2 * Copyright (C) 2009 Red Hat, Inc.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/sched.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/kthread.h>
20 #include <linux/khugepaged.h>
21 #include <linux/freezer.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/migrate.h>
25 #include <linux/hashtable.h>
26 #include <linux/userfaultfd_k.h>
29 #include <asm/pgalloc.h>
33 * By default transparent hugepage support is disabled in order that avoid
34 * to risk increase the memory footprint of applications without a guaranteed
35 * benefit. When transparent hugepage support is enabled, is for all mappings,
36 * and khugepaged scans all mappings.
37 * Defrag is invoked by khugepaged hugepage allocations and by page faults
38 * for all hugepage allocations.
40 unsigned long transparent_hugepage_flags __read_mostly =
41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
42 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
44 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
45 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
47 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
48 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
49 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
51 /* default scan 8*512 pte (or vmas) every 30 second */
52 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
53 static unsigned int khugepaged_pages_collapsed;
54 static unsigned int khugepaged_full_scans;
55 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
56 /* during fragmentation poll the hugepage allocator once every minute */
57 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
58 static struct task_struct *khugepaged_thread __read_mostly;
59 static DEFINE_MUTEX(khugepaged_mutex);
60 static DEFINE_SPINLOCK(khugepaged_mm_lock);
61 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
63 * default collapse hugepages if there is at least one pte mapped like
64 * it would have happened if the vma was large enough during page
67 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
69 static int khugepaged(void *none);
70 static int khugepaged_slab_init(void);
71 static void khugepaged_slab_exit(void);
73 #define MM_SLOTS_HASH_BITS 10
74 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
76 static struct kmem_cache *mm_slot_cache __read_mostly;
79 * struct mm_slot - hash lookup from mm to mm_slot
80 * @hash: hash collision list
81 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
82 * @mm: the mm that this information is valid for
85 struct hlist_node hash;
86 struct list_head mm_node;
91 * struct khugepaged_scan - cursor for scanning
92 * @mm_head: the head of the mm list to scan
93 * @mm_slot: the current mm_slot we are scanning
94 * @address: the next address inside that to be scanned
96 * There is only the one khugepaged_scan instance of this cursor structure.
98 struct khugepaged_scan {
99 struct list_head mm_head;
100 struct mm_slot *mm_slot;
101 unsigned long address;
103 static struct khugepaged_scan khugepaged_scan = {
104 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
108 static int set_recommended_min_free_kbytes(void)
112 unsigned long recommended_min;
114 for_each_populated_zone(zone)
117 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
118 recommended_min = pageblock_nr_pages * nr_zones * 2;
121 * Make sure that on average at least two pageblocks are almost free
122 * of another type, one for a migratetype to fall back to and a
123 * second to avoid subsequent fallbacks of other types There are 3
124 * MIGRATE_TYPES we care about.
126 recommended_min += pageblock_nr_pages * nr_zones *
127 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
129 /* don't ever allow to reserve more than 5% of the lowmem */
130 recommended_min = min(recommended_min,
131 (unsigned long) nr_free_buffer_pages() / 20);
132 recommended_min <<= (PAGE_SHIFT-10);
134 if (recommended_min > min_free_kbytes) {
135 if (user_min_free_kbytes >= 0)
136 pr_info("raising min_free_kbytes from %d to %lu "
137 "to help transparent hugepage allocations\n",
138 min_free_kbytes, recommended_min);
140 min_free_kbytes = recommended_min;
142 setup_per_zone_wmarks();
146 static int start_stop_khugepaged(void)
149 if (khugepaged_enabled()) {
150 if (!khugepaged_thread)
151 khugepaged_thread = kthread_run(khugepaged, NULL,
153 if (unlikely(IS_ERR(khugepaged_thread))) {
154 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
155 err = PTR_ERR(khugepaged_thread);
156 khugepaged_thread = NULL;
160 if (!list_empty(&khugepaged_scan.mm_head))
161 wake_up_interruptible(&khugepaged_wait);
163 set_recommended_min_free_kbytes();
164 } else if (khugepaged_thread) {
165 kthread_stop(khugepaged_thread);
166 khugepaged_thread = NULL;
172 static atomic_t huge_zero_refcount;
173 struct page *huge_zero_page __read_mostly;
175 static inline bool is_huge_zero_pmd(pmd_t pmd)
177 return is_huge_zero_page(pmd_page(pmd));
180 static struct page *get_huge_zero_page(void)
182 struct page *zero_page;
184 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
185 return READ_ONCE(huge_zero_page);
187 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
190 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
193 count_vm_event(THP_ZERO_PAGE_ALLOC);
195 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
197 __free_pages(zero_page, compound_order(zero_page));
201 /* We take additional reference here. It will be put back by shrinker */
202 atomic_set(&huge_zero_refcount, 2);
204 return READ_ONCE(huge_zero_page);
207 static void put_huge_zero_page(void)
210 * Counter should never go to zero here. Only shrinker can put
213 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
216 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
217 struct shrink_control *sc)
219 /* we can free zero page only if last reference remains */
220 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
223 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
224 struct shrink_control *sc)
226 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
227 struct page *zero_page = xchg(&huge_zero_page, NULL);
228 BUG_ON(zero_page == NULL);
229 __free_pages(zero_page, compound_order(zero_page));
236 static struct shrinker huge_zero_page_shrinker = {
237 .count_objects = shrink_huge_zero_page_count,
238 .scan_objects = shrink_huge_zero_page_scan,
239 .seeks = DEFAULT_SEEKS,
244 static ssize_t double_flag_show(struct kobject *kobj,
245 struct kobj_attribute *attr, char *buf,
246 enum transparent_hugepage_flag enabled,
247 enum transparent_hugepage_flag req_madv)
249 if (test_bit(enabled, &transparent_hugepage_flags)) {
250 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
251 return sprintf(buf, "[always] madvise never\n");
252 } else if (test_bit(req_madv, &transparent_hugepage_flags))
253 return sprintf(buf, "always [madvise] never\n");
255 return sprintf(buf, "always madvise [never]\n");
257 static ssize_t double_flag_store(struct kobject *kobj,
258 struct kobj_attribute *attr,
259 const char *buf, size_t count,
260 enum transparent_hugepage_flag enabled,
261 enum transparent_hugepage_flag req_madv)
263 if (!memcmp("always", buf,
264 min(sizeof("always")-1, count))) {
265 set_bit(enabled, &transparent_hugepage_flags);
266 clear_bit(req_madv, &transparent_hugepage_flags);
267 } else if (!memcmp("madvise", buf,
268 min(sizeof("madvise")-1, count))) {
269 clear_bit(enabled, &transparent_hugepage_flags);
270 set_bit(req_madv, &transparent_hugepage_flags);
271 } else if (!memcmp("never", buf,
272 min(sizeof("never")-1, count))) {
273 clear_bit(enabled, &transparent_hugepage_flags);
274 clear_bit(req_madv, &transparent_hugepage_flags);
281 static ssize_t enabled_show(struct kobject *kobj,
282 struct kobj_attribute *attr, char *buf)
284 return double_flag_show(kobj, attr, buf,
285 TRANSPARENT_HUGEPAGE_FLAG,
286 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
288 static ssize_t enabled_store(struct kobject *kobj,
289 struct kobj_attribute *attr,
290 const char *buf, size_t count)
294 ret = double_flag_store(kobj, attr, buf, count,
295 TRANSPARENT_HUGEPAGE_FLAG,
296 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
301 mutex_lock(&khugepaged_mutex);
302 err = start_stop_khugepaged();
303 mutex_unlock(&khugepaged_mutex);
311 static struct kobj_attribute enabled_attr =
312 __ATTR(enabled, 0644, enabled_show, enabled_store);
314 static ssize_t single_flag_show(struct kobject *kobj,
315 struct kobj_attribute *attr, char *buf,
316 enum transparent_hugepage_flag flag)
318 return sprintf(buf, "%d\n",
319 !!test_bit(flag, &transparent_hugepage_flags));
322 static ssize_t single_flag_store(struct kobject *kobj,
323 struct kobj_attribute *attr,
324 const char *buf, size_t count,
325 enum transparent_hugepage_flag flag)
330 ret = kstrtoul(buf, 10, &value);
337 set_bit(flag, &transparent_hugepage_flags);
339 clear_bit(flag, &transparent_hugepage_flags);
345 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
346 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
347 * memory just to allocate one more hugepage.
