2 * Generic hugetlb support.
3 * (C) William Irwin, 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>
26 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/hugetlb_cgroup.h>
32 #include <linux/node.h>
33 #include <linux/hugetlb_cgroup.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(vma
->vm_file
->f_dentry
->d_inode
);
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 if (list_empty(&h
->hugepage_freelists
[nid
]))
523 page
= list_entry(h
->hugepage_freelists
[nid
].next
, struct page
, lru
);
524 list_move(&page
->lru
, &h
->hugepage_activelist
);
525 set_page_refcounted(page
);
526 h
->free_huge_pages
--;
527 h
->free_huge_pages_node
[nid
]--;
531 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
532 struct vm_area_struct
*vma
,
533 unsigned long address
, int avoid_reserve
)
535 struct page
*page
= NULL
;
536 struct mempolicy
*mpol
;
537 nodemask_t
*nodemask
;
538 struct zonelist
*zonelist
;
541 unsigned int cpuset_mems_cookie
;
544 cpuset_mems_cookie
= get_mems_allowed();
545 zonelist
= huge_zonelist(vma
, address
,
546 htlb_alloc_mask
, &mpol
, &nodemask
);
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
552 if (!vma_has_reserves(vma
) &&
553 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
556 /* If reserves cannot be used, ensure enough pages are in the pool */
557 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
560 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
561 MAX_NR_ZONES
- 1, nodemask
) {
562 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
563 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
566 decrement_hugepage_resv_vma(h
, vma
);
573 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
582 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
586 VM_BUG_ON(h
->order
>= MAX_ORDER
);
589 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
590 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
591 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
592 1 << PG_referenced
| 1 << PG_dirty
|
593 1 << PG_active
| 1 << PG_reserved
|
594 1 << PG_private
| 1 << PG_writeback
);
596 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
597 set_compound_page_dtor(page
, NULL
);
598 set_page_refcounted(page
);
599 arch_release_hugepage(page
);
600 __free_pages(page
, huge_page_order(h
));
603 struct hstate
*size_to_hstate(unsigned long size
)
608 if (huge_page_size(h
) == size
)
614 static void free_huge_page(struct page
*page
)
617 * Can't pass hstate in here because it is called from the
618 * compound page destructor.
620 struct hstate
*h
= page_hstate(page
);
621 int nid
= page_to_nid(page
);
622 struct hugepage_subpool
*spool
=
623 (struct hugepage_subpool
*)page_private(page
);
625 set_page_private(page
, 0);
626 page
->mapping
= NULL
;
627 BUG_ON(page_count(page
));
628 BUG_ON(page_mapcount(page
));
630 spin_lock(&hugetlb_lock
);
631 hugetlb_cgroup_uncharge_page(hstate_index(h
),
632 pages_per_huge_page(h
), page
);
633 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
634 /* remove the page from active list */
635 list_del(&page
->lru
);
636 update_and_free_page(h
, page
);
637 h
->surplus_huge_pages
--;
638 h
->surplus_huge_pages_node
[nid
]--;
640 arch_clear_hugepage_flags(page
);
641 enqueue_huge_page(h
, page
);
643 spin_unlock(&hugetlb_lock
);
644 hugepage_subpool_put_pages(spool
, 1);
647 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
649 INIT_LIST_HEAD(&page
->lru
);
650 set_compound_page_dtor(page
, free_huge_page
);
651 spin_lock(&hugetlb_lock
);
652 set_hugetlb_cgroup(page
, NULL
);
654 h
->nr_huge_pages_node
[nid
]++;
655 spin_unlock(&hugetlb_lock
);
656 put_page(page
); /* free it into the hugepage allocator */
659 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
662 int nr_pages
= 1 << order
;
663 struct page
*p
= page
+ 1;
665 /* we rely on prep_new_huge_page to set the destructor */
666 set_compound_order(page
, order
);
668 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
670 set_page_count(p
, 0);
671 p
->first_page
= page
;
675 int PageHuge(struct page
*page
)
677 compound_page_dtor
*dtor
;
679 if (!PageCompound(page
))
682 page
= compound_head(page
);
683 dtor
= get_compound_page_dtor(page
);
685 return dtor
== free_huge_page
;
687 EXPORT_SYMBOL_GPL(PageHuge
);
689 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
693 if (h
->order
>= MAX_ORDER
)
696 page
= alloc_pages_exact_node(nid
,
697 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
698 __GFP_REPEAT
|__GFP_NOWARN
,
701 if (arch_prepare_hugepage(page
)) {
702 __free_pages(page
, huge_page_order(h
));
705 prep_new_huge_page(h
, page
, nid
);
712 * common helper functions for hstate_next_node_to_{alloc|free}.
