1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
5 #include <linux/sched.h>
6 #include <linux/sched/mm.h>
7 #include <linux/sched/coredump.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/mm_inline.h>
12 #include <linux/kthread.h>
13 #include <linux/khugepaged.h>
14 #include <linux/freezer.h>
15 #include <linux/mman.h>
16 #include <linux/hashtable.h>
17 #include <linux/userfaultfd_k.h>
18 #include <linux/page_idle.h>
19 #include <linux/swapops.h>
20 #include <linux/shmem_fs.h>
23 #include <asm/pgalloc.h>
32 SCAN_EXCEED_SHARED_PTE
,
36 SCAN_LACK_REFERENCED_PAGE
,
50 SCAN_ALLOC_HUGE_PAGE_FAIL
,
51 SCAN_CGROUP_CHARGE_FAIL
,
53 SCAN_PAGE_HAS_PRIVATE
,
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/huge_memory.h>
59 static struct task_struct
*khugepaged_thread __read_mostly
;
60 static DEFINE_MUTEX(khugepaged_mutex
);
62 /* default scan 8*512 pte (or vmas) every 30 second */
63 static unsigned int khugepaged_pages_to_scan __read_mostly
;
64 static unsigned int khugepaged_pages_collapsed
;
65 static unsigned int khugepaged_full_scans
;
66 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
67 /* during fragmentation poll the hugepage allocator once every minute */
68 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
69 static unsigned long khugepaged_sleep_expire
;
70 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
71 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
73 * default collapse hugepages if there is at least one pte mapped like
74 * it would have happened if the vma was large enough during page
77 static unsigned int khugepaged_max_ptes_none __read_mostly
;
78 static unsigned int khugepaged_max_ptes_swap __read_mostly
;
79 static unsigned int khugepaged_max_ptes_shared __read_mostly
;
81 #define MM_SLOTS_HASH_BITS 10
82 static __read_mostly
DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
84 static struct kmem_cache
*mm_slot_cache __read_mostly
;
86 #define MAX_PTE_MAPPED_THP 8
89 * struct mm_slot - hash lookup from mm to mm_slot
90 * @hash: hash collision list
91 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
92 * @mm: the mm that this information is valid for
93 * @nr_pte_mapped_thp: number of pte mapped THP
94 * @pte_mapped_thp: address array corresponding pte mapped THP
97 struct hlist_node hash
;
98 struct list_head mm_node
;
101 /* pte-mapped THP in this mm */
102 int nr_pte_mapped_thp
;
103 unsigned long pte_mapped_thp
[MAX_PTE_MAPPED_THP
];
107 * struct khugepaged_scan - cursor for scanning
108 * @mm_head: the head of the mm list to scan
109 * @mm_slot: the current mm_slot we are scanning
110 * @address: the next address inside that to be scanned
112 * There is only the one khugepaged_scan instance of this cursor structure.
114 struct khugepaged_scan
{
115 struct list_head mm_head
;
116 struct mm_slot
*mm_slot
;
117 unsigned long address
;
120 static struct khugepaged_scan khugepaged_scan
= {
121 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
125 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
126 struct kobj_attribute
*attr
,
129 return sysfs_emit(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
132 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
133 struct kobj_attribute
*attr
,
134 const char *buf
, size_t count
)
139 err
= kstrtouint(buf
, 10, &msecs
);
143 khugepaged_scan_sleep_millisecs
= msecs
;
144 khugepaged_sleep_expire
= 0;
145 wake_up_interruptible(&khugepaged_wait
);
149 static struct kobj_attribute scan_sleep_millisecs_attr
=
150 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
151 scan_sleep_millisecs_store
);
153 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
154 struct kobj_attribute
*attr
,
157 return sysfs_emit(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
160 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
161 struct kobj_attribute
*attr
,
162 const char *buf
, size_t count
)
167 err
= kstrtouint(buf
, 10, &msecs
);
171 khugepaged_alloc_sleep_millisecs
= msecs
;
172 khugepaged_sleep_expire
= 0;
173 wake_up_interruptible(&khugepaged_wait
);
177 static struct kobj_attribute alloc_sleep_millisecs_attr
=
178 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
179 alloc_sleep_millisecs_store
);
181 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
182 struct kobj_attribute
*attr
,
185 return sysfs_emit(buf
, "%u\n", khugepaged_pages_to_scan
);
187 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
188 struct kobj_attribute
*attr
,
189 const char *buf
, size_t count
)
194 err
= kstrtouint(buf
, 10, &pages
);
198 khugepaged_pages_to_scan
= pages
;
202 static struct kobj_attribute pages_to_scan_attr
=
203 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
204 pages_to_scan_store
);
206 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
207 struct kobj_attribute
*attr
,
210 return sysfs_emit(buf
, "%u\n", khugepaged_pages_collapsed
);
212 static struct kobj_attribute pages_collapsed_attr
=
213 __ATTR_RO(pages_collapsed
);
215 static ssize_t
full_scans_show(struct kobject
*kobj
,
216 struct kobj_attribute
*attr
,
219 return sysfs_emit(buf
, "%u\n", khugepaged_full_scans
);
221 static struct kobj_attribute full_scans_attr
=
222 __ATTR_RO(full_scans
);
224 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
225 struct kobj_attribute
*attr
, char *buf
)
227 return single_hugepage_flag_show(kobj
, attr
, buf
,
228 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
230 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
231 struct kobj_attribute
*attr
,
232 const char *buf
, size_t count
)
234 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
235 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
237 static struct kobj_attribute khugepaged_defrag_attr
=
238 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
239 khugepaged_defrag_store
);
242 * max_ptes_none controls if khugepaged should collapse hugepages over
243 * any unmapped ptes in turn potentially increasing the memory
244 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
245 * reduce the available free memory in the system as it
246 * runs. Increasing max_ptes_none will instead potentially reduce the
247 * free memory in the system during the khugepaged scan.
249 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
250 struct kobj_attribute
*attr
,
253 return sysfs_emit(buf
, "%u\n", khugepaged_max_ptes_none
);
255 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
256 struct kobj_attribute
*attr
,
257 const char *buf
, size_t count
)
260 unsigned long max_ptes_none
;
262 err
= kstrtoul(buf
, 10, &max_ptes_none
);
263 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
266 khugepaged_max_ptes_none
= max_ptes_none
;
270 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
271 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
272 khugepaged_max_ptes_none_store
);
274 static ssize_t
khugepaged_max_ptes_swap_show(struct kobject
*kobj
,
275 struct kobj_attribute
*attr
,
278 return sysfs_emit(buf
, "%u\n", khugepaged_max_ptes_swap
);
281 static ssize_t
khugepaged_max_ptes_swap_store(struct kobject
*kobj
,
282 struct kobj_attribute
*attr
,
283 const char *buf
, size_t count
)
286 unsigned long max_ptes_swap
;
288 err
= kstrtoul(buf
, 10, &max_ptes_swap
);
289 if (err
|| max_ptes_swap
> HPAGE_PMD_NR
-1)
292 khugepaged_max_ptes_swap
= max_ptes_swap
;
297 static struct kobj_attribute khugepaged_max_ptes_swap_attr
=
298 __ATTR(max_ptes_swap
, 0644, khugepaged_max_ptes_swap_show
,
299 khugepaged_max_ptes_swap_store
);
301 static ssize_t
khugepaged_max_ptes_shared_show(struct kobject
*kobj
,
302 struct kobj_attribute
*attr
,
305 return sysfs_emit(buf
, "%u\n", khugepaged_max_ptes_shared
);
308 static ssize_t
khugepaged_max_ptes_shared_store(struct kobject
*kobj
,
309 struct kobj_attribute
*attr
,
310 const char *buf
, size_t count
)
313 unsigned long max_ptes_shared
;
315 err
= kstrtoul(buf
, 10, &max_ptes_shared
);
316 if (err
|| max_ptes_shared
> HPAGE_PMD_NR
-1)
319 khugepaged_max_ptes_shared
= max_ptes_shared
;
324 static struct kobj_attribute khugepaged_max_ptes_shared_attr
=
325 __ATTR(max_ptes_shared
, 0644, khugepaged_max_ptes_shared_show
,
326 khugepaged_max_ptes_shared_store
);
328 static struct attribute
*khugepaged_attr
[] = {
329 &khugepaged_defrag_attr
.attr
,
330 &khugepaged_max_ptes_none_attr
.attr
,
331 &khugepaged_max_ptes_swap_attr
.attr
,
332 &khugepaged_max_ptes_shared_attr
.attr
,
333 &pages_to_scan_attr
.attr
,
334 &pages_collapsed_attr
.attr
,
335 &full_scans_attr
.attr
,
336 &scan_sleep_millisecs_attr
.attr
,
337 &alloc_sleep_millisecs_attr
.attr
,
341 struct attribute_group khugepaged_attr_group
= {
342 .attrs
= khugepaged_attr
,
343 .name
= "khugepaged",
345 #endif /* CONFIG_SYSFS */
347 int hugepage_madvise(struct vm_area_struct
*vma
,
348 unsigned long *vm_flags
, int advice
)
354 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
355 * can't handle this properly after s390_enable_sie, so we simply
356 * ignore the madvise to prevent qemu from causing a SIGSEGV.