349 static ssize_t defrag_show(struct kobject *kobj,
350 struct kobj_attribute *attr, char *buf)
352 return double_flag_show(kobj, attr, buf,
353 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
354 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
356 static ssize_t defrag_store(struct kobject *kobj,
357 struct kobj_attribute *attr,
358 const char *buf, size_t count)
360 return double_flag_store(kobj, attr, buf, count,
361 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
362 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
364 static struct kobj_attribute defrag_attr =
365 __ATTR(defrag, 0644, defrag_show, defrag_store);
367 static ssize_t use_zero_page_show(struct kobject *kobj,
368 struct kobj_attribute *attr, char *buf)
370 return single_flag_show(kobj, attr, buf,
371 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
373 static ssize_t use_zero_page_store(struct kobject *kobj,
374 struct kobj_attribute *attr, const char *buf, size_t count)
376 return single_flag_store(kobj, attr, buf, count,
377 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
379 static struct kobj_attribute use_zero_page_attr =
380 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
381 #ifdef CONFIG_DEBUG_VM
382 static ssize_t debug_cow_show(struct kobject *kobj,
383 struct kobj_attribute *attr, char *buf)
385 return single_flag_show(kobj, attr, buf,
386 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
388 static ssize_t debug_cow_store(struct kobject *kobj,
389 struct kobj_attribute *attr,
390 const char *buf, size_t count)
392 return single_flag_store(kobj, attr, buf, count,
393 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
395 static struct kobj_attribute debug_cow_attr =
396 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
397 #endif /* CONFIG_DEBUG_VM */
399 static struct attribute *hugepage_attr[] = {
402 &use_zero_page_attr.attr,
403 #ifdef CONFIG_DEBUG_VM
404 &debug_cow_attr.attr,
409 static struct attribute_group hugepage_attr_group = {
410 .attrs = hugepage_attr,
413 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
414 struct kobj_attribute *attr,
417 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
420 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
421 struct kobj_attribute *attr,
422 const char *buf, size_t count)
427 err = kstrtoul(buf, 10, &msecs);
428 if (err || msecs > UINT_MAX)
431 khugepaged_scan_sleep_millisecs = msecs;
432 wake_up_interruptible(&khugepaged_wait);
436 static struct kobj_attribute scan_sleep_millisecs_attr =
437 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
438 scan_sleep_millisecs_store);
440 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
441 struct kobj_attribute *attr,
444 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
447 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
448 struct kobj_attribute *attr,
449 const char *buf, size_t count)
454 err = kstrtoul(buf, 10, &msecs);
455 if (err || msecs > UINT_MAX)
458 khugepaged_alloc_sleep_millisecs = msecs;
459 wake_up_interruptible(&khugepaged_wait);
463 static struct kobj_attribute alloc_sleep_millisecs_attr =
464 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
465 alloc_sleep_millisecs_store);
467 static ssize_t pages_to_scan_show(struct kobject *kobj,
468 struct kobj_attribute *attr,
471 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
473 static ssize_t pages_to_scan_store(struct kobject *kobj,
474 struct kobj_attribute *attr,
475 const char *buf, size_t count)
480 err = kstrtoul(buf, 10, &pages);
481 if (err || !pages || pages > UINT_MAX)
484 khugepaged_pages_to_scan = pages;
488 static struct kobj_attribute pages_to_scan_attr =
489 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
490 pages_to_scan_store);
492 static ssize_t pages_collapsed_show(struct kobject *kobj,
493 struct kobj_attribute *attr,
496 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
498 static struct kobj_attribute pages_collapsed_attr =
499 __ATTR_RO(pages_collapsed);
501 static ssize_t full_scans_show(struct kobject *kobj,
502 struct kobj_attribute *attr,
505 return sprintf(buf, "%u\n", khugepaged_full_scans);
507 static struct kobj_attribute full_scans_attr =
508 __ATTR_RO(full_scans);
510 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
511 struct kobj_attribute *attr, char *buf)
513 return single_flag_show(kobj, attr, buf,
514 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
516 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
517 struct kobj_attribute *attr,
518 const char *buf, size_t count)
520 return single_flag_store(kobj, attr, buf, count,
521 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
523 static struct kobj_attribute khugepaged_defrag_attr =
524 __ATTR(defrag, 0644, khugepaged_defrag_show,
525 khugepaged_defrag_store);
528 * max_ptes_none controls if khugepaged should collapse hugepages over
529 * any unmapped ptes in turn potentially increasing the memory
530 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
531 * reduce the available free memory in the system as it
532 * runs. Increasing max_ptes_none will instead potentially reduce the
533 * free memory in the system during the khugepaged scan.
535 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
536 struct kobj_attribute *attr,
539 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
541 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
542 struct kobj_attribute *attr,
543 const char *buf, size_t count)
546 unsigned long max_ptes_none;
548 err = kstrtoul(buf, 10, &max_ptes_none);
549 if (err || max_ptes_none > HPAGE_PMD_NR-1)
552 khugepaged_max_ptes_none = max_ptes_none;
556 static struct kobj_attribute khugepaged_max_ptes_none_attr =
557 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
558 khugepaged_max_ptes_none_store);
560 static struct attribute *khugepaged_attr[] = {
561 &khugepaged_defrag_attr.attr,
562 &khugepaged_max_ptes_none_attr.attr,
563 &pages_to_scan_attr.attr,
564 &pages_collapsed_attr.attr,
565 &full_scans_attr.attr,
566 &scan_sleep_millisecs_attr.attr,
567 &alloc_sleep_millisecs_attr.attr,
571 static struct attribute_group khugepaged_attr_group = {
572 .attrs = khugepaged_attr,
573 .name = "khugepaged",
576 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
580 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
581 if (unlikely(!*hugepage_kobj)) {
582 pr_err("failed to create transparent hugepage kobject\n");
586 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
588 pr_err("failed to register transparent hugepage group\n");
592 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
594 pr_err("failed to register transparent hugepage group\n");
595 goto remove_hp_group;
601 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
603 kobject_put(*hugepage_kobj);
607 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
609 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
610 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
611 kobject_put(hugepage_kobj);
614 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
619 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
622 #endif /* CONFIG_SYSFS */
624 static int __init hugepage_init(void)
627 struct kobject *hugepage_kobj;
629 if (!has_transparent_hugepage()) {
630 transparent_hugepage_flags = 0;
634 err = hugepage_init_sysfs(&hugepage_kobj);
638 err = khugepaged_slab_init();
642 err = register_shrinker(&huge_zero_page_shrinker);
644 goto err_hzp_shrinker;
647 * By default disable transparent hugepages on smaller systems,
648 * where the extra memory used could hurt more than TLB overhead
649 * is likely to save. The admin can still enable it through /sys.
