2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
27 #include <asm/pgtable.h>
31 #include <linux/hugetlb.h>
32 #include <linux/hugetlb_cgroup.h>
33 #include <linux/node.h>
36 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
37 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
38 unsigned long hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 __initdata
LIST_HEAD(huge_boot_pages
);
46 /* for command line parsing */
47 static struct hstate
* __initdata parsed_hstate
;
48 static unsigned long __initdata default_hstate_max_huge_pages
;
49 static unsigned long __initdata default_hstate_size
;
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 DEFINE_SPINLOCK(hugetlb_lock
);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
58 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
60 spin_unlock(&spool
->lock
);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
70 struct hugepage_subpool
*spool
;
72 spool
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
76 spin_lock_init(&spool
->lock
);
78 spool
->max_hpages
= nr_blocks
;
79 spool
->used_hpages
= 0;
84 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
86 spin_lock(&spool
->lock
);
87 BUG_ON(!spool
->count
);
89 unlock_or_release_subpool(spool
);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
100 spin_lock(&spool
->lock
);
101 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
102 spool
->used_hpages
+= delta
;
106 spin_unlock(&spool
->lock
);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
117 spin_lock(&spool
->lock
);
118 spool
->used_hpages
-= delta
;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool
);
124 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
126 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
129 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
131 return subpool_inode(file_inode(vma
->vm_file
));
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link
;
154 static long region_add(struct list_head
*head
, long f
, long t
)
156 struct file_region
*rg
, *nrg
, *trg
;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg
, head
, link
)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
170 if (&rg
->link
== head
)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head
*head
, long f
, long t
)
192 struct file_region
*rg
, *nrg
;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg
, head
, link
)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg
->link
== head
|| t
< rg
->from
) {
204 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
209 INIT_LIST_HEAD(&nrg
->link
);
210 list_add(&nrg
->link
, rg
->link
.prev
);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
222 if (&rg
->link
== head
)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg
-= rg
->to
- rg
->from
;
239 static long region_truncate(struct list_head
*head
, long end
)
241 struct file_region
*rg
, *trg
;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg
, head
, link
)
248 if (&rg
->link
== head
)
251 /* If we are in the middle of a region then adjust it. */
252 if (end
> rg
->from
) {
255 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
260 if (&rg
->link
== head
)
262 chg
+= rg
->to
- rg
->from
;
269 static long region_count(struct list_head
*head
, long f
, long t
)
271 struct file_region
*rg
;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg
, head
, link
) {
284 seg_from
= max(rg
->from
, f
);
285 seg_to
= min(rg
->to
, t
);
287 chg
+= seg_to
- seg_from
;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
298 struct vm_area_struct
*vma
, unsigned long address
)
300 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
301 (vma
->vm_pgoff
>> huge_page_order(h
));
304 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
305 unsigned long address
)
307 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
316 struct hstate
*hstate
;
318 if (!is_vm_hugetlb_page(vma
))
321 hstate
= hstate_vma(vma
);
323 return 1UL << (hstate
->order
+ PAGE_SHIFT
);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
336 return vma_kernel_pagesize(vma
);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
370 return (unsigned long)vma
->vm_private_data
;
373 static void set_vma_private_data(struct vm_area_struct
*vma
,
376 vma
->vm_private_data
= (void *)value
;
381 struct list_head regions
;
384 static struct resv_map
*resv_map_alloc(void)
386 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
390 kref_init(&resv_map
->refs
);
391 INIT_LIST_HEAD(&resv_map
->regions
);
396 static void resv_map_release(struct kref
*ref
)
398 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map
->regions
, 0);
405 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
408 if (!(vma
->vm_flags
& VM_MAYSHARE
))
409 return (struct resv_map
*)(get_vma_private_data(vma
) &
414 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
417 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
419 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
420 HPAGE_RESV_MASK
) | (unsigned long)map
);
423 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
426 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
428 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
431 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
435 return (get_vma_private_data(vma
) & flag
) != 0;
438 /* Decrement the reserved pages in the hugepage pool by one */
439 static void decrement_hugepage_resv_vma(struct hstate
*h
,
440 struct vm_area_struct
*vma
)
442 if (vma
->vm_flags
& VM_NORESERVE
)
445 if (vma
->vm_flags
& VM_MAYSHARE
) {
446 /* Shared mappings always use reserves */
447 h
->resv_huge_pages
--;
448 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
450 * Only the process that called mmap() has reserves for
453 h
->resv_huge_pages
--;
457 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
458 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
460 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
461 if (!(vma
->vm_flags
& VM_MAYSHARE
))
462 vma
->vm_private_data
= (void *)0;
465 /* Returns true if the VMA has associated reserve pages */
466 static int vma_has_reserves(struct vm_area_struct
*vma
)
468 if (vma
->vm_flags
& VM_MAYSHARE
)
470 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
475 static void copy_gigantic_page(struct page
*dst
, struct page
*src
)
478 struct hstate
*h
= page_hstate(src
);
479 struct page
*dst_base
= dst
;
480 struct page
*src_base
= src
;
482 for (i
= 0; i
< pages_per_huge_page(h
); ) {
484 copy_highpage(dst
, src
);
487 dst
= mem_map_next(dst
, dst_base
, i
);
488 src
= mem_map_next(src
, src_base
, i
);
492 void copy_huge_page(struct page
*dst
, struct page
*src
)
495 struct hstate
*h
= page_hstate(src
);
497 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
498 copy_gigantic_page(dst
, src
);
503 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
505 copy_highpage(dst
+ i
, src
+ i
);
509 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
511 int nid
= page_to_nid(page
);
512 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
513 h
->free_huge_pages
++;
514 h
->free_huge_pages_node
[nid
]++;
517 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
521 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
522 if (!is_migrate_isolate_page(page
))
525 * if 'non-isolated free hugepage' not found on the list,
526 * the allocation fails.
528 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
530 list_move(&page
->lru
, &h
->hugepage_activelist
);
531 set_page_refcounted(page
);
532 h
->free_huge_pages
--;
533 h
->free_huge_pages_node
[nid
]--;
537 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
538 struct vm_area_struct
*vma
,
539 unsigned long address
, int avoid_reserve
)
541 struct page
*page
= NULL
;
542 struct mempolicy
*mpol
;
543 nodemask_t
*nodemask
;
544 struct zonelist
*zonelist
;
547 unsigned int cpuset_mems_cookie
;
550 cpuset_mems_cookie
= get_mems_allowed();
551 zonelist
= huge_zonelist(vma
, address
,
552 htlb_alloc_mask
, &mpol
, &nodemask
);
554 * A child process with MAP_PRIVATE mappings created by their parent
555 * have no page reserves. This check ensures that reservations are
556 * not "stolen". The child may still get SIGKILLed
558 if (!vma_has_reserves(vma
) &&
559 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
562 /* If reserves cannot be used, ensure enough pages are in the pool */
563 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
566 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
567 MAX_NR_ZONES
- 1, nodemask
) {
568 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
569 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
572 decrement_hugepage_resv_vma(h
, vma
);
579 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
588 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
592 VM_BUG_ON(h
->order
>= MAX_ORDER
);
595 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
596 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
597 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
598 1 << PG_referenced
| 1 << PG_dirty
|
599 1 << PG_active
| 1 << PG_reserved
|
600 1 << PG_private
| 1 << PG_writeback
);
602 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
603 set_compound_page_dtor(page
, NULL
);
604 set_page_refcounted(page
);
605 arch_release_hugepage(page
);
606 __free_pages(page
, huge_page_order(h
));
609 struct hstate
*size_to_hstate(unsigned long size
)
614 if (huge_page_size(h
) == size
)
620 static void free_huge_page(struct page
*page
)
623 * Can't pass hstate in here because it is called from the
624 * compound page destructor.
