2 * Copyright (C) 2009 Red Hat, Inc.
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/sched.h>
12 #include <linux/sched/coredump.h>
13 #include <linux/sched/numa_balancing.h>
14 #include <linux/highmem.h>
15 #include <linux/hugetlb.h>
16 #include <linux/mmu_notifier.h>
17 #include <linux/rmap.h>
18 #include <linux/swap.h>
19 #include <linux/shrinker.h>
20 #include <linux/mm_inline.h>
21 #include <linux/swapops.h>
22 #include <linux/dax.h>
23 #include <linux/khugepaged.h>
24 #include <linux/freezer.h>
25 #include <linux/pfn_t.h>
26 #include <linux/mman.h>
27 #include <linux/memremap.h>
28 #include <linux/pagemap.h>
29 #include <linux/debugfs.h>
30 #include <linux/migrate.h>
31 #include <linux/hashtable.h>
32 #include <linux/userfaultfd_k.h>
33 #include <linux/page_idle.h>
34 #include <linux/shmem_fs.h>
35 #include <linux/oom.h>
38 #include <asm/pgalloc.h>
42 * By default transparent hugepage support is disabled in order that avoid
43 * to risk increase the memory footprint of applications without a guaranteed
44 * benefit. When transparent hugepage support is enabled, is for all mappings,
45 * and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
49 unsigned long transparent_hugepage_flags __read_mostly
=
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
60 static struct shrinker deferred_split_shrinker
;
62 static atomic_t huge_zero_refcount
;
63 struct page
*huge_zero_page __read_mostly
;
65 static struct page
*get_huge_zero_page(void)
67 struct page
*zero_page
;
69 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
70 return READ_ONCE(huge_zero_page
);
72 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
78 count_vm_event(THP_ZERO_PAGE_ALLOC
);
80 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
82 __free_pages(zero_page
, compound_order(zero_page
));
86 /* We take additional reference here. It will be put back by shrinker */
87 atomic_set(&huge_zero_refcount
, 2);
89 return READ_ONCE(huge_zero_page
);
92 static void put_huge_zero_page(void)
95 * Counter should never go to zero here. Only shrinker can put
98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
101 struct page
*mm_get_huge_zero_page(struct mm_struct
*mm
)
103 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
104 return READ_ONCE(huge_zero_page
);
106 if (!get_huge_zero_page())
109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
110 put_huge_zero_page();
112 return READ_ONCE(huge_zero_page
);
115 void mm_put_huge_zero_page(struct mm_struct
*mm
)
117 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
118 put_huge_zero_page();
121 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
122 struct shrink_control
*sc
)
124 /* we can free zero page only if last reference remains */
125 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
128 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
129 struct shrink_control
*sc
)
131 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
132 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
133 BUG_ON(zero_page
== NULL
);
134 __free_pages(zero_page
, compound_order(zero_page
));
141 static struct shrinker huge_zero_page_shrinker
= {
142 .count_objects
= shrink_huge_zero_page_count
,
143 .scan_objects
= shrink_huge_zero_page_scan
,
144 .seeks
= DEFAULT_SEEKS
,
148 static ssize_t
enabled_show(struct kobject
*kobj
,
149 struct kobj_attribute
*attr
, char *buf
)
151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
))
152 return sprintf(buf
, "[always] madvise never\n");
153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
154 return sprintf(buf
, "always [madvise] never\n");
156 return sprintf(buf
, "always madvise [never]\n");
159 static ssize_t
enabled_store(struct kobject
*kobj
,
160 struct kobj_attribute
*attr
,
161 const char *buf
, size_t count
)
165 if (!memcmp("always", buf
,
166 min(sizeof("always")-1, count
))) {
167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
168 set_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
169 } else if (!memcmp("madvise", buf
,
170 min(sizeof("madvise")-1, count
))) {
171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
173 } else if (!memcmp("never", buf
,
174 min(sizeof("never")-1, count
))) {
175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
181 int err
= start_stop_khugepaged();
187 static struct kobj_attribute enabled_attr
=
188 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
190 ssize_t
single_hugepage_flag_show(struct kobject
*kobj
,
191 struct kobj_attribute
*attr
, char *buf
,
192 enum transparent_hugepage_flag flag
)
194 return sprintf(buf
, "%d\n",
195 !!test_bit(flag
, &transparent_hugepage_flags
));
198 ssize_t
single_hugepage_flag_store(struct kobject
*kobj
,
199 struct kobj_attribute
*attr
,
200 const char *buf
, size_t count
,
201 enum transparent_hugepage_flag flag
)
206 ret
= kstrtoul(buf
, 10, &value
);
213 set_bit(flag
, &transparent_hugepage_flags
);
215 clear_bit(flag
, &transparent_hugepage_flags
);
220 static ssize_t
defrag_show(struct kobject
*kobj
,
221 struct kobj_attribute
*attr
, char *buf
)
223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
224 return sprintf(buf
, "[always] defer defer+madvise madvise never\n");
225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
226 return sprintf(buf
, "always [defer] defer+madvise madvise never\n");
227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
228 return sprintf(buf
, "always defer [defer+madvise] madvise never\n");
229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
230 return sprintf(buf
, "always defer defer+madvise [madvise] never\n");
231 return sprintf(buf
, "always defer defer+madvise madvise [never]\n");
234 static ssize_t
defrag_store(struct kobject
*kobj
,
235 struct kobj_attribute
*attr
,
236 const char *buf
, size_t count
)
238 if (!memcmp("always", buf
,
239 min(sizeof("always")-1, count
))) {
240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
244 } else if (!memcmp("defer+madvise", buf
,
245 min(sizeof("defer+madvise")-1, count
))) {
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
250 } else if (!memcmp("defer", buf
,
251 min(sizeof("defer")-1, count
))) {
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
256 } else if (!memcmp("madvise", buf
,
257 min(sizeof("madvise")-1, count
))) {
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
262 } else if (!memcmp("never", buf
,
263 min(sizeof("never")-1, count
))) {
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
273 static struct kobj_attribute defrag_attr
=
274 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
276 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
277 struct kobj_attribute
*attr
, char *buf
)
279 return single_hugepage_flag_show(kobj
, attr
, buf
,
280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
282 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
283 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
285 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
288 static struct kobj_attribute use_zero_page_attr
=
289 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
291 static ssize_t
hpage_pmd_size_show(struct kobject
*kobj
,
292 struct kobj_attribute
*attr
, char *buf
)
294 return sprintf(buf
, "%lu\n", HPAGE_PMD_SIZE
);
296 static struct kobj_attribute hpage_pmd_size_attr
=
297 __ATTR_RO(hpage_pmd_size
);
299 #ifdef CONFIG_DEBUG_VM
300 static ssize_t
debug_cow_show(struct kobject
*kobj
,
301 struct kobj_attribute
*attr
, char *buf
)
303 return single_hugepage_flag_show(kobj
, attr
, buf
,
304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
306 static ssize_t
debug_cow_store(struct kobject
*kobj
,
307 struct kobj_attribute
*attr
,
308 const char *buf
, size_t count
)
310 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
313 static struct kobj_attribute debug_cow_attr
=
314 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
315 #endif /* CONFIG_DEBUG_VM */
317 static struct attribute
*hugepage_attr
[] = {
320 &use_zero_page_attr
.attr
,
321 &hpage_pmd_size_attr
.attr
,
322 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
323 &shmem_enabled_attr
.attr
,
325 #ifdef CONFIG_DEBUG_VM
326 &debug_cow_attr
.attr
,
331 static struct attribute_group hugepage_attr_group
= {
332 .attrs
= hugepage_attr
,
335 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
339 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
340 if (unlikely(!*hugepage_kobj
)) {
341 pr_err("failed to create transparent hugepage kobject\n");
345 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
347 pr_err("failed to register transparent hugepage group\n");
351 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
353 pr_err("failed to register transparent hugepage group\n");
354 goto remove_hp_group
;
360 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
362 kobject_put(*hugepage_kobj
);
366 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
368 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
369 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
370 kobject_put(hugepage_kobj
);
373 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
378 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
381 #endif /* CONFIG_SYSFS */
383 static int __init
hugepage_init(void)
386 struct kobject
*hugepage_kobj
;
388 if (!has_transparent_hugepage()) {
389 transparent_hugepage_flags
= 0;
394 * hugepages can't be allocated by the buddy allocator
396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
>= MAX_ORDER
);
398 * we use page->mapping and page->index in second tail page
399 * as list_head: assuming THP order >= 2
401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
403 err
= hugepage_init_sysfs(&hugepage_kobj
);
407 err
= khugepaged_init();
411 err
= register_shrinker(&huge_zero_page_shrinker
);
413 goto err_hzp_shrinker
;
414 err
= register_shrinker(&deferred_split_shrinker
);
416 goto err_split_shrinker
;
419 * By default disable transparent hugepages on smaller systems,
420 * where the extra memory used could hurt more than TLB overhead
421 * is likely to save. The admin can still enable it through /sys.