358 if (mm_has_pgste(vma
->vm_mm
))
361 *vm_flags
&= ~VM_NOHUGEPAGE
;
362 *vm_flags
|= VM_HUGEPAGE
;
364 * If the vma become good for khugepaged to scan,
365 * register it here without waiting a page fault that
366 * may not happen any time soon.
368 if (!(*vm_flags
& VM_NO_KHUGEPAGED
) &&
369 khugepaged_enter_vma_merge(vma
, *vm_flags
))
372 case MADV_NOHUGEPAGE
:
373 *vm_flags
&= ~VM_HUGEPAGE
;
374 *vm_flags
|= VM_NOHUGEPAGE
;
376 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
377 * this vma even if we leave the mm registered in khugepaged if
378 * it got registered before VM_NOHUGEPAGE was set.
386 int __init
khugepaged_init(void)
388 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
389 sizeof(struct mm_slot
),
390 __alignof__(struct mm_slot
), 0, NULL
);
394 khugepaged_pages_to_scan
= HPAGE_PMD_NR
* 8;
395 khugepaged_max_ptes_none
= HPAGE_PMD_NR
- 1;
396 khugepaged_max_ptes_swap
= HPAGE_PMD_NR
/ 8;
397 khugepaged_max_ptes_shared
= HPAGE_PMD_NR
/ 2;
402 void __init
khugepaged_destroy(void)
404 kmem_cache_destroy(mm_slot_cache
);
407 static inline struct mm_slot
*alloc_mm_slot(void)
409 if (!mm_slot_cache
) /* initialization failed */
411 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
414 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
416 kmem_cache_free(mm_slot_cache
, mm_slot
);
419 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
421 struct mm_slot
*mm_slot
;
423 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
424 if (mm
== mm_slot
->mm
)
430 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
431 struct mm_slot
*mm_slot
)
434 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
437 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
439 return atomic_read(&mm
->mm_users
) == 0;
442 static bool hugepage_vma_check(struct vm_area_struct
*vma
,
443 unsigned long vm_flags
)
445 /* Explicitly disabled through madvise. */
446 if ((vm_flags
& VM_NOHUGEPAGE
) ||
447 test_bit(MMF_DISABLE_THP
, &vma
->vm_mm
->flags
))
450 /* Enabled via shmem mount options or sysfs settings. */
451 if (shmem_file(vma
->vm_file
) && shmem_huge_enabled(vma
)) {
452 return IS_ALIGNED((vma
->vm_start
>> PAGE_SHIFT
) - vma
->vm_pgoff
,
456 /* THP settings require madvise. */
457 if (!(vm_flags
& VM_HUGEPAGE
) && !khugepaged_always())
460 /* Read-only file mappings need to be aligned for THP to work. */
461 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS
) && vma
->vm_file
&&
462 (vm_flags
& VM_DENYWRITE
)) {
463 return IS_ALIGNED((vma
->vm_start
>> PAGE_SHIFT
) - vma
->vm_pgoff
,
467 if (!vma
->anon_vma
|| vma
->vm_ops
)
469 if (vma_is_temporary_stack(vma
))
471 return !(vm_flags
& VM_NO_KHUGEPAGED
);
474 int __khugepaged_enter(struct mm_struct
*mm
)
476 struct mm_slot
*mm_slot
;
479 mm_slot
= alloc_mm_slot();
483 /* __khugepaged_exit() must not run from under us */
484 VM_BUG_ON_MM(atomic_read(&mm
->mm_users
) == 0, mm
);
485 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
486 free_mm_slot(mm_slot
);
490 spin_lock(&khugepaged_mm_lock
);
491 insert_to_mm_slots_hash(mm
, mm_slot
);
493 * Insert just behind the scanning cursor, to let the area settle
496 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
497 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
498 spin_unlock(&khugepaged_mm_lock
);
502 wake_up_interruptible(&khugepaged_wait
);
507 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
,
508 unsigned long vm_flags
)
510 unsigned long hstart
, hend
;
513 * khugepaged only supports read-only files for non-shmem files.
514 * khugepaged does not yet work on special mappings. And
515 * file-private shmem THP is not supported.
517 if (!hugepage_vma_check(vma
, vm_flags
))
520 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
521 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
523 return khugepaged_enter(vma
, vm_flags
);
527 void __khugepaged_exit(struct mm_struct
*mm
)
529 struct mm_slot
*mm_slot
;
532 spin_lock(&khugepaged_mm_lock
);
533 mm_slot
= get_mm_slot(mm
);
534 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
535 hash_del(&mm_slot
->hash
);
536 list_del(&mm_slot
->mm_node
);
539 spin_unlock(&khugepaged_mm_lock
);
542 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
543 free_mm_slot(mm_slot
);
545 } else if (mm_slot
) {
547 * This is required to serialize against
548 * khugepaged_test_exit() (which is guaranteed to run
549 * under mmap sem read mode). Stop here (after we
550 * return all pagetables will be destroyed) until
551 * khugepaged has finished working on the pagetables
552 * under the mmap_lock.
555 mmap_write_unlock(mm
);
559 static void release_pte_page(struct page
*page
)
561 mod_node_page_state(page_pgdat(page
),
562 NR_ISOLATED_ANON
+ page_is_file_lru(page
),
565 putback_lru_page(page
);
568 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
,
569 struct list_head
*compound_pagelist
)
571 struct page
*page
, *tmp
;
573 while (--_pte
>= pte
) {
574 pte_t pteval
= *_pte
;
576 page
= pte_page(pteval
);
577 if (!pte_none(pteval
) && !is_zero_pfn(pte_pfn(pteval
)) &&
579 release_pte_page(page
);
582 list_for_each_entry_safe(page
, tmp
, compound_pagelist
, lru
) {
583 list_del(&page
->lru
);
584 release_pte_page(page
);
588 static bool is_refcount_suitable(struct page
*page
)
590 int expected_refcount
;
592 expected_refcount
= total_mapcount(page
);
593 if (PageSwapCache(page
))
594 expected_refcount
+= compound_nr(page
);
596 return page_count(page
) == expected_refcount
;
599 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
600 unsigned long address
,
602 struct list_head
*compound_pagelist
)
604 struct page
*page
= NULL
;
606 int none_or_zero
= 0, shared
= 0, result
= 0, referenced
= 0;
607 bool writable
= false;
609 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
610 _pte
++, address
+= PAGE_SIZE
) {
611 pte_t pteval
= *_pte
;
612 if (pte_none(pteval
) || (pte_present(pteval
) &&
613 is_zero_pfn(pte_pfn(pteval
)))) {
614 if (!userfaultfd_armed(vma
) &&
615 ++none_or_zero
<= khugepaged_max_ptes_none
) {
618 result
= SCAN_EXCEED_NONE_PTE
;
622 if (!pte_present(pteval
)) {
623 result
= SCAN_PTE_NON_PRESENT
;
626 page
= vm_normal_page(vma
, address
, pteval
);
627 if (unlikely(!page
)) {
628 result
= SCAN_PAGE_NULL
;
632 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
634 if (page_mapcount(page
) > 1 &&
635 ++shared
> khugepaged_max_ptes_shared
) {
636 result
= SCAN_EXCEED_SHARED_PTE
;
640 if (PageCompound(page
)) {
642 page
= compound_head(page
);
645 * Check if we have dealt with the compound page
648 list_for_each_entry(p
, compound_pagelist
, lru
) {
655 * We can do it before isolate_lru_page because the
656 * page can't be freed from under us. NOTE: PG_lock
657 * is needed to serialize against split_huge_page
658 * when invoked from the VM.