626 struct hstate
*h
= page_hstate(page
);
627 int nid
= page_to_nid(page
);
628 struct hugepage_subpool
*spool
=
629 (struct hugepage_subpool
*)page_private(page
);
631 set_page_private(page
, 0);
632 page
->mapping
= NULL
;
633 BUG_ON(page_count(page
));
634 BUG_ON(page_mapcount(page
));
636 spin_lock(&hugetlb_lock
);
637 hugetlb_cgroup_uncharge_page(hstate_index(h
),
638 pages_per_huge_page(h
), page
);
639 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
640 /* remove the page from active list */
641 list_del(&page
->lru
);
642 update_and_free_page(h
, page
);
643 h
->surplus_huge_pages
--;
644 h
->surplus_huge_pages_node
[nid
]--;
646 arch_clear_hugepage_flags(page
);
647 enqueue_huge_page(h
, page
);
649 spin_unlock(&hugetlb_lock
);
650 hugepage_subpool_put_pages(spool
, 1);
653 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
655 INIT_LIST_HEAD(&page
->lru
);
656 set_compound_page_dtor(page
, free_huge_page
);
657 spin_lock(&hugetlb_lock
);
658 set_hugetlb_cgroup(page
, NULL
);
660 h
->nr_huge_pages_node
[nid
]++;
661 spin_unlock(&hugetlb_lock
);
662 put_page(page
); /* free it into the hugepage allocator */
665 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
668 int nr_pages
= 1 << order
;
669 struct page
*p
= page
+ 1;
671 /* we rely on prep_new_huge_page to set the destructor */
672 set_compound_order(page
, order
);
674 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
676 set_page_count(p
, 0);
677 p
->first_page
= page
;
682 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
683 * transparent huge pages. See the PageTransHuge() documentation for more
686 int PageHuge(struct page
*page
)
688 compound_page_dtor
*dtor
;
690 if (!PageCompound(page
))
693 page
= compound_head(page
);
694 dtor
= get_compound_page_dtor(page
);
696 return dtor
== free_huge_page
;
698 EXPORT_SYMBOL_GPL(PageHuge
);
701 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
702 * normal or transparent huge pages.
704 int PageHeadHuge(struct page
*page_head
)
706 compound_page_dtor
*dtor
;
708 if (!PageHead(page_head
))
711 dtor
= get_compound_page_dtor(page_head
);
713 return dtor
== free_huge_page
;
715 EXPORT_SYMBOL_GPL(PageHeadHuge
);
717 pgoff_t
__basepage_index(struct page
*page
)
719 struct page
*page_head
= compound_head(page
);
720 pgoff_t index
= page_index(page_head
);
721 unsigned long compound_idx
;
723 if (!PageHuge(page_head
))
724 return page_index(page
);
726 if (compound_order(page_head
) >= MAX_ORDER
)
727 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
729 compound_idx
= page
- page_head
;
731 return (index
<< compound_order(page_head
)) + compound_idx
;
734 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
738 if (h
->order
>= MAX_ORDER
)
741 page
= alloc_pages_exact_node(nid
,
742 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
743 __GFP_REPEAT
|__GFP_NOWARN
,
746 if (arch_prepare_hugepage(page
)) {
747 __free_pages(page
, huge_page_order(h
));
750 prep_new_huge_page(h
, page
, nid
);
757 * common helper functions for hstate_next_node_to_{alloc|free}.
758 * We may have allocated or freed a huge page based on a different
759 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
760 * be outside of *nodes_allowed. Ensure that we use an allowed
761 * node for alloc or free.
763 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
765 nid
= next_node(nid
, *nodes_allowed
);
766 if (nid
== MAX_NUMNODES
)
767 nid
= first_node(*nodes_allowed
);
768 VM_BUG_ON(nid
>= MAX_NUMNODES
);
773 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
775 if (!node_isset(nid
, *nodes_allowed
))
776 nid
= next_node_allowed(nid
, nodes_allowed
);
781 * returns the previously saved node ["this node"] from which to
782 * allocate a persistent huge page for the pool and advance the
783 * next node from which to allocate, handling wrap at end of node
786 static int hstate_next_node_to_alloc(struct hstate
*h
,
787 nodemask_t
*nodes_allowed
)
791 VM_BUG_ON(!nodes_allowed
);
793 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
794 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
799 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
806 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
807 next_nid
= start_nid
;
810 page
= alloc_fresh_huge_page_node(h
, next_nid
);
815 next_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
816 } while (next_nid
!= start_nid
);
819 count_vm_event(HTLB_BUDDY_PGALLOC
);
821 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
827 * helper for free_pool_huge_page() - return the previously saved
828 * node ["this node"] from which to free a huge page. Advance the
829 * next node id whether or not we find a free huge page to free so
830 * that the next attempt to free addresses the next node.
832 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
836 VM_BUG_ON(!nodes_allowed
);
838 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
839 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
845 * Free huge page from pool from next node to free.
846 * Attempt to keep persistent huge pages more or less
847 * balanced over allowed nodes.
848 * Called with hugetlb_lock locked.
850 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
857 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
858 next_nid
= start_nid
;
862 * If we're returning unused surplus pages, only examine
863 * nodes with surplus pages.
865 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[next_nid
]) &&
866 !list_empty(&h
->hugepage_freelists
[next_nid
])) {
868 list_entry(h
->hugepage_freelists
[next_nid
].next
,
870 list_del(&page
->lru
);
871 h
->free_huge_pages
--;
872 h
->free_huge_pages_node
[next_nid
]--;
874 h
->surplus_huge_pages
--;
875 h
->surplus_huge_pages_node
[next_nid
]--;
877 update_and_free_page(h
, page
);
881 next_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
882 } while (next_nid
!= start_nid
);
887 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
892 if (h
->order
>= MAX_ORDER
)
896 * Assume we will successfully allocate the surplus page to
897 * prevent racing processes from causing the surplus to exceed
900 * This however introduces a different race, where a process B
901 * tries to grow the static hugepage pool while alloc_pages() is
902 * called by process A. B will only examine the per-node
903 * counters in determining if surplus huge pages can be
904 * converted to normal huge pages in adjust_pool_surplus(). A
905 * won't be able to increment the per-node counter, until the
906 * lock is dropped by B, but B doesn't drop hugetlb_lock until
907 * no more huge pages can be converted from surplus to normal
908 * state (and doesn't try to convert again). Thus, we have a
909 * case where a surplus huge page exists, the pool is grown, and
910 * the surplus huge page still exists after, even though it
911 * should just have been converted to a normal huge page. This
912 * does not leak memory, though, as the hugepage will be freed
913 * once it is out of use. It also does not allow the counters to
914 * go out of whack in adjust_pool_surplus() as we don't modify
915 * the node values until we've gotten the hugepage and only the
916 * per-node value is checked there.
918 spin_lock(&hugetlb_lock
);
919 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
920 spin_unlock(&hugetlb_lock
);
924 h
->surplus_huge_pages
++;
926 spin_unlock(&hugetlb_lock
);
928 if (nid
== NUMA_NO_NODE
)
929 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
930 __GFP_REPEAT
|__GFP_NOWARN
,
933 page
= alloc_pages_exact_node(nid
,
934 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
935 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
937 if (page
&& arch_prepare_hugepage(page
)) {
938 __free_pages(page
, huge_page_order(h
));
942 spin_lock(&hugetlb_lock
);
944 INIT_LIST_HEAD(&page
->lru
);
945 r_nid
= page_to_nid(page
);
946 set_compound_page_dtor(page
, free_huge_page
);
947 set_hugetlb_cgroup(page
, NULL
);
949 * We incremented the global counters already
951 h
->nr_huge_pages_node
[r_nid
]++;
952 h
->surplus_huge_pages_node
[r_nid
]++;
953 __count_vm_event(HTLB_BUDDY_PGALLOC
);
956 h
->surplus_huge_pages
--;
957 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
959 spin_unlock(&hugetlb_lock
);
965 * This allocation function is useful in the context where vma is irrelevant.
966 * E.g. soft-offlining uses this function because it only cares physical
967 * address of error page.
969 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
973 spin_lock(&hugetlb_lock
);
974 page
= dequeue_huge_page_node(h
, nid
);
975 spin_unlock(&hugetlb_lock
);
978 page
= alloc_buddy_huge_page(h
, nid
);
984 * Increase the hugetlb pool such that it can accommodate a reservation
987 static int gather_surplus_pages(struct hstate
*h
, int delta
)
989 struct list_head surplus_list
;
990 struct page
*page
, *tmp
;
992 int needed
, allocated
;
993 bool alloc_ok
= true;
995 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
997 h
->resv_huge_pages
+= delta
;
1002 INIT_LIST_HEAD(&surplus_list
);
1006 spin_unlock(&hugetlb_lock
);
1007 for (i
= 0; i
< needed
; i
++) {
1008 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1013 list_add(&page
->lru
, &surplus_list
);
1018 * After retaking hugetlb_lock, we need to recalculate 'needed'
1019 * because either resv_huge_pages or free_huge_pages may have changed.
1021 spin_lock(&hugetlb_lock
);
1022 needed
= (h
->resv_huge_pages
+ delta
) -
1023 (h
->free_huge_pages
+ allocated
);
1028 * We were not able to allocate enough pages to
1029 * satisfy the entire reservation so we free what
1030 * we've allocated so far.
1035 * The surplus_list now contains _at_least_ the number of extra pages
1036 * needed to accommodate the reservation. Add the appropriate number
1037 * of pages to the hugetlb pool and free the extras back to the buddy
1038 * allocator. Commit the entire reservation here to prevent another
1039 * process from stealing the pages as they are added to the pool but
1040 * before they are reserved.
