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/backing-dev.h>
22 #include <linux/dax.h>
23 #include <linux/khugepaged.h>
24 #include <linux/freezer.h>
25 #include <linux/pfn_t.h>
26 #include <linux/mman.h>
27 #include <linux/memremap.h>
28 #include <linux/pagemap.h>
29 #include <linux/debugfs.h>
30 #include <linux/migrate.h>
31 #include <linux/hashtable.h>
32 #include <linux/userfaultfd_k.h>
33 #include <linux/page_idle.h>
34 #include <linux/shmem_fs.h>
35 #include <linux/oom.h>
36 #include <linux/numa.h>
37 #include <linux/page_owner.h>
38 #include <linux/sched/sysctl.h>
39 #include <linux/memory-tiers.h>
42 #include <asm/pgalloc.h>
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/thp.h>
50 * By default, transparent hugepage support is disabled in order to avoid
51 * risking an increased memory footprint for applications that are not
52 * guaranteed to benefit from it. When transparent hugepage support is
53 * enabled, it is for all mappings, and khugepaged scans all mappings.
54 * Defrag is invoked by khugepaged hugepage allocations and by page faults
55 * for all hugepage allocations.
57 unsigned long transparent_hugepage_flags __read_mostly
=
58 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
59 (1<<TRANSPARENT_HUGEPAGE_FLAG
)|
61 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
62 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
)|
64 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
)|
65 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG
)|
66 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
68 static struct shrinker deferred_split_shrinker
;
70 static atomic_t huge_zero_refcount
;
71 struct page
*huge_zero_page __read_mostly
;
72 unsigned long huge_zero_pfn __read_mostly
= ~0UL;
74 bool hugepage_vma_check(struct vm_area_struct
*vma
, unsigned long vm_flags
,
75 bool smaps
, bool in_pf
, bool enforce_sysfs
)
77 if (!vma
->vm_mm
) /* vdso */
81 * Explicitly disabled through madvise or prctl, or some
82 * architectures may disable THP for some mappings, for
85 if ((vm_flags
& VM_NOHUGEPAGE
) ||
86 test_bit(MMF_DISABLE_THP
, &vma
->vm_mm
->flags
))
89 * If the hardware/firmware marked hugepage support disabled.
91 if (transparent_hugepage_flags
& (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED
))
94 /* khugepaged doesn't collapse DAX vma, but page fault is fine. */
99 * Special VMA and hugetlb VMA.
100 * Must be checked after dax since some dax mappings may have
103 if (vm_flags
& VM_NO_KHUGEPAGED
)
107 * Check alignment for file vma and size for both file and anon vma.
109 * Skip the check for page fault. Huge fault does the check in fault
110 * handlers. And this check is not suitable for huge PUD fault.
113 !transhuge_vma_suitable(vma
, (vma
->vm_end
- HPAGE_PMD_SIZE
)))
117 * Enabled via shmem mount options or sysfs settings.
118 * Must be done before hugepage flags check since shmem has its
121 if (!in_pf
&& shmem_file(vma
->vm_file
))
122 return shmem_is_huge(file_inode(vma
->vm_file
), vma
->vm_pgoff
,
123 !enforce_sysfs
, vma
->vm_mm
, vm_flags
);
125 /* Enforce sysfs THP requirements as necessary */
127 (!hugepage_flags_enabled() || (!(vm_flags
& VM_HUGEPAGE
) &&
128 !hugepage_flags_always())))
131 /* Only regular file is valid */
132 if (!in_pf
&& file_thp_enabled(vma
))
135 if (!vma_is_anonymous(vma
))
138 if (vma_is_temporary_stack(vma
))
142 * THPeligible bit of smaps should show 1 for proper VMAs even
143 * though anon_vma is not initialized yet.
145 * Allow page fault since anon_vma may be not initialized until
146 * the first page fault.
149 return (smaps
|| in_pf
);
154 static bool get_huge_zero_page(void)
156 struct page
*zero_page
;
158 if (likely(atomic_inc_not_zero(&huge_zero_refcount
)))
161 zero_page
= alloc_pages((GFP_TRANSHUGE
| __GFP_ZERO
) & ~__GFP_MOVABLE
,
164 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED
);
168 if (cmpxchg(&huge_zero_page
, NULL
, zero_page
)) {
170 __free_pages(zero_page
, compound_order(zero_page
));
173 WRITE_ONCE(huge_zero_pfn
, page_to_pfn(zero_page
));
175 /* We take additional reference here. It will be put back by shrinker */
176 atomic_set(&huge_zero_refcount
, 2);
178 count_vm_event(THP_ZERO_PAGE_ALLOC
);
182 static void put_huge_zero_page(void)
185 * Counter should never go to zero here. Only shrinker can put
188 BUG_ON(atomic_dec_and_test(&huge_zero_refcount
));
191 struct page
*mm_get_huge_zero_page(struct mm_struct
*mm
)
193 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
194 return READ_ONCE(huge_zero_page
);
196 if (!get_huge_zero_page())
199 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
200 put_huge_zero_page();
202 return READ_ONCE(huge_zero_page
);
205 void mm_put_huge_zero_page(struct mm_struct
*mm
)
207 if (test_bit(MMF_HUGE_ZERO_PAGE
, &mm
->flags
))
208 put_huge_zero_page();
211 static unsigned long shrink_huge_zero_page_count(struct shrinker
*shrink
,
212 struct shrink_control
*sc
)
214 /* we can free zero page only if last reference remains */
215 return atomic_read(&huge_zero_refcount
) == 1 ? HPAGE_PMD_NR
: 0;
218 static unsigned long shrink_huge_zero_page_scan(struct shrinker
*shrink
,
219 struct shrink_control
*sc
)
221 if (atomic_cmpxchg(&huge_zero_refcount
, 1, 0) == 1) {
222 struct page
*zero_page
= xchg(&huge_zero_page
, NULL
);
223 BUG_ON(zero_page
== NULL
);
224 WRITE_ONCE(huge_zero_pfn
, ~0UL);
225 __free_pages(zero_page
, compound_order(zero_page
));
232 static struct shrinker huge_zero_page_shrinker
= {
233 .count_objects
= shrink_huge_zero_page_count
,
234 .scan_objects
= shrink_huge_zero_page_scan
,
235 .seeks
= DEFAULT_SEEKS
,
239 static ssize_t
enabled_show(struct kobject
*kobj
,
240 struct kobj_attribute
*attr
, char *buf
)
244 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
))
245 output
= "[always] madvise never";
246 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
247 &transparent_hugepage_flags
))
248 output
= "always [madvise] never";
250 output
= "always madvise [never]";
252 return sysfs_emit(buf
, "%s\n", output
);
255 static ssize_t
enabled_store(struct kobject
*kobj
,
256 struct kobj_attribute
*attr
,
257 const char *buf
, size_t count
)
261 if (sysfs_streq(buf
, "always")) {
262 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
263 set_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
264 } else if (sysfs_streq(buf
, "madvise")) {
265 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
266 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
267 } else if (sysfs_streq(buf
, "never")) {
268 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
, &transparent_hugepage_flags
);
269 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
274 int err
= start_stop_khugepaged();
281 static struct kobj_attribute enabled_attr
= __ATTR_RW(enabled
);
283 ssize_t
single_hugepage_flag_show(struct kobject
*kobj
,
284 struct kobj_attribute
*attr
, char *buf
,
285 enum transparent_hugepage_flag flag
)
287 return sysfs_emit(buf
, "%d\n",
288 !!test_bit(flag
, &transparent_hugepage_flags
));
291 ssize_t
single_hugepage_flag_store(struct kobject
*kobj
,
292 struct kobj_attribute
*attr
,
293 const char *buf
, size_t count
,
294 enum transparent_hugepage_flag flag
)
299 ret
= kstrtoul(buf
, 10, &value
);
306 set_bit(flag
, &transparent_hugepage_flags
);
308 clear_bit(flag
, &transparent_hugepage_flags
);
313 static ssize_t
defrag_show(struct kobject
*kobj
,
314 struct kobj_attribute
*attr
, char *buf
)
318 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
,
319 &transparent_hugepage_flags
))
320 output
= "[always] defer defer+madvise madvise never";
321 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
,
322 &transparent_hugepage_flags
))
323 output
= "always [defer] defer+madvise madvise never";
324 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
,
325 &transparent_hugepage_flags
))
326 output
= "always defer [defer+madvise] madvise never";
327 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
,
328 &transparent_hugepage_flags
))
329 output
= "always defer defer+madvise [madvise] never";
331 output
= "always defer defer+madvise madvise [never]";
333 return sysfs_emit(buf
, "%s\n", output
);
336 static ssize_t
defrag_store(struct kobject
*kobj
,
337 struct kobj_attribute
*attr
,
338 const char *buf
, size_t count
)
340 if (sysfs_streq(buf
, "always")) {
341 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
342 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
343 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
344 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
345 } else if (sysfs_streq(buf
, "defer+madvise")) {
346 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
347 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
348 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
349 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
350 } else if (sysfs_streq(buf
, "defer")) {
351 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
352 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
353 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
354 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
355 } else if (sysfs_streq(buf
, "madvise")) {
356 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
357 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
358 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
359 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
360 } else if (sysfs_streq(buf
, "never")) {
361 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
);
362 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
);
363 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
);
364 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
);
370 static struct kobj_attribute defrag_attr
= __ATTR_RW(defrag
);
372 static ssize_t
use_zero_page_show(struct kobject
*kobj
,
373 struct kobj_attribute
*attr
, char *buf
)
375 return single_hugepage_flag_show(kobj
, attr
, buf
,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
378 static ssize_t
use_zero_page_store(struct kobject
*kobj
,
379 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
381 return single_hugepage_flag_store(kobj
, attr
, buf
, count
,
382 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG
);
384 static struct kobj_attribute use_zero_page_attr
= __ATTR_RW(use_zero_page
);
386 static ssize_t
hpage_pmd_size_show(struct kobject
*kobj
,
387 struct kobj_attribute
*attr
, char *buf
)
389 return sysfs_emit(buf
, "%lu\n", HPAGE_PMD_SIZE
);
391 static struct kobj_attribute hpage_pmd_size_attr
=
392 __ATTR_RO(hpage_pmd_size
);
394 static struct attribute
*hugepage_attr
[] = {
397 &use_zero_page_attr
.attr
,
398 &hpage_pmd_size_attr
.attr
,
400 &shmem_enabled_attr
.attr
,
405 static const struct attribute_group hugepage_attr_group
= {
406 .attrs
= hugepage_attr
,
409 static int __init
hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
413 *hugepage_kobj
= kobject_create_and_add("transparent_hugepage", mm_kobj
);
414 if (unlikely(!*hugepage_kobj
)) {
415 pr_err("failed to create transparent hugepage kobject\n");
419 err
= sysfs_create_group(*hugepage_kobj
, &hugepage_attr_group
);
421 pr_err("failed to register transparent hugepage group\n");
425 err
= sysfs_create_group(*hugepage_kobj
, &khugepaged_attr_group
);
427 pr_err("failed to register transparent hugepage group\n");
428 goto remove_hp_group
;
434 sysfs_remove_group(*hugepage_kobj
, &hugepage_attr_group
);
436 kobject_put(*hugepage_kobj
);
440 static void __init
hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
442 sysfs_remove_group(hugepage_kobj
, &khugepaged_attr_group
);
443 sysfs_remove_group(hugepage_kobj
, &hugepage_attr_group
);
444 kobject_put(hugepage_kobj
);
447 static inline int hugepage_init_sysfs(struct kobject
**hugepage_kobj
)
452 static inline void hugepage_exit_sysfs(struct kobject
*hugepage_kobj
)
455 #endif /* CONFIG_SYSFS */
457 static int __init
hugepage_init(void)
460 struct kobject
*hugepage_kobj
;
462 if (!has_transparent_hugepage()) {
463 transparent_hugepage_flags
= 1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED
;
468 * hugepages can't be allocated by the buddy allocator
470 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
> MAX_ORDER
);
472 * we use page->mapping and page->index in second tail page
473 * as list_head: assuming THP order >= 2
475 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER
< 2);
477 err
= hugepage_init_sysfs(&hugepage_kobj
);
481 err
= khugepaged_init();
485 err
= register_shrinker(&huge_zero_page_shrinker
, "thp-zero");
487 goto err_hzp_shrinker
;
488 err
= register_shrinker(&deferred_split_shrinker
, "thp-deferred_split");
490 goto err_split_shrinker
;
493 * By default disable transparent hugepages on smaller systems,
494 * where the extra memory used could hurt more than TLB overhead
495 * is likely to save. The admin can still enable it through /sys.