423 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
))) {
424 transparent_hugepage_flags
= 0;
428 err
= start_stop_khugepaged();
434 unregister_shrinker(&deferred_split_shrinker
);
436 unregister_shrinker(&huge_zero_page_shrinker
);
438 khugepaged_destroy();
440 hugepage_exit_sysfs(hugepage_kobj
);
444 subsys_initcall(hugepage_init
);
446 static int __init
setup_transparent_hugepage(char *str
)
451 if (!strcmp(str
, "always")) {
452 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
453 &transparent_hugepage_flags
);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
455 &transparent_hugepage_flags
);
457 } else if (!strcmp(str
, "madvise")) {
458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
459 &transparent_hugepage_flags
);
460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
461 &transparent_hugepage_flags
);
463 } else if (!strcmp(str
, "never")) {
464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
465 &transparent_hugepage_flags
);
466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
467 &transparent_hugepage_flags
);
472 pr_warn("transparent_hugepage= cannot parse, ignored\n");
475 __setup("transparent_hugepage=", setup_transparent_hugepage
);
477 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
479 if (likely(vma
->vm_flags
& VM_WRITE
))
480 pmd
= pmd_mkwrite(pmd
);
484 static inline struct list_head
*page_deferred_list(struct page
*page
)
487 * ->lru in the tail pages is occupied by compound_head.
488 * Let's use ->mapping + ->index in the second tail page as list_head.
490 return (struct list_head
*)&page
[2].mapping
;
493 void prep_transhuge_page(struct page
*page
)
496 * we use page->mapping and page->indexlru in second tail page
497 * as list_head: assuming THP order >= 2
500 INIT_LIST_HEAD(page_deferred_list(page
));
501 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
504 unsigned long __thp_get_unmapped_area(struct file
*filp
, unsigned long len
,
505 loff_t off
, unsigned long flags
, unsigned long size
)
508 loff_t off_end
= off
+ len
;
509 loff_t off_align
= round_up(off
, size
);
510 unsigned long len_pad
;
512 if (off_end
<= off_align
|| (off_end
- off_align
) < size
)
515 len_pad
= len
+ size
;
516 if (len_pad
< len
|| (off
+ len_pad
) < off
)
519 addr
= current
->mm
->get_unmapped_area(filp
, 0, len_pad
,
520 off
>> PAGE_SHIFT
, flags
);
521 if (IS_ERR_VALUE(addr
))
524 addr
+= (off
- addr
) & (size
- 1);
528 unsigned long thp_get_unmapped_area(struct file
*filp
, unsigned long addr
,
529 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
531 loff_t off
= (loff_t
)pgoff
<< PAGE_SHIFT
;
535 if (!IS_DAX(filp
->f_mapping
->host
) || !IS_ENABLED(CONFIG_FS_DAX_PMD
))
538 addr
= __thp_get_unmapped_area(filp
, len
, off
, flags
, PMD_SIZE
);
543 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
545 EXPORT_SYMBOL_GPL(thp_get_unmapped_area
);
547 static int __do_huge_pmd_anonymous_page(struct vm_fault
*vmf
, struct page
*page
,
550 struct vm_area_struct
*vma
= vmf
->vma
;
551 struct mem_cgroup
*memcg
;
553 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
556 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
558 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, gfp
, &memcg
, true)) {
560 count_vm_event(THP_FAULT_FALLBACK
);
561 return VM_FAULT_FALLBACK
;
564 pgtable
= pte_alloc_one(vma
->vm_mm
, haddr
);
565 if (unlikely(!pgtable
)) {
570 clear_huge_page(page
, haddr
, HPAGE_PMD_NR
);
572 * The memory barrier inside __SetPageUptodate makes sure that
573 * clear_huge_page writes become visible before the set_pmd_at()
576 __SetPageUptodate(page
);
578 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
579 if (unlikely(!pmd_none(*vmf
->pmd
))) {
584 ret
= check_stable_address_space(vma
->vm_mm
);
588 /* Deliver the page fault to userland */
589 if (userfaultfd_missing(vma
)) {
592 spin_unlock(vmf
->ptl
);
593 mem_cgroup_cancel_charge(page
, memcg
, true);
595 pte_free(vma
->vm_mm
, pgtable
);
596 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
597 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
601 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
602 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
603 page_add_new_anon_rmap(page
, vma
, haddr
, true);
604 mem_cgroup_commit_charge(page
, memcg
, false, true);
605 lru_cache_add_active_or_unevictable(page
, vma
);
606 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
607 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
608 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
609 atomic_long_inc(&vma
->vm_mm
->nr_ptes
);
610 spin_unlock(vmf
->ptl
);
611 count_vm_event(THP_FAULT_ALLOC
);
616 spin_unlock(vmf
->ptl
);
619 pte_free(vma
->vm_mm
, pgtable
);
620 mem_cgroup_cancel_charge(page
, memcg
, true);
627 * always: directly stall for all thp allocations
628 * defer: wake kswapd and fail if not immediately available
629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
630 * fail if not immediately available
631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
633 * never: never stall for any thp allocation
635 static inline gfp_t
alloc_hugepage_direct_gfpmask(struct vm_area_struct
*vma
)
637 const bool vma_madvised
= !!(vma
->vm_flags
& VM_HUGEPAGE
);
639 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
640 return GFP_TRANSHUGE
| (vma_madvised
? 0 : __GFP_NORETRY
);
641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
642 return GFP_TRANSHUGE_LIGHT
| __GFP_KSWAPD_RECLAIM
;
643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
644 return GFP_TRANSHUGE_LIGHT
| (vma_madvised
? __GFP_DIRECT_RECLAIM
:
645 __GFP_KSWAPD_RECLAIM
);
646 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
647 return GFP_TRANSHUGE_LIGHT
| (vma_madvised
? __GFP_DIRECT_RECLAIM
:
649 return GFP_TRANSHUGE_LIGHT
;
652 /* Caller must hold page table lock. */
653 static bool set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
654 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
655 struct page
*zero_page
)
660 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
661 entry
= pmd_mkhuge(entry
);
663 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
664 set_pmd_at(mm
, haddr
, pmd
, entry
);
665 atomic_long_inc(&mm
->nr_ptes
);
669 int do_huge_pmd_anonymous_page(struct vm_fault
*vmf
)
671 struct vm_area_struct
*vma
= vmf
->vma
;
674 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
676 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
677 return VM_FAULT_FALLBACK
;
678 if (unlikely(anon_vma_prepare(vma
)))
680 if (unlikely(khugepaged_enter(vma
, vma
->vm_flags
)))
682 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
683 !mm_forbids_zeropage(vma
->vm_mm
) &&
684 transparent_hugepage_use_zero_page()) {
686 struct page
*zero_page
;
689 pgtable
= pte_alloc_one(vma
->vm_mm
, haddr
);
690 if (unlikely(!pgtable
))
692 zero_page
= mm_get_huge_zero_page(vma
->vm_mm
);
693 if (unlikely(!zero_page
)) {
694 pte_free(vma
->vm_mm
, pgtable
);
695 count_vm_event(THP_FAULT_FALLBACK
);
696 return VM_FAULT_FALLBACK
;
698 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
701 if (pmd_none(*vmf
->pmd
)) {
702 ret
= check_stable_address_space(vma
->vm_mm
);
704 spin_unlock(vmf
->ptl
);
705 } else if (userfaultfd_missing(vma
)) {
706 spin_unlock(vmf
->ptl
);
707 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
708 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
710 set_huge_zero_page(pgtable
, vma
->vm_mm
, vma
,
711 haddr
, vmf
->pmd
, zero_page
);
712 spin_unlock(vmf
->ptl
);
716 spin_unlock(vmf
->ptl
);
718 pte_free(vma
->vm_mm
, pgtable
);
721 gfp
= alloc_hugepage_direct_gfpmask(vma
);
722 page
= alloc_hugepage_vma(gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
723 if (unlikely(!page
)) {
724 count_vm_event(THP_FAULT_FALLBACK
);
725 return VM_FAULT_FALLBACK
;
727 prep_transhuge_page(page
);
728 return __do_huge_pmd_anonymous_page(vmf
, page
, gfp
);
731 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
732 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
,
735 struct mm_struct
*mm
= vma
->vm_mm
;
739 ptl
= pmd_lock(mm
, pmd
);
740 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
741 if (pfn_t_devmap(pfn
))
742 entry
= pmd_mkdevmap(entry
);
744 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
745 entry
= maybe_pmd_mkwrite(entry
, vma
);
749 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
750 atomic_long_inc(&mm
->nr_ptes
);
753 set_pmd_at(mm
, addr
, pmd
, entry
);
754 update_mmu_cache_pmd(vma
, addr
, pmd
);
758 int vmf_insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
759 pmd_t
*pmd
, pfn_t pfn
, bool write
)
761 pgprot_t pgprot
= vma
->vm_page_prot
;
762 pgtable_t pgtable
= NULL
;
764 * If we had pmd_special, we could avoid all these restrictions,
765 * but we need to be consistent with PTEs and architectures that
766 * can't support a 'special' bit.