1042 needed
+= allocated
;
1043 h
->resv_huge_pages
+= delta
;
1046 /* Free the needed pages to the hugetlb pool */
1047 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1051 * This page is now managed by the hugetlb allocator and has
1052 * no users -- drop the buddy allocator's reference.
1054 put_page_testzero(page
);
1055 VM_BUG_ON(page_count(page
));
1056 enqueue_huge_page(h
, page
);
1059 spin_unlock(&hugetlb_lock
);
1061 /* Free unnecessary surplus pages to the buddy allocator */
1062 if (!list_empty(&surplus_list
)) {
1063 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1067 spin_lock(&hugetlb_lock
);
1073 * This routine has two main purposes:
1074 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1075 * in unused_resv_pages. This corresponds to the prior adjustments made
1076 * to the associated reservation map.
1077 * 2) Free any unused surplus pages that may have been allocated to satisfy
1078 * the reservation. As many as unused_resv_pages may be freed.
1080 * Called with hugetlb_lock held. However, the lock could be dropped (and
1081 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1082 * we must make sure nobody else can claim pages we are in the process of
1083 * freeing. Do this by ensuring resv_huge_page always is greater than the
1084 * number of huge pages we plan to free when dropping the lock.
1086 static void return_unused_surplus_pages(struct hstate
*h
,
1087 unsigned long unused_resv_pages
)
1089 unsigned long nr_pages
;
1091 /* Cannot return gigantic pages currently */
1092 if (h
->order
>= MAX_ORDER
)
1096 * Part (or even all) of the reservation could have been backed
1097 * by pre-allocated pages. Only free surplus pages.
1099 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1102 * We want to release as many surplus pages as possible, spread
1103 * evenly across all nodes with memory. Iterate across these nodes
1104 * until we can no longer free unreserved surplus pages. This occurs
1105 * when the nodes with surplus pages have no free pages.
1106 * free_pool_huge_page() will balance the the freed pages across the
1107 * on-line nodes with memory and will handle the hstate accounting.
1109 * Note that we decrement resv_huge_pages as we free the pages. If
1110 * we drop the lock, resv_huge_pages will still be sufficiently large
1111 * to cover subsequent pages we may free.
1113 while (nr_pages
--) {
1114 h
->resv_huge_pages
--;
1115 unused_resv_pages
--;
1116 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1118 cond_resched_lock(&hugetlb_lock
);
1122 /* Fully uncommit the reservation */
1123 h
->resv_huge_pages
-= unused_resv_pages
;
1127 * Determine if the huge page at addr within the vma has an associated
1128 * reservation. Where it does not we will need to logically increase
1129 * reservation and actually increase subpool usage before an allocation
1130 * can occur. Where any new reservation would be required the
1131 * reservation change is prepared, but not committed. Once the page
1132 * has been allocated from the subpool and instantiated the change should
1133 * be committed via vma_commit_reservation. No action is required on
1136 static long vma_needs_reservation(struct hstate
*h
,
1137 struct vm_area_struct
*vma
, unsigned long addr
)
1139 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1140 struct inode
*inode
= mapping
->host
;
1142 if (vma
->vm_flags
& VM_MAYSHARE
) {
1143 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1144 return region_chg(&inode
->i_mapping
->private_list
,
1147 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1152 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1153 struct resv_map
*reservations
= vma_resv_map(vma
);
1155 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
1161 static void vma_commit_reservation(struct hstate
*h
,
1162 struct vm_area_struct
*vma
, unsigned long addr
)
1164 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1165 struct inode
*inode
= mapping
->host
;
1167 if (vma
->vm_flags
& VM_MAYSHARE
) {
1168 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1169 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1171 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1172 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1173 struct resv_map
*reservations
= vma_resv_map(vma
);
1175 /* Mark this page used in the map. */
1176 region_add(&reservations
->regions
, idx
, idx
+ 1);
1180 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1181 unsigned long addr
, int avoid_reserve
)
1183 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1184 struct hstate
*h
= hstate_vma(vma
);
1188 struct hugetlb_cgroup
*h_cg
;
1190 idx
= hstate_index(h
);
1192 * Processes that did not create the mapping will have no
1193 * reserves and will not have accounted against subpool
1194 * limit. Check that the subpool limit can be made before
1195 * satisfying the allocation MAP_NORESERVE mappings may also
1196 * need pages and subpool limit allocated allocated if no reserve
1199 chg
= vma_needs_reservation(h
, vma
, addr
);
1201 return ERR_PTR(-ENOMEM
);
1203 if (hugepage_subpool_get_pages(spool
, chg
))
1204 return ERR_PTR(-ENOSPC
);
1206 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1208 hugepage_subpool_put_pages(spool
, chg
);
1209 return ERR_PTR(-ENOSPC
);
1211 spin_lock(&hugetlb_lock
);
1212 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
1214 /* update page cgroup details */
1215 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
),
1217 spin_unlock(&hugetlb_lock
);
1219 spin_unlock(&hugetlb_lock
);
1220 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1222 hugetlb_cgroup_uncharge_cgroup(idx
,
1223 pages_per_huge_page(h
),
1225 hugepage_subpool_put_pages(spool
, chg
);
1226 return ERR_PTR(-ENOSPC
);
1228 spin_lock(&hugetlb_lock
);
1229 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
),
1231 list_move(&page
->lru
, &h
->hugepage_activelist
);
1232 spin_unlock(&hugetlb_lock
);
1235 set_page_private(page
, (unsigned long)spool
);
1237 vma_commit_reservation(h
, vma
, addr
);
1241 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1243 struct huge_bootmem_page
*m
;
1244 int nr_nodes
= nodes_weight(node_states
[N_MEMORY
]);
1249 addr
= __alloc_bootmem_node_nopanic(
1250 NODE_DATA(hstate_next_node_to_alloc(h
,
1251 &node_states
[N_MEMORY
])),
1252 huge_page_size(h
), huge_page_size(h
), 0);
1256 * Use the beginning of the huge page to store the
1257 * huge_bootmem_page struct (until gather_bootmem
1258 * puts them into the mem_map).
1268 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1269 /* Put them into a private list first because mem_map is not up yet */
1270 list_add(&m
->list
, &huge_boot_pages
);
1275 static void prep_compound_huge_page(struct page
*page
, int order
)
1277 if (unlikely(order
> (MAX_ORDER
- 1)))
1278 prep_compound_gigantic_page(page
, order
);
1280 prep_compound_page(page
, order
);
1283 /* Put bootmem huge pages into the standard lists after mem_map is up */
1284 static void __init
gather_bootmem_prealloc(void)
1286 struct huge_bootmem_page
*m
;
1288 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1289 struct hstate
*h
= m
->hstate
;
1292 #ifdef CONFIG_HIGHMEM
1293 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1294 free_bootmem_late((unsigned long)m
,
1295 sizeof(struct huge_bootmem_page
));
1297 page
= virt_to_page(m
);
1299 __ClearPageReserved(page
);
1300 WARN_ON(page_count(page
) != 1);
1301 prep_compound_huge_page(page
, h
->order
);
1302 prep_new_huge_page(h
, page
, page_to_nid(page
));
1304 * If we had gigantic hugepages allocated at boot time, we need
1305 * to restore the 'stolen' pages to totalram_pages in order to
1306 * fix confusing memory reports from free(1) and another
1307 * side-effects, like CommitLimit going negative.
1309 if (h
->order
> (MAX_ORDER
- 1))
1310 totalram_pages
+= 1 << h
->order
;
1314 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1318 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1319 if (h
->order
>= MAX_ORDER
) {
1320 if (!alloc_bootmem_huge_page(h
))
1322 } else if (!alloc_fresh_huge_page(h
,
1323 &node_states
[N_MEMORY
]))
1326 h
->max_huge_pages
= i
;
1329 static void __init
hugetlb_init_hstates(void)
1333 for_each_hstate(h
) {
1334 /* oversize hugepages were init'ed in early boot */
1335 if (h
->order
< MAX_ORDER
)
1336 hugetlb_hstate_alloc_pages(h
);
1340 static char * __init
memfmt(char *buf
, unsigned long n
)
1342 if (n
>= (1UL << 30))
1343 sprintf(buf
, "%lu GB", n
>> 30);
1344 else if (n
>= (1UL << 20))
1345 sprintf(buf
, "%lu MB", n
>> 20);
1347 sprintf(buf
, "%lu KB", n
>> 10);
1351 static void __init
report_hugepages(void)
1355 for_each_hstate(h
) {
1357 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1358 memfmt(buf
, huge_page_size(h
)),
1359 h
->free_huge_pages
);
1363 #ifdef CONFIG_HIGHMEM
1364 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1365 nodemask_t
*nodes_allowed
)
1369 if (h
->order
>= MAX_ORDER
)
1372 for_each_node_mask(i
, *nodes_allowed
) {
1373 struct page
*page
, *next
;
1374 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1375 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1376 if (count
>= h
->nr_huge_pages
)
1378 if (PageHighMem(page
))
1380 list_del(&page
->lru
);
1381 update_and_free_page(h
, page
);
1382 h
->free_huge_pages
--;
1383 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1388 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1389 nodemask_t
*nodes_allowed
)
1395 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1396 * balanced by operating on them in a round-robin fashion.
