1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2009 Red Hat, Inc.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/sched.h>
10 #include <linux/sched/coredump.h>
11 #include <linux/sched/numa_balancing.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/pfn_t.h>
24 #include <linux/mman.h>
25 #include <linux/memremap.h>
26 #include <linux/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/oom.h>
34 #include <linux/numa.h>
35 #include <linux/page_owner.h>
38 #include <asm/pgalloc.h>
42 * By default, transparent hugepage support is disabled in order to avoid
43 * risking an increased memory footprint for applications that are not
44 * guaranteed to benefit from it. When transparent hugepage support is
45 * enabled, it is for all mappings, and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
49 unsigned long transparent_hugepage_flags __read_mostly =
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
60 static struct shrinker deferred_split_shrinker;
62 static atomic_t huge_zero_refcount;
63 struct page *huge_zero_page __read_mostly;
65 bool transparent_hugepage_enabled(struct vm_area_struct *vma)
67 /* The addr is used to check if the vma size fits */
68 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
70 if (!transhuge_vma_suitable(vma, addr))
72 if (vma_is_anonymous(vma))
73 return __transparent_hugepage_enabled(vma);
74 if (vma_is_shmem(vma))
75 return shmem_huge_enabled(vma);
80 static struct page *get_huge_zero_page(void)
82 struct page *zero_page;
84 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
85 return READ_ONCE(huge_zero_page);
87 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
90 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
93 count_vm_event(THP_ZERO_PAGE_ALLOC);
95 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
97 __free_pages(zero_page, compound_order(zero_page));
101 /* We take additional reference here. It will be put back by shrinker */
102 atomic_set(&huge_zero_refcount, 2);
104 return READ_ONCE(huge_zero_page);
107 static void put_huge_zero_page(void)
110 * Counter should never go to zero here. Only shrinker can put
113 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
116 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
118 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
119 return READ_ONCE(huge_zero_page);
121 if (!get_huge_zero_page())
124 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
125 put_huge_zero_page();
127 return READ_ONCE(huge_zero_page);
130 void mm_put_huge_zero_page(struct mm_struct *mm)
132 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
133 put_huge_zero_page();
136 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
137 struct shrink_control *sc)
139 /* we can free zero page only if last reference remains */
140 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
143 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
144 struct shrink_control *sc)
146 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
147 struct page *zero_page = xchg(&huge_zero_page, NULL);
148 BUG_ON(zero_page == NULL);
149 __free_pages(zero_page, compound_order(zero_page));
156 static struct shrinker huge_zero_page_shrinker = {
157 .count_objects = shrink_huge_zero_page_count,
158 .scan_objects = shrink_huge_zero_page_scan,
159 .seeks = DEFAULT_SEEKS,
163 static ssize_t enabled_show(struct kobject *kobj,
164 struct kobj_attribute *attr, char *buf)
166 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
167 return sprintf(buf, "[always] madvise never\n");
168 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
169 return sprintf(buf, "always [madvise] never\n");
171 return sprintf(buf, "always madvise [never]\n");
174 static ssize_t enabled_store(struct kobject *kobj,
175 struct kobj_attribute *attr,
176 const char *buf, size_t count)
180 if (!memcmp("always", buf,
181 min(sizeof("always")-1, count))) {
182 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
183 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
184 } else if (!memcmp("madvise", buf,
185 min(sizeof("madvise")-1, count))) {
186 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
187 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
188 } else if (!memcmp("never", buf,
189 min(sizeof("never")-1, count))) {
190 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
191 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
196 int err = start_stop_khugepaged();
202 static struct kobj_attribute enabled_attr =
203 __ATTR(enabled, 0644, enabled_show, enabled_store);
205 ssize_t single_hugepage_flag_show(struct kobject *kobj,
206 struct kobj_attribute *attr, char *buf,
207 enum transparent_hugepage_flag flag)
209 return sprintf(buf, "%d\n",
210 !!test_bit(flag, &transparent_hugepage_flags));
213 ssize_t single_hugepage_flag_store(struct kobject *kobj,
214 struct kobj_attribute *attr,
215 const char *buf, size_t count,
216 enum transparent_hugepage_flag flag)
221 ret = kstrtoul(buf, 10, &value);
228 set_bit(flag, &transparent_hugepage_flags);
230 clear_bit(flag, &transparent_hugepage_flags);
235 static ssize_t defrag_show(struct kobject *kobj,
236 struct kobj_attribute *attr, char *buf)
238 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
239 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
240 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
241 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
242 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
243 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
244 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
245 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
246 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
249 static ssize_t defrag_store(struct kobject *kobj,
250 struct kobj_attribute *attr,
251 const char *buf, size_t count)
253 if (!memcmp("always", buf,
254 min(sizeof("always")-1, count))) {
255 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
256 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
258 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259 } else if (!memcmp("defer+madvise", buf,
260 min(sizeof("defer+madvise")-1, count))) {
261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
264 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
265 } else if (!memcmp("defer", buf,
266 min(sizeof("defer")-1, count))) {
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
269 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
270 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
271 } else if (!memcmp("madvise", buf,
272 min(sizeof("madvise")-1, count))) {
273 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
275 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
276 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
277 } else if (!memcmp("never", buf,
278 min(sizeof("never")-1, count))) {
279 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
280 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
281 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
282 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
288 static struct kobj_attribute defrag_attr =
289 __ATTR(defrag, 0644, defrag_show, defrag_store);
291 static ssize_t use_zero_page_show(struct kobject *kobj,
292 struct kobj_attribute *attr, char *buf)
294 return single_hugepage_flag_show(kobj, attr, buf,
295 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
297 static ssize_t use_zero_page_store(struct kobject *kobj,
298 struct kobj_attribute *attr, const char *buf, size_t count)
300 return single_hugepage_flag_store(kobj, attr, buf, count,
301 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
303 static struct kobj_attribute use_zero_page_attr =
304 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
306 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
307 struct kobj_attribute *attr, char *buf)
309 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
311 static struct kobj_attribute hpage_pmd_size_attr =
312 __ATTR_RO(hpage_pmd_size);
314 #ifdef CONFIG_DEBUG_VM
315 static ssize_t debug_cow_show(struct kobject *kobj,
316 struct kobj_attribute *attr, char *buf)
318 return single_hugepage_flag_show(kobj, attr, buf,
319 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
321 static ssize_t debug_cow_store(struct kobject *kobj,
322 struct kobj_attribute *attr,
323 const char *buf, size_t count)
325 return single_hugepage_flag_store(kobj, attr, buf, count,
326 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
328 static struct kobj_attribute debug_cow_attr =
329 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
330 #endif /* CONFIG_DEBUG_VM */
332 static struct attribute *hugepage_attr[] = {
335 &use_zero_page_attr.attr,
336 &hpage_pmd_size_attr.attr,
337 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
338 &shmem_enabled_attr.attr,
340 #ifdef CONFIG_DEBUG_VM
341 &debug_cow_attr.attr,
346 static const struct attribute_group hugepage_attr_group = {
347 .attrs = hugepage_attr,
350 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
354 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
355 if (unlikely(!*hugepage_kobj)) {
356 pr_err("failed to create transparent hugepage kobject\n");
360 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
362 pr_err("failed to register transparent hugepage group\n");
366 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
368 pr_err("failed to register transparent hugepage group\n");
369 goto remove_hp_group;
375 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
377 kobject_put(*hugepage_kobj);
381 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
383 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
384 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
385 kobject_put(hugepage_kobj);
388 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
393 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
396 #endif /* CONFIG_SYSFS */
398 static int __init hugepage_init(void)
401 struct kobject *hugepage_kobj;
403 if (!has_transparent_hugepage()) {
404 transparent_hugepage_flags = 0;
409 * hugepages can't be allocated by the buddy allocator
411 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
413 * we use page->mapping and page->index in second tail page
414 * as list_head: assuming THP order >= 2
416 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
418 err = hugepage_init_sysfs(&hugepage_kobj);
422 err = khugepaged_init();
426 err = register_shrinker(&huge_zero_page_shrinker);
428 goto err_hzp_shrinker;
429 err = register_shrinker(&deferred_split_shrinker);
431 goto err_split_shrinker;
434 * By default disable transparent hugepages on smaller systems,
435 * where the extra memory used could hurt more than TLB overhead
436 * is likely to save. The admin can still enable it through /sys.
