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>
36 #include <linux/page_owner.h>
39 #include <asm/pgalloc.h>
43 * By default, transparent hugepage support is disabled in order to avoid
44 * risking an increased memory footprint for applications that are not
45 * guaranteed to benefit from it. When transparent hugepage support is
46 * enabled, it is for all mappings, and khugepaged scans all mappings.
47 * Defrag is invoked by khugepaged hugepage allocations and by page faults
48 * for all hugepage allocations.
50 unsigned long transparent_hugepage_flags __read_mostly
=
51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
52 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
55 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
)|
58 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
59 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
61 static struct shrinker deferred_split_shrinker
;
63 static atomic_t huge_zero_refcount
;
64 struct page
*huge_zero_page __read_mostly
;
66 static struct page
*get_huge_zero_page(void)
68 struct page
*zero_page
;
70 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
71 return READ_ONCE(huge_zero_page
);
73 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
76 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
79 count_vm_event(THP_ZERO_PAGE_ALLOC
);
81 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
83 __free_pages(zero_page
, compound_order(zero_page
));
87 /* We take additional reference here. It will be put back by shrinker */
88 atomic_set(&huge_zero_refcount
, 2);
90 return READ_ONCE(huge_zero_page
);
93 static void put_huge_zero_page(void)
96 * Counter should never go to zero here. Only shrinker can put
99 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
102 struct page
*mm_get_huge_zero_page(struct mm_struct
*mm
)
104 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
105 return READ_ONCE(huge_zero_page
);
107 if (!get_huge_zero_page())
110 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
111 put_huge_zero_page();
113 return READ_ONCE(huge_zero_page
);
116 void mm_put_huge_zero_page(struct mm_struct
*mm
)
118 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
119 put_huge_zero_page();
122 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
123 struct shrink_control
*sc
)
125 /* we can free zero page only if last reference remains */
126 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
129 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
130 struct shrink_control
*sc
)
132 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
133 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
134 BUG_ON(zero_page
== NULL
);
135 __free_pages(zero_page
, compound_order(zero_page
));
142 static struct shrinker huge_zero_page_shrinker
= {
143 .count_objects
= shrink_huge_zero_page_count
,
144 .scan_objects
= shrink_huge_zero_page_scan
,
145 .seeks
= DEFAULT_SEEKS
,
149 static ssize_t
enabled_show(struct kobject
*kobj
,
150 struct kobj_attribute
*attr
, char *buf
)
152 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
))
153 return sprintf(buf
, "[always] madvise never\n");
154 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
155 return sprintf(buf
, "always [madvise] never\n");
157 return sprintf(buf
, "always madvise [never]\n");
160 static ssize_t
enabled_store(struct kobject
*kobj
,
161 struct kobj_attribute
*attr
,
162 const char *buf
, size_t count
)
166 if (!memcmp("always", buf
,
167 min(sizeof("always")-1, count
))) {
168 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
169 set_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
170 } else if (!memcmp("madvise", buf
,
171 min(sizeof("madvise")-1, count
))) {
172 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
173 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
174 } else if (!memcmp("never", buf
,
175 min(sizeof("never")-1, count
))) {
176 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
177 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
182 int err
= start_stop_khugepaged();
188 static struct kobj_attribute enabled_attr
=
189 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
191 ssize_t
single_hugepage_flag_show(struct kobject
*kobj
,
192 struct kobj_attribute
*attr
, char *buf
,
193 enum transparent_hugepage_flag flag
)
195 return sprintf(buf
, "%d\n",
196 !!test_bit(flag
, &transparent_hugepage_flags
));
199 ssize_t
single_hugepage_flag_store(struct kobject
*kobj
,
200 struct kobj_attribute
*attr
,
201 const char *buf
, size_t count
,
202 enum transparent_hugepage_flag flag
)
207 ret
= kstrtoul(buf
, 10, &value
);
214 set_bit(flag
, &transparent_hugepage_flags
);
216 clear_bit(flag
, &transparent_hugepage_flags
);
221 static ssize_t
defrag_show(struct kobject
*kobj
,
222 struct kobj_attribute
*attr
, char *buf
)
224 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
225 return sprintf(buf
, "[always] defer defer+madvise madvise never\n");
226 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
227 return sprintf(buf
, "always [defer] defer+madvise madvise never\n");
228 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
229 return sprintf(buf
, "always defer [defer+madvise] madvise never\n");
230 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
231 return sprintf(buf
, "always defer defer+madvise [madvise] never\n");
232 return sprintf(buf
, "always defer defer+madvise madvise [never]\n");
235 static ssize_t
defrag_store(struct kobject
*kobj
,
236 struct kobj_attribute
*attr
,
237 const char *buf
, size_t count
)
239 if (!memcmp("always", buf
,
240 min(sizeof("always")-1, count
))) {
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
243 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
244 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
245 } else if (!memcmp("defer+madvise", buf
,
246 min(sizeof("defer+madvise")-1, count
))) {
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
249 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
250 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
251 } else if (!memcmp("defer", buf
,
252 min(sizeof("defer")-1, count
))) {
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
255 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
256 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
257 } else if (!memcmp("madvise", buf
,
258 min(sizeof("madvise")-1, count
))) {
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
262 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
263 } else if (!memcmp("never", buf
,
264 min(sizeof("never")-1, count
))) {
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
274 static struct kobj_attribute defrag_attr
=
275 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
277 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
278 struct kobj_attribute
*attr
, char *buf
)
280 return single_hugepage_flag_show(kobj
, attr
, buf
,
281 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
283 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
284 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
286 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
287 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
289 static struct kobj_attribute use_zero_page_attr
=
290 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
292 static ssize_t
hpage_pmd_size_show(struct kobject
*kobj
,
293 struct kobj_attribute
*attr
, char *buf
)
295 return sprintf(buf
, "%lu\n", HPAGE_PMD_SIZE
);
297 static struct kobj_attribute hpage_pmd_size_attr
=
298 __ATTR_RO(hpage_pmd_size
);
300 #ifdef CONFIG_DEBUG_VM
301 static ssize_t
debug_cow_show(struct kobject
*kobj
,
302 struct kobj_attribute
*attr
, char *buf
)
304 return single_hugepage_flag_show(kobj
, attr
, buf
,
305 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
307 static ssize_t
debug_cow_store(struct kobject
*kobj
,
308 struct kobj_attribute
*attr
,
309 const char *buf
, size_t count
)
311 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
312 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG
);
314 static struct kobj_attribute debug_cow_attr
=
315 __ATTR(debug_cow
, 0644, debug_cow_show
, debug_cow_store
);
316 #endif /* CONFIG_DEBUG_VM */
318 static struct attribute
*hugepage_attr
[] = {
321 &use_zero_page_attr
.attr
,
322 &hpage_pmd_size_attr
.attr
,
323 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
324 &shmem_enabled_attr
.attr
,
326 #ifdef CONFIG_DEBUG_VM
327 &debug_cow_attr
.attr
,
332 static const struct attribute_group hugepage_attr_group
= {
333 .attrs
= hugepage_attr
,
336 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
340 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
341 if (unlikely(!*hugepage_kobj
)) {
342 pr_err("failed to create transparent hugepage kobject\n");
346 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
348 pr_err("failed to register transparent hugepage group\n");
352 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
354 pr_err("failed to register transparent hugepage group\n");
355 goto remove_hp_group
;
361 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
363 kobject_put(*hugepage_kobj
);
367 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
369 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
370 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
371 kobject_put(hugepage_kobj
);
374 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
379 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
382 #endif /* CONFIG_SYSFS */
384 static int __init
hugepage_init(void)
387 struct kobject
*hugepage_kobj
;
389 if (!has_transparent_hugepage()) {
390 transparent_hugepage_flags
= 0;
395 * hugepages can't be allocated by the buddy allocator
397 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
>= MAX_ORDER
);
399 * we use page->mapping and page->index in second tail page
400 * as list_head: assuming THP order >= 2
402 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
404 err
= hugepage_init_sysfs(&hugepage_kobj
);
408 err
= khugepaged_init();
412 err
= register_shrinker(&huge_zero_page_shrinker
);
414 goto err_hzp_shrinker
;
415 err
= register_shrinker(&deferred_split_shrinker
);
417 goto err_split_shrinker
;
420 * By default disable transparent hugepages on smaller systems,
421 * where the extra memory used could hurt more than TLB overhead
422 * is likely to save. The admin can still enable it through /sys.
424 if (totalram_pages
< (512 << (20 - PAGE_SHIFT
))) {
425 transparent_hugepage_flags
= 0;
429 err
= start_stop_khugepaged();
435 unregister_shrinker(&deferred_split_shrinker
);
437 unregister_shrinker(&huge_zero_page_shrinker
);
439 khugepaged_destroy();
441 hugepage_exit_sysfs(hugepage_kobj
);
445 subsys_initcall(hugepage_init
);
447 static int __init
setup_transparent_hugepage(char *str
)
452 if (!strcmp(str
, "always")) {
453 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
454 &transparent_hugepage_flags
);
455 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
456 &transparent_hugepage_flags
);
458 } else if (!strcmp(str
, "madvise")) {
459 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
460 &transparent_hugepage_flags
);
461 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
462 &transparent_hugepage_flags
);
464 } else if (!strcmp(str
, "never")) {
465 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
466 &transparent_hugepage_flags
);
467 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
468 &transparent_hugepage_flags
);
473 pr_warn("transparent_hugepage= cannot parse, ignored\n");
476 __setup("transparent_hugepage=", setup_transparent_hugepage
);
478 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
480 if (likely(vma
->vm_flags
& VM_WRITE
))
481 pmd
= pmd_mkwrite(pmd
);
485 static inline struct list_head
*page_deferred_list(struct page
*page
)
488 * ->lru in the tail pages is occupied by compound_head.
489 * Let's use ->mapping + ->index in the second tail page as list_head.
