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/coredump.h>
11 #include <linux/sched/numa_balancing.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/pfn_t.h>
24 #include <linux/mman.h>
25 #include <linux/memremap.h>
26 #include <linux/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/oom.h>
34 #include <linux/numa.h>
35 #include <linux/page_owner.h>
38 #include <asm/pgalloc.h>
42 * By default, transparent hugepage support is disabled in order to avoid
43 * risking an increased memory footprint for applications that are not
44 * guaranteed to benefit from it. When transparent hugepage support is
45 * enabled, it is for all mappings, and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
49 unsigned long transparent_hugepage_flags __read_mostly
=
50 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
53 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
60 static struct shrinker deferred_split_shrinker
;
62 static atomic_t huge_zero_refcount
;
63 struct page
*huge_zero_page __read_mostly
;
65 bool transparent_hugepage_enabled(struct vm_area_struct
*vma
)
67 /* The addr is used to check if the vma size fits */
68 unsigned long addr
= (vma
->vm_end
& HPAGE_PMD_MASK
) - HPAGE_PMD_SIZE
;
70 if (!transhuge_vma_suitable(vma
, addr
))
72 if (vma_is_anonymous(vma
))
73 return __transparent_hugepage_enabled(vma
);
74 if (vma_is_shmem(vma
))
75 return shmem_huge_enabled(vma
);
80 static struct page
*get_huge_zero_page(void)
82 struct page
*zero_page
;
84 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
85 return READ_ONCE(huge_zero_page
);
87 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
90 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
93 count_vm_event(THP_ZERO_PAGE_ALLOC
);
95 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
97 __free_pages(zero_page
, compound_order(zero_page
));
101 /* We take additional reference here. It will be put back by shrinker */
102 atomic_set(&huge_zero_refcount
, 2);
104 return READ_ONCE(huge_zero_page
);
107 static void put_huge_zero_page(void)
110 * Counter should never go to zero here. Only shrinker can put
113 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
116 struct page
*mm_get_huge_zero_page(struct mm_struct
*mm
)
118 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
119 return READ_ONCE(huge_zero_page
);
121 if (!get_huge_zero_page())
124 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
125 put_huge_zero_page();
127 return READ_ONCE(huge_zero_page
);
130 void mm_put_huge_zero_page(struct mm_struct
*mm
)
132 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
133 put_huge_zero_page();
136 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
137 struct shrink_control
*sc
)
139 /* we can free zero page only if last reference remains */
140 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
143 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
144 struct shrink_control
*sc
)
146 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
147 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
148 BUG_ON(zero_page
== NULL
);
149 __free_pages(zero_page
, compound_order(zero_page
));
156 static struct shrinker huge_zero_page_shrinker
= {
157 .count_objects
= shrink_huge_zero_page_count
,
158 .scan_objects
= shrink_huge_zero_page_scan
,
159 .seeks
= DEFAULT_SEEKS
,
163 static ssize_t
enabled_show(struct kobject
*kobj
,
164 struct kobj_attribute
*attr
, char *buf
)
166 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
))
167 return sprintf(buf
, "[always] madvise never\n");
168 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
169 return sprintf(buf
, "always [madvise] never\n");
171 return sprintf(buf
, "always madvise [never]\n");
174 static ssize_t
enabled_store(struct kobject
*kobj
,
175 struct kobj_attribute
*attr
,
176 const char *buf
, size_t count
)
180 if (sysfs_streq(buf
, "always")) {
181 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
182 set_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
183 } else if (sysfs_streq(buf
, "madvise")) {
184 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
185 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
186 } else if (sysfs_streq(buf
, "never")) {
187 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
188 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
193 int err
= start_stop_khugepaged();
199 static struct kobj_attribute enabled_attr
=
200 __ATTR(enabled
, 0644, enabled_show
, enabled_store
);
202 ssize_t
single_hugepage_flag_show(struct kobject
*kobj
,
203 struct kobj_attribute
*attr
, char *buf
,
204 enum transparent_hugepage_flag flag
)
206 return sprintf(buf
, "%d\n",
207 !!test_bit(flag
, &transparent_hugepage_flags
));
210 ssize_t
single_hugepage_flag_store(struct kobject
*kobj
,
211 struct kobj_attribute
*attr
,
212 const char *buf
, size_t count
,
213 enum transparent_hugepage_flag flag
)
218 ret
= kstrtoul(buf
, 10, &value
);
225 set_bit(flag
, &transparent_hugepage_flags
);
227 clear_bit(flag
, &transparent_hugepage_flags
);
232 static ssize_t
defrag_show(struct kobject
*kobj
,
233 struct kobj_attribute
*attr
, char *buf
)
235 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
236 return sprintf(buf
, "[always] defer defer+madvise madvise never\n");
237 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
238 return sprintf(buf
, "always [defer] defer+madvise madvise never\n");
239 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
240 return sprintf(buf
, "always defer [defer+madvise] madvise never\n");
241 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
242 return sprintf(buf
, "always defer defer+madvise [madvise] never\n");
243 return sprintf(buf
, "always defer defer+madvise madvise [never]\n");
246 static ssize_t
defrag_store(struct kobject
*kobj
,
247 struct kobj_attribute
*attr
,
248 const char *buf
, size_t count
)
250 if (sysfs_streq(buf
, "always")) {
251 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
254 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
255 } else if (sysfs_streq(buf
, "defer+madvise")) {
256 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
257 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
259 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
260 } else if (sysfs_streq(buf
, "defer")) {
261 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
262 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
263 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
264 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
265 } else if (sysfs_streq(buf
, "madvise")) {
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
268 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
269 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
270 } else if (sysfs_streq(buf
, "never")) {
271 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
272 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
273 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
274 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
280 static struct kobj_attribute defrag_attr
=
281 __ATTR(defrag
, 0644, defrag_show
, defrag_store
);
283 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
284 struct kobj_attribute
*attr
, char *buf
)
286 return single_hugepage_flag_show(kobj
, attr
, buf
,
287 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
289 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
290 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
292 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
293 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
295 static struct kobj_attribute use_zero_page_attr
=
296 __ATTR(use_zero_page
, 0644, use_zero_page_show
, use_zero_page_store
);
298 static ssize_t
hpage_pmd_size_show(struct kobject
*kobj
,
299 struct kobj_attribute
*attr
, char *buf
)
301 return sprintf(buf
, "%lu\n", HPAGE_PMD_SIZE
);
303 static struct kobj_attribute hpage_pmd_size_attr
=
304 __ATTR_RO(hpage_pmd_size
);
306 static struct attribute
*hugepage_attr
[] = {
309 &use_zero_page_attr
.attr
,
310 &hpage_pmd_size_attr
.attr
,
312 &shmem_enabled_attr
.attr
,
317 static const struct attribute_group hugepage_attr_group
= {
318 .attrs
= hugepage_attr
,
321 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
325 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
326 if (unlikely(!*hugepage_kobj
)) {
327 pr_err("failed to create transparent hugepage kobject\n");
331 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
333 pr_err("failed to register transparent hugepage group\n");
337 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
339 pr_err("failed to register transparent hugepage group\n");
340 goto remove_hp_group
;
346 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
348 kobject_put(*hugepage_kobj
);
352 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
354 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
355 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
356 kobject_put(hugepage_kobj
);
359 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
364 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
367 #endif /* CONFIG_SYSFS */
369 static int __init
hugepage_init(void)
372 struct kobject
*hugepage_kobj
;
374 if (!has_transparent_hugepage()) {
375 transparent_hugepage_flags
= 0;
380 * hugepages can't be allocated by the buddy allocator
382 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
>= MAX_ORDER
);
384 * we use page->mapping and page->index in second tail page
385 * as list_head: assuming THP order >= 2
387 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
389 err
= hugepage_init_sysfs(&hugepage_kobj
);
393 err
= khugepaged_init();
397 err
= register_shrinker(&huge_zero_page_shrinker
);
399 goto err_hzp_shrinker
;
400 err
= register_shrinker(&deferred_split_shrinker
);
402 goto err_split_shrinker
;
405 * By default disable transparent hugepages on smaller systems,
406 * where the extra memory used could hurt more than TLB overhead
407 * is likely to save. The admin can still enable it through /sys.
409 if (totalram_pages() < (512 << (20 - PAGE_SHIFT
))) {
410 transparent_hugepage_flags
= 0;
414 err
= start_stop_khugepaged();
420 unregister_shrinker(&deferred_split_shrinker
);
422 unregister_shrinker(&huge_zero_page_shrinker
);
424 khugepaged_destroy();
426 hugepage_exit_sysfs(hugepage_kobj
);
430 subsys_initcall(hugepage_init
);
432 static int __init
setup_transparent_hugepage(char *str
)
437 if (!strcmp(str
, "always")) {
438 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
439 &transparent_hugepage_flags
);
440 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
441 &transparent_hugepage_flags
);
443 } else if (!strcmp(str
, "madvise")) {
444 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
445 &transparent_hugepage_flags
);
446 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
447 &transparent_hugepage_flags
);
449 } else if (!strcmp(str
, "never")) {
450 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
451 &transparent_hugepage_flags
);
452 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
453 &transparent_hugepage_flags
);
458 pr_warn("transparent_hugepage= cannot parse, ignored\n");
461 __setup("transparent_hugepage=", setup_transparent_hugepage
);
463 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
465 if (likely(vma
->vm_flags
& VM_WRITE
))
466 pmd
= pmd_mkwrite(pmd
);
471 static inline struct deferred_split
*get_deferred_split_queue(struct page
*page
)
473 struct mem_cgroup
*memcg
= compound_head(page
)->mem_cgroup
;
474 struct pglist_data
*pgdat
= NODE_DATA(page_to_nid(page
));
477 return &memcg
->deferred_split_queue
;
479 return &pgdat
->deferred_split_queue
;
482 static inline struct deferred_split
*get_deferred_split_queue(struct page
*page
)
484 struct pglist_data
*pgdat
= NODE_DATA(page_to_nid(page
));
486 return &pgdat
->deferred_split_queue
;
490 void prep_transhuge_page(struct page
*page
)
493 * we use page->mapping and page->indexlru in second tail page
494 * as list_head: assuming THP order >= 2
497 INIT_LIST_HEAD(page_deferred_list(page
));
498 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
501 bool is_transparent_hugepage(struct page
*page
)
503 if (!PageCompound(page
))
506 page
= compound_head(page
);
507 return is_huge_zero_page(page
) ||
508 page
[1].compound_dtor
== TRANSHUGE_PAGE_DTOR
;
510 EXPORT_SYMBOL_GPL(is_transparent_hugepage
);
512 static unsigned long __thp_get_unmapped_area(struct file
*filp
,
513 unsigned long addr
, unsigned long len
,
514 loff_t off
, unsigned long flags
, unsigned long size
)
516 loff_t off_end
= off
+ len
;
517 loff_t off_align
= round_up(off
, size
);
518 unsigned long len_pad
, ret
;
520 if (off_end
<= off_align
|| (off_end
- off_align
) < size
)
523 len_pad
= len
+ size
;
524 if (len_pad
< len
|| (off
+ len_pad
) < off
)
527 ret
= current
->mm
->get_unmapped_area(filp
, addr
, len_pad
,
528 off
>> PAGE_SHIFT
, flags
);
531 * The failure might be due to length padding. The caller will retry
532 * without the padding.
