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 const 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
, vmf
->address
, 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
);
927 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
928 if (unlikely(is_swap_pmd(pmd
))) {
929 swp_entry_t entry
= pmd_to_swp_entry(pmd
);
931 VM_BUG_ON(!is_pmd_migration_entry(pmd
));
932 if (is_write_migration_entry(entry
)) {
933 make_migration_entry_read(&entry
);
934 pmd
= swp_entry_to_pmd(entry
);
935 if (pmd_swp_soft_dirty(*src_pmd
))
936 pmd
= pmd_swp_mksoft_dirty(pmd
);
937 set_pmd_at(src_mm
, addr
, src_pmd
, pmd
);
939 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
940 atomic_long_inc(&dst_mm
->nr_ptes
);
941 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
942 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
948 if (unlikely(!pmd_trans_huge(pmd
))) {
949 pte_free(dst_mm
, pgtable
);
953 * When page table lock is held, the huge zero pmd should not be
954 * under splitting since we don't split the page itself, only pmd to
957 if (is_huge_zero_pmd(pmd
)) {
958 struct page
*zero_page
;
960 * get_huge_zero_page() will never allocate a new page here,
961 * since we already have a zero page to copy. It just takes a
964 zero_page
= mm_get_huge_zero_page(dst_mm
);
965 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
971 src_page
= pmd_page(pmd
);
972 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
974 page_dup_rmap(src_page
, true);
975 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
976 atomic_long_inc(&dst_mm
->nr_ptes
);
977 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
979 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
980 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
981 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
985 spin_unlock(src_ptl
);
986 spin_unlock(dst_ptl
);
991 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
992 static void touch_pud(struct vm_area_struct
*vma
, unsigned long addr
,
993 pud_t
*pud
, int flags
)
997 _pud
= pud_mkyoung(*pud
);
998 if (flags
& FOLL_WRITE
)
999 _pud
= pud_mkdirty(_pud
);
1000 if (pudp_set_access_flags(vma
, addr
& HPAGE_PUD_MASK
,
1001 pud
, _pud
, flags
& FOLL_WRITE
))
1002 update_mmu_cache_pud(vma
, addr
, pud
);
1005 struct page
*follow_devmap_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1006 pud_t
*pud
, int flags
)
1008 unsigned long pfn
= pud_pfn(*pud
);
1009 struct mm_struct
*mm
= vma
->vm_mm
;
1010 struct dev_pagemap
*pgmap
;
1013 assert_spin_locked(pud_lockptr(mm
, pud
));
1015 if (flags
& FOLL_WRITE
&& !pud_write(*pud
))
1018 if (pud_present(*pud
) && pud_devmap(*pud
))
1023 if (flags
& FOLL_TOUCH
)
1024 touch_pud(vma
, addr
, pud
, flags
);
1027 * device mapped pages can only be returned if the
1028 * caller will manage the page reference count.
1030 if (!(flags
& FOLL_GET
))
1031 return ERR_PTR(-EEXIST
);
1033 pfn
+= (addr
& ~PUD_MASK
) >> PAGE_SHIFT
;
1034 pgmap
= get_dev_pagemap(pfn
, NULL
);
1036 return ERR_PTR(-EFAULT
);
1037 page
= pfn_to_page(pfn
);
1039 put_dev_pagemap(pgmap
);
1044 int copy_huge_pud(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1045 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1046 struct vm_area_struct
*vma
)
1048 spinlock_t
*dst_ptl
, *src_ptl
;
1052 dst_ptl
= pud_lock(dst_mm
, dst_pud
);
1053 src_ptl
= pud_lockptr(src_mm
, src_pud
);
1054 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1058 if (unlikely(!pud_trans_huge(pud
) && !pud_devmap(pud
)))
1062 * When page table lock is held, the huge zero pud should not be
1063 * under splitting since we don't split the page itself, only pud to
1066 if (is_huge_zero_pud(pud
)) {
1067 /* No huge zero pud yet */
1070 pudp_set_wrprotect(src_mm
, addr
, src_pud
);
1071 pud
= pud_mkold(pud_wrprotect(pud
));
1072 set_pud_at(dst_mm
, addr
, dst_pud
, pud
);
1076 spin_unlock(src_ptl
);
1077 spin_unlock(dst_ptl
);
1081 void huge_pud_set_accessed(struct vm_fault
*vmf
, pud_t orig_pud
)
1084 unsigned long haddr
;
1085 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1087 vmf
->ptl
= pud_lock(vmf
->vma
->vm_mm
, vmf
->pud
);
1088 if (unlikely(!pud_same(*vmf
->pud
, orig_pud
)))
1091 entry
= pud_mkyoung(orig_pud
);
1093 entry
= pud_mkdirty(entry
);
1094 haddr
= vmf
->address
& HPAGE_PUD_MASK
;
1095 if (pudp_set_access_flags(vmf
->vma
, haddr
, vmf
->pud
, entry
, write
))
1096 update_mmu_cache_pud(vmf
->vma
, vmf
->address
, vmf
->pud
);
1099 spin_unlock(vmf
->ptl
);
1101 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1103 void huge_pmd_set_accessed(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1106 unsigned long haddr
;
1107 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1109 vmf
->ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1110 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1113 entry
= pmd_mkyoung(orig_pmd
);
1115 entry
= pmd_mkdirty(entry
);
1116 haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1117 if (pmdp_set_access_flags(vmf
->vma
, haddr
, vmf
->pmd
, entry
, write
))
1118 update_mmu_cache_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
);
1121 spin_unlock(vmf
->ptl
);
1124 static int do_huge_pmd_wp_page_fallback(struct vm_fault
*vmf
, pmd_t orig_pmd
,
1127 struct vm_area_struct
*vma
= vmf
->vma
;
1128 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1129 struct mem_cgroup
*memcg
;
1133 struct page
**pages
;
1134 unsigned long mmun_start
; /* For mmu_notifiers */
1135 unsigned long mmun_end
; /* For mmu_notifiers */
1137 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1139 if (unlikely(!pages
)) {
1140 ret
|= VM_FAULT_OOM
;
1144 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1145 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
, vma
,
1146 vmf
->address
, page_to_nid(page
));
1147 if (unlikely(!pages
[i
] ||
1148 mem_cgroup_try_charge(pages
[i
], vma
->vm_mm
,
1149 GFP_KERNEL
, &memcg
, false))) {
1153 memcg
= (void *)page_private(pages
[i
]);
1154 set_page_private(pages
[i
], 0);
1155 mem_cgroup_cancel_charge(pages
[i
], memcg
,
1160 ret
|= VM_FAULT_OOM
;
1163 set_page_private(pages
[i
], (unsigned long)memcg
);
1166 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1167 copy_user_highpage(pages
[i
], page
+ i
,
1168 haddr
+ PAGE_SIZE
* i
, vma
);
1169 __SetPageUptodate(pages
[i
]);
1174 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1175 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1177 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1178 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1179 goto out_free_pages
;