491 return (struct list_head
*)&page
[2].mapping
;
494 void prep_transhuge_page(struct page
*page
)
497 * we use page->mapping and page->indexlru in second tail page
498 * as list_head: assuming THP order >= 2
501 INIT_LIST_HEAD(page_deferred_list(page
));
502 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
505 unsigned long __thp_get_unmapped_area(struct file
*filp
, unsigned long len
,
506 loff_t off
, unsigned long flags
, unsigned long size
)
509 loff_t off_end
= off
+ len
;
510 loff_t off_align
= round_up(off
, size
);
511 unsigned long len_pad
;
513 if (off_end
<= off_align
|| (off_end
- off_align
) < size
)
516 len_pad
= len
+ size
;
517 if (len_pad
< len
|| (off
+ len_pad
) < off
)
520 addr
= current
->mm
->get_unmapped_area(filp
, 0, len_pad
,
521 off
>> PAGE_SHIFT
, flags
);
522 if (IS_ERR_VALUE(addr
))
525 addr
+= (off
- addr
) & (size
- 1);
529 unsigned long thp_get_unmapped_area(struct file
*filp
, unsigned long addr
,
530 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
532 loff_t off
= (loff_t
)pgoff
<< PAGE_SHIFT
;
536 if (!IS_DAX(filp
->f_mapping
->host
) || !IS_ENABLED(CONFIG_FS_DAX_PMD
))
539 addr
= __thp_get_unmapped_area(filp
, len
, off
, flags
, PMD_SIZE
);
544 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
546 EXPORT_SYMBOL_GPL(thp_get_unmapped_area
);
548 static int __do_huge_pmd_anonymous_page(struct vm_fault
*vmf
, struct page
*page
,
551 struct vm_area_struct
*vma
= vmf
->vma
;
552 struct mem_cgroup
*memcg
;
554 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
557 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
559 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, gfp
| __GFP_NORETRY
, &memcg
,
562 count_vm_event(THP_FAULT_FALLBACK
);
563 return VM_FAULT_FALLBACK
;
566 pgtable
= pte_alloc_one(vma
->vm_mm
, haddr
);
567 if (unlikely(!pgtable
)) {
572 clear_huge_page(page
, vmf
->address
, HPAGE_PMD_NR
);
574 * The memory barrier inside __SetPageUptodate makes sure that
575 * clear_huge_page writes become visible before the set_pmd_at()
578 __SetPageUptodate(page
);
580 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
581 if (unlikely(!pmd_none(*vmf
->pmd
))) {
586 ret
= check_stable_address_space(vma
->vm_mm
);
590 /* Deliver the page fault to userland */
591 if (userfaultfd_missing(vma
)) {
594 spin_unlock(vmf
->ptl
);
595 mem_cgroup_cancel_charge(page
, memcg
, true);
597 pte_free(vma
->vm_mm
, pgtable
);
598 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
599 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
603 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
604 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
605 page_add_new_anon_rmap(page
, vma
, haddr
, true);
606 mem_cgroup_commit_charge(page
, memcg
, false, true);
607 lru_cache_add_active_or_unevictable(page
, vma
);
608 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
609 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
610 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
611 mm_inc_nr_ptes(vma
->vm_mm
);
612 spin_unlock(vmf
->ptl
);
613 count_vm_event(THP_FAULT_ALLOC
);
618 spin_unlock(vmf
->ptl
);
621 pte_free(vma
->vm_mm
, pgtable
);
622 mem_cgroup_cancel_charge(page
, memcg
, true);
629 * always: directly stall for all thp allocations
630 * defer: wake kswapd and fail if not immediately available
631 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
632 * fail if not immediately available
633 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
635 * never: never stall for any thp allocation
637 static inline gfp_t
alloc_hugepage_direct_gfpmask(struct vm_area_struct
*vma
)
639 const bool vma_madvised
= !!(vma
->vm_flags
& VM_HUGEPAGE
);
641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
642 return GFP_TRANSHUGE
| (vma_madvised
? 0 : __GFP_NORETRY
);
643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
644 return GFP_TRANSHUGE_LIGHT
| __GFP_KSWAPD_RECLAIM
;
645 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
646 return GFP_TRANSHUGE_LIGHT
| (vma_madvised
? __GFP_DIRECT_RECLAIM
:
647 __GFP_KSWAPD_RECLAIM
);
648 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
649 return GFP_TRANSHUGE_LIGHT
| (vma_madvised
? __GFP_DIRECT_RECLAIM
:
651 return GFP_TRANSHUGE_LIGHT
;
654 /* Caller must hold page table lock. */
655 static bool set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
656 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
657 struct page
*zero_page
)
662 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
663 entry
= pmd_mkhuge(entry
);
665 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
666 set_pmd_at(mm
, haddr
, pmd
, entry
);
671 int do_huge_pmd_anonymous_page(struct vm_fault
*vmf
)
673 struct vm_area_struct
*vma
= vmf
->vma
;
676 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
678 if (haddr
< vma
->vm_start
|| haddr
+ HPAGE_PMD_SIZE
> vma
->vm_end
)
679 return VM_FAULT_FALLBACK
;
680 if (unlikely(anon_vma_prepare(vma
)))
682 if (unlikely(khugepaged_enter(vma
, vma
->vm_flags
)))
684 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
685 !mm_forbids_zeropage(vma
->vm_mm
) &&
686 transparent_hugepage_use_zero_page()) {
688 struct page
*zero_page
;
691 pgtable
= pte_alloc_one(vma
->vm_mm
, haddr
);
692 if (unlikely(!pgtable
))
694 zero_page
= mm_get_huge_zero_page(vma
->vm_mm
);
695 if (unlikely(!zero_page
)) {
696 pte_free(vma
->vm_mm
, pgtable
);
697 count_vm_event(THP_FAULT_FALLBACK
);
698 return VM_FAULT_FALLBACK
;
700 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
703 if (pmd_none(*vmf
->pmd
)) {
704 ret
= check_stable_address_space(vma
->vm_mm
);
706 spin_unlock(vmf
->ptl
);
707 } else if (userfaultfd_missing(vma
)) {
708 spin_unlock(vmf
->ptl
);
709 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
710 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
712 set_huge_zero_page(pgtable
, vma
->vm_mm
, vma
,
713 haddr
, vmf
->pmd
, zero_page
);
714 spin_unlock(vmf
->ptl
);
718 spin_unlock(vmf
->ptl
);
720 pte_free(vma
->vm_mm
, pgtable
);
723 gfp
= alloc_hugepage_direct_gfpmask(vma
);
724 page
= alloc_hugepage_vma(gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
725 if (unlikely(!page
)) {
726 count_vm_event(THP_FAULT_FALLBACK
);
727 return VM_FAULT_FALLBACK
;
729 prep_transhuge_page(page
);
730 return __do_huge_pmd_anonymous_page(vmf
, page
, gfp
);
733 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
734 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
,
737 struct mm_struct
*mm
= vma
->vm_mm
;
741 ptl
= pmd_lock(mm
, pmd
);
742 if (!pmd_none(*pmd
)) {
744 if (pmd_pfn(*pmd
) != pfn_t_to_pfn(pfn
)) {
745 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd
));
748 entry
= pmd_mkyoung(*pmd
);
749 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
750 if (pmdp_set_access_flags(vma
, addr
, pmd
, entry
, 1))
751 update_mmu_cache_pmd(vma
, addr
, pmd
);
757 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
758 if (pfn_t_devmap(pfn
))
759 entry
= pmd_mkdevmap(entry
);
761 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
762 entry
= maybe_pmd_mkwrite(entry
, vma
);
766 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
771 set_pmd_at(mm
, addr
, pmd
, entry
);
772 update_mmu_cache_pmd(vma
, addr
, pmd
);
777 pte_free(mm
, pgtable
);
780 int vmf_insert_pfn_pmd(struct vm_fault
*vmf
, pfn_t pfn
, bool write
)
782 unsigned long addr
= vmf
->address
& PMD_MASK
;
783 struct vm_area_struct
*vma
= vmf
->vma
;
784 pgprot_t pgprot
= vma
->vm_page_prot
;
785 pgtable_t pgtable
= NULL
;
788 * If we had pmd_special, we could avoid all these restrictions,
789 * but we need to be consistent with PTEs and architectures that
790 * can't support a 'special' bit.
792 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
793 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
794 (VM_PFNMAP
|VM_MIXEDMAP
));
795 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
796 BUG_ON(!pfn_t_devmap(pfn
));
798 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
799 return VM_FAULT_SIGBUS
;
801 if (arch_needs_pgtable_deposit()) {
802 pgtable
= pte_alloc_one(vma
->vm_mm
, addr
);
807 track_pfn_insert(vma
, &pgprot
, pfn
);
809 insert_pfn_pmd(vma
, addr
, vmf
->pmd
, pfn
, pgprot
, write
, pgtable
);
810 return VM_FAULT_NOPAGE
;
812 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd
);
814 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
815 static pud_t
maybe_pud_mkwrite(pud_t pud
, struct vm_area_struct
*vma
)
817 if (likely(vma
->vm_flags
& VM_WRITE
))
818 pud
= pud_mkwrite(pud
);
822 static void insert_pfn_pud(struct vm_area_struct
*vma
, unsigned long addr
,
823 pud_t
*pud
, pfn_t pfn
, pgprot_t prot
, bool write
)
825 struct mm_struct
*mm
= vma
->vm_mm
;
829 ptl
= pud_lock(mm
, pud
);
830 if (!pud_none(*pud
)) {
832 if (pud_pfn(*pud
) != pfn_t_to_pfn(pfn
)) {
833 WARN_ON_ONCE(!is_huge_zero_pud(*pud
));
836 entry
= pud_mkyoung(*pud
);
837 entry
= maybe_pud_mkwrite(pud_mkdirty(entry
), vma
);
838 if (pudp_set_access_flags(vma
, addr
, pud
, entry
, 1))
839 update_mmu_cache_pud(vma
, addr
, pud
);
844 entry
= pud_mkhuge(pfn_t_pud(pfn
, prot
));
845 if (pfn_t_devmap(pfn
))
846 entry
= pud_mkdevmap(entry
);
848 entry
= pud_mkyoung(pud_mkdirty(entry
));
849 entry
= maybe_pud_mkwrite(entry
, vma
);
851 set_pud_at(mm
, addr
, pud
, entry
);
852 update_mmu_cache_pud(vma
, addr
, pud
);
858 int vmf_insert_pfn_pud(struct vm_fault
*vmf
, pfn_t pfn
, bool write
)
860 unsigned long addr
= vmf
->address
& PUD_MASK
;
861 struct vm_area_struct
*vma
= vmf
->vma
;
862 pgprot_t pgprot
= vma
->vm_page_prot
;
865 * If we had pud_special, we could avoid all these restrictions,
866 * but we need to be consistent with PTEs and architectures that
867 * can't support a 'special' bit.
