1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2009 Red Hat, Inc.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/sched.h>
10 #include <linux/sched/mm.h>
11 #include <linux/sched/coredump.h>
12 #include <linux/sched/numa_balancing.h>
13 #include <linux/highmem.h>
14 #include <linux/hugetlb.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/rmap.h>
17 #include <linux/swap.h>
18 #include <linux/shrinker.h>
19 #include <linux/mm_inline.h>
20 #include <linux/swapops.h>
21 #include <linux/dax.h>
22 #include <linux/khugepaged.h>
23 #include <linux/freezer.h>
24 #include <linux/pfn_t.h>
25 #include <linux/mman.h>
26 #include <linux/memremap.h>
27 #include <linux/pagemap.h>
28 #include <linux/debugfs.h>
29 #include <linux/migrate.h>
30 #include <linux/hashtable.h>
31 #include <linux/userfaultfd_k.h>
32 #include <linux/page_idle.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/oom.h>
35 #include <linux/numa.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
;
65 unsigned long huge_zero_pfn __read_mostly
= ~0UL;
67 static inline bool file_thp_enabled(struct vm_area_struct
*vma
)
69 return transhuge_vma_enabled(vma
, vma
->vm_flags
) && vma
->vm_file
&&
70 !inode_is_open_for_write(vma
->vm_file
->f_inode
) &&
71 (vma
->vm_flags
& VM_EXEC
);
74 bool transparent_hugepage_active(struct vm_area_struct
*vma
)
76 /* The addr is used to check if the vma size fits */
77 unsigned long addr
= (vma
->vm_end
& HPAGE_PMD_MASK
) - HPAGE_PMD_SIZE
;
79 if (!transhuge_vma_suitable(vma
, addr
))
81 if (vma_is_anonymous(vma
))
82 return __transparent_hugepage_enabled(vma
);
83 if (vma_is_shmem(vma
))
84 return shmem_huge_enabled(vma
);
85 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS
))
86 return file_thp_enabled(vma
);
91 static bool get_huge_zero_page(void)
93 struct page
*zero_page
;
95 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
98 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
101 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
104 count_vm_event(THP_ZERO_PAGE_ALLOC
);
106 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
108 __free_pages(zero_page
, compound_order(zero_page
));
111 WRITE_ONCE(huge_zero_pfn
, page_to_pfn(zero_page
));
113 /* We take additional reference here. It will be put back by shrinker */
114 atomic_set(&huge_zero_refcount
, 2);
119 static void put_huge_zero_page(void)
122 * Counter should never go to zero here. Only shrinker can put
125 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
128 struct page
*mm_get_huge_zero_page(struct mm_struct
*mm
)
130 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
131 return READ_ONCE(huge_zero_page
);
133 if (!get_huge_zero_page())
136 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
137 put_huge_zero_page();
139 return READ_ONCE(huge_zero_page
);
142 void mm_put_huge_zero_page(struct mm_struct
*mm
)
144 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
145 put_huge_zero_page();
148 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
149 struct shrink_control
*sc
)
151 /* we can free zero page only if last reference remains */
152 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
155 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
156 struct shrink_control
*sc
)
158 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
159 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
160 BUG_ON(zero_page
== NULL
);
161 WRITE_ONCE(huge_zero_pfn
, ~0UL);
162 __free_pages(zero_page
, compound_order(zero_page
));
169 static struct shrinker huge_zero_page_shrinker
= {
170 .count_objects
= shrink_huge_zero_page_count
,
171 .scan_objects
= shrink_huge_zero_page_scan
,
172 .seeks
= DEFAULT_SEEKS
,
176 static ssize_t
enabled_show(struct kobject
*kobj
,
177 struct kobj_attribute
*attr
, char *buf
)
181 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
))
182 output
= "[always] madvise never";
183 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
184 &transparent_hugepage_flags
))
185 output
= "always [madvise] never";
187 output
= "always madvise [never]";
189 return sysfs_emit(buf
, "%s\n", output
);
192 static ssize_t
enabled_store(struct kobject
*kobj
,
193 struct kobj_attribute
*attr
,
194 const char *buf
, size_t count
)
198 if (sysfs_streq(buf
, "always")) {
199 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
200 set_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
201 } else if (sysfs_streq(buf
, "madvise")) {
202 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
203 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
204 } else if (sysfs_streq(buf
, "never")) {
205 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
206 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
211 int err
= start_stop_khugepaged();
217 static struct kobj_attribute enabled_attr
=
218 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
220 ssize_t
single_hugepage_flag_show(struct kobject
*kobj
,
221 struct kobj_attribute
*attr
, char *buf
,
222 enum transparent_hugepage_flag flag
)
224 return sysfs_emit(buf
, "%d\n",
225 !!test_bit(flag
, &transparent_hugepage_flags
));
228 ssize_t
single_hugepage_flag_store(struct kobject
*kobj
,
229 struct kobj_attribute
*attr
,
230 const char *buf
, size_t count
,
231 enum transparent_hugepage_flag flag
)
236 ret
= kstrtoul(buf
, 10, &value
);
243 set_bit(flag
, &transparent_hugepage_flags
);
245 clear_bit(flag
, &transparent_hugepage_flags
);
250 static ssize_t
defrag_show(struct kobject
*kobj
,
251 struct kobj_attribute
*attr
, char *buf
)
255 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
,
256 &transparent_hugepage_flags
))
257 output
= "[always] defer defer+madvise madvise never";
258 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
,
259 &transparent_hugepage_flags
))
260 output
= "always [defer] defer+madvise madvise never";
261 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
,
262 &transparent_hugepage_flags
))
263 output
= "always defer [defer+madvise] madvise never";
264 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
,
265 &transparent_hugepage_flags
))
266 output
= "always defer defer+madvise [madvise] never";
268 output
= "always defer defer+madvise madvise [never]";
270 return sysfs_emit(buf
, "%s\n", output
);
273 static ssize_t
defrag_store(struct kobject
*kobj
,
274 struct kobj_attribute
*attr
,
275 const char *buf
, size_t count
)
277 if (sysfs_streq(buf
, "always")) {
278 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
279 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
280 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
281 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
282 } else if (sysfs_streq(buf
, "defer+madvise")) {
283 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
284 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
285 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
286 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
287 } else if (sysfs_streq(buf
, "defer")) {
288 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
289 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
290 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
291 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
292 } else if (sysfs_streq(buf
, "madvise")) {
293 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
294 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
295 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
296 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
297 } else if (sysfs_streq(buf
, "never")) {
298 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
299 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
300 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
301 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
307 static struct kobj_attribute defrag_attr
=
308 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
310 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
311 struct kobj_attribute
*attr
, char *buf
)
313 return single_hugepage_flag_show(kobj
, attr
, buf
,
314 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
316 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
317 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
319 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
320 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
322 static struct kobj_attribute use_zero_page_attr
=
323 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
325 static ssize_t
hpage_pmd_size_show(struct kobject
*kobj
,
326 struct kobj_attribute
*attr
, char *buf
)
328 return sysfs_emit(buf
, "%lu\n", HPAGE_PMD_SIZE
);
330 static struct kobj_attribute hpage_pmd_size_attr
=
331 __ATTR_RO(hpage_pmd_size
);
333 static struct attribute
*hugepage_attr
[] = {
336 &use_zero_page_attr
.attr
,
337 &hpage_pmd_size_attr
.attr
,
339 &shmem_enabled_attr
.attr
,
344 static const struct attribute_group hugepage_attr_group
= {
345 .attrs
= hugepage_attr
,
348 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
352 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
353 if (unlikely(!*hugepage_kobj
)) {
354 pr_err("failed to create transparent hugepage kobject\n");
358 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
360 pr_err("failed to register transparent hugepage group\n");
364 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
366 pr_err("failed to register transparent hugepage group\n");
367 goto remove_hp_group
;
373 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
375 kobject_put(*hugepage_kobj
);
379 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
381 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
382 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
383 kobject_put(hugepage_kobj
);
386 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
391 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
394 #endif /* CONFIG_SYSFS */
396 static int __init
hugepage_init(void)
399 struct kobject
*hugepage_kobj
;
401 if (!has_transparent_hugepage()) {
403 * Hardware doesn't support hugepages, hence disable
406 transparent_hugepage_flags
= 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX
;
411 * hugepages can't be allocated by the buddy allocator
413 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
>= MAX_ORDER
);
415 * we use page->mapping and page->index in second tail page
416 * as list_head: assuming THP order >= 2
418 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
420 err
= hugepage_init_sysfs(&hugepage_kobj
);
424 err
= khugepaged_init();
428 err
= register_shrinker(&huge_zero_page_shrinker
);
430 goto err_hzp_shrinker
;
431 err
= register_shrinker(&deferred_split_shrinker
);
433 goto err_split_shrinker
;
436 * By default disable transparent hugepages on smaller systems,
437 * where the extra memory used could hurt more than TLB overhead
438 * is likely to save. The admin can still enable it through /sys.
440 if (totalram_pages() < (512 << (20 - PAGE_SHIFT
))) {
441 transparent_hugepage_flags
= 0;
445 err
= start_stop_khugepaged();
451 unregister_shrinker(&deferred_split_shrinker
);
453 unregister_shrinker(&huge_zero_page_shrinker
);
455 khugepaged_destroy();
457 hugepage_exit_sysfs(hugepage_kobj
);
461 subsys_initcall(hugepage_init
);
463 static int __init
setup_transparent_hugepage(char *str
)
468 if (!strcmp(str
, "always")) {
469 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
470 &transparent_hugepage_flags
);
471 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
472 &transparent_hugepage_flags
);
474 } else if (!strcmp(str
, "madvise")) {
475 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
476 &transparent_hugepage_flags
);
477 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
478 &transparent_hugepage_flags
);
480 } else if (!strcmp(str
, "never")) {
481 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
482 &transparent_hugepage_flags
);
483 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
484 &transparent_hugepage_flags
);
489 pr_warn("transparent_hugepage= cannot parse, ignored\n");
492 __setup("transparent_hugepage=", setup_transparent_hugepage
);
494 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
496 if (likely(vma
->vm_flags
& VM_WRITE
))
497 pmd
= pmd_mkwrite(pmd
);
502 static inline struct deferred_split
*get_deferred_split_queue(struct page
*page
)
504 struct mem_cgroup
*memcg
= page_memcg(compound_head(page
));
505 struct pglist_data
*pgdat
= NODE_DATA(page_to_nid(page
));
508 return &memcg
->deferred_split_queue
;
510 return &pgdat
->deferred_split_queue
;
513 static inline struct deferred_split
*get_deferred_split_queue(struct page
*page
)
515 struct pglist_data
*pgdat
= NODE_DATA(page_to_nid(page
));
517 return &pgdat
->deferred_split_queue
;
521 void prep_transhuge_page(struct page
*page
)
524 * we use page->mapping and page->indexlru in second tail page
525 * as list_head: assuming THP order >= 2
528 INIT_LIST_HEAD(page_deferred_list(page
));
529 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
532 bool is_transparent_hugepage(struct page
*page
)
534 if (!PageCompound(page
))
537 page
= compound_head(page
);
538 return is_huge_zero_page(page
) ||
539 page
[1].compound_dtor
== TRANSHUGE_PAGE_DTOR
;
541 EXPORT_SYMBOL_GPL(is_transparent_hugepage
);
543 static unsigned long __thp_get_unmapped_area(struct file
*filp
,
544 unsigned long addr
, unsigned long len
,
545 loff_t off
, unsigned long flags
, unsigned long size
)
547 loff_t off_end
= off
+ len
;
548 loff_t off_align
= round_up(off
, size
);
549 unsigned long len_pad
, ret
;
551 if (off_end
<= off_align
|| (off_end
- off_align
) < size
)
554 len_pad
= len
+ size
;
555 if (len_pad
< len
|| (off
+ len_pad
) < off
)
558 ret
= current
->mm
->get_unmapped_area(filp
, addr
, len_pad
,
559 off
>> PAGE_SHIFT
, flags
);
562 * The failure might be due to length padding. The caller will retry
563 * without the padding.