1397 * Returns 1 if an adjustment was made.
1399 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1402 int start_nid
, next_nid
;
1405 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1408 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
1410 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
1411 next_nid
= start_nid
;
1417 * To shrink on this node, there must be a surplus page
1419 if (!h
->surplus_huge_pages_node
[nid
]) {
1420 next_nid
= hstate_next_node_to_alloc(h
,
1427 * Surplus cannot exceed the total number of pages
1429 if (h
->surplus_huge_pages_node
[nid
] >=
1430 h
->nr_huge_pages_node
[nid
]) {
1431 next_nid
= hstate_next_node_to_free(h
,
1437 h
->surplus_huge_pages
+= delta
;
1438 h
->surplus_huge_pages_node
[nid
] += delta
;
1441 } while (next_nid
!= start_nid
);
1446 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1447 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1448 nodemask_t
*nodes_allowed
)
1450 unsigned long min_count
, ret
;
1452 if (h
->order
>= MAX_ORDER
)
1453 return h
->max_huge_pages
;
1456 * Increase the pool size
1457 * First take pages out of surplus state. Then make up the
1458 * remaining difference by allocating fresh huge pages.
1460 * We might race with alloc_buddy_huge_page() here and be unable
1461 * to convert a surplus huge page to a normal huge page. That is
1462 * not critical, though, it just means the overall size of the
1463 * pool might be one hugepage larger than it needs to be, but
1464 * within all the constraints specified by the sysctls.
1466 spin_lock(&hugetlb_lock
);
1467 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1468 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1472 while (count
> persistent_huge_pages(h
)) {
1474 * If this allocation races such that we no longer need the
1475 * page, free_huge_page will handle it by freeing the page
1476 * and reducing the surplus.
1478 spin_unlock(&hugetlb_lock
);
1479 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1480 spin_lock(&hugetlb_lock
);
1484 /* Bail for signals. Probably ctrl-c from user */
1485 if (signal_pending(current
))
1490 * Decrease the pool size
1491 * First return free pages to the buddy allocator (being careful
1492 * to keep enough around to satisfy reservations). Then place
1493 * pages into surplus state as needed so the pool will shrink
1494 * to the desired size as pages become free.
1496 * By placing pages into the surplus state independent of the
1497 * overcommit value, we are allowing the surplus pool size to
1498 * exceed overcommit. There are few sane options here. Since
1499 * alloc_buddy_huge_page() is checking the global counter,
1500 * though, we'll note that we're not allowed to exceed surplus
1501 * and won't grow the pool anywhere else. Not until one of the
1502 * sysctls are changed, or the surplus pages go out of use.
1504 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1505 min_count
= max(count
, min_count
);
1506 try_to_free_low(h
, min_count
, nodes_allowed
);
1507 while (min_count
< persistent_huge_pages(h
)) {
1508 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1510 cond_resched_lock(&hugetlb_lock
);
1512 while (count
< persistent_huge_pages(h
)) {
1513 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1517 ret
= persistent_huge_pages(h
);
1518 spin_unlock(&hugetlb_lock
);
1522 #define HSTATE_ATTR_RO(_name) \
1523 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1525 #define HSTATE_ATTR(_name) \
1526 static struct kobj_attribute _name##_attr = \
1527 __ATTR(_name, 0644, _name##_show, _name##_store)
1529 static struct kobject
*hugepages_kobj
;
1530 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1532 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1534 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1538 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1539 if (hstate_kobjs
[i
] == kobj
) {
1541 *nidp
= NUMA_NO_NODE
;
1545 return kobj_to_node_hstate(kobj
, nidp
);
1548 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1549 struct kobj_attribute
*attr
, char *buf
)
1552 unsigned long nr_huge_pages
;
1555 h
= kobj_to_hstate(kobj
, &nid
);
1556 if (nid
== NUMA_NO_NODE
)
1557 nr_huge_pages
= h
->nr_huge_pages
;
1559 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1561 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1564 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1565 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1566 const char *buf
, size_t len
)
1570 unsigned long count
;
1572 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1574 err
= strict_strtoul(buf
, 10, &count
);
1578 h
= kobj_to_hstate(kobj
, &nid
);
1579 if (h
->order
>= MAX_ORDER
) {
1584 if (nid
== NUMA_NO_NODE
) {
1586 * global hstate attribute
1588 if (!(obey_mempolicy
&&
1589 init_nodemask_of_mempolicy(nodes_allowed
))) {
1590 NODEMASK_FREE(nodes_allowed
);
1591 nodes_allowed
= &node_states
[N_MEMORY
];
1593 } else if (nodes_allowed
) {
1595 * per node hstate attribute: adjust count to global,
1596 * but restrict alloc/free to the specified node.
1598 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1599 init_nodemask_of_node(nodes_allowed
, nid
);
1601 nodes_allowed
= &node_states
[N_MEMORY
];
1603 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1605 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1606 NODEMASK_FREE(nodes_allowed
);
1610 NODEMASK_FREE(nodes_allowed
);
1614 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1615 struct kobj_attribute
*attr
, char *buf
)
1617 return nr_hugepages_show_common(kobj
, attr
, buf
);
1620 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1621 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1623 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1625 HSTATE_ATTR(nr_hugepages
);
1630 * hstate attribute for optionally mempolicy-based constraint on persistent
1631 * huge page alloc/free.
1633 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1634 struct kobj_attribute
*attr
, char *buf
)
1636 return nr_hugepages_show_common(kobj
, attr
, buf
);
1639 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1640 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1642 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1644 HSTATE_ATTR(nr_hugepages_mempolicy
);
1648 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1649 struct kobj_attribute
*attr
, char *buf
)
1651 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1652 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1655 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1656 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1659 unsigned long input
;
1660 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1662 if (h
->order
>= MAX_ORDER
)
1665 err
= strict_strtoul(buf
, 10, &input
);
1669 spin_lock(&hugetlb_lock
);
1670 h
->nr_overcommit_huge_pages
= input
;
1671 spin_unlock(&hugetlb_lock
);
1675 HSTATE_ATTR(nr_overcommit_hugepages
);
1677 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1678 struct kobj_attribute
*attr
, char *buf
)
1681 unsigned long free_huge_pages
;
1684 h
= kobj_to_hstate(kobj
, &nid
);
1685 if (nid
== NUMA_NO_NODE
)
1686 free_huge_pages
= h
->free_huge_pages
;
1688 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1690 return sprintf(buf
, "%lu\n", free_huge_pages
);
1692 HSTATE_ATTR_RO(free_hugepages
);
1694 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1695 struct kobj_attribute
*attr
, char *buf
)
1697 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1698 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1700 HSTATE_ATTR_RO(resv_hugepages
);
1702 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1703 struct kobj_attribute
*attr
, char *buf
)
1706 unsigned long surplus_huge_pages
;
1709 h
= kobj_to_hstate(kobj
, &nid
);
1710 if (nid
== NUMA_NO_NODE
)
1711 surplus_huge_pages
= h
->surplus_huge_pages
;
1713 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1715 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1717 HSTATE_ATTR_RO(surplus_hugepages
);
1719 static struct attribute
*hstate_attrs
[] = {
1720 &nr_hugepages_attr
.attr
,
1721 &nr_overcommit_hugepages_attr
.attr
,
1722 &free_hugepages_attr
.attr
,
1723 &resv_hugepages_attr
.attr
,
1724 &surplus_hugepages_attr
.attr
,
1726 &nr_hugepages_mempolicy_attr
.attr
,
1731 static struct attribute_group hstate_attr_group
= {
1732 .attrs
= hstate_attrs
,
1735 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1736 struct kobject
**hstate_kobjs
,
1737 struct attribute_group
*hstate_attr_group
)
1740 int hi
= hstate_index(h
);
1742 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1743 if (!hstate_kobjs
[hi
])
1746 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1748 kobject_put(hstate_kobjs
[hi
]);
1753 static void __init
hugetlb_sysfs_init(void)
1758 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1759 if (!hugepages_kobj
)
1762 for_each_hstate(h
) {
1763 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1764 hstate_kobjs
, &hstate_attr_group
);
1766 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1773 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1774 * with node devices in node_devices[] using a parallel array. The array
1775 * index of a node device or _hstate == node id.
1776 * This is here to avoid any static dependency of the node device driver, in
1777 * the base kernel, on the hugetlb module.