438 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
439 transparent_hugepage_flags = 0;
443 err = start_stop_khugepaged();
449 unregister_shrinker(&deferred_split_shrinker);
451 unregister_shrinker(&huge_zero_page_shrinker);
453 khugepaged_destroy();
455 hugepage_exit_sysfs(hugepage_kobj);
459 subsys_initcall(hugepage_init);
461 static int __init setup_transparent_hugepage(char *str)
466 if (!strcmp(str, "always")) {
467 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
468 &transparent_hugepage_flags);
469 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
470 &transparent_hugepage_flags);
472 } else if (!strcmp(str, "madvise")) {
473 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
474 &transparent_hugepage_flags);
475 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
476 &transparent_hugepage_flags);
478 } else if (!strcmp(str, "never")) {
479 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
480 &transparent_hugepage_flags);
481 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
482 &transparent_hugepage_flags);
487 pr_warn("transparent_hugepage= cannot parse, ignored\n");
490 __setup("transparent_hugepage=", setup_transparent_hugepage);
492 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
494 if (likely(vma->vm_flags & VM_WRITE))
495 pmd = pmd_mkwrite(pmd);
499 static inline struct list_head *page_deferred_list(struct page *page)
501 /* ->lru in the tail pages is occupied by compound_head. */
502 return &page[2].deferred_list;
505 void prep_transhuge_page(struct page *page)
508 * we use page->mapping and page->indexlru in second tail page
509 * as list_head: assuming THP order >= 2
512 INIT_LIST_HEAD(page_deferred_list(page));
513 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
516 static unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
517 loff_t off, unsigned long flags, unsigned long size)
520 loff_t off_end = off + len;
521 loff_t off_align = round_up(off, size);
522 unsigned long len_pad;
524 if (off_end <= off_align || (off_end - off_align) < size)
527 len_pad = len + size;
528 if (len_pad < len || (off + len_pad) < off)
531 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
532 off >> PAGE_SHIFT, flags);
533 if (IS_ERR_VALUE(addr))
536 addr += (off - addr) & (size - 1);
540 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
541 unsigned long len, unsigned long pgoff, unsigned long flags)
543 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
547 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
550 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
555 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
557 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
559 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
560 struct page *page, gfp_t gfp)
562 struct vm_area_struct *vma = vmf->vma;
563 struct mem_cgroup *memcg;
565 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
568 VM_BUG_ON_PAGE(!PageCompound(page), page);
570 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, gfp, &memcg, true)) {
572 count_vm_event(THP_FAULT_FALLBACK);
573 return VM_FAULT_FALLBACK;
576 pgtable = pte_alloc_one(vma->vm_mm);
577 if (unlikely(!pgtable)) {
582 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
584 * The memory barrier inside __SetPageUptodate makes sure that
585 * clear_huge_page writes become visible before the set_pmd_at()
588 __SetPageUptodate(page);
590 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
591 if (unlikely(!pmd_none(*vmf->pmd))) {
596 ret = check_stable_address_space(vma->vm_mm);
600 /* Deliver the page fault to userland */
601 if (userfaultfd_missing(vma)) {
604 spin_unlock(vmf->ptl);
605 mem_cgroup_cancel_charge(page, memcg, true);
607 pte_free(vma->vm_mm, pgtable);
608 ret2 = handle_userfault(vmf, VM_UFFD_MISSING);
609 VM_BUG_ON(ret2 & VM_FAULT_FALLBACK);
613 entry = mk_huge_pmd(page, vma->vm_page_prot);
614 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
615 page_add_new_anon_rmap(page, vma, haddr, true);
616 mem_cgroup_commit_charge(page, memcg, false, true);
617 lru_cache_add_active_or_unevictable(page, vma);
618 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
619 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
620 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
621 mm_inc_nr_ptes(vma->vm_mm);
622 spin_unlock(vmf->ptl);
623 count_vm_event(THP_FAULT_ALLOC);
624 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
629 spin_unlock(vmf->ptl);
632 pte_free(vma->vm_mm, pgtable);
633 mem_cgroup_cancel_charge(page, memcg, true);
640 * always: directly stall for all thp allocations
641 * defer: wake kswapd and fail if not immediately available
642 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
643 * fail if not immediately available
644 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
646 * never: never stall for any thp allocation
648 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma, unsigned long addr)
650 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
651 const gfp_t gfp_mask = GFP_TRANSHUGE_LIGHT | __GFP_THISNODE;
653 /* Always do synchronous compaction */
654 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
655 return GFP_TRANSHUGE | __GFP_THISNODE |
656 (vma_madvised ? 0 : __GFP_NORETRY);
658 /* Kick kcompactd and fail quickly */
659 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
660 return gfp_mask | __GFP_KSWAPD_RECLAIM;
662 /* Synchronous compaction if madvised, otherwise kick kcompactd */
663 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
664 return gfp_mask | (vma_madvised ? __GFP_DIRECT_RECLAIM :
665 __GFP_KSWAPD_RECLAIM);
667 /* Only do synchronous compaction if madvised */
668 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
669 return gfp_mask | (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
674 /* Caller must hold page table lock. */
675 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
676 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
677 struct page *zero_page)
682 entry = mk_pmd(zero_page, vma->vm_page_prot);
683 entry = pmd_mkhuge(entry);
685 pgtable_trans_huge_deposit(mm, pmd, pgtable);
686 set_pmd_at(mm, haddr, pmd, entry);
691 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
693 struct vm_area_struct *vma = vmf->vma;
696 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
698 if (!transhuge_vma_suitable(vma, haddr))
699 return VM_FAULT_FALLBACK;
700 if (unlikely(anon_vma_prepare(vma)))
702 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
704 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
705 !mm_forbids_zeropage(vma->vm_mm) &&
706 transparent_hugepage_use_zero_page()) {
708 struct page *zero_page;
711 pgtable = pte_alloc_one(vma->vm_mm);
712 if (unlikely(!pgtable))
714 zero_page = mm_get_huge_zero_page(vma->vm_mm);
715 if (unlikely(!zero_page)) {
716 pte_free(vma->vm_mm, pgtable);
717 count_vm_event(THP_FAULT_FALLBACK);
718 return VM_FAULT_FALLBACK;
720 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
723 if (pmd_none(*vmf->pmd)) {
724 ret = check_stable_address_space(vma->vm_mm);
726 spin_unlock(vmf->ptl);
727 } else if (userfaultfd_missing(vma)) {
728 spin_unlock(vmf->ptl);
729 ret = handle_userfault(vmf, VM_UFFD_MISSING);
730 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
732 set_huge_zero_page(pgtable, vma->vm_mm, vma,
733 haddr, vmf->pmd, zero_page);
734 spin_unlock(vmf->ptl);
738 spin_unlock(vmf->ptl);
740 pte_free(vma->vm_mm, pgtable);
743 gfp = alloc_hugepage_direct_gfpmask(vma, haddr);
744 page = alloc_pages_vma(gfp, HPAGE_PMD_ORDER, vma, haddr, numa_node_id());
745 if (unlikely(!page)) {
746 count_vm_event(THP_FAULT_FALLBACK);
747 return VM_FAULT_FALLBACK;
749 prep_transhuge_page(page);
750 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
753 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
754 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
757 struct mm_struct *mm = vma->vm_mm;
761 ptl = pmd_lock(mm, pmd);
762 if (!pmd_none(*pmd)) {
764 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
765 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
768 entry = pmd_mkyoung(*pmd);
769 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
770 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
771 update_mmu_cache_pmd(vma, addr, pmd);
777 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
778 if (pfn_t_devmap(pfn))
779 entry = pmd_mkdevmap(entry);
781 entry = pmd_mkyoung(pmd_mkdirty(entry));
782 entry = maybe_pmd_mkwrite(entry, vma);
786 pgtable_trans_huge_deposit(mm, pmd, pgtable);
791 set_pmd_at(mm, addr, pmd, entry);
792 update_mmu_cache_pmd(vma, addr, pmd);
797 pte_free(mm, pgtable);
800 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write)
802 unsigned long addr = vmf->address & PMD_MASK;
803 struct vm_area_struct *vma = vmf->vma;
804 pgprot_t pgprot = vma->vm_page_prot;
805 pgtable_t pgtable = NULL;
808 * If we had pmd_special, we could avoid all these restrictions,
809 * but we need to be consistent with PTEs and architectures that
810 * can't support a 'special' bit.
