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/sched/coredump.h>
13 #include <linux/sched/numa_balancing.h>
14 #include <linux/highmem.h>
15 #include <linux/hugetlb.h>
16 #include <linux/mmu_notifier.h>
17 #include <linux/rmap.h>
18 #include <linux/swap.h>
19 #include <linux/shrinker.h>
20 #include <linux/mm_inline.h>
21 #include <linux/swapops.h>
22 #include <linux/dax.h>
23 #include <linux/khugepaged.h>
24 #include <linux/freezer.h>
25 #include <linux/pfn_t.h>
26 #include <linux/mman.h>
27 #include <linux/memremap.h>
28 #include <linux/pagemap.h>
29 #include <linux/debugfs.h>
30 #include <linux/migrate.h>
31 #include <linux/hashtable.h>
32 #include <linux/userfaultfd_k.h>
33 #include <linux/page_idle.h>
34 #include <linux/shmem_fs.h>
35 #include <linux/oom.h>
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;
65 static struct page *get_huge_zero_page(void)
67 struct page *zero_page;
69 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
70 return READ_ONCE(huge_zero_page);
72 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
78 count_vm_event(THP_ZERO_PAGE_ALLOC);
80 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
82 __free_pages(zero_page, compound_order(zero_page));
86 /* We take additional reference here. It will be put back by shrinker */
87 atomic_set(&huge_zero_refcount, 2);
89 return READ_ONCE(huge_zero_page);
92 static void put_huge_zero_page(void)
95 * Counter should never go to zero here. Only shrinker can put
98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
101 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
103 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
104 return READ_ONCE(huge_zero_page);
106 if (!get_huge_zero_page())
109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
110 put_huge_zero_page();
112 return READ_ONCE(huge_zero_page);
115 void mm_put_huge_zero_page(struct mm_struct *mm)
117 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
118 put_huge_zero_page();
121 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
122 struct shrink_control *sc)
124 /* we can free zero page only if last reference remains */
125 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
128 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
129 struct shrink_control *sc)
131 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
132 struct page *zero_page = xchg(&huge_zero_page, NULL);
133 BUG_ON(zero_page == NULL);
134 __free_pages(zero_page, compound_order(zero_page));
141 static struct shrinker huge_zero_page_shrinker = {
142 .count_objects = shrink_huge_zero_page_count,
143 .scan_objects = shrink_huge_zero_page_scan,
144 .seeks = DEFAULT_SEEKS,
148 static ssize_t enabled_show(struct kobject *kobj,
149 struct kobj_attribute *attr, char *buf)
151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
152 return sprintf(buf, "[always] madvise never\n");
153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
154 return sprintf(buf, "always [madvise] never\n");
156 return sprintf(buf, "always madvise [never]\n");
159 static ssize_t enabled_store(struct kobject *kobj,
160 struct kobj_attribute *attr,
161 const char *buf, size_t count)
165 if (!memcmp("always", buf,
166 min(sizeof("always")-1, count))) {
167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
168 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
169 } else if (!memcmp("madvise", buf,
170 min(sizeof("madvise")-1, count))) {
171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
173 } else if (!memcmp("never", buf,
174 min(sizeof("never")-1, count))) {
175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
181 int err = start_stop_khugepaged();
187 static struct kobj_attribute enabled_attr =
188 __ATTR(enabled, 0644, enabled_show, enabled_store);
190 ssize_t single_hugepage_flag_show(struct kobject *kobj,
191 struct kobj_attribute *attr, char *buf,
192 enum transparent_hugepage_flag flag)
194 return sprintf(buf, "%d\n",
195 !!test_bit(flag, &transparent_hugepage_flags));
198 ssize_t single_hugepage_flag_store(struct kobject *kobj,
199 struct kobj_attribute *attr,
200 const char *buf, size_t count,
201 enum transparent_hugepage_flag flag)
206 ret = kstrtoul(buf, 10, &value);
213 set_bit(flag, &transparent_hugepage_flags);
215 clear_bit(flag, &transparent_hugepage_flags);
220 static ssize_t defrag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf)
223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
224 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
226 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
228 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
230 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
231 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
234 static ssize_t defrag_store(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 const char *buf, size_t count)
238 if (!memcmp("always", buf,
239 min(sizeof("always")-1, count))) {
240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
244 } else if (!memcmp("defer+madvise", buf,
245 min(sizeof("defer+madvise")-1, count))) {
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
250 } else if (!memcmp("defer", buf,
251 min(sizeof("defer")-1, count))) {
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
256 } else if (!memcmp("madvise", buf,
257 min(sizeof("madvise")-1, count))) {
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
262 } else if (!memcmp("never", buf,
263 min(sizeof("never")-1, count))) {
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
273 static struct kobj_attribute defrag_attr =
274 __ATTR(defrag, 0644, defrag_show, defrag_store);
276 static ssize_t use_zero_page_show(struct kobject *kobj,
277 struct kobj_attribute *attr, char *buf)
279 return single_hugepage_flag_show(kobj, attr, buf,
280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
282 static ssize_t use_zero_page_store(struct kobject *kobj,
283 struct kobj_attribute *attr, const char *buf, size_t count)
285 return single_hugepage_flag_store(kobj, attr, buf, count,
286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
288 static struct kobj_attribute use_zero_page_attr =
289 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
291 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
292 struct kobj_attribute *attr, char *buf)
294 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
296 static struct kobj_attribute hpage_pmd_size_attr =
297 __ATTR_RO(hpage_pmd_size);
299 #ifdef CONFIG_DEBUG_VM
300 static ssize_t debug_cow_show(struct kobject *kobj,
301 struct kobj_attribute *attr, char *buf)
303 return single_hugepage_flag_show(kobj, attr, buf,
304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
306 static ssize_t debug_cow_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
310 return single_hugepage_flag_store(kobj, attr, buf, count,
311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
313 static struct kobj_attribute debug_cow_attr =
314 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
315 #endif /* CONFIG_DEBUG_VM */
317 static struct attribute *hugepage_attr[] = {
320 &use_zero_page_attr.attr,
321 &hpage_pmd_size_attr.attr,
322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
323 &shmem_enabled_attr.attr,
325 #ifdef CONFIG_DEBUG_VM
326 &debug_cow_attr.attr,
331 static const struct attribute_group hugepage_attr_group = {
332 .attrs = hugepage_attr,
335 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
339 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
340 if (unlikely(!*hugepage_kobj)) {
341 pr_err("failed to create transparent hugepage kobject\n");
345 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
347 pr_err("failed to register transparent hugepage group\n");
351 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
353 pr_err("failed to register transparent hugepage group\n");
354 goto remove_hp_group;
360 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
362 kobject_put(*hugepage_kobj);
366 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
368 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
369 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
370 kobject_put(hugepage_kobj);
373 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
378 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
381 #endif /* CONFIG_SYSFS */
383 static int __init hugepage_init(void)
386 struct kobject *hugepage_kobj;
388 if (!has_transparent_hugepage()) {
389 transparent_hugepage_flags = 0;
394 * hugepages can't be allocated by the buddy allocator
396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
398 * we use page->mapping and page->index in second tail page
399 * as list_head: assuming THP order >= 2
401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
403 err = hugepage_init_sysfs(&hugepage_kobj);
407 err = khugepaged_init();
411 err = register_shrinker(&huge_zero_page_shrinker);
413 goto err_hzp_shrinker;
414 err = register_shrinker(&deferred_split_shrinker);
416 goto err_split_shrinker;
419 * By default disable transparent hugepages on smaller systems,
420 * where the extra memory used could hurt more than TLB overhead
421 * is likely to save. The admin can still enable it through /sys.