660 if (!trylock_page(page
)) {
661 result
= SCAN_PAGE_LOCK
;
666 * Check if the page has any GUP (or other external) pins.
668 * The page table that maps the page has been already unlinked
669 * from the page table tree and this process cannot get
670 * an additinal pin on the page.
672 * New pins can come later if the page is shared across fork,
673 * but not from this process. The other process cannot write to
674 * the page, only trigger CoW.
676 if (!is_refcount_suitable(page
)) {
678 result
= SCAN_PAGE_COUNT
;
681 if (!pte_write(pteval
) && PageSwapCache(page
) &&
682 !reuse_swap_page(page
, NULL
)) {
684 * Page is in the swap cache and cannot be re-used.
685 * It cannot be collapsed into a THP.
688 result
= SCAN_SWAP_CACHE_PAGE
;
693 * Isolate the page to avoid collapsing an hugepage
694 * currently in use by the VM.
696 if (isolate_lru_page(page
)) {
698 result
= SCAN_DEL_PAGE_LRU
;
701 mod_node_page_state(page_pgdat(page
),
702 NR_ISOLATED_ANON
+ page_is_file_lru(page
),
704 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
705 VM_BUG_ON_PAGE(PageLRU(page
), page
);
707 if (PageCompound(page
))
708 list_add_tail(&page
->lru
, compound_pagelist
);
710 /* There should be enough young pte to collapse the page */
711 if (pte_young(pteval
) ||
712 page_is_young(page
) || PageReferenced(page
) ||
713 mmu_notifier_test_young(vma
->vm_mm
, address
))
716 if (pte_write(pteval
))
719 if (likely(writable
)) {
720 if (likely(referenced
)) {
721 result
= SCAN_SUCCEED
;
722 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
723 referenced
, writable
, result
);
727 result
= SCAN_PAGE_RO
;
731 release_pte_pages(pte
, _pte
, compound_pagelist
);
732 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
733 referenced
, writable
, result
);
737 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
738 struct vm_area_struct
*vma
,
739 unsigned long address
,
741 struct list_head
*compound_pagelist
)
743 struct page
*src_page
, *tmp
;
745 for (_pte
= pte
; _pte
< pte
+ HPAGE_PMD_NR
;
746 _pte
++, page
++, address
+= PAGE_SIZE
) {
747 pte_t pteval
= *_pte
;
749 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
750 clear_user_highpage(page
, address
);
751 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
752 if (is_zero_pfn(pte_pfn(pteval
))) {
754 * ptl mostly unnecessary.
758 * paravirt calls inside pte_clear here are
761 pte_clear(vma
->vm_mm
, address
, _pte
);
765 src_page
= pte_page(pteval
);
766 copy_user_highpage(page
, src_page
, address
, vma
);
767 if (!PageCompound(src_page
))
768 release_pte_page(src_page
);
770 * ptl mostly unnecessary, but preempt has to
771 * be disabled to update the per-cpu stats
772 * inside page_remove_rmap().
776 * paravirt calls inside pte_clear here are
779 pte_clear(vma
->vm_mm
, address
, _pte
);
780 page_remove_rmap(src_page
, false);
782 free_page_and_swap_cache(src_page
);
786 list_for_each_entry_safe(src_page
, tmp
, compound_pagelist
, lru
) {
787 list_del(&src_page
->lru
);
788 release_pte_page(src_page
);
792 static void khugepaged_alloc_sleep(void)
796 add_wait_queue(&khugepaged_wait
, &wait
);
797 freezable_schedule_timeout_interruptible(
798 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
799 remove_wait_queue(&khugepaged_wait
, &wait
);
802 static int khugepaged_node_load
[MAX_NUMNODES
];
804 static bool khugepaged_scan_abort(int nid
)
809 * If node_reclaim_mode is disabled, then no extra effort is made to
810 * allocate memory locally.
812 if (!node_reclaim_mode
)
815 /* If there is a count for this node already, it must be acceptable */
816 if (khugepaged_node_load
[nid
])
819 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
820 if (!khugepaged_node_load
[i
])
822 if (node_distance(nid
, i
) > node_reclaim_distance
)
828 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
829 static inline gfp_t
alloc_hugepage_khugepaged_gfpmask(void)
831 return khugepaged_defrag() ? GFP_TRANSHUGE
: GFP_TRANSHUGE_LIGHT
;
835 static int khugepaged_find_target_node(void)
837 static int last_khugepaged_target_node
= NUMA_NO_NODE
;
838 int nid
, target_node
= 0, max_value
= 0;
840 /* find first node with max normal pages hit */
841 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
842 if (khugepaged_node_load
[nid
] > max_value
) {
843 max_value
= khugepaged_node_load
[nid
];
847 /* do some balance if several nodes have the same hit record */
848 if (target_node
<= last_khugepaged_target_node
)
849 for (nid
= last_khugepaged_target_node
+ 1; nid
< MAX_NUMNODES
;
851 if (max_value
== khugepaged_node_load
[nid
]) {
856 last_khugepaged_target_node
= target_node
;
860 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
862 if (IS_ERR(*hpage
)) {
868 khugepaged_alloc_sleep();
878 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, int node
)
880 VM_BUG_ON_PAGE(*hpage
, *hpage
);
882 *hpage
= __alloc_pages_node(node
, gfp
, HPAGE_PMD_ORDER
);
883 if (unlikely(!*hpage
)) {
884 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
885 *hpage
= ERR_PTR(-ENOMEM
);
889 prep_transhuge_page(*hpage
);
890 count_vm_event(THP_COLLAPSE_ALLOC
);
894 static int khugepaged_find_target_node(void)
899 static inline struct page
*alloc_khugepaged_hugepage(void)
903 page
= alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
906 prep_transhuge_page(page
);
910 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
915 hpage
= alloc_khugepaged_hugepage();
917 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
922 khugepaged_alloc_sleep();
924 count_vm_event(THP_COLLAPSE_ALLOC
);
925 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
930 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
933 * If the hpage allocated earlier was briefly exposed in page cache
934 * before collapse_file() failed, it is possible that racing lookups
935 * have not yet completed, and would then be unpleasantly surprised by
936 * finding the hpage reused for the same mapping at a different offset.
937 * Just release the previous allocation if there is any danger of that.
939 if (*hpage
&& page_count(*hpage
) > 1) {
945 *hpage
= khugepaged_alloc_hugepage(wait
);
947 if (unlikely(!*hpage
))
954 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, int node
)
963 * If mmap_lock temporarily dropped, revalidate vma
964 * before taking mmap_lock.
