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/dax.h>
20 #include <linux/kthread.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/mman.h>
24 #include <linux/pagemap.h>
25 #include <linux/migrate.h>
26 #include <linux/hashtable.h>
27 #include <linux/userfaultfd_k.h>
30 #include <asm/pgalloc.h>
34 * By default transparent hugepage support is disabled in order that avoid
35 * to risk increase the memory footprint of applications without a guaranteed
36 * benefit. When transparent hugepage support is enabled, is for all mappings,
37 * and khugepaged scans all mappings.
38 * Defrag is invoked by khugepaged hugepage allocations and by page faults
39 * for all hugepage allocations.
41 unsigned long transparent_hugepage_flags __read_mostly =
42 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
43 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
45 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
46 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
48 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
49 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
50 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
52 /* default scan 8*512 pte (or vmas) every 30 second */
53 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
54 static unsigned int khugepaged_pages_collapsed;
55 static unsigned int khugepaged_full_scans;
56 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
57 /* during fragmentation poll the hugepage allocator once every minute */
58 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
59 static struct task_struct *khugepaged_thread __read_mostly;
60 static DEFINE_MUTEX(khugepaged_mutex);
61 static DEFINE_SPINLOCK(khugepaged_mm_lock);
62 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
64 * default collapse hugepages if there is at least one pte mapped like
65 * it would have happened if the vma was large enough during page
68 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
70 static int khugepaged(void *none);
71 static int khugepaged_slab_init(void);
72 static void khugepaged_slab_exit(void);
74 #define MM_SLOTS_HASH_BITS 10
75 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
77 static struct kmem_cache *mm_slot_cache __read_mostly;
80 * struct mm_slot - hash lookup from mm to mm_slot
81 * @hash: hash collision list
82 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
83 * @mm: the mm that this information is valid for
86 struct hlist_node hash;
87 struct list_head mm_node;
92 * struct khugepaged_scan - cursor for scanning
93 * @mm_head: the head of the mm list to scan
94 * @mm_slot: the current mm_slot we are scanning
95 * @address: the next address inside that to be scanned
97 * There is only the one khugepaged_scan instance of this cursor structure.
99 struct khugepaged_scan {
100 struct list_head mm_head;
101 struct mm_slot *mm_slot;
102 unsigned long address;
104 static struct khugepaged_scan khugepaged_scan = {
105 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
109 static int set_recommended_min_free_kbytes(void)
113 unsigned long recommended_min;
115 for_each_populated_zone(zone)
118 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
119 recommended_min = pageblock_nr_pages * nr_zones * 2;
122 * Make sure that on average at least two pageblocks are almost free
123 * of another type, one for a migratetype to fall back to and a
124 * second to avoid subsequent fallbacks of other types There are 3
125 * MIGRATE_TYPES we care about.
127 recommended_min += pageblock_nr_pages * nr_zones *
128 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
130 /* don't ever allow to reserve more than 5% of the lowmem */
131 recommended_min = min(recommended_min,
132 (unsigned long) nr_free_buffer_pages() / 20);
133 recommended_min <<= (PAGE_SHIFT-10);
135 if (recommended_min > min_free_kbytes) {
136 if (user_min_free_kbytes >= 0)
137 pr_info("raising min_free_kbytes from %d to %lu "
138 "to help transparent hugepage allocations\n",
139 min_free_kbytes, recommended_min);
141 min_free_kbytes = recommended_min;
143 setup_per_zone_wmarks();
147 static int start_stop_khugepaged(void)
150 if (khugepaged_enabled()) {
151 if (!khugepaged_thread)
152 khugepaged_thread = kthread_run(khugepaged, NULL,
154 if (unlikely(IS_ERR(khugepaged_thread))) {
155 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
156 err = PTR_ERR(khugepaged_thread);
157 khugepaged_thread = NULL;
161 if (!list_empty(&khugepaged_scan.mm_head))
162 wake_up_interruptible(&khugepaged_wait);
164 set_recommended_min_free_kbytes();
165 } else if (khugepaged_thread) {
166 kthread_stop(khugepaged_thread);
167 khugepaged_thread = NULL;
173 static atomic_t huge_zero_refcount;
174 struct page *huge_zero_page __read_mostly;
176 struct page *get_huge_zero_page(void)
178 struct page *zero_page;
180 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
181 return READ_ONCE(huge_zero_page);
183 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
186 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
189 count_vm_event(THP_ZERO_PAGE_ALLOC);
191 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
193 __free_pages(zero_page, compound_order(zero_page));
197 /* We take additional reference here. It will be put back by shrinker */
198 atomic_set(&huge_zero_refcount, 2);
200 return READ_ONCE(huge_zero_page);
203 static void put_huge_zero_page(void)
206 * Counter should never go to zero here. Only shrinker can put
209 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
212 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
213 struct shrink_control *sc)
215 /* we can free zero page only if last reference remains */
216 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
219 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
220 struct shrink_control *sc)
222 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
223 struct page *zero_page = xchg(&huge_zero_page, NULL);
224 BUG_ON(zero_page == NULL);
225 __free_pages(zero_page, compound_order(zero_page));
232 static struct shrinker huge_zero_page_shrinker = {
233 .count_objects = shrink_huge_zero_page_count,
234 .scan_objects = shrink_huge_zero_page_scan,
235 .seeks = DEFAULT_SEEKS,
240 static ssize_t double_flag_show(struct kobject *kobj,
241 struct kobj_attribute *attr, char *buf,
242 enum transparent_hugepage_flag enabled,
243 enum transparent_hugepage_flag req_madv)
245 if (test_bit(enabled, &transparent_hugepage_flags)) {
246 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
247 return sprintf(buf, "[always] madvise never\n");
248 } else if (test_bit(req_madv, &transparent_hugepage_flags))
249 return sprintf(buf, "always [madvise] never\n");
251 return sprintf(buf, "always madvise [never]\n");
253 static ssize_t double_flag_store(struct kobject *kobj,
254 struct kobj_attribute *attr,
255 const char *buf, size_t count,
256 enum transparent_hugepage_flag enabled,
257 enum transparent_hugepage_flag req_madv)
259 if (!memcmp("always", buf,
260 min(sizeof("always")-1, count))) {
261 set_bit(enabled, &transparent_hugepage_flags);
262 clear_bit(req_madv, &transparent_hugepage_flags);
263 } else if (!memcmp("madvise", buf,
264 min(sizeof("madvise")-1, count))) {
265 clear_bit(enabled, &transparent_hugepage_flags);
266 set_bit(req_madv, &transparent_hugepage_flags);
267 } else if (!memcmp("never", buf,
268 min(sizeof("never")-1, count))) {
269 clear_bit(enabled, &transparent_hugepage_flags);
270 clear_bit(req_madv, &transparent_hugepage_flags);
277 static ssize_t enabled_show(struct kobject *kobj,
278 struct kobj_attribute *attr, char *buf)
280 return double_flag_show(kobj, attr, buf,
281 TRANSPARENT_HUGEPAGE_FLAG,
282 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
284 static ssize_t enabled_store(struct kobject *kobj,
285 struct kobj_attribute *attr,
286 const char *buf, size_t count)
290 ret = double_flag_store(kobj, attr, buf, count,
291 TRANSPARENT_HUGEPAGE_FLAG,
292 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
297 mutex_lock(&khugepaged_mutex);
298 err = start_stop_khugepaged();
299 mutex_unlock(&khugepaged_mutex);
307 static struct kobj_attribute enabled_attr =
308 __ATTR(enabled, 0644, enabled_show, enabled_store);
310 static ssize_t single_flag_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf,
312 enum transparent_hugepage_flag flag)
314 return sprintf(buf, "%d\n",
315 !!test_bit(flag, &transparent_hugepage_flags));
318 static ssize_t single_flag_store(struct kobject *kobj,
319 struct kobj_attribute *attr,
320 const char *buf, size_t count,
321 enum transparent_hugepage_flag flag)
326 ret = kstrtoul(buf, 10, &value);
333 set_bit(flag, &transparent_hugepage_flags);
335 clear_bit(flag, &transparent_hugepage_flags);
341 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
342 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
343 * memory just to allocate one more hugepage.