651 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
652 transparent_hugepage_flags = 0;
656 err = start_stop_khugepaged();
662 unregister_shrinker(&huge_zero_page_shrinker);
664 khugepaged_slab_exit();
666 hugepage_exit_sysfs(hugepage_kobj);
670 subsys_initcall(hugepage_init);
672 static int __init setup_transparent_hugepage(char *str)
677 if (!strcmp(str, "always")) {
678 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
683 } else if (!strcmp(str, "madvise")) {
684 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
685 &transparent_hugepage_flags);
686 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
687 &transparent_hugepage_flags);
689 } else if (!strcmp(str, "never")) {
690 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
691 &transparent_hugepage_flags);
692 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
693 &transparent_hugepage_flags);
698 pr_warn("transparent_hugepage= cannot parse, ignored\n");
701 __setup("transparent_hugepage=", setup_transparent_hugepage);
703 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
705 if (likely(vma->vm_flags & VM_WRITE))
706 pmd = pmd_mkwrite(pmd);
710 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
713 entry = mk_pmd(page, prot);
714 entry = pmd_mkhuge(entry);
718 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
719 struct vm_area_struct *vma,
720 unsigned long haddr, pmd_t *pmd,
721 struct page *page, gfp_t gfp,
724 struct mem_cgroup *memcg;
728 VM_BUG_ON_PAGE(!PageCompound(page), page);
730 if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
732 count_vm_event(THP_FAULT_FALLBACK);
733 return VM_FAULT_FALLBACK;
736 pgtable = pte_alloc_one(mm, haddr);
737 if (unlikely(!pgtable)) {
738 mem_cgroup_cancel_charge(page, memcg);
743 clear_huge_page(page, haddr, HPAGE_PMD_NR);
745 * The memory barrier inside __SetPageUptodate makes sure that
746 * clear_huge_page writes become visible before the set_pmd_at()
749 __SetPageUptodate(page);
751 ptl = pmd_lock(mm, pmd);
752 if (unlikely(!pmd_none(*pmd))) {
754 mem_cgroup_cancel_charge(page, memcg);
756 pte_free(mm, pgtable);
760 /* Deliver the page fault to userland */
761 if (userfaultfd_missing(vma)) {
765 mem_cgroup_cancel_charge(page, memcg);
767 pte_free(mm, pgtable);
768 ret = handle_userfault(vma, haddr, flags,
770 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
774 entry = mk_huge_pmd(page, vma->vm_page_prot);
775 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
776 page_add_new_anon_rmap(page, vma, haddr);
777 mem_cgroup_commit_charge(page, memcg, false);
778 lru_cache_add_active_or_unevictable(page, vma);
779 pgtable_trans_huge_deposit(mm, pmd, pgtable);
780 set_pmd_at(mm, haddr, pmd, entry);
781 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
782 atomic_long_inc(&mm->nr_ptes);
784 count_vm_event(THP_FAULT_ALLOC);
790 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
792 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
795 /* Caller must hold page table lock. */
796 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
797 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
798 struct page *zero_page)
801 entry = mk_pmd(zero_page, vma->vm_page_prot);
802 entry = pmd_mkhuge(entry);
803 pgtable_trans_huge_deposit(mm, pmd, pgtable);
804 set_pmd_at(mm, haddr, pmd, entry);
805 atomic_long_inc(&mm->nr_ptes);
808 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
809 unsigned long address, pmd_t *pmd,
814 unsigned long haddr = address & HPAGE_PMD_MASK;
816 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
817 return VM_FAULT_FALLBACK;
818 if (unlikely(anon_vma_prepare(vma)))
820 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
822 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
823 transparent_hugepage_use_zero_page()) {
826 struct page *zero_page;
829 pgtable = pte_alloc_one(mm, haddr);
830 if (unlikely(!pgtable))
832 zero_page = get_huge_zero_page();
833 if (unlikely(!zero_page)) {
834 pte_free(mm, pgtable);
835 count_vm_event(THP_FAULT_FALLBACK);
836 return VM_FAULT_FALLBACK;
838 ptl = pmd_lock(mm, pmd);
841 if (pmd_none(*pmd)) {
842 if (userfaultfd_missing(vma)) {
844 ret = handle_userfault(vma, haddr, flags,
846 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
848 set_huge_zero_page(pgtable, mm, vma,
857 pte_free(mm, pgtable);
858 put_huge_zero_page();
862 gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
863 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
864 if (unlikely(!page)) {
865 count_vm_event(THP_FAULT_FALLBACK);
866 return VM_FAULT_FALLBACK;
868 return __do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page, gfp, flags);
871 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
872 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
873 struct vm_area_struct *vma)
875 spinlock_t *dst_ptl, *src_ptl;
876 struct page *src_page;
882 pgtable = pte_alloc_one(dst_mm, addr);
883 if (unlikely(!pgtable))
886 dst_ptl = pmd_lock(dst_mm, dst_pmd);
887 src_ptl = pmd_lockptr(src_mm, src_pmd);
888 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
892 if (unlikely(!pmd_trans_huge(pmd))) {
893 pte_free(dst_mm, pgtable);
897 * When page table lock is held, the huge zero pmd should not be
898 * under splitting since we don't split the page itself, only pmd to
901 if (is_huge_zero_pmd(pmd)) {
902 struct page *zero_page;
904 * get_huge_zero_page() will never allocate a new page here,
905 * since we already have a zero page to copy. It just takes a
908 zero_page = get_huge_zero_page();
909 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
915 if (unlikely(pmd_trans_splitting(pmd))) {
916 /* split huge page running from under us */
917 spin_unlock(src_ptl);
918 spin_unlock(dst_ptl);
919 pte_free(dst_mm, pgtable);
921 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
924 src_page = pmd_page(pmd);
925 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
927 page_dup_rmap(src_page);
928 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
930 pmdp_set_wrprotect(src_mm, addr, src_pmd);
931 pmd = pmd_mkold(pmd_wrprotect(pmd));
932 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
933 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
934 atomic_long_inc(&dst_mm->nr_ptes);
938 spin_unlock(src_ptl);
939 spin_unlock(dst_ptl);
944 void huge_pmd_set_accessed(struct mm_struct *mm,
945 struct vm_area_struct *vma,
946 unsigned long address,
947 pmd_t *pmd, pmd_t orig_pmd,
954 ptl = pmd_lock(mm, pmd);
955 if (unlikely(!pmd_same(*pmd, orig_pmd)))
958 entry = pmd_mkyoung(orig_pmd);
959 haddr = address & HPAGE_PMD_MASK;
960 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
961 update_mmu_cache_pmd(vma, address, pmd);
968 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
969 * during copy_user_huge_page()'s copy_page_rep(): in the case when
970 * the source page gets split and a tail freed before copy completes.
971 * Called under pmd_lock of checked pmd, so safe from splitting itself.
973 static void get_user_huge_page(struct page *page)
975 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
976 struct page *endpage = page + HPAGE_PMD_NR;
978 atomic_add(HPAGE_PMD_NR, &page->_count);
979 while (++page < endpage)
980 get_huge_page_tail(page);
986 static void put_user_huge_page(struct page *page)
988 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
989 struct page *endpage = page + HPAGE_PMD_NR;
991 while (page < endpage)
998 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
999 struct vm_area_struct *vma,
1000 unsigned long address,
1001 pmd_t *pmd, pmd_t orig_pmd,
1003 unsigned long haddr)
1005 struct mem_cgroup *memcg;
1010 struct page **pages;
1011 unsigned long mmun_start; /* For mmu_notifiers */
1012 unsigned long mmun_end; /* For mmu_notifiers */
1014 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1016 if (unlikely(!pages)) {
1017 ret |= VM_FAULT_OOM;
1021 for (i = 0; i < HPAGE_PMD_NR; i++) {
1022 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1024 vma, address, page_to_nid(page));
1025 if (unlikely(!pages[i] ||
1026 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1031 memcg = (void *)page_private(pages[i]);
1032 set_page_private(pages[i], 0);
1033 mem_cgroup_cancel_charge(pages[i], memcg);
1037 ret |= VM_FAULT_OOM;
1040 set_page_private(pages[i], (unsigned long)memcg);
1043 for (i = 0; i < HPAGE_PMD_NR; i++) {
1044 copy_user_highpage(pages[i], page + i,
1045 haddr + PAGE_SIZE * i, vma);
1046 __SetPageUptodate(pages[i]);
1051 mmun_end = haddr + HPAGE_PMD_SIZE;
1052 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1054 ptl = pmd_lock(mm, pmd);
1055 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1056 goto out_free_pages;
1057 VM_BUG_ON_PAGE(!PageHead(page), page);
1059 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1060 /* leave pmd empty until pte is filled */
1062 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1063 pmd_populate(mm, &_pmd, pgtable);
1065 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1067 entry = mk_pte(pages[i], vma->vm_page_prot);
1068 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1069 memcg = (void *)page_private(pages[i]);
1070 set_page_private(pages[i], 0);
1071 page_add_new_anon_rmap(pages[i], vma, haddr);
1072 mem_cgroup_commit_charge(pages[i], memcg, false);
1073 lru_cache_add_active_or_unevictable(pages[i], vma);
1074 pte = pte_offset_map(&_pmd, haddr);
1075 VM_BUG_ON(!pte_none(*pte));
1076 set_pte_at(mm, haddr, pte, entry);
1081 smp_wmb(); /* make pte visible before pmd */
1082 pmd_populate(mm, pmd, pgtable);
1083 page_remove_rmap(page);
1086 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1088 ret |= VM_FAULT_WRITE;
1096 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1097 for (i = 0; i < HPAGE_PMD_NR; i++) {
1098 memcg = (void *)page_private(pages[i]);
1099 set_page_private(pages[i], 0);
1100 mem_cgroup_cancel_charge(pages[i], memcg);
1107 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1108 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1112 struct page *page = NULL, *new_page;
1113 struct mem_cgroup *memcg;
1114 unsigned long haddr;
1115 unsigned long mmun_start; /* For mmu_notifiers */
1116 unsigned long mmun_end; /* For mmu_notifiers */
1117 gfp_t huge_gfp; /* for allocation and charge */
1119 ptl = pmd_lockptr(mm, pmd);
1120 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1121 haddr = address & HPAGE_PMD_MASK;
1122 if (is_huge_zero_pmd(orig_pmd))
1125 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1128 page = pmd_page(orig_pmd);
1129 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1130 if (page_mapcount(page) == 1) {
1132 entry = pmd_mkyoung(orig_pmd);
1133 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1134 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1135 update_mmu_cache_pmd(vma, address, pmd);
1136 ret |= VM_FAULT_WRITE;
1139 get_user_huge_page(page);
1142 if (transparent_hugepage_enabled(vma) &&
1143 !transparent_hugepage_debug_cow()) {
1144 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1145 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1149 if (unlikely(!new_page)) {
1151 split_huge_page_pmd(vma, address, pmd);
1152 ret |= VM_FAULT_FALLBACK;
1154 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1155 pmd, orig_pmd, page, haddr);
1156 if (ret & VM_FAULT_OOM) {
1157 split_huge_page(page);
1158 ret |= VM_FAULT_FALLBACK;
1160 put_user_huge_page(page);
1162 count_vm_event(THP_FAULT_FALLBACK);
1166 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1169 split_huge_page(page);
1170 put_user_huge_page(page);
1172 split_huge_page_pmd(vma, address, pmd);
1173 ret |= VM_FAULT_FALLBACK;
1174 count_vm_event(THP_FAULT_FALLBACK);
1178 count_vm_event(THP_FAULT_ALLOC);
1181 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1183 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1184 __SetPageUptodate(new_page);
1187 mmun_end = haddr + HPAGE_PMD_SIZE;
1188 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1192 put_user_huge_page(page);
1193 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1195 mem_cgroup_cancel_charge(new_page, memcg);
1200 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1201 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1202 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1203 page_add_new_anon_rmap(new_page, vma, haddr);
1204 mem_cgroup_commit_charge(new_page, memcg, false);
1205 lru_cache_add_active_or_unevictable(new_page, vma);
1206 set_pmd_at(mm, haddr, pmd, entry);
1207 update_mmu_cache_pmd(vma, address, pmd);
1209 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1210 put_huge_zero_page();
1212 VM_BUG_ON_PAGE(!PageHead(page), page);
1213 page_remove_rmap(page);
1216 ret |= VM_FAULT_WRITE;
1220 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1228 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1233 struct mm_struct *mm = vma->vm_mm;
1234 struct page *page = NULL;
1236 assert_spin_locked(pmd_lockptr(mm, pmd));
1238 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1241 /* Avoid dumping huge zero page */
1242 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1243 return ERR_PTR(-EFAULT);
1245 /* Full NUMA hinting faults to serialise migration in fault paths */
1246 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1249 page = pmd_page(*pmd);
1250 VM_BUG_ON_PAGE(!PageHead(page), page);
1251 if (flags & FOLL_TOUCH) {
1254 * We should set the dirty bit only for FOLL_WRITE but
1255 * for now the dirty bit in the pmd is meaningless.