965 * Return 0 if succeeds, otherwise return none-zero
969 static int hugepage_vma_revalidate(struct mm_struct
*mm
, unsigned long address
,
970 struct vm_area_struct
**vmap
)
972 struct vm_area_struct
*vma
;
973 unsigned long hstart
, hend
;
975 if (unlikely(khugepaged_test_exit(mm
)))
976 return SCAN_ANY_PROCESS
;
978 *vmap
= vma
= find_vma(mm
, address
);
980 return SCAN_VMA_NULL
;
982 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
983 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
984 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
985 return SCAN_ADDRESS_RANGE
;
986 if (!hugepage_vma_check(vma
, vma
->vm_flags
))
987 return SCAN_VMA_CHECK
;
988 /* Anon VMA expected */
989 if (!vma
->anon_vma
|| vma
->vm_ops
)
990 return SCAN_VMA_CHECK
;
995 * Bring missing pages in from swap, to complete THP collapse.
996 * Only done if khugepaged_scan_pmd believes it is worthwhile.
998 * Called and returns without pte mapped or spinlocks held,
999 * but with mmap_lock held to protect against vma changes.
1002 static bool __collapse_huge_page_swapin(struct mm_struct
*mm
,
1003 struct vm_area_struct
*vma
,
1004 unsigned long address
, pmd_t
*pmd
,
1009 struct vm_fault vmf
= {
1012 .flags
= FAULT_FLAG_ALLOW_RETRY
,
1014 .pgoff
= linear_page_index(vma
, address
),
1017 vmf
.pte
= pte_offset_map(pmd
, address
);
1018 for (; vmf
.address
< address
+ HPAGE_PMD_NR
*PAGE_SIZE
;
1019 vmf
.pte
++, vmf
.address
+= PAGE_SIZE
) {
1020 vmf
.orig_pte
= *vmf
.pte
;
1021 if (!is_swap_pte(vmf
.orig_pte
))
1024 ret
= do_swap_page(&vmf
);
1026 /* do_swap_page returns VM_FAULT_RETRY with released mmap_lock */
1027 if (ret
& VM_FAULT_RETRY
) {
1029 if (hugepage_vma_revalidate(mm
, address
, &vmf
.vma
)) {
1030 /* vma is no longer available, don't continue to swapin */
1031 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 0);
1034 /* check if the pmd is still valid */
1035 if (mm_find_pmd(mm
, address
) != pmd
) {
1036 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 0);
1040 if (ret
& VM_FAULT_ERROR
) {
1041 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 0);
1044 /* pte is unmapped now, we need to map it */
1045 vmf
.pte
= pte_offset_map(pmd
, vmf
.address
);
1050 /* Drain LRU add pagevec to remove extra pin on the swapped in pages */
1054 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 1);
1058 static void collapse_huge_page(struct mm_struct
*mm
,
1059 unsigned long address
,
1060 struct page
**hpage
,
1061 int node
, int referenced
, int unmapped
)
1063 LIST_HEAD(compound_pagelist
);
1067 struct page
*new_page
;
1068 spinlock_t
*pmd_ptl
, *pte_ptl
;
1069 int isolated
= 0, result
= 0;
1070 struct vm_area_struct
*vma
;
1071 struct mmu_notifier_range range
;
1074 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
1076 /* Only allocate from the target node */
1077 gfp
= alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE
;
1080 * Before allocating the hugepage, release the mmap_lock read lock.
1081 * The allocation can take potentially a long time if it involves
1082 * sync compaction, and we do not need to hold the mmap_lock during
1083 * that. We will recheck the vma after taking it again in write mode.
1085 mmap_read_unlock(mm
);
1086 new_page
= khugepaged_alloc_page(hpage
, gfp
, node
);
1088 result
= SCAN_ALLOC_HUGE_PAGE_FAIL
;
1092 if (unlikely(mem_cgroup_charge(new_page
, mm
, gfp
))) {
1093 result
= SCAN_CGROUP_CHARGE_FAIL
;
1096 count_memcg_page_event(new_page
, THP_COLLAPSE_ALLOC
);
1099 result
= hugepage_vma_revalidate(mm
, address
, &vma
);
1101 mmap_read_unlock(mm
);
1105 pmd
= mm_find_pmd(mm
, address
);
1107 result
= SCAN_PMD_NULL
;
1108 mmap_read_unlock(mm
);
1113 * __collapse_huge_page_swapin always returns with mmap_lock locked.
1114 * If it fails, we release mmap_lock and jump out_nolock.
1115 * Continuing to collapse causes inconsistency.
1117 if (unmapped
&& !__collapse_huge_page_swapin(mm
, vma
, address
,
1119 mmap_read_unlock(mm
);
1123 mmap_read_unlock(mm
);
1125 * Prevent all access to pagetables with the exception of
1126 * gup_fast later handled by the ptep_clear_flush and the VM
1127 * handled by the anon_vma lock + PG_lock.
1129 mmap_write_lock(mm
);
1130 result
= hugepage_vma_revalidate(mm
, address
, &vma
);
1133 /* check if the pmd is still valid */
1134 if (mm_find_pmd(mm
, address
) != pmd
)
1137 anon_vma_lock_write(vma
->anon_vma
);
1139 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm
,
1140 address
, address
+ HPAGE_PMD_SIZE
);
1141 mmu_notifier_invalidate_range_start(&range
);
1143 pte
= pte_offset_map(pmd
, address
);
1144 pte_ptl
= pte_lockptr(mm
, pmd
);
1146 pmd_ptl
= pmd_lock(mm
, pmd
); /* probably unnecessary */
1148 * After this gup_fast can't run anymore. This also removes
1149 * any huge TLB entry from the CPU so we won't allow
1150 * huge and small TLB entries for the same virtual address
1151 * to avoid the risk of CPU bugs in that area.
1153 _pmd
= pmdp_collapse_flush(vma
, address
, pmd
);
1154 spin_unlock(pmd_ptl
);
1155 mmu_notifier_invalidate_range_end(&range
);
1158 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
,
1159 &compound_pagelist
);
1160 spin_unlock(pte_ptl
);
1162 if (unlikely(!isolated
)) {
1165 BUG_ON(!pmd_none(*pmd
));
1167 * We can only use set_pmd_at when establishing
1168 * hugepmds and never for establishing regular pmds that
1169 * points to regular pagetables. Use pmd_populate for that
1171 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
1172 spin_unlock(pmd_ptl
);
1173 anon_vma_unlock_write(vma
->anon_vma
);
1179 * All pages are isolated and locked so anon_vma rmap
1180 * can't run anymore.
1182 anon_vma_unlock_write(vma
->anon_vma
);
1184 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, pte_ptl
,
1185 &compound_pagelist
);
1187 __SetPageUptodate(new_page
);
1188 pgtable
= pmd_pgtable(_pmd
);
1190 _pmd
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1191 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
1194 * spin_lock() below is not the equivalent of smp_wmb(), so
1195 * this is needed to avoid the copy_huge_page writes to become
1196 * visible after the set_pmd_at() write.