345 static ssize_t defrag_show(struct kobject *kobj,
346 struct kobj_attribute *attr, char *buf)
348 return double_flag_show(kobj, attr, buf,
349 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
350 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
352 static ssize_t defrag_store(struct kobject *kobj,
353 struct kobj_attribute *attr,
354 const char *buf, size_t count)
356 return double_flag_store(kobj, attr, buf, count,
357 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
358 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
360 static struct kobj_attribute defrag_attr =
361 __ATTR(defrag, 0644, defrag_show, defrag_store);
363 static ssize_t use_zero_page_show(struct kobject *kobj,
364 struct kobj_attribute *attr, char *buf)
366 return single_flag_show(kobj, attr, buf,
367 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
369 static ssize_t use_zero_page_store(struct kobject *kobj,
370 struct kobj_attribute *attr, const char *buf, size_t count)
372 return single_flag_store(kobj, attr, buf, count,
373 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
375 static struct kobj_attribute use_zero_page_attr =
376 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
377 #ifdef CONFIG_DEBUG_VM
378 static ssize_t debug_cow_show(struct kobject *kobj,
379 struct kobj_attribute *attr, char *buf)
381 return single_flag_show(kobj, attr, buf,
382 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
384 static ssize_t debug_cow_store(struct kobject *kobj,
385 struct kobj_attribute *attr,
386 const char *buf, size_t count)
388 return single_flag_store(kobj, attr, buf, count,
389 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
391 static struct kobj_attribute debug_cow_attr =
392 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
393 #endif /* CONFIG_DEBUG_VM */
395 static struct attribute *hugepage_attr[] = {
398 &use_zero_page_attr.attr,
399 #ifdef CONFIG_DEBUG_VM
400 &debug_cow_attr.attr,
405 static struct attribute_group hugepage_attr_group = {
406 .attrs = hugepage_attr,
409 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
410 struct kobj_attribute *attr,
413 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
416 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
417 struct kobj_attribute *attr,
418 const char *buf, size_t count)
423 err = kstrtoul(buf, 10, &msecs);
424 if (err || msecs > UINT_MAX)
427 khugepaged_scan_sleep_millisecs = msecs;
428 wake_up_interruptible(&khugepaged_wait);
432 static struct kobj_attribute scan_sleep_millisecs_attr =
433 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
434 scan_sleep_millisecs_store);
436 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
437 struct kobj_attribute *attr,
440 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
443 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
444 struct kobj_attribute *attr,
445 const char *buf, size_t count)
450 err = kstrtoul(buf, 10, &msecs);
451 if (err || msecs > UINT_MAX)
454 khugepaged_alloc_sleep_millisecs = msecs;
455 wake_up_interruptible(&khugepaged_wait);
459 static struct kobj_attribute alloc_sleep_millisecs_attr =
460 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
461 alloc_sleep_millisecs_store);
463 static ssize_t pages_to_scan_show(struct kobject *kobj,
464 struct kobj_attribute *attr,
467 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
469 static ssize_t pages_to_scan_store(struct kobject *kobj,
470 struct kobj_attribute *attr,
471 const char *buf, size_t count)
476 err = kstrtoul(buf, 10, &pages);
477 if (err || !pages || pages > UINT_MAX)
480 khugepaged_pages_to_scan = pages;
484 static struct kobj_attribute pages_to_scan_attr =
485 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
486 pages_to_scan_store);
488 static ssize_t pages_collapsed_show(struct kobject *kobj,
489 struct kobj_attribute *attr,
492 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
494 static struct kobj_attribute pages_collapsed_attr =
495 __ATTR_RO(pages_collapsed);
497 static ssize_t full_scans_show(struct kobject *kobj,
498 struct kobj_attribute *attr,
501 return sprintf(buf, "%u\n", khugepaged_full_scans);
503 static struct kobj_attribute full_scans_attr =
504 __ATTR_RO(full_scans);
506 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
507 struct kobj_attribute *attr, char *buf)
509 return single_flag_show(kobj, attr, buf,
510 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
512 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
513 struct kobj_attribute *attr,
514 const char *buf, size_t count)
516 return single_flag_store(kobj, attr, buf, count,
517 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
519 static struct kobj_attribute khugepaged_defrag_attr =
520 __ATTR(defrag, 0644, khugepaged_defrag_show,
521 khugepaged_defrag_store);
524 * max_ptes_none controls if khugepaged should collapse hugepages over
525 * any unmapped ptes in turn potentially increasing the memory
526 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
527 * reduce the available free memory in the system as it
528 * runs. Increasing max_ptes_none will instead potentially reduce the
529 * free memory in the system during the khugepaged scan.
531 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
532 struct kobj_attribute *attr,
535 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
537 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
538 struct kobj_attribute *attr,
539 const char *buf, size_t count)
542 unsigned long max_ptes_none;
544 err = kstrtoul(buf, 10, &max_ptes_none);
545 if (err || max_ptes_none > HPAGE_PMD_NR-1)
548 khugepaged_max_ptes_none = max_ptes_none;
552 static struct kobj_attribute khugepaged_max_ptes_none_attr =
553 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
554 khugepaged_max_ptes_none_store);
556 static struct attribute *khugepaged_attr[] = {
557 &khugepaged_defrag_attr.attr,
558 &khugepaged_max_ptes_none_attr.attr,
559 &pages_to_scan_attr.attr,
560 &pages_collapsed_attr.attr,
561 &full_scans_attr.attr,
562 &scan_sleep_millisecs_attr.attr,
563 &alloc_sleep_millisecs_attr.attr,
567 static struct attribute_group khugepaged_attr_group = {
568 .attrs = khugepaged_attr,
569 .name = "khugepaged",
572 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
576 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
577 if (unlikely(!*hugepage_kobj)) {
578 pr_err("failed to create transparent hugepage kobject\n");
582 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
584 pr_err("failed to register transparent hugepage group\n");
588 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
590 pr_err("failed to register transparent hugepage group\n");
591 goto remove_hp_group;
597 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
599 kobject_put(*hugepage_kobj);
603 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
605 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
606 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
607 kobject_put(hugepage_kobj);
610 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
615 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
618 #endif /* CONFIG_SYSFS */
620 static int __init hugepage_init(void)
623 struct kobject *hugepage_kobj;
625 if (!has_transparent_hugepage()) {
626 transparent_hugepage_flags = 0;
630 err = hugepage_init_sysfs(&hugepage_kobj);
634 err = khugepaged_slab_init();
638 err = register_shrinker(&huge_zero_page_shrinker);
640 goto err_hzp_shrinker;
643 * By default disable transparent hugepages on smaller systems,
644 * where the extra memory used could hurt more than TLB overhead
645 * is likely to save. The admin can still enable it through /sys.
647 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
648 transparent_hugepage_flags = 0;
652 err = start_stop_khugepaged();
658 unregister_shrinker(&huge_zero_page_shrinker);
660 khugepaged_slab_exit();
662 hugepage_exit_sysfs(hugepage_kobj);
666 subsys_initcall(hugepage_init);
668 static int __init setup_transparent_hugepage(char *str)
673 if (!strcmp(str, "always")) {
674 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
675 &transparent_hugepage_flags);
676 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
677 &transparent_hugepage_flags);
679 } else if (!strcmp(str, "madvise")) {
680 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
681 &transparent_hugepage_flags);
682 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
683 &transparent_hugepage_flags);
685 } else if (!strcmp(str, "never")) {
686 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
687 &transparent_hugepage_flags);
688 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
689 &transparent_hugepage_flags);
694 pr_warn("transparent_hugepage= cannot parse, ignored\n");
697 __setup("transparent_hugepage=", setup_transparent_hugepage);
699 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
701 if (likely(vma->vm_flags & VM_WRITE))
702 pmd = pmd_mkwrite(pmd);
706 static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
709 entry = mk_pmd(page, prot);
710 entry = pmd_mkhuge(entry);
714 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
715 struct vm_area_struct *vma,
716 unsigned long address, pmd_t *pmd,
717 struct page *page, gfp_t gfp,
720 struct mem_cgroup *memcg;
723 unsigned long haddr = address & HPAGE_PMD_MASK;
725 VM_BUG_ON_PAGE(!PageCompound(page), page);
727 if (mem_cgroup_try_charge(page, mm, gfp, &memcg)) {
729 count_vm_event(THP_FAULT_FALLBACK);
730 return VM_FAULT_FALLBACK;
733 pgtable = pte_alloc_one(mm, haddr);
734 if (unlikely(!pgtable)) {
735 mem_cgroup_cancel_charge(page, memcg);
740 clear_huge_page(page, haddr, HPAGE_PMD_NR);
742 * The memory barrier inside __SetPageUptodate makes sure that
743 * clear_huge_page writes become visible before the set_pmd_at()
746 __SetPageUptodate(page);
748 ptl = pmd_lock(mm, pmd);
749 if (unlikely(!pmd_none(*pmd))) {
751 mem_cgroup_cancel_charge(page, memcg);
753 pte_free(mm, pgtable);
757 /* Deliver the page fault to userland */
758 if (userfaultfd_missing(vma)) {
762 mem_cgroup_cancel_charge(page, memcg);
764 pte_free(mm, pgtable);
765 ret = handle_userfault(vma, address, flags,
767 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
771 entry = mk_huge_pmd(page, vma->vm_page_prot);
772 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
773 page_add_new_anon_rmap(page, vma, haddr);
774 mem_cgroup_commit_charge(page, memcg, false);
775 lru_cache_add_active_or_unevictable(page, vma);
776 pgtable_trans_huge_deposit(mm, pmd, pgtable);
777 set_pmd_at(mm, haddr, pmd, entry);
778 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
779 atomic_long_inc(&mm->nr_ptes);
781 count_vm_event(THP_FAULT_ALLOC);
787 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
789 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
792 /* Caller must hold page table lock. */
793 bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
794 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
795 struct page *zero_page)
800 entry = mk_pmd(zero_page, vma->vm_page_prot);
801 entry = pmd_mkhuge(entry);
802 pgtable_trans_huge_deposit(mm, pmd, pgtable);
803 set_pmd_at(mm, haddr, pmd, entry);
804 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, address, 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, address, pmd, page, gfp,
872 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
873 pmd_t *pmd, unsigned long pfn, pgprot_t prot, bool write)
875 struct mm_struct *mm = vma->vm_mm;
879 ptl = pmd_lock(mm, pmd);
880 if (pmd_none(*pmd)) {
881 entry = pmd_mkhuge(pfn_pmd(pfn, prot));
883 entry = pmd_mkyoung(pmd_mkdirty(entry));
884 entry = maybe_pmd_mkwrite(entry, vma);
886 set_pmd_at(mm, addr, pmd, entry);
887 update_mmu_cache_pmd(vma, addr, pmd);
892 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
893 pmd_t *pmd, unsigned long pfn, bool write)
895 pgprot_t pgprot = vma->vm_page_prot;
897 * If we had pmd_special, we could avoid all these restrictions,
898 * but we need to be consistent with PTEs and architectures that
899 * can't support a 'special' bit.