565 if (IS_ERR_VALUE(ret
))
569 * Do not try to align to THP boundary if allocation at the address
575 ret
+= (off
- ret
) & (size
- 1);
579 unsigned long thp_get_unmapped_area(struct file
*filp
, unsigned long addr
,
580 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
583 loff_t off
= (loff_t
)pgoff
<< PAGE_SHIFT
;
585 if (!IS_DAX(filp
->f_mapping
->host
) || !IS_ENABLED(CONFIG_FS_DAX_PMD
))
588 ret
= __thp_get_unmapped_area(filp
, addr
, len
, off
, flags
, PMD_SIZE
);
592 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
594 EXPORT_SYMBOL_GPL(thp_get_unmapped_area
);
596 static vm_fault_t
__do_huge_pmd_anonymous_page(struct vm_fault
*vmf
,
597 struct page
*page
, gfp_t gfp
)
599 struct vm_area_struct
*vma
= vmf
->vma
;
601 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
604 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
606 if (mem_cgroup_charge(page
, vma
->vm_mm
, gfp
)) {
608 count_vm_event(THP_FAULT_FALLBACK
);
609 count_vm_event(THP_FAULT_FALLBACK_CHARGE
);
610 return VM_FAULT_FALLBACK
;
612 cgroup_throttle_swaprate(page
, gfp
);
614 pgtable
= pte_alloc_one(vma
->vm_mm
);
615 if (unlikely(!pgtable
)) {
620 clear_huge_page(page
, vmf
->address
, HPAGE_PMD_NR
);
622 * The memory barrier inside __SetPageUptodate makes sure that
623 * clear_huge_page writes become visible before the set_pmd_at()
626 __SetPageUptodate(page
);
628 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
629 if (unlikely(!pmd_none(*vmf
->pmd
))) {
634 ret
= check_stable_address_space(vma
->vm_mm
);
638 /* Deliver the page fault to userland */
639 if (userfaultfd_missing(vma
)) {
640 spin_unlock(vmf
->ptl
);
642 pte_free(vma
->vm_mm
, pgtable
);
643 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
644 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
648 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
649 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
650 page_add_new_anon_rmap(page
, vma
, haddr
, true);
651 lru_cache_add_inactive_or_unevictable(page
, vma
);
652 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
653 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
654 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
655 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
656 mm_inc_nr_ptes(vma
->vm_mm
);
657 spin_unlock(vmf
->ptl
);
658 count_vm_event(THP_FAULT_ALLOC
);
659 count_memcg_event_mm(vma
->vm_mm
, THP_FAULT_ALLOC
);
664 spin_unlock(vmf
->ptl
);
667 pte_free(vma
->vm_mm
, pgtable
);
674 * always: directly stall for all thp allocations
675 * defer: wake kswapd and fail if not immediately available
676 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
677 * fail if not immediately available
678 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
680 * never: never stall for any thp allocation
682 gfp_t
vma_thp_gfp_mask(struct vm_area_struct
*vma
)
684 const bool vma_madvised
= vma
&& (vma
->vm_flags
& VM_HUGEPAGE
);
686 /* Always do synchronous compaction */
687 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
688 return GFP_TRANSHUGE
| (vma_madvised
? 0 : __GFP_NORETRY
);
690 /* Kick kcompactd and fail quickly */
691 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
692 return GFP_TRANSHUGE_LIGHT
| __GFP_KSWAPD_RECLAIM
;
694 /* Synchronous compaction if madvised, otherwise kick kcompactd */
695 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
696 return GFP_TRANSHUGE_LIGHT
|
697 (vma_madvised
? __GFP_DIRECT_RECLAIM
:
698 __GFP_KSWAPD_RECLAIM
);
700 /* Only do synchronous compaction if madvised */
701 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
702 return GFP_TRANSHUGE_LIGHT
|
703 (vma_madvised
? __GFP_DIRECT_RECLAIM
: 0);
705 return GFP_TRANSHUGE_LIGHT
;
708 /* Caller must hold page table lock. */
709 static void set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
710 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
711 struct page
*zero_page
)
716 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
717 entry
= pmd_mkhuge(entry
);
719 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
720 set_pmd_at(mm
, haddr
, pmd
, entry
);
724 vm_fault_t
do_huge_pmd_anonymous_page(struct vm_fault
*vmf
)
726 struct vm_area_struct
*vma
= vmf
->vma
;
729 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
731 if (!transhuge_vma_suitable(vma
, haddr
))
732 return VM_FAULT_FALLBACK
;
733 if (unlikely(anon_vma_prepare(vma
)))
735 if (unlikely(khugepaged_enter(vma
, vma
->vm_flags
)))
737 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
738 !mm_forbids_zeropage(vma
->vm_mm
) &&
739 transparent_hugepage_use_zero_page()) {
741 struct page
*zero_page
;
743 pgtable
= pte_alloc_one(vma
->vm_mm
);
744 if (unlikely(!pgtable
))
746 zero_page
= mm_get_huge_zero_page(vma
->vm_mm
);
747 if (unlikely(!zero_page
)) {
748 pte_free(vma
->vm_mm
, pgtable
);
749 count_vm_event(THP_FAULT_FALLBACK
);
750 return VM_FAULT_FALLBACK
;
752 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
754 if (pmd_none(*vmf
->pmd
)) {
755 ret
= check_stable_address_space(vma
->vm_mm
);
757 spin_unlock(vmf
->ptl
);
758 pte_free(vma
->vm_mm
, pgtable
);
759 } else if (userfaultfd_missing(vma
)) {
760 spin_unlock(vmf
->ptl
);
761 pte_free(vma
->vm_mm
, pgtable
);
762 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
763 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
765 set_huge_zero_page(pgtable
, vma
->vm_mm
, vma
,
766 haddr
, vmf
->pmd
, zero_page
);
767 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
768 spin_unlock(vmf
->ptl
);
771 spin_unlock(vmf
->ptl
);
772 pte_free(vma
->vm_mm
, pgtable
);
776 gfp
= vma_thp_gfp_mask(vma
);
777 page
= alloc_hugepage_vma(gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
778 if (unlikely(!page
)) {
779 count_vm_event(THP_FAULT_FALLBACK
);
780 return VM_FAULT_FALLBACK
;
782 prep_transhuge_page(page
);
783 return __do_huge_pmd_anonymous_page(vmf
, page
, gfp
);
786 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
787 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
,
790 struct mm_struct
*mm
= vma
->vm_mm
;
794 ptl
= pmd_lock(mm
, pmd
);
795 if (!pmd_none(*pmd
)) {
797 if (pmd_pfn(*pmd
) != pfn_t_to_pfn(pfn
)) {
798 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd
));
801 entry
= pmd_mkyoung(*pmd
);
802 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
803 if (pmdp_set_access_flags(vma
, addr
, pmd
, entry
, 1))
804 update_mmu_cache_pmd(vma
, addr
, pmd
);
810 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
811 if (pfn_t_devmap(pfn
))
812 entry
= pmd_mkdevmap(entry
);
814 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
815 entry
= maybe_pmd_mkwrite(entry
, vma
);
819 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
824 set_pmd_at(mm
, addr
, pmd
, entry
);
825 update_mmu_cache_pmd(vma
, addr
, pmd
);
830 pte_free(mm
, pgtable
);
834 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
835 * @vmf: Structure describing the fault
836 * @pfn: pfn to insert
837 * @pgprot: page protection to use
838 * @write: whether it's a write fault
840 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
841 * also consult the vmf_insert_mixed_prot() documentation when
842 * @pgprot != @vmf->vma->vm_page_prot.
844 * Return: vm_fault_t value.
846 vm_fault_t
vmf_insert_pfn_pmd_prot(struct vm_fault
*vmf
, pfn_t pfn
,
847 pgprot_t pgprot
, bool write
)
849 unsigned long addr
= vmf
->address
& PMD_MASK
;
850 struct vm_area_struct
*vma
= vmf
->vma
;
851 pgtable_t pgtable
= NULL
;
854 * If we had pmd_special, we could avoid all these restrictions,
855 * but we need to be consistent with PTEs and architectures that
856 * can't support a 'special' bit.
858 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) &&
860 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
861 (VM_PFNMAP
|VM_MIXEDMAP
));
862 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
864 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
865 return VM_FAULT_SIGBUS
;
867 if (arch_needs_pgtable_deposit()) {
868 pgtable
= pte_alloc_one(vma
->vm_mm
);
873 track_pfn_insert(vma
, &pgprot
, pfn
);
875 insert_pfn_pmd(vma
, addr
, vmf
->pmd
, pfn
, pgprot
, write
, pgtable
);
876 return VM_FAULT_NOPAGE
;
878 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot
);
880 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
881 static pud_t
maybe_pud_mkwrite(pud_t pud
, struct vm_area_struct
*vma
)
883 if (likely(vma
->vm_flags
& VM_WRITE
))
884 pud
= pud_mkwrite(pud
);
888 static void insert_pfn_pud(struct vm_area_struct
*vma
, unsigned long addr
,
889 pud_t
*pud
, pfn_t pfn
, pgprot_t prot
, bool write
)
891 struct mm_struct
*mm
= vma
->vm_mm
;
895 ptl
= pud_lock(mm
, pud
);
896 if (!pud_none(*pud
)) {
898 if (pud_pfn(*pud
) != pfn_t_to_pfn(pfn
)) {
899 WARN_ON_ONCE(!is_huge_zero_pud(*pud
));
902 entry
= pud_mkyoung(*pud
);
903 entry
= maybe_pud_mkwrite(pud_mkdirty(entry
), vma
);
904 if (pudp_set_access_flags(vma
, addr
, pud
, entry
, 1))
905 update_mmu_cache_pud(vma
, addr
, pud
);
910 entry
= pud_mkhuge(pfn_t_pud(pfn
, prot
));
911 if (pfn_t_devmap(pfn
))
912 entry
= pud_mkdevmap(entry
);
914 entry
= pud_mkyoung(pud_mkdirty(entry
));
915 entry
= maybe_pud_mkwrite(entry
, vma
);
917 set_pud_at(mm
, addr
, pud
, entry
);
918 update_mmu_cache_pud(vma
, addr
, pud
);
925 * vmf_insert_pfn_pud_prot - insert a pud size pfn
926 * @vmf: Structure describing the fault
927 * @pfn: pfn to insert
928 * @pgprot: page protection to use
929 * @write: whether it's a write fault
931 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
932 * also consult the vmf_insert_mixed_prot() documentation when
933 * @pgprot != @vmf->vma->vm_page_prot.