768 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
769 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
770 (VM_PFNMAP
|VM_MIXEDMAP
));
771 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
772 BUG_ON(!pfn_t_devmap(pfn
));
774 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
775 return VM_FAULT_SIGBUS
;
777 if (arch_needs_pgtable_deposit()) {
778 pgtable
= pte_alloc_one(vma
->vm_mm
, addr
);
783 track_pfn_insert(vma
, &pgprot
, pfn
);
785 insert_pfn_pmd(vma
, addr
, pmd
, pfn
, pgprot
, write
, pgtable
);
786 return VM_FAULT_NOPAGE
;
788 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd
);
790 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
791 static pud_t
maybe_pud_mkwrite(pud_t pud
, struct vm_area_struct
*vma
)
793 if (likely(vma
->vm_flags
& VM_WRITE
))
794 pud
= pud_mkwrite(pud
);
798 static void insert_pfn_pud(struct vm_area_struct
*vma
, unsigned long addr
,
799 pud_t
*pud
, pfn_t pfn
, pgprot_t prot
, bool write
)
801 struct mm_struct
*mm
= vma
->vm_mm
;
805 ptl
= pud_lock(mm
, pud
);
806 entry
= pud_mkhuge(pfn_t_pud(pfn
, prot
));
807 if (pfn_t_devmap(pfn
))
808 entry
= pud_mkdevmap(entry
);
810 entry
= pud_mkyoung(pud_mkdirty(entry
));
811 entry
= maybe_pud_mkwrite(entry
, vma
);
813 set_pud_at(mm
, addr
, pud
, entry
);
814 update_mmu_cache_pud(vma
, addr
, pud
);
818 int vmf_insert_pfn_pud(struct vm_area_struct
*vma
, unsigned long addr
,
819 pud_t
*pud
, pfn_t pfn
, bool write
)
821 pgprot_t pgprot
= vma
->vm_page_prot
;
823 * If we had pud_special, we could avoid all these restrictions,
824 * but we need to be consistent with PTEs and architectures that
825 * can't support a 'special' bit.
827 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
828 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
829 (VM_PFNMAP
|VM_MIXEDMAP
));
830 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
831 BUG_ON(!pfn_t_devmap(pfn
));
833 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
834 return VM_FAULT_SIGBUS
;
836 track_pfn_insert(vma
, &pgprot
, pfn
);
838 insert_pfn_pud(vma
, addr
, pud
, pfn
, pgprot
, write
);
839 return VM_FAULT_NOPAGE
;
841 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud
);
842 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
844 static void touch_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
845 pmd_t
*pmd
, int flags
)
849 _pmd
= pmd_mkyoung(*pmd
);
850 if (flags
& FOLL_WRITE
)
851 _pmd
= pmd_mkdirty(_pmd
);
852 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
853 pmd
, _pmd
, flags
& FOLL_WRITE
))
854 update_mmu_cache_pmd(vma
, addr
, pmd
);
857 struct page
*follow_devmap_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
858 pmd_t
*pmd
, int flags
)
860 unsigned long pfn
= pmd_pfn(*pmd
);
861 struct mm_struct
*mm
= vma
->vm_mm
;
862 struct dev_pagemap
*pgmap
;
865 assert_spin_locked(pmd_lockptr(mm
, pmd
));
868 * When we COW a devmap PMD entry, we split it into PTEs, so we should
869 * not be in this function with `flags & FOLL_COW` set.
871 WARN_ONCE(flags
& FOLL_COW
, "mm: In follow_devmap_pmd with FOLL_COW set");
873 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
876 if (pmd_present(*pmd
) && pmd_devmap(*pmd
))
881 if (flags
& FOLL_TOUCH
)
882 touch_pmd(vma
, addr
, pmd
, flags
);
885 * device mapped pages can only be returned if the
886 * caller will manage the page reference count.
888 if (!(flags
& FOLL_GET
))
889 return ERR_PTR(-EEXIST
);
891 pfn
+= (addr
& ~PMD_MASK
) >> PAGE_SHIFT
;
892 pgmap
= get_dev_pagemap(pfn
, NULL
);
894 return ERR_PTR(-EFAULT
);
895 page
= pfn_to_page(pfn
);
897 put_dev_pagemap(pgmap
);
902 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
903 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
904 struct vm_area_struct
*vma
)
906 spinlock_t
*dst_ptl
, *src_ptl
;
907 struct page
*src_page
;
909 pgtable_t pgtable
= NULL
;
912 /* Skip if can be re-fill on fault */
913 if (!vma_is_anonymous(vma
))
916 pgtable
= pte_alloc_one(dst_mm
, addr
);
917 if (unlikely(!pgtable
))
920 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
921 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
922 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
926 if (unlikely(!pmd_trans_huge(pmd
))) {
927 pte_free(dst_mm
, pgtable
);
931 * When page table lock is held, the huge zero pmd should not be
932 * under splitting since we don't split the page itself, only pmd to
935 if (is_huge_zero_pmd(pmd
)) {
936 struct page
*zero_page
;
938 * get_huge_zero_page() will never allocate a new page here,
939 * since we already have a zero page to copy. It just takes a
942 zero_page
= mm_get_huge_zero_page(dst_mm
);
943 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
949 src_page
= pmd_page(pmd
);
950 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
952 page_dup_rmap(src_page
, true);
953 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
954 atomic_long_inc(&dst_mm
->nr_ptes
);
955 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
957 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
958 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
959 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
963 spin_unlock(src_ptl
);
964 spin_unlock(dst_ptl
);
969 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
970 static void touch_pud(struct vm_area_struct
*vma
, unsigned long addr
,
971 pud_t
*pud
, int flags
)
975 _pud
= pud_mkyoung(*pud
);
976 if (flags
& FOLL_WRITE
)
977 _pud
= pud_mkdirty(_pud
);
978 if (pudp_set_access_flags(vma
, addr
& HPAGE_PUD_MASK
,
979 pud
, _pud
, flags
& FOLL_WRITE
))
980 update_mmu_cache_pud(vma
, addr
, pud
);
983 struct page
*follow_devmap_pud(struct vm_area_struct
*vma
, unsigned long addr
,
984 pud_t
*pud
, int flags
)
986 unsigned long pfn
= pud_pfn(*pud
);
987 struct mm_struct
*mm
= vma
->vm_mm
;
988 struct dev_pagemap
*pgmap
;
991 assert_spin_locked(pud_lockptr(mm
, pud
));
993 if (flags
& FOLL_WRITE
&& !pud_write(*pud
))
996 if (pud_present(*pud
) && pud_devmap(*pud
))
1001 if (flags
& FOLL_TOUCH
)
1002 touch_pud(vma
, addr
, pud
, flags
);
1005 * device mapped pages can only be returned if the
1006 * caller will manage the page reference count.
