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>
28 #include <linux/page_idle.h>
31 #include <asm/pgalloc.h>
35 * By default transparent hugepage support is disabled in order that avoid
36 * to risk increase the memory footprint of applications without a guaranteed
37 * benefit. When transparent hugepage support is enabled, is for all mappings,
38 * and khugepaged scans all mappings.
39 * Defrag is invoked by khugepaged hugepage allocations and by page faults
40 * for all hugepage allocations.
42 unsigned long transparent_hugepage_flags __read_mostly =
43 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
44 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
46 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
47 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
49 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
50 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
51 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
53 /* default scan 8*512 pte (or vmas) every 30 second */
54 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
55 static unsigned int khugepaged_pages_collapsed;
56 static unsigned int khugepaged_full_scans;
57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58 /* during fragmentation poll the hugepage allocator once every minute */
59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60 static struct task_struct *khugepaged_thread __read_mostly;
61 static DEFINE_MUTEX(khugepaged_mutex);
62 static DEFINE_SPINLOCK(khugepaged_mm_lock);
63 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
65 * default collapse hugepages if there is at least one pte mapped like
66 * it would have happened if the vma was large enough during page
69 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
71 static int khugepaged(void *none);
72 static int khugepaged_slab_init(void);
73 static void khugepaged_slab_exit(void);
75 #define MM_SLOTS_HASH_BITS 10
76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
78 static struct kmem_cache *mm_slot_cache __read_mostly;
81 * struct mm_slot - hash lookup from mm to mm_slot
82 * @hash: hash collision list
83 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
84 * @mm: the mm that this information is valid for
87 struct hlist_node hash;
88 struct list_head mm_node;
93 * struct khugepaged_scan - cursor for scanning
94 * @mm_head: the head of the mm list to scan
95 * @mm_slot: the current mm_slot we are scanning
96 * @address: the next address inside that to be scanned
98 * There is only the one khugepaged_scan instance of this cursor structure.
100 struct khugepaged_scan {
101 struct list_head mm_head;
102 struct mm_slot *mm_slot;
103 unsigned long address;
105 static struct khugepaged_scan khugepaged_scan = {
106 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
110 static void set_recommended_min_free_kbytes(void)
114 unsigned long recommended_min;
116 for_each_populated_zone(zone)
119 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
120 recommended_min = pageblock_nr_pages * nr_zones * 2;
123 * Make sure that on average at least two pageblocks are almost free
124 * of another type, one for a migratetype to fall back to and a
125 * second to avoid subsequent fallbacks of other types There are 3
126 * MIGRATE_TYPES we care about.
128 recommended_min += pageblock_nr_pages * nr_zones *
129 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
131 /* don't ever allow to reserve more than 5% of the lowmem */
132 recommended_min = min(recommended_min,
133 (unsigned long) nr_free_buffer_pages() / 20);
134 recommended_min <<= (PAGE_SHIFT-10);
136 if (recommended_min > min_free_kbytes) {
137 if (user_min_free_kbytes >= 0)
138 pr_info("raising min_free_kbytes from %d to %lu "
139 "to help transparent hugepage allocations\n",
140 min_free_kbytes, recommended_min);
142 min_free_kbytes = recommended_min;
144 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 (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_RECLAIM)) | extra_gfp;
792 /* Caller must hold page table lock. */
793 static 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);
1273 * FOLL_FORCE can write to even unwritable pmd's, but only
1274 * after we've gone through a COW cycle and they are dirty.
1276 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1278 return pmd_write(pmd) ||
1279 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1282 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1287 struct mm_struct *mm = vma->vm_mm;
1288 struct page *page = NULL;
1290 assert_spin_locked(pmd_lockptr(mm, pmd));
1292 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1295 /* Avoid dumping huge zero page */
1296 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1297 return ERR_PTR(-EFAULT);
1299 /* Full NUMA hinting faults to serialise migration in fault paths */
1300 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1303 page = pmd_page(*pmd);
1304 VM_BUG_ON_PAGE(!PageHead(page), page);
1305 if (flags & FOLL_TOUCH) {
1308 * We should set the dirty bit only for FOLL_WRITE but
1309 * for now the dirty bit in the pmd is meaningless.
1310 * And if the dirty bit will become meaningful and
1311 * we'll only set it with FOLL_WRITE, an atomic
1312 * set_bit will be required on the pmd to set the
1313 * young bit, instead of the current set_pmd_at.
1315 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1316 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1318 update_mmu_cache_pmd(vma, addr, pmd);
1320 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1321 if (page->mapping && trylock_page(page)) {
1324 mlock_vma_page(page);
1328 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1329 VM_BUG_ON_PAGE(!PageCompound(page), page);
1330 if (flags & FOLL_GET)
1331 get_page_foll(page);
1337 /* NUMA hinting page fault entry point for trans huge pmds */
1338 int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1339 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1342 struct anon_vma *anon_vma = NULL;
1344 unsigned long haddr = addr & HPAGE_PMD_MASK;
1345 int page_nid = -1, this_nid = numa_node_id();
1346 int target_nid, last_cpupid = -1;
1348 bool migrated = false;
1352 /* A PROT_NONE fault should not end up here */
1353 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1355 ptl = pmd_lock(mm, pmdp);
1356 if (unlikely(!pmd_same(pmd, *pmdp)))
1360 * If there are potential migrations, wait for completion and retry
1361 * without disrupting NUMA hinting information. Do not relock and
1362 * check_same as the page may no longer be mapped.