497 if (totalram_pages() < (512 << (20 - PAGE_SHIFT
))) {
498 transparent_hugepage_flags
= 0;
502 err
= start_stop_khugepaged();
508 unregister_shrinker(&deferred_split_shrinker
);
510 unregister_shrinker(&huge_zero_page_shrinker
);
512 khugepaged_destroy();
514 hugepage_exit_sysfs(hugepage_kobj
);
518 subsys_initcall(hugepage_init
);
520 static int __init
setup_transparent_hugepage(char *str
)
525 if (!strcmp(str
, "always")) {
526 set_bit(TRANSPARENT_HUGEPAGE_FLAG
,
527 &transparent_hugepage_flags
);
528 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
529 &transparent_hugepage_flags
);
531 } else if (!strcmp(str
, "madvise")) {
532 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
533 &transparent_hugepage_flags
);
534 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
535 &transparent_hugepage_flags
);
537 } else if (!strcmp(str
, "never")) {
538 clear_bit(TRANSPARENT_HUGEPAGE_FLAG
,
539 &transparent_hugepage_flags
);
540 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG
,
541 &transparent_hugepage_flags
);
546 pr_warn("transparent_hugepage= cannot parse, ignored\n");
549 __setup("transparent_hugepage=", setup_transparent_hugepage
);
551 pmd_t
maybe_pmd_mkwrite(pmd_t pmd
, struct vm_area_struct
*vma
)
553 if (likely(vma
->vm_flags
& VM_WRITE
))
554 pmd
= pmd_mkwrite(pmd
);
560 struct deferred_split
*get_deferred_split_queue(struct folio
*folio
)
562 struct mem_cgroup
*memcg
= folio_memcg(folio
);
563 struct pglist_data
*pgdat
= NODE_DATA(folio_nid(folio
));
566 return &memcg
->deferred_split_queue
;
568 return &pgdat
->deferred_split_queue
;
572 struct deferred_split
*get_deferred_split_queue(struct folio
*folio
)
574 struct pglist_data
*pgdat
= NODE_DATA(folio_nid(folio
));
576 return &pgdat
->deferred_split_queue
;
580 void prep_transhuge_page(struct page
*page
)
582 struct folio
*folio
= (struct folio
*)page
;
584 VM_BUG_ON_FOLIO(folio_order(folio
) < 2, folio
);
585 INIT_LIST_HEAD(&folio
->_deferred_list
);
586 set_compound_page_dtor(page
, TRANSHUGE_PAGE_DTOR
);
589 static inline bool is_transparent_hugepage(struct page
*page
)
593 if (!PageCompound(page
))
596 folio
= page_folio(page
);
597 return is_huge_zero_page(&folio
->page
) ||
598 folio
->_folio_dtor
== TRANSHUGE_PAGE_DTOR
;
601 static unsigned long __thp_get_unmapped_area(struct file
*filp
,
602 unsigned long addr
, unsigned long len
,
603 loff_t off
, unsigned long flags
, unsigned long size
)
605 loff_t off_end
= off
+ len
;
606 loff_t off_align
= round_up(off
, size
);
607 unsigned long len_pad
, ret
;
609 if (off_end
<= off_align
|| (off_end
- off_align
) < size
)
612 len_pad
= len
+ size
;
613 if (len_pad
< len
|| (off
+ len_pad
) < off
)
616 ret
= current
->mm
->get_unmapped_area(filp
, addr
, len_pad
,
617 off
>> PAGE_SHIFT
, flags
);
620 * The failure might be due to length padding. The caller will retry
621 * without the padding.
623 if (IS_ERR_VALUE(ret
))
627 * Do not try to align to THP boundary if allocation at the address
633 ret
+= (off
- ret
) & (size
- 1);
637 unsigned long thp_get_unmapped_area(struct file
*filp
, unsigned long addr
,
638 unsigned long len
, unsigned long pgoff
, unsigned long flags
)
641 loff_t off
= (loff_t
)pgoff
<< PAGE_SHIFT
;
643 ret
= __thp_get_unmapped_area(filp
, addr
, len
, off
, flags
, PMD_SIZE
);
647 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
649 EXPORT_SYMBOL_GPL(thp_get_unmapped_area
);
651 static vm_fault_t
__do_huge_pmd_anonymous_page(struct vm_fault
*vmf
,
652 struct page
*page
, gfp_t gfp
)
654 struct vm_area_struct
*vma
= vmf
->vma
;
655 struct folio
*folio
= page_folio(page
);
657 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
660 VM_BUG_ON_FOLIO(!folio_test_large(folio
), folio
);
662 if (mem_cgroup_charge(folio
, vma
->vm_mm
, gfp
)) {
664 count_vm_event(THP_FAULT_FALLBACK
);
665 count_vm_event(THP_FAULT_FALLBACK_CHARGE
);
666 return VM_FAULT_FALLBACK
;
668 folio_throttle_swaprate(folio
, gfp
);
670 pgtable
= pte_alloc_one(vma
->vm_mm
);
671 if (unlikely(!pgtable
)) {
676 clear_huge_page(page
, vmf
->address
, HPAGE_PMD_NR
);
678 * The memory barrier inside __folio_mark_uptodate makes sure that
679 * clear_huge_page writes become visible before the set_pmd_at()
682 __folio_mark_uptodate(folio
);
684 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
685 if (unlikely(!pmd_none(*vmf
->pmd
))) {
690 ret
= check_stable_address_space(vma
->vm_mm
);
694 /* Deliver the page fault to userland */
695 if (userfaultfd_missing(vma
)) {
696 spin_unlock(vmf
->ptl
);
698 pte_free(vma
->vm_mm
, pgtable
);
699 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
700 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
704 entry
= mk_huge_pmd(page
, vma
->vm_page_prot
);
705 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
706 folio_add_new_anon_rmap(folio
, vma
, haddr
);
707 folio_add_lru_vma(folio
, vma
);
708 pgtable_trans_huge_deposit(vma
->vm_mm
, vmf
->pmd
, pgtable
);
709 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, entry
);
710 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
711 add_mm_counter(vma
->vm_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
712 mm_inc_nr_ptes(vma
->vm_mm
);
713 spin_unlock(vmf
->ptl
);
714 count_vm_event(THP_FAULT_ALLOC
);
715 count_memcg_event_mm(vma
->vm_mm
, THP_FAULT_ALLOC
);
720 spin_unlock(vmf
->ptl
);
723 pte_free(vma
->vm_mm
, pgtable
);
730 * always: directly stall for all thp allocations
731 * defer: wake kswapd and fail if not immediately available
732 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
733 * fail if not immediately available
734 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
736 * never: never stall for any thp allocation
738 gfp_t
vma_thp_gfp_mask(struct vm_area_struct
*vma
)
740 const bool vma_madvised
= vma
&& (vma
->vm_flags
& VM_HUGEPAGE
);
742 /* Always do synchronous compaction */
743 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG
, &transparent_hugepage_flags
))
744 return GFP_TRANSHUGE
| (vma_madvised
? 0 : __GFP_NORETRY
);
746 /* Kick kcompactd and fail quickly */
747 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG
, &transparent_hugepage_flags
))
748 return GFP_TRANSHUGE_LIGHT
| __GFP_KSWAPD_RECLAIM
;
750 /* Synchronous compaction if madvised, otherwise kick kcompactd */
751 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG
, &transparent_hugepage_flags
))
752 return GFP_TRANSHUGE_LIGHT
|
753 (vma_madvised
? __GFP_DIRECT_RECLAIM
:
754 __GFP_KSWAPD_RECLAIM
);
756 /* Only do synchronous compaction if madvised */
757 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG
, &transparent_hugepage_flags
))
758 return GFP_TRANSHUGE_LIGHT
|
759 (vma_madvised
? __GFP_DIRECT_RECLAIM
: 0);
761 return GFP_TRANSHUGE_LIGHT
;
764 /* Caller must hold page table lock. */
765 static void set_huge_zero_page(pgtable_t pgtable
, struct mm_struct
*mm
,
766 struct vm_area_struct
*vma
, unsigned long haddr
, pmd_t
*pmd
,
767 struct page
*zero_page
)
772 entry
= mk_pmd(zero_page
, vma
->vm_page_prot
);
773 entry
= pmd_mkhuge(entry
);
774 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
775 set_pmd_at(mm
, haddr
, pmd
, entry
);
779 vm_fault_t
do_huge_pmd_anonymous_page(struct vm_fault
*vmf
)
781 struct vm_area_struct
*vma
= vmf
->vma
;
784 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
786 if (!transhuge_vma_suitable(vma
, haddr
))
787 return VM_FAULT_FALLBACK
;
788 if (unlikely(anon_vma_prepare(vma
)))
790 khugepaged_enter_vma(vma
, vma
->vm_flags
);
792 if (!(vmf
->flags
& FAULT_FLAG_WRITE
) &&
793 !mm_forbids_zeropage(vma
->vm_mm
) &&
794 transparent_hugepage_use_zero_page()) {
796 struct page
*zero_page
;
798 pgtable
= pte_alloc_one(vma
->vm_mm
);
799 if (unlikely(!pgtable
))
801 zero_page
= mm_get_huge_zero_page(vma
->vm_mm
);
802 if (unlikely(!zero_page
)) {
803 pte_free(vma
->vm_mm
, pgtable
);
804 count_vm_event(THP_FAULT_FALLBACK
);
805 return VM_FAULT_FALLBACK
;
807 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
809 if (pmd_none(*vmf
->pmd
)) {
810 ret
= check_stable_address_space(vma
->vm_mm
);
812 spin_unlock(vmf
->ptl
);
813 pte_free(vma
->vm_mm
, pgtable
);
814 } else if (userfaultfd_missing(vma
)) {
815 spin_unlock(vmf
->ptl
);
816 pte_free(vma
->vm_mm
, pgtable
);
817 ret
= handle_userfault(vmf
, VM_UFFD_MISSING
);
818 VM_BUG_ON(ret
& VM_FAULT_FALLBACK
);
820 set_huge_zero_page(pgtable
, vma
->vm_mm
, vma
,
821 haddr
, vmf
->pmd
, zero_page
);
822 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
823 spin_unlock(vmf
->ptl
);
826 spin_unlock(vmf
->ptl
);
827 pte_free(vma
->vm_mm
, pgtable
);
831 gfp
= vma_thp_gfp_mask(vma
);
832 folio
= vma_alloc_folio(gfp
, HPAGE_PMD_ORDER
, vma
, haddr
, true);
833 if (unlikely(!folio
)) {
834 count_vm_event(THP_FAULT_FALLBACK
);
835 return VM_FAULT_FALLBACK
;
837 return __do_huge_pmd_anonymous_page(vmf
, &folio
->page
, gfp
);
840 static void insert_pfn_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
841 pmd_t
*pmd
, pfn_t pfn
, pgprot_t prot
, bool write
,
844 struct mm_struct
*mm
= vma
->vm_mm
;
848 ptl
= pmd_lock(mm
, pmd
);
849 if (!pmd_none(*pmd
)) {
851 if (pmd_pfn(*pmd
) != pfn_t_to_pfn(pfn
)) {
852 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd
));
855 entry
= pmd_mkyoung(*pmd
);
856 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
857 if (pmdp_set_access_flags(vma
, addr
, pmd
, entry
, 1))
858 update_mmu_cache_pmd(vma
, addr
, pmd
);
864 entry
= pmd_mkhuge(pfn_t_pmd(pfn
, prot
));
865 if (pfn_t_devmap(pfn
))
866 entry
= pmd_mkdevmap(entry
);
868 entry
= pmd_mkyoung(pmd_mkdirty(entry
));
869 entry
= maybe_pmd_mkwrite(entry
, vma
);
873 pgtable_trans_huge_deposit(mm
, pmd
, pgtable
);
878 set_pmd_at(mm
, addr
, pmd
, entry
);
879 update_mmu_cache_pmd(vma
, addr
, pmd
);
884 pte_free(mm
, pgtable
);
888 * vmf_insert_pfn_pmd - insert a pmd size pfn
889 * @vmf: Structure describing the fault
890 * @pfn: pfn to insert
891 * @write: whether it's a write fault
893 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
895 * Return: vm_fault_t value.
897 vm_fault_t
vmf_insert_pfn_pmd(struct vm_fault
*vmf
, pfn_t pfn
, bool write
)
899 unsigned long addr
= vmf
->address
& PMD_MASK
;
900 struct vm_area_struct
*vma
= vmf
->vma
;
901 pgprot_t pgprot
= vma
->vm_page_prot
;
902 pgtable_t pgtable
= NULL
;
905 * If we had pmd_special, we could avoid all these restrictions,
906 * but we need to be consistent with PTEs and architectures that
907 * can't support a 'special' bit.