423 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
424 transparent_hugepage_flags = 0;
428 err = start_stop_khugepaged();
434 unregister_shrinker(&deferred_split_shrinker);
436 unregister_shrinker(&huge_zero_page_shrinker);
438 khugepaged_destroy();
440 hugepage_exit_sysfs(hugepage_kobj);
444 subsys_initcall(hugepage_init);
446 static int __init setup_transparent_hugepage(char *str)
451 if (!strcmp(str, "always")) {
452 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
453 &transparent_hugepage_flags);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455 &transparent_hugepage_flags);
457 } else if (!strcmp(str, "madvise")) {
458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
459 &transparent_hugepage_flags);
460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
461 &transparent_hugepage_flags);
463 } else if (!strcmp(str, "never")) {
464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
465 &transparent_hugepage_flags);
466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
467 &transparent_hugepage_flags);
472 pr_warn("transparent_hugepage= cannot parse, ignored\n");
475 __setup("transparent_hugepage=", setup_transparent_hugepage);
477 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
479 if (likely(vma->vm_flags & VM_WRITE))
480 pmd = pmd_mkwrite(pmd);
484 static inline struct list_head *page_deferred_list(struct page *page)
486 /* ->lru in the tail pages is occupied by compound_head. */
487 return &page[2].deferred_list;
490 void prep_transhuge_page(struct page *page)
493 * we use page->mapping and page->indexlru in second tail page
494 * as list_head: assuming THP order >= 2
497 INIT_LIST_HEAD(page_deferred_list(page));
498 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
501 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
502 loff_t off, unsigned long flags, unsigned long size)
505 loff_t off_end = off + len;
506 loff_t off_align = round_up(off, size);
507 unsigned long len_pad;
509 if (off_end <= off_align || (off_end - off_align) < size)
512 len_pad = len + size;
513 if (len_pad < len || (off + len_pad) < off)
516 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
517 off >> PAGE_SHIFT, flags);
518 if (IS_ERR_VALUE(addr))
521 addr += (off - addr) & (size - 1);
525 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
526 unsigned long len, unsigned long pgoff, unsigned long flags)
528 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
532 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
535 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
540 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
542 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
544 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
545 struct page *page, gfp_t gfp)
547 struct vm_area_struct *vma = vmf->vma;
548 struct mem_cgroup *memcg;
550 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
553 VM_BUG_ON_PAGE(!PageCompound(page), page);
555 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
557 count_vm_event(THP_FAULT_FALLBACK);
558 return VM_FAULT_FALLBACK;
561 pgtable = pte_alloc_one(vma->vm_mm, haddr);
562 if (unlikely(!pgtable)) {
567 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
569 * The memory barrier inside __SetPageUptodate makes sure that
570 * clear_huge_page writes become visible before the set_pmd_at()
573 __SetPageUptodate(page);
575 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
576 if (unlikely(!pmd_none(*vmf->pmd))) {
581 ret = check_stable_address_space(vma->vm_mm);
585 /* Deliver the page fault to userland */
586 if (userfaultfd_missing(vma)) {
589 spin_unlock(vmf->ptl);
590 mem_cgroup_cancel_charge(page, memcg, true);
592 pte_free(vma->vm_mm, pgtable);
593 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
594 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
598 entry = mk_huge_pmd(page, vma->vm_page_prot);
599 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
600 page_add_new_anon_rmap(page, vma, haddr, true);
601 mem_cgroup_commit_charge(page, memcg, false, true);
602 lru_cache_add_active_or_unevictable(page, vma);
603 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
604 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
605 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
606 mm_inc_nr_ptes(vma->vm_mm);
607 spin_unlock(vmf->ptl);
608 count_vm_event(THP_FAULT_ALLOC);
613 spin_unlock(vmf->ptl);
616 pte_free(vma->vm_mm, pgtable);
617 mem_cgroup_cancel_charge(page, memcg, true);
624 * always: directly stall for all thp allocations
625 * defer: wake kswapd and fail if not immediately available
626 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
627 * fail if not immediately available
628 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
630 * never: never stall for any thp allocation
632 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
634 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
636 /* Always do synchronous compaction */
637 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
638 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
640 /* Kick kcompactd and fail quickly */
641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
644 /* Synchronous compaction if madvised, otherwise kick kcompactd */
645 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
646 return GFP_TRANSHUGE_LIGHT |
647 (vma_madvised ? __GFP_DIRECT_RECLAIM :
648 __GFP_KSWAPD_RECLAIM);
650 /* Only do synchronous compaction if madvised */
651 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
652 return GFP_TRANSHUGE_LIGHT |
653 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
655 return GFP_TRANSHUGE_LIGHT;
658 /* Caller must hold page table lock. */
659 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
660 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
661 struct page *zero_page)
666 entry = mk_pmd(zero_page, vma->vm_page_prot);
667 entry = pmd_mkhuge(entry);
669 pgtable_trans_huge_deposit(mm, pmd, pgtable);
670 set_pmd_at(mm, haddr, pmd, entry);
675 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
677 struct vm_area_struct *vma = vmf->vma;
680 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
682 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
683 return VM_FAULT_FALLBACK;
684 if (unlikely(anon_vma_prepare(vma)))
686 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
688 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
689 !mm_forbids_zeropage(vma->vm_mm) &&
690 transparent_hugepage_use_zero_page()) {
692 struct page *zero_page;
695 pgtable = pte_alloc_one(vma->vm_mm, haddr);
696 if (unlikely(!pgtable))
698 zero_page = mm_get_huge_zero_page(vma->vm_mm);
699 if (unlikely(!zero_page)) {
700 pte_free(vma->vm_mm, pgtable);
701 count_vm_event(THP_FAULT_FALLBACK);
702 return VM_FAULT_FALLBACK;
704 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
707 if (pmd_none(*vmf->pmd)) {
708 ret = check_stable_address_space(vma->vm_mm);
710 spin_unlock(vmf->ptl);
711 } else if (userfaultfd_missing(vma)) {
712 spin_unlock(vmf->ptl);
713 ret = handle_userfault(vmf, VM_UFFD_MISSING);
714 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
716 set_huge_zero_page(pgtable, vma->vm_mm, vma,
717 haddr, vmf->pmd, zero_page);
718 spin_unlock(vmf->ptl);
722 spin_unlock(vmf->ptl);
724 pte_free(vma->vm_mm, pgtable);
727 gfp = alloc_hugepage_direct_gfpmask(vma);
728 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
729 if (unlikely(!page)) {
730 count_vm_event(THP_FAULT_FALLBACK);
731 return VM_FAULT_FALLBACK;
733 prep_transhuge_page(page);
734 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
737 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
738 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
741 struct mm_struct *mm = vma->vm_mm;
745 ptl = pmd_lock(mm, pmd);
746 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
747 if (pfn_t_devmap(pfn))
748 entry = pmd_mkdevmap(entry);
750 entry = pmd_mkyoung(pmd_mkdirty(entry));
751 entry = maybe_pmd_mkwrite(entry, vma);
755 pgtable_trans_huge_deposit(mm, pmd, pgtable);
759 set_pmd_at(mm, addr, pmd, entry);
760 update_mmu_cache_pmd(vma, addr, pmd);
764 vm_fault_t vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
765 pmd_t *pmd, pfn_t pfn, bool write)
767 pgprot_t pgprot = vma->vm_page_prot;
768 pgtable_t pgtable = NULL;
770 * If we had pmd_special, we could avoid all these restrictions,
771 * but we need to be consistent with PTEs and architectures that
772 * can't support a 'special' bit.
774 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
776 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
777 (VM_PFNMAP|VM_MIXEDMAP));
778 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
780 if (addr < vma->vm_start || addr >= vma->vm_end)
781 return VM_FAULT_SIGBUS;
783 if (arch_needs_pgtable_deposit()) {
784 pgtable = pte_alloc_one(vma->vm_mm, addr);
789 track_pfn_insert(vma, &pgprot, pfn);
791 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
792 return VM_FAULT_NOPAGE;
794 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
796 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
797 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
799 if (likely(vma->vm_flags & VM_WRITE))
800 pud = pud_mkwrite(pud);
804 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
805 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
807 struct mm_struct *mm = vma->vm_mm;
811 ptl = pud_lock(mm, pud);
812 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
813 if (pfn_t_devmap(pfn))
814 entry = pud_mkdevmap(entry);
816 entry = pud_mkyoung(pud_mkdirty(entry));
817 entry = maybe_pud_mkwrite(entry, vma);
819 set_pud_at(mm, addr, pud, entry);
820 update_mmu_cache_pud(vma, addr, pud);
824 vm_fault_t vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
825 pud_t *pud, pfn_t pfn, bool write)
827 pgprot_t pgprot = vma->vm_page_prot;
829 * If we had pud_special, we could avoid all these restrictions,
830 * but we need to be consistent with PTEs and architectures that
831 * can't support a 'special' bit.