1364 if (unlikely(pmd_trans_migrating(*pmdp))) {
1365 page = pmd_page(*pmdp);
1366 if (!get_page_unless_zero(page))
1369 wait_on_page_locked(page);
1374 page = pmd_page(pmd);
1375 BUG_ON(is_huge_zero_page(page));
1376 page_nid = page_to_nid(page);
1377 last_cpupid = page_cpupid_last(page);
1378 count_vm_numa_event(NUMA_HINT_FAULTS);
1379 if (page_nid == this_nid) {
1380 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1381 flags |= TNF_FAULT_LOCAL;
1384 /* See similar comment in do_numa_page for explanation */
1385 if (!(vma->vm_flags & VM_WRITE))
1386 flags |= TNF_NO_GROUP;
1389 * Acquire the page lock to serialise THP migrations but avoid dropping
1390 * page_table_lock if at all possible
1392 page_locked = trylock_page(page);
1393 target_nid = mpol_misplaced(page, vma, haddr);
1394 if (target_nid == -1) {
1395 /* If the page was locked, there are no parallel migrations */
1400 /* Migration could have started since the pmd_trans_migrating check */
1402 if (!get_page_unless_zero(page))
1405 wait_on_page_locked(page);
1412 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1413 * to serialises splits
1417 anon_vma = page_lock_anon_vma_read(page);
1419 /* Confirm the PMD did not change while page_table_lock was released */
1421 if (unlikely(!pmd_same(pmd, *pmdp))) {
1428 /* Bail if we fail to protect against THP splits for any reason */
1429 if (unlikely(!anon_vma)) {
1436 * Migrate the THP to the requested node, returns with page unlocked
1437 * and access rights restored.
1440 migrated = migrate_misplaced_transhuge_page(mm, vma,
1441 pmdp, pmd, addr, page, target_nid);
1443 flags |= TNF_MIGRATED;
1444 page_nid = target_nid;
1446 flags |= TNF_MIGRATE_FAIL;
1450 BUG_ON(!PageLocked(page));
1451 was_writable = pmd_write(pmd);
1452 pmd = pmd_modify(pmd, vma->vm_page_prot);
1453 pmd = pmd_mkyoung(pmd);
1455 pmd = pmd_mkwrite(pmd);
1456 set_pmd_at(mm, haddr, pmdp, pmd);
1457 update_mmu_cache_pmd(vma, addr, pmdp);
1464 page_unlock_anon_vma_read(anon_vma);
1467 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1472 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1473 pmd_t *pmd, unsigned long addr)
1478 if (__pmd_trans_huge_lock(pmd, vma, &ptl) != 1)
1481 * For architectures like ppc64 we look at deposited pgtable
1482 * when calling pmdp_huge_get_and_clear. So do the
1483 * pgtable_trans_huge_withdraw after finishing pmdp related
1486 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1488 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1489 if (vma_is_dax(vma)) {
1491 if (is_huge_zero_pmd(orig_pmd))
1492 put_huge_zero_page();
1493 } else if (is_huge_zero_pmd(orig_pmd)) {
1494 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1495 atomic_long_dec(&tlb->mm->nr_ptes);
1497 put_huge_zero_page();
1499 struct page *page = pmd_page(orig_pmd);
1500 page_remove_rmap(page);
1501 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1502 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1503 VM_BUG_ON_PAGE(!PageHead(page), page);
1504 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1505 atomic_long_dec(&tlb->mm->nr_ptes);
1507 tlb_remove_page(tlb, page);
1512 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1513 unsigned long old_addr,
1514 unsigned long new_addr, unsigned long old_end,
1515 pmd_t *old_pmd, pmd_t *new_pmd)
1517 spinlock_t *old_ptl, *new_ptl;
1521 struct mm_struct *mm = vma->vm_mm;
1523 if ((old_addr & ~HPAGE_PMD_MASK) ||
1524 (new_addr & ~HPAGE_PMD_MASK) ||
1525 old_end - old_addr < HPAGE_PMD_SIZE ||
1526 (new_vma->vm_flags & VM_NOHUGEPAGE))
1530 * The destination pmd shouldn't be established, free_pgtables()
1531 * should have release it.
1533 if (WARN_ON(!pmd_none(*new_pmd))) {
1534 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1539 * We don't have to worry about the ordering of src and dst
1540 * ptlocks because exclusive mmap_sem prevents deadlock.
1542 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1544 new_ptl = pmd_lockptr(mm, new_pmd);
1545 if (new_ptl != old_ptl)
1546 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1547 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1548 VM_BUG_ON(!pmd_none(*new_pmd));
1550 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1552 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1553 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1555 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1556 if (new_ptl != old_ptl)
1557 spin_unlock(new_ptl);
1558 spin_unlock(old_ptl);
1566 * - 0 if PMD could not be locked
1567 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1568 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1570 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1571 unsigned long addr, pgprot_t newprot, int prot_numa)
1573 struct mm_struct *mm = vma->vm_mm;
1577 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1579 bool preserve_write = prot_numa && pmd_write(*pmd);
1583 * Avoid trapping faults against the zero page. The read-only
1584 * data is likely to be read-cached on the local CPU and
1585 * local/remote hits to the zero page are not interesting.
1587 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1592 if (!prot_numa || !pmd_protnone(*pmd)) {
1593 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1594 entry = pmd_modify(entry, newprot);
1596 entry = pmd_mkwrite(entry);
1598 set_pmd_at(mm, addr, pmd, entry);
1599 BUG_ON(!preserve_write && pmd_write(entry));
1608 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1609 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1611 * Note that if it returns 1, this routine returns without unlocking page
1612 * table locks. So callers must unlock them.
1614 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1617 *ptl = pmd_lock(vma->vm_mm, pmd);
1618 if (likely(pmd_trans_huge(*pmd))) {
1619 if (unlikely(pmd_trans_splitting(*pmd))) {
1621 wait_split_huge_page(vma->anon_vma, pmd);
1624 /* Thp mapped by 'pmd' is stable, so we can
1625 * handle it as it is. */
1634 * This function returns whether a given @page is mapped onto the @address
1635 * in the virtual space of @mm.
1637 * When it's true, this function returns *pmd with holding the page table lock
1638 * and passing it back to the caller via @ptl.
1639 * If it's false, returns NULL without holding the page table lock.
