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>
33 SCAN_LACK_REFERENCED_PAGE
,
47 SCAN_ALLOC_HUGE_PAGE_FAIL
,
48 SCAN_CGROUP_CHARGE_FAIL
,
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/huge_memory.h>
56 /* default scan 8*512 pte (or vmas) every 30 second */
57 static unsigned int khugepaged_pages_to_scan __read_mostly
;
58 static unsigned int khugepaged_pages_collapsed
;
59 static unsigned int khugepaged_full_scans
;
60 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly
= 10000;
61 /* during fragmentation poll the hugepage allocator once every minute */
62 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly
= 60000;
63 static unsigned long khugepaged_sleep_expire
;
64 static DEFINE_SPINLOCK(khugepaged_mm_lock
);
65 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait
);
67 * default collapse hugepages if there is at least one pte mapped like
68 * it would have happened if the vma was large enough during page
71 static unsigned int khugepaged_max_ptes_none __read_mostly
;
72 static unsigned int khugepaged_max_ptes_swap __read_mostly
;
74 #define MM_SLOTS_HASH_BITS 10
75 static __read_mostly
DEFINE_HASHTABLE(mm_slots_hash
, MM_SLOTS_HASH_BITS
);
77 static struct kmem_cache
*mm_slot_cache __read_mostly
;
80 * struct mm_slot - hash lookup from mm to mm_slot
81 * @hash: hash collision list
82 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
83 * @mm: the mm that this information is valid for
86 struct hlist_node hash
;
87 struct list_head mm_node
;
92 * struct khugepaged_scan - cursor for scanning
93 * @mm_head: the head of the mm list to scan
94 * @mm_slot: the current mm_slot we are scanning
95 * @address: the next address inside that to be scanned
97 * There is only the one khugepaged_scan instance of this cursor structure.
99 struct khugepaged_scan
{
100 struct list_head mm_head
;
101 struct mm_slot
*mm_slot
;
102 unsigned long address
;
105 static struct khugepaged_scan khugepaged_scan
= {
106 .mm_head
= LIST_HEAD_INIT(khugepaged_scan
.mm_head
),
110 static ssize_t
scan_sleep_millisecs_show(struct kobject
*kobj
,
111 struct kobj_attribute
*attr
,
114 return sprintf(buf
, "%u\n", khugepaged_scan_sleep_millisecs
);
117 static ssize_t
scan_sleep_millisecs_store(struct kobject
*kobj
,
118 struct kobj_attribute
*attr
,
119 const char *buf
, size_t count
)
124 err
= kstrtoul(buf
, 10, &msecs
);
125 if (err
|| msecs
> UINT_MAX
)
128 khugepaged_scan_sleep_millisecs
= msecs
;
129 khugepaged_sleep_expire
= 0;
130 wake_up_interruptible(&khugepaged_wait
);
134 static struct kobj_attribute scan_sleep_millisecs_attr
=
135 __ATTR(scan_sleep_millisecs
, 0644, scan_sleep_millisecs_show
,
136 scan_sleep_millisecs_store
);
138 static ssize_t
alloc_sleep_millisecs_show(struct kobject
*kobj
,
139 struct kobj_attribute
*attr
,
142 return sprintf(buf
, "%u\n", khugepaged_alloc_sleep_millisecs
);
145 static ssize_t
alloc_sleep_millisecs_store(struct kobject
*kobj
,
146 struct kobj_attribute
*attr
,
147 const char *buf
, size_t count
)
152 err
= kstrtoul(buf
, 10, &msecs
);
153 if (err
|| msecs
> UINT_MAX
)
156 khugepaged_alloc_sleep_millisecs
= msecs
;
157 khugepaged_sleep_expire
= 0;
158 wake_up_interruptible(&khugepaged_wait
);
162 static struct kobj_attribute alloc_sleep_millisecs_attr
=
163 __ATTR(alloc_sleep_millisecs
, 0644, alloc_sleep_millisecs_show
,
164 alloc_sleep_millisecs_store
);
166 static ssize_t
pages_to_scan_show(struct kobject
*kobj
,
167 struct kobj_attribute
*attr
,
170 return sprintf(buf
, "%u\n", khugepaged_pages_to_scan
);
172 static ssize_t
pages_to_scan_store(struct kobject
*kobj
,
173 struct kobj_attribute
*attr
,
174 const char *buf
, size_t count
)
179 err
= kstrtoul(buf
, 10, &pages
);
180 if (err
|| !pages
|| pages
> UINT_MAX
)
183 khugepaged_pages_to_scan
= pages
;
187 static struct kobj_attribute pages_to_scan_attr
=
188 __ATTR(pages_to_scan
, 0644, pages_to_scan_show
,
189 pages_to_scan_store
);
191 static ssize_t
pages_collapsed_show(struct kobject
*kobj
,
192 struct kobj_attribute
*attr
,
195 return sprintf(buf
, "%u\n", khugepaged_pages_collapsed
);
197 static struct kobj_attribute pages_collapsed_attr
=
198 __ATTR_RO(pages_collapsed
);
200 static ssize_t
full_scans_show(struct kobject
*kobj
,
201 struct kobj_attribute
*attr
,
204 return sprintf(buf
, "%u\n", khugepaged_full_scans
);
206 static struct kobj_attribute full_scans_attr
=
207 __ATTR_RO(full_scans
);
209 static ssize_t
khugepaged_defrag_show(struct kobject
*kobj
,
210 struct kobj_attribute
*attr
, char *buf
)
212 return single_hugepage_flag_show(kobj
, attr
, buf
,
213 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
215 static ssize_t
khugepaged_defrag_store(struct kobject
*kobj
,
216 struct kobj_attribute
*attr
,
217 const char *buf
, size_t count
)
219 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
220 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
);
222 static struct kobj_attribute khugepaged_defrag_attr
=
223 __ATTR(defrag
, 0644, khugepaged_defrag_show
,
224 khugepaged_defrag_store
);
227 * max_ptes_none controls if khugepaged should collapse hugepages over
228 * any unmapped ptes in turn potentially increasing the memory
229 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
230 * reduce the available free memory in the system as it
231 * runs. Increasing max_ptes_none will instead potentially reduce the
232 * free memory in the system during the khugepaged scan.