833 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
835 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
836 (VM_PFNMAP|VM_MIXEDMAP));
837 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
839 if (addr < vma->vm_start || addr >= vma->vm_end)
840 return VM_FAULT_SIGBUS;
842 track_pfn_insert(vma, &pgprot, pfn);
844 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
845 return VM_FAULT_NOPAGE;
847 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
848 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
850 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
851 pmd_t *pmd, int flags)
855 _pmd = pmd_mkyoung(*pmd);
856 if (flags & FOLL_WRITE)
857 _pmd = pmd_mkdirty(_pmd);
858 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
859 pmd, _pmd, flags & FOLL_WRITE))
860 update_mmu_cache_pmd(vma, addr, pmd);
863 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
864 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
866 unsigned long pfn = pmd_pfn(*pmd);
867 struct mm_struct *mm = vma->vm_mm;
870 assert_spin_locked(pmd_lockptr(mm, pmd));
873 * When we COW a devmap PMD entry, we split it into PTEs, so we should
874 * not be in this function with `flags & FOLL_COW` set.
876 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
878 if (flags & FOLL_WRITE && !pmd_write(*pmd))
881 if (pmd_present(*pmd) && pmd_devmap(*pmd))
886 if (flags & FOLL_TOUCH)
887 touch_pmd(vma, addr, pmd, flags);
890 * device mapped pages can only be returned if the
891 * caller will manage the page reference count.
893 if (!(flags & FOLL_GET))
894 return ERR_PTR(-EEXIST);
896 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
897 *pgmap = get_dev_pagemap(pfn, *pgmap);
899 return ERR_PTR(-EFAULT);
900 page = pfn_to_page(pfn);
906 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
907 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
908 struct vm_area_struct *vma)
910 spinlock_t *dst_ptl, *src_ptl;
911 struct page *src_page;
913 pgtable_t pgtable = NULL;
916 /* Skip if can be re-fill on fault */
917 if (!vma_is_anonymous(vma))
920 pgtable = pte_alloc_one(dst_mm, addr);
921 if (unlikely(!pgtable))
924 dst_ptl = pmd_lock(dst_mm, dst_pmd);
925 src_ptl = pmd_lockptr(src_mm, src_pmd);
926 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
931 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
932 if (unlikely(is_swap_pmd(pmd))) {
933 swp_entry_t entry = pmd_to_swp_entry(pmd);
935 VM_BUG_ON(!is_pmd_migration_entry(pmd));
936 if (is_write_migration_entry(entry)) {
937 make_migration_entry_read(&entry);
938 pmd = swp_entry_to_pmd(entry);
939 if (pmd_swp_soft_dirty(*src_pmd))
940 pmd = pmd_swp_mksoft_dirty(pmd);
941 set_pmd_at(src_mm, addr, src_pmd, pmd);
943 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
944 mm_inc_nr_ptes(dst_mm);
945 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
946 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
952 if (unlikely(!pmd_trans_huge(pmd))) {
953 pte_free(dst_mm, pgtable);
957 * When page table lock is held, the huge zero pmd should not be
958 * under splitting since we don't split the page itself, only pmd to
961 if (is_huge_zero_pmd(pmd)) {
962 struct page *zero_page;
964 * get_huge_zero_page() will never allocate a new page here,
965 * since we already have a zero page to copy. It just takes a
968 zero_page = mm_get_huge_zero_page(dst_mm);
969 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
975 src_page = pmd_page(pmd);
976 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
978 page_dup_rmap(src_page, true);
979 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
980 mm_inc_nr_ptes(dst_mm);
981 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
983 pmdp_set_wrprotect(src_mm, addr, src_pmd);
984 pmd = pmd_mkold(pmd_wrprotect(pmd));
985 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
989 spin_unlock(src_ptl);
990 spin_unlock(dst_ptl);
995 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
996 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
997 pud_t *pud, int flags)
1001 _pud = pud_mkyoung(*pud);
1002 if (flags & FOLL_WRITE)
1003 _pud = pud_mkdirty(_pud);
1004 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1005 pud, _pud, flags & FOLL_WRITE))
1006 update_mmu_cache_pud(vma, addr, pud);
1009 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1010 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1012 unsigned long pfn = pud_pfn(*pud);
1013 struct mm_struct *mm = vma->vm_mm;
1016 assert_spin_locked(pud_lockptr(mm, pud));
1018 if (flags & FOLL_WRITE && !pud_write(*pud))
1021 if (pud_present(*pud) && pud_devmap(*pud))
1026 if (flags & FOLL_TOUCH)
1027 touch_pud(vma, addr, pud, flags);
1030 * device mapped pages can only be returned if the
1031 * caller will manage the page reference count.
1033 if (!(flags & FOLL_GET))
1034 return ERR_PTR(-EEXIST);
1036 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1037 *pgmap = get_dev_pagemap(pfn, *pgmap);
1039 return ERR_PTR(-EFAULT);
1040 page = pfn_to_page(pfn);
1046 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1047 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1048 struct vm_area_struct *vma)
1050 spinlock_t *dst_ptl, *src_ptl;
1054 dst_ptl = pud_lock(dst_mm, dst_pud);
1055 src_ptl = pud_lockptr(src_mm, src_pud);
1056 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1060 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1064 * When page table lock is held, the huge zero pud should not be
1065 * under splitting since we don't split the page itself, only pud to
1068 if (is_huge_zero_pud(pud)) {
1069 /* No huge zero pud yet */
1072 pudp_set_wrprotect(src_mm, addr, src_pud);
1073 pud = pud_mkold(pud_wrprotect(pud));
1074 set_pud_at(dst_mm, addr, dst_pud, pud);
1078 spin_unlock(src_ptl);
1079 spin_unlock(dst_ptl);
1083 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1086 unsigned long haddr;
1087 bool write = vmf->flags & FAULT_FLAG_WRITE;
1089 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1090 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1093 entry = pud_mkyoung(orig_pud);
1095 entry = pud_mkdirty(entry);
1096 haddr = vmf->address & HPAGE_PUD_MASK;
1097 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1098 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1101 spin_unlock(vmf->ptl);
1103 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1105 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1108 unsigned long haddr;
1109 bool write = vmf->flags & FAULT_FLAG_WRITE;
1111 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1112 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1115 entry = pmd_mkyoung(orig_pmd);
1117 entry = pmd_mkdirty(entry);
1118 haddr = vmf->address & HPAGE_PMD_MASK;
1119 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1120 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1123 spin_unlock(vmf->ptl);
1126 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1127 pmd_t orig_pmd, struct page *page)
1129 struct vm_area_struct *vma = vmf->vma;
1130 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1131 struct mem_cgroup *memcg;
1136 struct page **pages;
1137 struct mmu_notifier_range range;
1139 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1141 if (unlikely(!pages)) {
1142 ret |= VM_FAULT_OOM;
1146 for (i = 0; i < HPAGE_PMD_NR; i++) {
1147 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1148 vmf->address, page_to_nid(page));
1149 if (unlikely(!pages[i] ||
1150 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1151 GFP_KERNEL, &memcg, false))) {
1155 memcg = (void *)page_private(pages[i]);
1156 set_page_private(pages[i], 0);
1157 mem_cgroup_cancel_charge(pages[i], memcg,
1162 ret |= VM_FAULT_OOM;
1165 set_page_private(pages[i], (unsigned long)memcg);
1168 for (i = 0; i < HPAGE_PMD_NR; i++) {
1169 copy_user_highpage(pages[i], page + i,
1170 haddr + PAGE_SIZE * i, vma);
1171 __SetPageUptodate(pages[i]);
1175 mmu_notifier_range_init(&range, vma->vm_mm, haddr,
1176 haddr + HPAGE_PMD_SIZE);
1177 mmu_notifier_invalidate_range_start(&range);
1179 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1180 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1181 goto out_free_pages;
1182 VM_BUG_ON_PAGE(!PageHead(page), page);
1185 * Leave pmd empty until pte is filled note we must notify here as
1186 * concurrent CPU thread might write to new page before the call to
1187 * mmu_notifier_invalidate_range_end() happens which can lead to a
1188 * device seeing memory write in different order than CPU.