1641 pmd_t *page_check_address_pmd(struct page *page,
1642 struct mm_struct *mm,
1643 unsigned long address,
1644 enum page_check_address_pmd_flag flag,
1651 if (address & ~HPAGE_PMD_MASK)
1654 pgd = pgd_offset(mm, address);
1655 if (!pgd_present(*pgd))
1657 pud = pud_offset(pgd, address);
1658 if (!pud_present(*pud))
1660 pmd = pmd_offset(pud, address);
1662 *ptl = pmd_lock(mm, pmd);
1663 if (!pmd_present(*pmd))
1665 if (pmd_page(*pmd) != page)
1668 * split_vma() may create temporary aliased mappings. There is
1669 * no risk as long as all huge pmd are found and have their
1670 * splitting bit set before __split_huge_page_refcount
1671 * runs. Finding the same huge pmd more than once during the
1672 * same rmap walk is not a problem.
1674 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1675 pmd_trans_splitting(*pmd))
1677 if (pmd_trans_huge(*pmd)) {
1678 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1679 !pmd_trans_splitting(*pmd));
1687 static int __split_huge_page_splitting(struct page *page,
1688 struct vm_area_struct *vma,
1689 unsigned long address)
1691 struct mm_struct *mm = vma->vm_mm;
1695 /* For mmu_notifiers */
1696 const unsigned long mmun_start = address;
1697 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1699 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1700 pmd = page_check_address_pmd(page, mm, address,
1701 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1704 * We can't temporarily set the pmd to null in order
1705 * to split it, the pmd must remain marked huge at all
1706 * times or the VM won't take the pmd_trans_huge paths
1707 * and it won't wait on the anon_vma->root->rwsem to
1708 * serialize against split_huge_page*.
1710 pmdp_splitting_flush(vma, address, pmd);
1715 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1720 static void __split_huge_page_refcount(struct page *page,
1721 struct list_head *list)
1724 struct zone *zone = page_zone(page);
1725 struct lruvec *lruvec;
1728 /* prevent PageLRU to go away from under us, and freeze lru stats */
1729 spin_lock_irq(&zone->lru_lock);
1730 lruvec = mem_cgroup_page_lruvec(page, zone);
1732 compound_lock(page);
1733 /* complete memcg works before add pages to LRU */
1734 mem_cgroup_split_huge_fixup(page);
1736 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1737 struct page *page_tail = page + i;
1739 /* tail_page->_mapcount cannot change */
1740 BUG_ON(page_mapcount(page_tail) < 0);
1741 tail_count += page_mapcount(page_tail);
1742 /* check for overflow */
1743 BUG_ON(tail_count < 0);
1744 BUG_ON(atomic_read(&page_tail->_count) != 0);
1746 * tail_page->_count is zero and not changing from
1747 * under us. But get_page_unless_zero() may be running
1748 * from under us on the tail_page. If we used
1749 * atomic_set() below instead of atomic_add(), we
1750 * would then run atomic_set() concurrently with
1751 * get_page_unless_zero(), and atomic_set() is
1752 * implemented in C not using locked ops. spin_unlock
1753 * on x86 sometime uses locked ops because of PPro
1754 * errata 66, 92, so unless somebody can guarantee
1755 * atomic_set() here would be safe on all archs (and
1756 * not only on x86), it's safer to use atomic_add().
1758 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1759 &page_tail->_count);
1761 /* after clearing PageTail the gup refcount can be released */
1762 smp_mb__after_atomic();
1764 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1765 page_tail->flags |= (page->flags &
1766 ((1L << PG_referenced) |
1767 (1L << PG_swapbacked) |
1768 (1L << PG_mlocked) |
1769 (1L << PG_uptodate) |
1771 (1L << PG_unevictable)));
1772 page_tail->flags |= (1L << PG_dirty);
1774 clear_compound_head(page_tail);
1776 if (page_is_young(page))
1777 set_page_young(page_tail);
1778 if (page_is_idle(page))
1779 set_page_idle(page_tail);
1782 * __split_huge_page_splitting() already set the
1783 * splitting bit in all pmd that could map this
1784 * hugepage, that will ensure no CPU can alter the
1785 * mapcount on the head page. The mapcount is only
1786 * accounted in the head page and it has to be
1787 * transferred to all tail pages in the below code. So
1788 * for this code to be safe, the split the mapcount
1789 * can't change. But that doesn't mean userland can't
1790 * keep changing and reading the page contents while
1791 * we transfer the mapcount, so the pmd splitting
1792 * status is achieved setting a reserved bit in the
1793 * pmd, not by clearing the present bit.
1795 page_tail->_mapcount = page->_mapcount;
1797 BUG_ON(page_tail->mapping);
1798 page_tail->mapping = page->mapping;
1800 page_tail->index = page->index + i;
1801 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1803 BUG_ON(!PageAnon(page_tail));
1804 BUG_ON(!PageUptodate(page_tail));
1805 BUG_ON(!PageDirty(page_tail));
1806 BUG_ON(!PageSwapBacked(page_tail));
1808 lru_add_page_tail(page, page_tail, lruvec, list);
1810 atomic_sub(tail_count, &page->_count);
1811 BUG_ON(atomic_read(&page->_count) <= 0);
1813 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1815 ClearPageCompound(page);
1816 compound_unlock(page);
1817 spin_unlock_irq(&zone->lru_lock);
1819 for (i = 1; i < HPAGE_PMD_NR; i++) {
1820 struct page *page_tail = page + i;
1821 BUG_ON(page_count(page_tail) <= 0);
1823 * Tail pages may be freed if there wasn't any mapping
1824 * like if add_to_swap() is running on a lru page that
1825 * had its mapping zapped. And freeing these pages
1826 * requires taking the lru_lock so we do the put_page
1827 * of the tail pages after the split is complete.
1829 put_page(page_tail);
1833 * Only the head page (now become a regular page) is required
1834 * to be pinned by the caller.
1836 BUG_ON(page_count(page) <= 0);
1839 static int __split_huge_page_map(struct page *page,
1840 struct vm_area_struct *vma,
1841 unsigned long address)
1843 struct mm_struct *mm = vma->vm_mm;
1848 unsigned long haddr;
1850 pmd = page_check_address_pmd(page, mm, address,
1851 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1853 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1854 pmd_populate(mm, &_pmd, pgtable);
1855 if (pmd_write(*pmd))
1856 BUG_ON(page_mapcount(page) != 1);
1859 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1861 BUG_ON(PageCompound(page+i));
1863 * Note that NUMA hinting access restrictions are not
1864 * transferred to avoid any possibility of altering
1865 * permissions across VMAs.