1256 * And if the dirty bit will become meaningful and
1257 * we'll only set it with FOLL_WRITE, an atomic
1258 * set_bit will be required on the pmd to set the
1259 * young bit, instead of the current set_pmd_at.
1261 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1262 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1264 update_mmu_cache_pmd(vma, addr, pmd);
1266 if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1267 if (page->mapping && trylock_page(page)) {
1270 mlock_vma_page(page);
1274 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1275 VM_BUG_ON_PAGE(!PageCompound(page), page);
1276 if (flags & FOLL_GET)
1277 get_page_foll(page);
1283 /* NUMA hinting page fault entry point for trans huge pmds */
1284 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1285 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1288 struct anon_vma *anon_vma = NULL;
1290 unsigned long haddr = addr & HPAGE_PMD_MASK;
1291 int page_nid = -1, this_nid = numa_node_id();
1292 int target_nid, last_cpupid = -1;
1294 bool migrated = false;
1298 /* A PROT_NONE fault should not end up here */
1299 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1301 ptl = pmd_lock(mm, pmdp);
1302 if (unlikely(!pmd_same(pmd, *pmdp)))
1306 * If there are potential migrations, wait for completion and retry
1307 * without disrupting NUMA hinting information. Do not relock and
1308 * check_same as the page may no longer be mapped.
1310 if (unlikely(pmd_trans_migrating(*pmdp))) {
1311 page = pmd_page(*pmdp);
1313 wait_on_page_locked(page);
1317 page = pmd_page(pmd);
1318 BUG_ON(is_huge_zero_page(page));
1319 page_nid = page_to_nid(page);
1320 last_cpupid = page_cpupid_last(page);
1321 count_vm_numa_event(NUMA_HINT_FAULTS);
1322 if (page_nid == this_nid) {
1323 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1324 flags |= TNF_FAULT_LOCAL;
1327 /* See similar comment in do_numa_page for explanation */
1328 if (!(vma->vm_flags & VM_WRITE))
1329 flags |= TNF_NO_GROUP;
1332 * Acquire the page lock to serialise THP migrations but avoid dropping
1333 * page_table_lock if at all possible
1335 page_locked = trylock_page(page);
1336 target_nid = mpol_misplaced(page, vma, haddr);
1337 if (target_nid == -1) {
1338 /* If the page was locked, there are no parallel migrations */
1343 /* Migration could have started since the pmd_trans_migrating check */
1346 wait_on_page_locked(page);
1352 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1353 * to serialises splits
1357 anon_vma = page_lock_anon_vma_read(page);
1359 /* Confirm the PMD did not change while page_table_lock was released */
1361 if (unlikely(!pmd_same(pmd, *pmdp))) {
1368 /* Bail if we fail to protect against THP splits for any reason */
1369 if (unlikely(!anon_vma)) {
1376 * Migrate the THP to the requested node, returns with page unlocked
1377 * and access rights restored.
1380 migrated = migrate_misplaced_transhuge_page(mm, vma,
1381 pmdp, pmd, addr, page, target_nid);
1383 flags |= TNF_MIGRATED;
1384 page_nid = target_nid;
1386 flags |= TNF_MIGRATE_FAIL;
1390 BUG_ON(!PageLocked(page));
1391 was_writable = pmd_write(pmd);
1392 pmd = pmd_modify(pmd, vma->vm_page_prot);
1393 pmd = pmd_mkyoung(pmd);
1395 pmd = pmd_mkwrite(pmd);
1396 set_pmd_at(mm, haddr, pmdp, pmd);
1397 update_mmu_cache_pmd(vma, addr, pmdp);
1404 page_unlock_anon_vma_read(anon_vma);
1407 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1412 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1413 pmd_t *pmd, unsigned long addr)
1418 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1423 * For architectures like ppc64 we look at deposited pgtable
1424 * when calling pmdp_huge_get_and_clear. So do the
1425 * pgtable_trans_huge_withdraw after finishing pmdp related
1428 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1430 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1431 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1432 if (is_huge_zero_pmd(orig_pmd)) {
1433 atomic_long_dec(&tlb->mm->nr_ptes);
1435 put_huge_zero_page();
1437 page = pmd_page(orig_pmd);
1438 page_remove_rmap(page);
1439 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1440 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1441 VM_BUG_ON_PAGE(!PageHead(page), page);
1442 atomic_long_dec(&tlb->mm->nr_ptes);
1444 tlb_remove_page(tlb, page);
1446 pte_free(tlb->mm, pgtable);
1452 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1453 unsigned long old_addr,
1454 unsigned long new_addr, unsigned long old_end,
1455 pmd_t *old_pmd, pmd_t *new_pmd)
1457 spinlock_t *old_ptl, *new_ptl;
1461 struct mm_struct *mm = vma->vm_mm;
1463 if ((old_addr & ~HPAGE_PMD_MASK) ||
1464 (new_addr & ~HPAGE_PMD_MASK) ||
1465 old_end - old_addr < HPAGE_PMD_SIZE ||
1466 (new_vma->vm_flags & VM_NOHUGEPAGE))
1470 * The destination pmd shouldn't be established, free_pgtables()
1471 * should have release it.
1473 if (WARN_ON(!pmd_none(*new_pmd))) {
1474 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1479 * We don't have to worry about the ordering of src and dst
1480 * ptlocks because exclusive mmap_sem prevents deadlock.
1482 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1484 new_ptl = pmd_lockptr(mm, new_pmd);
1485 if (new_ptl != old_ptl)
1486 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1487 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1488 VM_BUG_ON(!pmd_none(*new_pmd));
1490 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1492 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1493 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1495 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1496 if (new_ptl != old_ptl)
1497 spin_unlock(new_ptl);
1498 spin_unlock(old_ptl);
1506 * - 0 if PMD could not be locked
1507 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1508 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1510 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1511 unsigned long addr, pgprot_t newprot, int prot_numa)
1513 struct mm_struct *mm = vma->vm_mm;
1517 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1519 bool preserve_write = prot_numa && pmd_write(*pmd);
1523 * Avoid trapping faults against the zero page. The read-only
1524 * data is likely to be read-cached on the local CPU and
1525 * local/remote hits to the zero page are not interesting.
1527 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1532 if (!prot_numa || !pmd_protnone(*pmd)) {
1533 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1534 entry = pmd_modify(entry, newprot);
1536 entry = pmd_mkwrite(entry);
1538 set_pmd_at(mm, addr, pmd, entry);
1539 BUG_ON(!preserve_write && pmd_write(entry));
1548 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1549 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1551 * Note that if it returns 1, this routine returns without unlocking page
1552 * table locks. So callers must unlock them.
1554 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1557 *ptl = pmd_lock(vma->vm_mm, pmd);
1558 if (likely(pmd_trans_huge(*pmd))) {
1559 if (unlikely(pmd_trans_splitting(*pmd))) {
1561 wait_split_huge_page(vma->anon_vma, pmd);
1564 /* Thp mapped by 'pmd' is stable, so we can
1565 * handle it as it is. */
1574 * This function returns whether a given @page is mapped onto the @address
1575 * in the virtual space of @mm.