713 * We may have allocated or freed a huge page based on a different
714 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
715 * be outside of *nodes_allowed. Ensure that we use an allowed
716 * node for alloc or free.
718 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
720 nid
= next_node(nid
, *nodes_allowed
);
721 if (nid
== MAX_NUMNODES
)
722 nid
= first_node(*nodes_allowed
);
723 VM_BUG_ON(nid
>= MAX_NUMNODES
);
728 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
730 if (!node_isset(nid
, *nodes_allowed
))
731 nid
= next_node_allowed(nid
, nodes_allowed
);
736 * returns the previously saved node ["this node"] from which to
737 * allocate a persistent huge page for the pool and advance the
738 * next node from which to allocate, handling wrap at end of node
741 static int hstate_next_node_to_alloc(struct hstate
*h
,
742 nodemask_t
*nodes_allowed
)
746 VM_BUG_ON(!nodes_allowed
);
748 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
749 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
754 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
761 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
762 next_nid
= start_nid
;
765 page
= alloc_fresh_huge_page_node(h
, next_nid
);
770 next_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
771 } while (next_nid
!= start_nid
);
774 count_vm_event(HTLB_BUDDY_PGALLOC
);
776 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
782 * helper for free_pool_huge_page() - return the previously saved
783 * node ["this node"] from which to free a huge page. Advance the
784 * next node id whether or not we find a free huge page to free so
785 * that the next attempt to free addresses the next node.
787 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
791 VM_BUG_ON(!nodes_allowed
);
793 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
794 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
800 * Free huge page from pool from next node to free.
801 * Attempt to keep persistent huge pages more or less
802 * balanced over allowed nodes.
803 * Called with hugetlb_lock locked.
805 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
812 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
813 next_nid
= start_nid
;
817 * If we're returning unused surplus pages, only examine
818 * nodes with surplus pages.
820 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[next_nid
]) &&
821 !list_empty(&h
->hugepage_freelists
[next_nid
])) {
823 list_entry(h
->hugepage_freelists
[next_nid
].next
,
825 list_del(&page
->lru
);
826 h
->free_huge_pages
--;
827 h
->free_huge_pages_node
[next_nid
]--;
829 h
->surplus_huge_pages
--;
830 h
->surplus_huge_pages_node
[next_nid
]--;
832 update_and_free_page(h
, page
);
836 next_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
837 } while (next_nid
!= start_nid
);
842 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
847 if (h
->order
>= MAX_ORDER
)
851 * Assume we will successfully allocate the surplus page to
852 * prevent racing processes from causing the surplus to exceed
855 * This however introduces a different race, where a process B
856 * tries to grow the static hugepage pool while alloc_pages() is
857 * called by process A. B will only examine the per-node
858 * counters in determining if surplus huge pages can be
859 * converted to normal huge pages in adjust_pool_surplus(). A
860 * won't be able to increment the per-node counter, until the
861 * lock is dropped by B, but B doesn't drop hugetlb_lock until
862 * no more huge pages can be converted from surplus to normal
863 * state (and doesn't try to convert again). Thus, we have a
864 * case where a surplus huge page exists, the pool is grown, and
865 * the surplus huge page still exists after, even though it
866 * should just have been converted to a normal huge page. This
867 * does not leak memory, though, as the hugepage will be freed
868 * once it is out of use. It also does not allow the counters to
869 * go out of whack in adjust_pool_surplus() as we don't modify
870 * the node values until we've gotten the hugepage and only the
871 * per-node value is checked there.