1867 entry = mk_pte(page + i, vma->vm_page_prot);
1868 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1869 if (!pmd_write(*pmd))
1870 entry = pte_wrprotect(entry);
1871 if (!pmd_young(*pmd))
1872 entry = pte_mkold(entry);
1873 pte = pte_offset_map(&_pmd, haddr);
1874 BUG_ON(!pte_none(*pte));
1875 set_pte_at(mm, haddr, pte, entry);
1879 smp_wmb(); /* make pte visible before pmd */
1881 * Up to this point the pmd is present and huge and
1882 * userland has the whole access to the hugepage
1883 * during the split (which happens in place). If we
1884 * overwrite the pmd with the not-huge version
1885 * pointing to the pte here (which of course we could
1886 * if all CPUs were bug free), userland could trigger
1887 * a small page size TLB miss on the small sized TLB
1888 * while the hugepage TLB entry is still established
1889 * in the huge TLB. Some CPU doesn't like that. See
1890 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1891 * Erratum 383 on page 93. Intel should be safe but is
1892 * also warns that it's only safe if the permission
1893 * and cache attributes of the two entries loaded in
1894 * the two TLB is identical (which should be the case
1895 * here). But it is generally safer to never allow
1896 * small and huge TLB entries for the same virtual
1897 * address to be loaded simultaneously. So instead of
1898 * doing "pmd_populate(); flush_pmd_tlb_range();" we first
1899 * mark the current pmd notpresent (atomically because
1900 * here the pmd_trans_huge and pmd_trans_splitting
1901 * must remain set at all times on the pmd until the
1902 * split is complete for this pmd), then we flush the
1903 * SMP TLB and finally we write the non-huge version
1904 * of the pmd entry with pmd_populate.
1906 pmdp_invalidate(vma, address, pmd);
1907 pmd_populate(mm, pmd, pgtable);
1915 /* must be called with anon_vma->root->rwsem held */
1916 static void __split_huge_page(struct page *page,
1917 struct anon_vma *anon_vma,
1918 struct list_head *list)
1920 int mapcount, mapcount2;
1921 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1922 struct anon_vma_chain *avc;
1924 BUG_ON(!PageHead(page));
1925 BUG_ON(PageTail(page));
1928 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1929 struct vm_area_struct *vma = avc->vma;
1930 unsigned long addr = vma_address(page, vma);
1931 BUG_ON(is_vma_temporary_stack(vma));
1932 mapcount += __split_huge_page_splitting(page, vma, addr);
1935 * It is critical that new vmas are added to the tail of the
1936 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1937 * and establishes a child pmd before
1938 * __split_huge_page_splitting() freezes the parent pmd (so if
1939 * we fail to prevent copy_huge_pmd() from running until the
1940 * whole __split_huge_page() is complete), we will still see
1941 * the newly established pmd of the child later during the
1942 * walk, to be able to set it as pmd_trans_splitting too.
1944 if (mapcount != page_mapcount(page)) {
1945 pr_err("mapcount %d page_mapcount %d\n",
1946 mapcount, page_mapcount(page));
1950 __split_huge_page_refcount(page, list);
1953 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1954 struct vm_area_struct *vma = avc->vma;
1955 unsigned long addr = vma_address(page, vma);
1956 BUG_ON(is_vma_temporary_stack(vma));
1957 mapcount2 += __split_huge_page_map(page, vma, addr);
1959 if (mapcount != mapcount2) {
1960 pr_err("mapcount %d mapcount2 %d page_mapcount %d\n",
1961 mapcount, mapcount2, page_mapcount(page));
1967 * Split a hugepage into normal pages. This doesn't change the position of head
1968 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1969 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1970 * from the hugepage.
1971 * Return 0 if the hugepage is split successfully otherwise return 1.
1973 int split_huge_page_to_list(struct page *page, struct list_head *list)
1975 struct anon_vma *anon_vma;
1978 BUG_ON(is_huge_zero_page(page));
1979 BUG_ON(!PageAnon(page));
1982 * The caller does not necessarily hold an mmap_sem that would prevent
1983 * the anon_vma disappearing so we first we take a reference to it
1984 * and then lock the anon_vma for write. This is similar to
1985 * page_lock_anon_vma_read except the write lock is taken to serialise
1986 * against parallel split or collapse operations.
1988 anon_vma = page_get_anon_vma(page);
1991 anon_vma_lock_write(anon_vma);
1994 if (!PageCompound(page))
1997 BUG_ON(!PageSwapBacked(page));
1998 __split_huge_page(page, anon_vma, list);
1999 count_vm_event(THP_SPLIT);
2001 BUG_ON(PageCompound(page));
2003 anon_vma_unlock_write(anon_vma);
2004 put_anon_vma(anon_vma);
2009 #define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
2011 int hugepage_madvise(struct vm_area_struct *vma,
2012 unsigned long *vm_flags, int advice)
2018 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
2019 * can't handle this properly after s390_enable_sie, so we simply
2020 * ignore the madvise to prevent qemu from causing a SIGSEGV.
2022 if (mm_has_pgste(vma->vm_mm))
2026 * Be somewhat over-protective like KSM for now!
2028 if (*vm_flags & VM_NO_THP)
2030 *vm_flags &= ~VM_NOHUGEPAGE;
2031 *vm_flags |= VM_HUGEPAGE;
2033 * If the vma become good for khugepaged to scan,
2034 * register it here without waiting a page fault that
2035 * may not happen any time soon.
2037 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
2040 case MADV_NOHUGEPAGE:
2042 * Be somewhat over-protective like KSM for now!
2044 if (*vm_flags & VM_NO_THP)
2046 *vm_flags &= ~VM_HUGEPAGE;
2047 *vm_flags |= VM_NOHUGEPAGE;
2049 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
2050 * this vma even if we leave the mm registered in khugepaged if
2051 * it got registered before VM_NOHUGEPAGE was set.