1190 * See Documentation/vm/mmu_notifier.rst
1192 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1194 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1195 pmd_populate(vma->vm_mm, &_pmd, pgtable);
1197 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1199 entry = mk_pte(pages[i], vma->vm_page_prot);
1200 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1201 memcg = (void *)page_private(pages[i]);
1202 set_page_private(pages[i], 0);
1203 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1204 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1205 lru_cache_add_active_or_unevictable(pages[i], vma);
1206 vmf->pte = pte_offset_map(&_pmd, haddr);
1207 VM_BUG_ON(!pte_none(*vmf->pte));
1208 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1209 pte_unmap(vmf->pte);
1213 smp_wmb(); /* make pte visible before pmd */
1214 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1215 page_remove_rmap(page, true);
1216 spin_unlock(vmf->ptl);
1219 * No need to double call mmu_notifier->invalidate_range() callback as
1220 * the above pmdp_huge_clear_flush_notify() did already call it.
1222 mmu_notifier_invalidate_range_only_end(&range);
1224 ret |= VM_FAULT_WRITE;
1231 spin_unlock(vmf->ptl);
1232 mmu_notifier_invalidate_range_end(&range);
1233 for (i = 0; i < HPAGE_PMD_NR; i++) {
1234 memcg = (void *)page_private(pages[i]);
1235 set_page_private(pages[i], 0);
1236 mem_cgroup_cancel_charge(pages[i], memcg, false);
1243 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1245 struct vm_area_struct *vma = vmf->vma;
1246 struct page *page = NULL, *new_page;
1247 struct mem_cgroup *memcg;
1248 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1249 struct mmu_notifier_range range;
1250 gfp_t huge_gfp; /* for allocation and charge */
1253 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1254 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1255 if (is_huge_zero_pmd(orig_pmd))
1257 spin_lock(vmf->ptl);
1258 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1261 page = pmd_page(orig_pmd);
1262 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1264 * We can only reuse the page if nobody else maps the huge page or it's
1267 if (!trylock_page(page)) {
1269 spin_unlock(vmf->ptl);
1271 spin_lock(vmf->ptl);
1272 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1279 if (reuse_swap_page(page, NULL)) {
1281 entry = pmd_mkyoung(orig_pmd);
1282 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1283 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1284 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1285 ret |= VM_FAULT_WRITE;
1291 spin_unlock(vmf->ptl);
1293 if (transparent_hugepage_enabled(vma) &&
1294 !transparent_hugepage_debug_cow()) {
1295 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1296 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1300 if (likely(new_page)) {
1301 prep_transhuge_page(new_page);
1304 split_huge_pmd(vma, vmf->pmd, vmf->address);
1305 ret |= VM_FAULT_FALLBACK;
1307 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1308 if (ret & VM_FAULT_OOM) {
1309 split_huge_pmd(vma, vmf->pmd, vmf->address);
1310 ret |= VM_FAULT_FALLBACK;
1314 count_vm_event(THP_FAULT_FALLBACK);
1318 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1319 huge_gfp, &memcg, true))) {
1321 split_huge_pmd(vma, vmf->pmd, vmf->address);
1324 ret |= VM_FAULT_FALLBACK;
1325 count_vm_event(THP_FAULT_FALLBACK);
1329 count_vm_event(THP_FAULT_ALLOC);
1332 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1334 copy_user_huge_page(new_page, page, vmf->address,
1336 __SetPageUptodate(new_page);
1338 mmu_notifier_range_init(&range, vma->vm_mm, haddr,
1339 haddr + HPAGE_PMD_SIZE);
1340 mmu_notifier_invalidate_range_start(&range);
1342 spin_lock(vmf->ptl);
1345 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1346 spin_unlock(vmf->ptl);
1347 mem_cgroup_cancel_charge(new_page, memcg, true);
1352 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1353 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1354 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1355 page_add_new_anon_rmap(new_page, vma, haddr, true);
1356 mem_cgroup_commit_charge(new_page, memcg, false, true);
1357 lru_cache_add_active_or_unevictable(new_page, vma);
1358 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1359 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1361 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1363 VM_BUG_ON_PAGE(!PageHead(page), page);
1364 page_remove_rmap(page, true);
1367 ret |= VM_FAULT_WRITE;
1369 spin_unlock(vmf->ptl);
1372 * No need to double call mmu_notifier->invalidate_range() callback as
1373 * the above pmdp_huge_clear_flush_notify() did already call it.
1375 mmu_notifier_invalidate_range_only_end(&range);
1379 spin_unlock(vmf->ptl);
1384 * FOLL_FORCE can write to even unwritable pmd's, but only
1385 * after we've gone through a COW cycle and they are dirty.
1387 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1389 return pmd_write(pmd) ||
1390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1393 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct page *page = NULL;
1401 assert_spin_locked(pmd_lockptr(mm, pmd));
1403 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1406 /* Avoid dumping huge zero page */
1407 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1408 return ERR_PTR(-EFAULT);
1410 /* Full NUMA hinting faults to serialise migration in fault paths */
1411 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1414 page = pmd_page(*pmd);
1415 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1416 if (flags & FOLL_TOUCH)
1417 touch_pmd(vma, addr, pmd, flags);
1418 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1420 * We don't mlock() pte-mapped THPs. This way we can avoid
1421 * leaking mlocked pages into non-VM_LOCKED VMAs.
1425 * In most cases the pmd is the only mapping of the page as we
1426 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1427 * writable private mappings in populate_vma_page_range().
1429 * The only scenario when we have the page shared here is if we
1430 * mlocking read-only mapping shared over fork(). We skip
1431 * mlocking such pages.
1435 * We can expect PageDoubleMap() to be stable under page lock:
1436 * for file pages we set it in page_add_file_rmap(), which
1437 * requires page to be locked.
1440 if (PageAnon(page) && compound_mapcount(page) != 1)
1442 if (PageDoubleMap(page) || !page->mapping)
1444 if (!trylock_page(page))
1447 if (page->mapping && !PageDoubleMap(page))
1448 mlock_vma_page(page);
1452 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1453 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1454 if (flags & FOLL_GET)
1461 /* NUMA hinting page fault entry point for trans huge pmds */
1462 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1464 struct vm_area_struct *vma = vmf->vma;
1465 struct anon_vma *anon_vma = NULL;
1467 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1468 int page_nid = -1, this_nid = numa_node_id();
1469 int target_nid, last_cpupid = -1;
1471 bool migrated = false;
1475 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1476 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1480 * If there are potential migrations, wait for completion and retry
1481 * without disrupting NUMA hinting information. Do not relock and
1482 * check_same as the page may no longer be mapped.
1484 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1485 page = pmd_page(*vmf->pmd);
1486 if (!get_page_unless_zero(page))
1488 spin_unlock(vmf->ptl);
1489 put_and_wait_on_page_locked(page);
1493 page = pmd_page(pmd);
1494 BUG_ON(is_huge_zero_page(page));
1495 page_nid = page_to_nid(page);
1496 last_cpupid = page_cpupid_last(page);
1497 count_vm_numa_event(NUMA_HINT_FAULTS);
1498 if (page_nid == this_nid) {
1499 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1500 flags |= TNF_FAULT_LOCAL;
1503 /* See similar comment in do_numa_page for explanation */
1504 if (!pmd_savedwrite(pmd))
1505 flags |= TNF_NO_GROUP;
1508 * Acquire the page lock to serialise THP migrations but avoid dropping
1509 * page_table_lock if at all possible
1511 page_locked = trylock_page(page);
1512 target_nid = mpol_misplaced(page, vma, haddr);
1513 if (target_nid == -1) {
1514 /* If the page was locked, there are no parallel migrations */
1519 /* Migration could have started since the pmd_trans_migrating check */
1522 if (!get_page_unless_zero(page))
1524 spin_unlock(vmf->ptl);
1525 put_and_wait_on_page_locked(page);
1530 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1531 * to serialises splits
1534 spin_unlock(vmf->ptl);
1535 anon_vma = page_lock_anon_vma_read(page);
1537 /* Confirm the PMD did not change while page_table_lock was released */
1538 spin_lock(vmf->ptl);
1539 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1546 /* Bail if we fail to protect against THP splits for any reason */
1547 if (unlikely(!anon_vma)) {
1554 * Since we took the NUMA fault, we must have observed the !accessible
1555 * bit. Make sure all other CPUs agree with that, to avoid them
1556 * modifying the page we're about to migrate.
1558 * Must be done under PTL such that we'll observe the relevant
1559 * inc_tlb_flush_pending().