812 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
814 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
815 (VM_PFNMAP|VM_MIXEDMAP));
816 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
818 if (addr < vma->vm_start || addr >= vma->vm_end)
819 return VM_FAULT_SIGBUS;
821 if (arch_needs_pgtable_deposit()) {
822 pgtable = pte_alloc_one(vma->vm_mm);
827 track_pfn_insert(vma, &pgprot, pfn);
829 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
830 return VM_FAULT_NOPAGE;
832 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
834 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
835 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
837 if (likely(vma->vm_flags & VM_WRITE))
838 pud = pud_mkwrite(pud);
842 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
843 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
845 struct mm_struct *mm = vma->vm_mm;
849 ptl = pud_lock(mm, pud);
850 if (!pud_none(*pud)) {
852 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
853 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
856 entry = pud_mkyoung(*pud);
857 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
858 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
859 update_mmu_cache_pud(vma, addr, pud);
864 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
865 if (pfn_t_devmap(pfn))
866 entry = pud_mkdevmap(entry);
868 entry = pud_mkyoung(pud_mkdirty(entry));
869 entry = maybe_pud_mkwrite(entry, vma);
871 set_pud_at(mm, addr, pud, entry);
872 update_mmu_cache_pud(vma, addr, pud);
878 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write)
880 unsigned long addr = vmf->address & PUD_MASK;
881 struct vm_area_struct *vma = vmf->vma;
882 pgprot_t pgprot = vma->vm_page_prot;
885 * If we had pud_special, we could avoid all these restrictions,
886 * but we need to be consistent with PTEs and architectures that
887 * can't support a 'special' bit.
889 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
891 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
892 (VM_PFNMAP|VM_MIXEDMAP));
893 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
895 if (addr < vma->vm_start || addr >= vma->vm_end)
896 return VM_FAULT_SIGBUS;
898 track_pfn_insert(vma, &pgprot, pfn);
900 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
901 return VM_FAULT_NOPAGE;
903 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
904 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
906 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
907 pmd_t *pmd, int flags)
911 _pmd = pmd_mkyoung(*pmd);
912 if (flags & FOLL_WRITE)
913 _pmd = pmd_mkdirty(_pmd);
914 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
915 pmd, _pmd, flags & FOLL_WRITE))
916 update_mmu_cache_pmd(vma, addr, pmd);
919 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
920 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
922 unsigned long pfn = pmd_pfn(*pmd);
923 struct mm_struct *mm = vma->vm_mm;
926 assert_spin_locked(pmd_lockptr(mm, pmd));
929 * When we COW a devmap PMD entry, we split it into PTEs, so we should
930 * not be in this function with `flags & FOLL_COW` set.
932 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
934 if (flags & FOLL_WRITE && !pmd_write(*pmd))
937 if (pmd_present(*pmd) && pmd_devmap(*pmd))
942 if (flags & FOLL_TOUCH)
943 touch_pmd(vma, addr, pmd, flags);
946 * device mapped pages can only be returned if the
947 * caller will manage the page reference count.
949 if (!(flags & FOLL_GET))
950 return ERR_PTR(-EEXIST);
952 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
953 *pgmap = get_dev_pagemap(pfn, *pgmap);
955 return ERR_PTR(-EFAULT);
956 page = pfn_to_page(pfn);
962 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
963 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
964 struct vm_area_struct *vma)
966 spinlock_t *dst_ptl, *src_ptl;
967 struct page *src_page;
969 pgtable_t pgtable = NULL;
972 /* Skip if can be re-fill on fault */
973 if (!vma_is_anonymous(vma))
976 pgtable = pte_alloc_one(dst_mm);
977 if (unlikely(!pgtable))
980 dst_ptl = pmd_lock(dst_mm, dst_pmd);
981 src_ptl = pmd_lockptr(src_mm, src_pmd);
982 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
987 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
988 if (unlikely(is_swap_pmd(pmd))) {
989 swp_entry_t entry = pmd_to_swp_entry(pmd);
991 VM_BUG_ON(!is_pmd_migration_entry(pmd));
992 if (is_write_migration_entry(entry)) {
993 make_migration_entry_read(&entry);
994 pmd = swp_entry_to_pmd(entry);
995 if (pmd_swp_soft_dirty(*src_pmd))
996 pmd = pmd_swp_mksoft_dirty(pmd);
997 set_pmd_at(src_mm, addr, src_pmd, pmd);
999 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1000 mm_inc_nr_ptes(dst_mm);
1001 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1002 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1008 if (unlikely(!pmd_trans_huge(pmd))) {
1009 pte_free(dst_mm, pgtable);
1013 * When page table lock is held, the huge zero pmd should not be
1014 * under splitting since we don't split the page itself, only pmd to
1017 if (is_huge_zero_pmd(pmd)) {
1018 struct page *zero_page;
1020 * get_huge_zero_page() will never allocate a new page here,
1021 * since we already have a zero page to copy. It just takes a
1024 zero_page = mm_get_huge_zero_page(dst_mm);
1025 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1031 src_page = pmd_page(pmd);
1032 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1034 page_dup_rmap(src_page, true);
1035 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1036 mm_inc_nr_ptes(dst_mm);
1037 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1039 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1040 pmd = pmd_mkold(pmd_wrprotect(pmd));
1041 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1045 spin_unlock(src_ptl);
1046 spin_unlock(dst_ptl);
1051 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1052 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1053 pud_t *pud, int flags)
1057 _pud = pud_mkyoung(*pud);
1058 if (flags & FOLL_WRITE)
1059 _pud = pud_mkdirty(_pud);
1060 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1061 pud, _pud, flags & FOLL_WRITE))
1062 update_mmu_cache_pud(vma, addr, pud);
1065 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1066 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1068 unsigned long pfn = pud_pfn(*pud);
1069 struct mm_struct *mm = vma->vm_mm;
1072 assert_spin_locked(pud_lockptr(mm, pud));
1074 if (flags & FOLL_WRITE && !pud_write(*pud))
1077 if (pud_present(*pud) && pud_devmap(*pud))
1082 if (flags & FOLL_TOUCH)
1083 touch_pud(vma, addr, pud, flags);
1086 * device mapped pages can only be returned if the
1087 * caller will manage the page reference count.
1089 if (!(flags & FOLL_GET))
1090 return ERR_PTR(-EEXIST);
1092 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1093 *pgmap = get_dev_pagemap(pfn, *pgmap);
1095 return ERR_PTR(-EFAULT);
1096 page = pfn_to_page(pfn);
1102 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1103 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1104 struct vm_area_struct *vma)
1106 spinlock_t *dst_ptl, *src_ptl;
1110 dst_ptl = pud_lock(dst_mm, dst_pud);
1111 src_ptl = pud_lockptr(src_mm, src_pud);
1112 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1116 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1120 * When page table lock is held, the huge zero pud should not be
1121 * under splitting since we don't split the page itself, only pud to
1124 if (is_huge_zero_pud(pud)) {
1125 /* No huge zero pud yet */
1128 pudp_set_wrprotect(src_mm, addr, src_pud);
1129 pud = pud_mkold(pud_wrprotect(pud));
1130 set_pud_at(dst_mm, addr, dst_pud, pud);
1134 spin_unlock(src_ptl);
1135 spin_unlock(dst_ptl);
1139 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1142 unsigned long haddr;
1143 bool write = vmf->flags & FAULT_FLAG_WRITE;
1145 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1146 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1149 entry = pud_mkyoung(orig_pud);
1151 entry = pud_mkdirty(entry);
1152 haddr = vmf->address & HPAGE_PUD_MASK;
1153 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1154 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1157 spin_unlock(vmf->ptl);
1159 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1161 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1164 unsigned long haddr;
1165 bool write = vmf->flags & FAULT_FLAG_WRITE;
1167 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1168 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1171 entry = pmd_mkyoung(orig_pmd);
1173 entry = pmd_mkdirty(entry);
1174 haddr = vmf->address & HPAGE_PMD_MASK;
1175 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1176 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1179 spin_unlock(vmf->ptl);
1182 static vm_fault_t do_huge_pmd_wp_page_fallback(struct vm_fault *vmf,
1183 pmd_t orig_pmd, struct page *page)
1185 struct vm_area_struct *vma = vmf->vma;
1186 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1187 struct mem_cgroup *memcg;
1192 struct page **pages;
1193 struct mmu_notifier_range range;
1195 pages = kmalloc_array(HPAGE_PMD_NR, sizeof(struct page *),
1197 if (unlikely(!pages)) {
1198 ret |= VM_FAULT_OOM;
1202 for (i = 0; i < HPAGE_PMD_NR; i++) {
1203 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1204 vmf->address, page_to_nid(page));
1205 if (unlikely(!pages[i] ||
1206 mem_cgroup_try_charge_delay(pages[i], vma->vm_mm,
1207 GFP_KERNEL, &memcg, false))) {
1211 memcg = (void *)page_private(pages[i]);
1212 set_page_private(pages[i], 0);
1213 mem_cgroup_cancel_charge(pages[i], memcg,
1218 ret |= VM_FAULT_OOM;
1221 set_page_private(pages[i], (unsigned long)memcg);
1224 for (i = 0; i < HPAGE_PMD_NR; i++) {
1225 copy_user_highpage(pages[i], page + i,
1226 haddr + PAGE_SIZE * i, vma);
1227 __SetPageUptodate(pages[i]);
1231 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1232 haddr, haddr + HPAGE_PMD_SIZE);
1233 mmu_notifier_invalidate_range_start(&range);
1235 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1236 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1237 goto out_free_pages;
1238 VM_BUG_ON_PAGE(!PageHead(page), page);
1241 * Leave pmd empty until pte is filled note we must notify here as
1242 * concurrent CPU thread might write to new page before the call to
1243 * mmu_notifier_invalidate_range_end() happens which can lead to a
1244 * device seeing memory write in different order than CPU.