901 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
902 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
903 (VM_PFNMAP|VM_MIXEDMAP));
904 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
905 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
907 if (addr < vma->vm_start || addr >= vma->vm_end)
908 return VM_FAULT_SIGBUS;
909 if (track_pfn_insert(vma, &pgprot, pfn))
910 return VM_FAULT_SIGBUS;
911 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
912 return VM_FAULT_NOPAGE;
915 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
917 struct vm_area_struct *vma)
919 spinlock_t *dst_ptl, *src_ptl;
920 struct page *src_page;
926 pgtable = pte_alloc_one(dst_mm, addr);
927 if (unlikely(!pgtable))
930 dst_ptl = pmd_lock(dst_mm, dst_pmd);
931 src_ptl = pmd_lockptr(src_mm, src_pmd);
932 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
936 if (unlikely(!pmd_trans_huge(pmd))) {
937 pte_free(dst_mm, pgtable);
941 * When page table lock is held, the huge zero pmd should not be
942 * under splitting since we don't split the page itself, only pmd to
945 if (is_huge_zero_pmd(pmd)) {
946 struct page *zero_page;
948 * get_huge_zero_page() will never allocate a new page here,
949 * since we already have a zero page to copy. It just takes a
952 zero_page = get_huge_zero_page();
953 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
959 if (unlikely(pmd_trans_splitting(pmd))) {
960 /* split huge page running from under us */
961 spin_unlock(src_ptl);
962 spin_unlock(dst_ptl);
963 pte_free(dst_mm, pgtable);
965 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
968 src_page = pmd_page(pmd);
969 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
971 page_dup_rmap(src_page);
972 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
974 pmdp_set_wrprotect(src_mm, addr, src_pmd);
975 pmd = pmd_mkold(pmd_wrprotect(pmd));
976 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
977 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
978 atomic_long_inc(&dst_mm->nr_ptes);
982 spin_unlock(src_ptl);
983 spin_unlock(dst_ptl);
988 void huge_pmd_set_accessed(struct mm_struct *mm,
989 struct vm_area_struct *vma,
990 unsigned long address,
991 pmd_t *pmd, pmd_t orig_pmd,
998 ptl = pmd_lock(mm, pmd);
999 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1002 entry = pmd_mkyoung(orig_pmd);
1003 haddr = address & HPAGE_PMD_MASK;
1004 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1005 update_mmu_cache_pmd(vma, address, pmd);
1012 * Save CONFIG_DEBUG_PAGEALLOC from faulting falsely on tail pages
1013 * during copy_user_huge_page()'s copy_page_rep(): in the case when
1014 * the source page gets split and a tail freed before copy completes.
1015 * Called under pmd_lock of checked pmd, so safe from splitting itself.
1017 static void get_user_huge_page(struct page *page)
1019 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1020 struct page *endpage = page + HPAGE_PMD_NR;
1022 atomic_add(HPAGE_PMD_NR, &page->_count);
1023 while (++page < endpage)
1024 get_huge_page_tail(page);
1030 static void put_user_huge_page(struct page *page)
1032 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC)) {
1033 struct page *endpage = page + HPAGE_PMD_NR;
1035 while (page < endpage)
1042 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1043 struct vm_area_struct *vma,
1044 unsigned long address,
1045 pmd_t *pmd, pmd_t orig_pmd,
1047 unsigned long haddr)
1049 struct mem_cgroup *memcg;
1054 struct page **pages;
1055 unsigned long mmun_start; /* For mmu_notifiers */
1056 unsigned long mmun_end; /* For mmu_notifiers */
1058 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1060 if (unlikely(!pages)) {
1061 ret |= VM_FAULT_OOM;
1065 for (i = 0; i < HPAGE_PMD_NR; i++) {
1066 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1068 vma, address, page_to_nid(page));
1069 if (unlikely(!pages[i] ||
1070 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1075 memcg = (void *)page_private(pages[i]);
1076 set_page_private(pages[i], 0);
1077 mem_cgroup_cancel_charge(pages[i], memcg);
1081 ret |= VM_FAULT_OOM;
1084 set_page_private(pages[i], (unsigned long)memcg);
1087 for (i = 0; i < HPAGE_PMD_NR; i++) {
1088 copy_user_highpage(pages[i], page + i,
1089 haddr + PAGE_SIZE * i, vma);
1090 __SetPageUptodate(pages[i]);
1095 mmun_end = haddr + HPAGE_PMD_SIZE;
1096 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1098 ptl = pmd_lock(mm, pmd);
1099 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1100 goto out_free_pages;
1101 VM_BUG_ON_PAGE(!PageHead(page), page);
1103 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1104 /* leave pmd empty until pte is filled */
1106 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1107 pmd_populate(mm, &_pmd, pgtable);
1109 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1111 entry = mk_pte(pages[i], vma->vm_page_prot);
1112 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1113 memcg = (void *)page_private(pages[i]);
1114 set_page_private(pages[i], 0);
1115 page_add_new_anon_rmap(pages[i], vma, haddr);
1116 mem_cgroup_commit_charge(pages[i], memcg, false);
1117 lru_cache_add_active_or_unevictable(pages[i], vma);
1118 pte = pte_offset_map(&_pmd, haddr);
1119 VM_BUG_ON(!pte_none(*pte));
1120 set_pte_at(mm, haddr, pte, entry);
1125 smp_wmb(); /* make pte visible before pmd */
1126 pmd_populate(mm, pmd, pgtable);
1127 page_remove_rmap(page);
1130 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1132 ret |= VM_FAULT_WRITE;
1140 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1141 for (i = 0; i < HPAGE_PMD_NR; i++) {
1142 memcg = (void *)page_private(pages[i]);
1143 set_page_private(pages[i], 0);
1144 mem_cgroup_cancel_charge(pages[i], memcg);
1151 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1152 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1156 struct page *page = NULL, *new_page;
1157 struct mem_cgroup *memcg;
1158 unsigned long haddr;
1159 unsigned long mmun_start; /* For mmu_notifiers */
1160 unsigned long mmun_end; /* For mmu_notifiers */
1161 gfp_t huge_gfp; /* for allocation and charge */
1163 ptl = pmd_lockptr(mm, pmd);
1164 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1165 haddr = address & HPAGE_PMD_MASK;
1166 if (is_huge_zero_pmd(orig_pmd))
1169 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1172 page = pmd_page(orig_pmd);
1173 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1174 if (page_mapcount(page) == 1) {
1176 entry = pmd_mkyoung(orig_pmd);
1177 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1178 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1179 update_mmu_cache_pmd(vma, address, pmd);
1180 ret |= VM_FAULT_WRITE;
1183 get_user_huge_page(page);
1186 if (transparent_hugepage_enabled(vma) &&
1187 !transparent_hugepage_debug_cow()) {
1188 huge_gfp = alloc_hugepage_gfpmask(transparent_hugepage_defrag(vma), 0);
1189 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1193 if (unlikely(!new_page)) {
1195 split_huge_page_pmd(vma, address, pmd);
1196 ret |= VM_FAULT_FALLBACK;
1198 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1199 pmd, orig_pmd, page, haddr);
1200 if (ret & VM_FAULT_OOM) {
1201 split_huge_page(page);
1202 ret |= VM_FAULT_FALLBACK;
1204 put_user_huge_page(page);
1206 count_vm_event(THP_FAULT_FALLBACK);
1210 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg))) {
1213 split_huge_page(page);
1214 put_user_huge_page(page);
1216 split_huge_page_pmd(vma, address, pmd);
1217 ret |= VM_FAULT_FALLBACK;
1218 count_vm_event(THP_FAULT_FALLBACK);
1222 count_vm_event(THP_FAULT_ALLOC);
1225 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1227 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1228 __SetPageUptodate(new_page);
1231 mmun_end = haddr + HPAGE_PMD_SIZE;
1232 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1236 put_user_huge_page(page);
1237 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1239 mem_cgroup_cancel_charge(new_page, memcg);
1244 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1245 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1246 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1247 page_add_new_anon_rmap(new_page, vma, haddr);
1248 mem_cgroup_commit_charge(new_page, memcg, false);
1249 lru_cache_add_active_or_unevictable(new_page, vma);
1250 set_pmd_at(mm, haddr, pmd, entry);
1251 update_mmu_cache_pmd(vma, address, pmd);
1253 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1254 put_huge_zero_page();
1256 VM_BUG_ON_PAGE(!PageHead(page), page);
1257 page_remove_rmap(page);
1260 ret |= VM_FAULT_WRITE;
1264 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1272 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1277 struct mm_struct *mm = vma->vm_mm;
1278 struct page *page = NULL;
1280 assert_spin_locked(pmd_lockptr(mm, pmd));
1282 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1285 /* Avoid dumping huge zero page */
1286 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1287 return ERR_PTR(-EFAULT);
1289 /* Full NUMA hinting faults to serialise migration in fault paths */
1290 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1293 page = pmd_page(*pmd);
1294 VM_BUG_ON_PAGE(!PageHead(page), page);
1295 if (flags & FOLL_TOUCH) {
1298 * We should set the dirty bit only for FOLL_WRITE but
1299 * for now the dirty bit in the pmd is meaningless.