1180 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1182 pmdp_huge_clear_flush_notify(vma
, haddr
, vmf
->pmd
);
1183 /* leave pmd empty until pte is filled */
1185 pgtable
= pgtable_trans_huge_withdraw(vma
->vm_mm
, vmf
->pmd
);
1186 pmd_populate(vma
->vm_mm
, &_pmd
, pgtable
);
1188 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1190 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1191 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1192 memcg
= (void *)page_private(pages
[i
]);
1193 set_page_private(pages
[i
], 0);
1194 page_add_new_anon_rmap(pages
[i
], vmf
->vma
, haddr
, false);
1195 mem_cgroup_commit_charge(pages
[i
], memcg
, false, false);
1196 lru_cache_add_active_or_unevictable(pages
[i
], vma
);
1197 vmf
->pte
= pte_offset_map(&_pmd
, haddr
);
1198 VM_BUG_ON(!pte_none(*vmf
->pte
));
1199 set_pte_at(vma
->vm_mm
, haddr
, vmf
->pte
, entry
);
1200 pte_unmap(vmf
->pte
);
1204 smp_wmb(); /* make pte visible before pmd */
1205 pmd_populate(vma
->vm_mm
, vmf
->pmd
, pgtable
);
1206 page_remove_rmap(page
, true);
1207 spin_unlock(vmf
->ptl
);
1209 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1211 ret
|= VM_FAULT_WRITE
;
1218 spin_unlock(vmf
->ptl
);
1219 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1220 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1221 memcg
= (void *)page_private(pages
[i
]);
1222 set_page_private(pages
[i
], 0);
1223 mem_cgroup_cancel_charge(pages
[i
], memcg
, false);
1230 int do_huge_pmd_wp_page(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1232 struct vm_area_struct
*vma
= vmf
->vma
;
1233 struct page
*page
= NULL
, *new_page
;
1234 struct mem_cgroup
*memcg
;
1235 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1236 unsigned long mmun_start
; /* For mmu_notifiers */
1237 unsigned long mmun_end
; /* For mmu_notifiers */
1238 gfp_t huge_gfp
; /* for allocation and charge */
1241 vmf
->ptl
= pmd_lockptr(vma
->vm_mm
, vmf
->pmd
);
1242 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1243 if (is_huge_zero_pmd(orig_pmd
))
1245 spin_lock(vmf
->ptl
);
1246 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1249 page
= pmd_page(orig_pmd
);
1250 VM_BUG_ON_PAGE(!PageCompound(page
) || !PageHead(page
), page
);
1252 * We can only reuse the page if nobody else maps the huge page or it's
1255 if (!trylock_page(page
)) {
1257 spin_unlock(vmf
->ptl
);
1259 spin_lock(vmf
->ptl
);
1260 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1267 if (reuse_swap_page(page
, NULL
)) {
1269 entry
= pmd_mkyoung(orig_pmd
);
1270 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1271 if (pmdp_set_access_flags(vma
, haddr
, vmf
->pmd
, entry
, 1))
1272 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1273 ret
|= VM_FAULT_WRITE
;
1279 spin_unlock(vmf
->ptl
);
1281 if (transparent_hugepage_enabled(vma
) &&
1282 !transparent_hugepage_debug_cow()) {
1283 huge_gfp
= alloc_hugepage_direct_gfpmask(vma
);
1284 new_page
= alloc_hugepage_vma(huge_gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
1288 if (likely(new_page
)) {
1289 prep_transhuge_page(new_page
);
1292 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1293 ret
|= VM_FAULT_FALLBACK
;
1295 ret
= do_huge_pmd_wp_page_fallback(vmf
, orig_pmd
, page
);
1296 if (ret
& VM_FAULT_OOM
) {
1297 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1298 ret
|= VM_FAULT_FALLBACK
;
1302 count_vm_event(THP_FAULT_FALLBACK
);
1306 if (unlikely(mem_cgroup_try_charge(new_page
, vma
->vm_mm
,
1307 huge_gfp
, &memcg
, true))) {
1309 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1312 ret
|= VM_FAULT_FALLBACK
;
1313 count_vm_event(THP_FAULT_FALLBACK
);
1317 count_vm_event(THP_FAULT_ALLOC
);
1320 clear_huge_page(new_page
, vmf
->address
, HPAGE_PMD_NR
);
1322 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1323 __SetPageUptodate(new_page
);
1326 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1327 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1329 spin_lock(vmf
->ptl
);
1332 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1333 spin_unlock(vmf
->ptl
);
1334 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1339 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1340 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1341 pmdp_huge_clear_flush_notify(vma
, haddr
, vmf
->pmd
);
1342 page_add_new_anon_rmap(new_page
, vma
, haddr
, true);
1343 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1344 lru_cache_add_active_or_unevictable(new_page
, vma
);
1345 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
1346 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1348 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1350 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1351 page_remove_rmap(page
, true);
1354 ret
|= VM_FAULT_WRITE
;
1356 spin_unlock(vmf
->ptl
);
1358 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1362 spin_unlock(vmf
->ptl
);
1367 * FOLL_FORCE can write to even unwritable pmd's, but only
1368 * after we've gone through a COW cycle and they are dirty.
1370 static inline bool can_follow_write_pmd(pmd_t pmd
, unsigned int flags
)
1372 return pmd_write(pmd
) ||
1373 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pmd_dirty(pmd
));
1376 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1381 struct mm_struct
*mm
= vma
->vm_mm
;
1382 struct page
*page
= NULL
;
1384 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1386 if (flags
& FOLL_WRITE
&& !can_follow_write_pmd(*pmd
, flags
))
1389 /* Avoid dumping huge zero page */
1390 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1391 return ERR_PTR(-EFAULT
);
1393 /* Full NUMA hinting faults to serialise migration in fault paths */
1394 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1397 page
= pmd_page(*pmd
);
1398 VM_BUG_ON_PAGE(!PageHead(page
) && !is_zone_device_page(page
), page
);
1399 if (flags
& FOLL_TOUCH
)
1400 touch_pmd(vma
, addr
, pmd
, flags
);
1401 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1403 * We don't mlock() pte-mapped THPs. This way we can avoid
1404 * leaking mlocked pages into non-VM_LOCKED VMAs.
1408 * In most cases the pmd is the only mapping of the page as we
1409 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1410 * writable private mappings in populate_vma_page_range().
1412 * The only scenario when we have the page shared here is if we
1413 * mlocking read-only mapping shared over fork(). We skip
1414 * mlocking such pages.
1418 * We can expect PageDoubleMap() to be stable under page lock:
1419 * for file pages we set it in page_add_file_rmap(), which
1420 * requires page to be locked.