1246 * See Documentation/vm/mmu_notifier.rst
1248 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1250 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1251 pmd_populate(vma->vm_mm, &_pmd, pgtable);
1253 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1255 entry = mk_pte(pages[i], vma->vm_page_prot);
1256 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1257 memcg = (void *)page_private(pages[i]);
1258 set_page_private(pages[i], 0);
1259 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1260 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1261 lru_cache_add_active_or_unevictable(pages[i], vma);
1262 vmf->pte = pte_offset_map(&_pmd, haddr);
1263 VM_BUG_ON(!pte_none(*vmf->pte));
1264 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1265 pte_unmap(vmf->pte);
1269 smp_wmb(); /* make pte visible before pmd */
1270 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1271 page_remove_rmap(page, true);
1272 spin_unlock(vmf->ptl);
1275 * No need to double call mmu_notifier->invalidate_range() callback as
1276 * the above pmdp_huge_clear_flush_notify() did already call it.
1278 mmu_notifier_invalidate_range_only_end(&range);
1280 ret |= VM_FAULT_WRITE;
1287 spin_unlock(vmf->ptl);
1288 mmu_notifier_invalidate_range_end(&range);
1289 for (i = 0; i < HPAGE_PMD_NR; i++) {
1290 memcg = (void *)page_private(pages[i]);
1291 set_page_private(pages[i], 0);
1292 mem_cgroup_cancel_charge(pages[i], memcg, false);
1299 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1301 struct vm_area_struct *vma = vmf->vma;
1302 struct page *page = NULL, *new_page;
1303 struct mem_cgroup *memcg;
1304 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1305 struct mmu_notifier_range range;
1306 gfp_t huge_gfp; /* for allocation and charge */
1309 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1310 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1311 if (is_huge_zero_pmd(orig_pmd))
1313 spin_lock(vmf->ptl);
1314 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1317 page = pmd_page(orig_pmd);
1318 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1320 * We can only reuse the page if nobody else maps the huge page or it's
1323 if (!trylock_page(page)) {
1325 spin_unlock(vmf->ptl);
1327 spin_lock(vmf->ptl);
1328 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1335 if (reuse_swap_page(page, NULL)) {
1337 entry = pmd_mkyoung(orig_pmd);
1338 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1339 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1340 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1341 ret |= VM_FAULT_WRITE;
1347 spin_unlock(vmf->ptl);
1349 if (__transparent_hugepage_enabled(vma) &&
1350 !transparent_hugepage_debug_cow()) {
1351 huge_gfp = alloc_hugepage_direct_gfpmask(vma, haddr);
1352 new_page = alloc_pages_vma(huge_gfp, HPAGE_PMD_ORDER, vma,
1353 haddr, numa_node_id());
1357 if (likely(new_page)) {
1358 prep_transhuge_page(new_page);
1361 split_huge_pmd(vma, vmf->pmd, vmf->address);
1362 ret |= VM_FAULT_FALLBACK;
1364 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1365 if (ret & VM_FAULT_OOM) {
1366 split_huge_pmd(vma, vmf->pmd, vmf->address);
1367 ret |= VM_FAULT_FALLBACK;
1371 count_vm_event(THP_FAULT_FALLBACK);
1375 if (unlikely(mem_cgroup_try_charge_delay(new_page, vma->vm_mm,
1376 huge_gfp, &memcg, true))) {
1378 split_huge_pmd(vma, vmf->pmd, vmf->address);
1381 ret |= VM_FAULT_FALLBACK;
1382 count_vm_event(THP_FAULT_FALLBACK);
1386 count_vm_event(THP_FAULT_ALLOC);
1387 count_memcg_events(memcg, THP_FAULT_ALLOC, 1);
1390 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1392 copy_user_huge_page(new_page, page, vmf->address,
1394 __SetPageUptodate(new_page);
1396 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1397 haddr, haddr + HPAGE_PMD_SIZE);
1398 mmu_notifier_invalidate_range_start(&range);
1400 spin_lock(vmf->ptl);
1403 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1404 spin_unlock(vmf->ptl);
1405 mem_cgroup_cancel_charge(new_page, memcg, true);
1410 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1411 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1412 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1413 page_add_new_anon_rmap(new_page, vma, haddr, true);
1414 mem_cgroup_commit_charge(new_page, memcg, false, true);
1415 lru_cache_add_active_or_unevictable(new_page, vma);
1416 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1417 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1419 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1421 VM_BUG_ON_PAGE(!PageHead(page), page);
1422 page_remove_rmap(page, true);
1425 ret |= VM_FAULT_WRITE;
1427 spin_unlock(vmf->ptl);
1430 * No need to double call mmu_notifier->invalidate_range() callback as
1431 * the above pmdp_huge_clear_flush_notify() did already call it.
1433 mmu_notifier_invalidate_range_only_end(&range);
1437 spin_unlock(vmf->ptl);
1442 * FOLL_FORCE can write to even unwritable pmd's, but only
1443 * after we've gone through a COW cycle and they are dirty.
1445 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1447 return pmd_write(pmd) ||
1448 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1451 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1456 struct mm_struct *mm = vma->vm_mm;
1457 struct page *page = NULL;
1459 assert_spin_locked(pmd_lockptr(mm, pmd));
1461 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1464 /* Avoid dumping huge zero page */
1465 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1466 return ERR_PTR(-EFAULT);
1468 /* Full NUMA hinting faults to serialise migration in fault paths */
1469 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1472 page = pmd_page(*pmd);
1473 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1474 if (flags & FOLL_TOUCH)
1475 touch_pmd(vma, addr, pmd, flags);
1476 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1478 * We don't mlock() pte-mapped THPs. This way we can avoid
1479 * leaking mlocked pages into non-VM_LOCKED VMAs.
1483 * In most cases the pmd is the only mapping of the page as we
1484 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1485 * writable private mappings in populate_vma_page_range().
1487 * The only scenario when we have the page shared here is if we
1488 * mlocking read-only mapping shared over fork(). We skip
1489 * mlocking such pages.
1493 * We can expect PageDoubleMap() to be stable under page lock:
1494 * for file pages we set it in page_add_file_rmap(), which
1495 * requires page to be locked.
1498 if (PageAnon(page) && compound_mapcount(page) != 1)
1500 if (PageDoubleMap(page) || !page->mapping)
1502 if (!trylock_page(page))
1505 if (page->mapping && !PageDoubleMap(page))
1506 mlock_vma_page(page);
1510 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1511 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1512 if (flags & FOLL_GET)
1519 /* NUMA hinting page fault entry point for trans huge pmds */
1520 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1522 struct vm_area_struct *vma = vmf->vma;
1523 struct anon_vma *anon_vma = NULL;
1525 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1526 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1527 int target_nid, last_cpupid = -1;
1529 bool migrated = false;
1533 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1534 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1538 * If there are potential migrations, wait for completion and retry
1539 * without disrupting NUMA hinting information. Do not relock and
1540 * check_same as the page may no longer be mapped.
1542 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1543 page = pmd_page(*vmf->pmd);
1544 if (!get_page_unless_zero(page))
1546 spin_unlock(vmf->ptl);
1547 put_and_wait_on_page_locked(page);
1551 page = pmd_page(pmd);
1552 BUG_ON(is_huge_zero_page(page));
1553 page_nid = page_to_nid(page);
1554 last_cpupid = page_cpupid_last(page);
1555 count_vm_numa_event(NUMA_HINT_FAULTS);
1556 if (page_nid == this_nid) {
1557 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1558 flags |= TNF_FAULT_LOCAL;
1561 /* See similar comment in do_numa_page for explanation */
1562 if (!pmd_savedwrite(pmd))
1563 flags |= TNF_NO_GROUP;
1566 * Acquire the page lock to serialise THP migrations but avoid dropping
1567 * page_table_lock if at all possible
1569 page_locked = trylock_page(page);
1570 target_nid = mpol_misplaced(page, vma, haddr);
1571 if (target_nid == NUMA_NO_NODE) {
1572 /* If the page was locked, there are no parallel migrations */
1577 /* Migration could have started since the pmd_trans_migrating check */
1579 page_nid = NUMA_NO_NODE;
1580 if (!get_page_unless_zero(page))
1582 spin_unlock(vmf->ptl);
1583 put_and_wait_on_page_locked(page);
1588 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1589 * to serialises splits
1592 spin_unlock(vmf->ptl);
1593 anon_vma = page_lock_anon_vma_read(page);
1595 /* Confirm the PMD did not change while page_table_lock was released */
1596 spin_lock(vmf->ptl);
1597 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1600 page_nid = NUMA_NO_NODE;
1604 /* Bail if we fail to protect against THP splits for any reason */
1605 if (unlikely(!anon_vma)) {
1607 page_nid = NUMA_NO_NODE;
1612 * Since we took the NUMA fault, we must have observed the !accessible
1613 * bit. Make sure all other CPUs agree with that, to avoid them
1614 * modifying the page we're about to migrate.