909 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) &&
911 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
912 (VM_PFNMAP
|VM_MIXEDMAP
));
913 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
915 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
916 return VM_FAULT_SIGBUS
;
918 if (arch_needs_pgtable_deposit()) {
919 pgtable
= pte_alloc_one(vma
->vm_mm
);
924 track_pfn_insert(vma
, &pgprot
, pfn
);
926 insert_pfn_pmd(vma
, addr
, vmf
->pmd
, pfn
, pgprot
, write
, pgtable
);
927 return VM_FAULT_NOPAGE
;
929 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd
);
931 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
932 static pud_t
maybe_pud_mkwrite(pud_t pud
, struct vm_area_struct
*vma
)
934 if (likely(vma
->vm_flags
& VM_WRITE
))
935 pud
= pud_mkwrite(pud
);
939 static void insert_pfn_pud(struct vm_area_struct
*vma
, unsigned long addr
,
940 pud_t
*pud
, pfn_t pfn
, bool write
)
942 struct mm_struct
*mm
= vma
->vm_mm
;
943 pgprot_t prot
= vma
->vm_page_prot
;
947 ptl
= pud_lock(mm
, pud
);
948 if (!pud_none(*pud
)) {
950 if (pud_pfn(*pud
) != pfn_t_to_pfn(pfn
)) {
951 WARN_ON_ONCE(!is_huge_zero_pud(*pud
));
954 entry
= pud_mkyoung(*pud
);
955 entry
= maybe_pud_mkwrite(pud_mkdirty(entry
), vma
);
956 if (pudp_set_access_flags(vma
, addr
, pud
, entry
, 1))
957 update_mmu_cache_pud(vma
, addr
, pud
);
962 entry
= pud_mkhuge(pfn_t_pud(pfn
, prot
));
963 if (pfn_t_devmap(pfn
))
964 entry
= pud_mkdevmap(entry
);
966 entry
= pud_mkyoung(pud_mkdirty(entry
));
967 entry
= maybe_pud_mkwrite(entry
, vma
);
969 set_pud_at(mm
, addr
, pud
, entry
);
970 update_mmu_cache_pud(vma
, addr
, pud
);
977 * vmf_insert_pfn_pud - insert a pud size pfn
978 * @vmf: Structure describing the fault
979 * @pfn: pfn to insert
980 * @write: whether it's a write fault
982 * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
984 * Return: vm_fault_t value.
986 vm_fault_t
vmf_insert_pfn_pud(struct vm_fault
*vmf
, pfn_t pfn
, bool write
)
988 unsigned long addr
= vmf
->address
& PUD_MASK
;
989 struct vm_area_struct
*vma
= vmf
->vma
;
990 pgprot_t pgprot
= vma
->vm_page_prot
;
993 * If we had pud_special, we could avoid all these restrictions,
994 * but we need to be consistent with PTEs and architectures that
995 * can't support a 'special' bit.
997 BUG_ON(!(vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) &&
999 BUG_ON((vma
->vm_flags
& (VM_PFNMAP
|VM_MIXEDMAP
)) ==
1000 (VM_PFNMAP
|VM_MIXEDMAP
));
1001 BUG_ON((vma
->vm_flags
& VM_PFNMAP
) && is_cow_mapping(vma
->vm_flags
));
1003 if (addr
< vma
->vm_start
|| addr
>= vma
->vm_end
)
1004 return VM_FAULT_SIGBUS
;
1006 track_pfn_insert(vma
, &pgprot
, pfn
);
1008 insert_pfn_pud(vma
, addr
, vmf
->pud
, pfn
, write
);
1009 return VM_FAULT_NOPAGE
;
1011 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud
);
1012 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1014 static void touch_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
1015 pmd_t
*pmd
, bool write
)
1019 _pmd
= pmd_mkyoung(*pmd
);
1021 _pmd
= pmd_mkdirty(_pmd
);
1022 if (pmdp_set_access_flags(vma
, addr
& HPAGE_PMD_MASK
,
1024 update_mmu_cache_pmd(vma
, addr
, pmd
);
1027 struct page
*follow_devmap_pmd(struct vm_area_struct
*vma
, unsigned long addr
,
1028 pmd_t
*pmd
, int flags
, struct dev_pagemap
**pgmap
)
1030 unsigned long pfn
= pmd_pfn(*pmd
);
1031 struct mm_struct
*mm
= vma
->vm_mm
;
1035 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1037 if (flags
& FOLL_WRITE
&& !pmd_write(*pmd
))
1040 if (pmd_present(*pmd
) && pmd_devmap(*pmd
))
1045 if (flags
& FOLL_TOUCH
)
1046 touch_pmd(vma
, addr
, pmd
, flags
& FOLL_WRITE
);
1049 * device mapped pages can only be returned if the
1050 * caller will manage the page reference count.
1052 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
1053 return ERR_PTR(-EEXIST
);
1055 pfn
+= (addr
& ~PMD_MASK
) >> PAGE_SHIFT
;
1056 *pgmap
= get_dev_pagemap(pfn
, *pgmap
);
1058 return ERR_PTR(-EFAULT
);
1059 page
= pfn_to_page(pfn
);
1060 ret
= try_grab_page(page
, flags
);
1062 page
= ERR_PTR(ret
);
1067 int copy_huge_pmd(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1068 pmd_t
*dst_pmd
, pmd_t
*src_pmd
, unsigned long addr
,
1069 struct vm_area_struct
*dst_vma
, struct vm_area_struct
*src_vma
)
1071 spinlock_t
*dst_ptl
, *src_ptl
;
1072 struct page
*src_page
;
1074 pgtable_t pgtable
= NULL
;
1077 /* Skip if can be re-fill on fault */
1078 if (!vma_is_anonymous(dst_vma
))
1081 pgtable
= pte_alloc_one(dst_mm
);
1082 if (unlikely(!pgtable
))
1085 dst_ptl
= pmd_lock(dst_mm
, dst_pmd
);
1086 src_ptl
= pmd_lockptr(src_mm
, src_pmd
);
1087 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1092 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1093 if (unlikely(is_swap_pmd(pmd
))) {
1094 swp_entry_t entry
= pmd_to_swp_entry(pmd
);
1096 VM_BUG_ON(!is_pmd_migration_entry(pmd
));
1097 if (!is_readable_migration_entry(entry
)) {
1098 entry
= make_readable_migration_entry(
1100 pmd
= swp_entry_to_pmd(entry
);
1101 if (pmd_swp_soft_dirty(*src_pmd
))
1102 pmd
= pmd_swp_mksoft_dirty(pmd
);
1103 if (pmd_swp_uffd_wp(*src_pmd
))
1104 pmd
= pmd_swp_mkuffd_wp(pmd
);
1105 set_pmd_at(src_mm
, addr
, src_pmd
, pmd
);
1107 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1108 mm_inc_nr_ptes(dst_mm
);
1109 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1110 if (!userfaultfd_wp(dst_vma
))
1111 pmd
= pmd_swp_clear_uffd_wp(pmd
);
1112 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1118 if (unlikely(!pmd_trans_huge(pmd
))) {
1119 pte_free(dst_mm
, pgtable
);
1123 * When page table lock is held, the huge zero pmd should not be
1124 * under splitting since we don't split the page itself, only pmd to
1127 if (is_huge_zero_pmd(pmd
)) {
1129 * get_huge_zero_page() will never allocate a new page here,
1130 * since we already have a zero page to copy. It just takes a
1133 mm_get_huge_zero_page(dst_mm
);
1137 src_page
= pmd_page(pmd
);
1138 VM_BUG_ON_PAGE(!PageHead(src_page
), src_page
);
1141 if (unlikely(page_try_dup_anon_rmap(src_page
, true, src_vma
))) {
1142 /* Page maybe pinned: split and retry the fault on PTEs. */
1144 pte_free(dst_mm
, pgtable
);
1145 spin_unlock(src_ptl
);
1146 spin_unlock(dst_ptl
);
1147 __split_huge_pmd(src_vma
, src_pmd
, addr
, false, NULL
);
1150 add_mm_counter(dst_mm
, MM_ANONPAGES
, HPAGE_PMD_NR
);
1152 mm_inc_nr_ptes(dst_mm
);
1153 pgtable_trans_huge_deposit(dst_mm
, dst_pmd
, pgtable
);
1154 pmdp_set_wrprotect(src_mm
, addr
, src_pmd
);
1155 if (!userfaultfd_wp(dst_vma
))
1156 pmd
= pmd_clear_uffd_wp(pmd
);
1157 pmd
= pmd_mkold(pmd_wrprotect(pmd
));
1158 set_pmd_at(dst_mm
, addr
, dst_pmd
, pmd
);
1162 spin_unlock(src_ptl
);
1163 spin_unlock(dst_ptl
);
1168 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1169 static void touch_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1170 pud_t
*pud
, bool write
)
1174 _pud
= pud_mkyoung(*pud
);
1176 _pud
= pud_mkdirty(_pud
);
1177 if (pudp_set_access_flags(vma
, addr
& HPAGE_PUD_MASK
,
1179 update_mmu_cache_pud(vma
, addr
, pud
);
1182 struct page
*follow_devmap_pud(struct vm_area_struct
*vma
, unsigned long addr
,
1183 pud_t
*pud
, int flags
, struct dev_pagemap
**pgmap
)
1185 unsigned long pfn
= pud_pfn(*pud
);
1186 struct mm_struct
*mm
= vma
->vm_mm
;
1190 assert_spin_locked(pud_lockptr(mm
, pud
));
1192 if (flags
& FOLL_WRITE
&& !pud_write(*pud
))
1195 if (pud_present(*pud
) && pud_devmap(*pud
))
1200 if (flags
& FOLL_TOUCH
)
1201 touch_pud(vma
, addr
, pud
, flags
& FOLL_WRITE
);
1204 * device mapped pages can only be returned if the
1205 * caller will manage the page reference count.
1207 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1209 if (!(flags
& (FOLL_GET
| FOLL_PIN
)))
1210 return ERR_PTR(-EEXIST
);
1212 pfn
+= (addr
& ~PUD_MASK
) >> PAGE_SHIFT
;
1213 *pgmap
= get_dev_pagemap(pfn
, *pgmap
);
1215 return ERR_PTR(-EFAULT
);
1216 page
= pfn_to_page(pfn
);
1218 ret
= try_grab_page(page
, flags
);
1220 page
= ERR_PTR(ret
);
1225 int copy_huge_pud(struct mm_struct
*dst_mm
, struct mm_struct
*src_mm
,
1226 pud_t
*dst_pud
, pud_t
*src_pud
, unsigned long addr
,
1227 struct vm_area_struct
*vma
)
1229 spinlock_t
*dst_ptl
, *src_ptl
;
1233 dst_ptl
= pud_lock(dst_mm
, dst_pud
);
1234 src_ptl
= pud_lockptr(src_mm
, src_pud
);
1235 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
1239 if (unlikely(!pud_trans_huge(pud
) && !pud_devmap(pud
)))
1243 * When page table lock is held, the huge zero pud should not be
1244 * under splitting since we don't split the page itself, only pud to
1247 if (is_huge_zero_pud(pud
)) {
1248 /* No huge zero pud yet */
1252 * TODO: once we support anonymous pages, use page_try_dup_anon_rmap()
1253 * and split if duplicating fails.