234 static ssize_t
khugepaged_max_ptes_none_show(struct kobject
*kobj
,
235 struct kobj_attribute
*attr
,
238 return sprintf(buf
, "%u\n", khugepaged_max_ptes_none
);
240 static ssize_t
khugepaged_max_ptes_none_store(struct kobject
*kobj
,
241 struct kobj_attribute
*attr
,
242 const char *buf
, size_t count
)
245 unsigned long max_ptes_none
;
247 err
= kstrtoul(buf
, 10, &max_ptes_none
);
248 if (err
|| max_ptes_none
> HPAGE_PMD_NR
-1)
251 khugepaged_max_ptes_none
= max_ptes_none
;
255 static struct kobj_attribute khugepaged_max_ptes_none_attr
=
256 __ATTR(max_ptes_none
, 0644, khugepaged_max_ptes_none_show
,
257 khugepaged_max_ptes_none_store
);
259 static ssize_t
khugepaged_max_ptes_swap_show(struct kobject
*kobj
,
260 struct kobj_attribute
*attr
,
263 return sprintf(buf
, "%u\n", khugepaged_max_ptes_swap
);
266 static ssize_t
khugepaged_max_ptes_swap_store(struct kobject
*kobj
,
267 struct kobj_attribute
*attr
,
268 const char *buf
, size_t count
)
271 unsigned long max_ptes_swap
;
273 err
= kstrtoul(buf
, 10, &max_ptes_swap
);
274 if (err
|| max_ptes_swap
> HPAGE_PMD_NR
-1)
277 khugepaged_max_ptes_swap
= max_ptes_swap
;
282 static struct kobj_attribute khugepaged_max_ptes_swap_attr
=
283 __ATTR(max_ptes_swap
, 0644, khugepaged_max_ptes_swap_show
,
284 khugepaged_max_ptes_swap_store
);
286 static struct attribute
*khugepaged_attr
[] = {
287 &khugepaged_defrag_attr
.attr
,
288 &khugepaged_max_ptes_none_attr
.attr
,
289 &pages_to_scan_attr
.attr
,
290 &pages_collapsed_attr
.attr
,
291 &full_scans_attr
.attr
,
292 &scan_sleep_millisecs_attr
.attr
,
293 &alloc_sleep_millisecs_attr
.attr
,
294 &khugepaged_max_ptes_swap_attr
.attr
,
298 struct attribute_group khugepaged_attr_group
= {
299 .attrs
= khugepaged_attr
,
300 .name
= "khugepaged",
302 #endif /* CONFIG_SYSFS */
304 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
306 int hugepage_madvise(struct vm_area_struct
*vma
,
307 unsigned long *vm_flags
, int advice
)
313 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
314 * can't handle this properly after s390_enable_sie, so we simply
315 * ignore the madvise to prevent qemu from causing a SIGSEGV.
317 if (mm_has_pgste(vma
->vm_mm
))
320 *vm_flags
&= ~VM_NOHUGEPAGE
;
321 *vm_flags
|= VM_HUGEPAGE
;
323 * If the vma become good for khugepaged to scan,
324 * register it here without waiting a page fault that
325 * may not happen any time soon.
327 if (!(*vm_flags
& VM_NO_KHUGEPAGED
) &&
328 khugepaged_enter_vma_merge(vma
, *vm_flags
))
331 case MADV_NOHUGEPAGE
:
332 *vm_flags
&= ~VM_HUGEPAGE
;
333 *vm_flags
|= VM_NOHUGEPAGE
;
335 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
336 * this vma even if we leave the mm registered in khugepaged if
337 * it got registered before VM_NOHUGEPAGE was set.
345 int __init
khugepaged_init(void)
347 mm_slot_cache
= kmem_cache_create("khugepaged_mm_slot",
348 sizeof(struct mm_slot
),
349 __alignof__(struct mm_slot
), 0, NULL
);
353 khugepaged_pages_to_scan
= HPAGE_PMD_NR
* 8;
354 khugepaged_max_ptes_none
= HPAGE_PMD_NR
- 1;
355 khugepaged_max_ptes_swap
= HPAGE_PMD_NR
/ 8;
360 void __init
khugepaged_destroy(void)
362 kmem_cache_destroy(mm_slot_cache
);
365 static inline struct mm_slot
*alloc_mm_slot(void)
367 if (!mm_slot_cache
) /* initialization failed */
369 return kmem_cache_zalloc(mm_slot_cache
, GFP_KERNEL
);
372 static inline void free_mm_slot(struct mm_slot
*mm_slot
)
374 kmem_cache_free(mm_slot_cache
, mm_slot
);
377 static struct mm_slot
*get_mm_slot(struct mm_struct
*mm
)
379 struct mm_slot
*mm_slot
;
381 hash_for_each_possible(mm_slots_hash
, mm_slot
, hash
, (unsigned long)mm
)
382 if (mm
== mm_slot
->mm
)
388 static void insert_to_mm_slots_hash(struct mm_struct
*mm
,
389 struct mm_slot
*mm_slot
)
392 hash_add(mm_slots_hash
, &mm_slot
->hash
, (long)mm
);
395 static inline int khugepaged_test_exit(struct mm_struct
*mm
)
397 return atomic_read(&mm
->mm_users
) == 0;
400 int __khugepaged_enter(struct mm_struct
*mm
)
402 struct mm_slot
*mm_slot
;
405 mm_slot
= alloc_mm_slot();
409 /* __khugepaged_exit() must not run from under us */
410 VM_BUG_ON_MM(khugepaged_test_exit(mm
), mm
);
411 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE
, &mm
->flags
))) {
412 free_mm_slot(mm_slot
);
416 spin_lock(&khugepaged_mm_lock
);
417 insert_to_mm_slots_hash(mm
, mm_slot
);
419 * Insert just behind the scanning cursor, to let the area settle
422 wakeup
= list_empty(&khugepaged_scan
.mm_head
);
423 list_add_tail(&mm_slot
->mm_node
, &khugepaged_scan
.mm_head
);
424 spin_unlock(&khugepaged_mm_lock
);
428 wake_up_interruptible(&khugepaged_wait
);
433 int khugepaged_enter_vma_merge(struct vm_area_struct
*vma
,
434 unsigned long vm_flags
)
436 unsigned long hstart
, hend
;
439 * Not yet faulted in so we will register later in the
440 * page fault if needed.
443 if (vma
->vm_ops
|| (vm_flags
& VM_NO_KHUGEPAGED
))
444 /* khugepaged not yet working on file or special mappings */
446 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
447 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
449 return khugepaged_enter(vma
, vm_flags
);
453 void __khugepaged_exit(struct mm_struct
*mm
)
455 struct mm_slot
*mm_slot
;
458 spin_lock(&khugepaged_mm_lock
);
459 mm_slot
= get_mm_slot(mm
);
460 if (mm_slot
&& khugepaged_scan
.mm_slot
!= mm_slot
) {
461 hash_del(&mm_slot
->hash
);
462 list_del(&mm_slot
->mm_node
);
465 spin_unlock(&khugepaged_mm_lock
);
468 clear_bit(MMF_VM_HUGEPAGE
, &mm
->flags
);
469 free_mm_slot(mm_slot
);
471 } else if (mm_slot
) {
473 * This is required to serialize against
474 * khugepaged_test_exit() (which is guaranteed to run
475 * under mmap sem read mode). Stop here (after we
476 * return all pagetables will be destroyed) until
477 * khugepaged has finished working on the pagetables
478 * under the mmap_sem.