534 if (IS_ERR_VALUE(ret
))
538 * Do not try to align to THP boundary if allocation at the address
544 ret
+= (off
- ret
) & (size
- 1);
548 unsigned long thp_get_unmapped_area(struct file
*filp
, unsigned long addr
,
549 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
552 loff_t off
= (loff_t
)pgoff
<< PAGE_SHIFT
;
554 if (!IS_DAX(filp
->f_mapping
->host
) || !IS_ENABLED(CONFIG_FS_DAX_PMD
))
557 ret
= __thp_get_unmapped_area(filp
, addr
, len
, off
, flags
, PMD_SIZE
);
561 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
563 EXPORT_SYMBOL_GPL(thp_get_unmapped_area
);
565 static vm_fault_t
__do_huge_pmd_anonymous_page(struct vm_fault
*vmf
,
566 struct page
*page
, gfp_t gfp
)
568 struct vm_area_struct
*vma
= vmf
->vma
;
570 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
573 VM_BUG_ON_PAGE(!PageCompound(page
), page
);
575 if (mem_cgroup_charge(page
, vma
->vm_mm
, gfp
)) {
577 count_vm_event(THP_FAULT_FALLBACK
);
578 count_vm_event(THP_FAULT_FALLBACK_CHARGE
);
579 return VM_FAULT_FALLBACK
;
581 cgroup_throttle_swaprate(page
, gfp
);
583 pgtable
= pte_alloc_one(vma
->vm_mm
);
584 if (unlikely(!pgtable
)) {
589 clear_huge_page(page
, vmf
->address
, HPAGE_PMD_NR
);
591 * The memory barrier inside __SetPageUptodate makes sure that
592 * clear_huge_page writes become visible before the set_pmd_at()
595 __SetPageUptodate(page
);
597 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
598 if (unlikely(!pmd_none(*vmf
->pmd
))) {
603 ret
= check_stable_address_space(vma
->vm_mm
);
607 /* Deliver the page fault to userland */
608 if (userfaultfd_missing(vma
)) {
611 spin_unlock(vmf
->ptl
);
613 pte_free(vma
->vm_mm
, pgtable
);
614 ret2
= handle_userfault(vmf
, VM_UFFD_MISSING
);
615 VM_BUG_ON(ret2
& VM_FAULT_FALLBACK
);
619 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
620 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
621 page_add_new_anon_rmap(page
, vma
, haddr
, true);
622 lru_cache_add_inactive_or_unevictable(page
, vma
);
623 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
624 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
625 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
626 mm_inc_nr_ptes(vma
->vm_mm
);
627 spin_unlock(vmf
->ptl
);
628 count_vm_event(THP_FAULT_ALLOC
);
629 count_memcg_event_mm(vma
->vm_mm
, THP_FAULT_ALLOC
);
634 spin_unlock(vmf
->ptl
);
637 pte_free(vma
->vm_mm
, pgtable
);
644 * always: directly stall for all thp allocations
645 * defer: wake kswapd and fail if not immediately available
646 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
647 * fail if not immediately available
648 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
650 * never: never stall for any thp allocation
652 static inline gfp_t
alloc_hugepage_direct_gfpmask(struct vm_area_struct
*vma
)
654 const bool vma_madvised
= !!(vma
->vm_flags
& VM_HUGEPAGE
);
656 /* Always do synchronous compaction */
657 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
658 return GFP_TRANSHUGE
| (vma_madvised
? 0 : __GFP_NORETRY
);
660 /* Kick kcompactd and fail quickly */
661 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
662 return GFP_TRANSHUGE_LIGHT
| __GFP_KSWAPD_RECLAIM
;
664 /* Synchronous compaction if madvised, otherwise kick kcompactd */
665 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
666 return GFP_TRANSHUGE_LIGHT
|
667 (vma_madvised
? __GFP_DIRECT_RECLAIM
:
668 __GFP_KSWAPD_RECLAIM
);
670 /* Only do synchronous compaction if madvised */
671 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
672 return GFP_TRANSHUGE_LIGHT
|
673 (vma_madvised
? __GFP_DIRECT_RECLAIM
: 0);
675 return GFP_TRANSHUGE_LIGHT
;
678 /* Caller must hold page table lock. */
679 static bool set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
680 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
681 struct page
*zero_page
)
686 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
687 entry
= pmd_mkhuge(entry
);
689 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
690 set_pmd_at(mm
, haddr
, pmd
, entry
);
695 vm_fault_t
do_huge_pmd_anonymous_page(struct vm_fault
*vmf
)
697 struct vm_area_struct
*vma
= vmf
->vma
;
700 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
702 if (!transhuge_vma_suitable(vma
, haddr
))
703 return VM_FAULT_FALLBACK
;
704 if (unlikely(anon_vma_prepare(vma
)))
706 if (unlikely(khugepaged_enter(vma
, vma
->vm_flags
)))
708 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
709 !mm_forbids_zeropage(vma
->vm_mm
) &&
710 transparent_hugepage_use_zero_page()) {
712 struct page
*zero_page
;
715 pgtable
= pte_alloc_one(vma
->vm_mm
);
716 if (unlikely(!pgtable
))
718 zero_page
= mm_get_huge_zero_page(vma
->vm_mm
);
719 if (unlikely(!zero_page
)) {
720 pte_free(vma
->vm_mm
, pgtable
);
721 count_vm_event(THP_FAULT_FALLBACK
);
722 return VM_FAULT_FALLBACK
;
724 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
727 if (pmd_none(*vmf
->pmd
)) {
728 ret
= check_stable_address_space(vma
->vm_mm
);
730 spin_unlock(vmf
->ptl
);
731 } else if (userfaultfd_missing(vma
)) {
732 spin_unlock(vmf
->ptl
);
733 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
734 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
736 set_huge_zero_page(pgtable
, vma
->vm_mm
, vma
,
737 haddr
, vmf
->pmd
, zero_page
);
738 spin_unlock(vmf
->ptl
);
742 spin_unlock(vmf
->ptl
);
744 pte_free(vma
->vm_mm
, pgtable
);
747 gfp
= alloc_hugepage_direct_gfpmask(vma
);
748 page
= alloc_hugepage_vma(gfp
, vma
, haddr
, HPAGE_PMD_ORDER
);
749 if (unlikely(!page
)) {
750 count_vm_event(THP_FAULT_FALLBACK
);
751 return VM_FAULT_FALLBACK
;
753 prep_transhuge_page(page
);
754 return __do_huge_pmd_anonymous_page(vmf
, page
, gfp
);
757 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
758 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
,
761 struct mm_struct
*mm
= vma
->vm_mm
;
765 ptl
= pmd_lock(mm
, pmd
);
766 if (!pmd_none(*pmd
)) {
768 if (pmd_pfn(*pmd
) != pfn_t_to_pfn(pfn
)) {
769 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd
));
772 entry
= pmd_mkyoung(*pmd
);
773 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
774 if (pmdp_set_access_flags(vma
, addr
, pmd
, entry
, 1))
775 update_mmu_cache_pmd(vma
, addr
, pmd
);
781 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
782 if (pfn_t_devmap(pfn
))
783 entry
= pmd_mkdevmap(entry
);
785 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
786 entry
= maybe_pmd_mkwrite(entry
, vma
);
790 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
795 set_pmd_at(mm
, addr
, pmd
, entry
);
796 update_mmu_cache_pmd(vma
, addr
, pmd
);
801 pte_free(mm
, pgtable
);
805 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
806 * @vmf: Structure describing the fault
807 * @pfn: pfn to insert
808 * @pgprot: page protection to use
809 * @write: whether it's a write fault
811 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
812 * also consult the vmf_insert_mixed_prot() documentation when
813 * @pgprot != @vmf->vma->vm_page_prot.
815 * Return: vm_fault_t value.
817 vm_fault_t
vmf_insert_pfn_pmd_prot(struct vm_fault
*vmf
, pfn_t pfn
,
818 pgprot_t pgprot
, bool write
)
820 unsigned long addr
= vmf
->address
& PMD_MASK
;
821 struct vm_area_struct
*vma
= vmf
->vma
;
822 pgtable_t pgtable
= NULL
;
825 * If we had pmd_special, we could avoid all these restrictions,
826 * but we need to be consistent with PTEs and architectures that
827 * can't support a 'special' bit.
829 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) &&
831 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
832 (VM_PFNMAP
|VM_MIXEDMAP
));
833 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
835 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
836 return VM_FAULT_SIGBUS
;
838 if (arch_needs_pgtable_deposit()) {
839 pgtable
= pte_alloc_one(vma
->vm_mm
);
844 track_pfn_insert(vma
, &pgprot
, pfn
);
846 insert_pfn_pmd(vma
, addr
, vmf
->pmd
, pfn
, pgprot
, write
, pgtable
);
847 return VM_FAULT_NOPAGE
;
849 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot
);
851 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
852 static pud_t
maybe_pud_mkwrite(pud_t pud
, struct vm_area_struct
*vma
)
854 if (likely(vma
->vm_flags
& VM_WRITE
))
855 pud
= pud_mkwrite(pud
);
859 static void insert_pfn_pud(struct vm_area_struct
*vma
, unsigned long addr
,
860 pud_t
*pud
, pfn_t pfn
, pgprot_t prot
, bool write
)
862 struct mm_struct
*mm
= vma
->vm_mm
;
866 ptl
= pud_lock(mm
, pud
);
867 if (!pud_none(*pud
)) {
869 if (pud_pfn(*pud
) != pfn_t_to_pfn(pfn
)) {
870 WARN_ON_ONCE(!is_huge_zero_pud(*pud
));
873 entry
= pud_mkyoung(*pud
);
874 entry
= maybe_pud_mkwrite(pud_mkdirty(entry
), vma
);
875 if (pudp_set_access_flags(vma
, addr
, pud
, entry
, 1))
876 update_mmu_cache_pud(vma
, addr
, pud
);
881 entry
= pud_mkhuge(pfn_t_pud(pfn
, prot
));
882 if (pfn_t_devmap(pfn
))
883 entry
= pud_mkdevmap(entry
);
885 entry
= pud_mkyoung(pud_mkdirty(entry
));
886 entry
= maybe_pud_mkwrite(entry
, vma
);
888 set_pud_at(mm
, addr
, pud
, entry
);
889 update_mmu_cache_pud(vma
, addr
, pud
);
896 * vmf_insert_pfn_pud_prot - insert a pud size pfn
897 * @vmf: Structure describing the fault
898 * @pfn: pfn to insert
899 * @pgprot: page protection to use
900 * @write: whether it's a write fault
902 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
903 * also consult the vmf_insert_mixed_prot() documentation when
904 * @pgprot != @vmf->vma->vm_page_prot.