935 * Return: vm_fault_t value.
937 vm_fault_t
vmf_insert_pfn_pud_prot(struct vm_fault
*vmf
, pfn_t pfn
,
938 pgprot_t pgprot
, bool write
)
940 unsigned long addr
= vmf
->address
& PUD_MASK
;
941 struct vm_area_struct
*vma
= vmf
->vma
;
944 * If we had pud_special, we could avoid all these restrictions,
945 * but we need to be consistent with PTEs and architectures that
946 * can't support a 'special' bit.
948 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) &&
950 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
951 (VM_PFNMAP
|VM_MIXEDMAP
));
952 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
954 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
955 return VM_FAULT_SIGBUS
;
957 track_pfn_insert(vma
, &pgprot
, pfn
);
959 insert_pfn_pud(vma
, addr
, vmf
->pud
, pfn
, pgprot
, write
);
960 return VM_FAULT_NOPAGE
;
962 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot
);
963 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
965 static void touch_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
966 pmd_t
*pmd
, int flags
)
970 _pmd
= pmd_mkyoung(*pmd
);
971 if (flags
& FOLL_WRITE
)
972 _pmd
= pmd_mkdirty(_pmd
);
973 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
974 pmd
, _pmd
, flags
& FOLL_WRITE
))
975 update_mmu_cache_pmd(vma
, addr
, pmd
);
978 struct page
*follow_devmap_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
979 pmd_t
*pmd
, int flags
, struct dev_pagemap
**pgmap
)
981 unsigned long pfn
= pmd_pfn(*pmd
);
982 struct mm_struct
*mm
= vma
->vm_mm
;
985 assert_spin_locked(pmd_lockptr(mm
, pmd
));
988 * When we COW a devmap PMD entry, we split it into PTEs, so we should
989 * not be in this function with `flags & FOLL_COW` set.
991 WARN_ONCE(flags
& FOLL_COW
, "mm: In follow_devmap_pmd with FOLL_COW set");
993 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
994 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
995 (FOLL_PIN
| FOLL_GET
)))
998 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1001 if (pmd_present(*pmd
) && pmd_devmap(*pmd
))
1006 if (flags
& FOLL_TOUCH
)
1007 touch_pmd(vma
, addr
, pmd
, flags
);
1010 * device mapped pages can only be returned if the
1011 * caller will manage the page reference count.
1013 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
1014 return ERR_PTR(-EEXIST
);
1016 pfn
+= (addr
& ~PMD_MASK
) >> PAGE_SHIFT
;
1017 *pgmap
= get_dev_pagemap(pfn
, *pgmap
);
1019 return ERR_PTR(-EFAULT
);
1020 page
= pfn_to_page(pfn
);
1021 if (!try_grab_page(page
, flags
))
1022 page
= ERR_PTR(-ENOMEM
);
1027 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1028 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1029 struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1031 spinlock_t
*dst_ptl
, *src_ptl
;
1032 struct page
*src_page
;
1034 pgtable_t pgtable
= NULL
;
1037 /* Skip if can be re-fill on fault */
1038 if (!vma_is_anonymous(dst_vma
))
1041 pgtable
= pte_alloc_one(dst_mm
);
1042 if (unlikely(!pgtable
))
1045 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
1046 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
1047 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1052 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1053 if (unlikely(is_swap_pmd(pmd
))) {
1054 swp_entry_t entry
= pmd_to_swp_entry(pmd
);
1056 VM_BUG_ON(!is_pmd_migration_entry(pmd
));
1057 if (is_write_migration_entry(entry
)) {
1058 make_migration_entry_read(&entry
);
1059 pmd
= swp_entry_to_pmd(entry
);
1060 if (pmd_swp_soft_dirty(*src_pmd
))
1061 pmd
= pmd_swp_mksoft_dirty(pmd
);
1062 if (pmd_swp_uffd_wp(*src_pmd
))
1063 pmd
= pmd_swp_mkuffd_wp(pmd
);
1064 set_pmd_at(src_mm
, addr
, src_pmd
, pmd
);
1066 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1067 mm_inc_nr_ptes(dst_mm
);
1068 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1069 if (!userfaultfd_wp(dst_vma
))
1070 pmd
= pmd_swp_clear_uffd_wp(pmd
);
1071 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1077 if (unlikely(!pmd_trans_huge(pmd
))) {
1078 pte_free(dst_mm
, pgtable
);
1082 * When page table lock is held, the huge zero pmd should not be
1083 * under splitting since we don't split the page itself, only pmd to
1086 if (is_huge_zero_pmd(pmd
)) {
1088 * get_huge_zero_page() will never allocate a new page here,
1089 * since we already have a zero page to copy. It just takes a
1092 mm_get_huge_zero_page(dst_mm
);
1096 src_page
= pmd_page(pmd
);
1097 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
1100 * If this page is a potentially pinned page, split and retry the fault
1101 * with smaller page size. Normally this should not happen because the
1102 * userspace should use MADV_DONTFORK upon pinned regions. This is a
1103 * best effort that the pinned pages won't be replaced by another
1104 * random page during the coming copy-on-write.
1106 if (unlikely(page_needs_cow_for_dma(src_vma
, src_page
))) {
1107 pte_free(dst_mm
, pgtable
);
1108 spin_unlock(src_ptl
);
1109 spin_unlock(dst_ptl
);
1110 __split_huge_pmd(src_vma
, src_pmd
, addr
, false, NULL
);
1115 page_dup_rmap(src_page
, true);
1116 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1118 mm_inc_nr_ptes(dst_mm
);
1119 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1120 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
1121 if (!userfaultfd_wp(dst_vma
))
1122 pmd
= pmd_clear_uffd_wp(pmd
);
1123 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
1124 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1128 spin_unlock(src_ptl
);
1129 spin_unlock(dst_ptl
);
1134 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1135 static void touch_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1136 pud_t
*pud
, int flags
)
1140 _pud
= pud_mkyoung(*pud
);
1141 if (flags
& FOLL_WRITE
)
1142 _pud
= pud_mkdirty(_pud
);
1143 if (pudp_set_access_flags(vma
, addr
& HPAGE_PUD_MASK
,
1144 pud
, _pud
, flags
& FOLL_WRITE
))
1145 update_mmu_cache_pud(vma
, addr
, pud
);
1148 struct page
*follow_devmap_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1149 pud_t
*pud
, int flags
, struct dev_pagemap
**pgmap
)
1151 unsigned long pfn
= pud_pfn(*pud
);
1152 struct mm_struct
*mm
= vma
->vm_mm
;
1155 assert_spin_locked(pud_lockptr(mm
, pud
));
1157 if (flags
& FOLL_WRITE
&& !pud_write(*pud
))
1160 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1161 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
1162 (FOLL_PIN
| FOLL_GET
)))
1165 if (pud_present(*pud
) && pud_devmap(*pud
))
1170 if (flags
& FOLL_TOUCH
)
1171 touch_pud(vma
, addr
, pud
, flags
);
1174 * device mapped pages can only be returned if the
1175 * caller will manage the page reference count.
1177 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1179 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
1180 return ERR_PTR(-EEXIST
);
1182 pfn
+= (addr
& ~PUD_MASK
) >> PAGE_SHIFT
;
1183 *pgmap
= get_dev_pagemap(pfn
, *pgmap
);
1185 return ERR_PTR(-EFAULT
);
1186 page
= pfn_to_page(pfn
);
1187 if (!try_grab_page(page
, flags
))
1188 page
= ERR_PTR(-ENOMEM
);
1193 int copy_huge_pud(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1194 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1195 struct vm_area_struct
*vma
)
1197 spinlock_t
*dst_ptl
, *src_ptl
;
1201 dst_ptl
= pud_lock(dst_mm
, dst_pud
);
1202 src_ptl
= pud_lockptr(src_mm
, src_pud
);
1203 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1207 if (unlikely(!pud_trans_huge(pud
) && !pud_devmap(pud
)))
1211 * When page table lock is held, the huge zero pud should not be
1212 * under splitting since we don't split the page itself, only pud to
1215 if (is_huge_zero_pud(pud
)) {
1216 /* No huge zero pud yet */
1219 /* Please refer to comments in copy_huge_pmd() */
1220 if (unlikely(page_needs_cow_for_dma(vma
, pud_page(pud
)))) {
1221 spin_unlock(src_ptl
);
1222 spin_unlock(dst_ptl
);
1223 __split_huge_pud(vma
, src_pud
, addr
);
1227 pudp_set_wrprotect(src_mm
, addr
, src_pud
);
1228 pud
= pud_mkold(pud_wrprotect(pud
));
1229 set_pud_at(dst_mm
, addr
, dst_pud
, pud
);
1233 spin_unlock(src_ptl
);
1234 spin_unlock(dst_ptl
);
1238 void huge_pud_set_accessed(struct vm_fault
*vmf
, pud_t orig_pud
)
1241 unsigned long haddr
;
1242 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1244 vmf
->ptl
= pud_lock(vmf
->vma
->vm_mm
, vmf
->pud
);
1245 if (unlikely(!pud_same(*vmf
->pud
, orig_pud
)))
1248 entry
= pud_mkyoung(orig_pud
);
1250 entry
= pud_mkdirty(entry
);
1251 haddr
= vmf
->address
& HPAGE_PUD_MASK
;
1252 if (pudp_set_access_flags(vmf
->vma
, haddr
, vmf
->pud
, entry
, write
))
1253 update_mmu_cache_pud(vmf
->vma
, vmf
->address
, vmf
->pud
);
1256 spin_unlock(vmf
->ptl
);
1258 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1260 void huge_pmd_set_accessed(struct vm_fault
*vmf
)
1263 unsigned long haddr
;
1264 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1265 pmd_t orig_pmd
= vmf
->orig_pmd
;
1267 vmf
->ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1268 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1271 entry
= pmd_mkyoung(orig_pmd
);
1273 entry
= pmd_mkdirty(entry
);
1274 haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1275 if (pmdp_set_access_flags(vmf
->vma
, haddr
, vmf
->pmd
, entry
, write
))
1276 update_mmu_cache_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
);
1279 spin_unlock(vmf
->ptl
);
1282 vm_fault_t
do_huge_pmd_wp_page(struct vm_fault
*vmf
)
1284 struct vm_area_struct
*vma
= vmf
->vma
;
1286 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1287 pmd_t orig_pmd
= vmf
->orig_pmd
;
1289 vmf
->ptl
= pmd_lockptr(vma
->vm_mm
, vmf
->pmd
);
1290 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1292 if (is_huge_zero_pmd(orig_pmd
))
1295 spin_lock(vmf
->ptl
);
1297 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1298 spin_unlock(vmf
->ptl
);
1302 page
= pmd_page(orig_pmd
);
1303 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1305 /* Lock page for reuse_swap_page() */
1306 if (!trylock_page(page
)) {
1308 spin_unlock(vmf
->ptl
);
1310 spin_lock(vmf
->ptl
);
1311 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1312 spin_unlock(vmf
->ptl
);
1321 * We can only reuse the page if nobody else maps the huge page or it's
1324 if (reuse_swap_page(page
, NULL
)) {
1326 entry
= pmd_mkyoung(orig_pmd
);
1327 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1328 if (pmdp_set_access_flags(vma
, haddr
, vmf
->pmd
, entry
, 1))
1329 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1331 spin_unlock(vmf
->ptl
);
1332 return VM_FAULT_WRITE
;
1336 spin_unlock(vmf
->ptl
);
1338 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
1339 return VM_FAULT_FALLBACK
;
1343 * FOLL_FORCE can write to even unwritable pmd's, but only
1344 * after we've gone through a COW cycle and they are dirty.