869 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)));
870 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
871 (VM_PFNMAP
|VM_MIXEDMAP
));
872 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
873 BUG_ON(!pfn_t_devmap(pfn
));
875 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
876 return VM_FAULT_SIGBUS
;
878 track_pfn_insert(vma
, &pgprot
, pfn
);
880 insert_pfn_pud(vma
, addr
, vmf
->pud
, pfn
, pgprot
, write
);
881 return VM_FAULT_NOPAGE
;
883 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud
);
884 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
886 static void touch_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
887 pmd_t
*pmd
, int flags
)
891 _pmd
= pmd_mkyoung(*pmd
);
892 if (flags
& FOLL_WRITE
)
893 _pmd
= pmd_mkdirty(_pmd
);
894 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
895 pmd
, _pmd
, flags
& FOLL_WRITE
))
896 update_mmu_cache_pmd(vma
, addr
, pmd
);
899 struct page
*follow_devmap_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
900 pmd_t
*pmd
, int flags
)
902 unsigned long pfn
= pmd_pfn(*pmd
);
903 struct mm_struct
*mm
= vma
->vm_mm
;
904 struct dev_pagemap
*pgmap
;
907 assert_spin_locked(pmd_lockptr(mm
, pmd
));
910 * When we COW a devmap PMD entry, we split it into PTEs, so we should
911 * not be in this function with `flags & FOLL_COW` set.
913 WARN_ONCE(flags
& FOLL_COW
, "mm: In follow_devmap_pmd with FOLL_COW set");
915 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
918 if (pmd_present(*pmd
) && pmd_devmap(*pmd
))
923 if (flags
& FOLL_TOUCH
)
924 touch_pmd(vma
, addr
, pmd
, flags
);
927 * device mapped pages can only be returned if the
928 * caller will manage the page reference count.
930 if (!(flags
& FOLL_GET
))
931 return ERR_PTR(-EEXIST
);
933 pfn
+= (addr
& ~PMD_MASK
) >> PAGE_SHIFT
;
934 pgmap
= get_dev_pagemap(pfn
, NULL
);
936 return ERR_PTR(-EFAULT
);
937 page
= pfn_to_page(pfn
);
939 put_dev_pagemap(pgmap
);
944 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
945 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
946 struct vm_area_struct
*vma
)
948 spinlock_t
*dst_ptl
, *src_ptl
;
949 struct page
*src_page
;
951 pgtable_t pgtable
= NULL
;
954 /* Skip if can be re-fill on fault */
955 if (!vma_is_anonymous(vma
))
958 pgtable
= pte_alloc_one(dst_mm
, addr
);
959 if (unlikely(!pgtable
))
962 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
963 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
964 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
969 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
970 if (unlikely(is_swap_pmd(pmd
))) {
971 swp_entry_t entry
= pmd_to_swp_entry(pmd
);
973 VM_BUG_ON(!is_pmd_migration_entry(pmd
));
974 if (is_write_migration_entry(entry
)) {
975 make_migration_entry_read(&entry
);
976 pmd
= swp_entry_to_pmd(entry
);
977 if (pmd_swp_soft_dirty(*src_pmd
))
978 pmd
= pmd_swp_mksoft_dirty(pmd
);
979 set_pmd_at(src_mm
, addr
, src_pmd
, pmd
);
981 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
982 mm_inc_nr_ptes(dst_mm
);
983 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
984 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
990 if (unlikely(!pmd_trans_huge(pmd
))) {
991 pte_free(dst_mm
, pgtable
);
995 * When page table lock is held, the huge zero pmd should not be
996 * under splitting since we don't split the page itself, only pmd to
999 if (is_huge_zero_pmd(pmd
)) {
1000 struct page
*zero_page
;
1002 * get_huge_zero_page() will never allocate a new page here,
1003 * since we already have a zero page to copy. It just takes a
1006 zero_page
= mm_get_huge_zero_page(dst_mm
);
1007 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
1013 src_page
= pmd_page(pmd
);
1014 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
1016 page_dup_rmap(src_page
, true);
1017 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1018 mm_inc_nr_ptes(dst_mm
);
1019 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1021 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
1022 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
1023 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1027 spin_unlock(src_ptl
);
1028 spin_unlock(dst_ptl
);
1033 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1034 static void touch_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1035 pud_t
*pud
, int flags
)
1039 _pud
= pud_mkyoung(*pud
);
1040 if (flags
& FOLL_WRITE
)
1041 _pud
= pud_mkdirty(_pud
);
1042 if (pudp_set_access_flags(vma
, addr
& HPAGE_PUD_MASK
,
1043 pud
, _pud
, flags
& FOLL_WRITE
))
1044 update_mmu_cache_pud(vma
, addr
, pud
);
1047 struct page
*follow_devmap_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1048 pud_t
*pud
, int flags
)
1050 unsigned long pfn
= pud_pfn(*pud
);
1051 struct mm_struct
*mm
= vma
->vm_mm
;
1052 struct dev_pagemap
*pgmap
;
1055 assert_spin_locked(pud_lockptr(mm
, pud
));
1057 if (flags
& FOLL_WRITE
&& !pud_write(*pud
))
1060 if (pud_present(*pud
) && pud_devmap(*pud
))
1065 if (flags
& FOLL_TOUCH
)
1066 touch_pud(vma
, addr
, pud
, flags
);
1069 * device mapped pages can only be returned if the
1070 * caller will manage the page reference count.
1072 if (!(flags
& FOLL_GET
))
1073 return ERR_PTR(-EEXIST
);
1075 pfn
+= (addr
& ~PUD_MASK
) >> PAGE_SHIFT
;
1076 pgmap
= get_dev_pagemap(pfn
, NULL
);
1078 return ERR_PTR(-EFAULT
);
1079 page
= pfn_to_page(pfn
);
1081 put_dev_pagemap(pgmap
);
1086 int copy_huge_pud(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1087 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1088 struct vm_area_struct
*vma
)
1090 spinlock_t
*dst_ptl
, *src_ptl
;
1094 dst_ptl
= pud_lock(dst_mm
, dst_pud
);
1095 src_ptl
= pud_lockptr(src_mm
, src_pud
);
1096 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1100 if (unlikely(!pud_trans_huge(pud
) && !pud_devmap(pud
)))
1104 * When page table lock is held, the huge zero pud should not be
1105 * under splitting since we don't split the page itself, only pud to
1108 if (is_huge_zero_pud(pud
)) {
1109 /* No huge zero pud yet */
1112 pudp_set_wrprotect(src_mm
, addr
, src_pud
);
1113 pud
= pud_mkold(pud_wrprotect(pud
));
1114 set_pud_at(dst_mm
, addr
, dst_pud
, pud
);
1118 spin_unlock(src_ptl
);
1119 spin_unlock(dst_ptl
);
1123 void huge_pud_set_accessed(struct vm_fault
*vmf
, pud_t orig_pud
)
1126 unsigned long haddr
;
1127 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1129 vmf
->ptl
= pud_lock(vmf
->vma
->vm_mm
, vmf
->pud
);
1130 if (unlikely(!pud_same(*vmf
->pud
, orig_pud
)))
1133 entry
= pud_mkyoung(orig_pud
);
1135 entry
= pud_mkdirty(entry
);
1136 haddr
= vmf
->address
& HPAGE_PUD_MASK
;
1137 if (pudp_set_access_flags(vmf
->vma
, haddr
, vmf
->pud
, entry
, write
))
1138 update_mmu_cache_pud(vmf
->vma
, vmf
->address
, vmf
->pud
);
1141 spin_unlock(vmf
->ptl
);
1143 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1145 void huge_pmd_set_accessed(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1148 unsigned long haddr
;
1149 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1151 vmf
->ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1152 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1155 entry
= pmd_mkyoung(orig_pmd
);
1157 entry
= pmd_mkdirty(entry
);
1158 haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1159 if (pmdp_set_access_flags(vmf
->vma
, haddr
, vmf
->pmd
, entry
, write
))
1160 update_mmu_cache_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
);
1163 spin_unlock(vmf
->ptl
);
1166 static int do_huge_pmd_wp_page_fallback(struct vm_fault
*vmf
, pmd_t orig_pmd
,
1169 struct vm_area_struct
*vma
= vmf
->vma
;
1170 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1171 struct mem_cgroup
*memcg
;
1175 struct page
**pages
;
1176 unsigned long mmun_start
; /* For mmu_notifiers */
1177 unsigned long mmun_end
; /* For mmu_notifiers */
1179 pages
= kmalloc(sizeof(struct page
*) * HPAGE_PMD_NR
,
1181 if (unlikely(!pages
)) {
1182 ret
|= VM_FAULT_OOM
;
1186 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1187 pages
[i
] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE
, vma
,
1188 vmf
->address
, page_to_nid(page
));
1189 if (unlikely(!pages
[i
] ||
1190 mem_cgroup_try_charge(pages
[i
], vma
->vm_mm
,
1191 GFP_KERNEL
, &memcg
, false))) {
1195 memcg
= (void *)page_private(pages
[i
]);
1196 set_page_private(pages
[i
], 0);
1197 mem_cgroup_cancel_charge(pages
[i
], memcg
,
1202 ret
|= VM_FAULT_OOM
;
1205 set_page_private(pages
[i
], (unsigned long)memcg
);
1208 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1209 copy_user_highpage(pages
[i
], page
+ i
,
1210 haddr
+ PAGE_SIZE
* i
, vma
);
1211 __SetPageUptodate(pages
[i
]);
1216 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1217 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1219 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1220 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1221 goto out_free_pages
;
1222 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1225 * Leave pmd empty until pte is filled note we must notify here as
1226 * concurrent CPU thread might write to new page before the call to
1227 * mmu_notifier_invalidate_range_end() happens which can lead to a
1228 * device seeing memory write in different order than CPU.
