge_page(page, list, end, flags);
2239 spin_unlock(&pgdata->split_queue_lock);
2242 spin_unlock(&mapping->tree_lock);
2243 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2250 anon_vma_unlock_write(anon_vma);
2251 put_anon_vma(anon_vma);
2254 i_mmap_unlock_read(mapping);
2256 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2260 void free_transhuge_page(struct page *page)
2262 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2263 unsigned long flags;
2265 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2266 if (!list_empty(page_deferred_list(page))) {
2267 pgdata->split_queue_len--;
2268 list_del(page_deferred_list(page));
2270 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2271 free_compound_page(page);
2274 void deferred_split_huge_page(struct page *page)
2276 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2277 unsigned long flags;
2279 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2281 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2282 if (list_empty(page_deferred_list(page))) {
2283 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2284 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2285 pgdata->split_queue_len++;
2287 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2290 static unsigned long deferred_split_count(struct shrinker *shrink,
2291 struct shrink_control *sc)
2293 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2294 return ACCESS_ONCE(pgdata->split_queue_len);
2297 static unsigned long deferred_split_scan(struct shrinker *shrink,
2298 struct shrink_control *sc)
2300 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2301 unsigned long flags;
2302 LIST_HEAD(list), *pos, *next;
2306 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2307 /* Take pin on all head pages to avoid freeing them under us */
2308 list_for_each_safe(pos, next, &pgdata->split_queue) {
2309 page = list_entry((void *)pos, struct page, mapping);
2310 page = compound_head(page);
2311 if (get_page_unless_zero(page)) {
2312 list_move(page_deferred_list(page), &list);
2314 /* We lost race with put_compound_page() */
2315 list_del_init(page_deferred_list(page));
2316 pgdata->split_queue_len--;
2318 if (!--sc->nr_to_scan)
2321 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2323 list_for_each_safe(pos, next, &list) {
2324 page = list_entry((void *)pos, struct page, mapping);
2325 if (!trylock_page(page))
2327 /* split_huge_page() removes page from list on success */
2328 if (!split_huge_page(page))
2335 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2336 list_splice_tail(&list, &pgdata->split_queue);
2337 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2340 * Stop shrinker if we didn't split any page, but the queue is empty.
2341 * This can happen if pages were freed under us.
2343 if (!split && list_empty(&pgdata->split_queue))
2348 static struct shrinker deferred_split_shrinker = {
2349 .count_objects = deferred_split_count,
2350 .scan_objects = deferred_split_scan,
2351 .seeks = DEFAULT_SEEKS,
2352 .flags = SHRINKER_NUMA_AWARE,
2355 #ifdef CONFIG_DEBUG_FS
2356 static int split_huge_pages_set(void *data, u64 val)
2360 unsigned long pfn, max_zone_pfn;
2361 unsigned long total = 0, split = 0;
2366 for_each_populated_zone(zone) {
2367 max_zone_pfn = zone_end_pfn(zone);
2368 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2369 if (!pfn_valid(pfn))
2372 page = pfn_to_page(pfn);
2373 if (!get_page_unless_zero(page))
2376 if (zone != page_zone(page))
2379 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2384 if (!split_huge_page(page))
2392 pr_info("%lu of %lu THP split\n", split, total);
2396 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2399 static int __init split_huge_pages_debugfs(void)
2403 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2404 &split_huge_pages_fops);
2406 pr_warn("Failed to create split_huge_pages in debugfs");
2409 late_initcall(split_huge_pages_debugfs);