1423 if (PageAnon(page
) && compound_mapcount(page
) != 1)
1425 if (PageDoubleMap(page
) || !page
->mapping
)
1427 if (!trylock_page(page
))
1430 if (page
->mapping
&& !PageDoubleMap(page
))
1431 mlock_vma_page(page
);
1435 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1436 VM_BUG_ON_PAGE(!PageCompound(page
) && !is_zone_device_page(page
), page
);
1437 if (flags
& FOLL_GET
)
1444 /* NUMA hinting page fault entry point for trans huge pmds */
1445 int do_huge_pmd_numa_page(struct vm_fault
*vmf
, pmd_t pmd
)
1447 struct vm_area_struct
*vma
= vmf
->vma
;
1448 struct anon_vma
*anon_vma
= NULL
;
1450 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1451 int page_nid
= -1, this_nid
= numa_node_id();
1452 int target_nid
, last_cpupid
= -1;
1454 bool migrated
= false;
1458 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1459 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
)))
1463 * If there are potential migrations, wait for completion and retry
1464 * without disrupting NUMA hinting information. Do not relock and
1465 * check_same as the page may no longer be mapped.
1467 if (unlikely(pmd_trans_migrating(*vmf
->pmd
))) {
1468 page
= pmd_page(*vmf
->pmd
);
1469 if (!get_page_unless_zero(page
))
1471 spin_unlock(vmf
->ptl
);
1472 wait_on_page_locked(page
);
1477 page
= pmd_page(pmd
);
1478 BUG_ON(is_huge_zero_page(page
));
1479 page_nid
= page_to_nid(page
);
1480 last_cpupid
= page_cpupid_last(page
);
1481 count_vm_numa_event(NUMA_HINT_FAULTS
);
1482 if (page_nid
== this_nid
) {
1483 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1484 flags
|= TNF_FAULT_LOCAL
;
1487 /* See similar comment in do_numa_page for explanation */
1488 if (!pmd_savedwrite(pmd
))
1489 flags
|= TNF_NO_GROUP
;
1492 * Acquire the page lock to serialise THP migrations but avoid dropping
1493 * page_table_lock if at all possible
1495 page_locked
= trylock_page(page
);
1496 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1497 if (target_nid
== -1) {
1498 /* If the page was locked, there are no parallel migrations */
1503 /* Migration could have started since the pmd_trans_migrating check */
1506 if (!get_page_unless_zero(page
))
1508 spin_unlock(vmf
->ptl
);
1509 wait_on_page_locked(page
);
1515 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1516 * to serialises splits
1519 spin_unlock(vmf
->ptl
);
1520 anon_vma
= page_lock_anon_vma_read(page
);
1522 /* Confirm the PMD did not change while page_table_lock was released */
1523 spin_lock(vmf
->ptl
);
1524 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
))) {
1531 /* Bail if we fail to protect against THP splits for any reason */
1532 if (unlikely(!anon_vma
)) {
1539 * Since we took the NUMA fault, we must have observed the !accessible
1540 * bit. Make sure all other CPUs agree with that, to avoid them
1541 * modifying the page we're about to migrate.
1543 * Must be done under PTL such that we'll observe the relevant
1544 * inc_tlb_flush_pending().
1546 * We are not sure a pending tlb flush here is for a huge page
1547 * mapping or not. Hence use the tlb range variant
1549 if (mm_tlb_flush_pending(vma
->vm_mm
))
1550 flush_tlb_range(vma
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
1553 * Migrate the THP to the requested node, returns with page unlocked
1554 * and access rights restored.
1556 spin_unlock(vmf
->ptl
);
1558 migrated
= migrate_misplaced_transhuge_page(vma
->vm_mm
, vma
,
1559 vmf
->pmd
, pmd
, vmf
->address
, page
, target_nid
);
1561 flags
|= TNF_MIGRATED
;
1562 page_nid
= target_nid
;
1564 flags
|= TNF_MIGRATE_FAIL
;
1568 BUG_ON(!PageLocked(page
));
1569 was_writable
= pmd_savedwrite(pmd
);
1570 pmd
= pmd_modify(pmd
, vma
->vm_page_prot
);
1571 pmd
= pmd_mkyoung(pmd
);
1573 pmd
= pmd_mkwrite(pmd
);
1574 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, pmd
);
1575 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1578 spin_unlock(vmf
->ptl
);
1582 page_unlock_anon_vma_read(anon_vma
);
1585 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
,
1592 * Return true if we do MADV_FREE successfully on entire pmd page.
1593 * Otherwise, return false.
1595 bool madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1596 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1601 struct mm_struct
*mm
= tlb
->mm
;
1604 tlb_remove_check_page_size_change(tlb
, HPAGE_PMD_SIZE
);
1606 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1611 if (is_huge_zero_pmd(orig_pmd
))
1614 if (unlikely(!pmd_present(orig_pmd
))) {
1615 VM_BUG_ON(thp_migration_supported() &&
1616 !is_pmd_migration_entry(orig_pmd
));
1620 page
= pmd_page(orig_pmd
);
1622 * If other processes are mapping this page, we couldn't discard
1623 * the page unless they all do MADV_FREE so let's skip the page.
1625 if (page_mapcount(page
) != 1)
1628 if (!trylock_page(page
))
1632 * If user want to discard part-pages of THP, split it so MADV_FREE
1633 * will deactivate only them.
1635 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1638 split_huge_page(page
);
1644 if (PageDirty(page
))
1645 ClearPageDirty(page
);
1648 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1649 pmdp_invalidate(vma
, addr
, pmd
);
1650 orig_pmd
= pmd_mkold(orig_pmd
);
1651 orig_pmd
= pmd_mkclean(orig_pmd
);
1653 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1654 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1657 mark_page_lazyfree(page
);
1665 static inline void zap_deposited_table(struct mm_struct
*mm
, pmd_t
*pmd
)
1669 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1670 pte_free(mm
, pgtable
);
1671 atomic_long_dec(&mm
->nr_ptes
);
1674 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1675 pmd_t
*pmd
, unsigned long addr
)
1680 tlb_remove_check_page_size_change(tlb
, HPAGE_PMD_SIZE
);
1682 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1686 * For architectures like ppc64 we look at deposited pgtable
1687 * when calling pmdp_huge_get_and_clear. So do the
1688 * pgtable_trans_huge_withdraw after finishing pmdp related
1691 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1693 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1694 if (vma_is_dax(vma
)) {
1695 if (arch_needs_pgtable_deposit())
1696 zap_deposited_table(tlb
->mm
, pmd
);
1698 if (is_huge_zero_pmd(orig_pmd
))
1699 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1700 } else if (is_huge_zero_pmd(orig_pmd
)) {
1701 zap_deposited_table(tlb
->mm
, pmd
);
1703 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1705 struct page
*page
= NULL
;
1706 int flush_needed
= 1;
1708 if (pmd_present(orig_pmd
)) {
1709 page
= pmd_page(orig_pmd
);
1710 page_remove_rmap(page
, true);
1711 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1712 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1713 } else if (thp_migration_supported()) {
1716 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd
));
1717 entry
= pmd_to_swp_entry(orig_pmd
);
1718 page
= pfn_to_page(swp_offset(entry
));
1721 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1723 if (PageAnon(page
)) {
1724 zap_deposited_table(tlb
->mm
, pmd
);
1725 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1727 if (arch_needs_pgtable_deposit())
1728 zap_deposited_table(tlb
->mm
, pmd
);
1729 add_mm_counter(tlb
->mm
, MM_FILEPAGES
, -HPAGE_PMD_NR
);
1734 tlb_remove_page_size(tlb
, page
, HPAGE_PMD_SIZE
);
1739 #ifndef pmd_move_must_withdraw
1740 static inline int pmd_move_must_withdraw(spinlock_t
*new_pmd_ptl
,
1741 spinlock_t
*old_pmd_ptl
,
1742 struct vm_area_struct
*vma
)
1745 * With split pmd lock we also need to move preallocated
1746 * PTE page table if new_pmd is on different PMD page table.