480 down_write(&mm
->mmap_sem
);
481 up_write(&mm
->mmap_sem
);
485 static void release_pte_page(struct page
*page
)
487 dec_node_page_state(page
, NR_ISOLATED_ANON
+ page_is_file_cache(page
));
489 putback_lru_page(page
);
492 static void release_pte_pages(pte_t
*pte
, pte_t
*_pte
)
494 while (--_pte
>= pte
) {
495 pte_t pteval
= *_pte
;
496 if (!pte_none(pteval
) && !is_zero_pfn(pte_pfn(pteval
)))
497 release_pte_page(pte_page(pteval
));
501 static int __collapse_huge_page_isolate(struct vm_area_struct
*vma
,
502 unsigned long address
,
505 struct page
*page
= NULL
;
507 int none_or_zero
= 0, result
= 0, referenced
= 0;
508 bool writable
= false;
510 for (_pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
511 _pte
++, address
+= PAGE_SIZE
) {
512 pte_t pteval
= *_pte
;
513 if (pte_none(pteval
) || (pte_present(pteval
) &&
514 is_zero_pfn(pte_pfn(pteval
)))) {
515 if (!userfaultfd_armed(vma
) &&
516 ++none_or_zero
<= khugepaged_max_ptes_none
) {
519 result
= SCAN_EXCEED_NONE_PTE
;
523 if (!pte_present(pteval
)) {
524 result
= SCAN_PTE_NON_PRESENT
;
527 page
= vm_normal_page(vma
, address
, pteval
);
528 if (unlikely(!page
)) {
529 result
= SCAN_PAGE_NULL
;
533 /* TODO: teach khugepaged to collapse THP mapped with pte */
534 if (PageCompound(page
)) {
535 result
= SCAN_PAGE_COMPOUND
;
539 VM_BUG_ON_PAGE(!PageAnon(page
), page
);
542 * We can do it before isolate_lru_page because the
543 * page can't be freed from under us. NOTE: PG_lock
544 * is needed to serialize against split_huge_page
545 * when invoked from the VM.
547 if (!trylock_page(page
)) {
548 result
= SCAN_PAGE_LOCK
;
553 * cannot use mapcount: can't collapse if there's a gup pin.
554 * The page must only be referenced by the scanned process
555 * and page swap cache.
557 if (page_count(page
) != 1 + PageSwapCache(page
)) {
559 result
= SCAN_PAGE_COUNT
;
562 if (pte_write(pteval
)) {
565 if (PageSwapCache(page
) &&
566 !reuse_swap_page(page
, NULL
)) {
568 result
= SCAN_SWAP_CACHE_PAGE
;
572 * Page is not in the swap cache. It can be collapsed
578 * Isolate the page to avoid collapsing an hugepage
579 * currently in use by the VM.
581 if (isolate_lru_page(page
)) {
583 result
= SCAN_DEL_PAGE_LRU
;
586 inc_node_page_state(page
,
587 NR_ISOLATED_ANON
+ page_is_file_cache(page
));
588 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
589 VM_BUG_ON_PAGE(PageLRU(page
), page
);
591 /* There should be enough young pte to collapse the page */
592 if (pte_young(pteval
) ||
593 page_is_young(page
) || PageReferenced(page
) ||
594 mmu_notifier_test_young(vma
->vm_mm
, address
))
597 if (likely(writable
)) {
598 if (likely(referenced
)) {
599 result
= SCAN_SUCCEED
;
600 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
601 referenced
, writable
, result
);
605 result
= SCAN_PAGE_RO
;
609 release_pte_pages(pte
, _pte
);
610 trace_mm_collapse_huge_page_isolate(page
, none_or_zero
,
611 referenced
, writable
, result
);
615 static void __collapse_huge_page_copy(pte_t
*pte
, struct page
*page
,
616 struct vm_area_struct
*vma
,
617 unsigned long address
,
621 for (_pte
= pte
; _pte
< pte
+ HPAGE_PMD_NR
;
622 _pte
++, page
++, address
+= PAGE_SIZE
) {
623 pte_t pteval
= *_pte
;
624 struct page
*src_page
;
626 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
627 clear_user_highpage(page
, address
);
628 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, 1);
629 if (is_zero_pfn(pte_pfn(pteval
))) {
631 * ptl mostly unnecessary.
635 * paravirt calls inside pte_clear here are
638 pte_clear(vma
->vm_mm
, address
, _pte
);
642 src_page
= pte_page(pteval
);
643 copy_user_highpage(page
, src_page
, address
, vma
);
644 VM_BUG_ON_PAGE(page_mapcount(src_page
) != 1, src_page
);
645 release_pte_page(src_page
);
647 * ptl mostly unnecessary, but preempt has to
648 * be disabled to update the per-cpu stats
649 * inside page_remove_rmap().
653 * paravirt calls inside pte_clear here are
656 pte_clear(vma
->vm_mm
, address
, _pte
);
657 page_remove_rmap(src_page
, false);
659 free_page_and_swap_cache(src_page
);
664 static void khugepaged_alloc_sleep(void)
668 add_wait_queue(&khugepaged_wait
, &wait
);
669 freezable_schedule_timeout_interruptible(
670 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs
));
671 remove_wait_queue(&khugepaged_wait
, &wait
);
674 static int khugepaged_node_load
[MAX_NUMNODES
];
676 static bool khugepaged_scan_abort(int nid
)
681 * If node_reclaim_mode is disabled, then no extra effort is made to
682 * allocate memory locally.
684 if (!node_reclaim_mode
)
687 /* If there is a count for this node already, it must be acceptable */
688 if (khugepaged_node_load
[nid
])
691 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
692 if (!khugepaged_node_load
[i
])
694 if (node_distance(nid
, i
) > RECLAIM_DISTANCE
)
700 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
701 static inline gfp_t
alloc_hugepage_khugepaged_gfpmask(void)
703 return khugepaged_defrag() ? GFP_TRANSHUGE
: GFP_TRANSHUGE_LIGHT
;
707 static int khugepaged_find_target_node(void)
709 static int last_khugepaged_target_node
= NUMA_NO_NODE
;
710 int nid
, target_node
= 0, max_value
= 0;
712 /* find first node with max normal pages hit */
713 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
714 if (khugepaged_node_load
[nid
] > max_value
) {
715 max_value
= khugepaged_node_load
[nid
];
719 /* do some balance if several nodes have the same hit record */
720 if (target_node
<= last_khugepaged_target_node
)
721 for (nid
= last_khugepaged_target_node
+ 1; nid
< MAX_NUMNODES
;
723 if (max_value
== khugepaged_node_load
[nid
]) {
728 last_khugepaged_target_node
= target_node
;
732 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
734 if (IS_ERR(*hpage
)) {
740 khugepaged_alloc_sleep();
750 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, int node
)
752 VM_BUG_ON_PAGE(*hpage
, *hpage
);
754 *hpage
= __alloc_pages_node(node
, gfp
, HPAGE_PMD_ORDER
);
755 if (unlikely(!*hpage
)) {
756 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
757 *hpage
= ERR_PTR(-ENOMEM
);
761 prep_transhuge_page(*hpage
);
762 count_vm_event(THP_COLLAPSE_ALLOC
);
766 static int khugepaged_find_target_node(void)
771 static inline struct page
*alloc_khugepaged_hugepage(void)
775 page
= alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
778 prep_transhuge_page(page
);
782 static struct page
*khugepaged_alloc_hugepage(bool *wait
)
787 hpage
= alloc_khugepaged_hugepage();
789 count_vm_event(THP_COLLAPSE_ALLOC_FAILED
);
794 khugepaged_alloc_sleep();
796 count_vm_event(THP_COLLAPSE_ALLOC
);
797 } while (unlikely(!hpage
) && likely(khugepaged_enabled()));
802 static bool khugepaged_prealloc_page(struct page
**hpage
, bool *wait
)
805 *hpage
= khugepaged_alloc_hugepage(wait
);
807 if (unlikely(!*hpage
))
814 khugepaged_alloc_page(struct page
**hpage
, gfp_t gfp
, int node
)
822 static bool hugepage_vma_check(struct vm_area_struct
*vma
)
824 if ((!(vma
->vm_flags
& VM_HUGEPAGE
) && !khugepaged_always()) ||
825 (vma
->vm_flags
& VM_NOHUGEPAGE
) ||
826 test_bit(MMF_DISABLE_THP
, &vma
->vm_mm
->flags
))
828 if (shmem_file(vma
->vm_file
)) {
829 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE
))
831 return IS_ALIGNED((vma
->vm_start
>> PAGE_SHIFT
) - vma
->vm_pgoff
,
834 if (!vma
->anon_vma
|| vma
->vm_ops
)
836 if (is_vma_temporary_stack(vma
))
838 return !(vma
->vm_flags
& VM_NO_KHUGEPAGED
);
842 * If mmap_sem temporarily dropped, revalidate vma
843 * before taking mmap_sem.