1779 struct node_hstate
{
1780 struct kobject
*hugepages_kobj
;
1781 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1783 struct node_hstate node_hstates
[MAX_NUMNODES
];
1786 * A subset of global hstate attributes for node devices
1788 static struct attribute
*per_node_hstate_attrs
[] = {
1789 &nr_hugepages_attr
.attr
,
1790 &free_hugepages_attr
.attr
,
1791 &surplus_hugepages_attr
.attr
,
1795 static struct attribute_group per_node_hstate_attr_group
= {
1796 .attrs
= per_node_hstate_attrs
,
1800 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1801 * Returns node id via non-NULL nidp.
1803 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1807 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1808 struct node_hstate
*nhs
= &node_hstates
[nid
];
1810 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1811 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1823 * Unregister hstate attributes from a single node device.
1824 * No-op if no hstate attributes attached.
1826 static void hugetlb_unregister_node(struct node
*node
)
1829 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1831 if (!nhs
->hugepages_kobj
)
1832 return; /* no hstate attributes */
1834 for_each_hstate(h
) {
1835 int idx
= hstate_index(h
);
1836 if (nhs
->hstate_kobjs
[idx
]) {
1837 kobject_put(nhs
->hstate_kobjs
[idx
]);
1838 nhs
->hstate_kobjs
[idx
] = NULL
;
1842 kobject_put(nhs
->hugepages_kobj
);
1843 nhs
->hugepages_kobj
= NULL
;
1847 * hugetlb module exit: unregister hstate attributes from node devices
1850 static void hugetlb_unregister_all_nodes(void)
1855 * disable node device registrations.
1857 register_hugetlbfs_with_node(NULL
, NULL
);
1860 * remove hstate attributes from any nodes that have them.
1862 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1863 hugetlb_unregister_node(node_devices
[nid
]);
1867 * Register hstate attributes for a single node device.
1868 * No-op if attributes already registered.
1870 static void hugetlb_register_node(struct node
*node
)
1873 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1876 if (nhs
->hugepages_kobj
)
1877 return; /* already allocated */
1879 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1881 if (!nhs
->hugepages_kobj
)
1884 for_each_hstate(h
) {
1885 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1887 &per_node_hstate_attr_group
);
1889 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1890 h
->name
, node
->dev
.id
);
1891 hugetlb_unregister_node(node
);
1898 * hugetlb init time: register hstate attributes for all registered node
1899 * devices of nodes that have memory. All on-line nodes should have
1900 * registered their associated device by this time.
1902 static void hugetlb_register_all_nodes(void)
1906 for_each_node_state(nid
, N_MEMORY
) {
1907 struct node
*node
= node_devices
[nid
];
1908 if (node
->dev
.id
== nid
)
1909 hugetlb_register_node(node
);
1913 * Let the node device driver know we're here so it can
1914 * [un]register hstate attributes on node hotplug.
1916 register_hugetlbfs_with_node(hugetlb_register_node
,
1917 hugetlb_unregister_node
);
1919 #else /* !CONFIG_NUMA */
1921 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1929 static void hugetlb_unregister_all_nodes(void) { }
1931 static void hugetlb_register_all_nodes(void) { }
1935 static void __exit
hugetlb_exit(void)
1939 hugetlb_unregister_all_nodes();
1941 for_each_hstate(h
) {
1942 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1945 kobject_put(hugepages_kobj
);
1947 module_exit(hugetlb_exit
);
1949 static int __init
hugetlb_init(void)
1951 /* Some platform decide whether they support huge pages at boot
1952 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1953 * there is no such support
1955 if (HPAGE_SHIFT
== 0)
1958 if (!size_to_hstate(default_hstate_size
)) {
1959 default_hstate_size
= HPAGE_SIZE
;
1960 if (!size_to_hstate(default_hstate_size
))
1961 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1963 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1964 if (default_hstate_max_huge_pages
)
1965 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1967 hugetlb_init_hstates();
1968 gather_bootmem_prealloc();
1971 hugetlb_sysfs_init();
1972 hugetlb_register_all_nodes();
1973 hugetlb_cgroup_file_init();
1977 module_init(hugetlb_init
);
1979 /* Should be called on processing a hugepagesz=... option */
1980 void __init
hugetlb_add_hstate(unsigned order
)
1985 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1986 pr_warning("hugepagesz= specified twice, ignoring\n");
1989 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1991 h
= &hstates
[hugetlb_max_hstate
++];
1993 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1994 h
->nr_huge_pages
= 0;
1995 h
->free_huge_pages
= 0;
1996 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1997 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1998 INIT_LIST_HEAD(&h
->hugepage_activelist
);
1999 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2000 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2001 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2002 huge_page_size(h
)/1024);
2007 static int __init
hugetlb_nrpages_setup(char *s
)
2010 static unsigned long *last_mhp
;
2013 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2014 * so this hugepages= parameter goes to the "default hstate".
2016 if (!hugetlb_max_hstate
)
2017 mhp
= &default_hstate_max_huge_pages
;
2019 mhp
= &parsed_hstate
->max_huge_pages
;
2021 if (mhp
== last_mhp
) {
2022 pr_warning("hugepages= specified twice without "
2023 "interleaving hugepagesz=, ignoring\n");
2027 if (sscanf(s
, "%lu", mhp
) <= 0)
2031 * Global state is always initialized later in hugetlb_init.
2032 * But we need to allocate >= MAX_ORDER hstates here early to still
2033 * use the bootmem allocator.
2035 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2036 hugetlb_hstate_alloc_pages(parsed_hstate
);
2042 __setup("hugepages=", hugetlb_nrpages_setup
);
2044 static int __init
hugetlb_default_setup(char *s
)
2046 default_hstate_size
= memparse(s
, &s
);
2049 __setup("default_hugepagesz=", hugetlb_default_setup
);
2051 static unsigned int cpuset_mems_nr(unsigned int *array
)
2054 unsigned int nr
= 0;
2056 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2062 #ifdef CONFIG_SYSCTL
2063 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2064 struct ctl_table
*table
, int write
,
2065 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2067 struct hstate
*h
= &default_hstate
;
2071 tmp
= h
->max_huge_pages
;
2073 if (write
&& h
->order
>= MAX_ORDER
)
2077 table
->maxlen
= sizeof(unsigned long);
2078 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2083 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2084 GFP_KERNEL
| __GFP_NORETRY
);
2085 if (!(obey_mempolicy
&&
2086 init_nodemask_of_mempolicy(nodes_allowed
))) {
2087 NODEMASK_FREE(nodes_allowed
);
2088 nodes_allowed
= &node_states
[N_MEMORY
];
2090 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2092 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2093 NODEMASK_FREE(nodes_allowed
);
2099 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2100 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2103 return hugetlb_sysctl_handler_common(false, table
, write
,
2104 buffer
, length
, ppos
);
2108 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2109 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2111 return hugetlb_sysctl_handler_common(true, table
, write
,
2112 buffer
, length
, ppos
);
2114 #endif /* CONFIG_NUMA */
2116 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2117 void __user
*buffer
,
2118 size_t *length
, loff_t
*ppos
)
2120 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2121 if (hugepages_treat_as_movable
)
2122 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2124 htlb_alloc_mask
= GFP_HIGHUSER
;
2128 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2129 void __user
*buffer
,
2130 size_t *length
, loff_t
*ppos
)
2132 struct hstate
*h
= &default_hstate
;
2136 tmp
= h
->nr_overcommit_huge_pages
;
2138 if (write
&& h
->order
>= MAX_ORDER
)
2142 table
->maxlen
= sizeof(unsigned long);
2143 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2148 spin_lock(&hugetlb_lock
);
2149 h
->nr_overcommit_huge_pages
= tmp
;
2150 spin_unlock(&hugetlb_lock
);
2156 #endif /* CONFIG_SYSCTL */
2158 void hugetlb_report_meminfo(struct seq_file
*m
)
2160 struct hstate
*h
= &default_hstate
;
2162 "HugePages_Total: %5lu\n"
2163 "HugePages_Free: %5lu\n"
2164 "HugePages_Rsvd: %5lu\n"
2165 "HugePages_Surp: %5lu\n"
2166 "Hugepagesize: %8lu kB\n",
2170 h
->surplus_huge_pages
,
2171 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2174 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2176 struct hstate
*h
= &default_hstate
;
2178 "Node %d HugePages_Total: %5u\n"
2179 "Node %d HugePages_Free: %5u\n"
2180 "Node %d HugePages_Surp: %5u\n",
2181 nid
, h
->nr_huge_pages_node
[nid
],
2182 nid
, h
->free_huge_pages_node
[nid
],
2183 nid
, h
->surplus_huge_pages_node
[nid
]);
2186 void hugetlb_show_meminfo(void)
2191 for_each_node_state(nid
, N_MEMORY
)
2193 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2195 h
->nr_huge_pages_node
[nid
],
2196 h
->free_huge_pages_node
[nid
],
2197 h
->surplus_huge_pages_node
[nid
],
2198 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2201 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2202 unsigned long hugetlb_total_pages(void)
2205 unsigned long nr_total_pages
= 0;
2208 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2209 return nr_total_pages
;
2212 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2216 spin_lock(&hugetlb_lock
);
2218 * When cpuset is configured, it breaks the strict hugetlb page
2219 * reservation as the accounting is done on a global variable. Such
2220 * reservation is completely rubbish in the presence of cpuset because
2221 * the reservation is not checked against page availability for the
2222 * current cpuset. Application can still potentially OOM'ed by kernel
2223 * with lack of free htlb page in cpuset that the task is in.