2059 static int __init khugepaged_slab_init(void)
2061 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
2062 sizeof(struct mm_slot),
2063 __alignof__(struct mm_slot), 0, NULL);
2070 static void __init khugepaged_slab_exit(void)
2072 kmem_cache_destroy(mm_slot_cache);
2075 static inline struct mm_slot *alloc_mm_slot(void)
2077 if (!mm_slot_cache) /* initialization failed */
2079 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
2082 static inline void free_mm_slot(struct mm_slot *mm_slot)
2084 kmem_cache_free(mm_slot_cache, mm_slot);
2087 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
2089 struct mm_slot *mm_slot;
2091 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
2092 if (mm == mm_slot->mm)
2098 static void insert_to_mm_slots_hash(struct mm_struct *mm,
2099 struct mm_slot *mm_slot)
2102 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
2105 static inline int khugepaged_test_exit(struct mm_struct *mm)
2107 return atomic_read(&mm->mm_users) == 0;
2110 int __khugepaged_enter(struct mm_struct *mm)
2112 struct mm_slot *mm_slot;
2115 mm_slot = alloc_mm_slot();
2119 /* __khugepaged_exit() must not run from under us */
2120 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
2121 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2122 free_mm_slot(mm_slot);
2126 spin_lock(&khugepaged_mm_lock);
2127 insert_to_mm_slots_hash(mm, mm_slot);
2129 * Insert just behind the scanning cursor, to let the area settle
2132 wakeup = list_empty(&khugepaged_scan.mm_head);
2133 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2134 spin_unlock(&khugepaged_mm_lock);
2136 atomic_inc(&mm->mm_count);
2138 wake_up_interruptible(&khugepaged_wait);
2143 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
2144 unsigned long vm_flags)
2146 unsigned long hstart, hend;
2149 * Not yet faulted in so we will register later in the
2150 * page fault if needed.
2153 if (vma->vm_ops || (vm_flags & VM_NO_THP))
2154 /* khugepaged not yet working on file or special mappings */
2156 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2157 hend = vma->vm_end & HPAGE_PMD_MASK;
2159 return khugepaged_enter(vma, vm_flags);
2163 void __khugepaged_exit(struct mm_struct *mm)
2165 struct mm_slot *mm_slot;
2168 spin_lock(&khugepaged_mm_lock);
2169 mm_slot = get_mm_slot(mm);
2170 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2171 hash_del(&mm_slot->hash);
2172 list_del(&mm_slot->mm_node);
2175 spin_unlock(&khugepaged_mm_lock);
2178 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2179 free_mm_slot(mm_slot);
2181 } else if (mm_slot) {
2183 * This is required to serialize against
2184 * khugepaged_test_exit() (which is guaranteed to run
2185 * under mmap sem read mode). Stop here (after we
2186 * return all pagetables will be destroyed) until
2187 * khugepaged has finished working on the pagetables
2188 * under the mmap_sem.
2190 down_write(&mm->mmap_sem);
2191 up_write(&mm->mmap_sem);
2195 static void release_pte_page(struct page *page)
2197 /* 0 stands for page_is_file_cache(page) == false */
2198 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2200 putback_lru_page(page);
2203 static void release_pte_pages(pte_t *pte, pte_t *_pte)
2205 while (--_pte >= pte) {
2206 pte_t pteval = *_pte;
2207 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2208 release_pte_page(pte_page(pteval));
2212 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2213 unsigned long address,
2218 int none_or_zero = 0;
2219 bool referenced = false, writable = false;
2220 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2221 _pte++, address += PAGE_SIZE) {
2222 pte_t pteval = *_pte;
2223 if (pte_none(pteval) || (pte_present(pteval) &&
2224 is_zero_pfn(pte_pfn(pteval)))) {
2225 if (!userfaultfd_armed(vma) &&
2226 ++none_or_zero <= khugepaged_max_ptes_none)
2231 if (!pte_present(pteval))
2233 page = vm_normal_page(vma, address, pteval);
2234 if (unlikely(!page))
2237 VM_BUG_ON_PAGE(PageCompound(page), page);
2238 VM_BUG_ON_PAGE(!PageAnon(page), page);
2239 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2242 * We can do it before isolate_lru_page because the
2243 * page can't be freed from under us. NOTE: PG_lock
2244 * is needed to serialize against split_huge_page
2245 * when invoked from the VM.
2247 if (!trylock_page(page))
2251 * cannot use mapcount: can't collapse if there's a gup pin.
2252 * The page must only be referenced by the scanned process
2253 * and page swap cache.
2255 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2259 if (pte_write(pteval)) {
2262 if (PageSwapCache(page) && !reuse_swap_page(page)) {
2267 * Page is not in the swap cache. It can be collapsed
2273 * Isolate the page to avoid collapsing an hugepage
2274 * currently in use by the VM.
2276 if (isolate_lru_page(page)) {
2280 /* 0 stands for page_is_file_cache(page) == false */
2281 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2282 VM_BUG_ON_PAGE(!PageLocked(page), page);
2283 VM_BUG_ON_PAGE(PageLRU(page), page);
2285 /* If there is no mapped pte young don't collapse the page */
2286 if (pte_young(pteval) ||
2287 page_is_young(page) || PageReferenced(page) ||
2288 mmu_notifier_test_young(vma->vm_mm, address))
2291 if (likely(referenced && writable))
2294 release_pte_pages(pte, _pte);
2298 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2299 struct vm_area_struct *vma,
2300 unsigned long address,
2304 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2305 pte_t pteval = *_pte;
2306 struct page *src_page;
2308 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2309 clear_user_highpage(page, address);
2310 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2311 if (is_zero_pfn(pte_pfn(pteval))) {
2313 * ptl mostly unnecessary.