1255 pudp_set_wrprotect(src_mm
, addr
, src_pud
);
1256 pud
= pud_mkold(pud_wrprotect(pud
));
1257 set_pud_at(dst_mm
, addr
, dst_pud
, pud
);
1261 spin_unlock(src_ptl
);
1262 spin_unlock(dst_ptl
);
1266 void huge_pud_set_accessed(struct vm_fault
*vmf
, pud_t orig_pud
)
1268 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1270 vmf
->ptl
= pud_lock(vmf
->vma
->vm_mm
, vmf
->pud
);
1271 if (unlikely(!pud_same(*vmf
->pud
, orig_pud
)))
1274 touch_pud(vmf
->vma
, vmf
->address
, vmf
->pud
, write
);
1276 spin_unlock(vmf
->ptl
);
1278 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1280 void huge_pmd_set_accessed(struct vm_fault
*vmf
)
1282 bool write
= vmf
->flags
& FAULT_FLAG_WRITE
;
1284 vmf
->ptl
= pmd_lock(vmf
->vma
->vm_mm
, vmf
->pmd
);
1285 if (unlikely(!pmd_same(*vmf
->pmd
, vmf
->orig_pmd
)))
1288 touch_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
, write
);
1291 spin_unlock(vmf
->ptl
);
1294 vm_fault_t
do_huge_pmd_wp_page(struct vm_fault
*vmf
)
1296 const bool unshare
= vmf
->flags
& FAULT_FLAG_UNSHARE
;
1297 struct vm_area_struct
*vma
= vmf
->vma
;
1298 struct folio
*folio
;
1300 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1301 pmd_t orig_pmd
= vmf
->orig_pmd
;
1303 vmf
->ptl
= pmd_lockptr(vma
->vm_mm
, vmf
->pmd
);
1304 VM_BUG_ON_VMA(!vma
->anon_vma
, vma
);
1306 if (is_huge_zero_pmd(orig_pmd
))
1309 spin_lock(vmf
->ptl
);
1311 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1312 spin_unlock(vmf
->ptl
);
1316 page
= pmd_page(orig_pmd
);
1317 folio
= page_folio(page
);
1318 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1320 /* Early check when only holding the PT lock. */
1321 if (PageAnonExclusive(page
))
1324 if (!folio_trylock(folio
)) {
1326 spin_unlock(vmf
->ptl
);
1328 spin_lock(vmf
->ptl
);
1329 if (unlikely(!pmd_same(*vmf
->pmd
, orig_pmd
))) {
1330 spin_unlock(vmf
->ptl
);
1331 folio_unlock(folio
);
1338 /* Recheck after temporarily dropping the PT lock. */
1339 if (PageAnonExclusive(page
)) {
1340 folio_unlock(folio
);
1345 * See do_wp_page(): we can only reuse the folio exclusively if
1346 * there are no additional references. Note that we always drain
1347 * the LRU pagevecs immediately after adding a THP.
1349 if (folio_ref_count(folio
) >
1350 1 + folio_test_swapcache(folio
) * folio_nr_pages(folio
))
1351 goto unlock_fallback
;
1352 if (folio_test_swapcache(folio
))
1353 folio_free_swap(folio
);
1354 if (folio_ref_count(folio
) == 1) {
1357 page_move_anon_rmap(page
, vma
);
1358 folio_unlock(folio
);
1360 if (unlikely(unshare
)) {
1361 spin_unlock(vmf
->ptl
);
1364 entry
= pmd_mkyoung(orig_pmd
);
1365 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1366 if (pmdp_set_access_flags(vma
, haddr
, vmf
->pmd
, entry
, 1))
1367 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1368 spin_unlock(vmf
->ptl
);
1373 folio_unlock(folio
);
1374 spin_unlock(vmf
->ptl
);
1376 __split_huge_pmd(vma
, vmf
->pmd
, vmf
->address
, false, NULL
);
1377 return VM_FAULT_FALLBACK
;
1380 static inline bool can_change_pmd_writable(struct vm_area_struct
*vma
,
1381 unsigned long addr
, pmd_t pmd
)
1385 if (WARN_ON_ONCE(!(vma
->vm_flags
& VM_WRITE
)))
1388 /* Don't touch entries that are not even readable (NUMA hinting). */
1389 if (pmd_protnone(pmd
))
1392 /* Do we need write faults for softdirty tracking? */
1393 if (vma_soft_dirty_enabled(vma
) && !pmd_soft_dirty(pmd
))
1396 /* Do we need write faults for uffd-wp tracking? */
1397 if (userfaultfd_huge_pmd_wp(vma
, pmd
))
1400 if (!(vma
->vm_flags
& VM_SHARED
)) {
1401 /* See can_change_pte_writable(). */
1402 page
= vm_normal_page_pmd(vma
, addr
, pmd
);
1403 return page
&& PageAnon(page
) && PageAnonExclusive(page
);
1406 /* See can_change_pte_writable(). */
1407 return pmd_dirty(pmd
);
1410 /* FOLL_FORCE can write to even unwritable PMDs in COW mappings. */
1411 static inline bool can_follow_write_pmd(pmd_t pmd
, struct page
*page
,
1412 struct vm_area_struct
*vma
,
1415 /* If the pmd is writable, we can write to the page. */
1419 /* Maybe FOLL_FORCE is set to override it? */
1420 if (!(flags
& FOLL_FORCE
))
1423 /* But FOLL_FORCE has no effect on shared mappings */
1424 if (vma
->vm_flags
& (VM_MAYSHARE
| VM_SHARED
))
1427 /* ... or read-only private ones */
1428 if (!(vma
->vm_flags
& VM_MAYWRITE
))
1431 /* ... or already writable ones that just need to take a write fault */
1432 if (vma
->vm_flags
& VM_WRITE
)
1436 * See can_change_pte_writable(): we broke COW and could map the page
1437 * writable if we have an exclusive anonymous page ...
1439 if (!page
|| !PageAnon(page
) || !PageAnonExclusive(page
))
1442 /* ... and a write-fault isn't required for other reasons. */
1443 if (vma_soft_dirty_enabled(vma
) && !pmd_soft_dirty(pmd
))
1445 return !userfaultfd_huge_pmd_wp(vma
, pmd
);
1448 struct page
*follow_trans_huge_pmd(struct vm_area_struct
*vma
,
1453 struct mm_struct
*mm
= vma
->vm_mm
;
1457 assert_spin_locked(pmd_lockptr(mm
, pmd
));
1459 page
= pmd_page(*pmd
);
1460 VM_BUG_ON_PAGE(!PageHead(page
) && !is_zone_device_page(page
), page
);
1462 if ((flags
& FOLL_WRITE
) &&
1463 !can_follow_write_pmd(*pmd
, page
, vma
, flags
))
1466 /* Avoid dumping huge zero page */
1467 if ((flags
& FOLL_DUMP
) && is_huge_zero_pmd(*pmd
))
1468 return ERR_PTR(-EFAULT
);
1470 /* Full NUMA hinting faults to serialise migration in fault paths */
1471 if (pmd_protnone(*pmd
) && !gup_can_follow_protnone(flags
))
1474 if (!pmd_write(*pmd
) && gup_must_unshare(vma
, flags
, page
))
1475 return ERR_PTR(-EMLINK
);
1477 VM_BUG_ON_PAGE((flags
& FOLL_PIN
) && PageAnon(page
) &&
1478 !PageAnonExclusive(page
), page
);
1480 ret
= try_grab_page(page
, flags
);
1482 return ERR_PTR(ret
);
1484 if (flags
& FOLL_TOUCH
)
1485 touch_pmd(vma
, addr
, pmd
, flags
& FOLL_WRITE
);
1487 page
+= (addr
& ~HPAGE_PMD_MASK
) >> PAGE_SHIFT
;
1488 VM_BUG_ON_PAGE(!PageCompound(page
) && !is_zone_device_page(page
), page
);
1493 /* NUMA hinting page fault entry point for trans huge pmds */
1494 vm_fault_t
do_huge_pmd_numa_page(struct vm_fault
*vmf
)
1496 struct vm_area_struct
*vma
= vmf
->vma
;
1497 pmd_t oldpmd
= vmf
->orig_pmd
;
1500 unsigned long haddr
= vmf
->address
& HPAGE_PMD_MASK
;
1501 int page_nid
= NUMA_NO_NODE
;
1502 int target_nid
, last_cpupid
= (-1 & LAST_CPUPID_MASK
);
1503 bool migrated
= false, writable
= false;
1506 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1507 if (unlikely(!pmd_same(oldpmd
, *vmf
->pmd
))) {
1508 spin_unlock(vmf
->ptl
);
1512 pmd
= pmd_modify(oldpmd
, vma
->vm_page_prot
);
1515 * Detect now whether the PMD could be writable; this information
1516 * is only valid while holding the PT lock.
1518 writable
= pmd_write(pmd
);
1519 if (!writable
&& vma_wants_manual_pte_write_upgrade(vma
) &&
1520 can_change_pmd_writable(vma
, vmf
->address
, pmd
))
1523 page
= vm_normal_page_pmd(vma
, haddr
, pmd
);
1527 /* See similar comment in do_numa_page for explanation */
1529 flags
|= TNF_NO_GROUP
;
1531 page_nid
= page_to_nid(page
);
1533 * For memory tiering mode, cpupid of slow memory page is used
1534 * to record page access time. So use default value.
1536 if (node_is_toptier(page_nid
))
1537 last_cpupid
= page_cpupid_last(page
);
1538 target_nid
= numa_migrate_prep(page
, vma
, haddr
, page_nid
,
1541 if (target_nid
== NUMA_NO_NODE
) {
1546 spin_unlock(vmf
->ptl
);
1549 migrated
= migrate_misplaced_page(page
, vma
, target_nid
);
1551 flags
|= TNF_MIGRATED
;
1552 page_nid
= target_nid
;
1554 flags
|= TNF_MIGRATE_FAIL
;
1555 vmf
->ptl
= pmd_lock(vma
->vm_mm
, vmf
->pmd
);
1556 if (unlikely(!pmd_same(oldpmd
, *vmf
->pmd
))) {
1557 spin_unlock(vmf
->ptl
);
1564 if (page_nid
!= NUMA_NO_NODE
)
1565 task_numa_fault(last_cpupid
, page_nid
, HPAGE_PMD_NR
,
1571 /* Restore the PMD */
1572 pmd
= pmd_modify(oldpmd
, vma
->vm_page_prot
);
1573 pmd
= pmd_mkyoung(pmd
);
1575 pmd
= pmd_mkwrite(pmd
);
1576 set_pmd_at(vma
->vm_mm
, haddr
, vmf
->pmd
, pmd
);
1577 update_mmu_cache_pmd(vma
, vmf
->address
, vmf
->pmd
);
1578 spin_unlock(vmf
->ptl
);
1583 * Return true if we do MADV_FREE successfully on entire pmd page.
1584 * Otherwise, return false.
1586 bool madvise_free_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1587 pmd_t
*pmd
, unsigned long addr
, unsigned long next
)
1591 struct folio
*folio
;
1592 struct mm_struct
*mm
= tlb
->mm
;
1595 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
1597 ptl
= pmd_trans_huge_lock(pmd
, vma
);
1602 if (is_huge_zero_pmd(orig_pmd
))
1605 if (unlikely(!pmd_present(orig_pmd
))) {
1606 VM_BUG_ON(thp_migration_supported() &&
1607 !is_pmd_migration_entry(orig_pmd
));
1611 folio
= pfn_folio(pmd_pfn(orig_pmd
));
1613 * If other processes are mapping this folio, we couldn't discard
1614 * the folio unless they all do MADV_FREE so let's skip the folio.
1616 if (folio_mapcount(folio
) != 1)
1619 if (!folio_trylock(folio
))
1623 * If user want to discard part-pages of THP, split it so MADV_FREE
1624 * will deactivate only them.
1626 if (next
- addr
!= HPAGE_PMD_SIZE
) {
1630 folio_unlock(folio
);
1635 if (folio_test_dirty(folio
))
1636 folio_clear_dirty(folio
);
1637 folio_unlock(folio
);
1639 if (pmd_young(orig_pmd
) || pmd_dirty(orig_pmd
)) {
1640 pmdp_invalidate(vma
, addr
, pmd
);
1641 orig_pmd
= pmd_mkold(orig_pmd
);
1642 orig_pmd
= pmd_mkclean(orig_pmd
);
1644 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
1645 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1648 folio_mark_lazyfree(folio
);
1656 static inline void zap_deposited_table(struct mm_struct
*mm
, pmd_t
*pmd
)
1660 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
1661 pte_free(mm
, pgtable
);
1665 int zap_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1666 pmd_t
*pmd
, unsigned long addr
)
1671 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
1673 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1677 * For architectures like ppc64 we look at deposited pgtable
1678 * when calling pmdp_huge_get_and_clear. So do the
1679 * pgtable_trans_huge_withdraw after finishing pmdp related
1682 orig_pmd
= pmdp_huge_get_and_clear_full(vma
, addr
, pmd
,
1684 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
1685 if (vma_is_special_huge(vma
)) {
1686 if (arch_needs_pgtable_deposit())
1687 zap_deposited_table(tlb
->mm
, pmd
);
1689 } else if (is_huge_zero_pmd(orig_pmd
)) {
1690 zap_deposited_table(tlb
->mm
, pmd
);
1693 struct page
*page
= NULL
;
1694 int flush_needed
= 1;
1696 if (pmd_present(orig_pmd
)) {
1697 page
= pmd_page(orig_pmd
);
1698 page_remove_rmap(page
, vma
, true);
1699 VM_BUG_ON_PAGE(page_mapcount(page
) < 0, page
);
1700 VM_BUG_ON_PAGE(!PageHead(page
), page
);
1701 } else if (thp_migration_supported()) {
1704 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd
));
1705 entry
= pmd_to_swp_entry(orig_pmd
);
1706 page
= pfn_swap_entry_to_page(entry
);
1709 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1711 if (PageAnon(page
)) {
1712 zap_deposited_table(tlb
->mm
, pmd
);
1713 add_mm_counter(tlb
->mm
, MM_ANONPAGES
, -HPAGE_PMD_NR
);
1715 if (arch_needs_pgtable_deposit())
1716 zap_deposited_table(tlb
->mm
, pmd
);
1717 add_mm_counter(tlb
->mm
, mm_counter_file(page
), -HPAGE_PMD_NR
);
1722 tlb_remove_page_size(tlb
, page
, HPAGE_PMD_SIZE
);
1727 #ifndef pmd_move_must_withdraw
1728 static inline int pmd_move_must_withdraw(spinlock_t
*new_pmd_ptl
,
1729 spinlock_t
*old_pmd_ptl
,
1730 struct vm_area_struct
*vma
)
1733 * With split pmd lock we also need to move preallocated
1734 * PTE page table if new_pmd is on different PMD page table.