844 * Return 0 if succeeds, otherwise return none-zero
848 static int hugepage_vma_revalidate(struct mm_struct
*mm
, unsigned long address
,
849 struct vm_area_struct
**vmap
)
851 struct vm_area_struct
*vma
;
852 unsigned long hstart
, hend
;
854 if (unlikely(khugepaged_test_exit(mm
)))
855 return SCAN_ANY_PROCESS
;
857 *vmap
= vma
= find_vma(mm
, address
);
859 return SCAN_VMA_NULL
;
861 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
862 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
863 if (address
< hstart
|| address
+ HPAGE_PMD_SIZE
> hend
)
864 return SCAN_ADDRESS_RANGE
;
865 if (!hugepage_vma_check(vma
))
866 return SCAN_VMA_CHECK
;
871 * Bring missing pages in from swap, to complete THP collapse.
872 * Only done if khugepaged_scan_pmd believes it is worthwhile.
874 * Called and returns without pte mapped or spinlocks held,
875 * but with mmap_sem held to protect against vma changes.
878 static bool __collapse_huge_page_swapin(struct mm_struct
*mm
,
879 struct vm_area_struct
*vma
,
880 unsigned long address
, pmd_t
*pmd
,
883 int swapped_in
= 0, ret
= 0;
884 struct vm_fault vmf
= {
887 .flags
= FAULT_FLAG_ALLOW_RETRY
,
889 .pgoff
= linear_page_index(vma
, address
),
892 /* we only decide to swapin, if there is enough young ptes */
893 if (referenced
< HPAGE_PMD_NR
/2) {
894 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 0);
897 vmf
.pte
= pte_offset_map(pmd
, address
);
898 for (; vmf
.address
< address
+ HPAGE_PMD_NR
*PAGE_SIZE
;
899 vmf
.pte
++, vmf
.address
+= PAGE_SIZE
) {
900 vmf
.orig_pte
= *vmf
.pte
;
901 if (!is_swap_pte(vmf
.orig_pte
))
904 ret
= do_swap_page(&vmf
);
906 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
907 if (ret
& VM_FAULT_RETRY
) {
908 down_read(&mm
->mmap_sem
);
909 if (hugepage_vma_revalidate(mm
, address
, &vmf
.vma
)) {
910 /* vma is no longer available, don't continue to swapin */
911 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 0);
914 /* check if the pmd is still valid */
915 if (mm_find_pmd(mm
, address
) != pmd
) {
916 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 0);
920 if (ret
& VM_FAULT_ERROR
) {
921 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 0);
924 /* pte is unmapped now, we need to map it */
925 vmf
.pte
= pte_offset_map(pmd
, vmf
.address
);
929 trace_mm_collapse_huge_page_swapin(mm
, swapped_in
, referenced
, 1);
933 static void collapse_huge_page(struct mm_struct
*mm
,
934 unsigned long address
,
936 int node
, int referenced
)
941 struct page
*new_page
;
942 spinlock_t
*pmd_ptl
, *pte_ptl
;
943 int isolated
= 0, result
= 0;
944 struct mem_cgroup
*memcg
;
945 struct vm_area_struct
*vma
;
946 unsigned long mmun_start
; /* For mmu_notifiers */
947 unsigned long mmun_end
; /* For mmu_notifiers */
950 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
952 /* Only allocate from the target node */
953 gfp
= alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE
;
956 * Before allocating the hugepage, release the mmap_sem read lock.
957 * The allocation can take potentially a long time if it involves
958 * sync compaction, and we do not need to hold the mmap_sem during
959 * that. We will recheck the vma after taking it again in write mode.
961 up_read(&mm
->mmap_sem
);
962 new_page
= khugepaged_alloc_page(hpage
, gfp
, node
);
964 result
= SCAN_ALLOC_HUGE_PAGE_FAIL
;
968 /* Do not oom kill for khugepaged charges */
969 if (unlikely(mem_cgroup_try_charge(new_page
, mm
, gfp
| __GFP_NORETRY
,
971 result
= SCAN_CGROUP_CHARGE_FAIL
;
975 down_read(&mm
->mmap_sem
);
976 result
= hugepage_vma_revalidate(mm
, address
, &vma
);
978 mem_cgroup_cancel_charge(new_page
, memcg
, true);
979 up_read(&mm
->mmap_sem
);
983 pmd
= mm_find_pmd(mm
, address
);
985 result
= SCAN_PMD_NULL
;
986 mem_cgroup_cancel_charge(new_page
, memcg
, true);
987 up_read(&mm
->mmap_sem
);
992 * __collapse_huge_page_swapin always returns with mmap_sem locked.
993 * If it fails, we release mmap_sem and jump out_nolock.
994 * Continuing to collapse causes inconsistency.
996 if (!__collapse_huge_page_swapin(mm
, vma
, address
, pmd
, referenced
)) {
997 mem_cgroup_cancel_charge(new_page
, memcg
, true);
998 up_read(&mm
->mmap_sem
);
1002 up_read(&mm
->mmap_sem
);
1004 * Prevent all access to pagetables with the exception of
1005 * gup_fast later handled by the ptep_clear_flush and the VM
1006 * handled by the anon_vma lock + PG_lock.
1008 down_write(&mm
->mmap_sem
);
1009 result
= hugepage_vma_revalidate(mm
, address
, &vma
);
1012 /* check if the pmd is still valid */
1013 if (mm_find_pmd(mm
, address
) != pmd
)
1016 anon_vma_lock_write(vma
->anon_vma
);
1018 pte
= pte_offset_map(pmd
, address
);
1019 pte_ptl
= pte_lockptr(mm
, pmd
);
1021 mmun_start
= address
;
1022 mmun_end
= address
+ HPAGE_PMD_SIZE
;
1023 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1024 pmd_ptl
= pmd_lock(mm
, pmd
); /* probably unnecessary */
1026 * After this gup_fast can't run anymore. This also removes
1027 * any huge TLB entry from the CPU so we won't allow
1028 * huge and small TLB entries for the same virtual address
1029 * to avoid the risk of CPU bugs in that area.