2317 * paravirt calls inside pte_clear here are
2320 pte_clear(vma->vm_mm, address, _pte);
2324 src_page = pte_page(pteval);
2325 copy_user_highpage(page, src_page, address, vma);
2326 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2327 release_pte_page(src_page);
2329 * ptl mostly unnecessary, but preempt has to
2330 * be disabled to update the per-cpu stats
2331 * inside page_remove_rmap().
2335 * paravirt calls inside pte_clear here are
2338 pte_clear(vma->vm_mm, address, _pte);
2339 page_remove_rmap(src_page);
2341 free_page_and_swap_cache(src_page);
2344 address += PAGE_SIZE;
2349 static void khugepaged_alloc_sleep(void)
2353 add_wait_queue(&khugepaged_wait, &wait);
2354 freezable_schedule_timeout_interruptible(
2355 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2356 remove_wait_queue(&khugepaged_wait, &wait);
2359 static int khugepaged_node_load[MAX_NUMNODES];
2361 static bool khugepaged_scan_abort(int nid)
2366 * If zone_reclaim_mode is disabled, then no extra effort is made to
2367 * allocate memory locally.
2369 if (!zone_reclaim_mode)
2372 /* If there is a count for this node already, it must be acceptable */
2373 if (khugepaged_node_load[nid])
2376 for (i = 0; i < MAX_NUMNODES; i++) {
2377 if (!khugepaged_node_load[i])
2379 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2386 static int khugepaged_find_target_node(void)
2388 static int last_khugepaged_target_node = NUMA_NO_NODE;
2389 int nid, target_node = 0, max_value = 0;
2391 /* find first node with max normal pages hit */
2392 for (nid = 0; nid < MAX_NUMNODES; nid++)
2393 if (khugepaged_node_load[nid] > max_value) {
2394 max_value = khugepaged_node_load[nid];
2398 /* do some balance if several nodes have the same hit record */
2399 if (target_node <= last_khugepaged_target_node)
2400 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2402 if (max_value == khugepaged_node_load[nid]) {
2407 last_khugepaged_target_node = target_node;
2411 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2413 if (IS_ERR(*hpage)) {
2419 khugepaged_alloc_sleep();
2420 } else if (*hpage) {
2428 static struct page *
2429 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2430 unsigned long address, int node)
2432 VM_BUG_ON_PAGE(*hpage, *hpage);
2435 * Before allocating the hugepage, release the mmap_sem read lock.
2436 * The allocation can take potentially a long time if it involves
2437 * sync compaction, and we do not need to hold the mmap_sem during
2438 * that. We will recheck the vma after taking it again in write mode.
2440 up_read(&mm->mmap_sem);
2442 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2443 if (unlikely(!*hpage)) {
2444 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2445 *hpage = ERR_PTR(-ENOMEM);
2449 count_vm_event(THP_COLLAPSE_ALLOC);
2453 static int khugepaged_find_target_node(void)
2458 static inline struct page *alloc_hugepage(int defrag)
2460 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2464 static struct page *khugepaged_alloc_hugepage(bool *wait)
2469 hpage = alloc_hugepage(khugepaged_defrag());
2471 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2476 khugepaged_alloc_sleep();
2478 count_vm_event(THP_COLLAPSE_ALLOC);
2479 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2484 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2487 *hpage = khugepaged_alloc_hugepage(wait);
2489 if (unlikely(!*hpage))
2495 static struct page *
2496 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2497 unsigned long address, int node)
2499 up_read(&mm->mmap_sem);
2506 static bool hugepage_vma_check(struct vm_area_struct *vma)
2508 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2509 (vma->vm_flags & VM_NOHUGEPAGE))
2512 if (!vma->anon_vma || vma->vm_ops)
2514 if (is_vma_temporary_stack(vma))
2516 return !(vma->vm_flags & VM_NO_THP);
2519 static void collapse_huge_page(struct mm_struct *mm,
2520 unsigned long address,
2521 struct page **hpage,
2522 struct vm_area_struct *vma,
2528 struct page *new_page;
2529 spinlock_t *pmd_ptl, *pte_ptl;
2531 unsigned long hstart, hend;
2532 struct mem_cgroup *memcg;
2533 unsigned long mmun_start; /* For mmu_notifiers */
2534 unsigned long mmun_end; /* For mmu_notifiers */
2537 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2539 /* Only allocate from the target node */
2540 gfp = alloc_hugepage_gfpmask(khugepaged_defrag(), __GFP_OTHER_NODE) |
2543 /* release the mmap_sem read lock. */
2544 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2548 if (unlikely(mem_cgroup_try_charge(new_page, mm,
2553 * Prevent all access to pagetables with the exception of
2554 * gup_fast later hanlded by the ptep_clear_flush and the VM
2555 * handled by the anon_vma lock + PG_lock.
2557 down_write(&mm->mmap_sem);
2558 if (unlikely(khugepaged_test_exit(mm)))
2561 vma = find_vma(mm, address);
2564 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2565 hend = vma->vm_end & HPAGE_PMD_MASK;
2566 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2568 if (!hugepage_vma_check(vma))
2570 pmd = mm_find_pmd(mm, address);
2574 anon_vma_lock_write(vma->anon_vma);
2576 pte = pte_offset_map(pmd, address);
2577 pte_ptl = pte_lockptr(mm, pmd);
2579 mmun_start = address;
2580 mmun_end = address + HPAGE_PMD_SIZE;
2581 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2582 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2584 * After this gup_fast can't run anymore. This also removes
2585 * any huge TLB entry from the CPU so we won't allow
2586 * huge and small TLB entries for the same virtual address
2587 * to avoid the risk of CPU bugs in that area.
2589 _pmd = pmdp_collapse_flush(vma, address, pmd);
2590 spin_unlock(pmd_ptl);
2591 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2594 isolated = __collapse_huge_page_isolate(vma, address, pte);
2595 spin_unlock(pte_ptl);
2597 if (unlikely(!isolated)) {
2600 BUG_ON(!pmd_none(*pmd));
2602 * We can only use set_pmd_at when establishing
2603 * hugepmds and never for establishing regular pmds that
2604 * points to regular pagetables. Use pmd_populate for that
2606 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2607 spin_unlock(pmd_ptl);
2608 anon_vma_unlock_write(vma->anon_vma);
2613 * All pages are isolated and locked so anon_vma rmap
2614 * can't run anymore.