1300 * And if the dirty bit will become meaningful and
1301 * we'll only set it with FOLL_WRITE, an atomic
1302 * set_bit will be required on the pmd to set the
1303 * young bit, instead of the current set_pmd_at.
1305 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1306 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1308 update_mmu_cache_pmd(vma, addr, pmd);
1310 if ((flags & FOLL_POPULATE) && (vma->vm_flags & VM_LOCKED)) {
1311 if (page->mapping && trylock_page(page)) {
1314 mlock_vma_page(page);
1318 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1319 VM_BUG_ON_PAGE(!PageCompound(page), page);
1320 if (flags & FOLL_GET)
1321 get_page_foll(page);
1327 /* NUMA hinting page fault entry point for trans huge pmds */
1328 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1329 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1332 struct anon_vma *anon_vma = NULL;
1334 unsigned long haddr = addr & HPAGE_PMD_MASK;
1335 int page_nid = -1, this_nid = numa_node_id();
1336 int target_nid, last_cpupid = -1;
1338 bool migrated = false;
1342 /* A PROT_NONE fault should not end up here */
1343 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1345 ptl = pmd_lock(mm, pmdp);
1346 if (unlikely(!pmd_same(pmd, *pmdp)))
1350 * If there are potential migrations, wait for completion and retry
1351 * without disrupting NUMA hinting information. Do not relock and
1352 * check_same as the page may no longer be mapped.
1354 if (unlikely(pmd_trans_migrating(*pmdp))) {
1355 page = pmd_page(*pmdp);
1357 wait_on_page_locked(page);
1361 page = pmd_page(pmd);
1362 BUG_ON(is_huge_zero_page(page));
1363 page_nid = page_to_nid(page);
1364 last_cpupid = page_cpupid_last(page);
1365 count_vm_numa_event(NUMA_HINT_FAULTS);
1366 if (page_nid == this_nid) {
1367 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1368 flags |= TNF_FAULT_LOCAL;
1371 /* See similar comment in do_numa_page for explanation */
1372 if (!(vma->vm_flags & VM_WRITE))
1373 flags |= TNF_NO_GROUP;
1376 * Acquire the page lock to serialise THP migrations but avoid dropping
1377 * page_table_lock if at all possible
1379 page_locked = trylock_page(page);
1380 target_nid = mpol_misplaced(page, vma, haddr);
1381 if (target_nid == -1) {
1382 /* If the page was locked, there are no parallel migrations */
1387 /* Migration could have started since the pmd_trans_migrating check */
1390 wait_on_page_locked(page);
1396 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1397 * to serialises splits
1401 anon_vma = page_lock_anon_vma_read(page);
1403 /* Confirm the PMD did not change while page_table_lock was released */
1405 if (unlikely(!pmd_same(pmd, *pmdp))) {
1412 /* Bail if we fail to protect against THP splits for any reason */
1413 if (unlikely(!anon_vma)) {
1420 * Migrate the THP to the requested node, returns with page unlocked
1421 * and access rights restored.
1424 migrated = migrate_misplaced_transhuge_page(mm, vma,
1425 pmdp, pmd, addr, page, target_nid);
1427 flags |= TNF_MIGRATED;
1428 page_nid = target_nid;
1430 flags |= TNF_MIGRATE_FAIL;
1434 BUG_ON(!PageLocked(page));
1435 was_writable = pmd_write(pmd);
1436 pmd = pmd_modify(pmd, vma->vm_page_prot);
1437 pmd = pmd_mkyoung(pmd);
1439 pmd = pmd_mkwrite(pmd);
1440 set_pmd_at(mm, haddr, pmdp, pmd);
1441 update_mmu_cache_pmd(vma, addr, pmdp);
1448 page_unlock_anon_vma_read(anon_vma);
1451 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1456 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1457 pmd_t *pmd, unsigned long addr)
1462 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1466 * For architectures like ppc64 we look at deposited pgtable
1467 * when calling pmdp_huge_get_and_clear. So do the
1468 * pgtable_trans_huge_withdraw after finishing pmdp related
1471 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1473 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1474 if (vma_is_dax(vma)) {
1475 if (is_huge_zero_pmd(orig_pmd)) {
1482 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1484 if (is_huge_zero_pmd(orig_pmd)) {
1485 atomic_long_dec(&tlb->mm->nr_ptes);
1487 put_huge_zero_page();
1489 struct page *page = pmd_page(orig_pmd);
1490 page_remove_rmap(page);
1491 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1492 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1493 VM_BUG_ON_PAGE(!PageHead(page), page);
1494 atomic_long_dec(&tlb->mm->nr_ptes);
1496 tlb_remove_page(tlb, page);
1499 pte_free(tlb->mm, pgtable);
1505 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1506 unsigned long old_addr,
1507 unsigned long new_addr, unsigned long old_end,
1508 pmd_t *old_pmd, pmd_t *new_pmd)
1510 spinlock_t *old_ptl, *new_ptl;
1514 struct mm_struct *mm = vma->vm_mm;
1516 if ((old_addr & ~HPAGE_PMD_MASK) ||
1517 (new_addr & ~HPAGE_PMD_MASK) ||
1518 old_end - old_addr < HPAGE_PMD_SIZE ||
1519 (new_vma->vm_flags & VM_NOHUGEPAGE))
1523 * The destination pmd shouldn't be established, free_pgtables()
1524 * should have release it.
1526 if (WARN_ON(!pmd_none(*new_pmd))) {
1527 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1532 * We don't have to worry about the ordering of src and dst
1533 * ptlocks because exclusive mmap_sem prevents deadlock.
1535 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1537 new_ptl = pmd_lockptr(mm, new_pmd);
1538 if (new_ptl != old_ptl)
1539 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1540 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1541 VM_BUG_ON(!pmd_none(*new_pmd));
1543 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1545 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1546 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1548 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1549 if (new_ptl != old_ptl)
1550 spin_unlock(new_ptl);
1551 spin_unlock(old_ptl);
1559 * - 0 if PMD could not be locked
1560 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1561 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1563 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1564 unsigned long addr, pgprot_t newprot, int prot_numa)
1566 struct mm_struct *mm = vma->vm_mm;
1570 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1572 bool preserve_write = prot_numa && pmd_write(*pmd);
1576 * Avoid trapping faults against the zero page. The read-only
1577 * data is likely to be read-cached on the local CPU and
1578 * local/remote hits to the zero page are not interesting.
1580 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1585 if (!prot_numa || !pmd_protnone(*pmd)) {
1586 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1587 entry = pmd_modify(entry, newprot);
1589 entry = pmd_mkwrite(entry);
1591 set_pmd_at(mm, addr, pmd, entry);
1592 BUG_ON(!preserve_write && pmd_write(entry));
1601 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1602 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1604 * Note that if it returns 1, this routine returns without unlocking page
1605 * table locks. So callers must unlock them.
1607 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1610 *ptl = pmd_lock(vma->vm_mm, pmd);
1611 if (likely(pmd_trans_huge(*pmd))) {
1612 if (unlikely(pmd_trans_splitting(*pmd))) {
1614 wait_split_huge_page(vma->anon_vma, pmd);
1617 /* Thp mapped by 'pmd' is stable, so we can
1618 * handle it as it is. */
1627 * This function returns whether a given @page is mapped onto the @address
1628 * in the virtual space of @mm.