1616 * Must be done under PTL such that we'll observe the relevant
1617 * inc_tlb_flush_pending().
1619 * We are not sure a pending tlb flush here is for a huge page
1620 * mapping or not. Hence use the tlb range variant
1622 if (mm_tlb_flush_pending(vma->vm_mm)) {
1623 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1625 * change_huge_pmd() released the pmd lock before
1626 * invalidating the secondary MMUs sharing the primary
1627 * MMU pagetables (with ->invalidate_range()). The
1628 * mmu_notifier_invalidate_range_end() (which
1629 * internally calls ->invalidate_range()) in
1630 * change_pmd_range() will run after us, so we can't
1631 * rely on it here and we need an explicit invalidate.
1633 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1634 haddr + HPAGE_PMD_SIZE);
1638 * Migrate the THP to the requested node, returns with page unlocked
1639 * and access rights restored.
1641 spin_unlock(vmf->ptl);
1643 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1644 vmf->pmd, pmd, vmf->address, page, target_nid);
1646 flags |= TNF_MIGRATED;
1647 page_nid = target_nid;
1649 flags |= TNF_MIGRATE_FAIL;
1653 BUG_ON(!PageLocked(page));
1654 was_writable = pmd_savedwrite(pmd);
1655 pmd = pmd_modify(pmd, vma->vm_page_prot);
1656 pmd = pmd_mkyoung(pmd);
1658 pmd = pmd_mkwrite(pmd);
1659 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1660 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1663 spin_unlock(vmf->ptl);
1667 page_unlock_anon_vma_read(anon_vma);
1669 if (page_nid != NUMA_NO_NODE)
1670 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1677 * Return true if we do MADV_FREE successfully on entire pmd page.
1678 * Otherwise, return false.
1680 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1681 pmd_t *pmd, unsigned long addr, unsigned long next)
1686 struct mm_struct *mm = tlb->mm;
1689 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1691 ptl = pmd_trans_huge_lock(pmd, vma);
1696 if (is_huge_zero_pmd(orig_pmd))
1699 if (unlikely(!pmd_present(orig_pmd))) {
1700 VM_BUG_ON(thp_migration_supported() &&
1701 !is_pmd_migration_entry(orig_pmd));
1705 page = pmd_page(orig_pmd);
1707 * If other processes are mapping this page, we couldn't discard
1708 * the page unless they all do MADV_FREE so let's skip the page.
1710 if (page_mapcount(page) != 1)
1713 if (!trylock_page(page))
1717 * If user want to discard part-pages of THP, split it so MADV_FREE
1718 * will deactivate only them.
1720 if (next - addr != HPAGE_PMD_SIZE) {
1723 split_huge_page(page);
1729 if (PageDirty(page))
1730 ClearPageDirty(page);
1733 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1734 pmdp_invalidate(vma, addr, pmd);
1735 orig_pmd = pmd_mkold(orig_pmd);
1736 orig_pmd = pmd_mkclean(orig_pmd);
1738 set_pmd_at(mm, addr, pmd, orig_pmd);
1739 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1742 mark_page_lazyfree(page);
1750 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1754 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1755 pte_free(mm, pgtable);
1759 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1760 pmd_t *pmd, unsigned long addr)
1765 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1767 ptl = __pmd_trans_huge_lock(pmd, vma);
1771 * For architectures like ppc64 we look at deposited pgtable
1772 * when calling pmdp_huge_get_and_clear. So do the
1773 * pgtable_trans_huge_withdraw after finishing pmdp related
1776 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1778 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1779 if (vma_is_dax(vma)) {
1780 if (arch_needs_pgtable_deposit())
1781 zap_deposited_table(tlb->mm, pmd);
1783 if (is_huge_zero_pmd(orig_pmd))
1784 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1785 } else if (is_huge_zero_pmd(orig_pmd)) {
1786 zap_deposited_table(tlb->mm, pmd);
1788 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1790 struct page *page = NULL;
1791 int flush_needed = 1;
1793 if (pmd_present(orig_pmd)) {
1794 page = pmd_page(orig_pmd);
1795 page_remove_rmap(page, true);
1796 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1797 VM_BUG_ON_PAGE(!PageHead(page), page);
1798 } else if (thp_migration_supported()) {
1801 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1802 entry = pmd_to_swp_entry(orig_pmd);
1803 page = pfn_to_page(swp_offset(entry));
1806 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1808 if (PageAnon(page)) {
1809 zap_deposited_table(tlb->mm, pmd);
1810 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1812 if (arch_needs_pgtable_deposit())
1813 zap_deposited_table(tlb->mm, pmd);
1814 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1819 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1824 #ifndef pmd_move_must_withdraw
1825 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1826 spinlock_t *old_pmd_ptl,
1827 struct vm_area_struct *vma)
1830 * With split pmd lock we also need to move preallocated
1831 * PTE page table if new_pmd is on different PMD page table.
1833 * We also don't deposit and withdraw tables for file pages.
1835 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1839 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1841 #ifdef CONFIG_MEM_SOFT_DIRTY
1842 if (unlikely(is_pmd_migration_entry(pmd)))
1843 pmd = pmd_swp_mksoft_dirty(pmd);
1844 else if (pmd_present(pmd))
1845 pmd = pmd_mksoft_dirty(pmd);
1850 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1851 unsigned long new_addr, unsigned long old_end,
1852 pmd_t *old_pmd, pmd_t *new_pmd)
1854 spinlock_t *old_ptl, *new_ptl;
1856 struct mm_struct *mm = vma->vm_mm;
1857 bool force_flush = false;
1859 if ((old_addr & ~HPAGE_PMD_MASK) ||
1860 (new_addr & ~HPAGE_PMD_MASK) ||
1861 old_end - old_addr < HPAGE_PMD_SIZE)
1865 * The destination pmd shouldn't be established, free_pgtables()
1866 * should have release it.
1868 if (WARN_ON(!pmd_none(*new_pmd))) {
1869 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1874 * We don't have to worry about the ordering of src and dst
1875 * ptlocks because exclusive mmap_sem prevents deadlock.
1877 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1879 new_ptl = pmd_lockptr(mm, new_pmd);
1880 if (new_ptl != old_ptl)
1881 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1882 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1883 if (pmd_present(pmd))
1885 VM_BUG_ON(!pmd_none(*new_pmd));
1887 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1889 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1890 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1892 pmd = move_soft_dirty_pmd(pmd);
1893 set_pmd_at(mm, new_addr, new_pmd, pmd);
1895 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1896 if (new_ptl != old_ptl)
1897 spin_unlock(new_ptl);
1898 spin_unlock(old_ptl);
1906 * - 0 if PMD could not be locked
1907 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1908 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1910 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1911 unsigned long addr, pgprot_t newprot, int prot_numa)
1913 struct mm_struct *mm = vma->vm_mm;
1916 bool preserve_write;
1919 ptl = __pmd_trans_huge_lock(pmd, vma);
1923 preserve_write = prot_numa && pmd_write(*pmd);
1926 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1927 if (is_swap_pmd(*pmd)) {
1928 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1930 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1931 if (is_write_migration_entry(entry)) {
1934 * A protection check is difficult so
1935 * just be safe and disable write
1937 make_migration_entry_read(&entry);
1938 newpmd = swp_entry_to_pmd(entry);
1939 if (pmd_swp_soft_dirty(*pmd))
1940 newpmd = pmd_swp_mksoft_dirty(newpmd);
1941 set_pmd_at(mm, addr, pmd, newpmd);
1948 * Avoid trapping faults against the zero page. The read-only
1949 * data is likely to be read-cached on the local CPU and
1950 * local/remote hits to the zero page are not interesting.
1952 if (prot_numa && is_huge_zero_pmd(*pmd))
1955 if (prot_numa && pmd_protnone(*pmd))
1959 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1960 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1961 * which is also under down_read(mmap_sem):
1964 * change_huge_pmd(prot_numa=1)
1965 * pmdp_huge_get_and_clear_notify()
1966 * madvise_dontneed()
1968 * pmd_trans_huge(*pmd) == 0 (without ptl)
1971 * // pmd is re-established
1973 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1974 * which may break userspace.