1201 BUG_ON(!pmd_none(*pmd
));
1202 page_add_new_anon_rmap(new_page
, vma
, address
, true);
1203 lru_cache_add_inactive_or_unevictable(new_page
, vma
);
1204 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
1205 set_pmd_at(mm
, address
, pmd
, _pmd
);
1206 update_mmu_cache_pmd(vma
, address
, pmd
);
1207 spin_unlock(pmd_ptl
);
1211 khugepaged_pages_collapsed
++;
1212 result
= SCAN_SUCCEED
;
1214 mmap_write_unlock(mm
);
1216 if (!IS_ERR_OR_NULL(*hpage
))
1217 mem_cgroup_uncharge(*hpage
);
1218 trace_mm_collapse_huge_page(mm
, isolated
, result
);
1224 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
1225 struct vm_area_struct
*vma
,
1226 unsigned long address
,
1227 struct page
**hpage
)
1231 int ret
= 0, result
= 0, referenced
= 0;
1232 int none_or_zero
= 0, shared
= 0;
1233 struct page
*page
= NULL
;
1234 unsigned long _address
;
1236 int node
= NUMA_NO_NODE
, unmapped
= 0;
1237 bool writable
= false;
1239 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
1241 pmd
= mm_find_pmd(mm
, address
);
1243 result
= SCAN_PMD_NULL
;
1247 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
1248 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1249 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
1250 _pte
++, _address
+= PAGE_SIZE
) {
1251 pte_t pteval
= *_pte
;
1252 if (is_swap_pte(pteval
)) {
1253 if (++unmapped
<= khugepaged_max_ptes_swap
) {
1255 * Always be strict with uffd-wp
1256 * enabled swap entries. Please see
1257 * comment below for pte_uffd_wp().
1259 if (pte_swp_uffd_wp(pteval
)) {
1260 result
= SCAN_PTE_UFFD_WP
;
1265 result
= SCAN_EXCEED_SWAP_PTE
;
1269 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
1270 if (!userfaultfd_armed(vma
) &&
1271 ++none_or_zero
<= khugepaged_max_ptes_none
) {
1274 result
= SCAN_EXCEED_NONE_PTE
;
1278 if (!pte_present(pteval
)) {
1279 result
= SCAN_PTE_NON_PRESENT
;
1282 if (pte_uffd_wp(pteval
)) {
1284 * Don't collapse the page if any of the small
1285 * PTEs are armed with uffd write protection.
1286 * Here we can also mark the new huge pmd as
1287 * write protected if any of the small ones is
1288 * marked but that could bring unknown
1289 * userfault messages that falls outside of
1290 * the registered range. So, just be simple.
1292 result
= SCAN_PTE_UFFD_WP
;
1295 if (pte_write(pteval
))
1298 page
= vm_normal_page(vma
, _address
, pteval
);
1299 if (unlikely(!page
)) {
1300 result
= SCAN_PAGE_NULL
;
1304 if (page_mapcount(page
) > 1 &&
1305 ++shared
> khugepaged_max_ptes_shared
) {
1306 result
= SCAN_EXCEED_SHARED_PTE
;
1310 page
= compound_head(page
);
1313 * Record which node the original page is from and save this
1314 * information to khugepaged_node_load[].
1315 * Khupaged will allocate hugepage from the node has the max
1318 node
= page_to_nid(page
);
1319 if (khugepaged_scan_abort(node
)) {
1320 result
= SCAN_SCAN_ABORT
;
1323 khugepaged_node_load
[node
]++;
1324 if (!PageLRU(page
)) {
1325 result
= SCAN_PAGE_LRU
;
1328 if (PageLocked(page
)) {
1329 result
= SCAN_PAGE_LOCK
;
1332 if (!PageAnon(page
)) {
1333 result
= SCAN_PAGE_ANON
;
1338 * Check if the page has any GUP (or other external) pins.
1340 * Here the check is racy it may see totmal_mapcount > refcount
1342 * For example, one process with one forked child process.
1343 * The parent has the PMD split due to MADV_DONTNEED, then
1344 * the child is trying unmap the whole PMD, but khugepaged
1345 * may be scanning the parent between the child has
1346 * PageDoubleMap flag cleared and dec the mapcount. So
1347 * khugepaged may see total_mapcount > refcount.
1349 * But such case is ephemeral we could always retry collapse
1350 * later. However it may report false positive if the page
1351 * has excessive GUP pins (i.e. 512). Anyway the same check
1352 * will be done again later the risk seems low.
1354 if (!is_refcount_suitable(page
)) {
1355 result
= SCAN_PAGE_COUNT
;
1358 if (pte_young(pteval
) ||
1359 page_is_young(page
) || PageReferenced(page
) ||
1360 mmu_notifier_test_young(vma
->vm_mm
, address
))
1364 result
= SCAN_PAGE_RO
;
1365 } else if (!referenced
|| (unmapped
&& referenced
< HPAGE_PMD_NR
/2)) {
1366 result
= SCAN_LACK_REFERENCED_PAGE
;
1368 result
= SCAN_SUCCEED
;
1372 pte_unmap_unlock(pte
, ptl
);
1374 node
= khugepaged_find_target_node();
1375 /* collapse_huge_page will return with the mmap_lock released */
1376 collapse_huge_page(mm
, address
, hpage
, node
,
1377 referenced
, unmapped
);
1380 trace_mm_khugepaged_scan_pmd(mm
, page
, writable
, referenced
,
1381 none_or_zero
, result
, unmapped
);
1385 static void collect_mm_slot(struct mm_slot
*mm_slot
)
1387 struct mm_struct
*mm
= mm_slot
->mm
;
1389 lockdep_assert_held(&khugepaged_mm_lock
);
1391 if (khugepaged_test_exit(mm
)) {
1393 hash_del(&mm_slot
->hash
);
1394 list_del(&mm_slot
->mm_node
);
1397 * Not strictly needed because the mm exited already.
1399 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1402 /* khugepaged_mm_lock actually not necessary for the below */
1403 free_mm_slot(mm_slot
);
1410 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1411 * khugepaged should try to collapse the page table.
1413 static int khugepaged_add_pte_mapped_thp(struct mm_struct
*mm
,
1416 struct mm_slot
*mm_slot
;
1418 VM_BUG_ON(addr
& ~HPAGE_PMD_MASK
);
1420 spin_lock(&khugepaged_mm_lock
);
1421 mm_slot
= get_mm_slot(mm
);
1422 if (likely(mm_slot
&& mm_slot
->nr_pte_mapped_thp
< MAX_PTE_MAPPED_THP
))
1423 mm_slot
->pte_mapped_thp
[mm_slot
->nr_pte_mapped_thp
++] = addr
;
1424 spin_unlock(&khugepaged_mm_lock
);
1429 * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at
1432 * @mm: process address space where collapse happens
1433 * @addr: THP collapse address
1435 * This function checks whether all the PTEs in the PMD are pointing to the
1436 * right THP. If so, retract the page table so the THP can refault in with
1439 void collapse_pte_mapped_thp(struct mm_struct
*mm
, unsigned long addr
)
1441 unsigned long haddr
= addr
& HPAGE_PMD_MASK
;
1442 struct vm_area_struct
*vma
= find_vma(mm
, haddr
);
1444 pte_t
*start_pte
, *pte
;
1450 if (!vma
|| !vma
->vm_file
||
1451 vma
->vm_start
> haddr
|| vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
)
1455 * This vm_flags may not have VM_HUGEPAGE if the page was not
1456 * collapsed by this mm. But we can still collapse if the page is
1457 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1458 * will not fail the vma for missing VM_HUGEPAGE
1460 if (!hugepage_vma_check(vma
, vma
->vm_flags
| VM_HUGEPAGE
))
1463 hpage
= find_lock_page(vma
->vm_file
->f_mapping
,
1464 linear_page_index(vma
, haddr
));
1468 if (!PageHead(hpage
))
1471 pmd
= mm_find_pmd(mm
, haddr
);
1475 start_pte
= pte_offset_map_lock(mm
, pmd
, haddr
, &ptl
);
1477 /* step 1: check all mapped PTEs are to the right huge page */
1478 for (i
= 0, addr
= haddr
, pte
= start_pte
;
1479 i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
, pte
++) {
1482 /* empty pte, skip */
1486 /* page swapped out, abort */
1487 if (!pte_present(*pte
))
1490 page
= vm_normal_page(vma
, addr
, *pte
);
1493 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1494 * page table, but the new page will not be a subpage of hpage.