906 * Return: vm_fault_t value.
908 vm_fault_t
vmf_insert_pfn_pud_prot(struct vm_fault
*vmf
, pfn_t pfn
,
909 pgprot_t pgprot
, bool write
)
911 unsigned long addr
= vmf
->address
& PUD_MASK
;
912 struct vm_area_struct
*vma
= vmf
->vma
;
915 * If we had pud_special, we could avoid all these restrictions,
916 * but we need to be consistent with PTEs and architectures that
917 * can't support a 'special' bit.
919 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) &&
921 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
922 (VM_PFNMAP
|VM_MIXEDMAP
));
923 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
925 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
926 return VM_FAULT_SIGBUS
;
928 track_pfn_insert(vma
, &pgprot
, pfn
);
930 insert_pfn_pud(vma
, addr
, vmf
->pud
, pfn
, pgprot
, write
);
931 return VM_FAULT_NOPAGE
;
933 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot
);
934 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
936 static void touch_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
937 pmd_t
*pmd
, int flags
)
941 _pmd
= pmd_mkyoung(*pmd
);
942 if (flags
& FOLL_WRITE
)
943 _pmd
= pmd_mkdirty(_pmd
);
944 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
945 pmd
, _pmd
, flags
& FOLL_WRITE
))
946 update_mmu_cache_pmd(vma
, addr
, pmd
);
949 struct page
*follow_devmap_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
950 pmd_t
*pmd
, int flags
, struct dev_pagemap
**pgmap
)
952 unsigned long pfn
= pmd_pfn(*pmd
);
953 struct mm_struct
*mm
= vma
->vm_mm
;
956 assert_spin_locked(pmd_lockptr(mm
, pmd
));
959 * When we COW a devmap PMD entry, we split it into PTEs, so we should
960 * not be in this function with `flags & FOLL_COW` set.
962 WARN_ONCE(flags
& FOLL_COW
, "mm: In follow_devmap_pmd with FOLL_COW set");
964 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
965 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
966 (FOLL_PIN
| FOLL_GET
)))
969 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
972 if (pmd_present(*pmd
) && pmd_devmap(*pmd
))
977 if (flags
& FOLL_TOUCH
)
978 touch_pmd(vma
, addr
, pmd
, flags
);
981 * device mapped pages can only be returned if the
982 * caller will manage the page reference count.
984 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
985 return ERR_PTR(-EEXIST
);
987 pfn
+= (addr
& ~PMD_MASK
) >> PAGE_SHIFT
;
988 *pgmap
= get_dev_pagemap(pfn
, *pgmap
);
990 return ERR_PTR(-EFAULT
);
991 page
= pfn_to_page(pfn
);
992 if (!try_grab_page(page
, flags
))
993 page
= ERR_PTR(-ENOMEM
);
998 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
999 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1000 struct vm_area_struct
*vma
)
1002 spinlock_t
*dst_ptl
, *src_ptl
;
1003 struct page
*src_page
;
1005 pgtable_t pgtable
= NULL
;
1008 /* Skip if can be re-fill on fault */
1009 if (!vma_is_anonymous(vma
))
1012 pgtable
= pte_alloc_one(dst_mm
);
1013 if (unlikely(!pgtable
))
1016 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
1017 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
1018 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1024 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1025 * does not have the VM_UFFD_WP, which means that the uffd
1026 * fork event is not enabled.
1028 if (!(vma
->vm_flags
& VM_UFFD_WP
))
1029 pmd
= pmd_clear_uffd_wp(pmd
);
1031 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1032 if (unlikely(is_swap_pmd(pmd
))) {
1033 swp_entry_t entry
= pmd_to_swp_entry(pmd
);
1035 VM_BUG_ON(!is_pmd_migration_entry(pmd
));
1036 if (is_write_migration_entry(entry
)) {
1037 make_migration_entry_read(&entry
);
1038 pmd
= swp_entry_to_pmd(entry
);
1039 if (pmd_swp_soft_dirty(*src_pmd
))
1040 pmd
= pmd_swp_mksoft_dirty(pmd
);
1041 set_pmd_at(src_mm
, addr
, src_pmd
, pmd
);
1043 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1044 mm_inc_nr_ptes(dst_mm
);
1045 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1046 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1052 if (unlikely(!pmd_trans_huge(pmd
))) {
1053 pte_free(dst_mm
, pgtable
);
1057 * When page table lock is held, the huge zero pmd should not be
1058 * under splitting since we don't split the page itself, only pmd to
1061 if (is_huge_zero_pmd(pmd
)) {
1062 struct page
*zero_page
;
1064 * get_huge_zero_page() will never allocate a new page here,
1065 * since we already have a zero page to copy. It just takes a
1068 zero_page
= mm_get_huge_zero_page(dst_mm
);
1069 set_huge_zero_page(pgtable
, dst_mm
, vma
, addr
, dst_pmd
,
1075 src_page
= pmd_page(pmd
);
1076 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
1079 * If this page is a potentially pinned page, split and retry the fault
1080 * with smaller page size. Normally this should not happen because the
1081 * userspace should use MADV_DONTFORK upon pinned regions. This is a
1082 * best effort that the pinned pages won't be replaced by another
1083 * random page during the coming copy-on-write.
1085 if (unlikely(is_cow_mapping(vma
->vm_flags
) &&
1086 atomic_read(&src_mm
->has_pinned
) &&
1087 page_maybe_dma_pinned(src_page
))) {
1088 pte_free(dst_mm
, pgtable
);
1089 spin_unlock(src_ptl
);
1090 spin_unlock(dst_ptl
);
1091 __split_huge_pmd(vma
, src_pmd
, addr
, false, NULL
);
1096 page_dup_rmap(src_page
, true);
1097 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1098 mm_inc_nr_ptes(dst_mm
);
1099 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1101 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
1102 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
1103 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1107 spin_unlock(src_ptl
);
1108 spin_unlock(dst_ptl
);
1113 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1114 static void touch_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1115 pud_t
*pud
, int flags
)
1119 _pud
= pud_mkyoung(*pud
);
1120 if (flags
& FOLL_WRITE
)
1121 _pud
= pud_mkdirty(_pud
);
1122 if (pudp_set_access_flags(vma
, addr
& HPAGE_PUD_MASK
,
1123 pud
, _pud
, flags
& FOLL_WRITE
))
1124 update_mmu_cache_pud(vma
, addr
, pud
);
1127 struct page
*follow_devmap_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1128 pud_t
*pud
, int flags
, struct dev_pagemap
**pgmap
)
1130 unsigned long pfn
= pud_pfn(*pud
);
1131 struct mm_struct
*mm
= vma
->vm_mm
;
1134 assert_spin_locked(pud_lockptr(mm
, pud
));
1136 if (flags
& FOLL_WRITE
&& !pud_write(*pud
))
1139 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1140 if (WARN_ON_ONCE((flags
& (FOLL_PIN
| FOLL_GET
)) ==
1141 (FOLL_PIN
| FOLL_GET
)))
1144 if (pud_present(*pud
) && pud_devmap(*pud
))
1149 if (flags
& FOLL_TOUCH
)
1150 touch_pud(vma
, addr
, pud
, flags
);
1153 * device mapped pages can only be returned if the
1154 * caller will manage the page reference count.