1561 * We are not sure a pending tlb flush here is for a huge page
1562 * mapping or not. Hence use the tlb range variant
1564 if (mm_tlb_flush_pending(vma->vm_mm)) {
1565 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1567 * change_huge_pmd() released the pmd lock before
1568 * invalidating the secondary MMUs sharing the primary
1569 * MMU pagetables (with ->invalidate_range()). The
1570 * mmu_notifier_invalidate_range_end() (which
1571 * internally calls ->invalidate_range()) in
1572 * change_pmd_range() will run after us, so we can't
1573 * rely on it here and we need an explicit invalidate.
1575 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1576 haddr + HPAGE_PMD_SIZE);
1580 * Migrate the THP to the requested node, returns with page unlocked
1581 * and access rights restored.
1583 spin_unlock(vmf->ptl);
1585 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1586 vmf->pmd, pmd, vmf->address, page, target_nid);
1588 flags |= TNF_MIGRATED;
1589 page_nid = target_nid;
1591 flags |= TNF_MIGRATE_FAIL;
1595 BUG_ON(!PageLocked(page));
1596 was_writable = pmd_savedwrite(pmd);
1597 pmd = pmd_modify(pmd, vma->vm_page_prot);
1598 pmd = pmd_mkyoung(pmd);
1600 pmd = pmd_mkwrite(pmd);
1601 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1602 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1605 spin_unlock(vmf->ptl);
1609 page_unlock_anon_vma_read(anon_vma);
1612 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1619 * Return true if we do MADV_FREE successfully on entire pmd page.
1620 * Otherwise, return false.
1622 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1623 pmd_t *pmd, unsigned long addr, unsigned long next)
1628 struct mm_struct *mm = tlb->mm;
1631 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1633 ptl = pmd_trans_huge_lock(pmd, vma);
1638 if (is_huge_zero_pmd(orig_pmd))
1641 if (unlikely(!pmd_present(orig_pmd))) {
1642 VM_BUG_ON(thp_migration_supported() &&
1643 !is_pmd_migration_entry(orig_pmd));
1647 page = pmd_page(orig_pmd);
1649 * If other processes are mapping this page, we couldn't discard
1650 * the page unless they all do MADV_FREE so let's skip the page.
1652 if (page_mapcount(page) != 1)
1655 if (!trylock_page(page))
1659 * If user want to discard part-pages of THP, split it so MADV_FREE
1660 * will deactivate only them.
1662 if (next - addr != HPAGE_PMD_SIZE) {
1665 split_huge_page(page);
1671 if (PageDirty(page))
1672 ClearPageDirty(page);
1675 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1676 pmdp_invalidate(vma, addr, pmd);
1677 orig_pmd = pmd_mkold(orig_pmd);
1678 orig_pmd = pmd_mkclean(orig_pmd);
1680 set_pmd_at(mm, addr, pmd, orig_pmd);
1681 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1684 mark_page_lazyfree(page);
1692 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1696 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1697 pte_free(mm, pgtable);
1701 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1702 pmd_t *pmd, unsigned long addr)
1707 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1709 ptl = __pmd_trans_huge_lock(pmd, vma);
1713 * For architectures like ppc64 we look at deposited pgtable
1714 * when calling pmdp_huge_get_and_clear. So do the
1715 * pgtable_trans_huge_withdraw after finishing pmdp related
1718 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1720 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1721 if (vma_is_dax(vma)) {
1722 if (arch_needs_pgtable_deposit())
1723 zap_deposited_table(tlb->mm, pmd);
1725 if (is_huge_zero_pmd(orig_pmd))
1726 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1727 } else if (is_huge_zero_pmd(orig_pmd)) {
1728 zap_deposited_table(tlb->mm, pmd);
1730 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1732 struct page *page = NULL;
1733 int flush_needed = 1;
1735 if (pmd_present(orig_pmd)) {
1736 page = pmd_page(orig_pmd);
1737 page_remove_rmap(page, true);
1738 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1739 VM_BUG_ON_PAGE(!PageHead(page), page);
1740 } else if (thp_migration_supported()) {
1743 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1744 entry = pmd_to_swp_entry(orig_pmd);
1745 page = pfn_to_page(swp_offset(entry));
1748 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1750 if (PageAnon(page)) {
1751 zap_deposited_table(tlb->mm, pmd);
1752 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1754 if (arch_needs_pgtable_deposit())
1755 zap_deposited_table(tlb->mm, pmd);
1756 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1761 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1766 #ifndef pmd_move_must_withdraw
1767 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1768 spinlock_t *old_pmd_ptl,
1769 struct vm_area_struct *vma)
1772 * With split pmd lock we also need to move preallocated
1773 * PTE page table if new_pmd is on different PMD page table.
1775 * We also don't deposit and withdraw tables for file pages.
1777 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1781 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1783 #ifdef CONFIG_MEM_SOFT_DIRTY
1784 if (unlikely(is_pmd_migration_entry(pmd)))
1785 pmd = pmd_swp_mksoft_dirty(pmd);
1786 else if (pmd_present(pmd))
1787 pmd = pmd_mksoft_dirty(pmd);
1792 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1793 unsigned long new_addr, unsigned long old_end,
1794 pmd_t *old_pmd, pmd_t *new_pmd)
1796 spinlock_t *old_ptl, *new_ptl;
1798 struct mm_struct *mm = vma->vm_mm;
1799 bool force_flush = false;
1801 if ((old_addr & ~HPAGE_PMD_MASK) ||
1802 (new_addr & ~HPAGE_PMD_MASK) ||
1803 old_end - old_addr < HPAGE_PMD_SIZE)
1807 * The destination pmd shouldn't be established, free_pgtables()
1808 * should have release it.
1810 if (WARN_ON(!pmd_none(*new_pmd))) {
1811 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1816 * We don't have to worry about the ordering of src and dst
1817 * ptlocks because exclusive mmap_sem prevents deadlock.
1819 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1821 new_ptl = pmd_lockptr(mm, new_pmd);
1822 if (new_ptl != old_ptl)
1823 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1824 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1825 if (pmd_present(pmd))
1827 VM_BUG_ON(!pmd_none(*new_pmd));
1829 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1831 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1832 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1834 pmd = move_soft_dirty_pmd(pmd);
1835 set_pmd_at(mm, new_addr, new_pmd, pmd);
1837 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1838 if (new_ptl != old_ptl)
1839 spin_unlock(new_ptl);
1840 spin_unlock(old_ptl);
1848 * - 0 if PMD could not be locked
1849 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1850 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1852 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1853 unsigned long addr, pgprot_t newprot, int prot_numa)
1855 struct mm_struct *mm = vma->vm_mm;
1858 bool preserve_write;
1861 ptl = __pmd_trans_huge_lock(pmd, vma);
1865 preserve_write = prot_numa && pmd_write(*pmd);
1868 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1869 if (is_swap_pmd(*pmd)) {
1870 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1872 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1873 if (is_write_migration_entry(entry)) {
1876 * A protection check is difficult so
1877 * just be safe and disable write
1879 make_migration_entry_read(&entry);
1880 newpmd = swp_entry_to_pmd(entry);
1881 if (pmd_swp_soft_dirty(*pmd))
1882 newpmd = pmd_swp_mksoft_dirty(newpmd);
1883 set_pmd_at(mm, addr, pmd, newpmd);
1890 * Avoid trapping faults against the zero page. The read-only
1891 * data is likely to be read-cached on the local CPU and
1892 * local/remote hits to the zero page are not interesting.
1894 if (prot_numa && is_huge_zero_pmd(*pmd))
1897 if (prot_numa && pmd_protnone(*pmd))
1901 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1902 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1903 * which is also under down_read(mmap_sem):
1906 * change_huge_pmd(prot_numa=1)
1907 * pmdp_huge_get_and_clear_notify()
1908 * madvise_dontneed()
1910 * pmd_trans_huge(*pmd) == 0 (without ptl)
1913 * // pmd is re-established
1915 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1916 * which may break userspace.
1918 * pmdp_invalidate() is required to make sure we don't miss
1919 * dirty/young flags set by hardware.
1921 entry = pmdp_invalidate(vma, addr, pmd);
1923 entry = pmd_modify(entry, newprot);
1925 entry = pmd_mk_savedwrite(entry);
1927 set_pmd_at(mm, addr, pmd, entry);
1928 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1935 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1937 * Note that if it returns page table lock pointer, this routine returns without
1938 * unlocking page table lock. So callers must unlock it.
1940 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1943 ptl = pmd_lock(vma->vm_mm, pmd);
1944 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1952 * Returns true if a given pud maps a thp, false otherwise.