1976 * pmdp_invalidate() is required to make sure we don't miss
1977 * dirty/young flags set by hardware.
1979 entry = pmdp_invalidate(vma, addr, pmd);
1981 entry = pmd_modify(entry, newprot);
1983 entry = pmd_mk_savedwrite(entry);
1985 set_pmd_at(mm, addr, pmd, entry);
1986 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1993 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1995 * Note that if it returns page table lock pointer, this routine returns without
1996 * unlocking page table lock. So callers must unlock it.
1998 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
2001 ptl = pmd_lock(vma->vm_mm, pmd);
2002 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
2010 * Returns true if a given pud maps a thp, false otherwise.
2012 * Note that if it returns true, this routine returns without unlocking page
2013 * table lock. So callers must unlock it.
2015 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
2019 ptl = pud_lock(vma->vm_mm, pud);
2020 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
2026 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2027 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
2028 pud_t *pud, unsigned long addr)
2032 ptl = __pud_trans_huge_lock(pud, vma);
2036 * For architectures like ppc64 we look at deposited pgtable
2037 * when calling pudp_huge_get_and_clear. So do the
2038 * pgtable_trans_huge_withdraw after finishing pudp related
2041 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
2042 tlb_remove_pud_tlb_entry(tlb, pud, addr);
2043 if (vma_is_dax(vma)) {
2045 /* No zero page support yet */
2047 /* No support for anonymous PUD pages yet */
2053 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
2054 unsigned long haddr)
2056 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
2057 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2058 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
2059 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
2061 count_vm_event(THP_SPLIT_PUD);
2063 pudp_huge_clear_flush_notify(vma, haddr, pud);
2066 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2067 unsigned long address)
2070 struct mmu_notifier_range range;
2072 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2073 address & HPAGE_PUD_MASK,
2074 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
2075 mmu_notifier_invalidate_range_start(&range);
2076 ptl = pud_lock(vma->vm_mm, pud);
2077 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2079 __split_huge_pud_locked(vma, pud, range.start);
2084 * No need to double call mmu_notifier->invalidate_range() callback as
2085 * the above pudp_huge_clear_flush_notify() did already call it.
2087 mmu_notifier_invalidate_range_only_end(&range);
2089 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2091 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2092 unsigned long haddr, pmd_t *pmd)
2094 struct mm_struct *mm = vma->vm_mm;
2100 * Leave pmd empty until pte is filled note that it is fine to delay
2101 * notification until mmu_notifier_invalidate_range_end() as we are
2102 * replacing a zero pmd write protected page with a zero pte write
2105 * See Documentation/vm/mmu_notifier.rst
2107 pmdp_huge_clear_flush(vma, haddr, pmd);
2109 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2110 pmd_populate(mm, &_pmd, pgtable);
2112 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2114 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2115 entry = pte_mkspecial(entry);
2116 pte = pte_offset_map(&_pmd, haddr);
2117 VM_BUG_ON(!pte_none(*pte));
2118 set_pte_at(mm, haddr, pte, entry);
2121 smp_wmb(); /* make pte visible before pmd */
2122 pmd_populate(mm, pmd, pgtable);
2125 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2126 unsigned long haddr, bool freeze)
2128 struct mm_struct *mm = vma->vm_mm;
2131 pmd_t old_pmd, _pmd;
2132 bool young, write, soft_dirty, pmd_migration = false;
2136 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2137 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2138 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2139 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2140 && !pmd_devmap(*pmd));
2142 count_vm_event(THP_SPLIT_PMD);
2144 if (!vma_is_anonymous(vma)) {
2145 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2147 * We are going to unmap this huge page. So
2148 * just go ahead and zap it
2150 if (arch_needs_pgtable_deposit())
2151 zap_deposited_table(mm, pmd);
2152 if (vma_is_dax(vma))
2154 page = pmd_page(_pmd);
2155 if (!PageDirty(page) && pmd_dirty(_pmd))
2156 set_page_dirty(page);
2157 if (!PageReferenced(page) && pmd_young(_pmd))
2158 SetPageReferenced(page);
2159 page_remove_rmap(page, true);
2161 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2163 } else if (is_huge_zero_pmd(*pmd)) {
2165 * FIXME: Do we want to invalidate secondary mmu by calling
2166 * mmu_notifier_invalidate_range() see comments below inside
2167 * __split_huge_pmd() ?
2169 * We are going from a zero huge page write protected to zero
2170 * small page also write protected so it does not seems useful
2171 * to invalidate secondary mmu at this time.
2173 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2177 * Up to this point the pmd is present and huge and userland has the
2178 * whole access to the hugepage during the split (which happens in
2179 * place). If we overwrite the pmd with the not-huge version pointing
2180 * to the pte here (which of course we could if all CPUs were bug
2181 * free), userland could trigger a small page size TLB miss on the
2182 * small sized TLB while the hugepage TLB entry is still established in
2183 * the huge TLB. Some CPU doesn't like that.
2184 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2185 * 383 on page 93. Intel should be safe but is also warns that it's
2186 * only safe if the permission and cache attributes of the two entries
2187 * loaded in the two TLB is identical (which should be the case here).
2188 * But it is generally safer to never allow small and huge TLB entries
2189 * for the same virtual address to be loaded simultaneously. So instead
2190 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2191 * current pmd notpresent (atomically because here the pmd_trans_huge
2192 * must remain set at all times on the pmd until the split is complete
2193 * for this pmd), then we flush the SMP TLB and finally we write the
2194 * non-huge version of the pmd entry with pmd_populate.
2196 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2198 pmd_migration = is_pmd_migration_entry(old_pmd);
2199 if (unlikely(pmd_migration)) {
2202 entry = pmd_to_swp_entry(old_pmd);
2203 page = pfn_to_page(swp_offset(entry));
2204 write = is_write_migration_entry(entry);
2206 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2208 page = pmd_page(old_pmd);
2209 if (pmd_dirty(old_pmd))
2211 write = pmd_write(old_pmd);
2212 young = pmd_young(old_pmd);
2213 soft_dirty = pmd_soft_dirty(old_pmd);
2215 VM_BUG_ON_PAGE(!page_count(page), page);
2216 page_ref_add(page, HPAGE_PMD_NR - 1);
2219 * Withdraw the table only after we mark the pmd entry invalid.
2220 * This's critical for some architectures (Power).
2222 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2223 pmd_populate(mm, &_pmd, pgtable);
2225 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2228 * Note that NUMA hinting access restrictions are not
2229 * transferred to avoid any possibility of altering
2230 * permissions across VMAs.
2232 if (freeze || pmd_migration) {
2233 swp_entry_t swp_entry;
2234 swp_entry = make_migration_entry(page + i, write);
2235 entry = swp_entry_to_pte(swp_entry);
2237 entry = pte_swp_mksoft_dirty(entry);
2239 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2240 entry = maybe_mkwrite(entry, vma);
2242 entry = pte_wrprotect(entry);
2244 entry = pte_mkold(entry);
2246 entry = pte_mksoft_dirty(entry);
2248 pte = pte_offset_map(&_pmd, addr);
2249 BUG_ON(!pte_none(*pte));
2250 set_pte_at(mm, addr, pte, entry);
2251 atomic_inc(&page[i]._mapcount);
2256 * Set PG_double_map before dropping compound_mapcount to avoid
2257 * false-negative page_mapped().
2259 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2260 for (i = 0; i < HPAGE_PMD_NR; i++)
2261 atomic_inc(&page[i]._mapcount);
2264 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2265 /* Last compound_mapcount is gone. */
2266 __dec_node_page_state(page, NR_ANON_THPS);
2267 if (TestClearPageDoubleMap(page)) {
2268 /* No need in mapcount reference anymore */
2269 for (i = 0; i < HPAGE_PMD_NR; i++)
2270 atomic_dec(&page[i]._mapcount);
2274 smp_wmb(); /* make pte visible before pmd */
2275 pmd_populate(mm, pmd, pgtable);
2278 for (i = 0; i < HPAGE_PMD_NR; i++) {
2279 page_remove_rmap(page + i, false);
2285 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2286 unsigned long address, bool freeze, struct page *page)
2289 struct mmu_notifier_range range;
2291 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2292 address & HPAGE_PMD_MASK,
2293 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2294 mmu_notifier_invalidate_range_start(&range);
2295 ptl = pmd_lock(vma->vm_mm, pmd);
2298 * If caller asks to setup a migration entries, we need a page to check
2299 * pmd against. Otherwise we can end up replacing wrong page.