1230 * See Documentation/vm/mmu_notifier.txt
1232 pmdp_huge_clear_flush_notify(vma
, haddr
, vmf
->pmd
);
1234 pgtable
= pgtable_trans_huge_withdraw(vma
->vm_mm
, vmf
->pmd
);
1235 pmd_populate(vma
->vm_mm
, &_pmd
, pgtable
);
1237 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1239 entry
= mk_pte(pages
[i
], vma
->vm_page_prot
);
1240 entry
= maybe_mkwrite(pte_mkdirty(entry
), vma
);
1241 memcg
= (void *)page_private(pages
[i
]);
1242 set_page_private(pages
[i
], 0);
1243 page_add_new_anon_rmap(pages
[i
], vmf
->vma
, haddr
, false);
1244 mem_cgroup_commit_charge(pages
[i
], memcg
, false, false);
1245 lru_cache_add_active_or_unevictable(pages
[i
], vma
);
1246 vmf
->pte
= pte_offset_map(&_pmd
, haddr
);
1247 VM_BUG_ON(!pte_none(*vmf
->pte
));
1248 set_pte_at(vma
->vm_mm
, haddr
, vmf
->pte
, entry
);
1249 pte_unmap(vmf
->pte
);
1253 smp_wmb(); /* make pte visible before pmd */
1254 pmd_populate(vma
->vm_mm
, vmf
->pmd
, pgtable
);
1255 page_remove_rmap(page
, true);
1256 spin_unlock(vmf
->ptl
);
1259 * No need to double call mmu_notifier->invalidate_range() callback as
1260 * the above pmdp_huge_clear_flush_notify() did already call it.
1262 mmu_notifier_invalidate_range_only_end(vma
->vm_mm
, mmun_start
,
1265 ret
|= VM_FAULT_WRITE
;
1272 spin_unlock(vmf
->ptl
);
1273 mmu_notifier_invalidate_range_end(vma
->vm_mm
, mmun_start
, mmun_end
);
1274 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
1275 memcg
= (void *)page_private(pages
[i
]);
1276 set_page_private(pages
[i
], 0);
1277 mem_cgroup_cancel_charge(pages
[i
], memcg
, false);
1284 int do_huge_pmd_wp_page(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1286 struct vm_area_struct
*vma
= vmf
->vma
;
1287 struct page
*page
= NULL
, *new_page
;
1288 struct mem_cgroup
*memcg
;
1289 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1290 unsigned long mmun_start
; /* For mmu_notifiers */
1291 unsigned long mmun_end
; /* For mmu_notifiers */
1292 gfp_t huge_gfp
; /* for allocation and charge */
1295 vmf
->ptl
= pmd_lockptr(vma
->vm_mm
, vmf
->pmd
);
1296 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1297 if (is_huge_zero_pmd(orig_pmd
))
1299 spin_lock(vmf
->ptl
);
1300 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1303 page
= pmd_page(orig_pmd
);
1304 VM_BUG_ON_PAGE(!PageCompound(page
) || !PageHead(page
), page
);
1306 * We can only reuse the page if nobody else maps the huge page or it's
1309 if (!trylock_page(page
)) {
1311 spin_unlock(vmf
->ptl
);
1313 spin_lock(vmf
->ptl
);
1314 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1321 if (reuse_swap_page(page
, NULL
)) {
1323 entry
= pmd_mkyoung(orig_pmd
);
1324 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1325 if (pmdp_set_access_flags(vma
, haddr
, vmf
->pmd
, entry
, 1))
1326 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1327 ret
|= VM_FAULT_WRITE
;
1333 spin_unlock(vmf
->ptl
);
1335 if (transparent_hugepage_enabled(vma
) &&
1336 !transparent_hugepage_debug_cow()) {
1337 huge_gfp
= alloc_hugepage_direct_gfpmask(vma
);
1338 new_page
= alloc_hugepage_vma(huge_gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
1342 if (likely(new_page
)) {
1343 prep_transhuge_page(new_page
);
1346 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1347 ret
|= VM_FAULT_FALLBACK
;
1349 ret
= do_huge_pmd_wp_page_fallback(vmf
, orig_pmd
, page
);
1350 if (ret
& VM_FAULT_OOM
) {
1351 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1352 ret
|= VM_FAULT_FALLBACK
;
1356 count_vm_event(THP_FAULT_FALLBACK
);
1360 if (unlikely(mem_cgroup_try_charge(new_page
, vma
->vm_mm
,
1361 huge_gfp
| __GFP_NORETRY
, &memcg
, true))) {
1363 split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
);
1366 ret
|= VM_FAULT_FALLBACK
;
1367 count_vm_event(THP_FAULT_FALLBACK
);
1371 count_vm_event(THP_FAULT_ALLOC
);
1374 clear_huge_page(new_page
, vmf
->address
, HPAGE_PMD_NR
);
1376 copy_user_huge_page(new_page
, page
, haddr
, vma
, HPAGE_PMD_NR
);
1377 __SetPageUptodate(new_page
);
1380 mmun_end
= haddr
+ HPAGE_PMD_SIZE
;
1381 mmu_notifier_invalidate_range_start(vma
->vm_mm
, mmun_start
, mmun_end
);
1383 spin_lock(vmf
->ptl
);
1386 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1387 spin_unlock(vmf
->ptl
);
1388 mem_cgroup_cancel_charge(new_page
, memcg
, true);
1393 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1394 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1395 pmdp_huge_clear_flush_notify(vma
, haddr
, vmf
->pmd
);
1396 page_add_new_anon_rmap(new_page
, vma
, haddr
, true);
1397 mem_cgroup_commit_charge(new_page
, memcg
, false, true);
1398 lru_cache_add_active_or_unevictable(new_page
, vma
);
1399 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
1400 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1402 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1404 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1405 page_remove_rmap(page
, true);
1408 ret
|= VM_FAULT_WRITE
;
1410 spin_unlock(vmf
->ptl
);
1413 * No need to double call mmu_notifier->invalidate_range() callback as
1414 * the above pmdp_huge_clear_flush_notify() did already call it.
1416 mmu_notifier_invalidate_range_only_end(vma
->vm_mm
, mmun_start
,
1421 spin_unlock(vmf
->ptl
);
1426 * FOLL_FORCE can write to even unwritable pmd's, but only
1427 * after we've gone through a COW cycle and they are dirty.
1429 static inline bool can_follow_write_pmd(pmd_t pmd
, unsigned int flags
)
1431 return pmd_write(pmd
) ||
1432 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pmd_dirty(pmd
));
1435 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1440 struct mm_struct
*mm
= vma
->vm_mm
;
1441 struct page
*page
= NULL
;
1443 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1445 if (flags
& FOLL_WRITE
&& !can_follow_write_pmd(*pmd
, flags
))
1448 /* Avoid dumping huge zero page */
1449 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1450 return ERR_PTR(-EFAULT
);
1452 /* Full NUMA hinting faults to serialise migration in fault paths */
1453 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1456 page
= pmd_page(*pmd
);
1457 VM_BUG_ON_PAGE(!PageHead(page
) && !is_zone_device_page(page
), page
);
1458 if (flags
& FOLL_TOUCH
)
1459 touch_pmd(vma
, addr
, pmd
, flags
);
1460 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1462 * We don't mlock() pte-mapped THPs. This way we can avoid
1463 * leaking mlocked pages into non-VM_LOCKED VMAs.
1467 * In most cases the pmd is the only mapping of the page as we
1468 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1469 * writable private mappings in populate_vma_page_range().
1471 * The only scenario when we have the page shared here is if we
1472 * mlocking read-only mapping shared over fork(). We skip
1473 * mlocking such pages.
1477 * We can expect PageDoubleMap() to be stable under page lock:
1478 * for file pages we set it in page_add_file_rmap(), which
1479 * requires page to be locked.
1482 if (PageAnon(page
) && compound_mapcount(page
) != 1)
1484 if (PageDoubleMap(page
) || !page
->mapping
)
1486 if (!trylock_page(page
))
1489 if (page
->mapping
&& !PageDoubleMap(page
))
1490 mlock_vma_page(page
);
1494 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1495 VM_BUG_ON_PAGE(!PageCompound(page
) && !is_zone_device_page(page
), page
);
1496 if (flags
& FOLL_GET
)
1503 /* NUMA hinting page fault entry point for trans huge pmds */
1504 int do_huge_pmd_numa_page(struct vm_fault
*vmf
, pmd_t pmd
)
1506 struct vm_area_struct
*vma
= vmf
->vma
;
1507 struct anon_vma
*anon_vma
= NULL
;
1509 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1510 int page_nid
= -1, this_nid
= numa_node_id();
1511 int target_nid
, last_cpupid
= -1;
1513 bool migrated
= false;
1517 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1518 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
)))
1522 * If there are potential migrations, wait for completion and retry
1523 * without disrupting NUMA hinting information. Do not relock and
1524 * check_same as the page may no longer be mapped.
1526 if (unlikely(pmd_trans_migrating(*vmf
->pmd
))) {
1527 page
= pmd_page(*vmf
->pmd
);
1528 if (!get_page_unless_zero(page
))
1530 spin_unlock(vmf
->ptl
);
1531 wait_on_page_locked(page
);
1536 page
= pmd_page(pmd
);
1537 BUG_ON(is_huge_zero_page(page
));
1538 page_nid
= page_to_nid(page
);
1539 last_cpupid
= page_cpupid_last(page
);
1540 count_vm_numa_event(NUMA_HINT_FAULTS
);
1541 if (page_nid
== this_nid
) {
1542 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1543 flags
|= TNF_FAULT_LOCAL
;
1546 /* See similar comment in do_numa_page for explanation */
1547 if (!pmd_savedwrite(pmd
))
1548 flags
|= TNF_NO_GROUP
;
1551 * Acquire the page lock to serialise THP migrations but avoid dropping
1552 * page_table_lock if at all possible
1554 page_locked
= trylock_page(page
);
1555 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1556 if (target_nid
== -1) {
1557 /* If the page was locked, there are no parallel migrations */
1562 /* Migration could have started since the pmd_trans_migrating check */
1565 if (!get_page_unless_zero(page
))
1567 spin_unlock(vmf
->ptl
);
1568 wait_on_page_locked(page
);
1574 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1575 * to serialises splits
1578 spin_unlock(vmf
->ptl
);
1579 anon_vma
= page_lock_anon_vma_read(page
);
1581 /* Confirm the PMD did not change while page_table_lock was released */
1582 spin_lock(vmf
->ptl
);
1583 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
))) {
1590 /* Bail if we fail to protect against THP splits for any reason */
1591 if (unlikely(!anon_vma
)) {
1598 * Since we took the NUMA fault, we must have observed the !accessible
1599 * bit. Make sure all other CPUs agree with that, to avoid them
1600 * modifying the page we're about to migrate.