1577 * When it's true, this function returns *pmd with holding the page table lock
1578 * and passing it back to the caller via @ptl.
1579 * If it's false, returns NULL without holding the page table lock.
1581 pmd_t *page_check_address_pmd(struct page *page,
1582 struct mm_struct *mm,
1583 unsigned long address,
1584 enum page_check_address_pmd_flag flag,
1591 if (address & ~HPAGE_PMD_MASK)
1594 pgd = pgd_offset(mm, address);
1595 if (!pgd_present(*pgd))
1597 pud = pud_offset(pgd, address);
1598 if (!pud_present(*pud))
1600 pmd = pmd_offset(pud, address);
1602 *ptl = pmd_lock(mm, pmd);
1603 if (!pmd_present(*pmd))
1605 if (pmd_page(*pmd) != page)
1608 * split_vma() may create temporary aliased mappings. There is
1609 * no risk as long as all huge pmd are found and have their
1610 * splitting bit set before __split_huge_page_refcount
1611 * runs. Finding the same huge pmd more than once during the
1612 * same rmap walk is not a problem.
1614 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1615 pmd_trans_splitting(*pmd))
1617 if (pmd_trans_huge(*pmd)) {
1618 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1619 !pmd_trans_splitting(*pmd));
1627 static int __split_huge_page_splitting(struct page *page,
1628 struct vm_area_struct *vma,
1629 unsigned long address)
1631 struct mm_struct *mm = vma->vm_mm;
1635 /* For mmu_notifiers */
1636 const unsigned long mmun_start = address;
1637 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1639 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1640 pmd = page_check_address_pmd(page, mm, address,
1641 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1644 * We can't temporarily set the pmd to null in order
1645 * to split it, the pmd must remain marked huge at all
1646 * times or the VM won't take the pmd_trans_huge paths
1647 * and it won't wait on the anon_vma->root->rwsem to
1648 * serialize against split_huge_page*.
1650 pmdp_splitting_flush(vma, address, pmd);
1655 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1660 static void __split_huge_page_refcount(struct page *page,
1661 struct list_head *list)
1664 struct zone *zone = page_zone(page);
1665 struct lruvec *lruvec;
1668 /* prevent PageLRU to go away from under us, and freeze lru stats */
1669 spin_lock_irq(&zone->lru_lock);
1670 lruvec = mem_cgroup_page_lruvec(page, zone);
1672 compound_lock(page);
1673 /* complete memcg works before add pages to LRU */
1674 mem_cgroup_split_huge_fixup(page);
1676 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1677 struct page *page_tail = page + i;
1679 /* tail_page->_mapcount cannot change */
1680 BUG_ON(page_mapcount(page_tail) < 0);
1681 tail_count += page_mapcount(page_tail);
1682 /* check for overflow */
1683 BUG_ON(tail_count < 0);
1684 BUG_ON(atomic_read(&page_tail->_count) != 0);
1686 * tail_page->_count is zero and not changing from
1687 * under us. But get_page_unless_zero() may be running
1688 * from under us on the tail_page. If we used
1689 * atomic_set() below instead of atomic_add(), we
1690 * would then run atomic_set() concurrently with
1691 * get_page_unless_zero(), and atomic_set() is
1692 * implemented in C not using locked ops. spin_unlock
1693 * on x86 sometime uses locked ops because of PPro
1694 * errata 66, 92, so unless somebody can guarantee
1695 * atomic_set() here would be safe on all archs (and
1696 * not only on x86), it's safer to use atomic_add().
1698 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1699 &page_tail->_count);
1701 /* after clearing PageTail the gup refcount can be released */
1702 smp_mb__after_atomic();
1704 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1705 page_tail->flags |= (page->flags &
1706 ((1L << PG_referenced) |
1707 (1L << PG_swapbacked) |
1708 (1L << PG_mlocked) |
1709 (1L << PG_uptodate) |
1711 (1L << PG_unevictable)));
1712 page_tail->flags |= (1L << PG_dirty);
1714 /* clear PageTail before overwriting first_page */
1718 * __split_huge_page_splitting() already set the
1719 * splitting bit in all pmd that could map this
1720 * hugepage, that will ensure no CPU can alter the
1721 * mapcount on the head page. The mapcount is only
1722 * accounted in the head page and it has to be
1723 * transferred to all tail pages in the below code. So
1724 * for this code to be safe, the split the mapcount
1725 * can't change. But that doesn't mean userland can't
1726 * keep changing and reading the page contents while
1727 * we transfer the mapcount, so the pmd splitting
1728 * status is achieved setting a reserved bit in the
1729 * pmd, not by clearing the present bit.
1731 page_tail->_mapcount = page->_mapcount;
1733 BUG_ON(page_tail->mapping);
1734 page_tail->mapping = page->mapping;
1736 page_tail->index = page->index + i;
1737 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1739 BUG_ON(!PageAnon(page_tail));
1740 BUG_ON(!PageUptodate(page_tail));
1741 BUG_ON(!PageDirty(page_tail));
1742 BUG_ON(!PageSwapBacked(page_tail));
1744 lru_add_page_tail(page, page_tail, lruvec, list);
1746 atomic_sub(tail_count, &page->_count);
1747 BUG_ON(atomic_read(&page->_count) <= 0);
1749 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1751 ClearPageCompound(page);
1752 compound_unlock(page);
1753 spin_unlock_irq(&zone->lru_lock);
1755 for (i = 1; i < HPAGE_PMD_NR; i++) {
1756 struct page *page_tail = page + i;
1757 BUG_ON(page_count(page_tail) <= 0);
1759 * Tail pages may be freed if there wasn't any mapping
1760 * like if add_to_swap() is running on a lru page that
1761 * had its mapping zapped. And freeing these pages
1762 * requires taking the lru_lock so we do the put_page
1763 * of the tail pages after the split is complete.
1765 put_page(page_tail);
1769 * Only the head page (now become a regular page) is required
1770 * to be pinned by the caller.
1772 BUG_ON(page_count(page) <= 0);
1775 static int __split_huge_page_map(struct page *page,
1776 struct vm_area_struct *vma,
1777 unsigned long address)
1779 struct mm_struct *mm = vma->vm_mm;
1784 unsigned long haddr;
1786 pmd = page_check_address_pmd(page, mm, address,
1787 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1789 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1790 pmd_populate(mm, &_pmd, pgtable);
1791 if (pmd_write(*pmd))
1792 BUG_ON(page_mapcount(page) != 1);
1795 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1797 BUG_ON(PageCompound(page+i));
1799 * Note that NUMA hinting access restrictions are not
1800 * transferred to avoid any possibility of altering
1801 * permissions across VMAs.
1803 entry = mk_pte(page + i, vma->vm_page_prot);
1804 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1805 if (!pmd_write(*pmd))
1806 entry = pte_wrprotect(entry);
1807 if (!pmd_young(*pmd))
1808 entry = pte_mkold(entry);
1809 pte = pte_offset_map(&_pmd, haddr);
1810 BUG_ON(!pte_none(*pte));
1811 set_pte_at(mm, haddr, pte, entry);
1815 smp_wmb(); /* make pte visible before pmd */
1817 * Up to this point the pmd is present and huge and
1818 * userland has the whole access to the hugepage
1819 * during the split (which happens in place). If we
1820 * overwrite the pmd with the not-huge version
1821 * pointing to the pte here (which of course we could
1822 * if all CPUs were bug free), userland could trigger
1823 * a small page size TLB miss on the small sized TLB
1824 * while the hugepage TLB entry is still established
1825 * in the huge TLB. Some CPU doesn't like that. See
1826 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1827 * Erratum 383 on page 93. Intel should be safe but is
1828 * also warns that it's only safe if the permission
1829 * and cache attributes of the two entries loaded in
1830 * the two TLB is identical (which should be the case
1831 * here). But it is generally safer to never allow
1832 * small and huge TLB entries for the same virtual
1833 * address to be loaded simultaneously. So instead of
1834 * doing "pmd_populate(); flush_tlb_range();" we first
1835 * mark the current pmd notpresent (atomically because
1836 * here the pmd_trans_huge and pmd_trans_splitting
1837 * must remain set at all times on the pmd until the
1838 * split is complete for this pmd), then we flush the
1839 * SMP TLB and finally we write the non-huge version
1840 * of the pmd entry with pmd_populate.