2301 VM_BUG_ON(freeze && !page);
2302 if (page && page != pmd_page(*pmd))
2305 if (pmd_trans_huge(*pmd)) {
2306 page = pmd_page(*pmd);
2307 if (PageMlocked(page))
2308 clear_page_mlock(page);
2309 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2311 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2315 * No need to double call mmu_notifier->invalidate_range() callback.
2316 * They are 3 cases to consider inside __split_huge_pmd_locked():
2317 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2318 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2319 * fault will trigger a flush_notify before pointing to a new page
2320 * (it is fine if the secondary mmu keeps pointing to the old zero
2321 * page in the meantime)
2322 * 3) Split a huge pmd into pte pointing to the same page. No need
2323 * to invalidate secondary tlb entry they are all still valid.
2324 * any further changes to individual pte will notify. So no need
2325 * to call mmu_notifier->invalidate_range()
2327 mmu_notifier_invalidate_range_only_end(&range);
2330 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2331 bool freeze, struct page *page)
2338 pgd = pgd_offset(vma->vm_mm, address);
2339 if (!pgd_present(*pgd))
2342 p4d = p4d_offset(pgd, address);
2343 if (!p4d_present(*p4d))
2346 pud = pud_offset(p4d, address);
2347 if (!pud_present(*pud))
2350 pmd = pmd_offset(pud, address);
2352 __split_huge_pmd(vma, pmd, address, freeze, page);
2355 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2356 unsigned long start,
2361 * If the new start address isn't hpage aligned and it could
2362 * previously contain an hugepage: check if we need to split
2365 if (start & ~HPAGE_PMD_MASK &&
2366 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2367 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2368 split_huge_pmd_address(vma, start, false, NULL);
2371 * If the new end address isn't hpage aligned and it could
2372 * previously contain an hugepage: check if we need to split
2375 if (end & ~HPAGE_PMD_MASK &&
2376 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2377 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2378 split_huge_pmd_address(vma, end, false, NULL);
2381 * If we're also updating the vma->vm_next->vm_start, if the new
2382 * vm_next->vm_start isn't page aligned and it could previously
2383 * contain an hugepage: check if we need to split an huge pmd.
2385 if (adjust_next > 0) {
2386 struct vm_area_struct *next = vma->vm_next;
2387 unsigned long nstart = next->vm_start;
2388 nstart += adjust_next << PAGE_SHIFT;
2389 if (nstart & ~HPAGE_PMD_MASK &&
2390 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2391 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2392 split_huge_pmd_address(next, nstart, false, NULL);
2396 static void unmap_page(struct page *page)
2398 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2399 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2402 VM_BUG_ON_PAGE(!PageHead(page), page);
2405 ttu_flags |= TTU_SPLIT_FREEZE;
2407 unmap_success = try_to_unmap(page, ttu_flags);
2408 VM_BUG_ON_PAGE(!unmap_success, page);
2411 static void remap_page(struct page *page)
2414 if (PageTransHuge(page)) {
2415 remove_migration_ptes(page, page, true);
2417 for (i = 0; i < HPAGE_PMD_NR; i++)
2418 remove_migration_ptes(page + i, page + i, true);
2422 static void __split_huge_page_tail(struct page *head, int tail,
2423 struct lruvec *lruvec, struct list_head *list)
2425 struct page *page_tail = head + tail;
2427 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2430 * Clone page flags before unfreezing refcount.
2432 * After successful get_page_unless_zero() might follow flags change,
2433 * for exmaple lock_page() which set PG_waiters.
2435 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2436 page_tail->flags |= (head->flags &
2437 ((1L << PG_referenced) |
2438 (1L << PG_swapbacked) |
2439 (1L << PG_swapcache) |
2440 (1L << PG_mlocked) |
2441 (1L << PG_uptodate) |
2443 (1L << PG_workingset) |
2445 (1L << PG_unevictable) |
2448 /* ->mapping in first tail page is compound_mapcount */
2449 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2451 page_tail->mapping = head->mapping;
2452 page_tail->index = head->index + tail;
2454 /* Page flags must be visible before we make the page non-compound. */
2458 * Clear PageTail before unfreezing page refcount.
2460 * After successful get_page_unless_zero() might follow put_page()
2461 * which needs correct compound_head().
2463 clear_compound_head(page_tail);
2465 /* Finally unfreeze refcount. Additional reference from page cache. */
2466 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2467 PageSwapCache(head)));
2469 if (page_is_young(head))
2470 set_page_young(page_tail);
2471 if (page_is_idle(head))
2472 set_page_idle(page_tail);
2474 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2477 * always add to the tail because some iterators expect new
2478 * pages to show after the currently processed elements - e.g.
2481 lru_add_page_tail(head, page_tail, lruvec, list);
2484 static void __split_huge_page(struct page *page, struct list_head *list,
2485 pgoff_t end, unsigned long flags)
2487 struct page *head = compound_head(page);
2488 pg_data_t *pgdat = page_pgdat(head);
2489 struct lruvec *lruvec;
2492 lruvec = mem_cgroup_page_lruvec(head, pgdat);
2494 /* complete memcg works before add pages to LRU */
2495 mem_cgroup_split_huge_fixup(head);
2497 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2498 __split_huge_page_tail(head, i, lruvec, list);
2499 /* Some pages can be beyond i_size: drop them from page cache */
2500 if (head[i].index >= end) {
2501 ClearPageDirty(head + i);
2502 __delete_from_page_cache(head + i, NULL);
2503 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2504 shmem_uncharge(head->mapping->host, 1);
2509 ClearPageCompound(head);
2511 split_page_owner(head, HPAGE_PMD_ORDER);
2513 /* See comment in __split_huge_page_tail() */
2514 if (PageAnon(head)) {
2515 /* Additional pin to swap cache */
2516 if (PageSwapCache(head))
2517 page_ref_add(head, 2);
2521 /* Additional pin to page cache */
2522 page_ref_add(head, 2);
2523 xa_unlock(&head->mapping->i_pages);
2526 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
2530 for (i = 0; i < HPAGE_PMD_NR; i++) {
2531 struct page *subpage = head + i;
2532 if (subpage == page)
2534 unlock_page(subpage);
2537 * Subpages may be freed if there wasn't any mapping
2538 * like if add_to_swap() is running on a lru page that
2539 * had its mapping zapped. And freeing these pages
2540 * requires taking the lru_lock so we do the put_page
2541 * of the tail pages after the split is complete.
2547 int total_mapcount(struct page *page)
2549 int i, compound, ret;
2551 VM_BUG_ON_PAGE(PageTail(page), page);
2553 if (likely(!PageCompound(page)))
2554 return atomic_read(&page->_mapcount) + 1;
2556 compound = compound_mapcount(page);
2560 for (i = 0; i < HPAGE_PMD_NR; i++)
2561 ret += atomic_read(&page[i]._mapcount) + 1;
2562 /* File pages has compound_mapcount included in _mapcount */
2563 if (!PageAnon(page))
2564 return ret - compound * HPAGE_PMD_NR;
2565 if (PageDoubleMap(page))
2566 ret -= HPAGE_PMD_NR;
2571 * This calculates accurately how many mappings a transparent hugepage
2572 * has (unlike page_mapcount() which isn't fully accurate). This full
2573 * accuracy is primarily needed to know if copy-on-write faults can
2574 * reuse the page and change the mapping to read-write instead of
2575 * copying them. At the same time this returns the total_mapcount too.
2577 * The function returns the highest mapcount any one of the subpages
2578 * has. If the return value is one, even if different processes are
2579 * mapping different subpages of the transparent hugepage, they can
2580 * all reuse it, because each process is reusing a different subpage.
2582 * The total_mapcount is instead counting all virtual mappings of the
2583 * subpages. If the total_mapcount is equal to "one", it tells the
2584 * caller all mappings belong to the same "mm" and in turn the
2585 * anon_vma of the transparent hugepage can become the vma->anon_vma
2586 * local one as no other process may be mapping any of the subpages.
2588 * It would be more accurate to replace page_mapcount() with
2589 * page_trans_huge_mapcount(), however we only use
2590 * page_trans_huge_mapcount() in the copy-on-write faults where we
2591 * need full accuracy to avoid breaking page pinning, because
2592 * page_trans_huge_mapcount() is slower than page_mapcount().