1346 static inline bool can_follow_write_pmd(pmd_t pmd
, unsigned int flags
)
1348 return pmd_write(pmd
) ||
1349 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pmd_dirty(pmd
));
1352 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1357 struct mm_struct
*mm
= vma
->vm_mm
;
1358 struct page
*page
= NULL
;
1360 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1362 if (flags
& FOLL_WRITE
&& !can_follow_write_pmd(*pmd
, flags
))
1365 /* Avoid dumping huge zero page */
1366 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1367 return ERR_PTR(-EFAULT
);
1369 /* Full NUMA hinting faults to serialise migration in fault paths */
1370 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1373 page
= pmd_page(*pmd
);
1374 VM_BUG_ON_PAGE(!PageHead(page
) && !is_zone_device_page(page
), page
);
1376 if (!try_grab_page(page
, flags
))
1377 return ERR_PTR(-ENOMEM
);
1379 if (flags
& FOLL_TOUCH
)
1380 touch_pmd(vma
, addr
, pmd
, flags
);
1382 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1384 * We don't mlock() pte-mapped THPs. This way we can avoid
1385 * leaking mlocked pages into non-VM_LOCKED VMAs.
1389 * In most cases the pmd is the only mapping of the page as we
1390 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1391 * writable private mappings in populate_vma_page_range().
1393 * The only scenario when we have the page shared here is if we
1394 * mlocking read-only mapping shared over fork(). We skip
1395 * mlocking such pages.
1399 * We can expect PageDoubleMap() to be stable under page lock:
1400 * for file pages we set it in page_add_file_rmap(), which
1401 * requires page to be locked.
1404 if (PageAnon(page
) && compound_mapcount(page
) != 1)
1406 if (PageDoubleMap(page
) || !page
->mapping
)
1408 if (!trylock_page(page
))
1410 if (page
->mapping
&& !PageDoubleMap(page
))
1411 mlock_vma_page(page
);
1415 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1416 VM_BUG_ON_PAGE(!PageCompound(page
) && !is_zone_device_page(page
), page
);
1422 /* NUMA hinting page fault entry point for trans huge pmds */
1423 vm_fault_t
do_huge_pmd_numa_page(struct vm_fault
*vmf
)
1425 struct vm_area_struct
*vma
= vmf
->vma
;
1426 pmd_t oldpmd
= vmf
->orig_pmd
;
1429 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1430 int page_nid
= NUMA_NO_NODE
;
1431 int target_nid
, last_cpupid
= -1;
1432 bool migrated
= false;
1433 bool was_writable
= pmd_savedwrite(oldpmd
);
1436 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1437 if (unlikely(!pmd_same(oldpmd
, *vmf
->pmd
))) {
1438 spin_unlock(vmf
->ptl
);
1443 * Since we took the NUMA fault, we must have observed the !accessible
1444 * bit. Make sure all other CPUs agree with that, to avoid them
1445 * modifying the page we're about to migrate.
1447 * Must be done under PTL such that we'll observe the relevant
1448 * inc_tlb_flush_pending().
1450 * We are not sure a pending tlb flush here is for a huge page
1451 * mapping or not. Hence use the tlb range variant
1453 if (mm_tlb_flush_pending(vma
->vm_mm
)) {
1454 flush_tlb_range(vma
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
1456 * change_huge_pmd() released the pmd lock before
1457 * invalidating the secondary MMUs sharing the primary
1458 * MMU pagetables (with ->invalidate_range()). The
1459 * mmu_notifier_invalidate_range_end() (which
1460 * internally calls ->invalidate_range()) in
1461 * change_pmd_range() will run after us, so we can't
1462 * rely on it here and we need an explicit invalidate.
1464 mmu_notifier_invalidate_range(vma
->vm_mm
, haddr
,
1465 haddr
+ HPAGE_PMD_SIZE
);
1468 pmd
= pmd_modify(oldpmd
, vma
->vm_page_prot
);
1469 page
= vm_normal_page_pmd(vma
, haddr
, pmd
);
1473 /* See similar comment in do_numa_page for explanation */
1475 flags
|= TNF_NO_GROUP
;
1477 page_nid
= page_to_nid(page
);
1478 last_cpupid
= page_cpupid_last(page
);
1479 target_nid
= numa_migrate_prep(page
, vma
, haddr
, page_nid
,
1482 if (target_nid
== NUMA_NO_NODE
) {
1487 spin_unlock(vmf
->ptl
);
1489 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
1491 flags
|= TNF_MIGRATED
;
1492 page_nid
= target_nid
;
1494 flags
|= TNF_MIGRATE_FAIL
;
1495 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1496 if (unlikely(!pmd_same(oldpmd
, *vmf
->pmd
))) {
1497 spin_unlock(vmf
->ptl
);
1504 if (page_nid
!= NUMA_NO_NODE
)
1505 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
,
1511 /* Restore the PMD */
1512 pmd
= pmd_modify(oldpmd
, vma
->vm_page_prot
);
1513 pmd
= pmd_mkyoung(pmd
);
1515 pmd
= pmd_mkwrite(pmd
);
1516 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, pmd
);
1517 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1518 spin_unlock(vmf
->ptl
);
1523 * Return true if we do MADV_FREE successfully on entire pmd page.
1524 * Otherwise, return false.
1526 bool madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1527 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1532 struct mm_struct
*mm
= tlb
->mm
;
1535 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
1537 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1542 if (is_huge_zero_pmd(orig_pmd
))
1545 if (unlikely(!pmd_present(orig_pmd
))) {
1546 VM_BUG_ON(thp_migration_supported() &&
1547 !is_pmd_migration_entry(orig_pmd
));
1551 page
= pmd_page(orig_pmd
);
1553 * If other processes are mapping this page, we couldn't discard
1554 * the page unless they all do MADV_FREE so let's skip the page.
1556 if (total_mapcount(page
) != 1)
1559 if (!trylock_page(page
))
1563 * If user want to discard part-pages of THP, split it so MADV_FREE
1564 * will deactivate only them.
1566 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1569 split_huge_page(page
);
1575 if (PageDirty(page
))
1576 ClearPageDirty(page
);
1579 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1580 pmdp_invalidate(vma
, addr
, pmd
);
1581 orig_pmd
= pmd_mkold(orig_pmd
);
1582 orig_pmd
= pmd_mkclean(orig_pmd
);
1584 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1585 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1588 mark_page_lazyfree(page
);
1596 static inline void zap_deposited_table(struct mm_struct
*mm
, pmd_t
*pmd
)
1600 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1601 pte_free(mm
, pgtable
);
1605 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1606 pmd_t
*pmd
, unsigned long addr
)
1611 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
1613 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1617 * For architectures like ppc64 we look at deposited pgtable
1618 * when calling pmdp_huge_get_and_clear. So do the
1619 * pgtable_trans_huge_withdraw after finishing pmdp related
1622 orig_pmd
= pmdp_huge_get_and_clear_full(vma
, addr
, pmd
,
1624 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1625 if (vma_is_special_huge(vma
)) {
1626 if (arch_needs_pgtable_deposit())
1627 zap_deposited_table(tlb
->mm
, pmd
);
1629 } else if (is_huge_zero_pmd(orig_pmd
)) {
1630 zap_deposited_table(tlb
->mm
, pmd
);
1633 struct page
*page
= NULL
;
1634 int flush_needed
= 1;
1636 if (pmd_present(orig_pmd
)) {
1637 page
= pmd_page(orig_pmd
);
1638 page_remove_rmap(page
, true);
1639 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1640 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1641 } else if (thp_migration_supported()) {
1644 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd
));
1645 entry
= pmd_to_swp_entry(orig_pmd
);
1646 page
= migration_entry_to_page(entry
);
1649 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1651 if (PageAnon(page
)) {
1652 zap_deposited_table(tlb
->mm
, pmd
);
1653 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1655 if (arch_needs_pgtable_deposit())
1656 zap_deposited_table(tlb
->mm
, pmd
);
1657 add_mm_counter(tlb
->mm
, mm_counter_file(page
), -HPAGE_PMD_NR
);
1662 tlb_remove_page_size(tlb
, page
, HPAGE_PMD_SIZE
);
1667 #ifndef pmd_move_must_withdraw
1668 static inline int pmd_move_must_withdraw(spinlock_t
*new_pmd_ptl
,
1669 spinlock_t
*old_pmd_ptl
,
1670 struct vm_area_struct
*vma
)
1673 * With split pmd lock we also need to move preallocated
1674 * PTE page table if new_pmd is on different PMD page table.
1676 * We also don't deposit and withdraw tables for file pages.
1678 return (new_pmd_ptl
!= old_pmd_ptl
) && vma_is_anonymous(vma
);
1682 static pmd_t
move_soft_dirty_pmd(pmd_t pmd
)
1684 #ifdef CONFIG_MEM_SOFT_DIRTY
1685 if (unlikely(is_pmd_migration_entry(pmd
)))
1686 pmd
= pmd_swp_mksoft_dirty(pmd
);
1687 else if (pmd_present(pmd
))
1688 pmd
= pmd_mksoft_dirty(pmd
);
1693 bool move_huge_pmd(struct vm_area_struct
*vma
, unsigned long old_addr
,
1694 unsigned long new_addr
, pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1696 spinlock_t
*old_ptl
, *new_ptl
;
1698 struct mm_struct
*mm
= vma
->vm_mm
;
1699 bool force_flush
= false;
1702 * The destination pmd shouldn't be established, free_pgtables()
1703 * should have release it.