1954 * Note that if it returns true, this routine returns without unlocking page
1955 * table lock. So callers must unlock it.
1957 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1961 ptl = pud_lock(vma->vm_mm, pud);
1962 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1968 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1969 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1970 pud_t *pud, unsigned long addr)
1975 ptl = __pud_trans_huge_lock(pud, vma);
1979 * For architectures like ppc64 we look at deposited pgtable
1980 * when calling pudp_huge_get_and_clear. So do the
1981 * pgtable_trans_huge_withdraw after finishing pudp related
1984 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1986 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1987 if (vma_is_dax(vma)) {
1989 /* No zero page support yet */
1991 /* No support for anonymous PUD pages yet */
1997 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1998 unsigned long haddr)
2000 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2001 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2002 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2003 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2005 count_vm_event(THP_SPLIT_PUD);
2007 pudp_huge_clear_flush_notify(vma, haddr, pud);
2010 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2011 unsigned long address)
2014 struct mmu_notifier_range range;
2016 mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PUD_MASK,
2017 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2018 mmu_notifier_invalidate_range_start(&range);
2019 ptl = pud_lock(vma->vm_mm, pud);
2020 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2022 __split_huge_pud_locked(vma, pud, range.start);
2027 * No need to double call mmu_notifier->invalidate_range() callback as
2028 * the above pudp_huge_clear_flush_notify() did already call it.
2030 mmu_notifier_invalidate_range_only_end(&range);
2032 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2034 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2035 unsigned long haddr, pmd_t *pmd)
2037 struct mm_struct *mm = vma->vm_mm;
2043 * Leave pmd empty until pte is filled note that it is fine to delay
2044 * notification until mmu_notifier_invalidate_range_end() as we are
2045 * replacing a zero pmd write protected page with a zero pte write
2048 * See Documentation/vm/mmu_notifier.rst
2050 pmdp_huge_clear_flush(vma, haddr, pmd);
2052 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2053 pmd_populate(mm, &_pmd, pgtable);
2055 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2057 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2058 entry = pte_mkspecial(entry);
2059 pte = pte_offset_map(&_pmd, haddr);
2060 VM_BUG_ON(!pte_none(*pte));
2061 set_pte_at(mm, haddr, pte, entry);
2064 smp_wmb(); /* make pte visible before pmd */
2065 pmd_populate(mm, pmd, pgtable);
2068 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2069 unsigned long haddr, bool freeze)
2071 struct mm_struct *mm = vma->vm_mm;
2074 pmd_t old_pmd, _pmd;
2075 bool young, write, soft_dirty, pmd_migration = false;
2079 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2080 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2081 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2082 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2083 && !pmd_devmap(*pmd));
2085 count_vm_event(THP_SPLIT_PMD);
2087 if (!vma_is_anonymous(vma)) {
2088 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2090 * We are going to unmap this huge page. So
2091 * just go ahead and zap it
2093 if (arch_needs_pgtable_deposit())
2094 zap_deposited_table(mm, pmd);
2095 if (vma_is_dax(vma))
2097 page = pmd_page(_pmd);
2098 if (!PageDirty(page) && pmd_dirty(_pmd))
2099 set_page_dirty(page);
2100 if (!PageReferenced(page) && pmd_young(_pmd))
2101 SetPageReferenced(page);
2102 page_remove_rmap(page, true);
2104 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2106 } else if (is_huge_zero_pmd(*pmd)) {
2108 * FIXME: Do we want to invalidate secondary mmu by calling
2109 * mmu_notifier_invalidate_range() see comments below inside
2110 * __split_huge_pmd() ?
2112 * We are going from a zero huge page write protected to zero
2113 * small page also write protected so it does not seems useful
2114 * to invalidate secondary mmu at this time.
2116 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2120 * Up to this point the pmd is present and huge and userland has the
2121 * whole access to the hugepage during the split (which happens in
2122 * place). If we overwrite the pmd with the not-huge version pointing
2123 * to the pte here (which of course we could if all CPUs were bug
2124 * free), userland could trigger a small page size TLB miss on the
2125 * small sized TLB while the hugepage TLB entry is still established in
2126 * the huge TLB. Some CPU doesn't like that.
2127 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2128 * 383 on page 93. Intel should be safe but is also warns that it's
2129 * only safe if the permission and cache attributes of the two entries
2130 * loaded in the two TLB is identical (which should be the case here).
2131 * But it is generally safer to never allow small and huge TLB entries
2132 * for the same virtual address to be loaded simultaneously. So instead
2133 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2134 * current pmd notpresent (atomically because here the pmd_trans_huge
2135 * must remain set at all times on the pmd until the split is complete
2136 * for this pmd), then we flush the SMP TLB and finally we write the
2137 * non-huge version of the pmd entry with pmd_populate.
2139 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2141 pmd_migration = is_pmd_migration_entry(old_pmd);
2142 if (unlikely(pmd_migration)) {
2145 entry = pmd_to_swp_entry(old_pmd);
2146 page = pfn_to_page(swp_offset(entry));
2147 write = is_write_migration_entry(entry);
2149 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2151 page = pmd_page(old_pmd);
2152 if (pmd_dirty(old_pmd))
2154 write = pmd_write(old_pmd);
2155 young = pmd_young(old_pmd);
2156 soft_dirty = pmd_soft_dirty(old_pmd);
2158 VM_BUG_ON_PAGE(!page_count(page), page);
2159 page_ref_add(page, HPAGE_PMD_NR - 1);
2162 * Withdraw the table only after we mark the pmd entry invalid.
2163 * This's critical for some architectures (Power).
2165 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2166 pmd_populate(mm, &_pmd, pgtable);
2168 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2171 * Note that NUMA hinting access restrictions are not
2172 * transferred to avoid any possibility of altering
2173 * permissions across VMAs.
2175 if (freeze || pmd_migration) {
2176 swp_entry_t swp_entry;
2177 swp_entry = make_migration_entry(page + i, write);
2178 entry = swp_entry_to_pte(swp_entry);
2180 entry = pte_swp_mksoft_dirty(entry);
2182 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2183 entry = maybe_mkwrite(entry, vma);
2185 entry = pte_wrprotect(entry);
2187 entry = pte_mkold(entry);
2189 entry = pte_mksoft_dirty(entry);
2191 pte = pte_offset_map(&_pmd, addr);
2192 BUG_ON(!pte_none(*pte));
2193 set_pte_at(mm, addr, pte, entry);
2194 atomic_inc(&page[i]._mapcount);
2199 * Set PG_double_map before dropping compound_mapcount to avoid
2200 * false-negative page_mapped().
2202 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2203 for (i = 0; i < HPAGE_PMD_NR; i++)
2204 atomic_inc(&page[i]._mapcount);
2207 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2208 /* Last compound_mapcount is gone. */
2209 __dec_node_page_state(page, NR_ANON_THPS);
2210 if (TestClearPageDoubleMap(page)) {
2211 /* No need in mapcount reference anymore */
2212 for (i = 0; i < HPAGE_PMD_NR; i++)
2213 atomic_dec(&page[i]._mapcount);
2217 smp_wmb(); /* make pte visible before pmd */
2218 pmd_populate(mm, pmd, pgtable);
2221 for (i = 0; i < HPAGE_PMD_NR; i++) {
2222 page_remove_rmap(page + i, false);
2228 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2229 unsigned long address, bool freeze, struct page *page)
2232 struct mmu_notifier_range range;
2234 mmu_notifier_range_init(&range, vma->vm_mm, address & HPAGE_PMD_MASK,
2235 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2236 mmu_notifier_invalidate_range_start(&range);
2237 ptl = pmd_lock(vma->vm_mm, pmd);
2240 * If caller asks to setup a migration entries, we need a page to check
2241 * pmd against. Otherwise we can end up replacing wrong page.
2243 VM_BUG_ON(freeze && !page);
2244 if (page && page != pmd_page(*pmd))
2247 if (pmd_trans_huge(*pmd)) {
2248 page = pmd_page(*pmd);
2249 if (PageMlocked(page))
2250 clear_page_mlock(page);
2251 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2253 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2257 * No need to double call mmu_notifier->invalidate_range() callback.
2258 * They are 3 cases to consider inside __split_huge_pmd_locked():
2259 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2260 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2261 * fault will trigger a flush_notify before pointing to a new page
2262 * (it is fine if the secondary mmu keeps pointing to the old zero
2263 * page in the meantime)
2264 * 3) Split a huge pmd into pte pointing to the same page. No need
2265 * to invalidate secondary tlb entry they are all still valid.