1008 if (!(flags
& FOLL_GET
))
1009 return ERR_PTR(-EEXIST
);
1011 pfn
+= (addr
& ~PUD_MASK
) >> PAGE_SHIFT
;
1012 pgmap
= get_dev_pagemap(pfn
, NULL
);
1014 return ERR_PTR(-EFAULT
);
1015 page
= pfn_to_page(pfn
);
1017 put_dev_pagemap(pgmap
);
1022 int copy_huge_pud(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1023 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1024 struct vm_area_struct
*vma
)
1026 spinlock_t
*dst_ptl
, *src_ptl
;
1030 dst_ptl
= pud_lock(dst_mm
, dst_pud
);
1031 src_ptl
= pud_lockptr(src_mm
, src_pud
);
1032 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1036 if (unlikely(!pud_trans_huge(pud
) && !pud_devmap(pud
)))
1040 * When page table lock is held, the huge zero pud should not be
1041 * under splitting since we don't split the page itself, only pud to
1044 if (is_huge_zero_pud(pud
)) {
1045 /* No huge zero pud yet */
1048 pudp_set_wrprotect(src_mm
, addr
, src_pud
);
1049 pud
= pud_mkold(pud_wrprotect(pud
));
1050 set_pud_at(dst_mm
, addr
, dst_pud
, pud
);
1054 spin_unlock(src_ptl
);
1055 spin_unlock(dst_ptl
);
1059 void huge_pud_set_accessed(struct vm_fault
*vmf
, pud_t orig_pud
)
1062 unsigned long haddr
;
1063 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1065 vmf
->ptl
= pud_lock(vmf
->vma
->vm_mm
, vmf
->pud
);
1066 if (unlikely(!pud_same(*vmf
->pud
, orig_pud
)))
1069 entry
= pud_mkyoung(orig_pud
);
1071 entry
= pud_mkdirty(entry
);
1072 haddr
= vmf
->address
& HPAGE_PUD_MASK
;
1073 if (pudp_set_access_flags(vmf
->vma
, haddr
, vmf
->pud
, entry
, write
))
1074 update_mmu_cache_pud(vmf
->vma
, vmf
->address
, vmf
->pud
);
1077 spin_unlock(vmf
->ptl
);
1079 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1081 void huge_pmd_set_accessed(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1084 unsigned long haddr
;
1085 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1087 vmf
->ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1088 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1091 entry
= pmd_mkyoung(orig_pmd
);
1093 entry
= pmd_mkdirty(entry
);
1094 haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1095 if (pmdp_set_access_flags(vmf
->vma
, haddr
, vmf
->pmd
, entry
, write
))
1096 update_mmu_cache_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
);
1099 spin_unlock(vmf
->ptl
);
1102 static int do_huge_pmd_wp_page_fallback(struct vm_fault
*vmf
, pmd_t orig_pmd
,
1105 struct vm_area_struct
*vma
= vmf
->vma
;
1106 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1107 struct mem_cgroup
*memcg
;
1111 struct page
**pages
;
1112 unsigned long mmun_start
; /* For mmu_notifiers */
1113 unsigned long mmun_end
; /* For mmu_notifiers */
1115 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1117 if (unlikely(!pages
)) {
1118 ret
|= VM_FAULT_OOM
;
1122 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1123 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
, vma
,
1124 vmf
->address
, page_to_nid(page
));
1125 if (unlikely(!pages
[i
] ||
1126 mem_cgroup_try_charge(pages
[i
], vma
->vm_mm
,
1127 GFP_KERNEL
, &memcg
, false))) {
1131 memcg
= (void *)page_private(pages
[i
]);
1132 set_page_private(pages
[i
], 0);
1133 mem_cgroup_cancel_charge(pages
[i
], memcg
,
1138 ret
|= VM_FAULT_OOM
;
1141 set_page_private(pages
[i
], (unsigned long)memcg
);
1144 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1145 copy_user_highpage(pages
[i
], page
+ i
,
1146 haddr
+ PAGE_SIZE
* i
, vma
);
1147 __SetPageUptodate(pages
[i
]);
1152 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1153 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1155 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1156 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1157 goto out_free_pages
;
1158 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1160 pmdp_huge_clear_flush_notify(vma
, haddr
, vmf
->pmd
);
1161 /* leave pmd empty until pte is filled */
1163 pgtable
= pgtable_trans_huge_withdraw(vma
->vm_mm
, vmf
->pmd
);
1164 pmd_populate(vma
->vm_mm
, &_pmd
, pgtable
);
1166 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1168 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1169 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1170 memcg
= (void *)page_private(pages
[i
]);
1171 set_page_private(pages
[i
], 0);
1172 page_add_new_anon_rmap(pages
[i
], vmf
->vma
, haddr
, false);
1173 mem_cgroup_commit_charge(pages
[i
], memcg
, false, false);
1174 lru_cache_add_active_or_unevictable(pages
[i
], vma
);
1175 vmf
->pte
= pte_offset_map(&_pmd
, haddr
);
1176 VM_BUG_ON(!pte_none(*vmf
->pte
));
1177 set_pte_at(vma
->vm_mm
, haddr
, vmf
->pte
, entry
);
1178 pte_unmap(vmf
->pte
);
1182 smp_wmb(); /* make pte visible before pmd */
1183 pmd_populate(vma
->vm_mm
, vmf
->pmd
, pgtable
);
1184 page_remove_rmap(page
, true);
1185 spin_unlock(vmf
->ptl
);
1187 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1189 ret
|= VM_FAULT_WRITE
;
1196 spin_unlock(vmf
->ptl
);
1197 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1198 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1199 memcg
= (void *)page_private(pages
[i
]);
1200 set_page_private(pages
[i
], 0);
1201 mem_cgroup_cancel_charge(pages
[i
], memcg
, false);
1208 int do_huge_pmd_wp_page(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1210 struct vm_area_struct
*vma
= vmf
->vma
;
1211 struct page
*page
= NULL
, *new_page
;
1212 struct mem_cgroup
*memcg
;
1213 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1214 unsigned long mmun_start
; /* For mmu_notifiers */
1215 unsigned long mmun_end
; /* For mmu_notifiers */
1216 gfp_t huge_gfp
; /* for allocation and charge */
1219 vmf
->ptl
= pmd_lockptr(vma
->vm_mm
, vmf
->pmd
);
1220 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1221 if (is_huge_zero_pmd(orig_pmd
))
1223 spin_lock(vmf
->ptl
);
1224 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1227 page
= pmd_page(orig_pmd
);
1228 VM_BUG_ON_PAGE(!PageCompound(page
) || !PageHead(page
), page
);
1230 * We can only reuse the page if nobody else maps the huge page or it's
1233 if (page_trans_huge_mapcount(page
, NULL
) == 1) {
1235 entry
= pmd_mkyoung(orig_pmd
);
1236 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1237 if (pmdp_set_access_flags(vma
, haddr
, vmf
->pmd
, entry
, 1))
1238 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1239 ret
|= VM_FAULT_WRITE
;
1243 spin_unlock(vmf
->ptl
);
1245 if (transparent_hugepage_enabled(vma
) &&
1246 !transparent_hugepage_debug_cow()) {
1247 huge_gfp
= alloc_hugepage_direct_gfpmask(vma
);
1248 new_page
= alloc_hugepage_vma(huge_gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
1252 if (likely(new_page
)) {
1253 prep_transhuge_page(new_page
);
1256 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1257 ret
|= VM_FAULT_FALLBACK
;
1259 ret
= do_huge_pmd_wp_page_fallback(vmf
, orig_pmd
, page
);
1260 if (ret
& VM_FAULT_OOM
) {
1261 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1262 ret
|= VM_FAULT_FALLBACK
;
1266 count_vm_event(THP_FAULT_FALLBACK
);
1270 if (unlikely(mem_cgroup_try_charge(new_page
, vma
->vm_mm
,
1271 huge_gfp
, &memcg
, true))) {
1273 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1276 ret
|= VM_FAULT_FALLBACK
;
1277 count_vm_event(THP_FAULT_FALLBACK
);
1281 count_vm_event(THP_FAULT_ALLOC
);
1284 clear_huge_page(new_page
, haddr
, HPAGE_PMD_NR
);
1286 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1287 __SetPageUptodate(new_page
);
1290 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1291 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1293 spin_lock(vmf
->ptl
);
1296 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1297 spin_unlock(vmf
->ptl
);
1298 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1303 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1304 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1305 pmdp_huge_clear_flush_notify(vma
, haddr
, vmf
->pmd
);
1306 page_add_new_anon_rmap(new_page
, vma
, haddr
, true);
1307 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1308 lru_cache_add_active_or_unevictable(new_page
, vma
);
1309 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
1310 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1312 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1314 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1315 page_remove_rmap(page
, true);
1318 ret
|= VM_FAULT_WRITE
;
1320 spin_unlock(vmf
->ptl
);
1322 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1326 spin_unlock(vmf
->ptl
);
1331 * FOLL_FORCE can write to even unwritable pmd's, but only
1332 * after we've gone through a COW cycle and they are dirty.
1334 static inline bool can_follow_write_pmd(pmd_t pmd
, unsigned int flags
)
1336 return pmd_write(pmd
) ||
1337 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pmd_dirty(pmd
));
1340 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1345 struct mm_struct
*mm
= vma
->vm_mm
;
1346 struct page
*page
= NULL
;
1348 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1350 if (flags
& FOLL_WRITE
&& !can_follow_write_pmd(*pmd
, flags
))
1353 /* Avoid dumping huge zero page */
1354 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1355 return ERR_PTR(-EFAULT
);
1357 /* Full NUMA hinting faults to serialise migration in fault paths */
1358 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1361 page
= pmd_page(*pmd
);
1362 VM_BUG_ON_PAGE(!PageHead(page
) && !is_zone_device_page(page
), page
);
1363 if (flags
& FOLL_TOUCH
)
1364 touch_pmd(vma
, addr
, pmd
, flags
);
1365 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1367 * We don't mlock() pte-mapped THPs. This way we can avoid
1368 * leaking mlocked pages into non-VM_LOCKED VMAs.
1372 * In most cases the pmd is the only mapping of the page as we
1373 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1374 * writable private mappings in populate_vma_page_range().
1376 * The only scenario when we have the page shared here is if we
1377 * mlocking read-only mapping shared over fork(). We skip
1378 * mlocking such pages.
1382 * We can expect PageDoubleMap() to be stable under page lock:
1383 * for file pages we set it in page_add_file_rmap(), which
1384 * requires page to be locked.