1705 if (WARN_ON(!pmd_none(*new_pmd
))) {
1706 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1711 * We don't have to worry about the ordering of src and dst
1712 * ptlocks because exclusive mmap_lock prevents deadlock.
1714 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1716 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1717 if (new_ptl
!= old_ptl
)
1718 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1719 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1720 if (pmd_present(pmd
))
1722 VM_BUG_ON(!pmd_none(*new_pmd
));
1724 if (pmd_move_must_withdraw(new_ptl
, old_ptl
, vma
)) {
1726 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1727 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1729 pmd
= move_soft_dirty_pmd(pmd
);
1730 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1732 flush_tlb_range(vma
, old_addr
, old_addr
+ PMD_SIZE
);
1733 if (new_ptl
!= old_ptl
)
1734 spin_unlock(new_ptl
);
1735 spin_unlock(old_ptl
);
1743 * - 0 if PMD could not be locked
1744 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1745 * - HPAGE_PMD_NR if protections changed and TLB flush necessary
1747 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1748 unsigned long addr
, pgprot_t newprot
, unsigned long cp_flags
)
1750 struct mm_struct
*mm
= vma
->vm_mm
;
1753 bool preserve_write
;
1755 bool prot_numa
= cp_flags
& MM_CP_PROT_NUMA
;
1756 bool uffd_wp
= cp_flags
& MM_CP_UFFD_WP
;
1757 bool uffd_wp_resolve
= cp_flags
& MM_CP_UFFD_WP_RESOLVE
;
1759 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1763 preserve_write
= prot_numa
&& pmd_write(*pmd
);
1766 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1767 if (is_swap_pmd(*pmd
)) {
1768 swp_entry_t entry
= pmd_to_swp_entry(*pmd
);
1770 VM_BUG_ON(!is_pmd_migration_entry(*pmd
));
1771 if (is_write_migration_entry(entry
)) {
1774 * A protection check is difficult so
1775 * just be safe and disable write
1777 make_migration_entry_read(&entry
);
1778 newpmd
= swp_entry_to_pmd(entry
);
1779 if (pmd_swp_soft_dirty(*pmd
))
1780 newpmd
= pmd_swp_mksoft_dirty(newpmd
);
1781 if (pmd_swp_uffd_wp(*pmd
))
1782 newpmd
= pmd_swp_mkuffd_wp(newpmd
);
1783 set_pmd_at(mm
, addr
, pmd
, newpmd
);
1790 * Avoid trapping faults against the zero page. The read-only
1791 * data is likely to be read-cached on the local CPU and
1792 * local/remote hits to the zero page are not interesting.
1794 if (prot_numa
&& is_huge_zero_pmd(*pmd
))
1797 if (prot_numa
&& pmd_protnone(*pmd
))
1801 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1802 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1803 * which is also under mmap_read_lock(mm):
1806 * change_huge_pmd(prot_numa=1)
1807 * pmdp_huge_get_and_clear_notify()
1808 * madvise_dontneed()
1810 * pmd_trans_huge(*pmd) == 0 (without ptl)
1813 * // pmd is re-established
1815 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1816 * which may break userspace.
1818 * pmdp_invalidate() is required to make sure we don't miss
1819 * dirty/young flags set by hardware.
1821 entry
= pmdp_invalidate(vma
, addr
, pmd
);
1823 entry
= pmd_modify(entry
, newprot
);
1825 entry
= pmd_mk_savedwrite(entry
);
1827 entry
= pmd_wrprotect(entry
);
1828 entry
= pmd_mkuffd_wp(entry
);
1829 } else if (uffd_wp_resolve
) {
1831 * Leave the write bit to be handled by PF interrupt
1832 * handler, then things like COW could be properly
1835 entry
= pmd_clear_uffd_wp(entry
);
1838 set_pmd_at(mm
, addr
, pmd
, entry
);
1839 BUG_ON(vma_is_anonymous(vma
) && !preserve_write
&& pmd_write(entry
));
1846 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1848 * Note that if it returns page table lock pointer, this routine returns without
1849 * unlocking page table lock. So callers must unlock it.
1851 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1854 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1855 if (likely(is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) ||
1863 * Returns true if a given pud maps a thp, false otherwise.
1865 * Note that if it returns true, this routine returns without unlocking page
1866 * table lock. So callers must unlock it.
1868 spinlock_t
*__pud_trans_huge_lock(pud_t
*pud
, struct vm_area_struct
*vma
)
1872 ptl
= pud_lock(vma
->vm_mm
, pud
);
1873 if (likely(pud_trans_huge(*pud
) || pud_devmap(*pud
)))
1879 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1880 int zap_huge_pud(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1881 pud_t
*pud
, unsigned long addr
)
1885 ptl
= __pud_trans_huge_lock(pud
, vma
);
1889 * For architectures like ppc64 we look at deposited pgtable
1890 * when calling pudp_huge_get_and_clear. So do the
1891 * pgtable_trans_huge_withdraw after finishing pudp related
1894 pudp_huge_get_and_clear_full(tlb
->mm
, addr
, pud
, tlb
->fullmm
);
1895 tlb_remove_pud_tlb_entry(tlb
, pud
, addr
);
1896 if (vma_is_special_huge(vma
)) {
1898 /* No zero page support yet */
1900 /* No support for anonymous PUD pages yet */
1906 static void __split_huge_pud_locked(struct vm_area_struct
*vma
, pud_t
*pud
,
1907 unsigned long haddr
)
1909 VM_BUG_ON(haddr
& ~HPAGE_PUD_MASK
);
1910 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
1911 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PUD_SIZE
, vma
);
1912 VM_BUG_ON(!pud_trans_huge(*pud
) && !pud_devmap(*pud
));
1914 count_vm_event(THP_SPLIT_PUD
);
1916 pudp_huge_clear_flush_notify(vma
, haddr
, pud
);
1919 void __split_huge_pud(struct vm_area_struct
*vma
, pud_t
*pud
,
1920 unsigned long address
)
1923 struct mmu_notifier_range range
;
1925 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1926 address
& HPAGE_PUD_MASK
,
1927 (address
& HPAGE_PUD_MASK
) + HPAGE_PUD_SIZE
);
1928 mmu_notifier_invalidate_range_start(&range
);
1929 ptl
= pud_lock(vma
->vm_mm
, pud
);
1930 if (unlikely(!pud_trans_huge(*pud
) && !pud_devmap(*pud
)))
1932 __split_huge_pud_locked(vma
, pud
, range
.start
);
1937 * No need to double call mmu_notifier->invalidate_range() callback as
1938 * the above pudp_huge_clear_flush_notify() did already call it.
1940 mmu_notifier_invalidate_range_only_end(&range
);
1942 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1944 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
1945 unsigned long haddr
, pmd_t
*pmd
)
1947 struct mm_struct
*mm
= vma
->vm_mm
;
1953 * Leave pmd empty until pte is filled note that it is fine to delay
1954 * notification until mmu_notifier_invalidate_range_end() as we are
1955 * replacing a zero pmd write protected page with a zero pte write
1958 * See Documentation/vm/mmu_notifier.rst
1960 pmdp_huge_clear_flush(vma
, haddr
, pmd
);
1962 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1963 pmd_populate(mm
, &_pmd
, pgtable
);
1965 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
1967 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
1968 entry
= pte_mkspecial(entry
);
1969 pte
= pte_offset_map(&_pmd
, haddr
);
1970 VM_BUG_ON(!pte_none(*pte
));
1971 set_pte_at(mm
, haddr
, pte
, entry
);
1974 smp_wmb(); /* make pte visible before pmd */
1975 pmd_populate(mm
, pmd
, pgtable
);
1978 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1979 unsigned long haddr
, bool freeze
)
1981 struct mm_struct
*mm
= vma
->vm_mm
;
1984 pmd_t old_pmd
, _pmd
;
1985 bool young
, write
, soft_dirty
, pmd_migration
= false, uffd_wp
= false;
1989 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
1990 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
1991 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
1992 VM_BUG_ON(!is_pmd_migration_entry(*pmd
) && !pmd_trans_huge(*pmd
)
1993 && !pmd_devmap(*pmd
));
1995 count_vm_event(THP_SPLIT_PMD
);
1997 if (!vma_is_anonymous(vma
)) {
1998 old_pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2000 * We are going to unmap this huge page. So
2001 * just go ahead and zap it
2003 if (arch_needs_pgtable_deposit())
2004 zap_deposited_table(mm
, pmd
);
2005 if (vma_is_special_huge(vma
))
2007 if (unlikely(is_pmd_migration_entry(old_pmd
))) {
2010 entry
= pmd_to_swp_entry(old_pmd
);
2011 page
= migration_entry_to_page(entry
);
2013 page
= pmd_page(old_pmd
);
2014 if (!PageDirty(page
) && pmd_dirty(old_pmd
))
2015 set_page_dirty(page
);
2016 if (!PageReferenced(page
) && pmd_young(old_pmd
))
2017 SetPageReferenced(page
);
2018 page_remove_rmap(page
, true);
2021 add_mm_counter(mm
, mm_counter_file(page
), -HPAGE_PMD_NR
);
2025 if (is_huge_zero_pmd(*pmd
)) {
2027 * FIXME: Do we want to invalidate secondary mmu by calling
2028 * mmu_notifier_invalidate_range() see comments below inside
2029 * __split_huge_pmd() ?
2031 * We are going from a zero huge page write protected to zero
2032 * small page also write protected so it does not seems useful
2033 * to invalidate secondary mmu at this time.
2035 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2039 * Up to this point the pmd is present and huge and userland has the
2040 * whole access to the hugepage during the split (which happens in
2041 * place). If we overwrite the pmd with the not-huge version pointing
2042 * to the pte here (which of course we could if all CPUs were bug
2043 * free), userland could trigger a small page size TLB miss on the
2044 * small sized TLB while the hugepage TLB entry is still established in
2045 * the huge TLB. Some CPU doesn't like that.
2046 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2047 * 383 on page 105. Intel should be safe but is also warns that it's
2048 * only safe if the permission and cache attributes of the two entries
2049 * loaded in the two TLB is identical (which should be the case here).
2050 * But it is generally safer to never allow small and huge TLB entries
2051 * for the same virtual address to be loaded simultaneously. So instead
2052 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2053 * current pmd notpresent (atomically because here the pmd_trans_huge
2054 * must remain set at all times on the pmd until the split is complete
2055 * for this pmd), then we flush the SMP TLB and finally we write the
2056 * non-huge version of the pmd entry with pmd_populate.