1736 * We also don't deposit and withdraw tables for file pages.
1738 return (new_pmd_ptl
!= old_pmd_ptl
) && vma_is_anonymous(vma
);
1742 static pmd_t
move_soft_dirty_pmd(pmd_t pmd
)
1744 #ifdef CONFIG_MEM_SOFT_DIRTY
1745 if (unlikely(is_pmd_migration_entry(pmd
)))
1746 pmd
= pmd_swp_mksoft_dirty(pmd
);
1747 else if (pmd_present(pmd
))
1748 pmd
= pmd_mksoft_dirty(pmd
);
1753 bool move_huge_pmd(struct vm_area_struct
*vma
, unsigned long old_addr
,
1754 unsigned long new_addr
, pmd_t
*old_pmd
, pmd_t
*new_pmd
)
1756 spinlock_t
*old_ptl
, *new_ptl
;
1758 struct mm_struct
*mm
= vma
->vm_mm
;
1759 bool force_flush
= false;
1762 * The destination pmd shouldn't be established, free_pgtables()
1763 * should have release it.
1765 if (WARN_ON(!pmd_none(*new_pmd
))) {
1766 VM_BUG_ON(pmd_trans_huge(*new_pmd
));
1771 * We don't have to worry about the ordering of src and dst
1772 * ptlocks because exclusive mmap_lock prevents deadlock.
1774 old_ptl
= __pmd_trans_huge_lock(old_pmd
, vma
);
1776 new_ptl
= pmd_lockptr(mm
, new_pmd
);
1777 if (new_ptl
!= old_ptl
)
1778 spin_lock_nested(new_ptl
, SINGLE_DEPTH_NESTING
);
1779 pmd
= pmdp_huge_get_and_clear(mm
, old_addr
, old_pmd
);
1780 if (pmd_present(pmd
))
1782 VM_BUG_ON(!pmd_none(*new_pmd
));
1784 if (pmd_move_must_withdraw(new_ptl
, old_ptl
, vma
)) {
1786 pgtable
= pgtable_trans_huge_withdraw(mm
, old_pmd
);
1787 pgtable_trans_huge_deposit(mm
, new_pmd
, pgtable
);
1789 pmd
= move_soft_dirty_pmd(pmd
);
1790 set_pmd_at(mm
, new_addr
, new_pmd
, pmd
);
1792 flush_pmd_tlb_range(vma
, old_addr
, old_addr
+ PMD_SIZE
);
1793 if (new_ptl
!= old_ptl
)
1794 spin_unlock(new_ptl
);
1795 spin_unlock(old_ptl
);
1803 * - 0 if PMD could not be locked
1804 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1805 * or if prot_numa but THP migration is not supported
1806 * - HPAGE_PMD_NR if protections changed and TLB flush necessary
1808 int change_huge_pmd(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1809 pmd_t
*pmd
, unsigned long addr
, pgprot_t newprot
,
1810 unsigned long cp_flags
)
1812 struct mm_struct
*mm
= vma
->vm_mm
;
1814 pmd_t oldpmd
, entry
;
1815 bool prot_numa
= cp_flags
& MM_CP_PROT_NUMA
;
1816 bool uffd_wp
= cp_flags
& MM_CP_UFFD_WP
;
1817 bool uffd_wp_resolve
= cp_flags
& MM_CP_UFFD_WP_RESOLVE
;
1820 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
1822 if (prot_numa
&& !thp_migration_supported())
1825 ptl
= __pmd_trans_huge_lock(pmd
, vma
);
1829 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1830 if (is_swap_pmd(*pmd
)) {
1831 swp_entry_t entry
= pmd_to_swp_entry(*pmd
);
1832 struct page
*page
= pfn_swap_entry_to_page(entry
);
1835 VM_BUG_ON(!is_pmd_migration_entry(*pmd
));
1836 if (is_writable_migration_entry(entry
)) {
1838 * A protection check is difficult so
1839 * just be safe and disable write
1842 entry
= make_readable_exclusive_migration_entry(swp_offset(entry
));
1844 entry
= make_readable_migration_entry(swp_offset(entry
));
1845 newpmd
= swp_entry_to_pmd(entry
);
1846 if (pmd_swp_soft_dirty(*pmd
))
1847 newpmd
= pmd_swp_mksoft_dirty(newpmd
);
1853 newpmd
= pmd_swp_mkuffd_wp(newpmd
);
1854 else if (uffd_wp_resolve
)
1855 newpmd
= pmd_swp_clear_uffd_wp(newpmd
);
1856 if (!pmd_same(*pmd
, newpmd
))
1857 set_pmd_at(mm
, addr
, pmd
, newpmd
);
1866 * Avoid trapping faults against the zero page. The read-only
1867 * data is likely to be read-cached on the local CPU and
1868 * local/remote hits to the zero page are not interesting.
1870 if (is_huge_zero_pmd(*pmd
))
1873 if (pmd_protnone(*pmd
))
1876 page
= pmd_page(*pmd
);
1877 toptier
= node_is_toptier(page_to_nid(page
));
1879 * Skip scanning top tier node if normal numa
1880 * balancing is disabled
1882 if (!(sysctl_numa_balancing_mode
& NUMA_BALANCING_NORMAL
) &&
1886 if (sysctl_numa_balancing_mode
& NUMA_BALANCING_MEMORY_TIERING
&&
1888 xchg_page_access_time(page
, jiffies_to_msecs(jiffies
));
1891 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1892 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1893 * which is also under mmap_read_lock(mm):
1896 * change_huge_pmd(prot_numa=1)
1897 * pmdp_huge_get_and_clear_notify()
1898 * madvise_dontneed()
1900 * pmd_trans_huge(*pmd) == 0 (without ptl)
1903 * // pmd is re-established
1905 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1906 * which may break userspace.
1908 * pmdp_invalidate_ad() is required to make sure we don't miss
1909 * dirty/young flags set by hardware.
1911 oldpmd
= pmdp_invalidate_ad(vma
, addr
, pmd
);
1913 entry
= pmd_modify(oldpmd
, newprot
);
1915 entry
= pmd_mkuffd_wp(entry
);
1916 else if (uffd_wp_resolve
)
1918 * Leave the write bit to be handled by PF interrupt
1919 * handler, then things like COW could be properly
1922 entry
= pmd_clear_uffd_wp(entry
);
1924 /* See change_pte_range(). */
1925 if ((cp_flags
& MM_CP_TRY_CHANGE_WRITABLE
) && !pmd_write(entry
) &&
1926 can_change_pmd_writable(vma
, addr
, entry
))
1927 entry
= pmd_mkwrite(entry
);
1930 set_pmd_at(mm
, addr
, pmd
, entry
);
1932 if (huge_pmd_needs_flush(oldpmd
, entry
))
1933 tlb_flush_pmd_range(tlb
, addr
, HPAGE_PMD_SIZE
);
1940 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1942 * Note that if it returns page table lock pointer, this routine returns without
1943 * unlocking page table lock. So callers must unlock it.
1945 spinlock_t
*__pmd_trans_huge_lock(pmd_t
*pmd
, struct vm_area_struct
*vma
)
1948 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
1949 if (likely(is_swap_pmd(*pmd
) || pmd_trans_huge(*pmd
) ||
1957 * Returns page table lock pointer if a given pud maps a thp, NULL otherwise.
1959 * Note that if it returns page table lock pointer, this routine returns without
1960 * unlocking page table lock. So callers must unlock it.
1962 spinlock_t
*__pud_trans_huge_lock(pud_t
*pud
, struct vm_area_struct
*vma
)
1966 ptl
= pud_lock(vma
->vm_mm
, pud
);
1967 if (likely(pud_trans_huge(*pud
) || pud_devmap(*pud
)))
1973 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1974 int zap_huge_pud(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
1975 pud_t
*pud
, unsigned long addr
)
1979 ptl
= __pud_trans_huge_lock(pud
, vma
);
1983 pudp_huge_get_and_clear_full(tlb
->mm
, addr
, pud
, tlb
->fullmm
);
1984 tlb_remove_pud_tlb_entry(tlb
, pud
, addr
);
1985 if (vma_is_special_huge(vma
)) {
1987 /* No zero page support yet */
1989 /* No support for anonymous PUD pages yet */
1995 static void __split_huge_pud_locked(struct vm_area_struct
*vma
, pud_t
*pud
,
1996 unsigned long haddr
)
1998 VM_BUG_ON(haddr
& ~HPAGE_PUD_MASK
);
1999 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2000 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PUD_SIZE
, vma
);
2001 VM_BUG_ON(!pud_trans_huge(*pud
) && !pud_devmap(*pud
));
2003 count_vm_event(THP_SPLIT_PUD
);
2005 pudp_huge_clear_flush_notify(vma
, haddr
, pud
);
2008 void __split_huge_pud(struct vm_area_struct
*vma
, pud_t
*pud
,
2009 unsigned long address
)
2012 struct mmu_notifier_range range
;
2014 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
->vm_mm
,
2015 address
& HPAGE_PUD_MASK
,
2016 (address
& HPAGE_PUD_MASK
) + HPAGE_PUD_SIZE
);
2017 mmu_notifier_invalidate_range_start(&range
);
2018 ptl
= pud_lock(vma
->vm_mm
, pud
);
2019 if (unlikely(!pud_trans_huge(*pud
) && !pud_devmap(*pud
)))
2021 __split_huge_pud_locked(vma
, pud
, range
.start
);
2026 * No need to double call mmu_notifier->invalidate_range() callback as
2027 * the above pudp_huge_clear_flush_notify() did already call it.
2029 mmu_notifier_invalidate_range_only_end(&range
);
2031 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2033 static void __split_huge_zero_page_pmd(struct vm_area_struct
*vma
,
2034 unsigned long haddr
, pmd_t
*pmd
)
2036 struct mm_struct
*mm
= vma
->vm_mm
;
2038 pmd_t _pmd
, old_pmd
;
2042 * Leave pmd empty until pte is filled note that it is fine to delay
2043 * notification until mmu_notifier_invalidate_range_end() as we are
2044 * replacing a zero pmd write protected page with a zero pte write
2047 * See Documentation/mm/mmu_notifier.rst
2049 old_pmd
= pmdp_huge_clear_flush(vma
, haddr
, pmd
);
2051 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2052 pmd_populate(mm
, &_pmd
, pgtable
);
2054 for (i
= 0; i
< HPAGE_PMD_NR
; i
++, haddr
+= PAGE_SIZE
) {
2056 entry
= pfn_pte(my_zero_pfn(haddr
), vma
->vm_page_prot
);
2057 entry
= pte_mkspecial(entry
);
2058 if (pmd_uffd_wp(old_pmd
))
2059 entry
= pte_mkuffd_wp(entry
);
2060 pte
= pte_offset_map(&_pmd
, haddr
);
2061 VM_BUG_ON(!pte_none(*pte
));
2062 set_pte_at(mm
, haddr
, pte
, entry
);
2065 smp_wmb(); /* make pte visible before pmd */
2066 pmd_populate(mm
, pmd
, pgtable
);
2069 static void __split_huge_pmd_locked(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2070 unsigned long haddr
, bool freeze
)
2072 struct mm_struct
*mm
= vma
->vm_mm
;
2075 pmd_t old_pmd
, _pmd
;
2076 bool young
, write
, soft_dirty
, pmd_migration
= false, uffd_wp
= false;
2077 bool anon_exclusive
= false, dirty
= false;
2081 VM_BUG_ON(haddr
& ~HPAGE_PMD_MASK
);
2082 VM_BUG_ON_VMA(vma
->vm_start
> haddr
, vma
);
2083 VM_BUG_ON_VMA(vma
->vm_end
< haddr
+ HPAGE_PMD_SIZE
, vma
);
2084 VM_BUG_ON(!is_pmd_migration_entry(*pmd
) && !pmd_trans_huge(*pmd
)
2085 && !pmd_devmap(*pmd
));
2087 count_vm_event(THP_SPLIT_PMD
);
2089 if (!vma_is_anonymous(vma
)) {
2090 old_pmd
= pmdp_huge_clear_flush_notify(vma
, haddr
, pmd
);
2092 * We are going to unmap this huge page. So
2093 * just go ahead and zap it
2095 if (arch_needs_pgtable_deposit())
2096 zap_deposited_table(mm
, pmd
);
2097 if (vma_is_special_huge(vma
))
2099 if (unlikely(is_pmd_migration_entry(old_pmd
))) {
2102 entry
= pmd_to_swp_entry(old_pmd
);
2103 page
= pfn_swap_entry_to_page(entry
);
2105 page
= pmd_page(old_pmd
);
2106 if (!PageDirty(page
) && pmd_dirty(old_pmd
))
2107 set_page_dirty(page
);
2108 if (!PageReferenced(page
) && pmd_young(old_pmd
))
2109 SetPageReferenced(page
);
2110 page_remove_rmap(page
, vma
, true);
2113 add_mm_counter(mm
, mm_counter_file(page
), -HPAGE_PMD_NR
);
2117 if (is_huge_zero_pmd(*pmd
)) {
2119 * FIXME: Do we want to invalidate secondary mmu by calling
2120 * mmu_notifier_invalidate_range() see comments below inside
2121 * __split_huge_pmd() ?