1748 * We also don't deposit and withdraw tables for file pages.
1750 return (new_pmd_ptl
!= old_pmd_ptl
) && vma_is_anonymous(vma
);
1754 static pmd_t
move_soft_dirty_pmd(pmd_t pmd
)
1756 #ifdef CONFIG_MEM_SOFT_DIRTY
1757 if (unlikely(is_pmd_migration_entry(pmd
)))
1758 pmd
= pmd_swp_mksoft_dirty(pmd
);
1759 else if (pmd_present(pmd
))
1760 pmd
= pmd_mksoft_dirty(pmd
);
1765 bool move_huge_pmd(struct vm_area_struct
*vma
, unsigned long old_addr
,
1766 unsigned long new_addr
, unsigned long old_end
,
1767 pmd_t
*old_pmd
, pmd_t
*new_pmd
, bool *need_flush
)
1769 spinlock_t
*old_ptl
, *new_ptl
;
1771 struct mm_struct
*mm
= vma
->vm_mm
;
1772 bool force_flush
= false;
1774 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1775 (new_addr
& ~HPAGE_PMD_MASK
) ||
1776 old_end
- old_addr
< HPAGE_PMD_SIZE
)
1780 * The destination pmd shouldn't be established, free_pgtables()
1781 * should have release it.
1783 if (WARN_ON(!pmd_none(*new_pmd
))) {
1784 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1789 * We don't have to worry about the ordering of src and dst
1790 * ptlocks because exclusive mmap_sem prevents deadlock.
1792 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1794 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1795 if (new_ptl
!= old_ptl
)
1796 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1797 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1798 if (pmd_present(pmd
) && pmd_dirty(pmd
))
1800 VM_BUG_ON(!pmd_none(*new_pmd
));
1802 if (pmd_move_must_withdraw(new_ptl
, old_ptl
, vma
)) {
1804 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1805 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1807 pmd
= move_soft_dirty_pmd(pmd
);
1808 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1809 if (new_ptl
!= old_ptl
)
1810 spin_unlock(new_ptl
);
1812 flush_tlb_range(vma
, old_addr
, old_addr
+ PMD_SIZE
);
1815 spin_unlock(old_ptl
);
1823 * - 0 if PMD could not be locked
1824 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1825 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1827 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1828 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1830 struct mm_struct
*mm
= vma
->vm_mm
;
1833 bool preserve_write
;
1836 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1840 preserve_write
= prot_numa
&& pmd_write(*pmd
);
1843 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1844 if (is_swap_pmd(*pmd
)) {
1845 swp_entry_t entry
= pmd_to_swp_entry(*pmd
);
1847 VM_BUG_ON(!is_pmd_migration_entry(*pmd
));
1848 if (is_write_migration_entry(entry
)) {
1851 * A protection check is difficult so
1852 * just be safe and disable write
1854 make_migration_entry_read(&entry
);
1855 newpmd
= swp_entry_to_pmd(entry
);
1856 if (pmd_swp_soft_dirty(*pmd
))
1857 newpmd
= pmd_swp_mksoft_dirty(newpmd
);
1858 set_pmd_at(mm
, addr
, pmd
, newpmd
);
1865 * Avoid trapping faults against the zero page. The read-only
1866 * data is likely to be read-cached on the local CPU and
1867 * local/remote hits to the zero page are not interesting.
1869 if (prot_numa
&& is_huge_zero_pmd(*pmd
))
1872 if (prot_numa
&& pmd_protnone(*pmd
))
1876 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1877 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1878 * which is also under down_read(mmap_sem):
1881 * change_huge_pmd(prot_numa=1)
1882 * pmdp_huge_get_and_clear_notify()
1883 * madvise_dontneed()
1885 * pmd_trans_huge(*pmd) == 0 (without ptl)
1888 * // pmd is re-established
1890 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1891 * which may break userspace.
1893 * pmdp_invalidate() is required to make sure we don't miss
1894 * dirty/young flags set by hardware.
1897 pmdp_invalidate(vma
, addr
, pmd
);
1900 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1903 if (pmd_dirty(*pmd
))
1904 entry
= pmd_mkdirty(entry
);
1905 if (pmd_young(*pmd
))
1906 entry
= pmd_mkyoung(entry
);
1908 entry
= pmd_modify(entry
, newprot
);
1910 entry
= pmd_mk_savedwrite(entry
);
1912 set_pmd_at(mm
, addr
, pmd
, entry
);
1913 BUG_ON(vma_is_anonymous(vma
) && !preserve_write
&& pmd_write(entry
));
1920 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1922 * Note that if it returns page table lock pointer, this routine returns without
1923 * unlocking page table lock. So callers must unlock it.
1925 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1928 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1929 if (likely(is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) ||
1937 * Returns true if a given pud maps a thp, false otherwise.
1939 * Note that if it returns true, this routine returns without unlocking page
1940 * table lock. So callers must unlock it.