2058 old_pmd
= pmdp_invalidate(vma
, haddr
, pmd
);
2060 pmd_migration
= is_pmd_migration_entry(old_pmd
);
2061 if (unlikely(pmd_migration
)) {
2064 entry
= pmd_to_swp_entry(old_pmd
);
2065 page
= migration_entry_to_page(entry
);
2066 write
= is_write_migration_entry(entry
);
2068 soft_dirty
= pmd_swp_soft_dirty(old_pmd
);
2069 uffd_wp
= pmd_swp_uffd_wp(old_pmd
);
2071 page
= pmd_page(old_pmd
);
2072 if (pmd_dirty(old_pmd
))
2074 write
= pmd_write(old_pmd
);
2075 young
= pmd_young(old_pmd
);
2076 soft_dirty
= pmd_soft_dirty(old_pmd
);
2077 uffd_wp
= pmd_uffd_wp(old_pmd
);
2079 VM_BUG_ON_PAGE(!page_count(page
), page
);
2080 page_ref_add(page
, HPAGE_PMD_NR
- 1);
2083 * Withdraw the table only after we mark the pmd entry invalid.
2084 * This's critical for some architectures (Power).
2086 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2087 pmd_populate(mm
, &_pmd
, pgtable
);
2089 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
2092 * Note that NUMA hinting access restrictions are not
2093 * transferred to avoid any possibility of altering
2094 * permissions across VMAs.
2096 if (freeze
|| pmd_migration
) {
2097 swp_entry_t swp_entry
;
2098 swp_entry
= make_migration_entry(page
+ i
, write
);
2099 entry
= swp_entry_to_pte(swp_entry
);
2101 entry
= pte_swp_mksoft_dirty(entry
);
2103 entry
= pte_swp_mkuffd_wp(entry
);
2105 entry
= mk_pte(page
+ i
, READ_ONCE(vma
->vm_page_prot
));
2106 entry
= maybe_mkwrite(entry
, vma
);
2108 entry
= pte_wrprotect(entry
);
2110 entry
= pte_mkold(entry
);
2112 entry
= pte_mksoft_dirty(entry
);
2114 entry
= pte_mkuffd_wp(entry
);
2116 pte
= pte_offset_map(&_pmd
, addr
);
2117 BUG_ON(!pte_none(*pte
));
2118 set_pte_at(mm
, addr
, pte
, entry
);
2120 atomic_inc(&page
[i
]._mapcount
);
2124 if (!pmd_migration
) {
2126 * Set PG_double_map before dropping compound_mapcount to avoid
2127 * false-negative page_mapped().
2129 if (compound_mapcount(page
) > 1 &&
2130 !TestSetPageDoubleMap(page
)) {
2131 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2132 atomic_inc(&page
[i
]._mapcount
);
2135 lock_page_memcg(page
);
2136 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
2137 /* Last compound_mapcount is gone. */
2138 __mod_lruvec_page_state(page
, NR_ANON_THPS
,
2140 if (TestClearPageDoubleMap(page
)) {
2141 /* No need in mapcount reference anymore */
2142 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2143 atomic_dec(&page
[i
]._mapcount
);
2146 unlock_page_memcg(page
);
2149 smp_wmb(); /* make pte visible before pmd */
2150 pmd_populate(mm
, pmd
, pgtable
);
2153 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2154 page_remove_rmap(page
+ i
, false);
2160 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2161 unsigned long address
, bool freeze
, struct page
*page
)
2164 struct mmu_notifier_range range
;
2165 bool do_unlock_page
= false;
2168 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
2169 address
& HPAGE_PMD_MASK
,
2170 (address
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
);
2171 mmu_notifier_invalidate_range_start(&range
);
2172 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
2175 * If caller asks to setup a migration entries, we need a page to check
2176 * pmd against. Otherwise we can end up replacing wrong page.
2178 VM_BUG_ON(freeze
&& !page
);
2180 VM_WARN_ON_ONCE(!PageLocked(page
));
2181 if (page
!= pmd_page(*pmd
))
2186 if (pmd_trans_huge(*pmd
)) {
2188 page
= pmd_page(*pmd
);
2190 * An anonymous page must be locked, to ensure that a
2191 * concurrent reuse_swap_page() sees stable mapcount;
2192 * but reuse_swap_page() is not used on shmem or file,
2193 * and page lock must not be taken when zap_pmd_range()
2194 * calls __split_huge_pmd() while i_mmap_lock is held.
2196 if (PageAnon(page
)) {
2197 if (unlikely(!trylock_page(page
))) {
2203 if (unlikely(!pmd_same(*pmd
, _pmd
))) {
2211 do_unlock_page
= true;
2214 if (PageMlocked(page
))
2215 clear_page_mlock(page
);
2216 } else if (!(pmd_devmap(*pmd
) || is_pmd_migration_entry(*pmd
)))
2218 __split_huge_pmd_locked(vma
, pmd
, range
.start
, freeze
);
2224 * No need to double call mmu_notifier->invalidate_range() callback.
2225 * They are 3 cases to consider inside __split_huge_pmd_locked():
2226 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2227 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2228 * fault will trigger a flush_notify before pointing to a new page
2229 * (it is fine if the secondary mmu keeps pointing to the old zero
2230 * page in the meantime)
2231 * 3) Split a huge pmd into pte pointing to the same page. No need
2232 * to invalidate secondary tlb entry they are all still valid.
2233 * any further changes to individual pte will notify. So no need
2234 * to call mmu_notifier->invalidate_range()
2236 mmu_notifier_invalidate_range_only_end(&range
);
2239 void split_huge_pmd_address(struct vm_area_struct
*vma
, unsigned long address
,
2240 bool freeze
, struct page
*page
)
2247 pgd
= pgd_offset(vma
->vm_mm
, address
);
2248 if (!pgd_present(*pgd
))
2251 p4d
= p4d_offset(pgd
, address
);
2252 if (!p4d_present(*p4d
))
2255 pud
= pud_offset(p4d
, address
);
2256 if (!pud_present(*pud
))
2259 pmd
= pmd_offset(pud
, address
);
2261 __split_huge_pmd(vma
, pmd
, address
, freeze
, page
);
2264 static inline void split_huge_pmd_if_needed(struct vm_area_struct
*vma
, unsigned long address
)
2267 * If the new address isn't hpage aligned and it could previously
2268 * contain an hugepage: check if we need to split an huge pmd.
2270 if (!IS_ALIGNED(address
, HPAGE_PMD_SIZE
) &&
2271 range_in_vma(vma
, ALIGN_DOWN(address
, HPAGE_PMD_SIZE
),
2272 ALIGN(address
, HPAGE_PMD_SIZE
)))
2273 split_huge_pmd_address(vma
, address
, false, NULL
);
2276 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2277 unsigned long start
,
2281 /* Check if we need to split start first. */
2282 split_huge_pmd_if_needed(vma
, start
);
2284 /* Check if we need to split end next. */
2285 split_huge_pmd_if_needed(vma
, end
);
2288 * If we're also updating the vma->vm_next->vm_start,
2289 * check if we need to split it.
2291 if (adjust_next
> 0) {
2292 struct vm_area_struct
*next
= vma
->vm_next
;
2293 unsigned long nstart
= next
->vm_start
;
2294 nstart
+= adjust_next
;
2295 split_huge_pmd_if_needed(next
, nstart
);
2299 static void unmap_page(struct page
*page
)
2301 enum ttu_flags ttu_flags
= TTU_IGNORE_MLOCK
| TTU_SYNC
|
2302 TTU_RMAP_LOCKED
| TTU_SPLIT_HUGE_PMD
;
2304 VM_BUG_ON_PAGE(!PageHead(page
), page
);
2307 ttu_flags
|= TTU_SPLIT_FREEZE
;
2309 try_to_unmap(page
, ttu_flags
);
2311 VM_WARN_ON_ONCE_PAGE(page_mapped(page
), page
);
2314 static void remap_page(struct page
*page
, unsigned int nr
)
2317 if (PageTransHuge(page
)) {
2318 remove_migration_ptes(page
, page
, true);
2320 for (i
= 0; i
< nr
; i
++)
2321 remove_migration_ptes(page
+ i
, page
+ i
, true);
2325 static void lru_add_page_tail(struct page
*head
, struct page
*tail
,
2326 struct lruvec
*lruvec
, struct list_head
*list
)
2328 VM_BUG_ON_PAGE(!PageHead(head
), head
);
2329 VM_BUG_ON_PAGE(PageCompound(tail
), head
);
2330 VM_BUG_ON_PAGE(PageLRU(tail
), head
);
2331 lockdep_assert_held(&lruvec
->lru_lock
);
2334 /* page reclaim is reclaiming a huge page */
2335 VM_WARN_ON(PageLRU(head
));
2337 list_add_tail(&tail
->lru
, list
);
2339 /* head is still on lru (and we have it frozen) */
2340 VM_WARN_ON(!PageLRU(head
));
2342 list_add_tail(&tail
->lru
, &head
->lru
);
2346 static void __split_huge_page_tail(struct page
*head
, int tail
,
2347 struct lruvec
*lruvec
, struct list_head
*list
)
2349 struct page
*page_tail
= head
+ tail
;
2351 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
2354 * Clone page flags before unfreezing refcount.
2356 * After successful get_page_unless_zero() might follow flags change,
2357 * for example lock_page() which set PG_waiters.
2359 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
2360 page_tail
->flags
|= (head
->flags
&
2361 ((1L << PG_referenced
) |
2362 (1L << PG_swapbacked
) |
2363 (1L << PG_swapcache
) |
2364 (1L << PG_mlocked
) |
2365 (1L << PG_uptodate
) |
2367 (1L << PG_workingset
) |
2369 (1L << PG_unevictable
) |
2375 /* ->mapping in first tail page is compound_mapcount */
2376 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
2378 page_tail
->mapping
= head
->mapping
;
2379 page_tail
->index
= head
->index
+ tail
;
2381 /* Page flags must be visible before we make the page non-compound. */
2385 * Clear PageTail before unfreezing page refcount.
2387 * After successful get_page_unless_zero() might follow put_page()
2388 * which needs correct compound_head().
2390 clear_compound_head(page_tail
);
2392 /* Finally unfreeze refcount. Additional reference from page cache. */
2393 page_ref_unfreeze(page_tail
, 1 + (!PageAnon(head
) ||
2394 PageSwapCache(head
)));
2396 if (page_is_young(head
))
2397 set_page_young(page_tail
);
2398 if (page_is_idle(head
))
2399 set_page_idle(page_tail
);
2401 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
2404 * always add to the tail because some iterators expect new
2405 * pages to show after the currently processed elements - e.g.