2224 * Attempt to enforce strict accounting with cpuset is almost
2225 * impossible (or too ugly) because cpuset is too fluid that
2226 * task or memory node can be dynamically moved between cpusets.
2228 * The change of semantics for shared hugetlb mapping with cpuset is
2229 * undesirable. However, in order to preserve some of the semantics,
2230 * we fall back to check against current free page availability as
2231 * a best attempt and hopefully to minimize the impact of changing
2232 * semantics that cpuset has.
2235 if (gather_surplus_pages(h
, delta
) < 0)
2238 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2239 return_unused_surplus_pages(h
, delta
);
2246 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2249 spin_unlock(&hugetlb_lock
);
2253 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2255 struct resv_map
*reservations
= vma_resv_map(vma
);
2258 * This new VMA should share its siblings reservation map if present.
2259 * The VMA will only ever have a valid reservation map pointer where
2260 * it is being copied for another still existing VMA. As that VMA
2261 * has a reference to the reservation map it cannot disappear until
2262 * after this open call completes. It is therefore safe to take a
2263 * new reference here without additional locking.
2266 kref_get(&reservations
->refs
);
2269 static void resv_map_put(struct vm_area_struct
*vma
)
2271 struct resv_map
*reservations
= vma_resv_map(vma
);
2275 kref_put(&reservations
->refs
, resv_map_release
);
2278 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2280 struct hstate
*h
= hstate_vma(vma
);
2281 struct resv_map
*reservations
= vma_resv_map(vma
);
2282 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2283 unsigned long reserve
;
2284 unsigned long start
;
2288 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2289 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2291 reserve
= (end
- start
) -
2292 region_count(&reservations
->regions
, start
, end
);
2297 hugetlb_acct_memory(h
, -reserve
);
2298 hugepage_subpool_put_pages(spool
, reserve
);
2304 * We cannot handle pagefaults against hugetlb pages at all. They cause
2305 * handle_mm_fault() to try to instantiate regular-sized pages in the
2306 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2309 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2315 const struct vm_operations_struct hugetlb_vm_ops
= {
2316 .fault
= hugetlb_vm_op_fault
,
2317 .open
= hugetlb_vm_op_open
,
2318 .close
= hugetlb_vm_op_close
,
2321 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2327 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2328 vma
->vm_page_prot
)));
2330 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2331 vma
->vm_page_prot
));
2333 entry
= pte_mkyoung(entry
);
2334 entry
= pte_mkhuge(entry
);
2335 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2340 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2341 unsigned long address
, pte_t
*ptep
)
2345 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2346 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2347 update_mmu_cache(vma
, address
, ptep
);
2350 static int is_hugetlb_entry_migration(pte_t pte
)
2354 if (huge_pte_none(pte
) || pte_present(pte
))
2356 swp
= pte_to_swp_entry(pte
);
2357 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2363 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2367 if (huge_pte_none(pte
) || pte_present(pte
))
2369 swp
= pte_to_swp_entry(pte
);
2370 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2376 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2377 struct vm_area_struct
*vma
)
2379 pte_t
*src_pte
, *dst_pte
, entry
;
2380 struct page
*ptepage
;
2383 struct hstate
*h
= hstate_vma(vma
);
2384 unsigned long sz
= huge_page_size(h
);
2386 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2388 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2389 src_pte
= huge_pte_offset(src
, addr
);
2392 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2396 /* If the pagetables are shared don't copy or take references */
2397 if (dst_pte
== src_pte
)
2400 spin_lock(&dst
->page_table_lock
);
2401 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2402 entry
= huge_ptep_get(src_pte
);
2403 if (huge_pte_none(entry
)) { /* skip none entry */
2405 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
2406 is_hugetlb_entry_hwpoisoned(entry
))) {
2407 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
2409 if (is_write_migration_entry(swp_entry
) && cow
) {
2411 * COW mappings require pages in both
2412 * parent and child to be set to read.
2414 make_migration_entry_read(&swp_entry
);
2415 entry
= swp_entry_to_pte(swp_entry
);
2416 set_huge_pte_at(src
, addr
, src_pte
, entry
);
2418 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2421 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2422 entry
= huge_ptep_get(src_pte
);
2423 ptepage
= pte_page(entry
);
2425 page_dup_rmap(ptepage
);
2426 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2428 spin_unlock(&src
->page_table_lock
);
2429 spin_unlock(&dst
->page_table_lock
);
2437 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2438 unsigned long start
, unsigned long end
,
2439 struct page
*ref_page
)
2441 int force_flush
= 0;
2442 struct mm_struct
*mm
= vma
->vm_mm
;
2443 unsigned long address
;
2447 struct hstate
*h
= hstate_vma(vma
);
2448 unsigned long sz
= huge_page_size(h
);
2449 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2450 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2452 WARN_ON(!is_vm_hugetlb_page(vma
));
2453 BUG_ON(start
& ~huge_page_mask(h
));
2454 BUG_ON(end
& ~huge_page_mask(h
));
2456 tlb_start_vma(tlb
, vma
);
2457 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2459 spin_lock(&mm
->page_table_lock
);
2460 for (address
= start
; address
< end
; address
+= sz
) {
2461 ptep
= huge_pte_offset(mm
, address
);
2465 if (huge_pmd_unshare(mm
, &address
, ptep
))
2468 pte
= huge_ptep_get(ptep
);
2469 if (huge_pte_none(pte
))
2473 * Migrating hugepage or HWPoisoned hugepage is already
2474 * unmapped and its refcount is dropped, so just clear pte here.
2476 if (unlikely(!pte_present(pte
))) {
2477 huge_pte_clear(mm
, address
, ptep
);
2481 page
= pte_page(pte
);
2483 * If a reference page is supplied, it is because a specific
2484 * page is being unmapped, not a range. Ensure the page we
2485 * are about to unmap is the actual page of interest.
2488 if (page
!= ref_page
)
2492 * Mark the VMA as having unmapped its page so that
2493 * future faults in this VMA will fail rather than
2494 * looking like data was lost
2496 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2499 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2500 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2501 if (huge_pte_dirty(pte
))
2502 set_page_dirty(page
);
2504 page_remove_rmap(page
);
2505 force_flush
= !__tlb_remove_page(tlb
, page
);
2508 /* Bail out after unmapping reference page if supplied */
2512 spin_unlock(&mm
->page_table_lock
);
2514 * mmu_gather ran out of room to batch pages, we break out of
2515 * the PTE lock to avoid doing the potential expensive TLB invalidate
2516 * and page-free while holding it.
2521 if (address
< end
&& !ref_page
)
2524 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2525 tlb_end_vma(tlb
, vma
);
2528 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2529 struct vm_area_struct
*vma
, unsigned long start
,
2530 unsigned long end
, struct page
*ref_page
)
2532 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2535 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2536 * test will fail on a vma being torn down, and not grab a page table
2537 * on its way out. We're lucky that the flag has such an appropriate
2538 * name, and can in fact be safely cleared here. We could clear it
2539 * before the __unmap_hugepage_range above, but all that's necessary
2540 * is to clear it before releasing the i_mmap_mutex. This works
2541 * because in the context this is called, the VMA is about to be
2542 * destroyed and the i_mmap_mutex is held.
2544 vma
->vm_flags
&= ~VM_MAYSHARE
;
2547 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2548 unsigned long end
, struct page
*ref_page
)
2550 struct mm_struct
*mm
;
2551 struct mmu_gather tlb
;
2555 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2556 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2557 tlb_finish_mmu(&tlb
, start
, end
);
2561 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2562 * mappping it owns the reserve page for. The intention is to unmap the page
2563 * from other VMAs and let the children be SIGKILLed if they are faulting the
2566 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2567 struct page
*page
, unsigned long address
)
2569 struct hstate
*h
= hstate_vma(vma
);
2570 struct vm_area_struct
*iter_vma
;
2571 struct address_space
*mapping
;
2575 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2576 * from page cache lookup which is in HPAGE_SIZE units.