1630 * When it's true, this function returns *pmd with holding the page table lock
1631 * and passing it back to the caller via @ptl.
1632 * If it's false, returns NULL without holding the page table lock.
1634 pmd_t *page_check_address_pmd(struct page *page,
1635 struct mm_struct *mm,
1636 unsigned long address,
1637 enum page_check_address_pmd_flag flag,
1644 if (address & ~HPAGE_PMD_MASK)
1647 pgd = pgd_offset(mm, address);
1648 if (!pgd_present(*pgd))
1650 pud = pud_offset(pgd, address);
1651 if (!pud_present(*pud))
1653 pmd = pmd_offset(pud, address);
1655 *ptl = pmd_lock(mm, pmd);
1656 if (!pmd_present(*pmd))
1658 if (pmd_page(*pmd) != page)
1661 * split_vma() may create temporary aliased mappings. There is
1662 * no risk as long as all huge pmd are found and have their
1663 * splitting bit set before __split_huge_page_refcount
1664 * runs. Finding the same huge pmd more than once during the
1665 * same rmap walk is not a problem.
1667 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1668 pmd_trans_splitting(*pmd))
1670 if (pmd_trans_huge(*pmd)) {
1671 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1672 !pmd_trans_splitting(*pmd));
1680 static int __split_huge_page_splitting(struct page *page,
1681 struct vm_area_struct *vma,
1682 unsigned long address)
1684 struct mm_struct *mm = vma->vm_mm;
1688 /* For mmu_notifiers */
1689 const unsigned long mmun_start = address;
1690 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1692 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1693 pmd = page_check_address_pmd(page, mm, address,
1694 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1697 * We can't temporarily set the pmd to null in order
1698 * to split it, the pmd must remain marked huge at all
1699 * times or the VM won't take the pmd_trans_huge paths
1700 * and it won't wait on the anon_vma->root->rwsem to
1701 * serialize against split_huge_page*.
1703 pmdp_splitting_flush(vma, address, pmd);
1708 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1713 static void __split_huge_page_refcount(struct page *page,
1714 struct list_head *list)
1717 struct zone *zone = page_zone(page);
1718 struct lruvec *lruvec;
1721 /* prevent PageLRU to go away from under us, and freeze lru stats */
1722 spin_lock_irq(&zone->lru_lock);
1723 lruvec = mem_cgroup_page_lruvec(page, zone);
1725 compound_lock(page);
1726 /* complete memcg works before add pages to LRU */
1727 mem_cgroup_split_huge_fixup(page);
1729 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1730 struct page *page_tail = page + i;
1732 /* tail_page->_mapcount cannot change */
1733 BUG_ON(page_mapcount(page_tail) < 0);
1734 tail_count += page_mapcount(page_tail);
1735 /* check for overflow */
1736 BUG_ON(tail_count < 0);
1737 BUG_ON(atomic_read(&page_tail->_count) != 0);
1739 * tail_page->_count is zero and not changing from
1740 * under us. But get_page_unless_zero() may be running
1741 * from under us on the tail_page. If we used
1742 * atomic_set() below instead of atomic_add(), we
1743 * would then run atomic_set() concurrently with
1744 * get_page_unless_zero(), and atomic_set() is
1745 * implemented in C not using locked ops. spin_unlock
1746 * on x86 sometime uses locked ops because of PPro
1747 * errata 66, 92, so unless somebody can guarantee
1748 * atomic_set() here would be safe on all archs (and
1749 * not only on x86), it's safer to use atomic_add().
1751 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1752 &page_tail->_count);
1754 /* after clearing PageTail the gup refcount can be released */
1755 smp_mb__after_atomic();
1757 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1758 page_tail->flags |= (page->flags &
1759 ((1L << PG_referenced) |
1760 (1L << PG_swapbacked) |
1761 (1L << PG_mlocked) |
1762 (1L << PG_uptodate) |
1764 (1L << PG_unevictable)));
1765 page_tail->flags |= (1L << PG_dirty);
1767 /* clear PageTail before overwriting first_page */
1771 * __split_huge_page_splitting() already set the
1772 * splitting bit in all pmd that could map this
1773 * hugepage, that will ensure no CPU can alter the
1774 * mapcount on the head page. The mapcount is only
1775 * accounted in the head page and it has to be
1776 * transferred to all tail pages in the below code. So
1777 * for this code to be safe, the split the mapcount
1778 * can't change. But that doesn't mean userland can't
1779 * keep changing and reading the page contents while
1780 * we transfer the mapcount, so the pmd splitting
1781 * status is achieved setting a reserved bit in the
1782 * pmd, not by clearing the present bit.
1784 page_tail->_mapcount = page->_mapcount;
1786 BUG_ON(page_tail->mapping);
1787 page_tail->mapping = page->mapping;
1789 page_tail->index = page->index + i;
1790 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1792 BUG_ON(!PageAnon(page_tail));
1793 BUG_ON(!PageUptodate(page_tail));
1794 BUG_ON(!PageDirty(page_tail));
1795 BUG_ON(!PageSwapBacked(page_tail));
1797 lru_add_page_tail(page, page_tail, lruvec, list);
1799 atomic_sub(tail_count, &page->_count);
1800 BUG_ON(atomic_read(&page->_count) <= 0);
1802 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1804 ClearPageCompound(page);
1805 compound_unlock(page);
1806 spin_unlock_irq(&zone->lru_lock);
1808 for (i = 1; i < HPAGE_PMD_NR; i++) {
1809 struct page *page_tail = page + i;
1810 BUG_ON(page_count(page_tail) <= 0);
1812 * Tail pages may be freed if there wasn't any mapping
1813 * like if add_to_swap() is running on a lru page that
1814 * had its mapping zapped. And freeing these pages
1815 * requires taking the lru_lock so we do the put_page
1816 * of the tail pages after the split is complete.
1818 put_page(page_tail);
1822 * Only the head page (now become a regular page) is required
1823 * to be pinned by the caller.
1825 BUG_ON(page_count(page) <= 0);
1828 static int __split_huge_page_map(struct page *page,
1829 struct vm_area_struct *vma,
1830 unsigned long address)
1832 struct mm_struct *mm = vma->vm_mm;
1837 unsigned long haddr;
1839 pmd = page_check_address_pmd(page, mm, address,
1840 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1842 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1843 pmd_populate(mm, &_pmd, pgtable);
1844 if (pmd_write(*pmd))
1845 BUG_ON(page_mapcount(page) != 1);
1848 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1850 BUG_ON(PageCompound(page+i));
1852 * Note that NUMA hinting access restrictions are not
1853 * transferred to avoid any possibility of altering
1854 * permissions across VMAs.
1856 entry = mk_pte(page + i, vma->vm_page_prot);
1857 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1858 if (!pmd_write(*pmd))
1859 entry = pte_wrprotect(entry);
1860 if (!pmd_young(*pmd))
1861 entry = pte_mkold(entry);
1862 pte = pte_offset_map(&_pmd, haddr);
1863 BUG_ON(!pte_none(*pte));
1864 set_pte_at(mm, haddr, pte, entry);
1868 smp_wmb(); /* make pte visible before pmd */
1870 * Up to this point the pmd is present and huge and
1871 * userland has the whole access to the hugepage
1872 * during the split (which happens in place). If we
1873 * overwrite the pmd with the not-huge version
1874 * pointing to the pte here (which of course we could
1875 * if all CPUs were bug free), userland could trigger
1876 * a small page size TLB miss on the small sized TLB
1877 * while the hugepage TLB entry is still established
1878 * in the huge TLB. Some CPU doesn't like that. See
1879 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1880 * Erratum 383 on page 93. Intel should be safe but is
1881 * also warns that it's only safe if the permission
1882 * and cache attributes of the two entries loaded in
1883 * the two TLB is identical (which should be the case
1884 * here). But it is generally safer to never allow
1885 * small and huge TLB entries for the same virtual
1886 * address to be loaded simultaneously. So instead of
1887 * doing "pmd_populate(); flush_tlb_range();" we first
1888 * mark the current pmd notpresent (atomically because
1889 * here the pmd_trans_huge and pmd_trans_splitting
1890 * must remain set at all times on the pmd until the
1891 * split is complete for this pmd), then we flush the
1892 * SMP TLB and finally we write the non-huge version
1893 * of the pmd entry with pmd_populate.