1156 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1158 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
1159 return ERR_PTR(-EEXIST
);
1161 pfn
+= (addr
& ~PUD_MASK
) >> PAGE_SHIFT
;
1162 *pgmap
= get_dev_pagemap(pfn
, *pgmap
);
1164 return ERR_PTR(-EFAULT
);
1165 page
= pfn_to_page(pfn
);
1166 if (!try_grab_page(page
, flags
))
1167 page
= ERR_PTR(-ENOMEM
);
1172 int copy_huge_pud(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1173 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1174 struct vm_area_struct
*vma
)
1176 spinlock_t
*dst_ptl
, *src_ptl
;
1180 dst_ptl
= pud_lock(dst_mm
, dst_pud
);
1181 src_ptl
= pud_lockptr(src_mm
, src_pud
);
1182 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1186 if (unlikely(!pud_trans_huge(pud
) && !pud_devmap(pud
)))
1190 * When page table lock is held, the huge zero pud should not be
1191 * under splitting since we don't split the page itself, only pud to
1194 if (is_huge_zero_pud(pud
)) {
1195 /* No huge zero pud yet */
1198 /* Please refer to comments in copy_huge_pmd() */
1199 if (unlikely(is_cow_mapping(vma
->vm_flags
) &&
1200 atomic_read(&src_mm
->has_pinned
) &&
1201 page_maybe_dma_pinned(pud_page(pud
)))) {
1202 spin_unlock(src_ptl
);
1203 spin_unlock(dst_ptl
);
1204 __split_huge_pud(vma
, src_pud
, addr
);
1208 pudp_set_wrprotect(src_mm
, addr
, src_pud
);
1209 pud
= pud_mkold(pud_wrprotect(pud
));
1210 set_pud_at(dst_mm
, addr
, dst_pud
, pud
);
1214 spin_unlock(src_ptl
);
1215 spin_unlock(dst_ptl
);
1219 void huge_pud_set_accessed(struct vm_fault
*vmf
, pud_t orig_pud
)
1222 unsigned long haddr
;
1223 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1225 vmf
->ptl
= pud_lock(vmf
->vma
->vm_mm
, vmf
->pud
);
1226 if (unlikely(!pud_same(*vmf
->pud
, orig_pud
)))
1229 entry
= pud_mkyoung(orig_pud
);
1231 entry
= pud_mkdirty(entry
);
1232 haddr
= vmf
->address
& HPAGE_PUD_MASK
;
1233 if (pudp_set_access_flags(vmf
->vma
, haddr
, vmf
->pud
, entry
, write
))
1234 update_mmu_cache_pud(vmf
->vma
, vmf
->address
, vmf
->pud
);
1237 spin_unlock(vmf
->ptl
);
1239 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1241 void huge_pmd_set_accessed(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1244 unsigned long haddr
;
1245 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1247 vmf
->ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1248 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
)))
1251 entry
= pmd_mkyoung(orig_pmd
);
1253 entry
= pmd_mkdirty(entry
);
1254 haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1255 if (pmdp_set_access_flags(vmf
->vma
, haddr
, vmf
->pmd
, entry
, write
))
1256 update_mmu_cache_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
);
1259 spin_unlock(vmf
->ptl
);
1262 vm_fault_t
do_huge_pmd_wp_page(struct vm_fault
*vmf
, pmd_t orig_pmd
)
1264 struct vm_area_struct
*vma
= vmf
->vma
;
1266 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1268 vmf
->ptl
= pmd_lockptr(vma
->vm_mm
, vmf
->pmd
);
1269 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1271 if (is_huge_zero_pmd(orig_pmd
))
1274 spin_lock(vmf
->ptl
);
1276 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1277 spin_unlock(vmf
->ptl
);
1281 page
= pmd_page(orig_pmd
);
1282 VM_BUG_ON_PAGE(!PageCompound(page
) || !PageHead(page
), page
);
1284 /* Lock page for reuse_swap_page() */
1285 if (!trylock_page(page
)) {
1287 spin_unlock(vmf
->ptl
);
1289 spin_lock(vmf
->ptl
);
1290 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1291 spin_unlock(vmf
->ptl
);
1300 * We can only reuse the page if nobody else maps the huge page or it's
1303 if (reuse_swap_page(page
, NULL
)) {
1305 entry
= pmd_mkyoung(orig_pmd
);
1306 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1307 if (pmdp_set_access_flags(vma
, haddr
, vmf
->pmd
, entry
, 1))
1308 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1310 spin_unlock(vmf
->ptl
);
1311 return VM_FAULT_WRITE
;
1315 spin_unlock(vmf
->ptl
);
1317 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
1318 return VM_FAULT_FALLBACK
;
1322 * FOLL_FORCE can write to even unwritable pmd's, but only
1323 * after we've gone through a COW cycle and they are dirty.
1325 static inline bool can_follow_write_pmd(pmd_t pmd
, unsigned int flags
)
1327 return pmd_write(pmd
) ||
1328 ((flags
& FOLL_FORCE
) && (flags
& FOLL_COW
) && pmd_dirty(pmd
));
1331 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1336 struct mm_struct
*mm
= vma
->vm_mm
;
1337 struct page
*page
= NULL
;
1339 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1341 if (flags
& FOLL_WRITE
&& !can_follow_write_pmd(*pmd
, flags
))
1344 /* Avoid dumping huge zero page */
1345 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1346 return ERR_PTR(-EFAULT
);
1348 /* Full NUMA hinting faults to serialise migration in fault paths */
1349 if ((flags
& FOLL_NUMA
) && pmd_protnone(*pmd
))
1352 page
= pmd_page(*pmd
);
1353 VM_BUG_ON_PAGE(!PageHead(page
) && !is_zone_device_page(page
), page
);
1355 if (!try_grab_page(page
, flags
))
1356 return ERR_PTR(-ENOMEM
);
1358 if (flags
& FOLL_TOUCH
)
1359 touch_pmd(vma
, addr
, pmd
, flags
);
1361 if ((flags
& FOLL_MLOCK
) && (vma
->vm_flags
& VM_LOCKED
)) {
1363 * We don't mlock() pte-mapped THPs. This way we can avoid
1364 * leaking mlocked pages into non-VM_LOCKED VMAs.
1368 * In most cases the pmd is the only mapping of the page as we
1369 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1370 * writable private mappings in populate_vma_page_range().
1372 * The only scenario when we have the page shared here is if we
1373 * mlocking read-only mapping shared over fork(). We skip
1374 * mlocking such pages.
1378 * We can expect PageDoubleMap() to be stable under page lock:
1379 * for file pages we set it in page_add_file_rmap(), which
1380 * requires page to be locked.
1383 if (PageAnon(page
) && compound_mapcount(page
) != 1)
1385 if (PageDoubleMap(page
) || !page
->mapping
)
1387 if (!trylock_page(page
))
1389 if (page
->mapping
&& !PageDoubleMap(page
))
1390 mlock_vma_page(page
);
1394 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1395 VM_BUG_ON_PAGE(!PageCompound(page
) && !is_zone_device_page(page
), page
);
1401 /* NUMA hinting page fault entry point for trans huge pmds */
1402 vm_fault_t
do_huge_pmd_numa_page(struct vm_fault
*vmf
, pmd_t pmd
)
1404 struct vm_area_struct
*vma
= vmf
->vma
;
1405 struct anon_vma
*anon_vma
= NULL
;
1407 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1408 int page_nid
= NUMA_NO_NODE
, this_nid
= numa_node_id();
1409 int target_nid
, last_cpupid
= -1;
1411 bool migrated
= false;
1415 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1416 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
)))
1420 * If there are potential migrations, wait for completion and retry
1421 * without disrupting NUMA hinting information. Do not relock and
1422 * check_same as the page may no longer be mapped.
1424 if (unlikely(pmd_trans_migrating(*vmf
->pmd
))) {
1425 page
= pmd_page(*vmf
->pmd
);
1426 if (!get_page_unless_zero(page
))
1428 spin_unlock(vmf
->ptl
);
1429 put_and_wait_on_page_locked(page
);
1433 page
= pmd_page(pmd
);
1434 BUG_ON(is_huge_zero_page(page
));
1435 page_nid
= page_to_nid(page
);
1436 last_cpupid
= page_cpupid_last(page
);
1437 count_vm_numa_event(NUMA_HINT_FAULTS
);
1438 if (page_nid
== this_nid
) {
1439 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL
);
1440 flags
|= TNF_FAULT_LOCAL
;
1443 /* See similar comment in do_numa_page for explanation */
1444 if (!pmd_savedwrite(pmd
))
1445 flags
|= TNF_NO_GROUP
;
1448 * Acquire the page lock to serialise THP migrations but avoid dropping
1449 * page_table_lock if at all possible
1451 page_locked
= trylock_page(page
);
1452 target_nid
= mpol_misplaced(page
, vma
, haddr
);
1453 if (target_nid
== NUMA_NO_NODE
) {
1454 /* If the page was locked, there are no parallel migrations */
1459 /* Migration could have started since the pmd_trans_migrating check */
1461 page_nid
= NUMA_NO_NODE
;
1462 if (!get_page_unless_zero(page
))
1464 spin_unlock(vmf
->ptl
);
1465 put_and_wait_on_page_locked(page
);
1470 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1471 * to serialises splits
1474 spin_unlock(vmf
->ptl
);
1475 anon_vma
= page_lock_anon_vma_read(page
);
1477 /* Confirm the PMD did not change while page_table_lock was released */
1478 spin_lock(vmf
->ptl
);
1479 if (unlikely(!pmd_same(pmd
, *vmf
->pmd
))) {
1482 page_nid
= NUMA_NO_NODE
;
1486 /* Bail if we fail to protect against THP splits for any reason */
1487 if (unlikely(!anon_vma
)) {
1489 page_nid
= NUMA_NO_NODE
;
1494 * Since we took the NUMA fault, we must have observed the !accessible
1495 * bit. Make sure all other CPUs agree with that, to avoid them
1496 * modifying the page we're about to migrate.
1498 * Must be done under PTL such that we'll observe the relevant
1499 * inc_tlb_flush_pending().
1501 * We are not sure a pending tlb flush here is for a huge page
1502 * mapping or not. Hence use the tlb range variant
1504 if (mm_tlb_flush_pending(vma
->vm_mm
)) {
1505 flush_tlb_range(vma
, haddr
, haddr
+ HPAGE_PMD_SIZE
);
1507 * change_huge_pmd() released the pmd lock before
1508 * invalidating the secondary MMUs sharing the primary
1509 * MMU pagetables (with ->invalidate_range()). The
1510 * mmu_notifier_invalidate_range_end() (which
1511 * internally calls ->invalidate_range()) in
1512 * change_pmd_range() will run after us, so we can't
1513 * rely on it here and we need an explicit invalidate.
1515 mmu_notifier_invalidate_range(vma
->vm_mm
, haddr
,
1516 haddr
+ HPAGE_PMD_SIZE
);
1520 * Migrate the THP to the requested node, returns with page unlocked
1521 * and access rights restored.
1523 spin_unlock(vmf
->ptl
);
1525 migrated
= migrate_misplaced_transhuge_page(vma
->vm_mm
, vma
,
1526 vmf
->pmd
, pmd
, vmf
->address
, page
, target_nid
);
1528 flags
|= TNF_MIGRATED
;
1529 page_nid
= target_nid
;
1531 flags
|= TNF_MIGRATE_FAIL
;
1535 BUG_ON(!PageLocked(page
));
1536 was_writable
= pmd_savedwrite(pmd
);
1537 pmd
= pmd_modify(pmd
, vma
->vm_page_prot
);
1538 pmd
= pmd_mkyoung(pmd
);
1540 pmd
= pmd_mkwrite(pmd
);
1541 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, pmd
);
1542 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1545 spin_unlock(vmf
->ptl
);
1549 page_unlock_anon_vma_read(anon_vma
);
1551 if (page_nid
!= NUMA_NO_NODE
)
1552 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
,
1559 * Return true if we do MADV_FREE successfully on entire pmd page.
1560 * Otherwise, return false.
1562 bool madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1563 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1568 struct mm_struct
*mm
= tlb
->mm
;
1571 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
1573 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1578 if (is_huge_zero_pmd(orig_pmd
))
1581 if (unlikely(!pmd_present(orig_pmd
))) {
1582 VM_BUG_ON(thp_migration_supported() &&
1583 !is_pmd_migration_entry(orig_pmd
));
1587 page
= pmd_page(orig_pmd
);
1589 * If other processes are mapping this page, we couldn't discard
1590 * the page unless they all do MADV_FREE so let's skip the page.