2578 address
= address
& huge_page_mask(h
);
2579 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2581 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2584 * Take the mapping lock for the duration of the table walk. As
2585 * this mapping should be shared between all the VMAs,
2586 * __unmap_hugepage_range() is called as the lock is already held
2588 mutex_lock(&mapping
->i_mmap_mutex
);
2589 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2590 /* Do not unmap the current VMA */
2591 if (iter_vma
== vma
)
2595 * Shared VMAs have their own reserves and do not affect
2596 * MAP_PRIVATE accounting but it is possible that a shared
2597 * VMA is using the same page so check and skip such VMAs.
2599 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
2603 * Unmap the page from other VMAs without their own reserves.
2604 * They get marked to be SIGKILLed if they fault in these
2605 * areas. This is because a future no-page fault on this VMA
2606 * could insert a zeroed page instead of the data existing
2607 * from the time of fork. This would look like data corruption
2609 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2610 unmap_hugepage_range(iter_vma
, address
,
2611 address
+ huge_page_size(h
), page
);
2613 mutex_unlock(&mapping
->i_mmap_mutex
);
2619 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2620 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2621 * cannot race with other handlers or page migration.
2622 * Keep the pte_same checks anyway to make transition from the mutex easier.
2624 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2625 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2626 struct page
*pagecache_page
)
2628 struct hstate
*h
= hstate_vma(vma
);
2629 struct page
*old_page
, *new_page
;
2631 int outside_reserve
= 0;
2632 unsigned long mmun_start
; /* For mmu_notifiers */
2633 unsigned long mmun_end
; /* For mmu_notifiers */
2635 old_page
= pte_page(pte
);
2638 /* If no-one else is actually using this page, avoid the copy
2639 * and just make the page writable */
2640 avoidcopy
= (page_mapcount(old_page
) == 1);
2642 if (PageAnon(old_page
))
2643 page_move_anon_rmap(old_page
, vma
, address
);
2644 set_huge_ptep_writable(vma
, address
, ptep
);
2649 * If the process that created a MAP_PRIVATE mapping is about to
2650 * perform a COW due to a shared page count, attempt to satisfy
2651 * the allocation without using the existing reserves. The pagecache
2652 * page is used to determine if the reserve at this address was
2653 * consumed or not. If reserves were used, a partial faulted mapping
2654 * at the time of fork() could consume its reserves on COW instead
2655 * of the full address range.
2657 if (!(vma
->vm_flags
& VM_MAYSHARE
) &&
2658 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2659 old_page
!= pagecache_page
)
2660 outside_reserve
= 1;
2662 page_cache_get(old_page
);
2664 /* Drop page_table_lock as buddy allocator may be called */
2665 spin_unlock(&mm
->page_table_lock
);
2666 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2668 if (IS_ERR(new_page
)) {
2669 long err
= PTR_ERR(new_page
);
2670 page_cache_release(old_page
);
2673 * If a process owning a MAP_PRIVATE mapping fails to COW,
2674 * it is due to references held by a child and an insufficient
2675 * huge page pool. To guarantee the original mappers
2676 * reliability, unmap the page from child processes. The child
2677 * may get SIGKILLed if it later faults.
2679 if (outside_reserve
) {
2680 BUG_ON(huge_pte_none(pte
));
2681 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2682 BUG_ON(huge_pte_none(pte
));
2683 spin_lock(&mm
->page_table_lock
);
2684 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2685 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2686 goto retry_avoidcopy
;
2688 * race occurs while re-acquiring page_table_lock, and
2696 /* Caller expects lock to be held */
2697 spin_lock(&mm
->page_table_lock
);
2699 return VM_FAULT_OOM
;
2701 return VM_FAULT_SIGBUS
;
2705 * When the original hugepage is shared one, it does not have
2706 * anon_vma prepared.
2708 if (unlikely(anon_vma_prepare(vma
))) {
2709 page_cache_release(new_page
);
2710 page_cache_release(old_page
);
2711 /* Caller expects lock to be held */
2712 spin_lock(&mm
->page_table_lock
);
2713 return VM_FAULT_OOM
;
2716 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2717 pages_per_huge_page(h
));
2718 __SetPageUptodate(new_page
);
2720 mmun_start
= address
& huge_page_mask(h
);
2721 mmun_end
= mmun_start
+ huge_page_size(h
);
2722 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2724 * Retake the page_table_lock to check for racing updates
2725 * before the page tables are altered
2727 spin_lock(&mm
->page_table_lock
);
2728 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2729 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2731 huge_ptep_clear_flush(vma
, address
, ptep
);
2732 set_huge_pte_at(mm
, address
, ptep
,
2733 make_huge_pte(vma
, new_page
, 1));
2734 page_remove_rmap(old_page
);
2735 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2736 /* Make the old page be freed below */
2737 new_page
= old_page
;
2739 spin_unlock(&mm
->page_table_lock
);
2740 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2741 /* Caller expects lock to be held */
2742 spin_lock(&mm
->page_table_lock
);
2743 page_cache_release(new_page
);
2744 page_cache_release(old_page
);
2748 /* Return the pagecache page at a given address within a VMA */
2749 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2750 struct vm_area_struct
*vma
, unsigned long address
)
2752 struct address_space
*mapping
;
2755 mapping
= vma
->vm_file
->f_mapping
;
2756 idx
= vma_hugecache_offset(h
, vma
, address
);
2758 return find_lock_page(mapping
, idx
);
2762 * Return whether there is a pagecache page to back given address within VMA.
2763 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2765 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2766 struct vm_area_struct
*vma
, unsigned long address
)
2768 struct address_space
*mapping
;
2772 mapping
= vma
->vm_file
->f_mapping
;
2773 idx
= vma_hugecache_offset(h
, vma
, address
);
2775 page
= find_get_page(mapping
, idx
);
2778 return page
!= NULL
;
2781 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2782 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2784 struct hstate
*h
= hstate_vma(vma
);
2785 int ret
= VM_FAULT_SIGBUS
;
2790 struct address_space
*mapping
;
2794 * Currently, we are forced to kill the process in the event the
2795 * original mapper has unmapped pages from the child due to a failed
2796 * COW. Warn that such a situation has occurred as it may not be obvious
2798 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2799 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2804 mapping
= vma
->vm_file
->f_mapping
;
2805 idx
= vma_hugecache_offset(h
, vma
, address
);
2808 * Use page lock to guard against racing truncation
2809 * before we get page_table_lock.
2812 page
= find_lock_page(mapping
, idx
);
2814 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2817 page
= alloc_huge_page(vma
, address
, 0);
2819 ret
= PTR_ERR(page
);
2823 ret
= VM_FAULT_SIGBUS
;
2826 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2827 __SetPageUptodate(page
);
2829 if (vma
->vm_flags
& VM_MAYSHARE
) {
2831 struct inode
*inode
= mapping
->host
;
2833 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2841 spin_lock(&inode
->i_lock
);
2842 inode
->i_blocks
+= blocks_per_huge_page(h
);
2843 spin_unlock(&inode
->i_lock
);
2846 if (unlikely(anon_vma_prepare(vma
))) {
2848 goto backout_unlocked
;
2854 * If memory error occurs between mmap() and fault, some process
2855 * don't have hwpoisoned swap entry for errored virtual address.
2856 * So we need to block hugepage fault by PG_hwpoison bit check.
2858 if (unlikely(PageHWPoison(page
))) {
2859 ret
= VM_FAULT_HWPOISON
|
2860 VM_FAULT_SET_HINDEX(hstate_index(h
));
2861 goto backout_unlocked
;
2866 * If we are going to COW a private mapping later, we examine the
2867 * pending reservations for this page now. This will ensure that
2868 * any allocations necessary to record that reservation occur outside
2871 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2872 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2874 goto backout_unlocked
;
2877 spin_lock(&mm
->page_table_lock
);
2878 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2883 if (!huge_pte_none(huge_ptep_get(ptep
)))
2887 hugepage_add_new_anon_rmap(page
, vma
, address
);
2889 page_dup_rmap(page
);
2890 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2891 && (vma
->vm_flags
& VM_SHARED
)));
2892 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2894 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2895 /* Optimization, do the COW without a second fault */
2896 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2899 spin_unlock(&mm
->page_table_lock
);
2905 spin_unlock(&mm
->page_table_lock
);
2912 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2913 unsigned long address
, unsigned int flags
)
2918 struct page
*page
= NULL
;
2919 struct page
*pagecache_page
= NULL
;
2920 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2921 struct hstate
*h
= hstate_vma(vma
);
2923 address
&= huge_page_mask(h
);
2925 ptep
= huge_pte_offset(mm
, address
);
2927 entry
= huge_ptep_get(ptep
);
2928 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2929 migration_entry_wait_huge(mm
, ptep
);
2931 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2932 return VM_FAULT_HWPOISON_LARGE
|
2933 VM_FAULT_SET_HINDEX(hstate_index(h
));
2936 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2938 return VM_FAULT_OOM
;
2941 * Serialize hugepage allocation and instantiation, so that we don't
2942 * get spurious allocation failures if two CPUs race to instantiate
2943 * the same page in the page cache.