1602 * Must be done under PTL such that we'll observe the relevant
1603 * inc_tlb_flush_pending().
1605 * We are not sure a pending tlb flush here is for a huge page
1606 * mapping or not. Hence use the tlb range variant
1608 if (mm_tlb_flush_pending(vma
->vm_mm
))
1609 flush_tlb_range(vma
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
1612 * Migrate the THP to the requested node, returns with page unlocked
1613 * and access rights restored.
1615 spin_unlock(vmf
->ptl
);
1617 migrated
= migrate_misplaced_transhuge_page(vma
->vm_mm
, vma
,
1618 vmf
->pmd
, pmd
, vmf
->address
, page
, target_nid
);
1620 flags
|= TNF_MIGRATED
;
1621 page_nid
= target_nid
;
1623 flags
|= TNF_MIGRATE_FAIL
;
1627 BUG_ON(!PageLocked(page
));
1628 was_writable
= pmd_savedwrite(pmd
);
1629 pmd
= pmd_modify(pmd
, vma
->vm_page_prot
);
1630 pmd
= pmd_mkyoung(pmd
);
1632 pmd
= pmd_mkwrite(pmd
);
1633 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, pmd
);
1634 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1637 spin_unlock(vmf
->ptl
);
1641 page_unlock_anon_vma_read(anon_vma
);
1644 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
,
1651 * Return true if we do MADV_FREE successfully on entire pmd page.
1652 * Otherwise, return false.
1654 bool madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1655 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1660 struct mm_struct
*mm
= tlb
->mm
;
1663 tlb_remove_check_page_size_change(tlb
, HPAGE_PMD_SIZE
);
1665 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1670 if (is_huge_zero_pmd(orig_pmd
))
1673 if (unlikely(!pmd_present(orig_pmd
))) {
1674 VM_BUG_ON(thp_migration_supported() &&
1675 !is_pmd_migration_entry(orig_pmd
));
1679 page
= pmd_page(orig_pmd
);
1681 * If other processes are mapping this page, we couldn't discard
1682 * the page unless they all do MADV_FREE so let's skip the page.
1684 if (page_mapcount(page
) != 1)
1687 if (!trylock_page(page
))
1691 * If user want to discard part-pages of THP, split it so MADV_FREE
1692 * will deactivate only them.
1694 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1697 split_huge_page(page
);
1703 if (PageDirty(page
))
1704 ClearPageDirty(page
);
1707 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1708 pmdp_invalidate(vma
, addr
, pmd
);
1709 orig_pmd
= pmd_mkold(orig_pmd
);
1710 orig_pmd
= pmd_mkclean(orig_pmd
);
1712 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1713 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1716 mark_page_lazyfree(page
);
1724 static inline void zap_deposited_table(struct mm_struct
*mm
, pmd_t
*pmd
)
1728 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1729 pte_free(mm
, pgtable
);
1733 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1734 pmd_t
*pmd
, unsigned long addr
)
1739 tlb_remove_check_page_size_change(tlb
, HPAGE_PMD_SIZE
);
1741 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1745 * For architectures like ppc64 we look at deposited pgtable
1746 * when calling pmdp_huge_get_and_clear. So do the
1747 * pgtable_trans_huge_withdraw after finishing pmdp related
1750 orig_pmd
= pmdp_huge_get_and_clear_full(tlb
->mm
, addr
, pmd
,
1752 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1753 if (vma_is_dax(vma
)) {
1754 if (arch_needs_pgtable_deposit())
1755 zap_deposited_table(tlb
->mm
, pmd
);
1757 if (is_huge_zero_pmd(orig_pmd
))
1758 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1759 } else if (is_huge_zero_pmd(orig_pmd
)) {
1760 zap_deposited_table(tlb
->mm
, pmd
);
1762 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1764 struct page
*page
= NULL
;
1765 int flush_needed
= 1;
1767 if (pmd_present(orig_pmd
)) {
1768 page
= pmd_page(orig_pmd
);
1769 page_remove_rmap(page
, true);
1770 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1771 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1772 } else if (thp_migration_supported()) {
1775 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd
));
1776 entry
= pmd_to_swp_entry(orig_pmd
);
1777 page
= pfn_to_page(swp_offset(entry
));
1780 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1782 if (PageAnon(page
)) {
1783 zap_deposited_table(tlb
->mm
, pmd
);
1784 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1786 if (arch_needs_pgtable_deposit())
1787 zap_deposited_table(tlb
->mm
, pmd
);
1788 add_mm_counter(tlb
->mm
, MM_FILEPAGES
, -HPAGE_PMD_NR
);
1793 tlb_remove_page_size(tlb
, page
, HPAGE_PMD_SIZE
);
1798 #ifndef pmd_move_must_withdraw
1799 static inline int pmd_move_must_withdraw(spinlock_t
*new_pmd_ptl
,
1800 spinlock_t
*old_pmd_ptl
,
1801 struct vm_area_struct
*vma
)
1804 * With split pmd lock we also need to move preallocated
1805 * PTE page table if new_pmd is on different PMD page table.
1807 * We also don't deposit and withdraw tables for file pages.
1809 return (new_pmd_ptl
!= old_pmd_ptl
) && vma_is_anonymous(vma
);
1813 static pmd_t
move_soft_dirty_pmd(pmd_t pmd
)
1815 #ifdef CONFIG_MEM_SOFT_DIRTY
1816 if (unlikely(is_pmd_migration_entry(pmd
)))
1817 pmd
= pmd_swp_mksoft_dirty(pmd
);
1818 else if (pmd_present(pmd
))
1819 pmd
= pmd_mksoft_dirty(pmd
);
1824 bool move_huge_pmd(struct vm_area_struct
*vma
, unsigned long old_addr
,
1825 unsigned long new_addr
, unsigned long old_end
,
1826 pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1828 spinlock_t
*old_ptl
, *new_ptl
;
1830 struct mm_struct
*mm
= vma
->vm_mm
;
1831 bool force_flush
= false;
1833 if ((old_addr
& ~HPAGE_PMD_MASK
) ||
1834 (new_addr
& ~HPAGE_PMD_MASK
) ||
1835 old_end
- old_addr
< HPAGE_PMD_SIZE
)
1839 * The destination pmd shouldn't be established, free_pgtables()
1840 * should have release it.
1842 if (WARN_ON(!pmd_none(*new_pmd
))) {
1843 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1848 * We don't have to worry about the ordering of src and dst
1849 * ptlocks because exclusive mmap_sem prevents deadlock.
1851 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1853 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1854 if (new_ptl
!= old_ptl
)
1855 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1856 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1857 if (pmd_present(pmd
))
1859 VM_BUG_ON(!pmd_none(*new_pmd
));
1861 if (pmd_move_must_withdraw(new_ptl
, old_ptl
, vma
)) {
1863 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1864 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1866 pmd
= move_soft_dirty_pmd(pmd
);
1867 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1869 flush_tlb_range(vma
, old_addr
, old_addr
+ PMD_SIZE
);
1870 if (new_ptl
!= old_ptl
)
1871 spin_unlock(new_ptl
);
1872 spin_unlock(old_ptl
);
1880 * - 0 if PMD could not be locked
1881 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1882 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1884 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1885 unsigned long addr
, pgprot_t newprot
, int prot_numa
)
1887 struct mm_struct
*mm
= vma
->vm_mm
;
1890 bool preserve_write
;
1893 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1897 preserve_write
= prot_numa
&& pmd_write(*pmd
);
1900 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1901 if (is_swap_pmd(*pmd
)) {
1902 swp_entry_t entry
= pmd_to_swp_entry(*pmd
);
1904 VM_BUG_ON(!is_pmd_migration_entry(*pmd
));
1905 if (is_write_migration_entry(entry
)) {
1908 * A protection check is difficult so
1909 * just be safe and disable write
1911 make_migration_entry_read(&entry
);
1912 newpmd
= swp_entry_to_pmd(entry
);
1913 if (pmd_swp_soft_dirty(*pmd
))
1914 newpmd
= pmd_swp_mksoft_dirty(newpmd
);
1915 set_pmd_at(mm
, addr
, pmd
, newpmd
);
1922 * Avoid trapping faults against the zero page. The read-only
1923 * data is likely to be read-cached on the local CPU and
1924 * local/remote hits to the zero page are not interesting.
1926 if (prot_numa
&& is_huge_zero_pmd(*pmd
))
1929 if (prot_numa
&& pmd_protnone(*pmd
))
1933 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1934 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1935 * which is also under down_read(mmap_sem):
1938 * change_huge_pmd(prot_numa=1)
1939 * pmdp_huge_get_and_clear_notify()
1940 * madvise_dontneed()
1942 * pmd_trans_huge(*pmd) == 0 (without ptl)
1945 * // pmd is re-established
1947 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1948 * which may break userspace.
1950 * pmdp_invalidate() is required to make sure we don't miss
1951 * dirty/young flags set by hardware.