1842 pmdp_invalidate(vma, address, pmd);
1843 pmd_populate(mm, pmd, pgtable);
1851 /* must be called with anon_vma->root->rwsem held */
1852 static void __split_huge_page(struct page *page,
1853 struct anon_vma *anon_vma,
1854 struct list_head *list)
1856 int mapcount, mapcount2;
1857 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1858 struct anon_vma_chain *avc;
1860 BUG_ON(!PageHead(page));
1861 BUG_ON(PageTail(page));
1864 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1865 struct vm_area_struct *vma = avc->vma;
1866 unsigned long addr = vma_address(page, vma);
1867 BUG_ON(is_vma_temporary_stack(vma));
1868 mapcount += __split_huge_page_splitting(page, vma, addr);
1871 * It is critical that new vmas are added to the tail of the
1872 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1873 * and establishes a child pmd before
1874 * __split_huge_page_splitting() freezes the parent pmd (so if
1875 * we fail to prevent copy_huge_pmd() from running until the
1876 * whole __split_huge_page() is complete), we will still see
1877 * the newly established pmd of the child later during the
1878 * walk, to be able to set it as pmd_trans_splitting too.
1880 if (mapcount != page_mapcount(page)) {
1881 pr_err("mapcount %d page_mapcount %d\n",
1882 mapcount, page_mapcount(page));
1886 __split_huge_page_refcount(page, list);
1889 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1890 struct vm_area_struct *vma = avc->vma;
1891 unsigned long addr = vma_address(page, vma);
1892 BUG_ON(is_vma_temporary_stack(vma));
1893 mapcount2 += __split_huge_page_map(page, vma, addr);
1895 if (mapcount != mapcount2) {
1896 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1897 mapcount, mapcount2, page_mapcount(page));
1903 * Split a hugepage into normal pages. This doesn't change the position of head
1904 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1905 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1906 * from the hugepage.
1907 * Return 0 if the hugepage is split successfully otherwise return 1.
1909 int split_huge_page_to_list(struct page *page, struct list_head *list)
1911 struct anon_vma *anon_vma;
1914 BUG_ON(is_huge_zero_page(page));
1915 BUG_ON(!PageAnon(page));
1918 * The caller does not necessarily hold an mmap_sem that would prevent
1919 * the anon_vma disappearing so we first we take a reference to it
1920 * and then lock the anon_vma for write. This is similar to
1921 * page_lock_anon_vma_read except the write lock is taken to serialise
1922 * against parallel split or collapse operations.
1924 anon_vma = page_get_anon_vma(page);
1927 anon_vma_lock_write(anon_vma);
1930 if (!PageCompound(page))
1933 BUG_ON(!PageSwapBacked(page));
1934 __split_huge_page(page, anon_vma, list);
1935 count_vm_event(THP_SPLIT);
1937 BUG_ON(PageCompound(page));
1939 anon_vma_unlock_write(anon_vma);
1940 put_anon_vma(anon_vma);
1945 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1947 int hugepage_madvise(struct vm_area_struct *vma,
1948 unsigned long *vm_flags, int advice)
1954 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1955 * can't handle this properly after s390_enable_sie, so we simply
1956 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1958 if (mm_has_pgste(vma->vm_mm))
1962 * Be somewhat over-protective like KSM for now!
1964 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1966 *vm_flags &= ~VM_NOHUGEPAGE;
1967 *vm_flags |= VM_HUGEPAGE;
1969 * If the vma become good for khugepaged to scan,
1970 * register it here without waiting a page fault that
1971 * may not happen any time soon.
1973 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1976 case MADV_NOHUGEPAGE:
1978 * Be somewhat over-protective like KSM for now!
1980 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1982 *vm_flags &= ~VM_HUGEPAGE;
1983 *vm_flags |= VM_NOHUGEPAGE;
1985 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1986 * this vma even if we leave the mm registered in khugepaged if
1987 * it got registered before VM_NOHUGEPAGE was set.
1995 static int __init khugepaged_slab_init(void)
1997 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1998 sizeof(struct mm_slot),
1999 __alignof__(struct mm_slot), 0, NULL);
2006 static void __init khugepaged_slab_exit(void)
2008 kmem_cache_destroy(mm_slot_cache);
2011 static inline struct mm_slot *alloc_mm_slot(void)
2013 if (!mm_slot_cache) /* initialization failed */
2015 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2018 static inline void free_mm_slot(struct mm_slot *mm_slot)
2020 kmem_cache_free(mm_slot_cache, mm_slot);
2023 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2025 struct mm_slot *mm_slot;
2027 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2028 if (mm == mm_slot->mm)
2034 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2035 struct mm_slot *mm_slot)
2038 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2041 static inline int khugepaged_test_exit(struct mm_struct *mm)
2043 return atomic_read(&mm->mm_users) == 0;
2046 int __khugepaged_enter(struct mm_struct *mm)
2048 struct mm_slot *mm_slot;
2051 mm_slot = alloc_mm_slot();
2055 /* __khugepaged_exit() must not run from under us */
2056 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2057 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2058 free_mm_slot(mm_slot);
2062 spin_lock(&khugepaged_mm_lock);
2063 insert_to_mm_slots_hash(mm, mm_slot);
2065 * Insert just behind the scanning cursor, to let the area settle
2068 wakeup = list_empty(&khugepaged_scan.mm_head);
2069 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2070 spin_unlock(&khugepaged_mm_lock);
2072 atomic_inc(&mm->mm_count);
2074 wake_up_interruptible(&khugepaged_wait);
2079 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2080 unsigned long vm_flags)
2082 unsigned long hstart, hend;
2085 * Not yet faulted in so we will register later in the
2086 * page fault if needed.
2090 /* khugepaged not yet working on file or special mappings */
2092 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2093 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2094 hend = vma->vm_end & HPAGE_PMD_MASK;
2096 return khugepaged_enter(vma, vm_flags);
2100 void __khugepaged_exit(struct mm_struct *mm)
2102 struct mm_slot *mm_slot;
2105 spin_lock(&khugepaged_mm_lock);
2106 mm_slot = get_mm_slot(mm);
2107 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2108 hash_del(&mm_slot->hash);
2109 list_del(&mm_slot->mm_node);
2112 spin_unlock(&khugepaged_mm_lock);
2115 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2116 free_mm_slot(mm_slot);
2118 } else if (mm_slot) {
2120 * This is required to serialize against
2121 * khugepaged_test_exit() (which is guaranteed to run
2122 * under mmap sem read mode). Stop here (after we
2123 * return all pagetables will be destroyed) until
2124 * khugepaged has finished working on the pagetables
2125 * under the mmap_sem.
2127 down_write(&mm->mmap_sem);
2128 up_write(&mm->mmap_sem);
2132 static void release_pte_page(struct page *page)
2134 /* 0 stands for page_is_file_cache(page) == false */
2135 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2137 putback_lru_page(page);
2140 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2142 while (--_pte >= pte) {
2143 pte_t pteval = *_pte;
2144 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2145 release_pte_page(pte_page(pteval));
2149 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2150 unsigned long address,
2155 int none_or_zero = 0;
2156 bool referenced = false, writable = false;
2157 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2158 _pte++, address += PAGE_SIZE) {
2159 pte_t pteval = *_pte;
2160 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2161 if (!userfaultfd_armed(vma) &&
2162 ++none_or_zero <= khugepaged_max_ptes_none)
2167 if (!pte_present(pteval))
2169 page = vm_normal_page(vma, address, pteval);
2170 if (unlikely(!page))
2173 VM_BUG_ON_PAGE(PageCompound(page), page);
2174 VM_BUG_ON_PAGE(!PageAnon(page), page);
2175 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2178 * We can do it before isolate_lru_page because the
2179 * page can't be freed from under us. NOTE: PG_lock
2180 * is needed to serialize against split_huge_page
2181 * when invoked from the VM.
2183 if (!trylock_page(page))
2187 * cannot use mapcount: can't collapse if there's a gup pin.
2188 * The page must only be referenced by the scanned process
2189 * and page swap cache.
2191 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2195 if (pte_write(pteval)) {
2198 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2203 * Page is not in the swap cache. It can be collapsed
2209 * Isolate the page to avoid collapsing an hugepage
2210 * currently in use by the VM.
2212 if (isolate_lru_page(page)) {
2216 /* 0 stands for page_is_file_cache(page) == false */
2217 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2218 VM_BUG_ON_PAGE(!PageLocked(page), page);
2219 VM_BUG_ON_PAGE(PageLRU(page), page);
2221 /* If there is no mapped pte young don't collapse the page */
2222 if (pte_young(pteval) || PageReferenced(page) ||
2223 mmu_notifier_test_young(vma->vm_mm, address))
2226 if (likely(referenced && writable))
2229 release_pte_pages(pte, _pte);
2233 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2234 struct vm_area_struct *vma,
2235 unsigned long address,
2239 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2240 pte_t pteval = *_pte;
2241 struct page *src_page;
2243 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2244 clear_user_highpage(page, address);
2245 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2246 if (is_zero_pfn(pte_pfn(pteval))) {
2248 * ptl mostly unnecessary.