1942 spinlock_t
*__pud_trans_huge_lock(pud_t
*pud
, struct vm_area_struct
*vma
)
1946 ptl
= pud_lock(vma
->vm_mm
, pud
);
1947 if (likely(pud_trans_huge(*pud
) || pud_devmap(*pud
)))
1953 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1954 int zap_huge_pud(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1955 pud_t
*pud
, unsigned long addr
)
1960 ptl
= __pud_trans_huge_lock(pud
, vma
);
1964 * For architectures like ppc64 we look at deposited pgtable
1965 * when calling pudp_huge_get_and_clear. So do the
1966 * pgtable_trans_huge_withdraw after finishing pudp related
1969 orig_pud
= pudp_huge_get_and_clear_full(tlb
->mm
, addr
, pud
,
1971 tlb_remove_pud_tlb_entry(tlb
, pud
, addr
);
1972 if (vma_is_dax(vma
)) {
1974 /* No zero page support yet */
1976 /* No support for anonymous PUD pages yet */
1982 static void __split_huge_pud_locked(struct vm_area_struct
*vma
, pud_t
*pud
,
1983 unsigned long haddr
)
1985 VM_BUG_ON(haddr
& ~HPAGE_PUD_MASK
);
1986 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
1987 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PUD_SIZE
, vma
);
1988 VM_BUG_ON(!pud_trans_huge(*pud
) && !pud_devmap(*pud
));
1990 count_vm_event(THP_SPLIT_PUD
);
1992 pudp_huge_clear_flush_notify(vma
, haddr
, pud
);
1995 void __split_huge_pud(struct vm_area_struct
*vma
, pud_t
*pud
,
1996 unsigned long address
)
1999 struct mm_struct
*mm
= vma
->vm_mm
;
2000 unsigned long haddr
= address
& HPAGE_PUD_MASK
;
2002 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PUD_SIZE
);
2003 ptl
= pud_lock(mm
, pud
);
2004 if (unlikely(!pud_trans_huge(*pud
) && !pud_devmap(*pud
)))
2006 __split_huge_pud_locked(vma
, pud
, haddr
);
2010 mmu_notifier_invalidate_range_end(mm
, haddr
, haddr
+ HPAGE_PUD_SIZE
);
2012 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2014 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2015 unsigned long haddr
, pmd_t
*pmd
)
2017 struct mm_struct
*mm
= vma
->vm_mm
;
2022 /* leave pmd empty until pte is filled */
2023 pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2025 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2026 pmd_populate(mm
, &_pmd
, pgtable
);
2028 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2030 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2031 entry
= pte_mkspecial(entry
);
2032 pte
= pte_offset_map(&_pmd
, haddr
);
2033 VM_BUG_ON(!pte_none(*pte
));
2034 set_pte_at(mm
, haddr
, pte
, entry
);
2037 smp_wmb(); /* make pte visible before pmd */
2038 pmd_populate(mm
, pmd
, pgtable
);
2041 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2042 unsigned long haddr
, bool freeze
)
2044 struct mm_struct
*mm
= vma
->vm_mm
;
2048 bool young
, write
, dirty
, soft_dirty
, pmd_migration
= false;
2052 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
2053 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2054 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
2055 VM_BUG_ON(!is_pmd_migration_entry(*pmd
) && !pmd_trans_huge(*pmd
)
2056 && !pmd_devmap(*pmd
));
2058 count_vm_event(THP_SPLIT_PMD
);
2060 if (!vma_is_anonymous(vma
)) {
2061 _pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2063 * We are going to unmap this huge page. So
2064 * just go ahead and zap it
2066 if (arch_needs_pgtable_deposit())
2067 zap_deposited_table(mm
, pmd
);
2068 if (vma_is_dax(vma
))
2070 page
= pmd_page(_pmd
);
2071 if (!PageReferenced(page
) && pmd_young(_pmd
))
2072 SetPageReferenced(page
);
2073 page_remove_rmap(page
, true);
2075 add_mm_counter(mm
, MM_FILEPAGES
, -HPAGE_PMD_NR
);
2077 } else if (is_huge_zero_pmd(*pmd
)) {
2078 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2081 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2082 pmd_migration
= is_pmd_migration_entry(*pmd
);
2083 if (pmd_migration
) {
2086 entry
= pmd_to_swp_entry(*pmd
);
2087 page
= pfn_to_page(swp_offset(entry
));
2090 page
= pmd_page(*pmd
);
2091 VM_BUG_ON_PAGE(!page_count(page
), page
);
2092 page_ref_add(page
, HPAGE_PMD_NR
- 1);
2093 write
= pmd_write(*pmd
);
2094 young
= pmd_young(*pmd
);
2095 dirty
= pmd_dirty(*pmd
);
2096 soft_dirty
= pmd_soft_dirty(*pmd
);
2098 pmdp_huge_split_prepare(vma
, haddr
, pmd
);
2099 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2100 pmd_populate(mm
, &_pmd
, pgtable
);
2102 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
2105 * Note that NUMA hinting access restrictions are not
2106 * transferred to avoid any possibility of altering
2107 * permissions across VMAs.
2109 if (freeze
|| pmd_migration
) {
2110 swp_entry_t swp_entry
;
2111 swp_entry
= make_migration_entry(page
+ i
, write
);
2112 entry
= swp_entry_to_pte(swp_entry
);
2114 entry
= pte_swp_mksoft_dirty(entry
);
2116 entry
= mk_pte(page
+ i
, READ_ONCE(vma
->vm_page_prot
));
2117 entry
= maybe_mkwrite(entry
, vma
);
2119 entry
= pte_wrprotect(entry
);
2121 entry
= pte_mkold(entry
);
2123 entry
= pte_mksoft_dirty(entry
);
2126 SetPageDirty(page
+ i
);
2127 pte
= pte_offset_map(&_pmd
, addr
);
2128 BUG_ON(!pte_none(*pte
));
2129 set_pte_at(mm
, addr
, pte
, entry
);
2130 atomic_inc(&page
[i
]._mapcount
);
2135 * Set PG_double_map before dropping compound_mapcount to avoid
2136 * false-negative page_mapped().
2138 if (compound_mapcount(page
) > 1 && !TestSetPageDoubleMap(page
)) {
2139 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2140 atomic_inc(&page
[i
]._mapcount
);
2143 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
2144 /* Last compound_mapcount is gone. */
2145 __dec_node_page_state(page
, NR_ANON_THPS
);
2146 if (TestClearPageDoubleMap(page
)) {
2147 /* No need in mapcount reference anymore */
2148 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2149 atomic_dec(&page
[i
]._mapcount
);
2153 smp_wmb(); /* make pte visible before pmd */
2155 * Up to this point the pmd is present and huge and userland has the
2156 * whole access to the hugepage during the split (which happens in
2157 * place). If we overwrite the pmd with the not-huge version pointing
2158 * to the pte here (which of course we could if all CPUs were bug
2159 * free), userland could trigger a small page size TLB miss on the
2160 * small sized TLB while the hugepage TLB entry is still established in
2161 * the huge TLB. Some CPU doesn't like that.
2162 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2163 * 383 on page 93. Intel should be safe but is also warns that it's
2164 * only safe if the permission and cache attributes of the two entries
2165 * loaded in the two TLB is identical (which should be the case here).
2166 * But it is generally safer to never allow small and huge TLB entries
2167 * for the same virtual address to be loaded simultaneously. So instead
2168 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2169 * current pmd notpresent (atomically because here the pmd_trans_huge
2170 * and pmd_trans_splitting must remain set at all times on the pmd
2171 * until the split is complete for this pmd), then we flush the SMP TLB
2172 * and finally we write the non-huge version of the pmd entry with
2175 pmdp_invalidate(vma
, haddr
, pmd
);
2176 pmd_populate(mm
, pmd
, pgtable
);
2179 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2180 page_remove_rmap(page
+ i
, false);
2186 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2187 unsigned long address
, bool freeze
, struct page
*page
)
2190 struct mm_struct
*mm
= vma
->vm_mm
;
2191 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2193 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2194 ptl
= pmd_lock(mm
, pmd
);
2197 * If caller asks to setup a migration entries, we need a page to check
2198 * pmd against. Otherwise we can end up replacing wrong page.