1895 pmdp_invalidate(vma, address, pmd);
1896 pmd_populate(mm, pmd, pgtable);
1904 /* must be called with anon_vma->root->rwsem held */
1905 static void __split_huge_page(struct page *page,
1906 struct anon_vma *anon_vma,
1907 struct list_head *list)
1909 int mapcount, mapcount2;
1910 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1911 struct anon_vma_chain *avc;
1913 BUG_ON(!PageHead(page));
1914 BUG_ON(PageTail(page));
1917 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1918 struct vm_area_struct *vma = avc->vma;
1919 unsigned long addr = vma_address(page, vma);
1920 BUG_ON(is_vma_temporary_stack(vma));
1921 mapcount += __split_huge_page_splitting(page, vma, addr);
1924 * It is critical that new vmas are added to the tail of the
1925 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1926 * and establishes a child pmd before
1927 * __split_huge_page_splitting() freezes the parent pmd (so if
1928 * we fail to prevent copy_huge_pmd() from running until the
1929 * whole __split_huge_page() is complete), we will still see
1930 * the newly established pmd of the child later during the
1931 * walk, to be able to set it as pmd_trans_splitting too.
1933 if (mapcount != page_mapcount(page)) {
1934 pr_err("mapcount %d page_mapcount %d\n",
1935 mapcount, page_mapcount(page));
1939 __split_huge_page_refcount(page, list);
1942 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1943 struct vm_area_struct *vma = avc->vma;
1944 unsigned long addr = vma_address(page, vma);
1945 BUG_ON(is_vma_temporary_stack(vma));
1946 mapcount2 += __split_huge_page_map(page, vma, addr);
1948 if (mapcount != mapcount2) {
1949 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1950 mapcount, mapcount2, page_mapcount(page));
1956 * Split a hugepage into normal pages. This doesn't change the position of head
1957 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1958 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1959 * from the hugepage.
1960 * Return 0 if the hugepage is split successfully otherwise return 1.
1962 int split_huge_page_to_list(struct page *page, struct list_head *list)
1964 struct anon_vma *anon_vma;
1967 BUG_ON(is_huge_zero_page(page));
1968 BUG_ON(!PageAnon(page));
1971 * The caller does not necessarily hold an mmap_sem that would prevent
1972 * the anon_vma disappearing so we first we take a reference to it
1973 * and then lock the anon_vma for write. This is similar to
1974 * page_lock_anon_vma_read except the write lock is taken to serialise
1975 * against parallel split or collapse operations.
1977 anon_vma = page_get_anon_vma(page);
1980 anon_vma_lock_write(anon_vma);
1983 if (!PageCompound(page))
1986 BUG_ON(!PageSwapBacked(page));
1987 __split_huge_page(page, anon_vma, list);
1988 count_vm_event(THP_SPLIT);
1990 BUG_ON(PageCompound(page));
1992 anon_vma_unlock_write(anon_vma);
1993 put_anon_vma(anon_vma);
1998 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
2000 int hugepage_madvise(struct vm_area_struct *vma,
2001 unsigned long *vm_flags, int advice)
2007 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2008 * can't handle this properly after s390_enable_sie, so we simply
2009 * ignore the madvise to prevent qemu from causing a SIGSEGV.
2011 if (mm_has_pgste(vma->vm_mm))
2015 * Be somewhat over-protective like KSM for now!
2017 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
2019 *vm_flags &= ~VM_NOHUGEPAGE;
2020 *vm_flags |= VM_HUGEPAGE;
2022 * If the vma become good for khugepaged to scan,
2023 * register it here without waiting a page fault that
2024 * may not happen any time soon.
2026 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2029 case MADV_NOHUGEPAGE:
2031 * Be somewhat over-protective like KSM for now!
2033 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
2035 *vm_flags &= ~VM_HUGEPAGE;
2036 *vm_flags |= VM_NOHUGEPAGE;
2038 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2039 * this vma even if we leave the mm registered in khugepaged if
2040 * it got registered before VM_NOHUGEPAGE was set.
2048 static int __init khugepaged_slab_init(void)
2050 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2051 sizeof(struct mm_slot),
2052 __alignof__(struct mm_slot), 0, NULL);
2059 static void __init khugepaged_slab_exit(void)
2061 kmem_cache_destroy(mm_slot_cache);
2064 static inline struct mm_slot *alloc_mm_slot(void)
2066 if (!mm_slot_cache) /* initialization failed */
2068 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2071 static inline void free_mm_slot(struct mm_slot *mm_slot)
2073 kmem_cache_free(mm_slot_cache, mm_slot);
2076 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2078 struct mm_slot *mm_slot;
2080 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2081 if (mm == mm_slot->mm)
2087 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2088 struct mm_slot *mm_slot)
2091 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2094 static inline int khugepaged_test_exit(struct mm_struct *mm)
2096 return atomic_read(&mm->mm_users) == 0;
2099 int __khugepaged_enter(struct mm_struct *mm)
2101 struct mm_slot *mm_slot;
2104 mm_slot = alloc_mm_slot();
2108 /* __khugepaged_exit() must not run from under us */
2109 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2110 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2111 free_mm_slot(mm_slot);
2115 spin_lock(&khugepaged_mm_lock);
2116 insert_to_mm_slots_hash(mm, mm_slot);
2118 * Insert just behind the scanning cursor, to let the area settle
2121 wakeup = list_empty(&khugepaged_scan.mm_head);
2122 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2123 spin_unlock(&khugepaged_mm_lock);
2125 atomic_inc(&mm->mm_count);
2127 wake_up_interruptible(&khugepaged_wait);
2132 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2133 unsigned long vm_flags)
2135 unsigned long hstart, hend;
2138 * Not yet faulted in so we will register later in the
2139 * page fault if needed.
2143 /* khugepaged not yet working on file or special mappings */
2145 VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2146 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2147 hend = vma->vm_end & HPAGE_PMD_MASK;
2149 return khugepaged_enter(vma, vm_flags);
2153 void __khugepaged_exit(struct mm_struct *mm)
2155 struct mm_slot *mm_slot;
2158 spin_lock(&khugepaged_mm_lock);
2159 mm_slot = get_mm_slot(mm);
2160 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2161 hash_del(&mm_slot->hash);
2162 list_del(&mm_slot->mm_node);
2165 spin_unlock(&khugepaged_mm_lock);
2168 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2169 free_mm_slot(mm_slot);
2171 } else if (mm_slot) {
2173 * This is required to serialize against
2174 * khugepaged_test_exit() (which is guaranteed to run
2175 * under mmap sem read mode). Stop here (after we
2176 * return all pagetables will be destroyed) until
2177 * khugepaged has finished working on the pagetables
2178 * under the mmap_sem.
2180 down_write(&mm->mmap_sem);
2181 up_write(&mm->mmap_sem);
2185 static void release_pte_page(struct page *page)
2187 /* 0 stands for page_is_file_cache(page) == false */
2188 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2190 putback_lru_page(page);
2193 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2195 while (--_pte >= pte) {
2196 pte_t pteval = *_pte;
2197 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2198 release_pte_page(pte_page(pteval));
2202 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2203 unsigned long address,
2208 int none_or_zero = 0;
2209 bool referenced = false, writable = false;
2210 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2211 _pte++, address += PAGE_SIZE) {
2212 pte_t pteval = *_pte;
2213 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2214 if (!userfaultfd_armed(vma) &&
2215 ++none_or_zero <= khugepaged_max_ptes_none)
2220 if (!pte_present(pteval))
2222 page = vm_normal_page(vma, address, pteval);
2223 if (unlikely(!page))
2226 VM_BUG_ON_PAGE(PageCompound(page), page);
2227 VM_BUG_ON_PAGE(!PageAnon(page), page);
2228 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2231 * We can do it before isolate_lru_page because the
2232 * page can't be freed from under us. NOTE: PG_lock
2233 * is needed to serialize against split_huge_page
2234 * when invoked from the VM.
2236 if (!trylock_page(page))
2240 * cannot use mapcount: can't collapse if there's a gup pin.
2241 * The page must only be referenced by the scanned process
2242 * and page swap cache.
2244 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2248 if (pte_write(pteval)) {
2251 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2256 * Page is not in the swap cache. It can be collapsed
2262 * Isolate the page to avoid collapsing an hugepage
2263 * currently in use by the VM.
2265 if (isolate_lru_page(page)) {
2269 /* 0 stands for page_is_file_cache(page) == false */
2270 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2271 VM_BUG_ON_PAGE(!PageLocked(page), page);
2272 VM_BUG_ON_PAGE(PageLRU(page), page);
2274 /* If there is no mapped pte young don't collapse the page */
2275 if (pte_young(pteval) || PageReferenced(page) ||
2276 mmu_notifier_test_young(vma->vm_mm, address))
2279 if (likely(referenced && writable))
2282 release_pte_pages(pte, _pte);
2286 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2287 struct vm_area_struct *vma,
2288 unsigned long address,
2292 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2293 pte_t pteval = *_pte;
2294 struct page *src_page;
2296 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2297 clear_user_highpage(page, address);
2298 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2299 if (is_zero_pfn(pte_pfn(pteval))) {
2301 * ptl mostly unnecessary.