1954 pmdp_invalidate(vma
, addr
, pmd
);
1957 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1960 if (pmd_dirty(*pmd
))
1961 entry
= pmd_mkdirty(entry
);
1962 if (pmd_young(*pmd
))
1963 entry
= pmd_mkyoung(entry
);
1965 entry
= pmd_modify(entry
, newprot
);
1967 entry
= pmd_mk_savedwrite(entry
);
1969 set_pmd_at(mm
, addr
, pmd
, entry
);
1970 BUG_ON(vma_is_anonymous(vma
) && !preserve_write
&& pmd_write(entry
));
1977 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1979 * Note that if it returns page table lock pointer, this routine returns without
1980 * unlocking page table lock. So callers must unlock it.
1982 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1985 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1986 if (likely(is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) ||
1994 * Returns true if a given pud maps a thp, false otherwise.
1996 * Note that if it returns true, this routine returns without unlocking page
1997 * table lock. So callers must unlock it.
1999 spinlock_t
*__pud_trans_huge_lock(pud_t
*pud
, struct vm_area_struct
*vma
)
2003 ptl
= pud_lock(vma
->vm_mm
, pud
);
2004 if (likely(pud_trans_huge(*pud
) || pud_devmap(*pud
)))
2010 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
2011 int zap_huge_pud(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2012 pud_t
*pud
, unsigned long addr
)
2017 ptl
= __pud_trans_huge_lock(pud
, vma
);
2021 * For architectures like ppc64 we look at deposited pgtable
2022 * when calling pudp_huge_get_and_clear. So do the
2023 * pgtable_trans_huge_withdraw after finishing pudp related
2026 orig_pud
= pudp_huge_get_and_clear_full(tlb
->mm
, addr
, pud
,
2028 tlb_remove_pud_tlb_entry(tlb
, pud
, addr
);
2029 if (vma_is_dax(vma
)) {
2031 /* No zero page support yet */
2033 /* No support for anonymous PUD pages yet */
2039 static void __split_huge_pud_locked(struct vm_area_struct
*vma
, pud_t
*pud
,
2040 unsigned long haddr
)
2042 VM_BUG_ON(haddr
& ~HPAGE_PUD_MASK
);
2043 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2044 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PUD_SIZE
, vma
);
2045 VM_BUG_ON(!pud_trans_huge(*pud
) && !pud_devmap(*pud
));
2047 count_vm_event(THP_SPLIT_PUD
);
2049 pudp_huge_clear_flush_notify(vma
, haddr
, pud
);
2052 void __split_huge_pud(struct vm_area_struct
*vma
, pud_t
*pud
,
2053 unsigned long address
)
2056 struct mm_struct
*mm
= vma
->vm_mm
;
2057 unsigned long haddr
= address
& HPAGE_PUD_MASK
;
2059 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PUD_SIZE
);
2060 ptl
= pud_lock(mm
, pud
);
2061 if (unlikely(!pud_trans_huge(*pud
) && !pud_devmap(*pud
)))
2063 __split_huge_pud_locked(vma
, pud
, haddr
);
2068 * No need to double call mmu_notifier->invalidate_range() callback as
2069 * the above pudp_huge_clear_flush_notify() did already call it.
2071 mmu_notifier_invalidate_range_only_end(mm
, haddr
, haddr
+
2074 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2076 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2077 unsigned long haddr
, pmd_t
*pmd
)
2079 struct mm_struct
*mm
= vma
->vm_mm
;
2085 * Leave pmd empty until pte is filled note that it is fine to delay
2086 * notification until mmu_notifier_invalidate_range_end() as we are
2087 * replacing a zero pmd write protected page with a zero pte write
2090 * See Documentation/vm/mmu_notifier.txt
2092 pmdp_huge_clear_flush(vma
, haddr
, pmd
);
2094 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2095 pmd_populate(mm
, &_pmd
, pgtable
);
2097 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2099 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2100 entry
= pte_mkspecial(entry
);
2101 pte
= pte_offset_map(&_pmd
, haddr
);
2102 VM_BUG_ON(!pte_none(*pte
));
2103 set_pte_at(mm
, haddr
, pte
, entry
);
2106 smp_wmb(); /* make pte visible before pmd */
2107 pmd_populate(mm
, pmd
, pgtable
);
2110 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2111 unsigned long haddr
, bool freeze
)
2113 struct mm_struct
*mm
= vma
->vm_mm
;
2117 bool young
, write
, dirty
, soft_dirty
, pmd_migration
= false;
2121 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
2122 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2123 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
2124 VM_BUG_ON(!is_pmd_migration_entry(*pmd
) && !pmd_trans_huge(*pmd
)
2125 && !pmd_devmap(*pmd
));
2127 count_vm_event(THP_SPLIT_PMD
);
2129 if (!vma_is_anonymous(vma
)) {
2130 _pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2132 * We are going to unmap this huge page. So
2133 * just go ahead and zap it
2135 if (arch_needs_pgtable_deposit())
2136 zap_deposited_table(mm
, pmd
);
2137 if (vma_is_dax(vma
))
2139 page
= pmd_page(_pmd
);
2140 if (!PageDirty(page
) && pmd_dirty(_pmd
))
2141 set_page_dirty(page
);
2142 if (!PageReferenced(page
) && pmd_young(_pmd
))
2143 SetPageReferenced(page
);
2144 page_remove_rmap(page
, true);
2146 add_mm_counter(mm
, MM_FILEPAGES
, -HPAGE_PMD_NR
);
2148 } else if (is_huge_zero_pmd(*pmd
)) {
2150 * FIXME: Do we want to invalidate secondary mmu by calling
2151 * mmu_notifier_invalidate_range() see comments below inside
2152 * __split_huge_pmd() ?
2154 * We are going from a zero huge page write protected to zero
2155 * small page also write protected so it does not seems useful
2156 * to invalidate secondary mmu at this time.
2158 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2161 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2162 pmd_migration
= is_pmd_migration_entry(*pmd
);
2163 if (pmd_migration
) {
2166 entry
= pmd_to_swp_entry(*pmd
);
2167 page
= pfn_to_page(swp_offset(entry
));
2170 page
= pmd_page(*pmd
);
2171 VM_BUG_ON_PAGE(!page_count(page
), page
);
2172 page_ref_add(page
, HPAGE_PMD_NR
- 1);
2173 write
= pmd_write(*pmd
);
2174 young
= pmd_young(*pmd
);
2175 dirty
= pmd_dirty(*pmd
);
2176 soft_dirty
= pmd_soft_dirty(*pmd
);
2178 pmdp_huge_split_prepare(vma
, haddr
, pmd
);
2179 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2180 pmd_populate(mm
, &_pmd
, pgtable
);
2182 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
2185 * Note that NUMA hinting access restrictions are not
2186 * transferred to avoid any possibility of altering
2187 * permissions across VMAs.
2189 if (freeze
|| pmd_migration
) {
2190 swp_entry_t swp_entry
;
2191 swp_entry
= make_migration_entry(page
+ i
, write
);
2192 entry
= swp_entry_to_pte(swp_entry
);
2194 entry
= pte_swp_mksoft_dirty(entry
);
2196 entry
= mk_pte(page
+ i
, READ_ONCE(vma
->vm_page_prot
));
2197 entry
= maybe_mkwrite(entry
, vma
);
2199 entry
= pte_wrprotect(entry
);
2201 entry
= pte_mkold(entry
);
2203 entry
= pte_mksoft_dirty(entry
);
2206 SetPageDirty(page
+ i
);
2207 pte
= pte_offset_map(&_pmd
, addr
);
2208 BUG_ON(!pte_none(*pte
));
2209 set_pte_at(mm
, addr
, pte
, entry
);
2210 atomic_inc(&page
[i
]._mapcount
);
2215 * Set PG_double_map before dropping compound_mapcount to avoid
2216 * false-negative page_mapped().
2218 if (compound_mapcount(page
) > 1 && !TestSetPageDoubleMap(page
)) {
2219 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2220 atomic_inc(&page
[i
]._mapcount
);
2223 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
2224 /* Last compound_mapcount is gone. */
2225 __dec_node_page_state(page
, NR_ANON_THPS
);
2226 if (TestClearPageDoubleMap(page
)) {
2227 /* No need in mapcount reference anymore */
2228 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2229 atomic_dec(&page
[i
]._mapcount
);
2233 smp_wmb(); /* make pte visible before pmd */
2235 * Up to this point the pmd is present and huge and userland has the
2236 * whole access to the hugepage during the split (which happens in
2237 * place). If we overwrite the pmd with the not-huge version pointing
2238 * to the pte here (which of course we could if all CPUs were bug
2239 * free), userland could trigger a small page size TLB miss on the
2240 * small sized TLB while the hugepage TLB entry is still established in
2241 * the huge TLB. Some CPU doesn't like that.
2242 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2243 * 383 on page 93. Intel should be safe but is also warns that it's
2244 * only safe if the permission and cache attributes of the two entries
2245 * loaded in the two TLB is identical (which should be the case here).
2246 * But it is generally safer to never allow small and huge TLB entries
2247 * for the same virtual address to be loaded simultaneously. So instead
2248 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2249 * current pmd notpresent (atomically because here the pmd_trans_huge
2250 * and pmd_trans_splitting must remain set at all times on the pmd
2251 * until the split is complete for this pmd), then we flush the SMP TLB
2252 * and finally we write the non-huge version of the pmd entry with
2255 pmdp_invalidate(vma
, haddr
, pmd
);
2256 pmd_populate(mm
, pmd
, pgtable
);
2259 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2260 page_remove_rmap(page
+ i
, false);
2266 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2267 unsigned long address
, bool freeze
, struct page
*page
)
2270 struct mm_struct
*mm
= vma
->vm_mm
;
2271 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
2273 mmu_notifier_invalidate_range_start(mm
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
2274 ptl
= pmd_lock(mm
, pmd
);
2277 * If caller asks to setup a migration entries, we need a page to check
2278 * pmd against. Otherwise we can end up replacing wrong page.
2280 VM_BUG_ON(freeze
&& !page
);
2281 if (page
&& page
!= pmd_page(*pmd
))
2284 if (pmd_trans_huge(*pmd
)) {
2285 page
= pmd_page(*pmd
);
2286 if (PageMlocked(page
))
2287 clear_page_mlock(page
);
2288 } else if (!(pmd_devmap(*pmd
) || is_pmd_migration_entry(*pmd
)))
2290 __split_huge_pmd_locked(vma
, pmd
, haddr
, freeze
);
2294 * No need to double call mmu_notifier->invalidate_range() callback.