2200 VM_BUG_ON(freeze
&& !page
);
2201 if (page
&& page
!= pmd_page(*pmd
))
2204 if (pmd_trans_huge(*pmd
)) {
2205 page
= pmd_page(*pmd
);
2206 if (PageMlocked(page
))
2207 clear_page_mlock(page
);
2208 } else if (!(pmd_devmap(*pmd
) || is_pmd_migration_entry(*pmd
)))
2210 __split_huge_pmd_locked(vma
, pmd
, haddr
, freeze
);
2213 mmu_notifier_invalidate_range_end(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2216 void split_huge_pmd_address(struct vm_area_struct
*vma
, unsigned long address
,
2217 bool freeze
, struct page
*page
)
2224 pgd
= pgd_offset(vma
->vm_mm
, address
);
2225 if (!pgd_present(*pgd
))
2228 p4d
= p4d_offset(pgd
, address
);
2229 if (!p4d_present(*p4d
))
2232 pud
= pud_offset(p4d
, address
);
2233 if (!pud_present(*pud
))
2236 pmd
= pmd_offset(pud
, address
);
2238 __split_huge_pmd(vma
, pmd
, address
, freeze
, page
);
2241 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2242 unsigned long start
,
2247 * If the new start address isn't hpage aligned and it could
2248 * previously contain an hugepage: check if we need to split
2251 if (start
& ~HPAGE_PMD_MASK
&&
2252 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2253 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2254 split_huge_pmd_address(vma
, start
, false, NULL
);
2257 * If the new end address isn't hpage aligned and it could
2258 * previously contain an hugepage: check if we need to split
2261 if (end
& ~HPAGE_PMD_MASK
&&
2262 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2263 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2264 split_huge_pmd_address(vma
, end
, false, NULL
);
2267 * If we're also updating the vma->vm_next->vm_start, if the new
2268 * vm_next->vm_start isn't page aligned and it could previously
2269 * contain an hugepage: check if we need to split an huge pmd.
2271 if (adjust_next
> 0) {
2272 struct vm_area_struct
*next
= vma
->vm_next
;
2273 unsigned long nstart
= next
->vm_start
;
2274 nstart
+= adjust_next
<< PAGE_SHIFT
;
2275 if (nstart
& ~HPAGE_PMD_MASK
&&
2276 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2277 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2278 split_huge_pmd_address(next
, nstart
, false, NULL
);
2282 static void freeze_page(struct page
*page
)
2284 enum ttu_flags ttu_flags
= TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
|
2285 TTU_RMAP_LOCKED
| TTU_SPLIT_HUGE_PMD
;
2288 VM_BUG_ON_PAGE(!PageHead(page
), page
);
2291 ttu_flags
|= TTU_SPLIT_FREEZE
;
2293 unmap_success
= try_to_unmap(page
, ttu_flags
);
2294 VM_BUG_ON_PAGE(!unmap_success
, page
);
2297 static void unfreeze_page(struct page
*page
)
2300 if (PageTransHuge(page
)) {
2301 remove_migration_ptes(page
, page
, true);
2303 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2304 remove_migration_ptes(page
+ i
, page
+ i
, true);
2308 static void __split_huge_page_tail(struct page
*head
, int tail
,
2309 struct lruvec
*lruvec
, struct list_head
*list
)
2311 struct page
*page_tail
= head
+ tail
;
2313 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
2314 VM_BUG_ON_PAGE(page_ref_count(page_tail
) != 0, page_tail
);
2317 * tail_page->_refcount is zero and not changing from under us. But
2318 * get_page_unless_zero() may be running from under us on the
2319 * tail_page. If we used atomic_set() below instead of atomic_inc() or
2320 * atomic_add(), we would then run atomic_set() concurrently with
2321 * get_page_unless_zero(), and atomic_set() is implemented in C not
2322 * using locked ops. spin_unlock on x86 sometime uses locked ops
2323 * because of PPro errata 66, 92, so unless somebody can guarantee
2324 * atomic_set() here would be safe on all archs (and not only on x86),
2325 * it's safer to use atomic_inc()/atomic_add().
2327 if (PageAnon(head
) && !PageSwapCache(head
)) {
2328 page_ref_inc(page_tail
);
2330 /* Additional pin to radix tree */
2331 page_ref_add(page_tail
, 2);
2334 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
2335 page_tail
->flags
|= (head
->flags
&
2336 ((1L << PG_referenced
) |
2337 (1L << PG_swapbacked
) |
2338 (1L << PG_swapcache
) |
2339 (1L << PG_mlocked
) |
2340 (1L << PG_uptodate
) |
2343 (1L << PG_unevictable
) |
2347 * After clearing PageTail the gup refcount can be released.
2348 * Page flags also must be visible before we make the page non-compound.
2352 clear_compound_head(page_tail
);
2354 if (page_is_young(head
))
2355 set_page_young(page_tail
);
2356 if (page_is_idle(head
))
2357 set_page_idle(page_tail
);
2359 /* ->mapping in first tail page is compound_mapcount */
2360 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
2362 page_tail
->mapping
= head
->mapping
;
2364 page_tail
->index
= head
->index
+ tail
;
2365 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
2366 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
2369 static void __split_huge_page(struct page
*page
, struct list_head
*list
,
2370 unsigned long flags
)
2372 struct page
*head
= compound_head(page
);
2373 struct zone
*zone
= page_zone(head
);
2374 struct lruvec
*lruvec
;
2378 lruvec
= mem_cgroup_page_lruvec(head
, zone
->zone_pgdat
);
2380 /* complete memcg works before add pages to LRU */
2381 mem_cgroup_split_huge_fixup(head
);
2383 if (!PageAnon(page
))
2384 end
= DIV_ROUND_UP(i_size_read(head
->mapping
->host
), PAGE_SIZE
);
2386 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
2387 __split_huge_page_tail(head
, i
, lruvec
, list
);
2388 /* Some pages can be beyond i_size: drop them from page cache */
2389 if (head
[i
].index
>= end
) {
2390 __ClearPageDirty(head
+ i
);
2391 __delete_from_page_cache(head
+ i
, NULL
);
2392 if (IS_ENABLED(CONFIG_SHMEM
) && PageSwapBacked(head
))
2393 shmem_uncharge(head
->mapping
->host
, 1);
2398 ClearPageCompound(head
);
2399 /* See comment in __split_huge_page_tail() */
2400 if (PageAnon(head
)) {
2401 /* Additional pin to radix tree of swap cache */
2402 if (PageSwapCache(head
))
2403 page_ref_add(head
, 2);
2407 /* Additional pin to radix tree */
2408 page_ref_add(head
, 2);
2409 spin_unlock(&head
->mapping
->tree_lock
);
2412 spin_unlock_irqrestore(zone_lru_lock(page_zone(head
)), flags
);
2414 unfreeze_page(head
);
2416 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2417 struct page
*subpage
= head
+ i
;
2418 if (subpage
== page
)
2420 unlock_page(subpage
);
2423 * Subpages may be freed if there wasn't any mapping
2424 * like if add_to_swap() is running on a lru page that
2425 * had its mapping zapped. And freeing these pages
2426 * requires taking the lru_lock so we do the put_page
2427 * of the tail pages after the split is complete.