2266 * any further changes to individual pte will notify. So no need
2267 * to call mmu_notifier->invalidate_range()
2269 mmu_notifier_invalidate_range_only_end(&range);
2272 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2273 bool freeze, struct page *page)
2280 pgd = pgd_offset(vma->vm_mm, address);
2281 if (!pgd_present(*pgd))
2284 p4d = p4d_offset(pgd, address);
2285 if (!p4d_present(*p4d))
2288 pud = pud_offset(p4d, address);
2289 if (!pud_present(*pud))
2292 pmd = pmd_offset(pud, address);
2294 __split_huge_pmd(vma, pmd, address, freeze, page);
2297 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2298 unsigned long start,
2303 * If the new start address isn't hpage aligned and it could
2304 * previously contain an hugepage: check if we need to split
2307 if (start & ~HPAGE_PMD_MASK &&
2308 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2309 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2310 split_huge_pmd_address(vma, start, false, NULL);
2313 * If the new end address isn't hpage aligned and it could
2314 * previously contain an hugepage: check if we need to split
2317 if (end & ~HPAGE_PMD_MASK &&
2318 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2319 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2320 split_huge_pmd_address(vma, end, false, NULL);
2323 * If we're also updating the vma->vm_next->vm_start, if the new
2324 * vm_next->vm_start isn't page aligned and it could previously
2325 * contain an hugepage: check if we need to split an huge pmd.
2327 if (adjust_next > 0) {
2328 struct vm_area_struct *next = vma->vm_next;
2329 unsigned long nstart = next->vm_start;
2330 nstart += adjust_next << PAGE_SHIFT;
2331 if (nstart & ~HPAGE_PMD_MASK &&
2332 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2333 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2334 split_huge_pmd_address(next, nstart, false, NULL);
2338 static void unmap_page(struct page *page)
2340 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2341 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2344 VM_BUG_ON_PAGE(!PageHead(page), page);
2347 ttu_flags |= TTU_SPLIT_FREEZE;
2349 unmap_success = try_to_unmap(page, ttu_flags);
2350 VM_BUG_ON_PAGE(!unmap_success, page);
2353 static void remap_page(struct page *page)
2356 if (PageTransHuge(page)) {
2357 remove_migration_ptes(page, page, true);
2359 for (i = 0; i < HPAGE_PMD_NR; i++)
2360 remove_migration_ptes(page + i, page + i, true);
2364 static void __split_huge_page_tail(struct page *head, int tail,
2365 struct lruvec *lruvec, struct list_head *list)
2367 struct page *page_tail = head + tail;
2369 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2372 * Clone page flags before unfreezing refcount.
2374 * After successful get_page_unless_zero() might follow flags change,
2375 * for exmaple lock_page() which set PG_waiters.
2377 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2378 page_tail->flags |= (head->flags &
2379 ((1L << PG_referenced) |
2380 (1L << PG_swapbacked) |
2381 (1L << PG_swapcache) |
2382 (1L << PG_mlocked) |
2383 (1L << PG_uptodate) |
2385 (1L << PG_workingset) |
2387 (1L << PG_unevictable) |
2390 /* ->mapping in first tail page is compound_mapcount */
2391 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2393 page_tail->mapping = head->mapping;
2394 page_tail->index = head->index + tail;
2396 /* Page flags must be visible before we make the page non-compound. */
2400 * Clear PageTail before unfreezing page refcount.
2402 * After successful get_page_unless_zero() might follow put_page()
2403 * which needs correct compound_head().
2405 clear_compound_head(page_tail);
2407 /* Finally unfreeze refcount. Additional reference from page cache. */
2408 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2409 PageSwapCache(head)));
2411 if (page_is_young(head))
2412 set_page_young(page_tail);
2413 if (page_is_idle(head))
2414 set_page_idle(page_tail);
2416 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2419 * always add to the tail because some iterators expect new
2420 * pages to show after the currently processed elements - e.g.
2423 lru_add_page_tail(head, page_tail, lruvec, list);
2426 static void __split_huge_page(struct page *page, struct list_head *list,
2427 pgoff_t end, unsigned long flags)
2429 struct page *head = compound_head(page);
2430 struct zone *zone = page_zone(head);
2431 struct lruvec *lruvec;
2434 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2436 /* complete memcg works before add pages to LRU */
2437 mem_cgroup_split_huge_fixup(head);
2439 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2440 __split_huge_page_tail(head, i, lruvec, list);
2441 /* Some pages can be beyond i_size: drop them from page cache */
2442 if (head[i].index >= end) {
2443 ClearPageDirty(head + i);
2444 __delete_from_page_cache(head + i, NULL);
2445 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2446 shmem_uncharge(head->mapping->host, 1);
2451 ClearPageCompound(head);
2452 /* See comment in __split_huge_page_tail() */
2453 if (PageAnon(head)) {
2454 /* Additional pin to swap cache */
2455 if (PageSwapCache(head))
2456 page_ref_add(head, 2);
2460 /* Additional pin to page cache */
2461 page_ref_add(head, 2);
2462 xa_unlock(&head->mapping->i_pages);
2465 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2469 for (i = 0; i < HPAGE_PMD_NR; i++) {
2470 struct page *subpage = head + i;
2471 if (subpage == page)
2473 unlock_page(subpage);
2476 * Subpages may be freed if there wasn't any mapping
2477 * like if add_to_swap() is running on a lru page that
2478 * had its mapping zapped. And freeing these pages
2479 * requires taking the lru_lock so we do the put_page
2480 * of the tail pages after the split is complete.
2486 int total_mapcount(struct page *page)
2488 int i, compound, ret;
2490 VM_BUG_ON_PAGE(PageTail(page), page);
2492 if (likely(!PageCompound(page)))
2493 return atomic_read(&page->_mapcount) + 1;
2495 compound = compound_mapcount(page);
2499 for (i = 0; i < HPAGE_PMD_NR; i++)
2500 ret += atomic_read(&page[i]._mapcount) + 1;
2501 /* File pages has compound_mapcount included in _mapcount */
2502 if (!PageAnon(page))
2503 return ret - compound * HPAGE_PMD_NR;
2504 if (PageDoubleMap(page))
2505 ret -= HPAGE_PMD_NR;
2510 * This calculates accurately how many mappings a transparent hugepage
2511 * has (unlike page_mapcount() which isn't fully accurate). This full
2512 * accuracy is primarily needed to know if copy-on-write faults can
2513 * reuse the page and change the mapping to read-write instead of
2514 * copying them. At the same time this returns the total_mapcount too.
2516 * The function returns the highest mapcount any one of the subpages
2517 * has. If the return value is one, even if different processes are
2518 * mapping different subpages of the transparent hugepage, they can
2519 * all reuse it, because each process is reusing a different subpage.
2521 * The total_mapcount is instead counting all virtual mappings of the
2522 * subpages. If the total_mapcount is equal to "one", it tells the
2523 * caller all mappings belong to the same "mm" and in turn the
2524 * anon_vma of the transparent hugepage can become the vma->anon_vma
2525 * local one as no other process may be mapping any of the subpages.
2527 * It would be more accurate to replace page_mapcount() with
2528 * page_trans_huge_mapcount(), however we only use
2529 * page_trans_huge_mapcount() in the copy-on-write faults where we
2530 * need full accuracy to avoid breaking page pinning, because
2531 * page_trans_huge_mapcount() is slower than page_mapcount().
2533 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2535 int i, ret, _total_mapcount, mapcount;
2537 /* hugetlbfs shouldn't call it */
2538 VM_BUG_ON_PAGE(PageHuge(page), page);
2540 if (likely(!PageTransCompound(page))) {
2541 mapcount = atomic_read(&page->_mapcount) + 1;
2543 *total_mapcount = mapcount;
2547 page = compound_head(page);
2549 _total_mapcount = ret = 0;
2550 for (i = 0; i < HPAGE_PMD_NR; i++) {
2551 mapcount = atomic_read(&page[i]._mapcount) + 1;
2552 ret = max(ret, mapcount);
2553 _total_mapcount += mapcount;
2555 if (PageDoubleMap(page)) {
2557 _total_mapcount -= HPAGE_PMD_NR;
2559 mapcount = compound_mapcount(page);
2561 _total_mapcount += mapcount;
2563 *total_mapcount = _total_mapcount;
2567 /* Racy check whether the huge page can be split */
2568 bool can_split_huge_page(struct page *page, int *pextra_pins)
2572 /* Additional pins from page cache */
2574 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2576 extra_pins = HPAGE_PMD_NR;
2578 *pextra_pins = extra_pins;
2579 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2583 * This function splits huge page into normal pages. @page can point to any
2584 * subpage of huge page to split. Split doesn't change the position of @page.