2295 * They are 3 cases to consider inside __split_huge_pmd_locked():
2296 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2297 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2298 * fault will trigger a flush_notify before pointing to a new page
2299 * (it is fine if the secondary mmu keeps pointing to the old zero
2300 * page in the meantime)
2301 * 3) Split a huge pmd into pte pointing to the same page. No need
2302 * to invalidate secondary tlb entry they are all still valid.
2303 * any further changes to individual pte will notify. So no need
2304 * to call mmu_notifier->invalidate_range()
2306 mmu_notifier_invalidate_range_only_end(mm
, haddr
, haddr
+
2310 void split_huge_pmd_address(struct vm_area_struct
*vma
, unsigned long address
,
2311 bool freeze
, struct page
*page
)
2318 pgd
= pgd_offset(vma
->vm_mm
, address
);
2319 if (!pgd_present(*pgd
))
2322 p4d
= p4d_offset(pgd
, address
);
2323 if (!p4d_present(*p4d
))
2326 pud
= pud_offset(p4d
, address
);
2327 if (!pud_present(*pud
))
2330 pmd
= pmd_offset(pud
, address
);
2332 __split_huge_pmd(vma
, pmd
, address
, freeze
, page
);
2335 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2336 unsigned long start
,
2341 * If the new start address isn't hpage aligned and it could
2342 * previously contain an hugepage: check if we need to split
2345 if (start
& ~HPAGE_PMD_MASK
&&
2346 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2347 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2348 split_huge_pmd_address(vma
, start
, false, NULL
);
2351 * If the new end address isn't hpage aligned and it could
2352 * previously contain an hugepage: check if we need to split
2355 if (end
& ~HPAGE_PMD_MASK
&&
2356 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2357 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2358 split_huge_pmd_address(vma
, end
, false, NULL
);
2361 * If we're also updating the vma->vm_next->vm_start, if the new
2362 * vm_next->vm_start isn't page aligned and it could previously
2363 * contain an hugepage: check if we need to split an huge pmd.
2365 if (adjust_next
> 0) {
2366 struct vm_area_struct
*next
= vma
->vm_next
;
2367 unsigned long nstart
= next
->vm_start
;
2368 nstart
+= adjust_next
<< PAGE_SHIFT
;
2369 if (nstart
& ~HPAGE_PMD_MASK
&&
2370 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2371 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2372 split_huge_pmd_address(next
, nstart
, false, NULL
);
2376 static void unmap_page(struct page
*page
)
2378 enum ttu_flags ttu_flags
= TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
|
2379 TTU_RMAP_LOCKED
| TTU_SPLIT_HUGE_PMD
;
2382 VM_BUG_ON_PAGE(!PageHead(page
), page
);
2385 ttu_flags
|= TTU_SPLIT_FREEZE
;
2387 unmap_success
= try_to_unmap(page
, ttu_flags
);
2388 VM_BUG_ON_PAGE(!unmap_success
, page
);
2391 static void remap_page(struct page
*page
)
2394 if (PageTransHuge(page
)) {
2395 remove_migration_ptes(page
, page
, true);
2397 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2398 remove_migration_ptes(page
+ i
, page
+ i
, true);
2402 static void __split_huge_page_tail(struct page
*head
, int tail
,
2403 struct lruvec
*lruvec
, struct list_head
*list
)
2405 struct page
*page_tail
= head
+ tail
;
2407 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
2410 * Clone page flags before unfreezing refcount.
2412 * After successful get_page_unless_zero() might follow flags change,
2413 * for exmaple lock_page() which set PG_waiters.
2415 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
2416 page_tail
->flags
|= (head
->flags
&
2417 ((1L << PG_referenced
) |
2418 (1L << PG_swapbacked
) |
2419 (1L << PG_swapcache
) |
2420 (1L << PG_mlocked
) |
2421 (1L << PG_uptodate
) |
2424 (1L << PG_unevictable
) |
2427 /* ->mapping in first tail page is compound_mapcount */
2428 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
2430 page_tail
->mapping
= head
->mapping
;
2431 page_tail
->index
= head
->index
+ tail
;
2433 /* Page flags must be visible before we make the page non-compound. */
2437 * Clear PageTail before unfreezing page refcount.
2439 * After successful get_page_unless_zero() might follow put_page()
2440 * which needs correct compound_head().
2442 clear_compound_head(page_tail
);
2444 /* Finally unfreeze refcount. Additional reference from page cache. */
2445 page_ref_unfreeze(page_tail
, 1 + (!PageAnon(head
) ||
2446 PageSwapCache(head
)));
2448 if (page_is_young(head
))
2449 set_page_young(page_tail
);
2450 if (page_is_idle(head
))
2451 set_page_idle(page_tail
);
2453 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
2454 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
2457 static void __split_huge_page(struct page
*page
, struct list_head
*list
,
2458 pgoff_t end
, unsigned long flags
)
2460 struct page
*head
= compound_head(page
);
2461 struct zone
*zone
= page_zone(head
);
2462 struct lruvec
*lruvec
;
2465 lruvec
= mem_cgroup_page_lruvec(head
, zone
->zone_pgdat
);
2467 /* complete memcg works before add pages to LRU */
2468 mem_cgroup_split_huge_fixup(head
);
2470 for (i
= HPAGE_PMD_NR
- 1; i
>= 1; i
--) {
2471 __split_huge_page_tail(head
, i
, lruvec
, list
);
2472 /* Some pages can be beyond i_size: drop them from page cache */
2473 if (head
[i
].index
>= end
) {
2474 ClearPageDirty(head
+ i
);
2475 __delete_from_page_cache(head
+ i
, NULL
);
2476 if (IS_ENABLED(CONFIG_SHMEM
) && PageSwapBacked(head
))
2477 shmem_uncharge(head
->mapping
->host
, 1);
2482 ClearPageCompound(head
);
2484 split_page_owner(head
, HPAGE_PMD_ORDER
);
2486 /* See comment in __split_huge_page_tail() */
2487 if (PageAnon(head
)) {
2488 /* Additional pin to radix tree of swap cache */
2489 if (PageSwapCache(head
))
2490 page_ref_add(head
, 2);
2494 /* Additional pin to radix tree */
2495 page_ref_add(head
, 2);
2496 spin_unlock(&head
->mapping
->tree_lock
);
2499 spin_unlock_irqrestore(zone_lru_lock(page_zone(head
)), flags
);
2503 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2504 struct page
*subpage
= head
+ i
;
2505 if (subpage
== page
)
2507 unlock_page(subpage
);
2510 * Subpages may be freed if there wasn't any mapping
2511 * like if add_to_swap() is running on a lru page that
2512 * had its mapping zapped. And freeing these pages
2513 * requires taking the lru_lock so we do the put_page
2514 * of the tail pages after the split is complete.
2520 int total_mapcount(struct page
*page
)
2522 int i
, compound
, ret
;
2524 VM_BUG_ON_PAGE(PageTail(page
), page
);
2526 if (likely(!PageCompound(page
)))
2527 return atomic_read(&page
->_mapcount
) + 1;
2529 compound
= compound_mapcount(page
);
2533 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2534 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
2535 /* File pages has compound_mapcount included in _mapcount */
2536 if (!PageAnon(page
))
2537 return ret
- compound
* HPAGE_PMD_NR
;
2538 if (PageDoubleMap(page
))
2539 ret
-= HPAGE_PMD_NR
;
2544 * This calculates accurately how many mappings a transparent hugepage
2545 * has (unlike page_mapcount() which isn't fully accurate). This full
2546 * accuracy is primarily needed to know if copy-on-write faults can
2547 * reuse the page and change the mapping to read-write instead of
2548 * copying them. At the same time this returns the total_mapcount too.
2550 * The function returns the highest mapcount any one of the subpages
2551 * has. If the return value is one, even if different processes are
2552 * mapping different subpages of the transparent hugepage, they can
2553 * all reuse it, because each process is reusing a different subpage.
2555 * The total_mapcount is instead counting all virtual mappings of the
2556 * subpages. If the total_mapcount is equal to "one", it tells the
2557 * caller all mappings belong to the same "mm" and in turn the
2558 * anon_vma of the transparent hugepage can become the vma->anon_vma
2559 * local one as no other process may be mapping any of the subpages.
2561 * It would be more accurate to replace page_mapcount() with
2562 * page_trans_huge_mapcount(), however we only use
2563 * page_trans_huge_mapcount() in the copy-on-write faults where we
2564 * need full accuracy to avoid breaking page pinning, because
2565 * page_trans_huge_mapcount() is slower than page_mapcount().
2567 int page_trans_huge_mapcount(struct page
*page
, int *total_mapcount
)
2569 int i
, ret
, _total_mapcount
, mapcount
;
2571 /* hugetlbfs shouldn't call it */
2572 VM_BUG_ON_PAGE(PageHuge(page
), page
);
2574 if (likely(!PageTransCompound(page
))) {
2575 mapcount
= atomic_read(&page
->_mapcount
) + 1;
2577 *total_mapcount
= mapcount
;
2581 page
= compound_head(page
);
2583 _total_mapcount
= ret
= 0;
2584 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2585 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
2586 ret
= max(ret
, mapcount
);
2587 _total_mapcount
+= mapcount
;
2589 if (PageDoubleMap(page
)) {
2591 _total_mapcount
-= HPAGE_PMD_NR
;
2593 mapcount
= compound_mapcount(page
);
2595 _total_mapcount
+= mapcount
;
2597 *total_mapcount
= _total_mapcount
;
2601 /* Racy check whether the huge page can be split */
2602 bool can_split_huge_page(struct page
*page
, int *pextra_pins
)
2606 /* Additional pins from radix tree */
2608 extra_pins
= PageSwapCache(page
) ? HPAGE_PMD_NR
: 0;
2610 extra_pins
= HPAGE_PMD_NR
;
2612 *pextra_pins
= extra_pins
;
2613 return total_mapcount(page
) == page_count(page
) - extra_pins
- 1;
2617 * This function splits huge page into normal pages. @page can point to any
2618 * subpage of huge page to split. Split doesn't change the position of @page.