2123 * We are going from a zero huge page write protected to zero
2124 * small page also write protected so it does not seems useful
2125 * to invalidate secondary mmu at this time.
2127 return __split_huge_zero_page_pmd(vma
, haddr
, pmd
);
2131 * Up to this point the pmd is present and huge and userland has the
2132 * whole access to the hugepage during the split (which happens in
2133 * place). If we overwrite the pmd with the not-huge version pointing
2134 * to the pte here (which of course we could if all CPUs were bug
2135 * free), userland could trigger a small page size TLB miss on the
2136 * small sized TLB while the hugepage TLB entry is still established in
2137 * the huge TLB. Some CPU doesn't like that.
2138 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2139 * 383 on page 105. Intel should be safe but is also warns that it's
2140 * only safe if the permission and cache attributes of the two entries
2141 * loaded in the two TLB is identical (which should be the case here).
2142 * But it is generally safer to never allow small and huge TLB entries
2143 * for the same virtual address to be loaded simultaneously. So instead
2144 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2145 * current pmd notpresent (atomically because here the pmd_trans_huge
2146 * must remain set at all times on the pmd until the split is complete
2147 * for this pmd), then we flush the SMP TLB and finally we write the
2148 * non-huge version of the pmd entry with pmd_populate.
2150 old_pmd
= pmdp_invalidate(vma
, haddr
, pmd
);
2152 pmd_migration
= is_pmd_migration_entry(old_pmd
);
2153 if (unlikely(pmd_migration
)) {
2156 entry
= pmd_to_swp_entry(old_pmd
);
2157 page
= pfn_swap_entry_to_page(entry
);
2158 write
= is_writable_migration_entry(entry
);
2160 anon_exclusive
= is_readable_exclusive_migration_entry(entry
);
2161 young
= is_migration_entry_young(entry
);
2162 dirty
= is_migration_entry_dirty(entry
);
2163 soft_dirty
= pmd_swp_soft_dirty(old_pmd
);
2164 uffd_wp
= pmd_swp_uffd_wp(old_pmd
);
2166 page
= pmd_page(old_pmd
);
2167 if (pmd_dirty(old_pmd
)) {
2171 write
= pmd_write(old_pmd
);
2172 young
= pmd_young(old_pmd
);
2173 soft_dirty
= pmd_soft_dirty(old_pmd
);
2174 uffd_wp
= pmd_uffd_wp(old_pmd
);
2176 VM_BUG_ON_PAGE(!page_count(page
), page
);
2179 * Without "freeze", we'll simply split the PMD, propagating the
2180 * PageAnonExclusive() flag for each PTE by setting it for
2181 * each subpage -- no need to (temporarily) clear.
2183 * With "freeze" we want to replace mapped pages by
2184 * migration entries right away. This is only possible if we
2185 * managed to clear PageAnonExclusive() -- see
2186 * set_pmd_migration_entry().
2188 * In case we cannot clear PageAnonExclusive(), split the PMD
2189 * only and let try_to_migrate_one() fail later.
2191 * See page_try_share_anon_rmap(): invalidate PMD first.
2193 anon_exclusive
= PageAnon(page
) && PageAnonExclusive(page
);
2194 if (freeze
&& anon_exclusive
&& page_try_share_anon_rmap(page
))
2197 page_ref_add(page
, HPAGE_PMD_NR
- 1);
2201 * Withdraw the table only after we mark the pmd entry invalid.
2202 * This's critical for some architectures (Power).
2204 pgtable
= pgtable_trans_huge_withdraw(mm
, pmd
);
2205 pmd_populate(mm
, &_pmd
, pgtable
);
2207 for (i
= 0, addr
= haddr
; i
< HPAGE_PMD_NR
; i
++, addr
+= PAGE_SIZE
) {
2210 * Note that NUMA hinting access restrictions are not
2211 * transferred to avoid any possibility of altering
2212 * permissions across VMAs.
2214 if (freeze
|| pmd_migration
) {
2215 swp_entry_t swp_entry
;
2217 swp_entry
= make_writable_migration_entry(
2218 page_to_pfn(page
+ i
));
2219 else if (anon_exclusive
)
2220 swp_entry
= make_readable_exclusive_migration_entry(
2221 page_to_pfn(page
+ i
));
2223 swp_entry
= make_readable_migration_entry(
2224 page_to_pfn(page
+ i
));
2226 swp_entry
= make_migration_entry_young(swp_entry
);
2228 swp_entry
= make_migration_entry_dirty(swp_entry
);
2229 entry
= swp_entry_to_pte(swp_entry
);
2231 entry
= pte_swp_mksoft_dirty(entry
);
2233 entry
= pte_swp_mkuffd_wp(entry
);
2235 entry
= mk_pte(page
+ i
, READ_ONCE(vma
->vm_page_prot
));
2237 entry
= pte_mkwrite(entry
);
2239 SetPageAnonExclusive(page
+ i
);
2241 entry
= pte_mkold(entry
);
2242 /* NOTE: this may set soft-dirty too on some archs */
2244 entry
= pte_mkdirty(entry
);
2246 entry
= pte_mksoft_dirty(entry
);
2248 entry
= pte_mkuffd_wp(entry
);
2249 page_add_anon_rmap(page
+ i
, vma
, addr
, false);
2251 pte
= pte_offset_map(&_pmd
, addr
);
2252 BUG_ON(!pte_none(*pte
));
2253 set_pte_at(mm
, addr
, pte
, entry
);
2258 page_remove_rmap(page
, vma
, true);
2262 smp_wmb(); /* make pte visible before pmd */
2263 pmd_populate(mm
, pmd
, pgtable
);
2266 void __split_huge_pmd(struct vm_area_struct
*vma
, pmd_t
*pmd
,
2267 unsigned long address
, bool freeze
, struct folio
*folio
)
2270 struct mmu_notifier_range range
;
2272 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
->vm_mm
,
2273 address
& HPAGE_PMD_MASK
,
2274 (address
& HPAGE_PMD_MASK
) + HPAGE_PMD_SIZE
);
2275 mmu_notifier_invalidate_range_start(&range
);
2276 ptl
= pmd_lock(vma
->vm_mm
, pmd
);
2279 * If caller asks to setup a migration entry, we need a folio to check
2280 * pmd against. Otherwise we can end up replacing wrong folio.
2282 VM_BUG_ON(freeze
&& !folio
);
2283 VM_WARN_ON_ONCE(folio
&& !folio_test_locked(folio
));
2285 if (pmd_trans_huge(*pmd
) || pmd_devmap(*pmd
) ||
2286 is_pmd_migration_entry(*pmd
)) {
2288 * It's safe to call pmd_page when folio is set because it's
2289 * guaranteed that pmd is present.
2291 if (folio
&& folio
!= page_folio(pmd_page(*pmd
)))
2293 __split_huge_pmd_locked(vma
, pmd
, range
.start
, freeze
);
2299 * No need to double call mmu_notifier->invalidate_range() callback.
2300 * They are 3 cases to consider inside __split_huge_pmd_locked():
2301 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2302 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2303 * fault will trigger a flush_notify before pointing to a new page
2304 * (it is fine if the secondary mmu keeps pointing to the old zero
2305 * page in the meantime)
2306 * 3) Split a huge pmd into pte pointing to the same page. No need
2307 * to invalidate secondary tlb entry they are all still valid.
2308 * any further changes to individual pte will notify. So no need
2309 * to call mmu_notifier->invalidate_range()
2311 mmu_notifier_invalidate_range_only_end(&range
);
2314 void split_huge_pmd_address(struct vm_area_struct
*vma
, unsigned long address
,
2315 bool freeze
, struct folio
*folio
)
2317 pmd_t
*pmd
= mm_find_pmd(vma
->vm_mm
, address
);
2322 __split_huge_pmd(vma
, pmd
, address
, freeze
, folio
);
2325 static inline void split_huge_pmd_if_needed(struct vm_area_struct
*vma
, unsigned long address
)
2328 * If the new address isn't hpage aligned and it could previously
2329 * contain an hugepage: check if we need to split an huge pmd.
2331 if (!IS_ALIGNED(address
, HPAGE_PMD_SIZE
) &&
2332 range_in_vma(vma
, ALIGN_DOWN(address
, HPAGE_PMD_SIZE
),
2333 ALIGN(address
, HPAGE_PMD_SIZE
)))
2334 split_huge_pmd_address(vma
, address
, false, NULL
);
2337 void vma_adjust_trans_huge(struct vm_area_struct
*vma
,
2338 unsigned long start
,
2342 /* Check if we need to split start first. */
2343 split_huge_pmd_if_needed(vma
, start
);
2345 /* Check if we need to split end next. */
2346 split_huge_pmd_if_needed(vma
, end
);
2349 * If we're also updating the next vma vm_start,
2350 * check if we need to split it.
2352 if (adjust_next
> 0) {
2353 struct vm_area_struct
*next
= find_vma(vma
->vm_mm
, vma
->vm_end
);
2354 unsigned long nstart
= next
->vm_start
;
2355 nstart
+= adjust_next
;
2356 split_huge_pmd_if_needed(next
, nstart
);
2360 static void unmap_folio(struct folio
*folio
)
2362 enum ttu_flags ttu_flags
= TTU_RMAP_LOCKED
| TTU_SPLIT_HUGE_PMD
|
2365 VM_BUG_ON_FOLIO(!folio_test_large(folio
), folio
);
2368 * Anon pages need migration entries to preserve them, but file
2369 * pages can simply be left unmapped, then faulted back on demand.
2370 * If that is ever changed (perhaps for mlock), update remap_page().
2372 if (folio_test_anon(folio
))
2373 try_to_migrate(folio
, ttu_flags
);
2375 try_to_unmap(folio
, ttu_flags
| TTU_IGNORE_MLOCK
);
2378 static void remap_page(struct folio
*folio
, unsigned long nr
)
2382 /* If unmap_folio() uses try_to_migrate() on file, remove this check */
2383 if (!folio_test_anon(folio
))
2386 remove_migration_ptes(folio
, folio
, true);
2387 i
+= folio_nr_pages(folio
);
2390 folio
= folio_next(folio
);
2394 static void lru_add_page_tail(struct page
*head
, struct page
*tail
,
2395 struct lruvec
*lruvec
, struct list_head
*list
)
2397 VM_BUG_ON_PAGE(!PageHead(head
), head
);
2398 VM_BUG_ON_PAGE(PageCompound(tail
), head
);
2399 VM_BUG_ON_PAGE(PageLRU(tail
), head
);
2400 lockdep_assert_held(&lruvec
->lru_lock
);
2403 /* page reclaim is reclaiming a huge page */
2404 VM_WARN_ON(PageLRU(head
));
2406 list_add_tail(&tail
->lru
, list
);
2408 /* head is still on lru (and we have it frozen) */
2409 VM_WARN_ON(!PageLRU(head
));
2410 if (PageUnevictable(tail
))
2411 tail
->mlock_count
= 0;
2413 list_add_tail(&tail
->lru
, &head
->lru
);
2418 static void __split_huge_page_tail(struct page
*head
, int tail
,
2419 struct lruvec
*lruvec
, struct list_head
*list
)
2421 struct page
*page_tail
= head
+ tail
;
2423 VM_BUG_ON_PAGE(atomic_read(&page_tail
->_mapcount
) != -1, page_tail
);
2426 * Clone page flags before unfreezing refcount.