1592 if (page_mapcount(page
) != 1)
1595 if (!trylock_page(page
))
1599 * If user want to discard part-pages of THP, split it so MADV_FREE
1600 * will deactivate only them.
1602 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1605 split_huge_page(page
);
1611 if (PageDirty(page
))
1612 ClearPageDirty(page
);
1615 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1616 pmdp_invalidate(vma
, addr
, pmd
);
1617 orig_pmd
= pmd_mkold(orig_pmd
);
1618 orig_pmd
= pmd_mkclean(orig_pmd
);
1620 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1621 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1624 mark_page_lazyfree(page
);
1632 static inline void zap_deposited_table(struct mm_struct
*mm
, pmd_t
*pmd
)
1636 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1637 pte_free(mm
, pgtable
);
1641 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1642 pmd_t
*pmd
, unsigned long addr
)
1647 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
1649 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1653 * For architectures like ppc64 we look at deposited pgtable
1654 * when calling pmdp_huge_get_and_clear. So do the
1655 * pgtable_trans_huge_withdraw after finishing pmdp related
1658 orig_pmd
= pmdp_huge_get_and_clear_full(vma
, addr
, pmd
,
1660 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1661 if (vma_is_special_huge(vma
)) {
1662 if (arch_needs_pgtable_deposit())
1663 zap_deposited_table(tlb
->mm
, pmd
);
1665 if (is_huge_zero_pmd(orig_pmd
))
1666 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1667 } else if (is_huge_zero_pmd(orig_pmd
)) {
1668 zap_deposited_table(tlb
->mm
, pmd
);
1670 tlb_remove_page_size(tlb
, pmd_page(orig_pmd
), HPAGE_PMD_SIZE
);
1672 struct page
*page
= NULL
;
1673 int flush_needed
= 1;
1675 if (pmd_present(orig_pmd
)) {
1676 page
= pmd_page(orig_pmd
);
1677 page_remove_rmap(page
, true);
1678 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1679 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1680 } else if (thp_migration_supported()) {
1683 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd
));
1684 entry
= pmd_to_swp_entry(orig_pmd
);
1685 page
= pfn_to_page(swp_offset(entry
));
1688 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1690 if (PageAnon(page
)) {
1691 zap_deposited_table(tlb
->mm
, pmd
);
1692 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1694 if (arch_needs_pgtable_deposit())
1695 zap_deposited_table(tlb
->mm
, pmd
);
1696 add_mm_counter(tlb
->mm
, mm_counter_file(page
), -HPAGE_PMD_NR
);
1701 tlb_remove_page_size(tlb
, page
, HPAGE_PMD_SIZE
);
1706 #ifndef pmd_move_must_withdraw
1707 static inline int pmd_move_must_withdraw(spinlock_t
*new_pmd_ptl
,
1708 spinlock_t
*old_pmd_ptl
,
1709 struct vm_area_struct
*vma
)
1712 * With split pmd lock we also need to move preallocated
1713 * PTE page table if new_pmd is on different PMD page table.
1715 * We also don't deposit and withdraw tables for file pages.
1717 return (new_pmd_ptl
!= old_pmd_ptl
) && vma_is_anonymous(vma
);
1721 static pmd_t
move_soft_dirty_pmd(pmd_t pmd
)
1723 #ifdef CONFIG_MEM_SOFT_DIRTY
1724 if (unlikely(is_pmd_migration_entry(pmd
)))
1725 pmd
= pmd_swp_mksoft_dirty(pmd
);
1726 else if (pmd_present(pmd
))
1727 pmd
= pmd_mksoft_dirty(pmd
);
1732 bool move_huge_pmd(struct vm_area_struct
*vma
, unsigned long old_addr
,
1733 unsigned long new_addr
, pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1735 spinlock_t
*old_ptl
, *new_ptl
;
1737 struct mm_struct
*mm
= vma
->vm_mm
;
1738 bool force_flush
= false;
1741 * The destination pmd shouldn't be established, free_pgtables()
1742 * should have release it.
1744 if (WARN_ON(!pmd_none(*new_pmd
))) {
1745 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1750 * We don't have to worry about the ordering of src and dst
1751 * ptlocks because exclusive mmap_lock prevents deadlock.
1753 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1755 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1756 if (new_ptl
!= old_ptl
)
1757 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1758 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1759 if (pmd_present(pmd
))
1761 VM_BUG_ON(!pmd_none(*new_pmd
));
1763 if (pmd_move_must_withdraw(new_ptl
, old_ptl
, vma
)) {
1765 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1766 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1768 pmd
= move_soft_dirty_pmd(pmd
);
1769 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1771 flush_tlb_range(vma
, old_addr
, old_addr
+ PMD_SIZE
);
1772 if (new_ptl
!= old_ptl
)
1773 spin_unlock(new_ptl
);
1774 spin_unlock(old_ptl
);
1782 * - 0 if PMD could not be locked
1783 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1784 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1786 int change_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
1787 unsigned long addr
, pgprot_t newprot
, unsigned long cp_flags
)
1789 struct mm_struct
*mm
= vma
->vm_mm
;
1792 bool preserve_write
;
1794 bool prot_numa
= cp_flags
& MM_CP_PROT_NUMA
;
1795 bool uffd_wp
= cp_flags
& MM_CP_UFFD_WP
;
1796 bool uffd_wp_resolve
= cp_flags
& MM_CP_UFFD_WP_RESOLVE
;
1798 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1802 preserve_write
= prot_numa
&& pmd_write(*pmd
);
1805 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1806 if (is_swap_pmd(*pmd
)) {
1807 swp_entry_t entry
= pmd_to_swp_entry(*pmd
);
1809 VM_BUG_ON(!is_pmd_migration_entry(*pmd
));
1810 if (is_write_migration_entry(entry
)) {
1813 * A protection check is difficult so
1814 * just be safe and disable write
1816 make_migration_entry_read(&entry
);
1817 newpmd
= swp_entry_to_pmd(entry
);
1818 if (pmd_swp_soft_dirty(*pmd
))
1819 newpmd
= pmd_swp_mksoft_dirty(newpmd
);
1820 set_pmd_at(mm
, addr
, pmd
, newpmd
);
1827 * Avoid trapping faults against the zero page. The read-only
1828 * data is likely to be read-cached on the local CPU and
1829 * local/remote hits to the zero page are not interesting.
1831 if (prot_numa
&& is_huge_zero_pmd(*pmd
))
1834 if (prot_numa
&& pmd_protnone(*pmd
))
1838 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1839 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1840 * which is also under mmap_read_lock(mm):
1843 * change_huge_pmd(prot_numa=1)
1844 * pmdp_huge_get_and_clear_notify()
1845 * madvise_dontneed()
1847 * pmd_trans_huge(*pmd) == 0 (without ptl)
1850 * // pmd is re-established
1852 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1853 * which may break userspace.
1855 * pmdp_invalidate() is required to make sure we don't miss
1856 * dirty/young flags set by hardware.
1858 entry
= pmdp_invalidate(vma
, addr
, pmd
);
1860 entry
= pmd_modify(entry
, newprot
);
1862 entry
= pmd_mk_savedwrite(entry
);
1864 entry
= pmd_wrprotect(entry
);
1865 entry
= pmd_mkuffd_wp(entry
);
1866 } else if (uffd_wp_resolve
) {
1868 * Leave the write bit to be handled by PF interrupt
1869 * handler, then things like COW could be properly
1872 entry
= pmd_clear_uffd_wp(entry
);
1875 set_pmd_at(mm
, addr
, pmd
, entry
);
1876 BUG_ON(vma_is_anonymous(vma
) && !preserve_write
&& pmd_write(entry
));
1883 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1885 * Note that if it returns page table lock pointer, this routine returns without
1886 * unlocking page table lock. So callers must unlock it.
1888 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1891 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1892 if (likely(is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) ||
1900 * Returns true if a given pud maps a thp, false otherwise.
1902 * Note that if it returns true, this routine returns without unlocking page
1903 * table lock. So callers must unlock it.
1905 spinlock_t
*__pud_trans_huge_lock(pud_t
*pud
, struct vm_area_struct
*vma
)
1909 ptl
= pud_lock(vma
->vm_mm
, pud
);
1910 if (likely(pud_trans_huge(*pud
) || pud_devmap(*pud
)))
1916 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1917 int zap_huge_pud(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1918 pud_t
*pud
, unsigned long addr
)
1922 ptl
= __pud_trans_huge_lock(pud
, vma
);
1926 * For architectures like ppc64 we look at deposited pgtable
1927 * when calling pudp_huge_get_and_clear. So do the
1928 * pgtable_trans_huge_withdraw after finishing pudp related
1931 pudp_huge_get_and_clear_full(tlb
->mm
, addr
, pud
, tlb
->fullmm
);
1932 tlb_remove_pud_tlb_entry(tlb
, pud
, addr
);
1933 if (vma_is_special_huge(vma
)) {
1935 /* No zero page support yet */
1937 /* No support for anonymous PUD pages yet */
1943 static void __split_huge_pud_locked(struct vm_area_struct
*vma
, pud_t
*pud
,
1944 unsigned long haddr
)
1946 VM_BUG_ON(haddr
& ~HPAGE_PUD_MASK
);
1947 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
1948 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PUD_SIZE
, vma
);
1949 VM_BUG_ON(!pud_trans_huge(*pud
) && !pud_devmap(*pud
));
1951 count_vm_event(THP_SPLIT_PUD
);
1953 pudp_huge_clear_flush_notify(vma
, haddr
, pud
);
1956 void __split_huge_pud(struct vm_area_struct
*vma
, pud_t
*pud
,
1957 unsigned long address
)
1960 struct mmu_notifier_range range
;
1962 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
1963 address
& HPAGE_PUD_MASK
,
1964 (address
& HPAGE_PUD_MASK
) + HPAGE_PUD_SIZE
);
1965 mmu_notifier_invalidate_range_start(&range
);
1966 ptl
= pud_lock(vma
->vm_mm
, pud
);
1967 if (unlikely(!pud_trans_huge(*pud
) && !pud_devmap(*pud
)))
1969 __split_huge_pud_locked(vma
, pud
, range
.start
);
1974 * No need to double call mmu_notifier->invalidate_range() callback as
1975 * the above pudp_huge_clear_flush_notify() did already call it.