2594 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2596 int i, ret, _total_mapcount, mapcount;
2598 /* hugetlbfs shouldn't call it */
2599 VM_BUG_ON_PAGE(PageHuge(page), page);
2601 if (likely(!PageTransCompound(page))) {
2602 mapcount = atomic_read(&page->_mapcount) + 1;
2604 *total_mapcount = mapcount;
2608 page = compound_head(page);
2610 _total_mapcount = ret = 0;
2611 for (i = 0; i < HPAGE_PMD_NR; i++) {
2612 mapcount = atomic_read(&page[i]._mapcount) + 1;
2613 ret = max(ret, mapcount);
2614 _total_mapcount += mapcount;
2616 if (PageDoubleMap(page)) {
2618 _total_mapcount -= HPAGE_PMD_NR;
2620 mapcount = compound_mapcount(page);
2622 _total_mapcount += mapcount;
2624 *total_mapcount = _total_mapcount;
2628 /* Racy check whether the huge page can be split */
2629 bool can_split_huge_page(struct page *page, int *pextra_pins)
2633 /* Additional pins from page cache */
2635 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2637 extra_pins = HPAGE_PMD_NR;
2639 *pextra_pins = extra_pins;
2640 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2644 * This function splits huge page into normal pages. @page can point to any
2645 * subpage of huge page to split. Split doesn't change the position of @page.
2647 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2648 * The huge page must be locked.
2650 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2652 * Both head page and tail pages will inherit mapping, flags, and so on from
2655 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2656 * they are not mapped.
2658 * Returns 0 if the hugepage is split successfully.
2659 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2662 int split_huge_page_to_list(struct page *page, struct list_head *list)
2664 struct page *head = compound_head(page);
2665 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2666 struct anon_vma *anon_vma = NULL;
2667 struct address_space *mapping = NULL;
2668 int count, mapcount, extra_pins, ret;
2670 unsigned long flags;
2673 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2674 VM_BUG_ON_PAGE(!PageLocked(page), page);
2675 VM_BUG_ON_PAGE(!PageCompound(page), page);
2677 if (PageWriteback(page))
2680 if (PageAnon(head)) {
2682 * The caller does not necessarily hold an mmap_sem that would
2683 * prevent the anon_vma disappearing so we first we take a
2684 * reference to it and then lock the anon_vma for write. This
2685 * is similar to page_lock_anon_vma_read except the write lock
2686 * is taken to serialise against parallel split or collapse
2689 anon_vma = page_get_anon_vma(head);
2696 anon_vma_lock_write(anon_vma);
2698 mapping = head->mapping;
2707 i_mmap_lock_read(mapping);
2710 *__split_huge_page() may need to trim off pages beyond EOF:
2711 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2712 * which cannot be nested inside the page tree lock. So note
2713 * end now: i_size itself may be changed at any moment, but
2714 * head page lock is good enough to serialize the trimming.
2716 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2720 * Racy check if we can split the page, before unmap_page() will
2723 if (!can_split_huge_page(head, &extra_pins)) {
2728 mlocked = PageMlocked(page);
2730 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2732 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2736 /* prevent PageLRU to go away from under us, and freeze lru stats */
2737 spin_lock_irqsave(&pgdata->lru_lock, flags);
2740 XA_STATE(xas, &mapping->i_pages, page_index(head));
2743 * Check if the head page is present in page cache.
2744 * We assume all tail are present too, if head is there.
2746 xa_lock(&mapping->i_pages);
2747 if (xas_load(&xas) != head)
2751 /* Prevent deferred_split_scan() touching ->_refcount */
2752 spin_lock(&pgdata->split_queue_lock);
2753 count = page_count(head);
2754 mapcount = total_mapcount(head);
2755 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2756 if (!list_empty(page_deferred_list(head))) {
2757 pgdata->split_queue_len--;
2758 list_del(page_deferred_list(head));
2761 __dec_node_page_state(page, NR_SHMEM_THPS);
2762 spin_unlock(&pgdata->split_queue_lock);
2763 __split_huge_page(page, list, end, flags);
2764 if (PageSwapCache(head)) {
2765 swp_entry_t entry = { .val = page_private(head) };
2767 ret = split_swap_cluster(entry);
2771 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2772 pr_alert("total_mapcount: %u, page_count(): %u\n",
2775 dump_page(head, NULL);
2776 dump_page(page, "total_mapcount(head) > 0");
2779 spin_unlock(&pgdata->split_queue_lock);
2781 xa_unlock(&mapping->i_pages);
2782 spin_unlock_irqrestore(&pgdata->lru_lock, flags);
2789 anon_vma_unlock_write(anon_vma);
2790 put_anon_vma(anon_vma);
2793 i_mmap_unlock_read(mapping);
2795 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2799 void free_transhuge_page(struct page *page)
2801 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2802 unsigned long flags;
2804 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2805 if (!list_empty(page_deferred_list(page))) {
2806 pgdata->split_queue_len--;
2807 list_del(page_deferred_list(page));
2809 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2810 free_compound_page(page);
2813 void deferred_split_huge_page(struct page *page)
2815 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2816 unsigned long flags;
2818 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2820 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2821 if (list_empty(page_deferred_list(page))) {
2822 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2823 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2824 pgdata->split_queue_len++;
2826 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2829 static unsigned long deferred_split_count(struct shrinker *shrink,
2830 struct shrink_control *sc)
2832 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2833 return READ_ONCE(pgdata->split_queue_len);
2836 static unsigned long deferred_split_scan(struct shrinker *shrink,
2837 struct shrink_control *sc)
2839 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2840 unsigned long flags;
2841 LIST_HEAD(list), *pos, *next;
2845 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2846 /* Take pin on all head pages to avoid freeing them under us */
2847 list_for_each_safe(pos, next, &pgdata->split_queue) {
2848 page = list_entry((void *)pos, struct page, mapping);
2849 page = compound_head(page);
2850 if (get_page_unless_zero(page)) {
2851 list_move(page_deferred_list(page), &list);
2853 /* We lost race with put_compound_page() */
2854 list_del_init(page_deferred_list(page));
2855 pgdata->split_queue_len--;
2857 if (!--sc->nr_to_scan)
2860 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2862 list_for_each_safe(pos, next, &list) {
2863 page = list_entry((void *)pos, struct page, mapping);
2864 if (!trylock_page(page))
2866 /* split_huge_page() removes page from list on success */
2867 if (!split_huge_page(page))
2874 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2875 list_splice_tail(&list, &pgdata->split_queue);
2876 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2879 * Stop shrinker if we didn't split any page, but the queue is empty.
2880 * This can happen if pages were freed under us.
2882 if (!split && list_empty(&pgdata->split_queue))
2887 static struct shrinker deferred_split_shrinker = {
2888 .count_objects = deferred_split_count,
2889 .scan_objects = deferred_split_scan,
2890 .seeks = DEFAULT_SEEKS,
2891 .flags = SHRINKER_NUMA_AWARE,
2894 #ifdef CONFIG_DEBUG_FS
2895 static int split_huge_pages_set(void *data, u64 val)
2899 unsigned long pfn, max_zone_pfn;
2900 unsigned long total = 0, split = 0;
2905 for_each_populated_zone(zone) {
2906 max_zone_pfn = zone_end_pfn(zone);
2907 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2908 if (!pfn_valid(pfn))
2911 page = pfn_to_page(pfn);
2912 if (!get_page_unless_zero(page))
2915 if (zone != page_zone(page))
2918 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2923 if (!split_huge_page(page))
2931 pr_info("%lu of %lu THP split\n", split, total);
2935 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2938 static int __init split_huge_pages_debugfs(void)
2940 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2941 &split_huge_pages_fops);
2944 late_initcall(split_huge_pages_debugfs);
2947 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2948 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2951 struct vm_area_struct *vma = pvmw->vma;
2952 struct mm_struct *mm = vma->vm_mm;
2953 unsigned long address = pvmw->address;
2958 if (!(pvmw->pmd && !pvmw->pte))
2961 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2962 pmdval = *pvmw->pmd;
2963 pmdp_invalidate(vma, address, pvmw->pmd);
2964 if (pmd_dirty(pmdval))
2965 set_page_dirty(page);
2966 entry = make_migration_entry(page, pmd_write(pmdval));
2967 pmdswp = swp_entry_to_pmd(entry);
2968 if (pmd_soft_dirty(pmdval))
2969 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2970 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2971 page_remove_rmap(page, true);
2975 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2977 struct vm_area_struct *vma = pvmw->vma;
2978 struct mm_struct *mm = vma->vm_mm;
2979 unsigned long address = pvmw->address;
2980 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2984 if (!(pvmw->pmd && !pvmw->pte))
2987 entry = pmd_to_swp_entry(*pvmw->pmd);
2989 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2990 if (pmd_swp_soft_dirty(*pvmw->pmd))
2991 pmde = pmd_mksoft_dirty(pmde);
2992 if (is_write_migration_entry(entry))
2993 pmde = maybe_pmd_mkwrite(pmde, vma);
2995 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2997 page_add_anon_rmap(new, vma, mmun_start, true);
2999 page_add_file_rmap(new, true);
3000 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3001 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3002 mlock_vma_page(new);
3003 update_mmu_cache_pmd(vma, address, pvmw->pmd);