1387 if (PageAnon(page
) && compound_mapcount(page
) != 1)
1389 if (PageDoubleMap(page
) || !page
->mapping
)
1391 if (!trylock_page(page
))
1394 if (page
->mapping
&& !PageDoubleMap(page
))
1395 mlock_vma_page(page
);
1399 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1400 VM_BUG_ON_PAGE(!PageCompound(page
) && !is_zone_device_page(page
), page
);
1401 if (flags
& FOLL_GET
)
1408 /* NUMA hinting page fault entry point for trans huge pmds */
1409 int do_huge_pmd_numa_page(struct vm_fault
*vmf
, pmd_t pmd
)
1411 struct vm_area_struct
*vma
= vmf
->vma
;
1412 struct anon_vma
*anon_vma
= NULL
;
1414 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1415 int page_nid
= -1, this_nid
= numa_node_id();
1416 int target_nid
, last_cpupid
= -1;
1418 bool migrated
= false;
1422 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1423 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
)))
1427 * If there are potential migrations, wait for completion and retry
1428 * without disrupting NUMA hinting information. Do not relock and
1429 * check_same as the page may no longer be mapped.
1431 if (unlikely(pmd_trans_migrating(*vmf
->pmd
))) {
1432 page
= pmd_page(*vmf
->pmd
);
1433 if (!get_page_unless_zero(page
))
1435 spin_unlock(vmf
->ptl
);
1436 wait_on_page_locked(page
);
1441 page
= pmd_page(pmd
);
1442 BUG_ON(is_huge_zero_page(page
));
1443 page_nid
= page_to_nid(page
);
1444 last_cpupid
= page_cpupid_last(page
);
1445 count_vm_numa_event(NUMA_HINT_FAULTS
);
1446 if (page_nid
== this_nid
) {
1447 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1448 flags
|= TNF_FAULT_LOCAL
;
1451 /* See similar comment in do_numa_page for explanation */
1452 if (!pmd_savedwrite(pmd
))
1453 flags
|= TNF_NO_GROUP
;
1456 * Acquire the page lock to serialise THP migrations but avoid dropping
1457 * page_table_lock if at all possible
1459 page_locked
= trylock_page(page
);
1460 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1461 if (target_nid
== -1) {
1462 /* If the page was locked, there are no parallel migrations */
1467 /* Migration could have started since the pmd_trans_migrating check */
1470 if (!get_page_unless_zero(page
))
1472 spin_unlock(vmf
->ptl
);
1473 wait_on_page_locked(page
);
1479 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1480 * to serialises splits
1483 spin_unlock(vmf
->ptl
);
1484 anon_vma
= page_lock_anon_vma_read(page
);
1486 /* Confirm the PMD did not change while page_table_lock was released */
1487 spin_lock(vmf
->ptl
);
1488 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
))) {
1495 /* Bail if we fail to protect against THP splits for any reason */
1496 if (unlikely(!anon_vma
)) {
1503 * The page_table_lock above provides a memory barrier
1504 * with change_protection_range.
1506 if (mm_tlb_flush_pending(vma
->vm_mm
))
1507 flush_tlb_range(vma
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
1510 * Migrate the THP to the requested node, returns with page unlocked
1511 * and access rights restored.
1513 spin_unlock(vmf
->ptl
);
1514 migrated
= migrate_misplaced_transhuge_page(vma
->vm_mm
, vma
,
1515 vmf
->pmd
, pmd
, vmf
->address
, page
, target_nid
);
1517 flags
|= TNF_MIGRATED
;
1518 page_nid
= target_nid
;
1520 flags
|= TNF_MIGRATE_FAIL
;
1524 BUG_ON(!PageLocked(page
));
1525 was_writable
= pmd_savedwrite(pmd
);
1526 pmd
= pmd_modify(pmd
, vma
->vm_page_prot
);
1527 pmd
= pmd_mkyoung(pmd
);
1529 pmd
= pmd_mkwrite(pmd
);
1530 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, pmd
);
1531 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1534 spin_unlock(vmf
->ptl
);
1538 page_unlock_anon_vma_read(anon_vma
);
1541 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
,
1548 * Return true if we do MADV_FREE successfully on entire pmd page.
1549 * Otherwise, return false.
1551 bool madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1552 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1557 struct mm_struct
*mm
= tlb
->mm
;
1560 tlb_remove_check_page_size_change(tlb
, HPAGE_PMD_SIZE
);
1562 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1567 if (is_huge_zero_pmd(orig_pmd
))
1570 page
= pmd_page(orig_pmd
);
1572 * If other processes are mapping this page, we couldn't discard
1573 * the page unless they all do MADV_FREE so let's skip the page.
1575 if (page_mapcount(page
) != 1)
1578 if (!trylock_page(page
))
1582 * If user want to discard part-pages of THP, split it so MADV_FREE
1583 * will deactivate only them.
1585 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1588 split_huge_page(page
);
1594 if (PageDirty(page
))
1595 ClearPageDirty(page
);
1598 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1599 pmdp_invalidate(vma
, addr
, pmd
);
1600 orig_pmd
= pmd_mkold(orig_pmd
);
1601 orig_pmd
= pmd_mkclean(orig_pmd
);
1603 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1604 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1607 mark_page_lazyfree(page
);
1615 static inline void zap_deposited_table(struct mm_struct
*mm
, pmd_t
*pmd
)
1619 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1620 pte_free(mm
, pgtable
);
1621 atomic_long_dec(&mm
->nr_ptes
);
1624 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1625 pmd_t
*pmd
, unsigned long addr
)
1630 tlb_remove_check_page_size_change(tlb
, HPAGE_PMD_SIZE
);
1632 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1636 * For architectures like ppc64 we look at deposited pgtable
1637 * when calling pmdp_huge_get_and_clear. So do the
1638 * pgtable_trans_huge_withdraw after finishing pmdp related
1641 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1643 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1644 if (vma_is_dax(vma
)) {
1645 if (arch_needs_pgtable_deposit())
1646 zap_deposited_table(tlb
->mm
, pmd
);
1648 if (is_huge_zero_pmd(orig_pmd
))
1649 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1650 } else if (is_huge_zero_pmd(orig_pmd
)) {
1651 zap_deposited_table(tlb
->mm
, pmd
);
1653 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1655 struct page
*page
= pmd_page(orig_pmd
);
1656 page_remove_rmap(page
, true);
1657 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1658 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1659 if (PageAnon(page
)) {
1660 zap_deposited_table(tlb
->mm
, pmd
);
1661 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1663 if (arch_needs_pgtable_deposit())
1664 zap_deposited_table(tlb
->mm
, pmd
);
1665 add_mm_counter(tlb
->mm
, MM_FILEPAGES
, -HPAGE_PMD_NR
);
1668 tlb_remove_page_size(tlb
, page
, HPAGE_PMD_SIZE
);
1673 #ifndef pmd_move_must_withdraw
1674 static inline int pmd_move_must_withdraw(spinlock_t
*new_pmd_ptl
,
1675 spinlock_t
*old_pmd_ptl
,
1676 struct vm_area_struct
*vma
)
1679 * With split pmd lock we also need to move preallocated
1680 * PTE page table if new_pmd is on different PMD page table.
1682 * We also don't deposit and withdraw tables for file pages.
1684 return (new_pmd_ptl
!= old_pmd_ptl
) && vma_is_anonymous(vma
);
1688 bool move_huge_pmd(struct vm_area_struct
*vma
, unsigned long old_addr
,
1689 unsigned long new_addr
, unsigned long old_end
,
1690 pmd_t
*old_pmd
, pmd_t
*new_pmd
, bool *need_flush
)
1692 spinlock_t
*old_ptl
, *new_ptl
;
1694 struct mm_struct
*mm
= vma
->vm_mm
;
1695 bool force_flush
= false;
1697 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1698 (new_addr
& ~HPAGE_PMD_MASK
) ||
1699 old_end
- old_addr
< HPAGE_PMD_SIZE
)
1703 * The destination pmd shouldn't be established, free_pgtables()
1704 * should have release it.
1706 if (WARN_ON(!pmd_none(*new_pmd
))) {
1707 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1712 * We don't have to worry about the ordering of src and dst
1713 * ptlocks because exclusive mmap_sem prevents deadlock.
1715 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1717 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1718 if (new_ptl
!= old_ptl
)
1719 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1720 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1721 if (pmd_present(pmd
) && pmd_dirty(pmd
))
1723 VM_BUG_ON(!pmd_none(*new_pmd
));
1725 if (pmd_move_must_withdraw(new_ptl
, old_ptl
, vma
)) {
1727 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1728 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1730 set_pmd_at(mm
, new_addr
, new_pmd
, pmd_mksoft_dirty(pmd
));
1731 if (new_ptl
!= old_ptl
)
1732 spin_unlock(new_ptl
);
1734 flush_tlb_range(vma
, old_addr
, old_addr
+ PMD_SIZE
);
1737 spin_unlock(old_ptl
);
1745 * - 0 if PMD could not be locked
1746 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1747 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1749 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1750 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1752 struct mm_struct
*mm
= vma
->vm_mm
;
1755 bool preserve_write
;
1758 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1762 preserve_write
= prot_numa
&& pmd_write(*pmd
);
1766 * Avoid trapping faults against the zero page. The read-only
1767 * data is likely to be read-cached on the local CPU and
1768 * local/remote hits to the zero page are not interesting.