1977 mmu_notifier_invalidate_range_only_end(&range
);
1979 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1981 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
1982 unsigned long haddr
, pmd_t
*pmd
)
1984 struct mm_struct
*mm
= vma
->vm_mm
;
1990 * Leave pmd empty until pte is filled note that it is fine to delay
1991 * notification until mmu_notifier_invalidate_range_end() as we are
1992 * replacing a zero pmd write protected page with a zero pte write
1995 * See Documentation/vm/mmu_notifier.rst
1997 pmdp_huge_clear_flush(vma
, haddr
, pmd
);
1999 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2000 pmd_populate(mm
, &_pmd
, pgtable
);
2002 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2004 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2005 entry
= pte_mkspecial(entry
);
2006 pte
= pte_offset_map(&_pmd
, haddr
);
2007 VM_BUG_ON(!pte_none(*pte
));
2008 set_pte_at(mm
, haddr
, pte
, entry
);
2011 smp_wmb(); /* make pte visible before pmd */
2012 pmd_populate(mm
, pmd
, pgtable
);
2015 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2016 unsigned long haddr
, bool freeze
)
2018 struct mm_struct
*mm
= vma
->vm_mm
;
2021 pmd_t old_pmd
, _pmd
;
2022 bool young
, write
, soft_dirty
, pmd_migration
= false, uffd_wp
= false;
2026 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
2027 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2028 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
2029 VM_BUG_ON(!is_pmd_migration_entry(*pmd
) && !pmd_trans_huge(*pmd
)
2030 && !pmd_devmap(*pmd
));
2032 count_vm_event(THP_SPLIT_PMD
);
2034 if (!vma_is_anonymous(vma
)) {
2035 _pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2037 * We are going to unmap this huge page. So
2038 * just go ahead and zap it
2040 if (arch_needs_pgtable_deposit())
2041 zap_deposited_table(mm
, pmd
);
2042 if (vma_is_special_huge(vma
))
2044 page
= pmd_page(_pmd
);
2045 if (!PageDirty(page
) && pmd_dirty(_pmd
))
2046 set_page_dirty(page
);
2047 if (!PageReferenced(page
) && pmd_young(_pmd
))
2048 SetPageReferenced(page
);
2049 page_remove_rmap(page
, true);
2051 add_mm_counter(mm
, mm_counter_file(page
), -HPAGE_PMD_NR
);
2053 } else if (pmd_trans_huge(*pmd
) && is_huge_zero_pmd(*pmd
)) {
2055 * FIXME: Do we want to invalidate secondary mmu by calling
2056 * mmu_notifier_invalidate_range() see comments below inside
2057 * __split_huge_pmd() ?
2059 * We are going from a zero huge page write protected to zero
2060 * small page also write protected so it does not seems useful
2061 * to invalidate secondary mmu at this time.
2063 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2067 * Up to this point the pmd is present and huge and userland has the
2068 * whole access to the hugepage during the split (which happens in
2069 * place). If we overwrite the pmd with the not-huge version pointing
2070 * to the pte here (which of course we could if all CPUs were bug
2071 * free), userland could trigger a small page size TLB miss on the
2072 * small sized TLB while the hugepage TLB entry is still established in
2073 * the huge TLB. Some CPU doesn't like that.
2074 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2075 * 383 on page 105. Intel should be safe but is also warns that it's
2076 * only safe if the permission and cache attributes of the two entries
2077 * loaded in the two TLB is identical (which should be the case here).
2078 * But it is generally safer to never allow small and huge TLB entries
2079 * for the same virtual address to be loaded simultaneously. So instead
2080 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2081 * current pmd notpresent (atomically because here the pmd_trans_huge
2082 * must remain set at all times on the pmd until the split is complete
2083 * for this pmd), then we flush the SMP TLB and finally we write the
2084 * non-huge version of the pmd entry with pmd_populate.
2086 old_pmd
= pmdp_invalidate(vma
, haddr
, pmd
);
2088 pmd_migration
= is_pmd_migration_entry(old_pmd
);
2089 if (unlikely(pmd_migration
)) {
2092 entry
= pmd_to_swp_entry(old_pmd
);
2093 page
= pfn_to_page(swp_offset(entry
));
2094 write
= is_write_migration_entry(entry
);
2096 soft_dirty
= pmd_swp_soft_dirty(old_pmd
);
2097 uffd_wp
= pmd_swp_uffd_wp(old_pmd
);
2099 page
= pmd_page(old_pmd
);
2100 if (pmd_dirty(old_pmd
))
2102 write
= pmd_write(old_pmd
);
2103 young
= pmd_young(old_pmd
);
2104 soft_dirty
= pmd_soft_dirty(old_pmd
);
2105 uffd_wp
= pmd_uffd_wp(old_pmd
);
2107 VM_BUG_ON_PAGE(!page_count(page
), page
);
2108 page_ref_add(page
, HPAGE_PMD_NR
- 1);
2111 * Withdraw the table only after we mark the pmd entry invalid.
2112 * This's critical for some architectures (Power).
2114 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2115 pmd_populate(mm
, &_pmd
, pgtable
);
2117 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
2120 * Note that NUMA hinting access restrictions are not
2121 * transferred to avoid any possibility of altering
2122 * permissions across VMAs.
2124 if (freeze
|| pmd_migration
) {
2125 swp_entry_t swp_entry
;
2126 swp_entry
= make_migration_entry(page
+ i
, write
);
2127 entry
= swp_entry_to_pte(swp_entry
);
2129 entry
= pte_swp_mksoft_dirty(entry
);
2131 entry
= pte_swp_mkuffd_wp(entry
);
2133 entry
= mk_pte(page
+ i
, READ_ONCE(vma
->vm_page_prot
));
2134 entry
= maybe_mkwrite(entry
, vma
);
2136 entry
= pte_wrprotect(entry
);
2138 entry
= pte_mkold(entry
);
2140 entry
= pte_mksoft_dirty(entry
);
2142 entry
= pte_mkuffd_wp(entry
);
2144 pte
= pte_offset_map(&_pmd
, addr
);
2145 BUG_ON(!pte_none(*pte
));
2146 set_pte_at(mm
, addr
, pte
, entry
);
2148 atomic_inc(&page
[i
]._mapcount
);
2152 if (!pmd_migration
) {
2154 * Set PG_double_map before dropping compound_mapcount to avoid
2155 * false-negative page_mapped().
2157 if (compound_mapcount(page
) > 1 &&
2158 !TestSetPageDoubleMap(page
)) {
2159 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2160 atomic_inc(&page
[i
]._mapcount
);
2163 lock_page_memcg(page
);
2164 if (atomic_add_negative(-1, compound_mapcount_ptr(page
))) {
2165 /* Last compound_mapcount is gone. */
2166 __dec_lruvec_page_state(page
, NR_ANON_THPS
);
2167 if (TestClearPageDoubleMap(page
)) {
2168 /* No need in mapcount reference anymore */
2169 for (i
= 0; i
< HPAGE_PMD_NR
; i
++)
2170 atomic_dec(&page
[i
]._mapcount
);
2173 unlock_page_memcg(page
);
2176 smp_wmb(); /* make pte visible before pmd */
2177 pmd_populate(mm
, pmd
, pgtable
);
2180 for (i
= 0; i
< HPAGE_PMD_NR
; i
++) {
2181 page_remove_rmap(page
+ i
, false);
2187 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2188 unsigned long address
, bool freeze
, struct page
*page
)
2191 struct mmu_notifier_range range
;
2192 bool was_locked
= false;
2195 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, vma
->vm_mm
,
2196 address
& HPAGE_PMD_MASK
,
2197 (address
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
);
2198 mmu_notifier_invalidate_range_start(&range
);
2199 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
2202 * If caller asks to setup a migration entries, we need a page to check
2203 * pmd against. Otherwise we can end up replacing wrong page.
2205 VM_BUG_ON(freeze
&& !page
);
2207 VM_WARN_ON_ONCE(!PageLocked(page
));
2209 if (page
!= pmd_page(*pmd
))
2214 if (pmd_trans_huge(*pmd
)) {
2216 page
= pmd_page(*pmd
);
2217 if (unlikely(!trylock_page(page
))) {
2223 if (unlikely(!pmd_same(*pmd
, _pmd
))) {
2232 if (PageMlocked(page
))
2233 clear_page_mlock(page
);
2234 } else if (!(pmd_devmap(*pmd
) || is_pmd_migration_entry(*pmd
)))
2236 __split_huge_pmd_locked(vma
, pmd
, range
.start
, freeze
);
2239 if (!was_locked
&& page
)
2242 * No need to double call mmu_notifier->invalidate_range() callback.
2243 * They are 3 cases to consider inside __split_huge_pmd_locked():
2244 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2245 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2246 * fault will trigger a flush_notify before pointing to a new page
2247 * (it is fine if the secondary mmu keeps pointing to the old zero
2248 * page in the meantime)
2249 * 3) Split a huge pmd into pte pointing to the same page. No need
2250 * to invalidate secondary tlb entry they are all still valid.
2251 * any further changes to individual pte will notify. So no need
2252 * to call mmu_notifier->invalidate_range()
2254 mmu_notifier_invalidate_range_only_end(&range
);
2257 void split_huge_pmd_address(struct vm_area_struct
*vma
, unsigned long address
,
2258 bool freeze
, struct page
*page
)
2265 pgd
= pgd_offset(vma
->vm_mm
, address
);
2266 if (!pgd_present(*pgd
))
2269 p4d
= p4d_offset(pgd
, address
);
2270 if (!p4d_present(*p4d
))
2273 pud
= pud_offset(p4d
, address
);
2274 if (!pud_present(*pud
))
2277 pmd
= pmd_offset(pud
, address
);
2279 __split_huge_pmd(vma
, pmd
, address
, freeze
, page
);
2282 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2283 unsigned long start
,
2288 * If the new start address isn't hpage aligned and it could
2289 * previously contain an hugepage: check if we need to split
2292 if (start
& ~HPAGE_PMD_MASK
&&
2293 (start
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2294 (start
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2295 split_huge_pmd_address(vma
, start
, false, NULL
);
2298 * If the new end address isn't hpage aligned and it could
2299 * previously contain an hugepage: check if we need to split
2302 if (end
& ~HPAGE_PMD_MASK
&&
2303 (end
& HPAGE_PMD_MASK
) >= vma
->vm_start
&&
2304 (end
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= vma
->vm_end
)
2305 split_huge_pmd_address(vma
, end
, false, NULL
);
2308 * If we're also updating the vma->vm_next->vm_start, if the new
2309 * vm_next->vm_start isn't hpage aligned and it could previously
2310 * contain an hugepage: check if we need to split an huge pmd.