2620 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2621 * The huge page must be locked.
2623 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2625 * Both head page and tail pages will inherit mapping, flags, and so on from
2628 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2629 * they are not mapped.
2631 * Returns 0 if the hugepage is split successfully.
2632 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2635 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
2637 struct page
*head
= compound_head(page
);
2638 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(head
));
2639 struct anon_vma
*anon_vma
= NULL
;
2640 struct address_space
*mapping
= NULL
;
2641 int count
, mapcount
, extra_pins
, ret
;
2643 unsigned long flags
;
2646 VM_BUG_ON_PAGE(is_huge_zero_page(page
), page
);
2647 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2648 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
2650 if (PageWriteback(page
))
2653 if (PageAnon(head
)) {
2655 * The caller does not necessarily hold an mmap_sem that would
2656 * prevent the anon_vma disappearing so we first we take a
2657 * reference to it and then lock the anon_vma for write. This
2658 * is similar to page_lock_anon_vma_read except the write lock
2659 * is taken to serialise against parallel split or collapse
2662 anon_vma
= page_get_anon_vma(head
);
2669 anon_vma_lock_write(anon_vma
);
2671 mapping
= head
->mapping
;
2680 i_mmap_lock_read(mapping
);
2683 *__split_huge_page() may need to trim off pages beyond EOF:
2684 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2685 * which cannot be nested inside the page tree lock. So note
2686 * end now: i_size itself may be changed at any moment, but
2687 * head page lock is good enough to serialize the trimming.
2689 end
= DIV_ROUND_UP(i_size_read(mapping
->host
), PAGE_SIZE
);
2693 * Racy check if we can split the page, before unmap_page() will
2696 if (!can_split_huge_page(head
, &extra_pins
)) {
2701 mlocked
= PageMlocked(page
);
2703 VM_BUG_ON_PAGE(compound_mapcount(head
), head
);
2705 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2709 /* prevent PageLRU to go away from under us, and freeze lru stats */
2710 spin_lock_irqsave(zone_lru_lock(page_zone(head
)), flags
);
2715 spin_lock(&mapping
->tree_lock
);
2716 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
2719 * Check if the head page is present in radix tree.
2720 * We assume all tail are present too, if head is there.
2722 if (radix_tree_deref_slot_protected(pslot
,
2723 &mapping
->tree_lock
) != head
)
2727 /* Prevent deferred_split_scan() touching ->_refcount */
2728 spin_lock(&pgdata
->split_queue_lock
);
2729 count
= page_count(head
);
2730 mapcount
= total_mapcount(head
);
2731 if (!mapcount
&& page_ref_freeze(head
, 1 + extra_pins
)) {
2732 if (!list_empty(page_deferred_list(head
))) {
2733 pgdata
->split_queue_len
--;
2734 list_del(page_deferred_list(head
));
2737 __dec_node_page_state(page
, NR_SHMEM_THPS
);
2738 spin_unlock(&pgdata
->split_queue_lock
);
2739 __split_huge_page(page
, list
, end
, flags
);
2740 if (PageSwapCache(head
)) {
2741 swp_entry_t entry
= { .val
= page_private(head
) };
2743 ret
= split_swap_cluster(entry
);
2747 if (IS_ENABLED(CONFIG_DEBUG_VM
) && mapcount
) {
2748 pr_alert("total_mapcount: %u, page_count(): %u\n",
2751 dump_page(head
, NULL
);
2752 dump_page(page
, "total_mapcount(head) > 0");
2755 spin_unlock(&pgdata
->split_queue_lock
);
2757 spin_unlock(&mapping
->tree_lock
);
2758 spin_unlock_irqrestore(zone_lru_lock(page_zone(head
)), flags
);
2765 anon_vma_unlock_write(anon_vma
);
2766 put_anon_vma(anon_vma
);
2769 i_mmap_unlock_read(mapping
);
2771 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
2775 void free_transhuge_page(struct page
*page
)
2777 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
2778 unsigned long flags
;
2780 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2781 if (!list_empty(page_deferred_list(page
))) {
2782 pgdata
->split_queue_len
--;
2783 list_del(page_deferred_list(page
));
2785 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2786 free_compound_page(page
);
2789 void deferred_split_huge_page(struct page
*page
)
2791 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(page
));
2792 unsigned long flags
;
2794 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
2796 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2797 if (list_empty(page_deferred_list(page
))) {
2798 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
2799 list_add_tail(page_deferred_list(page
), &pgdata
->split_queue
);
2800 pgdata
->split_queue_len
++;
2802 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2805 static unsigned long deferred_split_count(struct shrinker
*shrink
,
2806 struct shrink_control
*sc
)
2808 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2809 return READ_ONCE(pgdata
->split_queue_len
);
2812 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
2813 struct shrink_control
*sc
)
2815 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2816 unsigned long flags
;
2817 LIST_HEAD(list
), *pos
, *next
;
2821 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2822 /* Take pin on all head pages to avoid freeing them under us */
2823 list_for_each_safe(pos
, next
, &pgdata
->split_queue
) {
2824 page
= list_entry((void *)pos
, struct page
, mapping
);
2825 page
= compound_head(page
);
2826 if (get_page_unless_zero(page
)) {
2827 list_move(page_deferred_list(page
), &list
);
2829 /* We lost race with put_compound_page() */
2830 list_del_init(page_deferred_list(page
));
2831 pgdata
->split_queue_len
--;
2833 if (!--sc
->nr_to_scan
)
2836 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2838 list_for_each_safe(pos
, next
, &list
) {
2839 page
= list_entry((void *)pos
, struct page
, mapping
);
2840 if (!trylock_page(page
))
2842 /* split_huge_page() removes page from list on success */
2843 if (!split_huge_page(page
))
2850 spin_lock_irqsave(&pgdata
->split_queue_lock
, flags
);
2851 list_splice_tail(&list
, &pgdata
->split_queue
);
2852 spin_unlock_irqrestore(&pgdata
->split_queue_lock
, flags
);
2855 * Stop shrinker if we didn't split any page, but the queue is empty.
2856 * This can happen if pages were freed under us.
2858 if (!split
&& list_empty(&pgdata
->split_queue
))
2863 static struct shrinker deferred_split_shrinker
= {
2864 .count_objects
= deferred_split_count
,
2865 .scan_objects
= deferred_split_scan
,
2866 .seeks
= DEFAULT_SEEKS
,
2867 .flags
= SHRINKER_NUMA_AWARE
,
2870 #ifdef CONFIG_DEBUG_FS
2871 static int split_huge_pages_set(void *data
, u64 val
)
2875 unsigned long pfn
, max_zone_pfn
;
2876 unsigned long total
= 0, split
= 0;
2881 for_each_populated_zone(zone
) {
2882 max_zone_pfn
= zone_end_pfn(zone
);
2883 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
2884 if (!pfn_valid(pfn
))
2887 page
= pfn_to_page(pfn
);
2888 if (!get_page_unless_zero(page
))
2891 if (zone
!= page_zone(page
))
2894 if (!PageHead(page
) || PageHuge(page
) || !PageLRU(page
))
2899 if (!split_huge_page(page
))
2907 pr_info("%lu of %lu THP split\n", split
, total
);
2911 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops
, NULL
, split_huge_pages_set
,
2914 static int __init
split_huge_pages_debugfs(void)
2918 ret
= debugfs_create_file("split_huge_pages", 0200, NULL
, NULL
,
2919 &split_huge_pages_fops
);
2921 pr_warn("Failed to create split_huge_pages in debugfs");
2924 late_initcall(split_huge_pages_debugfs
);
2927 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2928 void set_pmd_migration_entry(struct page_vma_mapped_walk
*pvmw
,
2931 struct vm_area_struct
*vma
= pvmw
->vma
;
2932 struct mm_struct
*mm
= vma
->vm_mm
;
2933 unsigned long address
= pvmw
->address
;
2938 if (!(pvmw
->pmd
&& !pvmw
->pte
))
2941 flush_cache_range(vma
, address
, address
+ HPAGE_PMD_SIZE
);
2942 pmdval
= *pvmw
->pmd
;
2943 pmdp_invalidate(vma
, address
, pvmw
->pmd
);
2944 if (pmd_dirty(pmdval
))
2945 set_page_dirty(page
);
2946 entry
= make_migration_entry(page
, pmd_write(pmdval
));
2947 pmdswp
= swp_entry_to_pmd(entry
);
2948 if (pmd_soft_dirty(pmdval
))
2949 pmdswp
= pmd_swp_mksoft_dirty(pmdswp
);
2950 set_pmd_at(mm
, address
, pvmw
->pmd
, pmdswp
);
2951 page_remove_rmap(page
, true);
2955 void remove_migration_pmd(struct page_vma_mapped_walk
*pvmw
, struct page
*new)
2957 struct vm_area_struct
*vma
= pvmw
->vma
;
2958 struct mm_struct
*mm
= vma
->vm_mm
;
2959 unsigned long address
= pvmw
->address
;
2960 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
2964 if (!(pvmw
->pmd
&& !pvmw
->pte
))
2967 entry
= pmd_to_swp_entry(*pvmw
->pmd
);
2969 pmde
= pmd_mkold(mk_huge_pmd(new, vma
->vm_page_prot
));
2970 if (pmd_swp_soft_dirty(*pvmw
->pmd
))
2971 pmde
= pmd_mksoft_dirty(pmde
);
2972 if (is_write_migration_entry(entry
))
2973 pmde
= maybe_pmd_mkwrite(pmde
, vma
);
2975 flush_cache_range(vma
, mmun_start
, mmun_start
+ HPAGE_PMD_SIZE
);
2976 page_add_anon_rmap(new, vma
, mmun_start
, true);
2977 set_pmd_at(mm
, mmun_start
, pvmw
->pmd
, pmde
);
2978 if ((vma
->vm_flags
& VM_LOCKED
) && !PageDoubleMap(new))
2979 mlock_vma_page(new);
2980 update_mmu_cache_pmd(vma
, address
, pvmw
->pmd
);