2433 int total_mapcount(struct page
*page
)
2435 int i
, compound
, ret
;
2437 VM_BUG_ON_PAGE(PageTail(page
), page
);
2439 if (likely(!PageCompound(page
)))
2440 return atomic_read(&page
->_mapcount
) + 1;
2442 compound
= compound_mapcount(page
);
2446 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2447 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
2448 /* File pages has compound_mapcount included in _mapcount */
2449 if (!PageAnon(page
))
2450 return ret
- compound
* HPAGE_PMD_NR
;
2451 if (PageDoubleMap(page
))
2452 ret
-= HPAGE_PMD_NR
;
2457 * This calculates accurately how many mappings a transparent hugepage
2458 * has (unlike page_mapcount() which isn't fully accurate). This full
2459 * accuracy is primarily needed to know if copy-on-write faults can
2460 * reuse the page and change the mapping to read-write instead of
2461 * copying them. At the same time this returns the total_mapcount too.
2463 * The function returns the highest mapcount any one of the subpages
2464 * has. If the return value is one, even if different processes are
2465 * mapping different subpages of the transparent hugepage, they can
2466 * all reuse it, because each process is reusing a different subpage.
2468 * The total_mapcount is instead counting all virtual mappings of the
2469 * subpages. If the total_mapcount is equal to "one", it tells the
2470 * caller all mappings belong to the same "mm" and in turn the
2471 * anon_vma of the transparent hugepage can become the vma->anon_vma
2472 * local one as no other process may be mapping any of the subpages.
2474 * It would be more accurate to replace page_mapcount() with
2475 * page_trans_huge_mapcount(), however we only use
2476 * page_trans_huge_mapcount() in the copy-on-write faults where we
2477 * need full accuracy to avoid breaking page pinning, because
2478 * page_trans_huge_mapcount() is slower than page_mapcount().
2480 int page_trans_huge_mapcount(struct page
*page
, int *total_mapcount
)
2482 int i
, ret
, _total_mapcount
, mapcount
;
2484 /* hugetlbfs shouldn't call it */
2485 VM_BUG_ON_PAGE(PageHuge(page
), page
);
2487 if (likely(!PageTransCompound(page
))) {
2488 mapcount
= atomic_read(&page
->_mapcount
) + 1;
2490 *total_mapcount
= mapcount
;
2494 page
= compound_head(page
);
2496 _total_mapcount
= ret
= 0;
2497 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2498 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
2499 ret
= max(ret
, mapcount
);
2500 _total_mapcount
+= mapcount
;
2502 if (PageDoubleMap(page
)) {
2504 _total_mapcount
-= HPAGE_PMD_NR
;
2506 mapcount
= compound_mapcount(page
);
2508 _total_mapcount
+= mapcount
;
2510 *total_mapcount
= _total_mapcount
;
2514 /* Racy check whether the huge page can be split */
2515 bool can_split_huge_page(struct page
*page
, int *pextra_pins
)
2519 /* Additional pins from radix tree */
2521 extra_pins
= PageSwapCache(page
) ? HPAGE_PMD_NR
: 0;
2523 extra_pins
= HPAGE_PMD_NR
;
2525 *pextra_pins
= extra_pins
;
2526 return total_mapcount(page
) == page_count(page
) - extra_pins
- 1;
2530 * This function splits huge page into normal pages. @page can point to any
2531 * subpage of huge page to split. Split doesn't change the position of @page.
2533 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2534 * The huge page must be locked.
2536 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2538 * Both head page and tail pages will inherit mapping, flags, and so on from
2541 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2542 * they are not mapped.
2544 * Returns 0 if the hugepage is split successfully.
2545 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2548 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
2550 struct page
*head
= compound_head(page
);
2551 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(head
));
2552 struct anon_vma
*anon_vma
= NULL
;
2553 struct address_space
*mapping
= NULL
;
2554 int count
, mapcount
, extra_pins
, ret
;
2556 unsigned long flags
;
2558 VM_BUG_ON_PAGE(is_huge_zero_page(page
), page
);
2559 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2560 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
2562 if (PageWriteback(page
))
2565 if (PageAnon(head
)) {
2567 * The caller does not necessarily hold an mmap_sem that would
2568 * prevent the anon_vma disappearing so we first we take a
2569 * reference to it and then lock the anon_vma for write. This
2570 * is similar to page_lock_anon_vma_read except the write lock
2571 * is taken to serialise against parallel split or collapse
2574 anon_vma
= page_get_anon_vma(head
);
2580 anon_vma_lock_write(anon_vma
);
2582 mapping
= head
->mapping
;
2591 i_mmap_lock_read(mapping
);
2595 * Racy check if we can split the page, before freeze_page() will
2598 if (!can_split_huge_page(head
, &extra_pins
)) {
2603 mlocked
= PageMlocked(page
);
2605 VM_BUG_ON_PAGE(compound_mapcount(head
), head
);
2607 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2611 /* prevent PageLRU to go away from under us, and freeze lru stats */
2612 spin_lock_irqsave(zone_lru_lock(page_zone(head
)), flags
);
2617 spin_lock(&mapping
->tree_lock
);
2618 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
2621 * Check if the head page is present in radix tree.
2622 * We assume all tail are present too, if head is there.