2252 * paravirt calls inside pte_clear here are
2255 pte_clear(vma->vm_mm, address, _pte);
2259 src_page = pte_page(pteval);
2260 copy_user_highpage(page, src_page, address, vma);
2261 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2262 release_pte_page(src_page);
2264 * ptl mostly unnecessary, but preempt has to
2265 * be disabled to update the per-cpu stats
2266 * inside page_remove_rmap().
2270 * paravirt calls inside pte_clear here are
2273 pte_clear(vma->vm_mm, address, _pte);
2274 page_remove_rmap(src_page);
2276 free_page_and_swap_cache(src_page);
2279 address += PAGE_SIZE;
2284 static void khugepaged_alloc_sleep(void)
2286 wait_event_freezable_timeout(khugepaged_wait, false,
2287 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2290 static int khugepaged_node_load[MAX_NUMNODES];
2292 static bool khugepaged_scan_abort(int nid)
2297 * If zone_reclaim_mode is disabled, then no extra effort is made to
2298 * allocate memory locally.
2300 if (!zone_reclaim_mode)
2303 /* If there is a count for this node already, it must be acceptable */
2304 if (khugepaged_node_load[nid])
2307 for (i = 0; i < MAX_NUMNODES; i++) {
2308 if (!khugepaged_node_load[i])
2310 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2317 static int khugepaged_find_target_node(void)
2319 static int last_khugepaged_target_node = NUMA_NO_NODE;
2320 int nid, target_node = 0, max_value = 0;
2322 /* find first node with max normal pages hit */
2323 for (nid = 0; nid < MAX_NUMNODES; nid++)
2324 if (khugepaged_node_load[nid] > max_value) {
2325 max_value = khugepaged_node_load[nid];
2329 /* do some balance if several nodes have the same hit record */
2330 if (target_node <= last_khugepaged_target_node)
2331 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2333 if (max_value == khugepaged_node_load[nid]) {
2338 last_khugepaged_target_node = target_node;
2342 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2344 if (IS_ERR(*hpage)) {
2350 khugepaged_alloc_sleep();
2351 } else if (*hpage) {
2359 static struct page *
2360 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2361 struct vm_area_struct *vma, unsigned long address,
2364 VM_BUG_ON_PAGE(*hpage, *hpage);
2367 * Before allocating the hugepage, release the mmap_sem read lock.
2368 * The allocation can take potentially a long time if it involves
2369 * sync compaction, and we do not need to hold the mmap_sem during
2370 * that. We will recheck the vma after taking it again in write mode.
2372 up_read(&mm->mmap_sem);
2374 *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2375 if (unlikely(!*hpage)) {
2376 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2377 *hpage = ERR_PTR(-ENOMEM);
2381 count_vm_event(THP_COLLAPSE_ALLOC);
2385 static int khugepaged_find_target_node(void)
2390 static inline struct page *alloc_hugepage(int defrag)
2392 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2396 static struct page *khugepaged_alloc_hugepage(bool *wait)
2401 hpage = alloc_hugepage(khugepaged_defrag());
2403 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2408 khugepaged_alloc_sleep();
2410 count_vm_event(THP_COLLAPSE_ALLOC);
2411 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2416 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2419 *hpage = khugepaged_alloc_hugepage(wait);
2421 if (unlikely(!*hpage))
2427 static struct page *
2428 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2429 struct vm_area_struct *vma, unsigned long address,
2432 up_read(&mm->mmap_sem);
2439 static bool hugepage_vma_check(struct vm_area_struct *vma)
2441 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2442 (vma->vm_flags & VM_NOHUGEPAGE))
2445 if (!vma->anon_vma || vma->vm_ops)
2447 if (is_vma_temporary_stack(vma))
2449 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2453 static void collapse_huge_page(struct mm_struct *mm,
2454 unsigned long address,
2455 struct page **hpage,
2456 struct vm_area_struct *vma,
2462 struct page *new_page;
2463 spinlock_t *pmd_ptl, *pte_ptl;
2465 unsigned long hstart, hend;
2466 struct mem_cgroup *memcg;
2467 unsigned long mmun_start; /* For mmu_notifiers */
2468 unsigned long mmun_end; /* For mmu_notifiers */
2471 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2473 /* Only allocate from the target node */
2474 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2477 /* release the mmap_sem read lock. */
2478 new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2482 if (unlikely(mem_cgroup_try_charge(new_page, mm,
2487 * Prevent all access to pagetables with the exception of
2488 * gup_fast later hanlded by the ptep_clear_flush and the VM
2489 * handled by the anon_vma lock + PG_lock.
2491 down_write(&mm->mmap_sem);
2492 if (unlikely(khugepaged_test_exit(mm)))
2495 vma = find_vma(mm, address);
2498 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2499 hend = vma->vm_end & HPAGE_PMD_MASK;
2500 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2502 if (!hugepage_vma_check(vma))
2504 pmd = mm_find_pmd(mm, address);
2508 anon_vma_lock_write(vma->anon_vma);
2510 pte = pte_offset_map(pmd, address);
2511 pte_ptl = pte_lockptr(mm, pmd);
2513 mmun_start = address;
2514 mmun_end = address + HPAGE_PMD_SIZE;
2515 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2516 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2518 * After this gup_fast can't run anymore. This also removes
2519 * any huge TLB entry from the CPU so we won't allow
2520 * huge and small TLB entries for the same virtual address
2521 * to avoid the risk of CPU bugs in that area.
2523 _pmd = pmdp_collapse_flush(vma, address, pmd);
2524 spin_unlock(pmd_ptl);
2525 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2528 isolated = __collapse_huge_page_isolate(vma, address, pte);
2529 spin_unlock(pte_ptl);
2531 if (unlikely(!isolated)) {
2534 BUG_ON(!pmd_none(*pmd));
2536 * We can only use set_pmd_at when establishing
2537 * hugepmds and never for establishing regular pmds that
2538 * points to regular pagetables. Use pmd_populate for that
2540 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2541 spin_unlock(pmd_ptl);
2542 anon_vma_unlock_write(vma->anon_vma);
2547 * All pages are isolated and locked so anon_vma rmap
2548 * can't run anymore.
2550 anon_vma_unlock_write(vma->anon_vma);
2552 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2554 __SetPageUptodate(new_page);
2555 pgtable = pmd_pgtable(_pmd);
2557 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2558 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2561 * spin_lock() below is not the equivalent of smp_wmb(), so
2562 * this is needed to avoid the copy_huge_page writes to become
2563 * visible after the set_pmd_at() write.
2568 BUG_ON(!pmd_none(*pmd));
2569 page_add_new_anon_rmap(new_page, vma, address);
2570 mem_cgroup_commit_charge(new_page, memcg, false);
2571 lru_cache_add_active_or_unevictable(new_page, vma);
2572 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2573 set_pmd_at(mm, address, pmd, _pmd);
2574 update_mmu_cache_pmd(vma, address, pmd);
2575 spin_unlock(pmd_ptl);
2579 khugepaged_pages_collapsed++;
2581 up_write(&mm->mmap_sem);
2585 mem_cgroup_cancel_charge(new_page, memcg);
2589 static int khugepaged_scan_pmd(struct mm_struct *mm,
2590 struct vm_area_struct *vma,
2591 unsigned long address,
2592 struct page **hpage)
2596 int ret = 0, none_or_zero = 0;
2598 unsigned long _address;
2600 int node = NUMA_NO_NODE;
2601 bool writable = false, referenced = false;
2603 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2605 pmd = mm_find_pmd(mm, address);
2609 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2610 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2611 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2612 _pte++, _address += PAGE_SIZE) {
2613 pte_t pteval = *_pte;
2614 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2615 if (!userfaultfd_armed(vma) &&
2616 ++none_or_zero <= khugepaged_max_ptes_none)
2621 if (!pte_present(pteval))
2623 if (pte_write(pteval))
2626 page = vm_normal_page(vma, _address, pteval);
2627 if (unlikely(!page))
2630 * Record which node the original page is from and save this
2631 * information to khugepaged_node_load[].
2632 * Khupaged will allocate hugepage from the node has the max
2635 node = page_to_nid(page);
2636 if (khugepaged_scan_abort(node))
2638 khugepaged_node_load[node]++;
2639 VM_BUG_ON_PAGE(PageCompound(page), page);
2640 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2643 * cannot use mapcount: can't collapse if there's a gup pin.
2644 * The page must only be referenced by the scanned process
2645 * and page swap cache.
2647 if (page_count(page) != 1 + !!PageSwapCache(page))
2649 if (pte_young(pteval) || PageReferenced(page) ||
2650 mmu_notifier_test_young(vma->vm_mm, address))
2653 if (referenced && writable)
2656 pte_unmap_unlock(pte, ptl);
2658 node = khugepaged_find_target_node();
2659 /* collapse_huge_page will return with the mmap_sem released */
2660 collapse_huge_page(mm, address, hpage, vma, node);
2666 static void collect_mm_slot(struct mm_slot *mm_slot)
2668 struct mm_struct *mm = mm_slot->mm;
2670 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2672 if (khugepaged_test_exit(mm)) {
2674 hash_del(&mm_slot->hash);
2675 list_del(&mm_slot->mm_node);
2678 * Not strictly needed because the mm exited already.