1496 if (hpage
+ i
!= page
)
1501 /* step 2: adjust rmap */
1502 for (i
= 0, addr
= haddr
, pte
= start_pte
;
1503 i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
, pte
++) {
1508 page
= vm_normal_page(vma
, addr
, *pte
);
1509 page_remove_rmap(page
, false);
1512 pte_unmap_unlock(start_pte
, ptl
);
1514 /* step 3: set proper refcount and mm_counters. */
1516 page_ref_sub(hpage
, count
);
1517 add_mm_counter(vma
->vm_mm
, mm_counter_file(hpage
), -count
);
1520 /* step 4: collapse pmd */
1521 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1522 _pmd
= pmdp_collapse_flush(vma
, haddr
, pmd
);
1525 pte_free(mm
, pmd_pgtable(_pmd
));
1533 pte_unmap_unlock(start_pte
, ptl
);
1537 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot
*mm_slot
)
1539 struct mm_struct
*mm
= mm_slot
->mm
;
1542 if (likely(mm_slot
->nr_pte_mapped_thp
== 0))
1545 if (!mmap_write_trylock(mm
))
1548 if (unlikely(khugepaged_test_exit(mm
)))
1551 for (i
= 0; i
< mm_slot
->nr_pte_mapped_thp
; i
++)
1552 collapse_pte_mapped_thp(mm
, mm_slot
->pte_mapped_thp
[i
]);
1555 mm_slot
->nr_pte_mapped_thp
= 0;
1556 mmap_write_unlock(mm
);
1560 static void retract_page_tables(struct address_space
*mapping
, pgoff_t pgoff
)
1562 struct vm_area_struct
*vma
;
1563 struct mm_struct
*mm
;
1567 i_mmap_lock_write(mapping
);
1568 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1570 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1571 * got written to. These VMAs are likely not worth investing
1572 * mmap_write_lock(mm) as PMD-mapping is likely to be split
1575 * Not that vma->anon_vma check is racy: it can be set up after
1576 * the check but before we took mmap_lock by the fault path.
1577 * But page lock would prevent establishing any new ptes of the
1578 * page, so we are safe.
1580 * An alternative would be drop the check, but check that page
1581 * table is clear before calling pmdp_collapse_flush() under
1582 * ptl. It has higher chance to recover THP for the VMA, but
1583 * has higher cost too.
1587 addr
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
1588 if (addr
& ~HPAGE_PMD_MASK
)
1590 if (vma
->vm_end
< addr
+ HPAGE_PMD_SIZE
)
1593 pmd
= mm_find_pmd(mm
, addr
);
1597 * We need exclusive mmap_lock to retract page table.
1599 * We use trylock due to lock inversion: we need to acquire
1600 * mmap_lock while holding page lock. Fault path does it in
1601 * reverse order. Trylock is a way to avoid deadlock.
1603 if (mmap_write_trylock(mm
)) {
1604 if (!khugepaged_test_exit(mm
)) {
1605 spinlock_t
*ptl
= pmd_lock(mm
, pmd
);
1606 /* assume page table is clear */
1607 _pmd
= pmdp_collapse_flush(vma
, addr
, pmd
);
1610 pte_free(mm
, pmd_pgtable(_pmd
));
1612 mmap_write_unlock(mm
);
1614 /* Try again later */
1615 khugepaged_add_pte_mapped_thp(mm
, addr
);
1618 i_mmap_unlock_write(mapping
);
1622 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1624 * @mm: process address space where collapse happens
1625 * @file: file that collapse on
1626 * @start: collapse start address
1627 * @hpage: new allocated huge page for collapse
1628 * @node: appointed node the new huge page allocate from
1630 * Basic scheme is simple, details are more complex:
1631 * - allocate and lock a new huge page;
1632 * - scan page cache replacing old pages with the new one
1633 * + swap/gup in pages if necessary;
1635 * + keep old pages around in case rollback is required;
1636 * - if replacing succeeds:
1639 * + unlock huge page;
1640 * - if replacing failed;
1641 * + put all pages back and unfreeze them;
1642 * + restore gaps in the page cache;
1643 * + unlock and free huge page;
1645 static void collapse_file(struct mm_struct
*mm
,
1646 struct file
*file
, pgoff_t start
,
1647 struct page
**hpage
, int node
)
1649 struct address_space
*mapping
= file
->f_mapping
;
1651 struct page
*new_page
;
1652 pgoff_t index
, end
= start
+ HPAGE_PMD_NR
;
1653 LIST_HEAD(pagelist
);
1654 XA_STATE_ORDER(xas
, &mapping
->i_pages
, start
, HPAGE_PMD_ORDER
);
1655 int nr_none
= 0, result
= SCAN_SUCCEED
;
1656 bool is_shmem
= shmem_file(file
);
1658 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS
) && !is_shmem
);
1659 VM_BUG_ON(start
& (HPAGE_PMD_NR
- 1));
1661 /* Only allocate from the target node */
1662 gfp
= alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE
;
1664 new_page
= khugepaged_alloc_page(hpage
, gfp
, node
);
1666 result
= SCAN_ALLOC_HUGE_PAGE_FAIL
;
1670 if (unlikely(mem_cgroup_charge(new_page
, mm
, gfp
))) {
1671 result
= SCAN_CGROUP_CHARGE_FAIL
;
1674 count_memcg_page_event(new_page
, THP_COLLAPSE_ALLOC
);
1676 /* This will be less messy when we use multi-index entries */
1679 xas_create_range(&xas
);
1680 if (!xas_error(&xas
))
1682 xas_unlock_irq(&xas
);
1683 if (!xas_nomem(&xas
, GFP_KERNEL
)) {
1689 __SetPageLocked(new_page
);
1691 __SetPageSwapBacked(new_page
);
1692 new_page
->index
= start
;
1693 new_page
->mapping
= mapping
;
1696 * At this point the new_page is locked and not up-to-date.
1697 * It's safe to insert it into the page cache, because nobody would
1698 * be able to map it or use it in another way until we unlock it.
1701 xas_set(&xas
, start
);
1702 for (index
= start
; index
< end
; index
++) {
1703 struct page
*page
= xas_next(&xas
);
1705 VM_BUG_ON(index
!= xas
.xa_index
);
1709 * Stop if extent has been truncated or
1710 * hole-punched, and is now completely
1713 if (index
== start
) {
1714 if (!xas_next_entry(&xas
, end
- 1)) {
1715 result
= SCAN_TRUNCATED
;
1718 xas_set(&xas
, index
);
1720 if (!shmem_charge(mapping
->host
, 1)) {
1724 xas_store(&xas
, new_page
);
1729 if (xa_is_value(page
) || !PageUptodate(page
)) {
1730 xas_unlock_irq(&xas
);
1731 /* swap in or instantiate fallocated page */
1732 if (shmem_getpage(mapping
->host
, index
, &page
,
1737 } else if (trylock_page(page
)) {
1739 xas_unlock_irq(&xas
);
1741 result
= SCAN_PAGE_LOCK
;
1744 } else { /* !is_shmem */
1745 if (!page
|| xa_is_value(page
)) {
1746 xas_unlock_irq(&xas
);
1747 page_cache_sync_readahead(mapping
, &file
->f_ra
,
1750 /* drain pagevecs to help isolate_lru_page() */
1752 page
= find_lock_page(mapping
, index
);
1753 if (unlikely(page
== NULL
)) {
1757 } else if (PageDirty(page
)) {
1759 * khugepaged only works on read-only fd,
1760 * so this page is dirty because it hasn't
1761 * been flushed since first write. There
1762 * won't be new dirty pages.
1764 * Trigger async flush here and hope the
1765 * writeback is done when khugepaged
1766 * revisits this page.
1768 * This is a one-off situation. We are not
1769 * forcing writeback in loop.