2408 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
2411 static void __split_huge_page(struct page
*page
, struct list_head
*list
,
2414 struct page
*head
= compound_head(page
);
2415 struct lruvec
*lruvec
;
2416 struct address_space
*swap_cache
= NULL
;
2417 unsigned long offset
= 0;
2418 unsigned int nr
= thp_nr_pages(head
);
2421 /* complete memcg works before add pages to LRU */
2422 split_page_memcg(head
, nr
);
2424 if (PageAnon(head
) && PageSwapCache(head
)) {
2425 swp_entry_t entry
= { .val
= page_private(head
) };
2427 offset
= swp_offset(entry
);
2428 swap_cache
= swap_address_space(entry
);
2429 xa_lock(&swap_cache
->i_pages
);
2432 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2433 lruvec
= lock_page_lruvec(head
);
2435 for (i
= nr
- 1; i
>= 1; i
--) {
2436 __split_huge_page_tail(head
, i
, lruvec
, list
);
2437 /* Some pages can be beyond i_size: drop them from page cache */
2438 if (head
[i
].index
>= end
) {
2439 ClearPageDirty(head
+ i
);
2440 __delete_from_page_cache(head
+ i
, NULL
);
2441 if (IS_ENABLED(CONFIG_SHMEM
) && PageSwapBacked(head
))
2442 shmem_uncharge(head
->mapping
->host
, 1);
2444 } else if (!PageAnon(page
)) {
2445 __xa_store(&head
->mapping
->i_pages
, head
[i
].index
,
2447 } else if (swap_cache
) {
2448 __xa_store(&swap_cache
->i_pages
, offset
+ i
,
2453 ClearPageCompound(head
);
2454 unlock_page_lruvec(lruvec
);
2455 /* Caller disabled irqs, so they are still disabled here */
2457 split_page_owner(head
, nr
);
2459 /* See comment in __split_huge_page_tail() */
2460 if (PageAnon(head
)) {
2461 /* Additional pin to swap cache */
2462 if (PageSwapCache(head
)) {
2463 page_ref_add(head
, 2);
2464 xa_unlock(&swap_cache
->i_pages
);
2469 /* Additional pin to page cache */
2470 page_ref_add(head
, 2);
2471 xa_unlock(&head
->mapping
->i_pages
);
2475 remap_page(head
, nr
);
2477 if (PageSwapCache(head
)) {
2478 swp_entry_t entry
= { .val
= page_private(head
) };
2480 split_swap_cluster(entry
);
2483 for (i
= 0; i
< nr
; i
++) {
2484 struct page
*subpage
= head
+ i
;
2485 if (subpage
== page
)
2487 unlock_page(subpage
);
2490 * Subpages may be freed if there wasn't any mapping
2491 * like if add_to_swap() is running on a lru page that
2492 * had its mapping zapped. And freeing these pages
2493 * requires taking the lru_lock so we do the put_page
2494 * of the tail pages after the split is complete.
2500 int total_mapcount(struct page
*page
)
2502 int i
, compound
, nr
, ret
;
2504 VM_BUG_ON_PAGE(PageTail(page
), page
);
2506 if (likely(!PageCompound(page
)))
2507 return atomic_read(&page
->_mapcount
) + 1;
2509 compound
= compound_mapcount(page
);
2510 nr
= compound_nr(page
);
2514 for (i
= 0; i
< nr
; i
++)
2515 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
2516 /* File pages has compound_mapcount included in _mapcount */
2517 if (!PageAnon(page
))
2518 return ret
- compound
* nr
;
2519 if (PageDoubleMap(page
))
2525 * This calculates accurately how many mappings a transparent hugepage
2526 * has (unlike page_mapcount() which isn't fully accurate). This full
2527 * accuracy is primarily needed to know if copy-on-write faults can
2528 * reuse the page and change the mapping to read-write instead of
2529 * copying them. At the same time this returns the total_mapcount too.
2531 * The function returns the highest mapcount any one of the subpages
2532 * has. If the return value is one, even if different processes are
2533 * mapping different subpages of the transparent hugepage, they can
2534 * all reuse it, because each process is reusing a different subpage.
2536 * The total_mapcount is instead counting all virtual mappings of the
2537 * subpages. If the total_mapcount is equal to "one", it tells the
2538 * caller all mappings belong to the same "mm" and in turn the
2539 * anon_vma of the transparent hugepage can become the vma->anon_vma
2540 * local one as no other process may be mapping any of the subpages.
2542 * It would be more accurate to replace page_mapcount() with
2543 * page_trans_huge_mapcount(), however we only use
2544 * page_trans_huge_mapcount() in the copy-on-write faults where we
2545 * need full accuracy to avoid breaking page pinning, because
2546 * page_trans_huge_mapcount() is slower than page_mapcount().
2548 int page_trans_huge_mapcount(struct page
*page
, int *total_mapcount
)
2550 int i
, ret
, _total_mapcount
, mapcount
;
2552 /* hugetlbfs shouldn't call it */
2553 VM_BUG_ON_PAGE(PageHuge(page
), page
);
2555 if (likely(!PageTransCompound(page
))) {
2556 mapcount
= atomic_read(&page
->_mapcount
) + 1;
2558 *total_mapcount
= mapcount
;
2562 page
= compound_head(page
);
2564 _total_mapcount
= ret
= 0;
2565 for (i
= 0; i
< thp_nr_pages(page
); i
++) {
2566 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
2567 ret
= max(ret
, mapcount
);
2568 _total_mapcount
+= mapcount
;
2570 if (PageDoubleMap(page
)) {
2572 _total_mapcount
-= thp_nr_pages(page
);
2574 mapcount
= compound_mapcount(page
);
2576 _total_mapcount
+= mapcount
;
2578 *total_mapcount
= _total_mapcount
;
2582 /* Racy check whether the huge page can be split */
2583 bool can_split_huge_page(struct page
*page
, int *pextra_pins
)
2587 /* Additional pins from page cache */
2589 extra_pins
= PageSwapCache(page
) ? thp_nr_pages(page
) : 0;
2591 extra_pins
= thp_nr_pages(page
);
2593 *pextra_pins
= extra_pins
;
2594 return total_mapcount(page
) == page_count(page
) - extra_pins
- 1;
2598 * This function splits huge page into normal pages. @page can point to any
2599 * subpage of huge page to split. Split doesn't change the position of @page.
2601 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2602 * The huge page must be locked.
2604 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2606 * Both head page and tail pages will inherit mapping, flags, and so on from
2609 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2610 * they are not mapped.
2612 * Returns 0 if the hugepage is split successfully.
2613 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2616 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
2618 struct page
*head
= compound_head(page
);
2619 struct deferred_split
*ds_queue
= get_deferred_split_queue(head
);
2620 struct anon_vma
*anon_vma
= NULL
;
2621 struct address_space
*mapping
= NULL
;
2622 int extra_pins
, ret
;
2625 VM_BUG_ON_PAGE(is_huge_zero_page(head
), head
);
2626 VM_BUG_ON_PAGE(!PageLocked(head
), head
);
2627 VM_BUG_ON_PAGE(!PageCompound(head
), head
);
2629 if (PageWriteback(head
))
2632 if (PageAnon(head
)) {
2634 * The caller does not necessarily hold an mmap_lock that would
2635 * prevent the anon_vma disappearing so we first we take a
2636 * reference to it and then lock the anon_vma for write. This
2637 * is similar to page_lock_anon_vma_read except the write lock
2638 * is taken to serialise against parallel split or collapse
2641 anon_vma
= page_get_anon_vma(head
);
2648 anon_vma_lock_write(anon_vma
);
2650 mapping
= head
->mapping
;
2659 i_mmap_lock_read(mapping
);
2662 *__split_huge_page() may need to trim off pages beyond EOF:
2663 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2664 * which cannot be nested inside the page tree lock. So note
2665 * end now: i_size itself may be changed at any moment, but
2666 * head page lock is good enough to serialize the trimming.
2668 end
= DIV_ROUND_UP(i_size_read(mapping
->host
), PAGE_SIZE
);
2672 * Racy check if we can split the page, before unmap_page() will
2675 if (!can_split_huge_page(head
, &extra_pins
)) {
2682 /* block interrupt reentry in xa_lock and spinlock */
2683 local_irq_disable();
2685 XA_STATE(xas
, &mapping
->i_pages
, page_index(head
));
2688 * Check if the head page is present in page cache.
2689 * We assume all tail are present too, if head is there.
2691 xa_lock(&mapping
->i_pages
);
2692 if (xas_load(&xas
) != head
)
2696 /* Prevent deferred_split_scan() touching ->_refcount */
2697 spin_lock(&ds_queue
->split_queue_lock
);
2698 if (page_ref_freeze(head
, 1 + extra_pins
)) {
2699 if (!list_empty(page_deferred_list(head
))) {
2700 ds_queue
->split_queue_len
--;
2701 list_del(page_deferred_list(head
));
2703 spin_unlock(&ds_queue
->split_queue_lock
);
2705 int nr
= thp_nr_pages(head
);
2707 if (PageSwapBacked(head
))
2708 __mod_lruvec_page_state(head
, NR_SHMEM_THPS
,
2711 __mod_lruvec_page_state(head
, NR_FILE_THPS
,
2715 __split_huge_page(page
, list
, end
);
2718 spin_unlock(&ds_queue
->split_queue_lock
);
2721 xa_unlock(&mapping
->i_pages
);
2723 remap_page(head
, thp_nr_pages(head
));
2729 anon_vma_unlock_write(anon_vma
);
2730 put_anon_vma(anon_vma
);
2733 i_mmap_unlock_read(mapping
);
2735 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
2739 void free_transhuge_page(struct page
*page
)
2741 struct deferred_split
*ds_queue
= get_deferred_split_queue(page
);
2742 unsigned long flags
;
2744 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2745 if (!list_empty(page_deferred_list(page
))) {
2746 ds_queue
->split_queue_len
--;
2747 list_del(page_deferred_list(page
));
2749 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2750 free_compound_page(page
);
2753 void deferred_split_huge_page(struct page
*page
)
2755 struct deferred_split
*ds_queue
= get_deferred_split_queue(page
);
2757 struct mem_cgroup
*memcg
= page_memcg(compound_head(page
));
2759 unsigned long flags
;
2761 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
2764 * The try_to_unmap() in page reclaim path might reach here too,
2765 * this may cause a race condition to corrupt deferred split queue.
2766 * And, if page reclaim is already handling the same page, it is
2767 * unnecessary to handle it again in shrinker.
2769 * Check PageSwapCache to determine if the page is being
2770 * handled by page reclaim since THP swap would add the page into
2771 * swap cache before calling try_to_unmap().