2312 if (adjust_next
> 0) {
2313 struct vm_area_struct
*next
= vma
->vm_next
;
2314 unsigned long nstart
= next
->vm_start
;
2315 nstart
+= adjust_next
;
2316 if (nstart
& ~HPAGE_PMD_MASK
&&
2317 (nstart
& HPAGE_PMD_MASK
) >= next
->vm_start
&&
2318 (nstart
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
<= next
->vm_end
)
2319 split_huge_pmd_address(next
, nstart
, false, NULL
);
2323 static void unmap_page(struct page
*page
)
2325 enum ttu_flags ttu_flags
= TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
|
2326 TTU_RMAP_LOCKED
| TTU_SPLIT_HUGE_PMD
;
2329 VM_BUG_ON_PAGE(!PageHead(page
), page
);
2332 ttu_flags
|= TTU_SPLIT_FREEZE
;
2334 unmap_success
= try_to_unmap(page
, ttu_flags
);
2335 VM_BUG_ON_PAGE(!unmap_success
, page
);
2338 static void remap_page(struct page
*page
, unsigned int nr
)
2341 if (PageTransHuge(page
)) {
2342 remove_migration_ptes(page
, page
, true);
2344 for (i
= 0; i
< nr
; i
++)
2345 remove_migration_ptes(page
+ i
, page
+ i
, true);
2349 static void __split_huge_page_tail(struct page
*head
, int tail
,
2350 struct lruvec
*lruvec
, struct list_head
*list
)
2352 struct page
*page_tail
= head
+ tail
;
2354 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
2357 * Clone page flags before unfreezing refcount.
2359 * After successful get_page_unless_zero() might follow flags change,
2360 * for exmaple lock_page() which set PG_waiters.
2362 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
2363 page_tail
->flags
|= (head
->flags
&
2364 ((1L << PG_referenced
) |
2365 (1L << PG_swapbacked
) |
2366 (1L << PG_swapcache
) |
2367 (1L << PG_mlocked
) |
2368 (1L << PG_uptodate
) |
2370 (1L << PG_workingset
) |
2372 (1L << PG_unevictable
) |
2378 /* ->mapping in first tail page is compound_mapcount */
2379 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
2381 page_tail
->mapping
= head
->mapping
;
2382 page_tail
->index
= head
->index
+ tail
;
2384 /* Page flags must be visible before we make the page non-compound. */
2388 * Clear PageTail before unfreezing page refcount.
2390 * After successful get_page_unless_zero() might follow put_page()
2391 * which needs correct compound_head().
2393 clear_compound_head(page_tail
);
2395 /* Finally unfreeze refcount. Additional reference from page cache. */
2396 page_ref_unfreeze(page_tail
, 1 + (!PageAnon(head
) ||
2397 PageSwapCache(head
)));
2399 if (page_is_young(head
))
2400 set_page_young(page_tail
);
2401 if (page_is_idle(head
))
2402 set_page_idle(page_tail
);
2404 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
2407 * always add to the tail because some iterators expect new
2408 * pages to show after the currently processed elements - e.g.
2411 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
2414 static void __split_huge_page(struct page
*page
, struct list_head
*list
,
2415 pgoff_t end
, unsigned long flags
)
2417 struct page
*head
= compound_head(page
);
2418 pg_data_t
*pgdat
= page_pgdat(head
);
2419 struct lruvec
*lruvec
;
2420 struct address_space
*swap_cache
= NULL
;
2421 unsigned long offset
= 0;
2422 unsigned int nr
= thp_nr_pages(head
);
2425 lruvec
= mem_cgroup_page_lruvec(head
, pgdat
);
2427 /* complete memcg works before add pages to LRU */
2428 mem_cgroup_split_huge_fixup(head
);
2430 if (PageAnon(head
) && PageSwapCache(head
)) {
2431 swp_entry_t entry
= { .val
= page_private(head
) };
2433 offset
= swp_offset(entry
);
2434 swap_cache
= swap_address_space(entry
);
2435 xa_lock(&swap_cache
->i_pages
);
2438 for (i
= nr
- 1; i
>= 1; i
--) {
2439 __split_huge_page_tail(head
, i
, lruvec
, list
);
2440 /* Some pages can be beyond i_size: drop them from page cache */
2441 if (head
[i
].index
>= end
) {
2442 ClearPageDirty(head
+ i
);
2443 __delete_from_page_cache(head
+ i
, NULL
);
2444 if (IS_ENABLED(CONFIG_SHMEM
) && PageSwapBacked(head
))
2445 shmem_uncharge(head
->mapping
->host
, 1);
2447 } else if (!PageAnon(page
)) {
2448 __xa_store(&head
->mapping
->i_pages
, head
[i
].index
,
2450 } else if (swap_cache
) {
2451 __xa_store(&swap_cache
->i_pages
, offset
+ i
,
2456 ClearPageCompound(head
);
2458 split_page_owner(head
, nr
);
2460 /* See comment in __split_huge_page_tail() */
2461 if (PageAnon(head
)) {
2462 /* Additional pin to swap cache */
2463 if (PageSwapCache(head
)) {
2464 page_ref_add(head
, 2);
2465 xa_unlock(&swap_cache
->i_pages
);
2470 /* Additional pin to page cache */
2471 page_ref_add(head
, 2);
2472 xa_unlock(&head
->mapping
->i_pages
);
2475 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
2477 remap_page(head
, nr
);
2479 if (PageSwapCache(head
)) {
2480 swp_entry_t entry
= { .val
= page_private(head
) };
2482 split_swap_cluster(entry
);
2485 for (i
= 0; i
< nr
; i
++) {
2486 struct page
*subpage
= head
+ i
;
2487 if (subpage
== page
)
2489 unlock_page(subpage
);
2492 * Subpages may be freed if there wasn't any mapping
2493 * like if add_to_swap() is running on a lru page that
2494 * had its mapping zapped. And freeing these pages
2495 * requires taking the lru_lock so we do the put_page
2496 * of the tail pages after the split is complete.
2502 int total_mapcount(struct page
*page
)
2504 int i
, compound
, nr
, ret
;
2506 VM_BUG_ON_PAGE(PageTail(page
), page
);
2508 if (likely(!PageCompound(page
)))
2509 return atomic_read(&page
->_mapcount
) + 1;
2511 compound
= compound_mapcount(page
);
2512 nr
= compound_nr(page
);
2516 for (i
= 0; i
< nr
; i
++)
2517 ret
+= atomic_read(&page
[i
]._mapcount
) + 1;
2518 /* File pages has compound_mapcount included in _mapcount */
2519 if (!PageAnon(page
))
2520 return ret
- compound
* nr
;
2521 if (PageDoubleMap(page
))
2527 * This calculates accurately how many mappings a transparent hugepage
2528 * has (unlike page_mapcount() which isn't fully accurate). This full
2529 * accuracy is primarily needed to know if copy-on-write faults can
2530 * reuse the page and change the mapping to read-write instead of
2531 * copying them. At the same time this returns the total_mapcount too.
2533 * The function returns the highest mapcount any one of the subpages
2534 * has. If the return value is one, even if different processes are
2535 * mapping different subpages of the transparent hugepage, they can
2536 * all reuse it, because each process is reusing a different subpage.
2538 * The total_mapcount is instead counting all virtual mappings of the
2539 * subpages. If the total_mapcount is equal to "one", it tells the
2540 * caller all mappings belong to the same "mm" and in turn the
2541 * anon_vma of the transparent hugepage can become the vma->anon_vma
2542 * local one as no other process may be mapping any of the subpages.
2544 * It would be more accurate to replace page_mapcount() with
2545 * page_trans_huge_mapcount(), however we only use
2546 * page_trans_huge_mapcount() in the copy-on-write faults where we
2547 * need full accuracy to avoid breaking page pinning, because
2548 * page_trans_huge_mapcount() is slower than page_mapcount().
2550 int page_trans_huge_mapcount(struct page
*page
, int *total_mapcount
)
2552 int i
, ret
, _total_mapcount
, mapcount
;
2554 /* hugetlbfs shouldn't call it */
2555 VM_BUG_ON_PAGE(PageHuge(page
), page
);
2557 if (likely(!PageTransCompound(page
))) {
2558 mapcount
= atomic_read(&page
->_mapcount
) + 1;
2560 *total_mapcount
= mapcount
;
2564 page
= compound_head(page
);
2566 _total_mapcount
= ret
= 0;
2567 for (i
= 0; i
< thp_nr_pages(page
); i
++) {
2568 mapcount
= atomic_read(&page
[i
]._mapcount
) + 1;
2569 ret
= max(ret
, mapcount
);
2570 _total_mapcount
+= mapcount
;
2572 if (PageDoubleMap(page
)) {
2574 _total_mapcount
-= thp_nr_pages(page
);
2576 mapcount
= compound_mapcount(page
);
2578 _total_mapcount
+= mapcount
;
2580 *total_mapcount
= _total_mapcount
;
2584 /* Racy check whether the huge page can be split */
2585 bool can_split_huge_page(struct page
*page
, int *pextra_pins
)
2589 /* Additional pins from page cache */
2591 extra_pins
= PageSwapCache(page
) ? thp_nr_pages(page
) : 0;
2593 extra_pins
= thp_nr_pages(page
);
2595 *pextra_pins
= extra_pins
;
2596 return total_mapcount(page
) == page_count(page
) - extra_pins
- 1;
2600 * This function splits huge page into normal pages. @page can point to any
2601 * subpage of huge page to split. Split doesn't change the position of @page.
2603 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2604 * The huge page must be locked.
2606 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2608 * Both head page and tail pages will inherit mapping, flags, and so on from
2611 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2612 * they are not mapped.
2614 * Returns 0 if the hugepage is split successfully.