1770 if (prot_numa
&& is_huge_zero_pmd(*pmd
))
1773 if (prot_numa
&& pmd_protnone(*pmd
))
1777 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1778 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1779 * which is also under down_read(mmap_sem):
1782 * change_huge_pmd(prot_numa=1)
1783 * pmdp_huge_get_and_clear_notify()
1784 * madvise_dontneed()
1786 * pmd_trans_huge(*pmd) == 0 (without ptl)
1789 * // pmd is re-established
1791 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1792 * which may break userspace.
1794 * pmdp_invalidate() is required to make sure we don't miss
1795 * dirty/young flags set by hardware.
1798 pmdp_invalidate(vma
, addr
, pmd
);
1801 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1804 if (pmd_dirty(*pmd
))
1805 entry
= pmd_mkdirty(entry
);
1806 if (pmd_young(*pmd
))
1807 entry
= pmd_mkyoung(entry
);
1809 entry
= pmd_modify(entry
, newprot
);
1811 entry
= pmd_mk_savedwrite(entry
);
1813 set_pmd_at(mm
, addr
, pmd
, entry
);
1814 BUG_ON(vma_is_anonymous(vma
) && !preserve_write
&& pmd_write(entry
));
1821 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1823 * Note that if it returns page table lock pointer, this routine returns without
1824 * unlocking page table lock. So callers must unlock it.
1826 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1829 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1830 if (likely(pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
)))
1837 * Returns true if a given pud maps a thp, false otherwise.
1839 * Note that if it returns true, this routine returns without unlocking page
1840 * table lock. So callers must unlock it.
1842 spinlock_t
*__pud_trans_huge_lock(pud_t
*pud
, struct vm_area_struct
*vma
)
1846 ptl
= pud_lock(vma
->vm_mm
, pud
);
1847 if (likely(pud_trans_huge(*pud
) || pud_devmap(*pud
)))
1853 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1854 int zap_huge_pud(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1855 pud_t
*pud
, unsigned long addr
)
1860 ptl
= __pud_trans_huge_lock(pud
, vma
);
1864 * For architectures like ppc64 we look at deposited pgtable
1865 * when calling pudp_huge_get_and_clear. So do the
1866 * pgtable_trans_huge_withdraw after finishing pudp related
1869 orig_pud
= pudp_huge_get_and_clear_full(tlb
->mm
, addr
, pud
,
1871 tlb_remove_pud_tlb_entry(tlb
, pud
, addr
);
1872 if (vma_is_dax(vma
)) {
1874 /* No zero page support yet */
1876 /* No support for anonymous PUD pages yet */
1882 static void __split_huge_pud_locked(struct vm_area_struct
*vma
, pud_t
*pud
,
1883 unsigned long haddr
)
1885 VM_BUG_ON(haddr
& ~HPAGE_PUD_MASK
);
1886 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
1887 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PUD_SIZE
, vma
);
1888 VM_BUG_ON(!pud_trans_huge(*pud
) && !pud_devmap(*pud
));
1890 count_vm_event(THP_SPLIT_PUD
);
1892 pudp_huge_clear_flush_notify(vma
, haddr
, pud
);
1895 void __split_huge_pud(struct vm_area_struct
*vma
, pud_t
*pud
,
1896 unsigned long address
)
1899 struct mm_struct
*mm
= vma
->vm_mm
;
1900 unsigned long haddr
= address
& HPAGE_PUD_MASK
;
1902 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PUD_SIZE
);
1903 ptl
= pud_lock(mm
, pud
);
1904 if (unlikely(!pud_trans_huge(*pud
) && !pud_devmap(*pud
)))
1906 __split_huge_pud_locked(vma
, pud
, haddr
);
1910 mmu_notifier_invalidate_range_end(mm
, haddr
, haddr
+ HPAGE_PUD_SIZE
);
1912 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1914 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
1915 unsigned long haddr
, pmd_t
*pmd
)
1917 struct mm_struct
*mm
= vma
->vm_mm
;
1922 /* leave pmd empty until pte is filled */
1923 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
1925 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1926 pmd_populate(mm
, &_pmd
, pgtable
);
1928 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1930 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
1931 entry
= pte_mkspecial(entry
);
1932 pte
= pte_offset_map(&_pmd
, haddr
);
1933 VM_BUG_ON(!pte_none(*pte
));
1934 set_pte_at(mm
, haddr
, pte
, entry
);
1937 smp_wmb(); /* make pte visible before pmd */
1938 pmd_populate(mm
, pmd
, pgtable
);
1941 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1942 unsigned long haddr
, bool freeze
)
1944 struct mm_struct
*mm
= vma
->vm_mm
;
1948 bool young
, write
, dirty
, soft_dirty
;
1952 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
1953 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
1954 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
1955 VM_BUG_ON(!pmd_trans_huge(*pmd
) && !pmd_devmap(*pmd
));
1957 count_vm_event(THP_SPLIT_PMD
);
1959 if (!vma_is_anonymous(vma
)) {
1960 _pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
1962 * We are going to unmap this huge page. So
1963 * just go ahead and zap it
1965 if (arch_needs_pgtable_deposit())
1966 zap_deposited_table(mm
, pmd
);
1967 if (vma_is_dax(vma
))
1969 page
= pmd_page(_pmd
);
1970 if (!PageReferenced(page
) && pmd_young(_pmd
))
1971 SetPageReferenced(page
);
1972 page_remove_rmap(page
, true);
1974 add_mm_counter(mm
, MM_FILEPAGES
, -HPAGE_PMD_NR
);
1976 } else if (is_huge_zero_pmd(*pmd
)) {
1977 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
1980 page
= pmd_page(*pmd
);
1981 VM_BUG_ON_PAGE(!page_count(page
), page
);
1982 page_ref_add(page
, HPAGE_PMD_NR
- 1);
1983 write
= pmd_write(*pmd
);
1984 young
= pmd_young(*pmd
);
1985 dirty
= pmd_dirty(*pmd
);
1986 soft_dirty
= pmd_soft_dirty(*pmd
);
1988 pmdp_huge_split_prepare(vma
, haddr
, pmd
);
1989 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1990 pmd_populate(mm
, &_pmd
, pgtable
);
1992 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
1995 * Note that NUMA hinting access restrictions are not
1996 * transferred to avoid any possibility of altering
1997 * permissions across VMAs.
2000 swp_entry_t swp_entry
;
2001 swp_entry
= make_migration_entry(page
+ i
, write
);
2002 entry
= swp_entry_to_pte(swp_entry
);
2004 entry
= pte_swp_mksoft_dirty(entry
);
2006 entry
= mk_pte(page
+ i
, READ_ONCE(vma
->vm_page_prot
));
2007 entry
= maybe_mkwrite(entry
, vma
);
2009 entry
= pte_wrprotect(entry
);
2011 entry
= pte_mkold(entry
);
2013 entry
= pte_mksoft_dirty(entry
);
2016 SetPageDirty(page
+ i
);
2017 pte
= pte_offset_map(&_pmd
, addr
);
2018 BUG_ON(!pte_none(*pte
));
2019 set_pte_at(mm
, addr
, pte
, entry
);
2020 atomic_inc(&page
[i
]._mapcount
);
2025 * Set PG_double_map before dropping compound_mapcount to avoid
2026 * false-negative page_mapped().
2028 if (compound_mapcount(page
) > 1 && !TestSetPageDoubleMap(page
)) {
2029 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2030 atomic_inc(&page
[i
]._mapcount
);
2033 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
2034 /* Last compound_mapcount is gone. */
2035 __dec_node_page_state(page
, NR_ANON_THPS
);
2036 if (TestClearPageDoubleMap(page
)) {
2037 /* No need in mapcount reference anymore */
2038 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2039 atomic_dec(&page
[i
]._mapcount
);
2043 smp_wmb(); /* make pte visible before pmd */
2045 * Up to this point the pmd is present and huge and userland has the
2046 * whole access to the hugepage during the split (which happens in
2047 * place). If we overwrite the pmd with the not-huge version pointing
2048 * to the pte here (which of course we could if all CPUs were bug
2049 * free), userland could trigger a small page size TLB miss on the
2050 * small sized TLB while the hugepage TLB entry is still established in
2051 * the huge TLB. Some CPU doesn't like that.
2052 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2053 * 383 on page 93. Intel should be safe but is also warns that it's
2054 * only safe if the permission and cache attributes of the two entries
2055 * loaded in the two TLB is identical (which should be the case here).