2428 * After successful get_page_unless_zero() might follow flags change,
2429 * for example lock_page() which set PG_waiters.
2431 * Note that for mapped sub-pages of an anonymous THP,
2432 * PG_anon_exclusive has been cleared in unmap_folio() and is stored in
2433 * the migration entry instead from where remap_page() will restore it.
2434 * We can still have PG_anon_exclusive set on effectively unmapped and
2435 * unreferenced sub-pages of an anonymous THP: we can simply drop
2436 * PG_anon_exclusive (-> PG_mappedtodisk) for these here.
2438 page_tail
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
2439 page_tail
->flags
|= (head
->flags
&
2440 ((1L << PG_referenced
) |
2441 (1L << PG_swapbacked
) |
2442 (1L << PG_swapcache
) |
2443 (1L << PG_mlocked
) |
2444 (1L << PG_uptodate
) |
2446 (1L << PG_workingset
) |
2448 (1L << PG_unevictable
) |
2449 #ifdef CONFIG_ARCH_USES_PG_ARCH_X
2454 LRU_GEN_MASK
| LRU_REFS_MASK
));
2456 /* ->mapping in first and second tail page is replaced by other uses */
2457 VM_BUG_ON_PAGE(tail
> 2 && page_tail
->mapping
!= TAIL_MAPPING
,
2459 page_tail
->mapping
= head
->mapping
;
2460 page_tail
->index
= head
->index
+ tail
;
2463 * page->private should not be set in tail pages with the exception
2464 * of swap cache pages that store the swp_entry_t in tail pages.
2465 * Fix up and warn once if private is unexpectedly set.
2467 * What of 32-bit systems, on which folio->_pincount overlays
2468 * head[1].private? No problem: THP_SWAP is not enabled on 32-bit, and
2469 * pincount must be 0 for folio_ref_freeze() to have succeeded.
2471 if (!folio_test_swapcache(page_folio(head
))) {
2472 VM_WARN_ON_ONCE_PAGE(page_tail
->private != 0, page_tail
);
2473 page_tail
->private = 0;
2476 /* Page flags must be visible before we make the page non-compound. */
2480 * Clear PageTail before unfreezing page refcount.
2482 * After successful get_page_unless_zero() might follow put_page()
2483 * which needs correct compound_head().
2485 clear_compound_head(page_tail
);
2487 /* Finally unfreeze refcount. Additional reference from page cache. */
2488 page_ref_unfreeze(page_tail
, 1 + (!PageAnon(head
) ||
2489 PageSwapCache(head
)));
2491 if (page_is_young(head
))
2492 set_page_young(page_tail
);
2493 if (page_is_idle(head
))
2494 set_page_idle(page_tail
);
2496 page_cpupid_xchg_last(page_tail
, page_cpupid_last(head
));
2499 * always add to the tail because some iterators expect new
2500 * pages to show after the currently processed elements - e.g.
2503 lru_add_page_tail(head
, page_tail
, lruvec
, list
);
2506 static void __split_huge_page(struct page
*page
, struct list_head
*list
,
2509 struct folio
*folio
= page_folio(page
);
2510 struct page
*head
= &folio
->page
;
2511 struct lruvec
*lruvec
;
2512 struct address_space
*swap_cache
= NULL
;
2513 unsigned long offset
= 0;
2514 unsigned int nr
= thp_nr_pages(head
);
2517 /* complete memcg works before add pages to LRU */
2518 split_page_memcg(head
, nr
);
2520 if (PageAnon(head
) && PageSwapCache(head
)) {
2521 swp_entry_t entry
= { .val
= page_private(head
) };
2523 offset
= swp_offset(entry
);
2524 swap_cache
= swap_address_space(entry
);
2525 xa_lock(&swap_cache
->i_pages
);
2528 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2529 lruvec
= folio_lruvec_lock(folio
);
2531 ClearPageHasHWPoisoned(head
);
2533 for (i
= nr
- 1; i
>= 1; i
--) {
2534 __split_huge_page_tail(head
, i
, lruvec
, list
);
2535 /* Some pages can be beyond EOF: drop them from page cache */
2536 if (head
[i
].index
>= end
) {
2537 struct folio
*tail
= page_folio(head
+ i
);
2539 if (shmem_mapping(head
->mapping
))
2540 shmem_uncharge(head
->mapping
->host
, 1);
2541 else if (folio_test_clear_dirty(tail
))
2542 folio_account_cleaned(tail
,
2543 inode_to_wb(folio
->mapping
->host
));
2544 __filemap_remove_folio(tail
, NULL
);
2546 } else if (!PageAnon(page
)) {
2547 __xa_store(&head
->mapping
->i_pages
, head
[i
].index
,
2549 } else if (swap_cache
) {
2550 __xa_store(&swap_cache
->i_pages
, offset
+ i
,
2555 ClearPageCompound(head
);
2556 unlock_page_lruvec(lruvec
);
2557 /* Caller disabled irqs, so they are still disabled here */
2559 split_page_owner(head
, nr
);
2561 /* See comment in __split_huge_page_tail() */
2562 if (PageAnon(head
)) {
2563 /* Additional pin to swap cache */
2564 if (PageSwapCache(head
)) {
2565 page_ref_add(head
, 2);
2566 xa_unlock(&swap_cache
->i_pages
);
2571 /* Additional pin to page cache */
2572 page_ref_add(head
, 2);
2573 xa_unlock(&head
->mapping
->i_pages
);
2577 remap_page(folio
, nr
);
2579 if (PageSwapCache(head
)) {
2580 swp_entry_t entry
= { .val
= page_private(head
) };
2582 split_swap_cluster(entry
);
2585 for (i
= 0; i
< nr
; i
++) {
2586 struct page
*subpage
= head
+ i
;
2587 if (subpage
== page
)
2589 unlock_page(subpage
);
2592 * Subpages may be freed if there wasn't any mapping
2593 * like if add_to_swap() is running on a lru page that
2594 * had its mapping zapped. And freeing these pages
2595 * requires taking the lru_lock so we do the put_page
2596 * of the tail pages after the split is complete.
2598 free_page_and_swap_cache(subpage
);
2602 /* Racy check whether the huge page can be split */
2603 bool can_split_folio(struct folio
*folio
, int *pextra_pins
)
2607 /* Additional pins from page cache */
2608 if (folio_test_anon(folio
))
2609 extra_pins
= folio_test_swapcache(folio
) ?
2610 folio_nr_pages(folio
) : 0;
2612 extra_pins
= folio_nr_pages(folio
);
2614 *pextra_pins
= extra_pins
;
2615 return folio_mapcount(folio
) == folio_ref_count(folio
) - extra_pins
- 1;
2619 * This function splits huge page into normal pages. @page can point to any
2620 * subpage of huge page to split. Split doesn't change the position of @page.
2622 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2623 * The huge page must be locked.
2625 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2627 * Both head page and tail pages will inherit mapping, flags, and so on from
2630 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2631 * they are not mapped.
2633 * Returns 0 if the hugepage is split successfully.
2634 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2637 int split_huge_page_to_list(struct page
*page
, struct list_head
*list
)
2639 struct folio
*folio
= page_folio(page
);
2640 struct deferred_split
*ds_queue
= get_deferred_split_queue(folio
);
2641 XA_STATE(xas
, &folio
->mapping
->i_pages
, folio
->index
);
2642 struct anon_vma
*anon_vma
= NULL
;
2643 struct address_space
*mapping
= NULL
;
2644 int extra_pins
, ret
;
2648 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
2649 VM_BUG_ON_FOLIO(!folio_test_large(folio
), folio
);
2651 is_hzp
= is_huge_zero_page(&folio
->page
);
2653 pr_warn_ratelimited("Called split_huge_page for huge zero page\n");
2657 if (folio_test_writeback(folio
))
2660 if (folio_test_anon(folio
)) {
2662 * The caller does not necessarily hold an mmap_lock that would
2663 * prevent the anon_vma disappearing so we first we take a
2664 * reference to it and then lock the anon_vma for write. This
2665 * is similar to folio_lock_anon_vma_read except the write lock
2666 * is taken to serialise against parallel split or collapse
2669 anon_vma
= folio_get_anon_vma(folio
);
2676 anon_vma_lock_write(anon_vma
);
2680 mapping
= folio
->mapping
;
2688 gfp
= current_gfp_context(mapping_gfp_mask(mapping
) &
2691 if (folio_test_private(folio
) &&
2692 !filemap_release_folio(folio
, gfp
)) {
2697 xas_split_alloc(&xas
, folio
, folio_order(folio
), gfp
);
2698 if (xas_error(&xas
)) {
2699 ret
= xas_error(&xas
);
2704 i_mmap_lock_read(mapping
);
2707 *__split_huge_page() may need to trim off pages beyond EOF:
2708 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2709 * which cannot be nested inside the page tree lock. So note
2710 * end now: i_size itself may be changed at any moment, but
2711 * folio lock is good enough to serialize the trimming.
2713 end
= DIV_ROUND_UP(i_size_read(mapping
->host
), PAGE_SIZE
);
2714 if (shmem_mapping(mapping
))
2715 end
= shmem_fallocend(mapping
->host
, end
);
2719 * Racy check if we can split the page, before unmap_folio() will
2722 if (!can_split_folio(folio
, &extra_pins
)) {
2729 /* block interrupt reentry in xa_lock and spinlock */
2730 local_irq_disable();
2733 * Check if the folio is present in page cache.
2734 * We assume all tail are present too, if folio is there.
2738 if (xas_load(&xas
) != folio
)
2742 /* Prevent deferred_split_scan() touching ->_refcount */
2743 spin_lock(&ds_queue
->split_queue_lock
);
2744 if (folio_ref_freeze(folio
, 1 + extra_pins
)) {
2745 if (!list_empty(&folio
->_deferred_list
)) {
2746 ds_queue
->split_queue_len
--;
2747 list_del(&folio
->_deferred_list
);
2749 spin_unlock(&ds_queue
->split_queue_lock
);
2751 int nr
= folio_nr_pages(folio
);
2753 xas_split(&xas
, folio
, folio_order(folio
));
2754 if (folio_test_swapbacked(folio
)) {
2755 __lruvec_stat_mod_folio(folio
, NR_SHMEM_THPS
,
2758 __lruvec_stat_mod_folio(folio
, NR_FILE_THPS
,
2760 filemap_nr_thps_dec(mapping
);
2764 __split_huge_page(page
, list
, end
);
2767 spin_unlock(&ds_queue
->split_queue_lock
);
2772 remap_page(folio
, folio_nr_pages(folio
));
2778 anon_vma_unlock_write(anon_vma
);
2779 put_anon_vma(anon_vma
);
2782 i_mmap_unlock_read(mapping
);
2785 count_vm_event(!ret
? THP_SPLIT_PAGE
: THP_SPLIT_PAGE_FAILED
);
2789 void free_transhuge_page(struct page
*page
)
2791 struct folio
*folio
= (struct folio
*)page
;
2792 struct deferred_split
*ds_queue
= get_deferred_split_queue(folio
);
2793 unsigned long flags
;
2795 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2796 if (!list_empty(&folio
->_deferred_list
)) {
2797 ds_queue
->split_queue_len
--;
2798 list_del(&folio
->_deferred_list
);
2800 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2801 free_compound_page(page
);
2804 void deferred_split_folio(struct folio
*folio
)
2806 struct deferred_split
*ds_queue
= get_deferred_split_queue(folio
);
2808 struct mem_cgroup
*memcg
= folio_memcg(folio
);
2810 unsigned long flags
;
2812 VM_BUG_ON_FOLIO(folio_order(folio
) < 2, folio
);
2815 * The try_to_unmap() in page reclaim path might reach here too,
2816 * this may cause a race condition to corrupt deferred split queue.
2817 * And, if page reclaim is already handling the same folio, it is
2818 * unnecessary to handle it again in shrinker.
2820 * Check the swapcache flag to determine if the folio is being
2821 * handled by page reclaim since THP swap would add the folio into
2822 * swap cache before calling try_to_unmap().