2945 mutex_lock(&hugetlb_instantiation_mutex
);
2946 entry
= huge_ptep_get(ptep
);
2947 if (huge_pte_none(entry
)) {
2948 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2955 * If we are going to COW the mapping later, we examine the pending
2956 * reservations for this page now. This will ensure that any
2957 * allocations necessary to record that reservation occur outside the
2958 * spinlock. For private mappings, we also lookup the pagecache
2959 * page now as it is used to determine if a reservation has been
2962 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2963 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2968 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2969 pagecache_page
= hugetlbfs_pagecache_page(h
,
2974 * hugetlb_cow() requires page locks of pte_page(entry) and
2975 * pagecache_page, so here we need take the former one
2976 * when page != pagecache_page or !pagecache_page.
2977 * Note that locking order is always pagecache_page -> page,
2978 * so no worry about deadlock.
2980 page
= pte_page(entry
);
2982 if (page
!= pagecache_page
)
2985 spin_lock(&mm
->page_table_lock
);
2986 /* Check for a racing update before calling hugetlb_cow */
2987 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2988 goto out_page_table_lock
;
2991 if (flags
& FAULT_FLAG_WRITE
) {
2992 if (!huge_pte_write(entry
)) {
2993 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2995 goto out_page_table_lock
;
2997 entry
= huge_pte_mkdirty(entry
);
2999 entry
= pte_mkyoung(entry
);
3000 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3001 flags
& FAULT_FLAG_WRITE
))
3002 update_mmu_cache(vma
, address
, ptep
);
3004 out_page_table_lock
:
3005 spin_unlock(&mm
->page_table_lock
);
3007 if (pagecache_page
) {
3008 unlock_page(pagecache_page
);
3009 put_page(pagecache_page
);
3011 if (page
!= pagecache_page
)
3016 mutex_unlock(&hugetlb_instantiation_mutex
);
3021 /* Can be overriden by architectures */
3022 __attribute__((weak
)) struct page
*
3023 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3024 pud_t
*pud
, int write
)
3030 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3031 struct page
**pages
, struct vm_area_struct
**vmas
,
3032 unsigned long *position
, unsigned long *nr_pages
,
3033 long i
, unsigned int flags
)
3035 unsigned long pfn_offset
;
3036 unsigned long vaddr
= *position
;
3037 unsigned long remainder
= *nr_pages
;
3038 struct hstate
*h
= hstate_vma(vma
);
3040 spin_lock(&mm
->page_table_lock
);
3041 while (vaddr
< vma
->vm_end
&& remainder
) {
3047 * Some archs (sparc64, sh*) have multiple pte_ts to
3048 * each hugepage. We have to make sure we get the
3049 * first, for the page indexing below to work.
3051 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3052 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3055 * When coredumping, it suits get_dump_page if we just return
3056 * an error where there's an empty slot with no huge pagecache
3057 * to back it. This way, we avoid allocating a hugepage, and
3058 * the sparse dumpfile avoids allocating disk blocks, but its
3059 * huge holes still show up with zeroes where they need to be.
3061 if (absent
&& (flags
& FOLL_DUMP
) &&
3062 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3068 * We need call hugetlb_fault for both hugepages under migration
3069 * (in which case hugetlb_fault waits for the migration,) and
3070 * hwpoisoned hugepages (in which case we need to prevent the
3071 * caller from accessing to them.) In order to do this, we use
3072 * here is_swap_pte instead of is_hugetlb_entry_migration and
3073 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3074 * both cases, and because we can't follow correct pages
3075 * directly from any kind of swap entries.
3077 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3078 ((flags
& FOLL_WRITE
) &&
3079 !huge_pte_write(huge_ptep_get(pte
)))) {
3082 spin_unlock(&mm
->page_table_lock
);
3083 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3084 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3085 spin_lock(&mm
->page_table_lock
);
3086 if (!(ret
& VM_FAULT_ERROR
))
3093 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3094 page
= pte_page(huge_ptep_get(pte
));
3097 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3108 if (vaddr
< vma
->vm_end
&& remainder
&&
3109 pfn_offset
< pages_per_huge_page(h
)) {
3111 * We use pfn_offset to avoid touching the pageframes
3112 * of this compound page.
3117 spin_unlock(&mm
->page_table_lock
);
3118 *nr_pages
= remainder
;
3121 return i
? i
: -EFAULT
;
3124 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3125 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3127 struct mm_struct
*mm
= vma
->vm_mm
;
3128 unsigned long start
= address
;
3131 struct hstate
*h
= hstate_vma(vma
);
3132 unsigned long pages
= 0;
3134 BUG_ON(address
>= end
);
3135 flush_cache_range(vma
, address
, end
);
3137 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3138 spin_lock(&mm
->page_table_lock
);
3139 for (; address
< end
; address
+= huge_page_size(h
)) {
3140 ptep
= huge_pte_offset(mm
, address
);
3143 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3147 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3148 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3149 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3150 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3151 set_huge_pte_at(mm
, address
, ptep
, pte
);
3155 spin_unlock(&mm
->page_table_lock
);
3157 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3158 * may have cleared our pud entry and done put_page on the page table:
3159 * once we release i_mmap_mutex, another task can do the final put_page
3160 * and that page table be reused and filled with junk.
3162 flush_tlb_range(vma
, start
, end
);
3163 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3165 return pages
<< h
->order
;
3168 int hugetlb_reserve_pages(struct inode
*inode
,
3170 struct vm_area_struct
*vma
,
3171 vm_flags_t vm_flags
)
3174 struct hstate
*h
= hstate_inode(inode
);
3175 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3178 * Only apply hugepage reservation if asked. At fault time, an
3179 * attempt will be made for VM_NORESERVE to allocate a page
3180 * without using reserves
3182 if (vm_flags
& VM_NORESERVE
)
3186 * Shared mappings base their reservation on the number of pages that
3187 * are already allocated on behalf of the file. Private mappings need
3188 * to reserve the full area even if read-only as mprotect() may be
3189 * called to make the mapping read-write. Assume !vma is a shm mapping
3191 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3192 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3194 struct resv_map
*resv_map
= resv_map_alloc();
3200 set_vma_resv_map(vma
, resv_map
);
3201 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3209 /* There must be enough pages in the subpool for the mapping */
3210 if (hugepage_subpool_get_pages(spool
, chg
)) {
3216 * Check enough hugepages are available for the reservation.
3217 * Hand the pages back to the subpool if there are not
3219 ret
= hugetlb_acct_memory(h
, chg
);
3221 hugepage_subpool_put_pages(spool
, chg
);
3226 * Account for the reservations made. Shared mappings record regions
3227 * that have reservations as they are shared by multiple VMAs.
3228 * When the last VMA disappears, the region map says how much
3229 * the reservation was and the page cache tells how much of
3230 * the reservation was consumed. Private mappings are per-VMA and
3231 * only the consumed reservations are tracked. When the VMA
3232 * disappears, the original reservation is the VMA size and the
3233 * consumed reservations are stored in the map. Hence, nothing
3234 * else has to be done for private mappings here
3236 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3237 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3245 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3247 struct hstate
*h
= hstate_inode(inode
);
3248 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3249 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3251 spin_lock(&inode
->i_lock
);
3252 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3253 spin_unlock(&inode
->i_lock
);
3255 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3256 hugetlb_acct_memory(h
, -(chg
- freed
));
3259 #ifdef CONFIG_MEMORY_FAILURE
3261 /* Should be called in hugetlb_lock */
3262 static int is_hugepage_on_freelist(struct page
*hpage
)
3266 struct hstate
*h
= page_hstate(hpage
);
3267 int nid
= page_to_nid(hpage
);
3269 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3276 * This function is called from memory failure code.
3277 * Assume the caller holds page lock of the head page.
3279 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3281 struct hstate
*h
= page_hstate(hpage
);
3282 int nid
= page_to_nid(hpage
);
3285 spin_lock(&hugetlb_lock
);
3286 if (is_hugepage_on_freelist(hpage
)) {
3288 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3289 * but dangling hpage->lru can trigger list-debug warnings
3290 * (this happens when we call unpoison_memory() on it),
3291 * so let it point to itself with list_del_init().
3293 list_del_init(&hpage
->lru
);
3294 set_page_refcounted(hpage
);
3295 h
->free_huge_pages
--;
3296 h
->free_huge_pages_node
[nid
]--;
3299 spin_unlock(&hugetlb_lock
);