2616 anon_vma_unlock_write(vma->anon_vma);
2618 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2620 __SetPageUptodate(new_page);
2621 pgtable = pmd_pgtable(_pmd);
2623 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2624 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2627 * spin_lock() below is not the equivalent of smp_wmb(), so
2628 * this is needed to avoid the copy_huge_page writes to become
2629 * visible after the set_pmd_at() write.
2634 BUG_ON(!pmd_none(*pmd));
2635 page_add_new_anon_rmap(new_page, vma, address);
2636 mem_cgroup_commit_charge(new_page, memcg, false);
2637 lru_cache_add_active_or_unevictable(new_page, vma);
2638 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2639 set_pmd_at(mm, address, pmd, _pmd);
2640 update_mmu_cache_pmd(vma, address, pmd);
2641 spin_unlock(pmd_ptl);
2645 khugepaged_pages_collapsed++;
2647 up_write(&mm->mmap_sem);
2651 mem_cgroup_cancel_charge(new_page, memcg);
2655 static int khugepaged_scan_pmd(struct mm_struct *mm,
2656 struct vm_area_struct *vma,
2657 unsigned long address,
2658 struct page **hpage)
2662 int ret = 0, none_or_zero = 0;
2664 unsigned long _address;
2666 int node = NUMA_NO_NODE;
2667 bool writable = false, referenced = false;
2669 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2671 pmd = mm_find_pmd(mm, address);
2675 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2676 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2677 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2678 _pte++, _address += PAGE_SIZE) {
2679 pte_t pteval = *_pte;
2680 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2681 if (!userfaultfd_armed(vma) &&
2682 ++none_or_zero <= khugepaged_max_ptes_none)
2687 if (!pte_present(pteval))
2689 if (pte_write(pteval))
2692 page = vm_normal_page(vma, _address, pteval);
2693 if (unlikely(!page))
2696 * Record which node the original page is from and save this
2697 * information to khugepaged_node_load[].
2698 * Khupaged will allocate hugepage from the node has the max
2701 node = page_to_nid(page);
2702 if (khugepaged_scan_abort(node))
2704 khugepaged_node_load[node]++;
2705 VM_BUG_ON_PAGE(PageCompound(page), page);
2706 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2709 * cannot use mapcount: can't collapse if there's a gup pin.
2710 * The page must only be referenced by the scanned process
2711 * and page swap cache.
2713 if (page_count(page) != 1 + !!PageSwapCache(page))
2715 if (pte_young(pteval) ||
2716 page_is_young(page) || PageReferenced(page) ||
2717 mmu_notifier_test_young(vma->vm_mm, address))
2720 if (referenced && writable)
2723 pte_unmap_unlock(pte, ptl);
2725 node = khugepaged_find_target_node();
2726 /* collapse_huge_page will return with the mmap_sem released */
2727 collapse_huge_page(mm, address, hpage, vma, node);
2733 static void collect_mm_slot(struct mm_slot *mm_slot)
2735 struct mm_struct *mm = mm_slot->mm;
2737 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2739 if (khugepaged_test_exit(mm)) {
2741 hash_del(&mm_slot->hash);
2742 list_del(&mm_slot->mm_node);
2745 * Not strictly needed because the mm exited already.
2747 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2750 /* khugepaged_mm_lock actually not necessary for the below */
2751 free_mm_slot(mm_slot);
2756 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2757 struct page **hpage)
2758 __releases(&khugepaged_mm_lock)
2759 __acquires(&khugepaged_mm_lock)
2761 struct mm_slot *mm_slot;
2762 struct mm_struct *mm;
2763 struct vm_area_struct *vma;
2767 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2769 if (khugepaged_scan.mm_slot)
2770 mm_slot = khugepaged_scan.mm_slot;
2772 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2773 struct mm_slot, mm_node);
2774 khugepaged_scan.address = 0;
2775 khugepaged_scan.mm_slot = mm_slot;
2777 spin_unlock(&khugepaged_mm_lock);
2780 down_read(&mm->mmap_sem);
2781 if (unlikely(khugepaged_test_exit(mm)))
2784 vma = find_vma(mm, khugepaged_scan.address);
2787 for (; vma; vma = vma->vm_next) {
2788 unsigned long hstart, hend;
2791 if (unlikely(khugepaged_test_exit(mm))) {
2795 if (!hugepage_vma_check(vma)) {
2800 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2801 hend = vma->vm_end & HPAGE_PMD_MASK;
2804 if (khugepaged_scan.address > hend)
2806 if (khugepaged_scan.address < hstart)
2807 khugepaged_scan.address = hstart;
2808 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2810 while (khugepaged_scan.address < hend) {
2813 if (unlikely(khugepaged_test_exit(mm)))
2814 goto breakouterloop;
2816 VM_BUG_ON(khugepaged_scan.address < hstart ||
2817 khugepaged_scan.address + HPAGE_PMD_SIZE >
2819 ret = khugepaged_scan_pmd(mm, vma,
2820 khugepaged_scan.address,
2822 /* move to next address */
2823 khugepaged_scan.address += HPAGE_PMD_SIZE;
2824 progress += HPAGE_PMD_NR;
2826 /* we released mmap_sem so break loop */
2827 goto breakouterloop_mmap_sem;
2828 if (progress >= pages)
2829 goto breakouterloop;
2833 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2834 breakouterloop_mmap_sem:
2836 spin_lock(&khugepaged_mm_lock);
2837 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2839 * Release the current mm_slot if this mm is about to die, or
2840 * if we scanned all vmas of this mm.
2842 if (khugepaged_test_exit(mm) || !vma) {
2844 * Make sure that if mm_users is reaching zero while
2845 * khugepaged runs here, khugepaged_exit will find
2846 * mm_slot not pointing to the exiting mm.