2824 if (folio_test_swapcache(folio
))
2827 if (!list_empty(&folio
->_deferred_list
))
2830 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2831 if (list_empty(&folio
->_deferred_list
)) {
2832 count_vm_event(THP_DEFERRED_SPLIT_PAGE
);
2833 list_add_tail(&folio
->_deferred_list
, &ds_queue
->split_queue
);
2834 ds_queue
->split_queue_len
++;
2837 set_shrinker_bit(memcg
, folio_nid(folio
),
2838 deferred_split_shrinker
.id
);
2841 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2844 static unsigned long deferred_split_count(struct shrinker
*shrink
,
2845 struct shrink_control
*sc
)
2847 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2848 struct deferred_split
*ds_queue
= &pgdata
->deferred_split_queue
;
2852 ds_queue
= &sc
->memcg
->deferred_split_queue
;
2854 return READ_ONCE(ds_queue
->split_queue_len
);
2857 static unsigned long deferred_split_scan(struct shrinker
*shrink
,
2858 struct shrink_control
*sc
)
2860 struct pglist_data
*pgdata
= NODE_DATA(sc
->nid
);
2861 struct deferred_split
*ds_queue
= &pgdata
->deferred_split_queue
;
2862 unsigned long flags
;
2864 struct folio
*folio
, *next
;
2869 ds_queue
= &sc
->memcg
->deferred_split_queue
;
2872 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2873 /* Take pin on all head pages to avoid freeing them under us */
2874 list_for_each_entry_safe(folio
, next
, &ds_queue
->split_queue
,
2876 if (folio_try_get(folio
)) {
2877 list_move(&folio
->_deferred_list
, &list
);
2879 /* We lost race with folio_put() */
2880 list_del_init(&folio
->_deferred_list
);
2881 ds_queue
->split_queue_len
--;
2883 if (!--sc
->nr_to_scan
)
2886 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2888 list_for_each_entry_safe(folio
, next
, &list
, _deferred_list
) {
2889 if (!folio_trylock(folio
))
2891 /* split_huge_page() removes page from list on success */
2892 if (!split_folio(folio
))
2894 folio_unlock(folio
);
2899 spin_lock_irqsave(&ds_queue
->split_queue_lock
, flags
);
2900 list_splice_tail(&list
, &ds_queue
->split_queue
);
2901 spin_unlock_irqrestore(&ds_queue
->split_queue_lock
, flags
);
2904 * Stop shrinker if we didn't split any page, but the queue is empty.
2905 * This can happen if pages were freed under us.
2907 if (!split
&& list_empty(&ds_queue
->split_queue
))
2912 static struct shrinker deferred_split_shrinker
= {
2913 .count_objects
= deferred_split_count
,
2914 .scan_objects
= deferred_split_scan
,
2915 .seeks
= DEFAULT_SEEKS
,
2916 .flags
= SHRINKER_NUMA_AWARE
| SHRINKER_MEMCG_AWARE
|
2920 #ifdef CONFIG_DEBUG_FS
2921 static void split_huge_pages_all(void)
2925 struct folio
*folio
;
2926 unsigned long pfn
, max_zone_pfn
;
2927 unsigned long total
= 0, split
= 0;
2929 pr_debug("Split all THPs\n");
2930 for_each_zone(zone
) {
2931 if (!managed_zone(zone
))
2933 max_zone_pfn
= zone_end_pfn(zone
);
2934 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++) {
2937 page
= pfn_to_online_page(pfn
);
2938 if (!page
|| PageTail(page
))
2940 folio
= page_folio(page
);
2941 if (!folio_try_get(folio
))
2944 if (unlikely(page_folio(page
) != folio
))
2947 if (zone
!= folio_zone(folio
))
2950 if (!folio_test_large(folio
)
2951 || folio_test_hugetlb(folio
)
2952 || !folio_test_lru(folio
))
2957 nr_pages
= folio_nr_pages(folio
);
2958 if (!split_folio(folio
))
2960 pfn
+= nr_pages
- 1;
2961 folio_unlock(folio
);
2968 pr_debug("%lu of %lu THP split\n", split
, total
);
2971 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct
*vma
)
2973 return vma_is_special_huge(vma
) || (vma
->vm_flags
& VM_IO
) ||
2974 is_vm_hugetlb_page(vma
);
2977 static int split_huge_pages_pid(int pid
, unsigned long vaddr_start
,
2978 unsigned long vaddr_end
)
2981 struct task_struct
*task
;
2982 struct mm_struct
*mm
;
2983 unsigned long total
= 0, split
= 0;
2986 vaddr_start
&= PAGE_MASK
;
2987 vaddr_end
&= PAGE_MASK
;
2989 /* Find the task_struct from pid */
2991 task
= find_task_by_vpid(pid
);
2997 get_task_struct(task
);
3000 /* Find the mm_struct */
3001 mm
= get_task_mm(task
);
3002 put_task_struct(task
);
3009 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3010 pid
, vaddr_start
, vaddr_end
);
3014 * always increase addr by PAGE_SIZE, since we could have a PTE page
3015 * table filled with PTE-mapped THPs, each of which is distinct.
3017 for (addr
= vaddr_start
; addr
< vaddr_end
; addr
+= PAGE_SIZE
) {
3018 struct vm_area_struct
*vma
= vma_lookup(mm
, addr
);
3024 /* skip special VMA and hugetlb VMA */
3025 if (vma_not_suitable_for_thp_split(vma
)) {
3030 /* FOLL_DUMP to ignore special (like zero) pages */
3031 page
= follow_page(vma
, addr
, FOLL_GET
| FOLL_DUMP
);
3033 if (IS_ERR_OR_NULL(page
))
3036 if (!is_transparent_hugepage(page
))
3040 if (!can_split_folio(page_folio(page
), NULL
))
3043 if (!trylock_page(page
))
3046 if (!split_huge_page(page
))
3054 mmap_read_unlock(mm
);
3057 pr_debug("%lu of %lu THP split\n", split
, total
);
3063 static int split_huge_pages_in_file(const char *file_path
, pgoff_t off_start
,
3066 struct filename
*file
;
3067 struct file
*candidate
;
3068 struct address_space
*mapping
;
3072 unsigned long total
= 0, split
= 0;
3074 file
= getname_kernel(file_path
);
3078 candidate
= file_open_name(file
, O_RDONLY
, 0);
3079 if (IS_ERR(candidate
))
3082 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3083 file_path
, off_start
, off_end
);
3085 mapping
= candidate
->f_mapping
;
3087 for (index
= off_start
; index
< off_end
; index
+= nr_pages
) {
3088 struct folio
*folio
= filemap_get_folio(mapping
, index
);
3094 if (!folio_test_large(folio
))
3098 nr_pages
= folio_nr_pages(folio
);
3100 if (!folio_trylock(folio
))
3103 if (!split_folio(folio
))
3106 folio_unlock(folio
);
3112 filp_close(candidate
, NULL
);
3115 pr_debug("%lu of %lu file-backed THP split\n", split
, total
);
3121 #define MAX_INPUT_BUF_SZ 255
3123 static ssize_t
split_huge_pages_write(struct file
*file
, const char __user
*buf
,
3124 size_t count
, loff_t
*ppops
)
3126 static DEFINE_MUTEX(split_debug_mutex
);
3128 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3129 char input_buf
[MAX_INPUT_BUF_SZ
];
3131 unsigned long vaddr_start
, vaddr_end
;
3133 ret
= mutex_lock_interruptible(&split_debug_mutex
);
3139 memset(input_buf
, 0, MAX_INPUT_BUF_SZ
);
3140 if (copy_from_user(input_buf
, buf
, min_t(size_t, count
, MAX_INPUT_BUF_SZ
)))
3143 input_buf
[MAX_INPUT_BUF_SZ
- 1] = '\0';
3145 if (input_buf
[0] == '/') {
3147 char *buf
= input_buf
;
3148 char file_path
[MAX_INPUT_BUF_SZ
];
3149 pgoff_t off_start
= 0, off_end
= 0;
3150 size_t input_len
= strlen(input_buf
);
3152 tok
= strsep(&buf
, ",");
3154 strcpy(file_path
, tok
);
3160 ret
= sscanf(buf
, "0x%lx,0x%lx", &off_start
, &off_end
);
3165 ret
= split_huge_pages_in_file(file_path
, off_start
, off_end
);
3172 ret
= sscanf(input_buf
, "%d,0x%lx,0x%lx", &pid
, &vaddr_start
, &vaddr_end
);
3173 if (ret
== 1 && pid
== 1) {
3174 split_huge_pages_all();
3175 ret
= strlen(input_buf
);
3177 } else if (ret
!= 3) {
3182 ret
= split_huge_pages_pid(pid
, vaddr_start
, vaddr_end
);
3184 ret
= strlen(input_buf
);
3186 mutex_unlock(&split_debug_mutex
);
3191 static const struct file_operations split_huge_pages_fops
= {
3192 .owner
= THIS_MODULE
,
3193 .write
= split_huge_pages_write
,
3194 .llseek
= no_llseek
,
3197 static int __init
split_huge_pages_debugfs(void)
3199 debugfs_create_file("split_huge_pages", 0200, NULL
, NULL
,
3200 &split_huge_pages_fops
);
3203 late_initcall(split_huge_pages_debugfs
);
3206 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3207 int set_pmd_migration_entry(struct page_vma_mapped_walk
*pvmw
,
3210 struct vm_area_struct
*vma
= pvmw
->vma
;
3211 struct mm_struct
*mm
= vma
->vm_mm
;
3212 unsigned long address
= pvmw
->address
;
3213 bool anon_exclusive
;
3218 if (!(pvmw
->pmd
&& !pvmw
->pte
))
3221 flush_cache_range(vma
, address
, address
+ HPAGE_PMD_SIZE
);
3222 pmdval
= pmdp_invalidate(vma
, address
, pvmw
->pmd
);
3224 /* See page_try_share_anon_rmap(): invalidate PMD first. */
3225 anon_exclusive
= PageAnon(page
) && PageAnonExclusive(page
);
3226 if (anon_exclusive
&& page_try_share_anon_rmap(page
)) {
3227 set_pmd_at(mm
, address
, pvmw
->pmd
, pmdval
);
3231 if (pmd_dirty(pmdval
))
3232 set_page_dirty(page
);
3233 if (pmd_write(pmdval
))
3234 entry
= make_writable_migration_entry(page_to_pfn(page
));
3235 else if (anon_exclusive
)
3236 entry
= make_readable_exclusive_migration_entry(page_to_pfn(page
));
3238 entry
= make_readable_migration_entry(page_to_pfn(page
));
3239 if (pmd_young(pmdval
))
3240 entry
= make_migration_entry_young(entry
);
3241 if (pmd_dirty(pmdval
))
3242 entry
= make_migration_entry_dirty(entry
);
3243 pmdswp
= swp_entry_to_pmd(entry
);
3244 if (pmd_soft_dirty(pmdval
))
3245 pmdswp
= pmd_swp_mksoft_dirty(pmdswp
);
3246 if (pmd_uffd_wp(pmdval
))
3247 pmdswp
= pmd_swp_mkuffd_wp(pmdswp
);
3248 set_pmd_at(mm
, address
, pvmw
->pmd
, pmdswp
);
3249 page_remove_rmap(page
, vma
, true);
3251 trace_set_migration_pmd(address
, pmd_val(pmdswp
));
3256 void remove_migration_pmd(struct page_vma_mapped_walk
*pvmw
, struct page
*new)
3258 struct vm_area_struct
*vma
= pvmw
->vma
;
3259 struct mm_struct
*mm
= vma
->vm_mm
;
3260 unsigned long address
= pvmw
->address
;
3261 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
3265 if (!(pvmw
->pmd
&& !pvmw
->pte
))
3268 entry
= pmd_to_swp_entry(*pvmw
->pmd
);
3270 pmde
= mk_huge_pmd(new, READ_ONCE(vma
->vm_page_prot
));
3271 if (pmd_swp_soft_dirty(*pvmw
->pmd
))
3272 pmde
= pmd_mksoft_dirty(pmde
);
3273 if (is_writable_migration_entry(entry
))
3274 pmde
= pmd_mkwrite(pmde
);
3275 if (pmd_swp_uffd_wp(*pvmw
->pmd
))
3276 pmde
= pmd_mkuffd_wp(pmde
);
3277 if (!is_migration_entry_young(entry
))
3278 pmde
= pmd_mkold(pmde
);
3279 /* NOTE: this may contain setting soft-dirty on some archs */
3280 if (PageDirty(new) && is_migration_entry_dirty(entry
))
3281 pmde
= pmd_mkdirty(pmde
);
3283 if (PageAnon(new)) {
3284 rmap_t rmap_flags
= RMAP_COMPOUND
;
3286 if (!is_readable_migration_entry(entry
))
3287 rmap_flags
|= RMAP_EXCLUSIVE
;
3289 page_add_anon_rmap(new, vma
, haddr
, rmap_flags
);
3291 page_add_file_rmap(new, vma
, true);
3293 VM_BUG_ON(pmd_write(pmde
) && PageAnon(new) && !PageAnonExclusive(new));
3294 set_pmd_at(mm
, haddr
, pvmw
->pmd
, pmde
);
3296 /* No need to invalidate - it was non-present before */
3297 update_mmu_cache_pmd(vma
, address
, pvmw
->pmd
);
3298 trace_remove_migration_pmd(address
, pmd_val(pmde
));