2624 if (radix_tree_deref_slot_protected(pslot
,
2625 &mapping
->tree_lock
) != head
)
2629 /* Prevent deferred_split_scan() touching ->_refcount */
2630 spin_lock(&pgdata
->split_queue_lock
);
2631 count
= page_count(head
);
2632 mapcount
= total_mapcount(head
);
2633 if (!mapcount
&& page_ref_freeze(head
, 1 + extra_pins
)) {
2634 if (!list_empty(page_deferred_list(head
))) {
2635 pgdata
->split_queue_len
--;
2636 list_del(page_deferred_list(head
));
2639 __dec_node_page_state(page
, NR_SHMEM_THPS
);
2640 spin_unlock(&pgdata
->split_queue_lock
);
2641 __split_huge_page(page
, list
, flags
);
2642 if (PageSwapCache(head
)) {
2643 swp_entry_t entry
= { .val
= page_private(head
) };
2645 ret
= split_swap_cluster(entry
);
2649 if (IS_ENABLED(CONFIG_DEBUG_VM
) && mapcount
) {
2650 pr_alert("total_mapcount: %u, page_count(): %u\n",
2653 dump_page(head
, NULL
);
2654 dump_page(page
, "total_mapcount(head) > 0");
2657 spin_unlock(&pgdata
->split_queue_lock
);
2659 spin_unlock(&mapping
->tree_lock
);
2660 spin_unlock_irqrestore(zone_lru_lock(page_zone(head
)), flags
);
2661 unfreeze_page(head
);
2667 anon_vma_unlock_write(anon_vma
);
2668 put_anon_vma(anon_vma
);
2671 i_mmap_unlock_read(mapping
);
2673 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
2677 void free_transhuge_page(struct page
*page
)
2679 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
2680 unsigned long flags
;
2682 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2683 if (!list_empty(page_deferred_list(page
))) {
2684 pgdata
->split_queue_len
--;
2685 list_del(page_deferred_list(page
));
2687 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2688 free_compound_page(page
);
2691 void deferred_split_huge_page(struct page
*page
)
2693 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
2694 unsigned long flags
;
2696 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
2698 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2699 if (list_empty(page_deferred_list(page
))) {
2700 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
2701 list_add_tail(page_deferred_list(page
), &pgdata
->split_queue
);
2702 pgdata
->split_queue_len
++;
2704 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2707 static unsigned long deferred_split_count(struct shrinker
*shrink
,
2708 struct shrink_control
*sc
)
2710 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2711 return ACCESS_ONCE(pgdata
->split_queue_len
);
2714 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
2715 struct shrink_control
*sc
)
2717 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2718 unsigned long flags
;
2719 LIST_HEAD(list
), *pos
, *next
;
2723 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2724 /* Take pin on all head pages to avoid freeing them under us */
2725 list_for_each_safe(pos
, next
, &pgdata
->split_queue
) {
2726 page
= list_entry((void *)pos
, struct page
, mapping
);
2727 page
= compound_head(page
);
2728 if (get_page_unless_zero(page
)) {
2729 list_move(page_deferred_list(page
), &list
);
2731 /* We lost race with put_compound_page() */
2732 list_del_init(page_deferred_list(page
));
2733 pgdata
->split_queue_len
--;
2735 if (!--sc
->nr_to_scan
)
2738 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2740 list_for_each_safe(pos
, next
, &list
) {
2741 page
= list_entry((void *)pos
, struct page
, mapping
);
2743 /* split_huge_page() removes page from list on success */
2744 if (!split_huge_page(page
))
2750 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2751 list_splice_tail(&list
, &pgdata
->split_queue
);
2752 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2755 * Stop shrinker if we didn't split any page, but the queue is empty.
2756 * This can happen if pages were freed under us.
2758 if (!split
&& list_empty(&pgdata
->split_queue
))
2763 static struct shrinker deferred_split_shrinker
= {
2764 .count_objects
= deferred_split_count
,
2765 .scan_objects
= deferred_split_scan
,
2766 .seeks
= DEFAULT_SEEKS
,
2767 .flags
= SHRINKER_NUMA_AWARE
,
2770 #ifdef CONFIG_DEBUG_FS
2771 static int split_huge_pages_set(void *data
, u64 val
)
2775 unsigned long pfn
, max_zone_pfn
;
2776 unsigned long total
= 0, split
= 0;
2781 for_each_populated_zone(zone
) {
2782 max_zone_pfn
= zone_end_pfn(zone
);
2783 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
2784 if (!pfn_valid(pfn
))
2787 page
= pfn_to_page(pfn
);
2788 if (!get_page_unless_zero(page
))
2791 if (zone
!= page_zone(page
))
2794 if (!PageHead(page
) || PageHuge(page
) || !PageLRU(page
))
2799 if (!split_huge_page(page
))
2807 pr_info("%lu of %lu THP split\n", split
, total
);
2811 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops
, NULL
, split_huge_pages_set
,
2814 static int __init
split_huge_pages_debugfs(void)
2818 ret
= debugfs_create_file("split_huge_pages", 0200, NULL
, NULL
,
2819 &split_huge_pages_fops
);
2821 pr_warn("Failed to create split_huge_pages in debugfs");
2824 late_initcall(split_huge_pages_debugfs
);
2827 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2828 void set_pmd_migration_entry(struct page_vma_mapped_walk
*pvmw
,
2831 struct vm_area_struct
*vma
= pvmw
->vma
;
2832 struct mm_struct
*mm
= vma
->vm_mm
;
2833 unsigned long address
= pvmw
->address
;
2838 if (!(pvmw
->pmd
&& !pvmw
->pte
))
2841 mmu_notifier_invalidate_range_start(mm
, address
,
2842 address
+ HPAGE_PMD_SIZE
);
2844 flush_cache_range(vma
, address
, address
+ HPAGE_PMD_SIZE
);
2845 pmdval
= *pvmw
->pmd
;
2846 pmdp_invalidate(vma
, address
, pvmw
->pmd
);
2847 if (pmd_dirty(pmdval
))
2848 set_page_dirty(page
);
2849 entry
= make_migration_entry(page
, pmd_write(pmdval
));
2850 pmdswp
= swp_entry_to_pmd(entry
);
2851 if (pmd_soft_dirty(pmdval
))
2852 pmdswp
= pmd_swp_mksoft_dirty(pmdswp
);
2853 set_pmd_at(mm
, address
, pvmw
->pmd
, pmdswp
);
2854 page_remove_rmap(page
, true);
2857 mmu_notifier_invalidate_range_end(mm
, address
,
2858 address
+ HPAGE_PMD_SIZE
);
2861 void remove_migration_pmd(struct page_vma_mapped_walk
*pvmw
, struct page
*new)
2863 struct vm_area_struct
*vma
= pvmw
->vma
;
2864 struct mm_struct
*mm
= vma
->vm_mm
;
2865 unsigned long address
= pvmw
->address
;
2866 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
2870 if (!(pvmw
->pmd
&& !pvmw
->pte
))
2873 entry
= pmd_to_swp_entry(*pvmw
->pmd
);
2875 pmde
= pmd_mkold(mk_huge_pmd(new, vma
->vm_page_prot
));
2876 if (pmd_swp_soft_dirty(*pvmw
->pmd
))
2877 pmde
= pmd_mksoft_dirty(pmde
);
2878 if (is_write_migration_entry(entry
))
2879 pmde
= maybe_pmd_mkwrite(pmde
, vma
);
2881 flush_cache_range(vma
, mmun_start
, mmun_start
+ HPAGE_PMD_SIZE
);
2882 page_add_anon_rmap(new, vma
, mmun_start
, true);
2883 set_pmd_at(mm
, mmun_start
, pvmw
->pmd
, pmde
);
2884 if (vma
->vm_flags
& VM_LOCKED
)
2885 mlock_vma_page(new);
2886 update_mmu_cache_pmd(vma
, address
, pvmw
->pmd
);