2848 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2849 khugepaged_scan.mm_slot = list_entry(
2850 mm_slot->mm_node.next,
2851 struct mm_slot, mm_node);
2852 khugepaged_scan.address = 0;
2854 khugepaged_scan.mm_slot = NULL;
2855 khugepaged_full_scans++;
2858 collect_mm_slot(mm_slot);
2864 static int khugepaged_has_work(void)
2866 return !list_empty(&khugepaged_scan.mm_head) &&
2867 khugepaged_enabled();
2870 static int khugepaged_wait_event(void)
2872 return !list_empty(&khugepaged_scan.mm_head) ||
2873 kthread_should_stop();
2876 static void khugepaged_do_scan(void)
2878 struct page *hpage = NULL;
2879 unsigned int progress = 0, pass_through_head = 0;
2880 unsigned int pages = khugepaged_pages_to_scan;
2883 barrier(); /* write khugepaged_pages_to_scan to local stack */
2885 while (progress < pages) {
2886 if (!khugepaged_prealloc_page(&hpage, &wait))
2891 if (unlikely(kthread_should_stop() || try_to_freeze()))
2894 spin_lock(&khugepaged_mm_lock);
2895 if (!khugepaged_scan.mm_slot)
2896 pass_through_head++;
2897 if (khugepaged_has_work() &&
2898 pass_through_head < 2)
2899 progress += khugepaged_scan_mm_slot(pages - progress,
2903 spin_unlock(&khugepaged_mm_lock);
2906 if (!IS_ERR_OR_NULL(hpage))
2910 static void khugepaged_wait_work(void)
2912 if (khugepaged_has_work()) {
2913 if (!khugepaged_scan_sleep_millisecs)
2916 wait_event_freezable_timeout(khugepaged_wait,
2917 kthread_should_stop(),
2918 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2922 if (khugepaged_enabled())
2923 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2926 static int khugepaged(void *none)
2928 struct mm_slot *mm_slot;
2931 set_user_nice(current, MAX_NICE);
2933 while (!kthread_should_stop()) {
2934 khugepaged_do_scan();
2935 khugepaged_wait_work();
2938 spin_lock(&khugepaged_mm_lock);
2939 mm_slot = khugepaged_scan.mm_slot;
2940 khugepaged_scan.mm_slot = NULL;
2942 collect_mm_slot(mm_slot);
2943 spin_unlock(&khugepaged_mm_lock);
2947 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2948 unsigned long haddr, pmd_t *pmd)
2950 struct mm_struct *mm = vma->vm_mm;
2955 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2956 /* leave pmd empty until pte is filled */
2958 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2959 pmd_populate(mm, &_pmd, pgtable);
2961 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2963 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2964 entry = pte_mkspecial(entry);
2965 pte = pte_offset_map(&_pmd, haddr);
2966 VM_BUG_ON(!pte_none(*pte));
2967 set_pte_at(mm, haddr, pte, entry);
2970 smp_wmb(); /* make pte visible before pmd */
2971 pmd_populate(mm, pmd, pgtable);
2972 put_huge_zero_page();
2975 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2979 struct page *page = NULL;
2980 struct mm_struct *mm = vma->vm_mm;
2981 unsigned long haddr = address & HPAGE_PMD_MASK;
2982 unsigned long mmun_start; /* For mmu_notifiers */
2983 unsigned long mmun_end; /* For mmu_notifiers */
2985 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2988 mmun_end = haddr + HPAGE_PMD_SIZE;
2990 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2991 ptl = pmd_lock(mm, pmd);
2992 if (unlikely(!pmd_trans_huge(*pmd)))
2994 if (vma_is_dax(vma)) {
2995 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2996 if (is_huge_zero_pmd(_pmd))
2997 put_huge_zero_page();
2998 } else if (is_huge_zero_pmd(*pmd)) {
2999 __split_huge_zero_page_pmd(vma, haddr, pmd);
3001 page = pmd_page(*pmd);
3002 VM_BUG_ON_PAGE(!page_count(page), page);
3007 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3012 split_huge_page(page);
3016 * We don't always have down_write of mmap_sem here: a racing
3017 * do_huge_pmd_wp_page() might have copied-on-write to another
3018 * huge page before our split_huge_page() got the anon_vma lock.
3020 if (unlikely(pmd_trans_huge(*pmd)))
3024 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
3027 struct vm_area_struct *vma;
3029 vma = find_vma(mm, address);
3030 BUG_ON(vma == NULL);
3031 split_huge_page_pmd(vma, address, pmd);
3034 static void split_huge_page_address(struct mm_struct *mm,
3035 unsigned long address)
3041 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
3043 pgd = pgd_offset(mm, address);
3044 if (!pgd_present(*pgd))
3047 pud = pud_offset(pgd, address);
3048 if (!pud_present(*pud))
3051 pmd = pmd_offset(pud, address);
3052 if (!pmd_present(*pmd))
3055 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3056 * materialize from under us.
3058 split_huge_page_pmd_mm(mm, address, pmd);
3061 void vma_adjust_trans_huge(struct vm_area_struct *vma,
3062 unsigned long start,
3067 * If the new start address isn't hpage aligned and it could
3068 * previously contain an hugepage: check if we need to split
3071 if (start & ~HPAGE_PMD_MASK &&
3072 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3073 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3074 split_huge_page_address(vma->vm_mm, start);
3077 * If the new end address isn't hpage aligned and it could
3078 * previously contain an hugepage: check if we need to split
3081 if (end & ~HPAGE_PMD_MASK &&
3082 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3083 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3084 split_huge_page_address(vma->vm_mm, end);
3087 * If we're also updating the vma->vm_next->vm_start, if the new
3088 * vm_next->vm_start isn't page aligned and it could previously
3089 * contain an hugepage: check if we need to split an huge pmd.
3091 if (adjust_next > 0) {
3092 struct vm_area_struct *next = vma->vm_next;
3093 unsigned long nstart = next->vm_start;
3094 nstart += adjust_next << PAGE_SHIFT;
3095 if (nstart & ~HPAGE_PMD_MASK &&
3096 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3097 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3098 split_huge_page_address(next->vm_mm, nstart);