2305 * paravirt calls inside pte_clear here are
2308 pte_clear(vma->vm_mm, address, _pte);
2312 src_page = pte_page(pteval);
2313 copy_user_highpage(page, src_page, address, vma);
2314 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2315 release_pte_page(src_page);
2317 * ptl mostly unnecessary, but preempt has to
2318 * be disabled to update the per-cpu stats
2319 * inside page_remove_rmap().
2323 * paravirt calls inside pte_clear here are
2326 pte_clear(vma->vm_mm, address, _pte);
2327 page_remove_rmap(src_page);
2329 free_page_and_swap_cache(src_page);
2332 address += PAGE_SIZE;
2337 static void khugepaged_alloc_sleep(void)
2339 wait_event_freezable_timeout(khugepaged_wait, false,
2340 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2343 static int khugepaged_node_load[MAX_NUMNODES];
2345 static bool khugepaged_scan_abort(int nid)
2350 * If zone_reclaim_mode is disabled, then no extra effort is made to
2351 * allocate memory locally.
2353 if (!zone_reclaim_mode)
2356 /* If there is a count for this node already, it must be acceptable */
2357 if (khugepaged_node_load[nid])
2360 for (i = 0; i < MAX_NUMNODES; i++) {
2361 if (!khugepaged_node_load[i])
2363 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2370 static int khugepaged_find_target_node(void)
2372 static int last_khugepaged_target_node = NUMA_NO_NODE;
2373 int nid, target_node = 0, max_value = 0;
2375 /* find first node with max normal pages hit */
2376 for (nid = 0; nid < MAX_NUMNODES; nid++)
2377 if (khugepaged_node_load[nid] > max_value) {
2378 max_value = khugepaged_node_load[nid];
2382 /* do some balance if several nodes have the same hit record */
2383 if (target_node <= last_khugepaged_target_node)
2384 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2386 if (max_value == khugepaged_node_load[nid]) {
2391 last_khugepaged_target_node = target_node;
2395 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2397 if (IS_ERR(*hpage)) {
2403 khugepaged_alloc_sleep();
2404 } else if (*hpage) {
2412 static struct page *
2413 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2414 struct vm_area_struct *vma, unsigned long address,
2417 VM_BUG_ON_PAGE(*hpage, *hpage);
2420 * Before allocating the hugepage, release the mmap_sem read lock.
2421 * The allocation can take potentially a long time if it involves
2422 * sync compaction, and we do not need to hold the mmap_sem during
2423 * that. We will recheck the vma after taking it again in write mode.
2425 up_read(&mm->mmap_sem);
2427 *hpage = alloc_pages_exact_node(node, gfp, HPAGE_PMD_ORDER);
2428 if (unlikely(!*hpage)) {
2429 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2430 *hpage = ERR_PTR(-ENOMEM);
2434 count_vm_event(THP_COLLAPSE_ALLOC);
2438 static int khugepaged_find_target_node(void)
2443 static inline struct page *alloc_hugepage(int defrag)
2445 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2449 static struct page *khugepaged_alloc_hugepage(bool *wait)
2454 hpage = alloc_hugepage(khugepaged_defrag());
2456 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2461 khugepaged_alloc_sleep();
2463 count_vm_event(THP_COLLAPSE_ALLOC);
2464 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2469 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2472 *hpage = khugepaged_alloc_hugepage(wait);
2474 if (unlikely(!*hpage))
2480 static struct page *
2481 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2482 struct vm_area_struct *vma, unsigned long address,
2485 up_read(&mm->mmap_sem);
2492 static bool hugepage_vma_check(struct vm_area_struct *vma)
2494 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2495 (vma->vm_flags & VM_NOHUGEPAGE))
2498 if (!vma->anon_vma || vma->vm_ops)
2500 if (is_vma_temporary_stack(vma))
2502 VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
2506 static void collapse_huge_page(struct mm_struct *mm,
2507 unsigned long address,
2508 struct page **hpage,
2509 struct vm_area_struct *vma,
2515 struct page *new_page;
2516 spinlock_t *pmd_ptl, *pte_ptl;
2518 unsigned long hstart, hend;
2519 struct mem_cgroup *memcg;
2520 unsigned long mmun_start; /* For mmu_notifiers */
2521 unsigned long mmun_end; /* For mmu_notifiers */
2524 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2526 /* Only allocate from the target node */
2527 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2530 /* release the mmap_sem read lock. */
2531 new_page = khugepaged_alloc_page(hpage, gfp, mm, vma, address, node);
2535 if (unlikely(mem_cgroup_try_charge(new_page, mm,
2540 * Prevent all access to pagetables with the exception of
2541 * gup_fast later hanlded by the ptep_clear_flush and the VM
2542 * handled by the anon_vma lock + PG_lock.
2544 down_write(&mm->mmap_sem);
2545 if (unlikely(khugepaged_test_exit(mm)))
2548 vma = find_vma(mm, address);
2551 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2552 hend = vma->vm_end & HPAGE_PMD_MASK;
2553 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2555 if (!hugepage_vma_check(vma))
2557 pmd = mm_find_pmd(mm, address);
2561 anon_vma_lock_write(vma->anon_vma);
2563 pte = pte_offset_map(pmd, address);
2564 pte_ptl = pte_lockptr(mm, pmd);
2566 mmun_start = address;
2567 mmun_end = address + HPAGE_PMD_SIZE;
2568 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2569 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2571 * After this gup_fast can't run anymore. This also removes
2572 * any huge TLB entry from the CPU so we won't allow
2573 * huge and small TLB entries for the same virtual address
2574 * to avoid the risk of CPU bugs in that area.
2576 _pmd = pmdp_collapse_flush(vma, address, pmd);
2577 spin_unlock(pmd_ptl);
2578 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2581 isolated = __collapse_huge_page_isolate(vma, address, pte);
2582 spin_unlock(pte_ptl);
2584 if (unlikely(!isolated)) {
2587 BUG_ON(!pmd_none(*pmd));
2589 * We can only use set_pmd_at when establishing
2590 * hugepmds and never for establishing regular pmds that
2591 * points to regular pagetables. Use pmd_populate for that
2593 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2594 spin_unlock(pmd_ptl);
2595 anon_vma_unlock_write(vma->anon_vma);
2600 * All pages are isolated and locked so anon_vma rmap
2601 * can't run anymore.
2603 anon_vma_unlock_write(vma->anon_vma);
2605 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2607 __SetPageUptodate(new_page);
2608 pgtable = pmd_pgtable(_pmd);
2610 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2611 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2614 * spin_lock() below is not the equivalent of smp_wmb(), so
2615 * this is needed to avoid the copy_huge_page writes to become
2616 * visible after the set_pmd_at() write.
2621 BUG_ON(!pmd_none(*pmd));
2622 page_add_new_anon_rmap(new_page, vma, address);
2623 mem_cgroup_commit_charge(new_page, memcg, false);
2624 lru_cache_add_active_or_unevictable(new_page, vma);
2625 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2626 set_pmd_at(mm, address, pmd, _pmd);
2627 update_mmu_cache_pmd(vma, address, pmd);
2628 spin_unlock(pmd_ptl);
2632 khugepaged_pages_collapsed++;
2634 up_write(&mm->mmap_sem);
2638 mem_cgroup_cancel_charge(new_page, memcg);
2642 static int khugepaged_scan_pmd(struct mm_struct *mm,
2643 struct vm_area_struct *vma,
2644 unsigned long address,
2645 struct page **hpage)
2649 int ret = 0, none_or_zero = 0;
2651 unsigned long _address;
2653 int node = NUMA_NO_NODE;
2654 bool writable = false, referenced = false;
2656 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2658 pmd = mm_find_pmd(mm, address);
2662 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2663 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2664 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2665 _pte++, _address += PAGE_SIZE) {
2666 pte_t pteval = *_pte;
2667 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2668 if (!userfaultfd_armed(vma) &&
2669 ++none_or_zero <= khugepaged_max_ptes_none)
2674 if (!pte_present(pteval))
2676 if (pte_write(pteval))
2679 page = vm_normal_page(vma, _address, pteval);
2680 if (unlikely(!page))
2683 * Record which node the original page is from and save this
2684 * information to khugepaged_node_load[].
2685 * Khupaged will allocate hugepage from the node has the max
2688 node = page_to_nid(page);
2689 if (khugepaged_scan_abort(node))
2691 khugepaged_node_load[node]++;
2692 VM_BUG_ON_PAGE(PageCompound(page), page);
2693 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2696 * cannot use mapcount: can't collapse if there's a gup pin.
2697 * The page must only be referenced by the scanned process
2698 * and page swap cache.
2700 if (page_count(page) != 1 + !!PageSwapCache(page))
2702 if (pte_young(pteval) || PageReferenced(page) ||
2703 mmu_notifier_test_young(vma->vm_mm, address))
2706 if (referenced && writable)
2709 pte_unmap_unlock(pte, ptl);
2711 node = khugepaged_find_target_node();
2712 /* collapse_huge_page will return with the mmap_sem released */
2713 collapse_huge_page(mm, address, hpage, vma, node);
2719 static void collect_mm_slot(struct mm_slot *mm_slot)
2721 struct mm_struct *mm = mm_slot->mm;
2723 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2725 if (khugepaged_test_exit(mm)) {
2727 hash_del(&mm_slot->hash);
2728 list_del(&mm_slot->mm_node);
2731 * Not strictly needed because the mm exited already.