1771 xas_unlock_irq(&xas
);
1772 filemap_flush(mapping
);
1775 } else if (trylock_page(page
)) {
1777 xas_unlock_irq(&xas
);
1779 result
= SCAN_PAGE_LOCK
;
1785 * The page must be locked, so we can drop the i_pages lock
1786 * without racing with truncate.
1788 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1790 /* make sure the page is up to date */
1791 if (unlikely(!PageUptodate(page
))) {
1797 * If file was truncated then extended, or hole-punched, before
1798 * we locked the first page, then a THP might be there already.
1800 if (PageTransCompound(page
)) {
1801 result
= SCAN_PAGE_COMPOUND
;
1805 if (page_mapping(page
) != mapping
) {
1806 result
= SCAN_TRUNCATED
;
1810 if (!is_shmem
&& PageDirty(page
)) {
1812 * khugepaged only works on read-only fd, so this
1813 * page is dirty because it hasn't been flushed
1814 * since first write.
1820 if (isolate_lru_page(page
)) {
1821 result
= SCAN_DEL_PAGE_LRU
;
1825 if (page_has_private(page
) &&
1826 !try_to_release_page(page
, GFP_KERNEL
)) {
1827 result
= SCAN_PAGE_HAS_PRIVATE
;
1828 putback_lru_page(page
);
1832 if (page_mapped(page
))
1833 unmap_mapping_pages(mapping
, index
, 1, false);
1836 xas_set(&xas
, index
);
1838 VM_BUG_ON_PAGE(page
!= xas_load(&xas
), page
);
1839 VM_BUG_ON_PAGE(page_mapped(page
), page
);
1842 * The page is expected to have page_count() == 3:
1843 * - we hold a pin on it;
1844 * - one reference from page cache;
1845 * - one from isolate_lru_page;
1847 if (!page_ref_freeze(page
, 3)) {
1848 result
= SCAN_PAGE_COUNT
;
1849 xas_unlock_irq(&xas
);
1850 putback_lru_page(page
);
1855 * Add the page to the list to be able to undo the collapse if
1856 * something go wrong.
1858 list_add_tail(&page
->lru
, &pagelist
);
1860 /* Finally, replace with the new page. */
1861 xas_store(&xas
, new_page
);
1870 __inc_lruvec_page_state(new_page
, NR_SHMEM_THPS
);
1872 __inc_lruvec_page_state(new_page
, NR_FILE_THPS
);
1873 filemap_nr_thps_inc(mapping
);
1877 __mod_lruvec_page_state(new_page
, NR_FILE_PAGES
, nr_none
);
1879 __mod_lruvec_page_state(new_page
, NR_SHMEM
, nr_none
);
1883 xas_unlock_irq(&xas
);
1886 if (result
== SCAN_SUCCEED
) {
1887 struct page
*page
, *tmp
;
1890 * Replacing old pages with new one has succeeded, now we
1891 * need to copy the content and free the old pages.
1894 list_for_each_entry_safe(page
, tmp
, &pagelist
, lru
) {
1895 while (index
< page
->index
) {
1896 clear_highpage(new_page
+ (index
% HPAGE_PMD_NR
));
1899 copy_highpage(new_page
+ (page
->index
% HPAGE_PMD_NR
),
1901 list_del(&page
->lru
);
1902 page
->mapping
= NULL
;
1903 page_ref_unfreeze(page
, 1);
1904 ClearPageActive(page
);
1905 ClearPageUnevictable(page
);
1910 while (index
< end
) {
1911 clear_highpage(new_page
+ (index
% HPAGE_PMD_NR
));
1915 SetPageUptodate(new_page
);
1916 page_ref_add(new_page
, HPAGE_PMD_NR
- 1);
1918 set_page_dirty(new_page
);
1919 lru_cache_add(new_page
);
1922 * Remove pte page tables, so we can re-fault the page as huge.
1924 retract_page_tables(mapping
, start
);
1927 khugepaged_pages_collapsed
++;
1931 /* Something went wrong: roll back page cache changes */
1933 mapping
->nrpages
-= nr_none
;
1936 shmem_uncharge(mapping
->host
, nr_none
);
1938 xas_set(&xas
, start
);
1939 xas_for_each(&xas
, page
, end
- 1) {
1940 page
= list_first_entry_or_null(&pagelist
,
1942 if (!page
|| xas
.xa_index
< page
->index
) {
1946 /* Put holes back where they were */
1947 xas_store(&xas
, NULL
);
1951 VM_BUG_ON_PAGE(page
->index
!= xas
.xa_index
, page
);
1953 /* Unfreeze the page. */
1954 list_del(&page
->lru
);
1955 page_ref_unfreeze(page
, 2);
1956 xas_store(&xas
, page
);
1958 xas_unlock_irq(&xas
);
1960 putback_lru_page(page
);
1964 xas_unlock_irq(&xas
);
1966 new_page
->mapping
= NULL
;
1969 unlock_page(new_page
);
1971 VM_BUG_ON(!list_empty(&pagelist
));
1972 if (!IS_ERR_OR_NULL(*hpage
))
1973 mem_cgroup_uncharge(*hpage
);
1974 /* TODO: tracepoints */
1977 static void khugepaged_scan_file(struct mm_struct
*mm
,
1978 struct file
*file
, pgoff_t start
, struct page
**hpage
)
1980 struct page
*page
= NULL
;
1981 struct address_space
*mapping
= file
->f_mapping
;
1982 XA_STATE(xas
, &mapping
->i_pages
, start
);
1984 int node
= NUMA_NO_NODE
;
1985 int result
= SCAN_SUCCEED
;
1989 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
1991 xas_for_each(&xas
, page
, start
+ HPAGE_PMD_NR
- 1) {
1992 if (xas_retry(&xas
, page
))
1995 if (xa_is_value(page
)) {
1996 if (++swap
> khugepaged_max_ptes_swap
) {
1997 result
= SCAN_EXCEED_SWAP_PTE
;
2003 if (PageTransCompound(page
)) {
2004 result
= SCAN_PAGE_COMPOUND
;
2008 node
= page_to_nid(page
);
2009 if (khugepaged_scan_abort(node
)) {
2010 result
= SCAN_SCAN_ABORT
;
2013 khugepaged_node_load
[node
]++;
2015 if (!PageLRU(page
)) {
2016 result
= SCAN_PAGE_LRU
;
2020 if (page_count(page
) !=
2021 1 + page_mapcount(page
) + page_has_private(page
)) {
2022 result
= SCAN_PAGE_COUNT
;
2027 * We probably should check if the page is referenced here, but
2028 * nobody would transfer pte_young() to PageReferenced() for us.
2029 * And rmap walk here is just too costly...
2034 if (need_resched()) {
2041 if (result
== SCAN_SUCCEED
) {
2042 if (present
< HPAGE_PMD_NR
- khugepaged_max_ptes_none
) {
2043 result
= SCAN_EXCEED_NONE_PTE
;
2045 node
= khugepaged_find_target_node();
2046 collapse_file(mm
, file
, start
, hpage
, node
);
2050 /* TODO: tracepoints */
2053 static void khugepaged_scan_file(struct mm_struct
*mm
,
2054 struct file
*file
, pgoff_t start
, struct page
**hpage
)
2059 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot
*mm_slot
)
2065 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
2066 struct page
**hpage
)
2067 __releases(&khugepaged_mm_lock
)
2068 __acquires(&khugepaged_mm_lock
)
2070 struct mm_slot
*mm_slot
;
2071 struct mm_struct
*mm
;
2072 struct vm_area_struct
*vma
;
2076 lockdep_assert_held(&khugepaged_mm_lock
);
2078 if (khugepaged_scan
.mm_slot
)
2079 mm_slot
= khugepaged_scan
.mm_slot
;
2081 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
2082 struct mm_slot
, mm_node
);
2083 khugepaged_scan
.address
= 0;
2084 khugepaged_scan
.mm_slot
= mm_slot
;
2086 spin_unlock(&khugepaged_mm_lock
);
2087 khugepaged_collapse_pte_mapped_thps(mm_slot
);
2091 * Don't wait for semaphore (to avoid long wait times). Just move to
2092 * the next mm on the list.