1031 _pmd
= pmdp_collapse_flush(vma
, address
, pmd
);
1032 spin_unlock(pmd_ptl
);
1033 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1036 isolated
= __collapse_huge_page_isolate(vma
, address
, pte
);
1037 spin_unlock(pte_ptl
);
1039 if (unlikely(!isolated
)) {
1042 BUG_ON(!pmd_none(*pmd
));
1044 * We can only use set_pmd_at when establishing
1045 * hugepmds and never for establishing regular pmds that
1046 * points to regular pagetables. Use pmd_populate for that
1048 pmd_populate(mm
, pmd
, pmd_pgtable(_pmd
));
1049 spin_unlock(pmd_ptl
);
1050 anon_vma_unlock_write(vma
->anon_vma
);
1056 * All pages are isolated and locked so anon_vma rmap
1057 * can't run anymore.
1059 anon_vma_unlock_write(vma
->anon_vma
);
1061 __collapse_huge_page_copy(pte
, new_page
, vma
, address
, pte_ptl
);
1063 __SetPageUptodate(new_page
);
1064 pgtable
= pmd_pgtable(_pmd
);
1066 _pmd
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1067 _pmd
= maybe_pmd_mkwrite(pmd_mkdirty(_pmd
), vma
);
1070 * spin_lock() below is not the equivalent of smp_wmb(), so
1071 * this is needed to avoid the copy_huge_page writes to become
1072 * visible after the set_pmd_at() write.
1077 BUG_ON(!pmd_none(*pmd
));
1078 page_add_new_anon_rmap(new_page
, vma
, address
, true);
1079 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1080 lru_cache_add_active_or_unevictable(new_page
, vma
);
1081 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
1082 set_pmd_at(mm
, address
, pmd
, _pmd
);
1083 update_mmu_cache_pmd(vma
, address
, pmd
);
1084 spin_unlock(pmd_ptl
);
1088 khugepaged_pages_collapsed
++;
1089 result
= SCAN_SUCCEED
;
1091 up_write(&mm
->mmap_sem
);
1093 trace_mm_collapse_huge_page(mm
, isolated
, result
);
1096 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1100 static int khugepaged_scan_pmd(struct mm_struct
*mm
,
1101 struct vm_area_struct
*vma
,
1102 unsigned long address
,
1103 struct page
**hpage
)
1107 int ret
= 0, none_or_zero
= 0, result
= 0, referenced
= 0;
1108 struct page
*page
= NULL
;
1109 unsigned long _address
;
1111 int node
= NUMA_NO_NODE
, unmapped
= 0;
1112 bool writable
= false;
1114 VM_BUG_ON(address
& ~HPAGE_PMD_MASK
);
1116 pmd
= mm_find_pmd(mm
, address
);
1118 result
= SCAN_PMD_NULL
;
1122 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
1123 pte
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
1124 for (_address
= address
, _pte
= pte
; _pte
< pte
+HPAGE_PMD_NR
;
1125 _pte
++, _address
+= PAGE_SIZE
) {
1126 pte_t pteval
= *_pte
;
1127 if (is_swap_pte(pteval
)) {
1128 if (++unmapped
<= khugepaged_max_ptes_swap
) {
1131 result
= SCAN_EXCEED_SWAP_PTE
;
1135 if (pte_none(pteval
) || is_zero_pfn(pte_pfn(pteval
))) {
1136 if (!userfaultfd_armed(vma
) &&
1137 ++none_or_zero
<= khugepaged_max_ptes_none
) {
1140 result
= SCAN_EXCEED_NONE_PTE
;
1144 if (!pte_present(pteval
)) {
1145 result
= SCAN_PTE_NON_PRESENT
;
1148 if (pte_write(pteval
))
1151 page
= vm_normal_page(vma
, _address
, pteval
);
1152 if (unlikely(!page
)) {
1153 result
= SCAN_PAGE_NULL
;
1157 /* TODO: teach khugepaged to collapse THP mapped with pte */
1158 if (PageCompound(page
)) {
1159 result
= SCAN_PAGE_COMPOUND
;
1164 * Record which node the original page is from and save this
1165 * information to khugepaged_node_load[].
1166 * Khupaged will allocate hugepage from the node has the max
1169 node
= page_to_nid(page
);
1170 if (khugepaged_scan_abort(node
)) {
1171 result
= SCAN_SCAN_ABORT
;
1174 khugepaged_node_load
[node
]++;
1175 if (!PageLRU(page
)) {
1176 result
= SCAN_PAGE_LRU
;
1179 if (PageLocked(page
)) {
1180 result
= SCAN_PAGE_LOCK
;
1183 if (!PageAnon(page
)) {
1184 result
= SCAN_PAGE_ANON
;
1189 * cannot use mapcount: can't collapse if there's a gup pin.
1190 * The page must only be referenced by the scanned process
1191 * and page swap cache.
1193 if (page_count(page
) != 1 + PageSwapCache(page
)) {
1194 result
= SCAN_PAGE_COUNT
;
1197 if (pte_young(pteval
) ||
1198 page_is_young(page
) || PageReferenced(page
) ||
1199 mmu_notifier_test_young(vma
->vm_mm
, address
))
1204 result
= SCAN_SUCCEED
;
1207 result
= SCAN_LACK_REFERENCED_PAGE
;
1210 result
= SCAN_PAGE_RO
;
1213 pte_unmap_unlock(pte
, ptl
);
1215 node
= khugepaged_find_target_node();
1216 /* collapse_huge_page will return with the mmap_sem released */
1217 collapse_huge_page(mm
, address
, hpage
, node
, referenced
);
1220 trace_mm_khugepaged_scan_pmd(mm
, page
, writable
, referenced
,
1221 none_or_zero
, result
, unmapped
);
1225 static void collect_mm_slot(struct mm_slot
*mm_slot
)
1227 struct mm_struct
*mm
= mm_slot
->mm
;
1229 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
1231 if (khugepaged_test_exit(mm
)) {
1233 hash_del(&mm_slot
->hash
);
1234 list_del(&mm_slot
->mm_node
);
1237 * Not strictly needed because the mm exited already.
1239 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1242 /* khugepaged_mm_lock actually not necessary for the below */
1243 free_mm_slot(mm_slot
);
1248 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1249 static void retract_page_tables(struct address_space
*mapping
, pgoff_t pgoff
)
1251 struct vm_area_struct
*vma
;
1255 i_mmap_lock_write(mapping
);
1256 vma_interval_tree_foreach(vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1257 /* probably overkill */
1260 addr
= vma
->vm_start
+ ((pgoff
- vma
->vm_pgoff
) << PAGE_SHIFT
);
1261 if (addr
& ~HPAGE_PMD_MASK
)
1263 if (vma
->vm_end
< addr
+ HPAGE_PMD_SIZE
)
1265 pmd
= mm_find_pmd(vma
->vm_mm
, addr
);
1269 * We need exclusive mmap_sem to retract page table.