2615 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2618 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
2620 struct page
*head
= compound_head(page
);
2621 struct pglist_data
*pgdata
= NODE_DATA(page_to_nid(head
));
2622 struct deferred_split
*ds_queue
= get_deferred_split_queue(head
);
2623 struct anon_vma
*anon_vma
= NULL
;
2624 struct address_space
*mapping
= NULL
;
2625 int count
, mapcount
, extra_pins
, ret
;
2626 unsigned long flags
;
2629 VM_BUG_ON_PAGE(is_huge_zero_page(head
), head
);
2630 VM_BUG_ON_PAGE(!PageLocked(head
), head
);
2631 VM_BUG_ON_PAGE(!PageCompound(head
), head
);
2633 if (PageWriteback(head
))
2636 if (PageAnon(head
)) {
2638 * The caller does not necessarily hold an mmap_lock that would
2639 * prevent the anon_vma disappearing so we first we take a
2640 * reference to it and then lock the anon_vma for write. This
2641 * is similar to page_lock_anon_vma_read except the write lock
2642 * is taken to serialise against parallel split or collapse
2645 anon_vma
= page_get_anon_vma(head
);
2652 anon_vma_lock_write(anon_vma
);
2654 mapping
= head
->mapping
;
2663 i_mmap_lock_read(mapping
);
2666 *__split_huge_page() may need to trim off pages beyond EOF:
2667 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2668 * which cannot be nested inside the page tree lock. So note
2669 * end now: i_size itself may be changed at any moment, but
2670 * head page lock is good enough to serialize the trimming.
2672 end
= DIV_ROUND_UP(i_size_read(mapping
->host
), PAGE_SIZE
);
2676 * Racy check if we can split the page, before unmap_page() will
2679 if (!can_split_huge_page(head
, &extra_pins
)) {
2685 VM_BUG_ON_PAGE(compound_mapcount(head
), head
);
2687 /* prevent PageLRU to go away from under us, and freeze lru stats */
2688 spin_lock_irqsave(&pgdata
->lru_lock
, flags
);
2691 XA_STATE(xas
, &mapping
->i_pages
, page_index(head
));
2694 * Check if the head page is present in page cache.
2695 * We assume all tail are present too, if head is there.
2697 xa_lock(&mapping
->i_pages
);
2698 if (xas_load(&xas
) != head
)
2702 /* Prevent deferred_split_scan() touching ->_refcount */
2703 spin_lock(&ds_queue
->split_queue_lock
);
2704 count
= page_count(head
);
2705 mapcount
= total_mapcount(head
);
2706 if (!mapcount
&& page_ref_freeze(head
, 1 + extra_pins
)) {
2707 if (!list_empty(page_deferred_list(head
))) {
2708 ds_queue
->split_queue_len
--;
2709 list_del(page_deferred_list(head
));
2711 spin_unlock(&ds_queue
->split_queue_lock
);
2713 if (PageSwapBacked(head
))
2714 __dec_node_page_state(head
, NR_SHMEM_THPS
);
2716 __dec_node_page_state(head
, NR_FILE_THPS
);
2719 __split_huge_page(page
, list
, end
, flags
);
2722 if (IS_ENABLED(CONFIG_DEBUG_VM
) && mapcount
) {
2723 pr_alert("total_mapcount: %u, page_count(): %u\n",
2726 dump_page(head
, NULL
);
2727 dump_page(page
, "total_mapcount(head) > 0");
2730 spin_unlock(&ds_queue
->split_queue_lock
);
2732 xa_unlock(&mapping
->i_pages
);
2733 spin_unlock_irqrestore(&pgdata
->lru_lock
, flags
);
2734 remap_page(head
, thp_nr_pages(head
));
2740 anon_vma_unlock_write(anon_vma
);
2741 put_anon_vma(anon_vma
);
2744 i_mmap_unlock_read(mapping
);
2746 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
2750 void free_transhuge_page(struct page
*page
)
2752 struct deferred_split
*ds_queue
= get_deferred_split_queue(page
);
2753 unsigned long flags
;
2755 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2756 if (!list_empty(page_deferred_list(page
))) {
2757 ds_queue
->split_queue_len
--;
2758 list_del(page_deferred_list(page
));
2760 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2761 free_compound_page(page
);
2764 void deferred_split_huge_page(struct page
*page
)
2766 struct deferred_split
*ds_queue
= get_deferred_split_queue(page
);
2768 struct mem_cgroup
*memcg
= compound_head(page
)->mem_cgroup
;
2770 unsigned long flags
;
2772 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
2775 * The try_to_unmap() in page reclaim path might reach here too,
2776 * this may cause a race condition to corrupt deferred split queue.
2777 * And, if page reclaim is already handling the same page, it is
2778 * unnecessary to handle it again in shrinker.
2780 * Check PageSwapCache to determine if the page is being
2781 * handled by page reclaim since THP swap would add the page into
2782 * swap cache before calling try_to_unmap().
2784 if (PageSwapCache(page
))
2787 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2788 if (list_empty(page_deferred_list(page
))) {
2789 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
2790 list_add_tail(page_deferred_list(page
), &ds_queue
->split_queue
);
2791 ds_queue
->split_queue_len
++;
2794 memcg_set_shrinker_bit(memcg
, page_to_nid(page
),
2795 deferred_split_shrinker
.id
);
2798 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2801 static unsigned long deferred_split_count(struct shrinker
*shrink
,
2802 struct shrink_control
*sc
)
2804 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2805 struct deferred_split
*ds_queue
= &pgdata
->deferred_split_queue
;
2809 ds_queue
= &sc
->memcg
->deferred_split_queue
;
2811 return READ_ONCE(ds_queue
->split_queue_len
);
2814 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
2815 struct shrink_control
*sc
)
2817 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2818 struct deferred_split
*ds_queue
= &pgdata
->deferred_split_queue
;
2819 unsigned long flags
;
2820 LIST_HEAD(list
), *pos
, *next
;
2826 ds_queue
= &sc
->memcg
->deferred_split_queue
;
2829 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2830 /* Take pin on all head pages to avoid freeing them under us */
2831 list_for_each_safe(pos
, next
, &ds_queue
->split_queue
) {
2832 page
= list_entry((void *)pos
, struct page
, mapping
);
2833 page
= compound_head(page
);
2834 if (get_page_unless_zero(page
)) {
2835 list_move(page_deferred_list(page
), &list
);
2837 /* We lost race with put_compound_page() */
2838 list_del_init(page_deferred_list(page
));
2839 ds_queue
->split_queue_len
--;
2841 if (!--sc
->nr_to_scan
)
2844 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2846 list_for_each_safe(pos
, next
, &list
) {
2847 page
= list_entry((void *)pos
, struct page
, mapping
);
2848 if (!trylock_page(page
))
2850 /* split_huge_page() removes page from list on success */
2851 if (!split_huge_page(page
))
2858 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2859 list_splice_tail(&list
, &ds_queue
->split_queue
);
2860 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2863 * Stop shrinker if we didn't split any page, but the queue is empty.
2864 * This can happen if pages were freed under us.
2866 if (!split
&& list_empty(&ds_queue
->split_queue
))
2871 static struct shrinker deferred_split_shrinker
= {
2872 .count_objects
= deferred_split_count
,
2873 .scan_objects
= deferred_split_scan
,
2874 .seeks
= DEFAULT_SEEKS
,
2875 .flags
= SHRINKER_NUMA_AWARE
| SHRINKER_MEMCG_AWARE
|
2879 #ifdef CONFIG_DEBUG_FS
2880 static int split_huge_pages_set(void *data
, u64 val
)
2884 unsigned long pfn
, max_zone_pfn
;
2885 unsigned long total
= 0, split
= 0;
2890 for_each_populated_zone(zone
) {
2891 max_zone_pfn
= zone_end_pfn(zone
);
2892 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
2893 if (!pfn_valid(pfn
))
2896 page
= pfn_to_page(pfn
);
2897 if (!get_page_unless_zero(page
))
2900 if (zone
!= page_zone(page
))
2903 if (!PageHead(page
) || PageHuge(page
) || !PageLRU(page
))
2908 if (!split_huge_page(page
))
2916 pr_info("%lu of %lu THP split\n", split
, total
);
2920 DEFINE_DEBUGFS_ATTRIBUTE(split_huge_pages_fops
, NULL
, split_huge_pages_set
,
2923 static int __init
split_huge_pages_debugfs(void)
2925 debugfs_create_file("split_huge_pages", 0200, NULL
, NULL
,
2926 &split_huge_pages_fops
);
2929 late_initcall(split_huge_pages_debugfs
);
2932 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2933 void set_pmd_migration_entry(struct page_vma_mapped_walk
*pvmw
,
2936 struct vm_area_struct
*vma
= pvmw
->vma
;
2937 struct mm_struct
*mm
= vma
->vm_mm
;
2938 unsigned long address
= pvmw
->address
;
2943 if (!(pvmw
->pmd
&& !pvmw
->pte
))
2946 flush_cache_range(vma
, address
, address
+ HPAGE_PMD_SIZE
);
2947 pmdval
= pmdp_invalidate(vma
, address
, pvmw
->pmd
);
2948 if (pmd_dirty(pmdval
))
2949 set_page_dirty(page
);
2950 entry
= make_migration_entry(page
, pmd_write(pmdval
));
2951 pmdswp
= swp_entry_to_pmd(entry
);
2952 if (pmd_soft_dirty(pmdval
))
2953 pmdswp
= pmd_swp_mksoft_dirty(pmdswp
);
2954 set_pmd_at(mm
, address
, pvmw
->pmd
, pmdswp
);
2955 page_remove_rmap(page
, true);
2959 void remove_migration_pmd(struct page_vma_mapped_walk
*pvmw
, struct page
*new)
2961 struct vm_area_struct
*vma
= pvmw
->vma
;
2962 struct mm_struct
*mm
= vma
->vm_mm
;
2963 unsigned long address
= pvmw
->address
;
2964 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
2968 if (!(pvmw
->pmd
&& !pvmw
->pte
))
2971 entry
= pmd_to_swp_entry(*pvmw
->pmd
);
2973 pmde
= pmd_mkold(mk_huge_pmd(new, vma
->vm_page_prot
));
2974 if (pmd_swp_soft_dirty(*pvmw
->pmd
))
2975 pmde
= pmd_mksoft_dirty(pmde
);
2976 if (is_write_migration_entry(entry
))
2977 pmde
= maybe_pmd_mkwrite(pmde
, vma
);
2979 flush_cache_range(vma
, mmun_start
, mmun_start
+ HPAGE_PMD_SIZE
);
2981 page_add_anon_rmap(new, vma
, mmun_start
, true);
2983 page_add_file_rmap(new, true);
2984 set_pmd_at(mm
, mmun_start
, pvmw
->pmd
, pmde
);
2985 if ((vma
->vm_flags
& VM_LOCKED
) && !PageDoubleMap(new))
2986 mlock_vma_page(new);
2987 update_mmu_cache_pmd(vma
, address
, pvmw
->pmd
);