2733 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2736 /* khugepaged_mm_lock actually not necessary for the below */
2737 free_mm_slot(mm_slot);
2742 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2743 struct page **hpage)
2744 __releases(&khugepaged_mm_lock)
2745 __acquires(&khugepaged_mm_lock)
2747 struct mm_slot *mm_slot;
2748 struct mm_struct *mm;
2749 struct vm_area_struct *vma;
2753 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2755 if (khugepaged_scan.mm_slot)
2756 mm_slot = khugepaged_scan.mm_slot;
2758 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2759 struct mm_slot, mm_node);
2760 khugepaged_scan.address = 0;
2761 khugepaged_scan.mm_slot = mm_slot;
2763 spin_unlock(&khugepaged_mm_lock);
2766 down_read(&mm->mmap_sem);
2767 if (unlikely(khugepaged_test_exit(mm)))
2770 vma = find_vma(mm, khugepaged_scan.address);
2773 for (; vma; vma = vma->vm_next) {
2774 unsigned long hstart, hend;
2777 if (unlikely(khugepaged_test_exit(mm))) {
2781 if (!hugepage_vma_check(vma)) {
2786 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2787 hend = vma->vm_end & HPAGE_PMD_MASK;
2790 if (khugepaged_scan.address > hend)
2792 if (khugepaged_scan.address < hstart)
2793 khugepaged_scan.address = hstart;
2794 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2796 while (khugepaged_scan.address < hend) {
2799 if (unlikely(khugepaged_test_exit(mm)))
2800 goto breakouterloop;
2802 VM_BUG_ON(khugepaged_scan.address < hstart ||
2803 khugepaged_scan.address + HPAGE_PMD_SIZE >
2805 ret = khugepaged_scan_pmd(mm, vma,
2806 khugepaged_scan.address,
2808 /* move to next address */
2809 khugepaged_scan.address += HPAGE_PMD_SIZE;
2810 progress += HPAGE_PMD_NR;
2812 /* we released mmap_sem so break loop */
2813 goto breakouterloop_mmap_sem;
2814 if (progress >= pages)
2815 goto breakouterloop;
2819 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2820 breakouterloop_mmap_sem:
2822 spin_lock(&khugepaged_mm_lock);
2823 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2825 * Release the current mm_slot if this mm is about to die, or
2826 * if we scanned all vmas of this mm.
2828 if (khugepaged_test_exit(mm) || !vma) {
2830 * Make sure that if mm_users is reaching zero while
2831 * khugepaged runs here, khugepaged_exit will find
2832 * mm_slot not pointing to the exiting mm.
2834 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2835 khugepaged_scan.mm_slot = list_entry(
2836 mm_slot->mm_node.next,
2837 struct mm_slot, mm_node);
2838 khugepaged_scan.address = 0;
2840 khugepaged_scan.mm_slot = NULL;
2841 khugepaged_full_scans++;
2844 collect_mm_slot(mm_slot);
2850 static int khugepaged_has_work(void)
2852 return !list_empty(&khugepaged_scan.mm_head) &&
2853 khugepaged_enabled();
2856 static int khugepaged_wait_event(void)
2858 return !list_empty(&khugepaged_scan.mm_head) ||
2859 kthread_should_stop();
2862 static void khugepaged_do_scan(void)
2864 struct page *hpage = NULL;
2865 unsigned int progress = 0, pass_through_head = 0;
2866 unsigned int pages = khugepaged_pages_to_scan;
2869 barrier(); /* write khugepaged_pages_to_scan to local stack */
2871 while (progress < pages) {
2872 if (!khugepaged_prealloc_page(&hpage, &wait))
2877 if (unlikely(kthread_should_stop() || try_to_freeze()))
2880 spin_lock(&khugepaged_mm_lock);
2881 if (!khugepaged_scan.mm_slot)
2882 pass_through_head++;
2883 if (khugepaged_has_work() &&
2884 pass_through_head < 2)
2885 progress += khugepaged_scan_mm_slot(pages - progress,
2889 spin_unlock(&khugepaged_mm_lock);
2892 if (!IS_ERR_OR_NULL(hpage))
2896 static void khugepaged_wait_work(void)
2898 if (khugepaged_has_work()) {
2899 if (!khugepaged_scan_sleep_millisecs)
2902 wait_event_freezable_timeout(khugepaged_wait,
2903 kthread_should_stop(),
2904 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2908 if (khugepaged_enabled())
2909 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2912 static int khugepaged(void *none)
2914 struct mm_slot *mm_slot;
2917 set_user_nice(current, MAX_NICE);
2919 while (!kthread_should_stop()) {
2920 khugepaged_do_scan();
2921 khugepaged_wait_work();
2924 spin_lock(&khugepaged_mm_lock);
2925 mm_slot = khugepaged_scan.mm_slot;
2926 khugepaged_scan.mm_slot = NULL;
2928 collect_mm_slot(mm_slot);
2929 spin_unlock(&khugepaged_mm_lock);
2933 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2934 unsigned long haddr, pmd_t *pmd)
2936 struct mm_struct *mm = vma->vm_mm;
2941 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2942 /* leave pmd empty until pte is filled */
2944 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2945 pmd_populate(mm, &_pmd, pgtable);
2947 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2949 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2950 entry = pte_mkspecial(entry);
2951 pte = pte_offset_map(&_pmd, haddr);
2952 VM_BUG_ON(!pte_none(*pte));
2953 set_pte_at(mm, haddr, pte, entry);
2956 smp_wmb(); /* make pte visible before pmd */
2957 pmd_populate(mm, pmd, pgtable);
2958 put_huge_zero_page();
2961 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2965 struct page *page = NULL;
2966 struct mm_struct *mm = vma->vm_mm;
2967 unsigned long haddr = address & HPAGE_PMD_MASK;
2968 unsigned long mmun_start; /* For mmu_notifiers */
2969 unsigned long mmun_end; /* For mmu_notifiers */
2971 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2974 mmun_end = haddr + HPAGE_PMD_SIZE;
2976 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2977 ptl = pmd_lock(mm, pmd);
2978 if (unlikely(!pmd_trans_huge(*pmd)))
2980 if (vma_is_dax(vma)) {
2981 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2982 if (is_huge_zero_pmd(_pmd))
2983 put_huge_zero_page();
2984 } else if (is_huge_zero_pmd(*pmd)) {
2985 __split_huge_zero_page_pmd(vma, haddr, pmd);
2987 page = pmd_page(*pmd);
2988 VM_BUG_ON_PAGE(!page_count(page), page);
2993 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2998 split_huge_page(page);
3002 * We don't always have down_write of mmap_sem here: a racing
3003 * do_huge_pmd_wp_page() might have copied-on-write to another
3004 * huge page before our split_huge_page() got the anon_vma lock.
3006 if (unlikely(pmd_trans_huge(*pmd)))
3010 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3013 struct vm_area_struct *vma;
3015 vma = find_vma(mm, address);
3016 BUG_ON(vma == NULL);
3017 split_huge_page_pmd(vma, address, pmd);
3020 static void split_huge_page_address(struct mm_struct *mm,
3021 unsigned long address)
3027 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3029 pgd = pgd_offset(mm, address);
3030 if (!pgd_present(*pgd))
3033 pud = pud_offset(pgd, address);
3034 if (!pud_present(*pud))
3037 pmd = pmd_offset(pud, address);
3038 if (!pmd_present(*pmd))
3041 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3042 * materialize from under us.
3044 split_huge_page_pmd_mm(mm, address, pmd);
3047 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3048 unsigned long start,
3053 * If the new start address isn't hpage aligned and it could
3054 * previously contain an hugepage: check if we need to split
3057 if (start & ~HPAGE_PMD_MASK &&
3058 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3059 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3060 split_huge_page_address(vma->vm_mm, start);
3063 * If the new end address isn't hpage aligned and it could
3064 * previously contain an hugepage: check if we need to split
3067 if (end & ~HPAGE_PMD_MASK &&
3068 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3069 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3070 split_huge_page_address(vma->vm_mm, end);
3073 * If we're also updating the vma->vm_next->vm_start, if the new
3074 * vm_next->vm_start isn't page aligned and it could previously
3075 * contain an hugepage: check if we need to split an huge pmd.
3077 if (adjust_next > 0) {
3078 struct vm_area_struct *next = vma->vm_next;
3079 unsigned long nstart = next->vm_start;
3080 nstart += adjust_next << PAGE_SHIFT;
3081 if (nstart & ~HPAGE_PMD_MASK &&
3082 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3083 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3084 split_huge_page_address(next->vm_mm, nstart);