2056 * But it is generally safer to never allow small and huge TLB entries
2057 * for the same virtual address to be loaded simultaneously. So instead
2058 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2059 * current pmd notpresent (atomically because here the pmd_trans_huge
2060 * and pmd_trans_splitting must remain set at all times on the pmd
2061 * until the split is complete for this pmd), then we flush the SMP TLB
2062 * and finally we write the non-huge version of the pmd entry with
2065 pmdp_invalidate(vma
, haddr
, pmd
);
2066 pmd_populate(mm
, pmd
, pgtable
);
2069 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2070 page_remove_rmap(page
+ i
, false);
2076 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2077 unsigned long address
, bool freeze
, struct page
*page
)
2080 struct mm_struct
*mm
= vma
->vm_mm
;
2081 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2083 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2084 ptl
= pmd_lock(mm
, pmd
);
2087 * If caller asks to setup a migration entries, we need a page to check
2088 * pmd against. Otherwise we can end up replacing wrong page.
2090 VM_BUG_ON(freeze
&& !page
);
2091 if (page
&& page
!= pmd_page(*pmd
))
2094 if (pmd_trans_huge(*pmd
)) {
2095 page
= pmd_page(*pmd
);
2096 if (PageMlocked(page
))
2097 clear_page_mlock(page
);
2098 } else if (!pmd_devmap(*pmd
))
2100 __split_huge_pmd_locked(vma
, pmd
, haddr
, freeze
);
2103 mmu_notifier_invalidate_range_end(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2106 void split_huge_pmd_address(struct vm_area_struct
*vma
, unsigned long address
,
2107 bool freeze
, struct page
*page
)
2114 pgd
= pgd_offset(vma
->vm_mm
, address
);
2115 if (!pgd_present(*pgd
))
2118 p4d
= p4d_offset(pgd
, address
);
2119 if (!p4d_present(*p4d
))
2122 pud
= pud_offset(p4d
, address
);
2123 if (!pud_present(*pud
))
2126 pmd
= pmd_offset(pud
, address
);
2128 __split_huge_pmd(vma
, pmd
, address
, freeze
, page
);
2131 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2132 unsigned long start
,
2137 * If the new start address isn't hpage aligned and it could
2138 * previously contain an hugepage: check if we need to split
2141 if (start
& ~HPAGE_PMD_MASK
&&
2142 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2143 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2144 split_huge_pmd_address(vma
, start
, false, NULL
);
2147 * If the new end address isn't hpage aligned and it could
2148 * previously contain an hugepage: check if we need to split
2151 if (end
& ~HPAGE_PMD_MASK
&&
2152 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2153 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2154 split_huge_pmd_address(vma
, end
, false, NULL
);
2157 * If we're also updating the vma->vm_next->vm_start, if the new
2158 * vm_next->vm_start isn't page aligned and it could previously
2159 * contain an hugepage: check if we need to split an huge pmd.
2161 if (adjust_next
> 0) {
2162 struct vm_area_struct
*next
= vma
->vm_next
;
2163 unsigned long nstart
= next
->vm_start
;
2164 nstart
+= adjust_next
<< PAGE_SHIFT
;
2165 if (nstart
& ~HPAGE_PMD_MASK
&&
2166 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2167 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2168 split_huge_pmd_address(next
, nstart
, false, NULL
);
2172 static void freeze_page(struct page
*page
)
2174 enum ttu_flags ttu_flags
= TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
|
2175 TTU_RMAP_LOCKED
| TTU_SPLIT_HUGE_PMD
;
2178 VM_BUG_ON_PAGE(!PageHead(page
), page
);
2181 ttu_flags
|= TTU_MIGRATION
;
2183 unmap_success
= try_to_unmap(page
, ttu_flags
);
2184 VM_BUG_ON_PAGE(!unmap_success
, page
);
2187 static void unfreeze_page(struct page
*page
)
2190 if (PageTransHuge(page
)) {
2191 remove_migration_ptes(page
, page
, true);
2193 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2194 remove_migration_ptes(page
+ i
, page
+ i
, true);
2198 static void __split_huge_page_tail(struct page
*head
, int tail
,
2199 struct lruvec
*lruvec
, struct list_head
*list
)
2201 struct page
*page_tail
= head
+ tail
;
2203 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
2204 VM_BUG_ON_PAGE(page_ref_count(page_tail
) != 0, page_tail
);
2207 * tail_page->_refcount is zero and not changing from under us. But
2208 * get_page_unless_zero() may be running from under us on the
2209 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2210 * atomic_add(), we would then run atomic_set() concurrently with
2211 * get_page_unless_zero(), and atomic_set() is implemented in C not
2212 * using locked ops. spin_unlock on x86 sometime uses locked ops
2213 * because of PPro errata 66, 92, so unless somebody can guarantee
2214 * atomic_set() here would be safe on all archs (and not only on x86),
2215 * it's safer to use atomic_inc()/atomic_add().
2217 if (PageAnon(head
) && !PageSwapCache(head
)) {
2218 page_ref_inc(page_tail
);
2220 /* Additional pin to radix tree */
2221 page_ref_add(page_tail
, 2);
2224 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
2225 page_tail
->flags
|= (head
->flags
&
2226 ((1L << PG_referenced
) |
2227 (1L << PG_swapbacked
) |
2228 (1L << PG_swapcache
) |
2229 (1L << PG_mlocked
) |
2230 (1L << PG_uptodate
) |
2233 (1L << PG_unevictable
) |
2237 * After clearing PageTail the gup refcount can be released.
2238 * Page flags also must be visible before we make the page non-compound.
2242 clear_compound_head(page_tail
);
2244 if (page_is_young(head
))
2245 set_page_young(page_tail
);
2246 if (page_is_idle(head
))
2247 set_page_idle(page_tail
);
2249 /* ->mapping in first tail page is compound_mapcount */
2250 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
2252 page_tail
->mapping
= head
->mapping
;
2254 page_tail
->index
= head
->index
+ tail
;
2255 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
2256 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
2259 static void __split_huge_page(struct page
*page
, struct list_head
*list
,
2260 unsigned long flags
)
2262 struct page
*head
= compound_head(page
);
2263 struct zone
*zone
= page_zone(head
);
2264 struct lruvec
*lruvec
;
2268 lruvec
= mem_cgroup_page_lruvec(head
, zone
->zone_pgdat
);
2270 /* complete memcg works before add pages to LRU */
2271 mem_cgroup_split_huge_fixup(head
);
2273 if (!PageAnon(page
))
2274 end
= DIV_ROUND_UP(i_size_read(head
->mapping
->host
), PAGE_SIZE
);
2276 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
2277 __split_huge_page_tail(head
, i
, lruvec
, list
);
2278 /* Some pages can be beyond i_size: drop them from page cache */
2279 if (head
[i
].index
>= end
) {
2280 __ClearPageDirty(head
+ i
);
2281 __delete_from_page_cache(head
+ i
, NULL
);
2282 if (IS_ENABLED(CONFIG_SHMEM
) && PageSwapBacked(head
))
2283 shmem_uncharge(head
->mapping
->host
, 1);
2288 ClearPageCompound(head
);
2289 /* See comment in __split_huge_page_tail() */
2290 if (PageAnon(head
)) {
2291 /* Additional pin to radix tree of swap cache */
2292 if (PageSwapCache(head
))
2293 page_ref_add(head
, 2);
2297 /* Additional pin to radix tree */
2298 page_ref_add(head
, 2);
2299 spin_unlock(&head
->mapping
->tree_lock
);
2302 spin_unlock_irqrestore(zone_lru_lock(page_zone(head
)), flags
);
2304 unfreeze_page(head
);
2306 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2307 struct page
*subpage
= head
+ i
;
2308 if (subpage
== page
)
2310 unlock_page(subpage
);
2313 * Subpages may be freed if there wasn't any mapping
2314 * like if add_to_swap() is running on a lru page that
2315 * had its mapping zapped. And freeing these pages
2316 * requires taking the lru_lock so we do the put_page
2317 * of the tail pages after the split is complete.
2323 int total_mapcount(struct page
*page
)
2325 int i
, compound
, ret
;
2327 VM_BUG_ON_PAGE(PageTail(page
), page
);
2329 if (likely(!PageCompound(page
)))
2330 return atomic_read(&page
->_mapcount
) + 1;
2332 compound
= compound_mapcount(page
);
2336 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2337 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
2338 /* File pages has compound_mapcount included in _mapcount */
2339 if (!PageAnon(page
))
2340 return ret
- compound
* HPAGE_PMD_NR
;
2341 if (PageDoubleMap(page
))
2342 ret
-= HPAGE_PMD_NR
;
2347 * This calculates accurately how many mappings a transparent hugepage
2348 * has (unlike page_mapcount() which isn't fully accurate). This full
2349 * accuracy is primarily needed to know if copy-on-write faults can
2350 * reuse the page and change the mapping to read-write instead of
2351 * copying them. At the same time this returns the total_mapcount too.
2353 * The function returns the highest mapcount any one of the subpages
2354 * has. If the return value is one, even if different processes are
2355 * mapping different subpages of the transparent hugepage, they can
2356 * all reuse it, because each process is reusing a different subpage.