1270 * If trylock fails we would end up with pte-mapped THP after
1271 * re-fault. Not ideal, but it's more important to not disturb
1272 * the system too much.
1274 if (down_write_trylock(&vma
->vm_mm
->mmap_sem
)) {
1275 spinlock_t
*ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1276 /* assume page table is clear */
1277 _pmd
= pmdp_collapse_flush(vma
, addr
, pmd
);
1279 up_write(&vma
->vm_mm
->mmap_sem
);
1280 atomic_long_dec(&vma
->vm_mm
->nr_ptes
);
1281 pte_free(vma
->vm_mm
, pmd_pgtable(_pmd
));
1284 i_mmap_unlock_write(mapping
);
1288 * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1290 * Basic scheme is simple, details are more complex:
1291 * - allocate and lock a new huge page;
1292 * - scan over radix tree replacing old pages the new one
1293 * + swap in pages if necessary;
1295 * + keep old pages around in case if rollback is required;
1296 * - if replacing succeed:
1299 * + unlock huge page;
1300 * - if replacing failed;
1301 * + put all pages back and unfreeze them;
1302 * + restore gaps in the radix-tree;
1303 * + unlock and free huge page;
1305 static void collapse_shmem(struct mm_struct
*mm
,
1306 struct address_space
*mapping
, pgoff_t start
,
1307 struct page
**hpage
, int node
)
1310 struct page
*page
, *new_page
, *tmp
;
1311 struct mem_cgroup
*memcg
;
1312 pgoff_t index
, end
= start
+ HPAGE_PMD_NR
;
1313 LIST_HEAD(pagelist
);
1314 struct radix_tree_iter iter
;
1316 int nr_none
= 0, result
= SCAN_SUCCEED
;
1318 VM_BUG_ON(start
& (HPAGE_PMD_NR
- 1));
1320 /* Only allocate from the target node */
1321 gfp
= alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE
;
1323 new_page
= khugepaged_alloc_page(hpage
, gfp
, node
);
1325 result
= SCAN_ALLOC_HUGE_PAGE_FAIL
;
1329 /* Do not oom kill for khugepaged charges */
1330 if (unlikely(mem_cgroup_try_charge(new_page
, mm
, gfp
| __GFP_NORETRY
,
1332 result
= SCAN_CGROUP_CHARGE_FAIL
;
1336 __SetPageLocked(new_page
);
1337 __SetPageSwapBacked(new_page
);
1338 new_page
->index
= start
;
1339 new_page
->mapping
= mapping
;
1342 * At this point the new_page is locked and not up-to-date.
1343 * It's safe to insert it into the page cache, because nobody would
1344 * be able to map it or use it in another way until we unlock it.
1348 spin_lock_irq(&mapping
->tree_lock
);
1349 radix_tree_for_each_slot(slot
, &mapping
->page_tree
, &iter
, start
) {
1350 int n
= min(iter
.index
, end
) - index
;
1353 * Stop if extent has been hole-punched, and is now completely
1354 * empty (the more obvious i_size_read() check would take an
1355 * irq-unsafe seqlock on 32-bit).
1357 if (n
>= HPAGE_PMD_NR
) {
1358 result
= SCAN_TRUNCATED
;
1363 * Handle holes in the radix tree: charge it from shmem and
1364 * insert relevant subpage of new_page into the radix-tree.
1366 if (n
&& !shmem_charge(mapping
->host
, n
)) {
1370 for (; index
< min(iter
.index
, end
); index
++) {
1371 radix_tree_insert(&mapping
->page_tree
, index
,
1372 new_page
+ (index
% HPAGE_PMD_NR
));
1380 page
= radix_tree_deref_slot_protected(slot
,
1381 &mapping
->tree_lock
);
1382 if (radix_tree_exceptional_entry(page
) || !PageUptodate(page
)) {
1383 spin_unlock_irq(&mapping
->tree_lock
);
1384 /* swap in or instantiate fallocated page */
1385 if (shmem_getpage(mapping
->host
, index
, &page
,
1390 } else if (trylock_page(page
)) {
1392 spin_unlock_irq(&mapping
->tree_lock
);
1394 result
= SCAN_PAGE_LOCK
;
1399 * The page must be locked, so we can drop the tree_lock
1400 * without racing with truncate.
1402 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1403 VM_BUG_ON_PAGE(!PageUptodate(page
), page
);
1406 * If file was truncated then extended, or hole-punched, before
1407 * we locked the first page, then a THP might be there already.
1409 if (PageTransCompound(page
)) {
1410 result
= SCAN_PAGE_COMPOUND
;
1414 if (page_mapping(page
) != mapping
) {
1415 result
= SCAN_TRUNCATED
;
1419 if (isolate_lru_page(page
)) {
1420 result
= SCAN_DEL_PAGE_LRU
;
1424 if (page_mapped(page
))
1425 unmap_mapping_range(mapping
, index
<< PAGE_SHIFT
,
1428 spin_lock_irq(&mapping
->tree_lock
);
1430 slot
= radix_tree_lookup_slot(&mapping
->page_tree
, index
);
1431 VM_BUG_ON_PAGE(page
!= radix_tree_deref_slot_protected(slot
,
1432 &mapping
->tree_lock
), page
);
1433 VM_BUG_ON_PAGE(page_mapped(page
), page
);
1436 * The page is expected to have page_count() == 3:
1437 * - we hold a pin on it;
1438 * - one reference from radix tree;
1439 * - one from isolate_lru_page;
1441 if (!page_ref_freeze(page
, 3)) {
1442 result
= SCAN_PAGE_COUNT
;
1443 spin_unlock_irq(&mapping
->tree_lock
);
1444 putback_lru_page(page
);
1449 * Add the page to the list to be able to undo the collapse if
1450 * something go wrong.
1452 list_add_tail(&page
->lru
, &pagelist
);
1454 /* Finally, replace with the new page. */
1455 radix_tree_replace_slot(&mapping
->page_tree
, slot
,
1456 new_page
+ (index
% HPAGE_PMD_NR
));
1458 slot
= radix_tree_iter_resume(slot
, &iter
);
1468 * Handle hole in radix tree at the end of the range.
1469 * This code only triggers if there's nothing in radix tree
1473 int n
= end
- index
;
1475 /* Stop if extent has been truncated, and is now empty */
1476 if (n
>= HPAGE_PMD_NR
) {
1477 result
= SCAN_TRUNCATED
;
1480 if (!shmem_charge(mapping
->host
, n
)) {
1484 for (; index
< end
; index
++) {
1485 radix_tree_insert(&mapping
->page_tree
, index
,
1486 new_page
+ (index
% HPAGE_PMD_NR
));
1491 __inc_node_page_state(new_page
, NR_SHMEM_THPS
);
1493 struct zone
*zone
= page_zone(new_page
);
1495 __mod_node_page_state(zone
->zone_pgdat
, NR_FILE_PAGES
, nr_none
);
1496 __mod_node_page_state(zone
->zone_pgdat
, NR_SHMEM
, nr_none
);
1500 spin_unlock_irq(&mapping
->tree_lock
);
1503 if (result
== SCAN_SUCCEED
) {
1505 * Replacing old pages with new one has succeed, now we need to
1506 * copy the content and free old pages.