2773 if (PageSwapCache(page
))
2776 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2777 if (list_empty(page_deferred_list(page
))) {
2778 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
2779 list_add_tail(page_deferred_list(page
), &ds_queue
->split_queue
);
2780 ds_queue
->split_queue_len
++;
2783 set_shrinker_bit(memcg
, page_to_nid(page
),
2784 deferred_split_shrinker
.id
);
2787 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2790 static unsigned long deferred_split_count(struct shrinker
*shrink
,
2791 struct shrink_control
*sc
)
2793 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2794 struct deferred_split
*ds_queue
= &pgdata
->deferred_split_queue
;
2798 ds_queue
= &sc
->memcg
->deferred_split_queue
;
2800 return READ_ONCE(ds_queue
->split_queue_len
);
2803 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
2804 struct shrink_control
*sc
)
2806 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2807 struct deferred_split
*ds_queue
= &pgdata
->deferred_split_queue
;
2808 unsigned long flags
;
2809 LIST_HEAD(list
), *pos
, *next
;
2815 ds_queue
= &sc
->memcg
->deferred_split_queue
;
2818 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2819 /* Take pin on all head pages to avoid freeing them under us */
2820 list_for_each_safe(pos
, next
, &ds_queue
->split_queue
) {
2821 page
= list_entry((void *)pos
, struct page
, deferred_list
);
2822 page
= compound_head(page
);
2823 if (get_page_unless_zero(page
)) {
2824 list_move(page_deferred_list(page
), &list
);
2826 /* We lost race with put_compound_page() */
2827 list_del_init(page_deferred_list(page
));
2828 ds_queue
->split_queue_len
--;
2830 if (!--sc
->nr_to_scan
)
2833 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2835 list_for_each_safe(pos
, next
, &list
) {
2836 page
= list_entry((void *)pos
, struct page
, deferred_list
);
2837 if (!trylock_page(page
))
2839 /* split_huge_page() removes page from list on success */
2840 if (!split_huge_page(page
))
2847 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2848 list_splice_tail(&list
, &ds_queue
->split_queue
);
2849 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2852 * Stop shrinker if we didn't split any page, but the queue is empty.
2853 * This can happen if pages were freed under us.
2855 if (!split
&& list_empty(&ds_queue
->split_queue
))
2860 static struct shrinker deferred_split_shrinker
= {
2861 .count_objects
= deferred_split_count
,
2862 .scan_objects
= deferred_split_scan
,
2863 .seeks
= DEFAULT_SEEKS
,
2864 .flags
= SHRINKER_NUMA_AWARE
| SHRINKER_MEMCG_AWARE
|
2868 #ifdef CONFIG_DEBUG_FS
2869 static void split_huge_pages_all(void)
2873 unsigned long pfn
, max_zone_pfn
;
2874 unsigned long total
= 0, split
= 0;
2876 pr_debug("Split all THPs\n");
2877 for_each_populated_zone(zone
) {
2878 max_zone_pfn
= zone_end_pfn(zone
);
2879 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
2880 if (!pfn_valid(pfn
))
2883 page
= pfn_to_page(pfn
);
2884 if (!get_page_unless_zero(page
))
2887 if (zone
!= page_zone(page
))
2890 if (!PageHead(page
) || PageHuge(page
) || !PageLRU(page
))
2895 if (!split_huge_page(page
))
2904 pr_debug("%lu of %lu THP split\n", split
, total
);
2907 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct
*vma
)
2909 return vma_is_special_huge(vma
) || (vma
->vm_flags
& VM_IO
) ||
2910 is_vm_hugetlb_page(vma
);
2913 static int split_huge_pages_pid(int pid
, unsigned long vaddr_start
,
2914 unsigned long vaddr_end
)
2917 struct task_struct
*task
;
2918 struct mm_struct
*mm
;
2919 unsigned long total
= 0, split
= 0;
2922 vaddr_start
&= PAGE_MASK
;
2923 vaddr_end
&= PAGE_MASK
;
2925 /* Find the task_struct from pid */
2927 task
= find_task_by_vpid(pid
);
2933 get_task_struct(task
);
2936 /* Find the mm_struct */
2937 mm
= get_task_mm(task
);
2938 put_task_struct(task
);
2945 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
2946 pid
, vaddr_start
, vaddr_end
);
2950 * always increase addr by PAGE_SIZE, since we could have a PTE page
2951 * table filled with PTE-mapped THPs, each of which is distinct.
2953 for (addr
= vaddr_start
; addr
< vaddr_end
; addr
+= PAGE_SIZE
) {
2954 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
2955 unsigned int follflags
;
2958 if (!vma
|| addr
< vma
->vm_start
)
2961 /* skip special VMA and hugetlb VMA */
2962 if (vma_not_suitable_for_thp_split(vma
)) {
2967 /* FOLL_DUMP to ignore special (like zero) pages */
2968 follflags
= FOLL_GET
| FOLL_DUMP
;
2969 page
= follow_page(vma
, addr
, follflags
);
2976 if (!is_transparent_hugepage(page
))
2980 if (!can_split_huge_page(compound_head(page
), NULL
))
2983 if (!trylock_page(page
))
2986 if (!split_huge_page(page
))
2994 mmap_read_unlock(mm
);
2997 pr_debug("%lu of %lu THP split\n", split
, total
);
3003 static int split_huge_pages_in_file(const char *file_path
, pgoff_t off_start
,
3006 struct filename
*file
;
3007 struct file
*candidate
;
3008 struct address_space
*mapping
;
3012 unsigned long total
= 0, split
= 0;
3014 file
= getname_kernel(file_path
);
3018 candidate
= file_open_name(file
, O_RDONLY
, 0);
3019 if (IS_ERR(candidate
))
3022 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3023 file_path
, off_start
, off_end
);
3025 mapping
= candidate
->f_mapping
;
3027 for (index
= off_start
; index
< off_end
; index
+= nr_pages
) {
3028 struct page
*fpage
= pagecache_get_page(mapping
, index
,
3029 FGP_ENTRY
| FGP_HEAD
, 0);
3032 if (xa_is_value(fpage
) || !fpage
)
3035 if (!is_transparent_hugepage(fpage
))
3039 nr_pages
= thp_nr_pages(fpage
);
3041 if (!trylock_page(fpage
))
3044 if (!split_huge_page(fpage
))
3053 filp_close(candidate
, NULL
);
3056 pr_debug("%lu of %lu file-backed THP split\n", split
, total
);
3062 #define MAX_INPUT_BUF_SZ 255
3064 static ssize_t
split_huge_pages_write(struct file
*file
, const char __user
*buf
,
3065 size_t count
, loff_t
*ppops
)
3067 static DEFINE_MUTEX(split_debug_mutex
);
3069 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3070 char input_buf
[MAX_INPUT_BUF_SZ
];
3072 unsigned long vaddr_start
, vaddr_end
;
3074 ret
= mutex_lock_interruptible(&split_debug_mutex
);
3080 memset(input_buf
, 0, MAX_INPUT_BUF_SZ
);
3081 if (copy_from_user(input_buf
, buf
, min_t(size_t, count
, MAX_INPUT_BUF_SZ
)))
3084 input_buf
[MAX_INPUT_BUF_SZ
- 1] = '\0';
3086 if (input_buf
[0] == '/') {
3088 char *buf
= input_buf
;
3089 char file_path
[MAX_INPUT_BUF_SZ
];
3090 pgoff_t off_start
= 0, off_end
= 0;
3091 size_t input_len
= strlen(input_buf
);
3093 tok
= strsep(&buf
, ",");
3095 strncpy(file_path
, tok
, MAX_INPUT_BUF_SZ
);
3101 ret
= sscanf(buf
, "0x%lx,0x%lx", &off_start
, &off_end
);
3106 ret
= split_huge_pages_in_file(file_path
, off_start
, off_end
);
3113 ret
= sscanf(input_buf
, "%d,0x%lx,0x%lx", &pid
, &vaddr_start
, &vaddr_end
);
3114 if (ret
== 1 && pid
== 1) {
3115 split_huge_pages_all();
3116 ret
= strlen(input_buf
);
3118 } else if (ret
!= 3) {
3123 ret
= split_huge_pages_pid(pid
, vaddr_start
, vaddr_end
);
3125 ret
= strlen(input_buf
);
3127 mutex_unlock(&split_debug_mutex
);
3132 static const struct file_operations split_huge_pages_fops
= {
3133 .owner
= THIS_MODULE
,
3134 .write
= split_huge_pages_write
,
3135 .llseek
= no_llseek
,
3138 static int __init
split_huge_pages_debugfs(void)
3140 debugfs_create_file("split_huge_pages", 0200, NULL
, NULL
,
3141 &split_huge_pages_fops
);
3144 late_initcall(split_huge_pages_debugfs
);
3147 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3148 void set_pmd_migration_entry(struct page_vma_mapped_walk
*pvmw
,
3151 struct vm_area_struct
*vma
= pvmw
->vma
;
3152 struct mm_struct
*mm
= vma
->vm_mm
;
3153 unsigned long address
= pvmw
->address
;
3158 if (!(pvmw
->pmd
&& !pvmw
->pte
))
3161 flush_cache_range(vma
, address
, address
+ HPAGE_PMD_SIZE
);
3162 pmdval
= pmdp_invalidate(vma
, address
, pvmw
->pmd
);
3163 if (pmd_dirty(pmdval
))
3164 set_page_dirty(page
);
3165 entry
= make_migration_entry(page
, pmd_write(pmdval
));
3166 pmdswp
= swp_entry_to_pmd(entry
);
3167 if (pmd_soft_dirty(pmdval
))
3168 pmdswp
= pmd_swp_mksoft_dirty(pmdswp
);
3169 set_pmd_at(mm
, address
, pvmw
->pmd
, pmdswp
);
3170 page_remove_rmap(page
, true);
3174 void remove_migration_pmd(struct page_vma_mapped_walk
*pvmw
, struct page
*new)
3176 struct vm_area_struct
*vma
= pvmw
->vma
;
3177 struct mm_struct
*mm
= vma
->vm_mm
;
3178 unsigned long address
= pvmw
->address
;
3179 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
3183 if (!(pvmw
->pmd
&& !pvmw
->pte
))
3186 entry
= pmd_to_swp_entry(*pvmw
->pmd
);
3188 pmde
= pmd_mkold(mk_huge_pmd(new, vma
->vm_page_prot
));
3189 if (pmd_swp_soft_dirty(*pvmw
->pmd
))
3190 pmde
= pmd_mksoft_dirty(pmde
);
3191 if (is_write_migration_entry(entry
))
3192 pmde
= maybe_pmd_mkwrite(pmde
, vma
);
3193 if (pmd_swp_uffd_wp(*pvmw
->pmd
))
3194 pmde
= pmd_wrprotect(pmd_mkuffd_wp(pmde
));
3196 flush_cache_range(vma
, mmun_start
, mmun_start
+ HPAGE_PMD_SIZE
);
3198 page_add_anon_rmap(new, vma
, mmun_start
, true);
3200 page_add_file_rmap(new, true);
3201 set_pmd_at(mm
, mmun_start
, pvmw
->pmd
, pmde
);
3202 if ((vma
->vm_flags
& VM_LOCKED
) && !PageDoubleMap(new))
3203 mlock_vma_page(new);
3204 update_mmu_cache_pmd(vma
, address
, pvmw
->pmd
);