2358 * The total_mapcount is instead counting all virtual mappings of the
2359 * subpages. If the total_mapcount is equal to "one", it tells the
2360 * caller all mappings belong to the same "mm" and in turn the
2361 * anon_vma of the transparent hugepage can become the vma->anon_vma
2362 * local one as no other process may be mapping any of the subpages.
2364 * It would be more accurate to replace page_mapcount() with
2365 * page_trans_huge_mapcount(), however we only use
2366 * page_trans_huge_mapcount() in the copy-on-write faults where we
2367 * need full accuracy to avoid breaking page pinning, because
2368 * page_trans_huge_mapcount() is slower than page_mapcount().
2370 int page_trans_huge_mapcount(struct page
*page
, int *total_mapcount
)
2372 int i
, ret
, _total_mapcount
, mapcount
;
2374 /* hugetlbfs shouldn't call it */
2375 VM_BUG_ON_PAGE(PageHuge(page
), page
);
2377 if (likely(!PageTransCompound(page
))) {
2378 mapcount
= atomic_read(&page
->_mapcount
) + 1;
2380 *total_mapcount
= mapcount
;
2384 page
= compound_head(page
);
2386 _total_mapcount
= ret
= 0;
2387 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2388 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
2389 ret
= max(ret
, mapcount
);
2390 _total_mapcount
+= mapcount
;
2392 if (PageDoubleMap(page
)) {
2394 _total_mapcount
-= HPAGE_PMD_NR
;
2396 mapcount
= compound_mapcount(page
);
2398 _total_mapcount
+= mapcount
;
2400 *total_mapcount
= _total_mapcount
;
2404 /* Racy check whether the huge page can be split */
2405 bool can_split_huge_page(struct page
*page
, int *pextra_pins
)
2409 /* Additional pins from radix tree */
2411 extra_pins
= PageSwapCache(page
) ? HPAGE_PMD_NR
: 0;
2413 extra_pins
= HPAGE_PMD_NR
;
2415 *pextra_pins
= extra_pins
;
2416 return total_mapcount(page
) == page_count(page
) - extra_pins
- 1;
2420 * This function splits huge page into normal pages. @page can point to any
2421 * subpage of huge page to split. Split doesn't change the position of @page.
2423 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2424 * The huge page must be locked.
2426 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2428 * Both head page and tail pages will inherit mapping, flags, and so on from
2431 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2432 * they are not mapped.
2434 * Returns 0 if the hugepage is split successfully.
2435 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2438 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
2440 struct page
*head
= compound_head(page
);
2441 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(head
));
2442 struct anon_vma
*anon_vma
= NULL
;
2443 struct address_space
*mapping
= NULL
;
2444 int count
, mapcount
, extra_pins
, ret
;
2446 unsigned long flags
;
2448 VM_BUG_ON_PAGE(is_huge_zero_page(page
), page
);
2449 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2450 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
2452 if (PageAnon(head
)) {
2454 * The caller does not necessarily hold an mmap_sem that would
2455 * prevent the anon_vma disappearing so we first we take a
2456 * reference to it and then lock the anon_vma for write. This
2457 * is similar to page_lock_anon_vma_read except the write lock
2458 * is taken to serialise against parallel split or collapse
2461 anon_vma
= page_get_anon_vma(head
);
2467 anon_vma_lock_write(anon_vma
);
2469 mapping
= head
->mapping
;
2478 i_mmap_lock_read(mapping
);
2482 * Racy check if we can split the page, before freeze_page() will
2485 if (!can_split_huge_page(head
, &extra_pins
)) {
2490 mlocked
= PageMlocked(page
);
2492 VM_BUG_ON_PAGE(compound_mapcount(head
), head
);
2494 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2498 /* prevent PageLRU to go away from under us, and freeze lru stats */
2499 spin_lock_irqsave(zone_lru_lock(page_zone(head
)), flags
);
2504 spin_lock(&mapping
->tree_lock
);
2505 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
2508 * Check if the head page is present in radix tree.
2509 * We assume all tail are present too, if head is there.
2511 if (radix_tree_deref_slot_protected(pslot
,
2512 &mapping
->tree_lock
) != head
)
2516 /* Prevent deferred_split_scan() touching ->_refcount */
2517 spin_lock(&pgdata
->split_queue_lock
);
2518 count
= page_count(head
);
2519 mapcount
= total_mapcount(head
);
2520 if (!mapcount
&& page_ref_freeze(head
, 1 + extra_pins
)) {
2521 if (!list_empty(page_deferred_list(head
))) {
2522 pgdata
->split_queue_len
--;
2523 list_del(page_deferred_list(head
));
2526 __dec_node_page_state(page
, NR_SHMEM_THPS
);
2527 spin_unlock(&pgdata
->split_queue_lock
);
2528 __split_huge_page(page
, list
, flags
);
2531 if (IS_ENABLED(CONFIG_DEBUG_VM
) && mapcount
) {
2532 pr_alert("total_mapcount: %u, page_count(): %u\n",
2535 dump_page(head
, NULL
);
2536 dump_page(page
, "total_mapcount(head) > 0");
2539 spin_unlock(&pgdata
->split_queue_lock
);
2541 spin_unlock(&mapping
->tree_lock
);
2542 spin_unlock_irqrestore(zone_lru_lock(page_zone(head
)), flags
);
2543 unfreeze_page(head
);
2549 anon_vma_unlock_write(anon_vma
);
2550 put_anon_vma(anon_vma
);
2553 i_mmap_unlock_read(mapping
);
2555 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
2559 void free_transhuge_page(struct page
*page
)
2561 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
2562 unsigned long flags
;
2564 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2565 if (!list_empty(page_deferred_list(page
))) {
2566 pgdata
->split_queue_len
--;
2567 list_del(page_deferred_list(page
));
2569 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2570 free_compound_page(page
);
2573 void deferred_split_huge_page(struct page
*page
)
2575 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
2576 unsigned long flags
;
2578 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
2580 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2581 if (list_empty(page_deferred_list(page
))) {
2582 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
2583 list_add_tail(page_deferred_list(page
), &pgdata
->split_queue
);
2584 pgdata
->split_queue_len
++;
2586 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2589 static unsigned long deferred_split_count(struct shrinker
*shrink
,
2590 struct shrink_control
*sc
)
2592 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2593 return ACCESS_ONCE(pgdata
->split_queue_len
);
2596 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
2597 struct shrink_control
*sc
)
2599 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2600 unsigned long flags
;
2601 LIST_HEAD(list
), *pos
, *next
;
2605 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2606 /* Take pin on all head pages to avoid freeing them under us */
2607 list_for_each_safe(pos
, next
, &pgdata
->split_queue
) {
2608 page
= list_entry((void *)pos
, struct page
, mapping
);
2609 page
= compound_head(page
);
2610 if (get_page_unless_zero(page
)) {
2611 list_move(page_deferred_list(page
), &list
);
2613 /* We lost race with put_compound_page() */
2614 list_del_init(page_deferred_list(page
));
2615 pgdata
->split_queue_len
--;
2617 if (!--sc
->nr_to_scan
)
2620 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2622 list_for_each_safe(pos
, next
, &list
) {
2623 page
= list_entry((void *)pos
, struct page
, mapping
);
2625 /* split_huge_page() removes page from list on success */
2626 if (!split_huge_page(page
))
2632 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2633 list_splice_tail(&list
, &pgdata
->split_queue
);
2634 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2637 * Stop shrinker if we didn't split any page, but the queue is empty.
2638 * This can happen if pages were freed under us.
2640 if (!split
&& list_empty(&pgdata
->split_queue
))
2645 static struct shrinker deferred_split_shrinker
= {
2646 .count_objects
= deferred_split_count
,
2647 .scan_objects
= deferred_split_scan
,
2648 .seeks
= DEFAULT_SEEKS
,
2649 .flags
= SHRINKER_NUMA_AWARE
,
2652 #ifdef CONFIG_DEBUG_FS
2653 static int split_huge_pages_set(void *data
, u64 val
)
2657 unsigned long pfn
, max_zone_pfn
;
2658 unsigned long total
= 0, split
= 0;
2663 for_each_populated_zone(zone
) {
2664 max_zone_pfn
= zone_end_pfn(zone
);
2665 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
2666 if (!pfn_valid(pfn
))
2669 page
= pfn_to_page(pfn
);
2670 if (!get_page_unless_zero(page
))
2673 if (zone
!= page_zone(page
))
2676 if (!PageHead(page
) || PageHuge(page
) || !PageLRU(page
))
2681 if (!split_huge_page(page
))
2689 pr_info("%lu of %lu THP split\n", split
, total
);
2693 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops
, NULL
, split_huge_pages_set
,
2696 static int __init
split_huge_pages_debugfs(void)
2700 ret
= debugfs_create_file("split_huge_pages", 0200, NULL
, NULL
,
2701 &split_huge_pages_fops
);
2703 pr_warn("Failed to create split_huge_pages in debugfs");
2706 late_initcall(split_huge_pages_debugfs
);