1509 list_for_each_entry_safe(page
, tmp
, &pagelist
, lru
) {
1510 while (index
< page
->index
) {
1511 clear_highpage(new_page
+ (index
% HPAGE_PMD_NR
));
1514 copy_highpage(new_page
+ (page
->index
% HPAGE_PMD_NR
),
1516 list_del(&page
->lru
);
1517 page
->mapping
= NULL
;
1518 page_ref_unfreeze(page
, 1);
1519 ClearPageActive(page
);
1520 ClearPageUnevictable(page
);
1525 while (index
< end
) {
1526 clear_highpage(new_page
+ (index
% HPAGE_PMD_NR
));
1530 SetPageUptodate(new_page
);
1531 page_ref_add(new_page
, HPAGE_PMD_NR
- 1);
1532 set_page_dirty(new_page
);
1533 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1534 lru_cache_add_anon(new_page
);
1537 * Remove pte page tables, so we can re-fault the page as huge.
1539 retract_page_tables(mapping
, start
);
1542 /* Something went wrong: rollback changes to the radix-tree */
1543 spin_lock_irq(&mapping
->tree_lock
);
1544 mapping
->nrpages
-= nr_none
;
1545 shmem_uncharge(mapping
->host
, nr_none
);
1547 radix_tree_for_each_slot(slot
, &mapping
->page_tree
, &iter
,
1549 if (iter
.index
>= end
)
1551 page
= list_first_entry_or_null(&pagelist
,
1553 if (!page
|| iter
.index
< page
->index
) {
1557 /* Put holes back where they were */
1558 radix_tree_delete(&mapping
->page_tree
,
1563 VM_BUG_ON_PAGE(page
->index
!= iter
.index
, page
);
1565 /* Unfreeze the page. */
1566 list_del(&page
->lru
);
1567 page_ref_unfreeze(page
, 2);
1568 radix_tree_replace_slot(&mapping
->page_tree
,
1570 slot
= radix_tree_iter_resume(slot
, &iter
);
1571 spin_unlock_irq(&mapping
->tree_lock
);
1573 putback_lru_page(page
);
1574 spin_lock_irq(&mapping
->tree_lock
);
1577 spin_unlock_irq(&mapping
->tree_lock
);
1579 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1580 new_page
->mapping
= NULL
;
1583 unlock_page(new_page
);
1585 VM_BUG_ON(!list_empty(&pagelist
));
1586 /* TODO: tracepoints */
1589 static void khugepaged_scan_shmem(struct mm_struct
*mm
,
1590 struct address_space
*mapping
,
1591 pgoff_t start
, struct page
**hpage
)
1593 struct page
*page
= NULL
;
1594 struct radix_tree_iter iter
;
1597 int node
= NUMA_NO_NODE
;
1598 int result
= SCAN_SUCCEED
;
1602 memset(khugepaged_node_load
, 0, sizeof(khugepaged_node_load
));
1604 radix_tree_for_each_slot(slot
, &mapping
->page_tree
, &iter
, start
) {
1605 if (iter
.index
>= start
+ HPAGE_PMD_NR
)
1608 page
= radix_tree_deref_slot(slot
);
1609 if (radix_tree_deref_retry(page
)) {
1610 slot
= radix_tree_iter_retry(&iter
);
1614 if (radix_tree_exception(page
)) {
1615 if (++swap
> khugepaged_max_ptes_swap
) {
1616 result
= SCAN_EXCEED_SWAP_PTE
;
1622 if (PageTransCompound(page
)) {
1623 result
= SCAN_PAGE_COMPOUND
;
1627 node
= page_to_nid(page
);
1628 if (khugepaged_scan_abort(node
)) {
1629 result
= SCAN_SCAN_ABORT
;
1632 khugepaged_node_load
[node
]++;
1634 if (!PageLRU(page
)) {
1635 result
= SCAN_PAGE_LRU
;
1639 if (page_count(page
) != 1 + page_mapcount(page
)) {
1640 result
= SCAN_PAGE_COUNT
;
1645 * We probably should check if the page is referenced here, but
1646 * nobody would transfer pte_young() to PageReferenced() for us.
1647 * And rmap walk here is just too costly...
1652 if (need_resched()) {
1653 slot
= radix_tree_iter_resume(slot
, &iter
);
1659 if (result
== SCAN_SUCCEED
) {
1660 if (present
< HPAGE_PMD_NR
- khugepaged_max_ptes_none
) {
1661 result
= SCAN_EXCEED_NONE_PTE
;
1663 node
= khugepaged_find_target_node();
1664 collapse_shmem(mm
, mapping
, start
, hpage
, node
);
1668 /* TODO: tracepoints */
1671 static void khugepaged_scan_shmem(struct mm_struct
*mm
,
1672 struct address_space
*mapping
,
1673 pgoff_t start
, struct page
**hpage
)
1679 static unsigned int khugepaged_scan_mm_slot(unsigned int pages
,
1680 struct page
**hpage
)
1681 __releases(&khugepaged_mm_lock
)
1682 __acquires(&khugepaged_mm_lock
)
1684 struct mm_slot
*mm_slot
;
1685 struct mm_struct
*mm
;
1686 struct vm_area_struct
*vma
;
1690 VM_BUG_ON(NR_CPUS
!= 1 && !spin_is_locked(&khugepaged_mm_lock
));
1692 if (khugepaged_scan
.mm_slot
)
1693 mm_slot
= khugepaged_scan
.mm_slot
;
1695 mm_slot
= list_entry(khugepaged_scan
.mm_head
.next
,
1696 struct mm_slot
, mm_node
);
1697 khugepaged_scan
.address
= 0;
1698 khugepaged_scan
.mm_slot
= mm_slot
;
1700 spin_unlock(&khugepaged_mm_lock
);
1704 * Don't wait for semaphore (to avoid long wait times). Just move to
1705 * the next mm on the list.