2680 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2683 /* khugepaged_mm_lock actually not necessary for the below */
2684 free_mm_slot(mm_slot);
2689 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2690 struct page **hpage)
2691 __releases(&khugepaged_mm_lock)
2692 __acquires(&khugepaged_mm_lock)
2694 struct mm_slot *mm_slot;
2695 struct mm_struct *mm;
2696 struct vm_area_struct *vma;
2700 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2702 if (khugepaged_scan.mm_slot)
2703 mm_slot = khugepaged_scan.mm_slot;
2705 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2706 struct mm_slot, mm_node);
2707 khugepaged_scan.address = 0;
2708 khugepaged_scan.mm_slot = mm_slot;
2710 spin_unlock(&khugepaged_mm_lock);
2713 down_read(&mm->mmap_sem);
2714 if (unlikely(khugepaged_test_exit(mm)))
2717 vma = find_vma(mm, khugepaged_scan.address);
2720 for (; vma; vma = vma->vm_next) {
2721 unsigned long hstart, hend;
2724 if (unlikely(khugepaged_test_exit(mm))) {
2728 if (!hugepage_vma_check(vma)) {
2733 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2734 hend = vma->vm_end & HPAGE_PMD_MASK;
2737 if (khugepaged_scan.address > hend)
2739 if (khugepaged_scan.address < hstart)
2740 khugepaged_scan.address = hstart;
2741 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2743 while (khugepaged_scan.address < hend) {
2746 if (unlikely(khugepaged_test_exit(mm)))
2747 goto breakouterloop;
2749 VM_BUG_ON(khugepaged_scan.address < hstart ||
2750 khugepaged_scan.address + HPAGE_PMD_SIZE >
2752 ret = khugepaged_scan_pmd(mm, vma,
2753 khugepaged_scan.address,
2755 /* move to next address */
2756 khugepaged_scan.address += HPAGE_PMD_SIZE;
2757 progress += HPAGE_PMD_NR;
2759 /* we released mmap_sem so break loop */
2760 goto breakouterloop_mmap_sem;
2761 if (progress >= pages)
2762 goto breakouterloop;
2766 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2767 breakouterloop_mmap_sem:
2769 spin_lock(&khugepaged_mm_lock);
2770 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2772 * Release the current mm_slot if this mm is about to die, or
2773 * if we scanned all vmas of this mm.
2775 if (khugepaged_test_exit(mm) || !vma) {
2777 * Make sure that if mm_users is reaching zero while
2778 * khugepaged runs here, khugepaged_exit will find
2779 * mm_slot not pointing to the exiting mm.
2781 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2782 khugepaged_scan.mm_slot = list_entry(
2783 mm_slot->mm_node.next,
2784 struct mm_slot, mm_node);
2785 khugepaged_scan.address = 0;
2787 khugepaged_scan.mm_slot = NULL;
2788 khugepaged_full_scans++;
2791 collect_mm_slot(mm_slot);
2797 static int khugepaged_has_work(void)
2799 return !list_empty(&khugepaged_scan.mm_head) &&
2800 khugepaged_enabled();
2803 static int khugepaged_wait_event(void)
2805 return !list_empty(&khugepaged_scan.mm_head) ||
2806 kthread_should_stop();
2809 static void khugepaged_do_scan(void)
2811 struct page *hpage = NULL;
2812 unsigned int progress = 0, pass_through_head = 0;
2813 unsigned int pages = khugepaged_pages_to_scan;
2816 barrier(); /* write khugepaged_pages_to_scan to local stack */
2818 while (progress < pages) {
2819 if (!khugepaged_prealloc_page(&hpage, &wait))
2824 if (unlikely(kthread_should_stop() || try_to_freeze()))
2827 spin_lock(&khugepaged_mm_lock);
2828 if (!khugepaged_scan.mm_slot)
2829 pass_through_head++;
2830 if (khugepaged_has_work() &&
2831 pass_through_head < 2)
2832 progress += khugepaged_scan_mm_slot(pages - progress,
2836 spin_unlock(&khugepaged_mm_lock);
2839 if (!IS_ERR_OR_NULL(hpage))
2843 static void khugepaged_wait_work(void)
2845 if (khugepaged_has_work()) {
2846 if (!khugepaged_scan_sleep_millisecs)
2849 wait_event_freezable_timeout(khugepaged_wait,
2850 kthread_should_stop(),
2851 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2855 if (khugepaged_enabled())
2856 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2859 static int khugepaged(void *none)
2861 struct mm_slot *mm_slot;
2864 set_user_nice(current, MAX_NICE);
2866 while (!kthread_should_stop()) {
2867 khugepaged_do_scan();
2868 khugepaged_wait_work();
2871 spin_lock(&khugepaged_mm_lock);
2872 mm_slot = khugepaged_scan.mm_slot;
2873 khugepaged_scan.mm_slot = NULL;
2875 collect_mm_slot(mm_slot);
2876 spin_unlock(&khugepaged_mm_lock);
2880 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2881 unsigned long haddr, pmd_t *pmd)
2883 struct mm_struct *mm = vma->vm_mm;
2888 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2889 /* leave pmd empty until pte is filled */
2891 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2892 pmd_populate(mm, &_pmd, pgtable);
2894 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2896 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2897 entry = pte_mkspecial(entry);
2898 pte = pte_offset_map(&_pmd, haddr);
2899 VM_BUG_ON(!pte_none(*pte));
2900 set_pte_at(mm, haddr, pte, entry);
2903 smp_wmb(); /* make pte visible before pmd */
2904 pmd_populate(mm, pmd, pgtable);
2905 put_huge_zero_page();
2908 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2913 struct mm_struct *mm = vma->vm_mm;
2914 unsigned long haddr = address & HPAGE_PMD_MASK;
2915 unsigned long mmun_start; /* For mmu_notifiers */
2916 unsigned long mmun_end; /* For mmu_notifiers */
2918 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2921 mmun_end = haddr + HPAGE_PMD_SIZE;
2923 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2924 ptl = pmd_lock(mm, pmd);
2925 if (unlikely(!pmd_trans_huge(*pmd))) {
2927 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2930 if (is_huge_zero_pmd(*pmd)) {
2931 __split_huge_zero_page_pmd(vma, haddr, pmd);
2933 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2936 page = pmd_page(*pmd);
2937 VM_BUG_ON_PAGE(!page_count(page), page);
2940 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2942 split_huge_page(page);
2947 * We don't always have down_write of mmap_sem here: a racing
2948 * do_huge_pmd_wp_page() might have copied-on-write to another
2949 * huge page before our split_huge_page() got the anon_vma lock.
2951 if (unlikely(pmd_trans_huge(*pmd)))
2955 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2958 struct vm_area_struct *vma;
2960 vma = find_vma(mm, address);
2961 BUG_ON(vma == NULL);
2962 split_huge_page_pmd(vma, address, pmd);
2965 static void split_huge_page_address(struct mm_struct *mm,
2966 unsigned long address)
2972 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2974 pgd = pgd_offset(mm, address);
2975 if (!pgd_present(*pgd))
2978 pud = pud_offset(pgd, address);
2979 if (!pud_present(*pud))
2982 pmd = pmd_offset(pud, address);
2983 if (!pmd_present(*pmd))
2986 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2987 * materialize from under us.
2989 split_huge_page_pmd_mm(mm, address, pmd);
2992 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2993 unsigned long start,
2998 * If the new start address isn't hpage aligned and it could
2999 * previously contain an hugepage: check if we need to split
3002 if (start & ~HPAGE_PMD_MASK &&
3003 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3004 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3005 split_huge_page_address(vma->vm_mm, start);
3008 * If the new end address isn't hpage aligned and it could
3009 * previously contain an hugepage: check if we need to split
3012 if (end & ~HPAGE_PMD_MASK &&
3013 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3014 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3015 split_huge_page_address(vma->vm_mm, end);
3018 * If we're also updating the vma->vm_next->vm_start, if the new
3019 * vm_next->vm_start isn't page aligned and it could previously
3020 * contain an hugepage: check if we need to split an huge pmd.
3022 if (adjust_next > 0) {
3023 struct vm_area_struct *next = vma->vm_next;
3024 unsigned long nstart = next->vm_start;
3025 nstart += adjust_next << PAGE_SHIFT;
3026 if (nstart & ~HPAGE_PMD_MASK &&
3027 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3028 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3029 split_huge_page_address(next->vm_mm, nstart);