2095 if (unlikely(!mmap_read_trylock(mm
)))
2096 goto breakouterloop_mmap_lock
;
2097 if (likely(!khugepaged_test_exit(mm
)))
2098 vma
= find_vma(mm
, khugepaged_scan
.address
);
2101 for (; vma
; vma
= vma
->vm_next
) {
2102 unsigned long hstart
, hend
;
2105 if (unlikely(khugepaged_test_exit(mm
))) {
2109 if (!hugepage_vma_check(vma
, vma
->vm_flags
)) {
2114 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
2115 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
2118 if (khugepaged_scan
.address
> hend
)
2120 if (khugepaged_scan
.address
< hstart
)
2121 khugepaged_scan
.address
= hstart
;
2122 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
2123 if (shmem_file(vma
->vm_file
) && !shmem_huge_enabled(vma
))
2126 while (khugepaged_scan
.address
< hend
) {
2129 if (unlikely(khugepaged_test_exit(mm
)))
2130 goto breakouterloop
;
2132 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
2133 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
2135 if (IS_ENABLED(CONFIG_SHMEM
) && vma
->vm_file
) {
2136 struct file
*file
= get_file(vma
->vm_file
);
2137 pgoff_t pgoff
= linear_page_index(vma
,
2138 khugepaged_scan
.address
);
2140 mmap_read_unlock(mm
);
2142 khugepaged_scan_file(mm
, file
, pgoff
, hpage
);
2145 ret
= khugepaged_scan_pmd(mm
, vma
,
2146 khugepaged_scan
.address
,
2149 /* move to next address */
2150 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
2151 progress
+= HPAGE_PMD_NR
;
2153 /* we released mmap_lock so break loop */
2154 goto breakouterloop_mmap_lock
;
2155 if (progress
>= pages
)
2156 goto breakouterloop
;
2160 mmap_read_unlock(mm
); /* exit_mmap will destroy ptes after this */
2161 breakouterloop_mmap_lock
:
2163 spin_lock(&khugepaged_mm_lock
);
2164 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
2166 * Release the current mm_slot if this mm is about to die, or
2167 * if we scanned all vmas of this mm.
2169 if (khugepaged_test_exit(mm
) || !vma
) {
2171 * Make sure that if mm_users is reaching zero while
2172 * khugepaged runs here, khugepaged_exit will find
2173 * mm_slot not pointing to the exiting mm.
2175 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
2176 khugepaged_scan
.mm_slot
= list_entry(
2177 mm_slot
->mm_node
.next
,
2178 struct mm_slot
, mm_node
);
2179 khugepaged_scan
.address
= 0;
2181 khugepaged_scan
.mm_slot
= NULL
;
2182 khugepaged_full_scans
++;
2185 collect_mm_slot(mm_slot
);
2191 static int khugepaged_has_work(void)
2193 return !list_empty(&khugepaged_scan
.mm_head
) &&
2194 khugepaged_enabled();
2197 static int khugepaged_wait_event(void)
2199 return !list_empty(&khugepaged_scan
.mm_head
) ||
2200 kthread_should_stop();
2203 static void khugepaged_do_scan(void)
2205 struct page
*hpage
= NULL
;
2206 unsigned int progress
= 0, pass_through_head
= 0;
2207 unsigned int pages
= khugepaged_pages_to_scan
;
2210 barrier(); /* write khugepaged_pages_to_scan to local stack */
2212 lru_add_drain_all();
2214 while (progress
< pages
) {
2215 if (!khugepaged_prealloc_page(&hpage
, &wait
))
2220 if (unlikely(kthread_should_stop() || try_to_freeze()))
2223 spin_lock(&khugepaged_mm_lock
);
2224 if (!khugepaged_scan
.mm_slot
)
2225 pass_through_head
++;
2226 if (khugepaged_has_work() &&
2227 pass_through_head
< 2)
2228 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
2232 spin_unlock(&khugepaged_mm_lock
);
2235 if (!IS_ERR_OR_NULL(hpage
))
2239 static bool khugepaged_should_wakeup(void)
2241 return kthread_should_stop() ||
2242 time_after_eq(jiffies
, khugepaged_sleep_expire
);
2245 static void khugepaged_wait_work(void)
2247 if (khugepaged_has_work()) {
2248 const unsigned long scan_sleep_jiffies
=
2249 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
);
2251 if (!scan_sleep_jiffies
)
2254 khugepaged_sleep_expire
= jiffies
+ scan_sleep_jiffies
;
2255 wait_event_freezable_timeout(khugepaged_wait
,
2256 khugepaged_should_wakeup(),
2257 scan_sleep_jiffies
);
2261 if (khugepaged_enabled())
2262 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
2265 static int khugepaged(void *none
)
2267 struct mm_slot
*mm_slot
;
2270 set_user_nice(current
, MAX_NICE
);
2272 while (!kthread_should_stop()) {
2273 khugepaged_do_scan();
2274 khugepaged_wait_work();
2277 spin_lock(&khugepaged_mm_lock
);
2278 mm_slot
= khugepaged_scan
.mm_slot
;
2279 khugepaged_scan
.mm_slot
= NULL
;
2281 collect_mm_slot(mm_slot
);
2282 spin_unlock(&khugepaged_mm_lock
);
2286 static void set_recommended_min_free_kbytes(void)
2290 unsigned long recommended_min
;
2292 for_each_populated_zone(zone
) {
2294 * We don't need to worry about fragmentation of
2295 * ZONE_MOVABLE since it only has movable pages.
2297 if (zone_idx(zone
) > gfp_zone(GFP_USER
))
2303 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2304 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
2307 * Make sure that on average at least two pageblocks are almost free
2308 * of another type, one for a migratetype to fall back to and a
2309 * second to avoid subsequent fallbacks of other types There are 3
2310 * MIGRATE_TYPES we care about.
2312 recommended_min
+= pageblock_nr_pages
* nr_zones
*
2313 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
2315 /* don't ever allow to reserve more than 5% of the lowmem */
2316 recommended_min
= min(recommended_min
,
2317 (unsigned long) nr_free_buffer_pages() / 20);
2318 recommended_min
<<= (PAGE_SHIFT
-10);
2320 if (recommended_min
> min_free_kbytes
) {
2321 if (user_min_free_kbytes
>= 0)
2322 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2323 min_free_kbytes
, recommended_min
);
2325 min_free_kbytes
= recommended_min
;
2327 setup_per_zone_wmarks();
2330 int start_stop_khugepaged(void)
2334 mutex_lock(&khugepaged_mutex
);
2335 if (khugepaged_enabled()) {
2336 if (!khugepaged_thread
)
2337 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
2339 if (IS_ERR(khugepaged_thread
)) {
2340 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2341 err
= PTR_ERR(khugepaged_thread
);
2342 khugepaged_thread
= NULL
;
2346 if (!list_empty(&khugepaged_scan
.mm_head
))
2347 wake_up_interruptible(&khugepaged_wait
);
2349 set_recommended_min_free_kbytes();
2350 } else if (khugepaged_thread
) {
2351 kthread_stop(khugepaged_thread
);
2352 khugepaged_thread
= NULL
;
2355 mutex_unlock(&khugepaged_mutex
);
2359 void khugepaged_min_free_kbytes_update(void)
2361 mutex_lock(&khugepaged_mutex
);
2362 if (khugepaged_enabled() && khugepaged_thread
)
2363 set_recommended_min_free_kbytes();
2364 mutex_unlock(&khugepaged_mutex
);