1708 if (unlikely(!down_read_trylock(&mm
->mmap_sem
)))
1709 goto breakouterloop_mmap_sem
;
1710 if (likely(!khugepaged_test_exit(mm
)))
1711 vma
= find_vma(mm
, khugepaged_scan
.address
);
1714 for (; vma
; vma
= vma
->vm_next
) {
1715 unsigned long hstart
, hend
;
1718 if (unlikely(khugepaged_test_exit(mm
))) {
1722 if (!hugepage_vma_check(vma
)) {
1727 hstart
= (vma
->vm_start
+ ~HPAGE_PMD_MASK
) & HPAGE_PMD_MASK
;
1728 hend
= vma
->vm_end
& HPAGE_PMD_MASK
;
1731 if (khugepaged_scan
.address
> hend
)
1733 if (khugepaged_scan
.address
< hstart
)
1734 khugepaged_scan
.address
= hstart
;
1735 VM_BUG_ON(khugepaged_scan
.address
& ~HPAGE_PMD_MASK
);
1737 while (khugepaged_scan
.address
< hend
) {
1740 if (unlikely(khugepaged_test_exit(mm
)))
1741 goto breakouterloop
;
1743 VM_BUG_ON(khugepaged_scan
.address
< hstart
||
1744 khugepaged_scan
.address
+ HPAGE_PMD_SIZE
>
1746 if (shmem_file(vma
->vm_file
)) {
1748 pgoff_t pgoff
= linear_page_index(vma
,
1749 khugepaged_scan
.address
);
1750 if (!shmem_huge_enabled(vma
))
1752 file
= get_file(vma
->vm_file
);
1753 up_read(&mm
->mmap_sem
);
1755 khugepaged_scan_shmem(mm
, file
->f_mapping
,
1759 ret
= khugepaged_scan_pmd(mm
, vma
,
1760 khugepaged_scan
.address
,
1763 /* move to next address */
1764 khugepaged_scan
.address
+= HPAGE_PMD_SIZE
;
1765 progress
+= HPAGE_PMD_NR
;
1767 /* we released mmap_sem so break loop */
1768 goto breakouterloop_mmap_sem
;
1769 if (progress
>= pages
)
1770 goto breakouterloop
;
1774 up_read(&mm
->mmap_sem
); /* exit_mmap will destroy ptes after this */
1775 breakouterloop_mmap_sem
:
1777 spin_lock(&khugepaged_mm_lock
);
1778 VM_BUG_ON(khugepaged_scan
.mm_slot
!= mm_slot
);
1780 * Release the current mm_slot if this mm is about to die, or
1781 * if we scanned all vmas of this mm.
1783 if (khugepaged_test_exit(mm
) || !vma
) {
1785 * Make sure that if mm_users is reaching zero while
1786 * khugepaged runs here, khugepaged_exit will find
1787 * mm_slot not pointing to the exiting mm.
1789 if (mm_slot
->mm_node
.next
!= &khugepaged_scan
.mm_head
) {
1790 khugepaged_scan
.mm_slot
= list_entry(
1791 mm_slot
->mm_node
.next
,
1792 struct mm_slot
, mm_node
);
1793 khugepaged_scan
.address
= 0;
1795 khugepaged_scan
.mm_slot
= NULL
;
1796 khugepaged_full_scans
++;
1799 collect_mm_slot(mm_slot
);
1805 static int khugepaged_has_work(void)
1807 return !list_empty(&khugepaged_scan
.mm_head
) &&
1808 khugepaged_enabled();
1811 static int khugepaged_wait_event(void)
1813 return !list_empty(&khugepaged_scan
.mm_head
) ||
1814 kthread_should_stop();
1817 static void khugepaged_do_scan(void)
1819 struct page
*hpage
= NULL
;
1820 unsigned int progress
= 0, pass_through_head
= 0;
1821 unsigned int pages
= khugepaged_pages_to_scan
;
1824 barrier(); /* write khugepaged_pages_to_scan to local stack */
1826 while (progress
< pages
) {
1827 if (!khugepaged_prealloc_page(&hpage
, &wait
))
1832 if (unlikely(kthread_should_stop() || try_to_freeze()))
1835 spin_lock(&khugepaged_mm_lock
);
1836 if (!khugepaged_scan
.mm_slot
)
1837 pass_through_head
++;
1838 if (khugepaged_has_work() &&
1839 pass_through_head
< 2)
1840 progress
+= khugepaged_scan_mm_slot(pages
- progress
,
1844 spin_unlock(&khugepaged_mm_lock
);
1847 if (!IS_ERR_OR_NULL(hpage
))
1851 static bool khugepaged_should_wakeup(void)
1853 return kthread_should_stop() ||
1854 time_after_eq(jiffies
, khugepaged_sleep_expire
);
1857 static void khugepaged_wait_work(void)
1859 if (khugepaged_has_work()) {
1860 const unsigned long scan_sleep_jiffies
=
1861 msecs_to_jiffies(khugepaged_scan_sleep_millisecs
);
1863 if (!scan_sleep_jiffies
)
1866 khugepaged_sleep_expire
= jiffies
+ scan_sleep_jiffies
;
1867 wait_event_freezable_timeout(khugepaged_wait
,
1868 khugepaged_should_wakeup(),
1869 scan_sleep_jiffies
);
1873 if (khugepaged_enabled())
1874 wait_event_freezable(khugepaged_wait
, khugepaged_wait_event());
1877 static int khugepaged(void *none
)
1879 struct mm_slot
*mm_slot
;
1882 set_user_nice(current
, MAX_NICE
);
1884 while (!kthread_should_stop()) {
1885 khugepaged_do_scan();
1886 khugepaged_wait_work();
1889 spin_lock(&khugepaged_mm_lock
);
1890 mm_slot
= khugepaged_scan
.mm_slot
;
1891 khugepaged_scan
.mm_slot
= NULL
;
1893 collect_mm_slot(mm_slot
);
1894 spin_unlock(&khugepaged_mm_lock
);
1898 static void set_recommended_min_free_kbytes(void)
1902 unsigned long recommended_min
;
1904 for_each_populated_zone(zone
)
1907 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1908 recommended_min
= pageblock_nr_pages
* nr_zones
* 2;
1911 * Make sure that on average at least two pageblocks are almost free
1912 * of another type, one for a migratetype to fall back to and a
1913 * second to avoid subsequent fallbacks of other types There are 3
1914 * MIGRATE_TYPES we care about.
1916 recommended_min
+= pageblock_nr_pages
* nr_zones
*
1917 MIGRATE_PCPTYPES
* MIGRATE_PCPTYPES
;
1919 /* don't ever allow to reserve more than 5% of the lowmem */
1920 recommended_min
= min(recommended_min
,
1921 (unsigned long) nr_free_buffer_pages() / 20);
1922 recommended_min
<<= (PAGE_SHIFT
-10);
1924 if (recommended_min
> min_free_kbytes
) {
1925 if (user_min_free_kbytes
>= 0)
1926 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1927 min_free_kbytes
, recommended_min
);
1929 min_free_kbytes
= recommended_min
;
1931 setup_per_zone_wmarks();
1934 int start_stop_khugepaged(void)
1936 static struct task_struct
*khugepaged_thread __read_mostly
;
1937 static DEFINE_MUTEX(khugepaged_mutex
);
1940 mutex_lock(&khugepaged_mutex
);
1941 if (khugepaged_enabled()) {
1942 if (!khugepaged_thread
)
1943 khugepaged_thread
= kthread_run(khugepaged
, NULL
,
1945 if (IS_ERR(khugepaged_thread
)) {
1946 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1947 err
= PTR_ERR(khugepaged_thread
);
1948 khugepaged_thread
= NULL
;
1952 if (!list_empty(&khugepaged_scan
.mm_head
))
1953 wake_up_interruptible(&khugepaged_wait
);
1955 set_recommended_min_free_kbytes();
1956 } else if (khugepaged_thread
) {
1957 kthread_stop(khugepaged_thread
);
1958 khugepaged_thread
= NULL
;
1961 mutex_unlock(&khugepaged_mutex
);