2586 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2587 * The huge page must be locked.
2589 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2591 * Both head page and tail pages will inherit mapping, flags, and so on from
2594 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2595 * they are not mapped.
2597 * Returns 0 if the hugepage is split successfully.
2598 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2601 int split_huge_page_to_list(struct page *page, struct list_head *list)
2603 struct page *head = compound_head(page);
2604 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2605 struct anon_vma *anon_vma = NULL;
2606 struct address_space *mapping = NULL;
2607 int count, mapcount, extra_pins, ret;
2609 unsigned long flags;
2612 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2613 VM_BUG_ON_PAGE(!PageLocked(page), page);
2614 VM_BUG_ON_PAGE(!PageCompound(page), page);
2616 if (PageWriteback(page))
2619 if (PageAnon(head)) {
2621 * The caller does not necessarily hold an mmap_sem that would
2622 * prevent the anon_vma disappearing so we first we take a
2623 * reference to it and then lock the anon_vma for write. This
2624 * is similar to page_lock_anon_vma_read except the write lock
2625 * is taken to serialise against parallel split or collapse
2628 anon_vma = page_get_anon_vma(head);
2635 anon_vma_lock_write(anon_vma);
2637 mapping = head->mapping;
2646 i_mmap_lock_read(mapping);
2649 *__split_huge_page() may need to trim off pages beyond EOF:
2650 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2651 * which cannot be nested inside the page tree lock. So note
2652 * end now: i_size itself may be changed at any moment, but
2653 * head page lock is good enough to serialize the trimming.
2655 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2659 * Racy check if we can split the page, before unmap_page() will
2662 if (!can_split_huge_page(head, &extra_pins)) {
2667 mlocked = PageMlocked(page);
2669 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2671 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2675 /* prevent PageLRU to go away from under us, and freeze lru stats */
2676 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2679 XA_STATE(xas, &mapping->i_pages, page_index(head));
2682 * Check if the head page is present in page cache.
2683 * We assume all tail are present too, if head is there.
2685 xa_lock(&mapping->i_pages);
2686 if (xas_load(&xas) != head)
2690 /* Prevent deferred_split_scan() touching ->_refcount */
2691 spin_lock(&pgdata->split_queue_lock);
2692 count = page_count(head);
2693 mapcount = total_mapcount(head);
2694 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2695 if (!list_empty(page_deferred_list(head))) {
2696 pgdata->split_queue_len--;
2697 list_del(page_deferred_list(head));
2700 __dec_node_page_state(page, NR_SHMEM_THPS);
2701 spin_unlock(&pgdata->split_queue_lock);
2702 __split_huge_page(page, list, end, flags);
2703 if (PageSwapCache(head)) {
2704 swp_entry_t entry = { .val = page_private(head) };
2706 ret = split_swap_cluster(entry);
2710 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2711 pr_alert("total_mapcount: %u, page_count(): %u\n",
2714 dump_page(head, NULL);
2715 dump_page(page, "total_mapcount(head) > 0");
2718 spin_unlock(&pgdata->split_queue_lock);
2720 xa_unlock(&mapping->i_pages);
2721 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2728 anon_vma_unlock_write(anon_vma);
2729 put_anon_vma(anon_vma);
2732 i_mmap_unlock_read(mapping);
2734 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2738 void free_transhuge_page(struct page *page)
2740 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2741 unsigned long flags;
2743 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2744 if (!list_empty(page_deferred_list(page))) {
2745 pgdata->split_queue_len--;
2746 list_del(page_deferred_list(page));
2748 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2749 free_compound_page(page);
2752 void deferred_split_huge_page(struct page *page)
2754 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2755 unsigned long flags;
2757 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2759 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2760 if (list_empty(page_deferred_list(page))) {
2761 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2762 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2763 pgdata->split_queue_len++;
2765 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2768 static unsigned long deferred_split_count(struct shrinker *shrink,
2769 struct shrink_control *sc)
2771 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2772 return READ_ONCE(pgdata->split_queue_len);
2775 static unsigned long deferred_split_scan(struct shrinker *shrink,
2776 struct shrink_control *sc)
2778 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2779 unsigned long flags;
2780 LIST_HEAD(list), *pos, *next;
2784 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2785 /* Take pin on all head pages to avoid freeing them under us */
2786 list_for_each_safe(pos, next, &pgdata->split_queue) {
2787 page = list_entry((void *)pos, struct page, mapping);
2788 page = compound_head(page);
2789 if (get_page_unless_zero(page)) {
2790 list_move(page_deferred_list(page), &list);
2792 /* We lost race with put_compound_page() */
2793 list_del_init(page_deferred_list(page));
2794 pgdata->split_queue_len--;
2796 if (!--sc->nr_to_scan)
2799 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2801 list_for_each_safe(pos, next, &list) {
2802 page = list_entry((void *)pos, struct page, mapping);
2803 if (!trylock_page(page))
2805 /* split_huge_page() removes page from list on success */
2806 if (!split_huge_page(page))
2813 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2814 list_splice_tail(&list, &pgdata->split_queue);
2815 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2818 * Stop shrinker if we didn't split any page, but the queue is empty.
2819 * This can happen if pages were freed under us.
2821 if (!split && list_empty(&pgdata->split_queue))
2826 static struct shrinker deferred_split_shrinker = {
2827 .count_objects = deferred_split_count,
2828 .scan_objects = deferred_split_scan,
2829 .seeks = DEFAULT_SEEKS,
2830 .flags = SHRINKER_NUMA_AWARE,
2833 #ifdef CONFIG_DEBUG_FS
2834 static int split_huge_pages_set(void *data, u64 val)
2838 unsigned long pfn, max_zone_pfn;
2839 unsigned long total = 0, split = 0;
2844 for_each_populated_zone(zone) {
2845 max_zone_pfn = zone_end_pfn(zone);
2846 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2847 if (!pfn_valid(pfn))
2850 page = pfn_to_page(pfn);
2851 if (!get_page_unless_zero(page))
2854 if (zone != page_zone(page))
2857 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2862 if (!split_huge_page(page))
2870 pr_info("%lu of %lu THP split\n", split, total);
2874 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2877 static int __init split_huge_pages_debugfs(void)
2881 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2882 &split_huge_pages_fops);
2884 pr_warn("Failed to create split_huge_pages in debugfs");
2887 late_initcall(split_huge_pages_debugfs);
2890 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2891 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2894 struct vm_area_struct *vma = pvmw->vma;
2895 struct mm_struct *mm = vma->vm_mm;
2896 unsigned long address = pvmw->address;
2901 if (!(pvmw->pmd && !pvmw->pte))
2904 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2905 pmdval = *pvmw->pmd;
2906 pmdp_invalidate(vma, address, pvmw->pmd);
2907 if (pmd_dirty(pmdval))
2908 set_page_dirty(page);
2909 entry = make_migration_entry(page, pmd_write(pmdval));
2910 pmdswp = swp_entry_to_pmd(entry);
2911 if (pmd_soft_dirty(pmdval))
2912 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2913 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2914 page_remove_rmap(page, true);
2918 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2920 struct vm_area_struct *vma = pvmw->vma;
2921 struct mm_struct *mm = vma->vm_mm;
2922 unsigned long address = pvmw->address;
2923 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2927 if (!(pvmw->pmd && !pvmw->pte))
2930 entry = pmd_to_swp_entry(*pvmw->pmd);
2932 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2933 if (pmd_swp_soft_dirty(*pvmw->pmd))
2934 pmde = pmd_mksoft_dirty(pmde);
2935 if (is_write_migration_entry(entry))
2936 pmde = maybe_pmd_mkwrite(pmde, vma);
2938 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2940 page_add_anon_rmap(new, vma, mmun_start, true);
2942 page_add_file_rmap(new, true);
2943 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2944 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
2945 mlock_vma_page(new);
2946 update_mmu_cache_pmd(vma, address, pvmw->pmd);