873 spin_lock(&hugetlb_lock
);
874 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
875 spin_unlock(&hugetlb_lock
);
879 h
->surplus_huge_pages
++;
881 spin_unlock(&hugetlb_lock
);
883 if (nid
== NUMA_NO_NODE
)
884 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
885 __GFP_REPEAT
|__GFP_NOWARN
,
888 page
= alloc_pages_exact_node(nid
,
889 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
890 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
892 if (page
&& arch_prepare_hugepage(page
)) {
893 __free_pages(page
, huge_page_order(h
));
897 spin_lock(&hugetlb_lock
);
899 INIT_LIST_HEAD(&page
->lru
);
900 r_nid
= page_to_nid(page
);
901 set_compound_page_dtor(page
, free_huge_page
);
902 set_hugetlb_cgroup(page
, NULL
);
904 * We incremented the global counters already
906 h
->nr_huge_pages_node
[r_nid
]++;
907 h
->surplus_huge_pages_node
[r_nid
]++;
908 __count_vm_event(HTLB_BUDDY_PGALLOC
);
911 h
->surplus_huge_pages
--;
912 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
914 spin_unlock(&hugetlb_lock
);
920 * This allocation function is useful in the context where vma is irrelevant.
921 * E.g. soft-offlining uses this function because it only cares physical
922 * address of error page.
924 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
928 spin_lock(&hugetlb_lock
);
929 page
= dequeue_huge_page_node(h
, nid
);
930 spin_unlock(&hugetlb_lock
);
933 page
= alloc_buddy_huge_page(h
, nid
);
939 * Increase the hugetlb pool such that it can accommodate a reservation
942 static int gather_surplus_pages(struct hstate
*h
, int delta
)
944 struct list_head surplus_list
;
945 struct page
*page
, *tmp
;
947 int needed
, allocated
;
948 bool alloc_ok
= true;
950 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
952 h
->resv_huge_pages
+= delta
;
957 INIT_LIST_HEAD(&surplus_list
);
961 spin_unlock(&hugetlb_lock
);
962 for (i
= 0; i
< needed
; i
++) {
963 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
968 list_add(&page
->lru
, &surplus_list
);
973 * After retaking hugetlb_lock, we need to recalculate 'needed'
974 * because either resv_huge_pages or free_huge_pages may have changed.
976 spin_lock(&hugetlb_lock
);
977 needed
= (h
->resv_huge_pages
+ delta
) -
978 (h
->free_huge_pages
+ allocated
);
983 * We were not able to allocate enough pages to
984 * satisfy the entire reservation so we free what
985 * we've allocated so far.
990 * The surplus_list now contains _at_least_ the number of extra pages
991 * needed to accommodate the reservation. Add the appropriate number
992 * of pages to the hugetlb pool and free the extras back to the buddy
993 * allocator. Commit the entire reservation here to prevent another
994 * process from stealing the pages as they are added to the pool but
995 * before they are reserved.
998 h
->resv_huge_pages
+= delta
;
1001 /* Free the needed pages to the hugetlb pool */
1002 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1006 * This page is now managed by the hugetlb allocator and has
1007 * no users -- drop the buddy allocator's reference.
1009 put_page_testzero(page
);
1010 VM_BUG_ON(page_count(page
));
1011 enqueue_huge_page(h
, page
);
1014 spin_unlock(&hugetlb_lock
);
1016 /* Free unnecessary surplus pages to the buddy allocator */
1017 if (!list_empty(&surplus_list
)) {
1018 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1022 spin_lock(&hugetlb_lock
);
1028 * When releasing a hugetlb pool reservation, any surplus pages that were
1029 * allocated to satisfy the reservation must be explicitly freed if they were
1031 * Called with hugetlb_lock held.
1033 static void return_unused_surplus_pages(struct hstate
*h
,
1034 unsigned long unused_resv_pages
)
1036 unsigned long nr_pages
;
1038 /* Uncommit the reservation */
1039 h
->resv_huge_pages
-= unused_resv_pages
;
1041 /* Cannot return gigantic pages currently */
1042 if (h
->order
>= MAX_ORDER
)
1045 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1048 * We want to release as many surplus pages as possible, spread
1049 * evenly across all nodes with memory. Iterate across these nodes
1050 * until we can no longer free unreserved surplus pages. This occurs
1051 * when the nodes with surplus pages have no free pages.
1052 * free_pool_huge_page() will balance the the freed pages across the
1053 * on-line nodes with memory and will handle the hstate accounting.
1055 while (nr_pages
--) {
1056 if (!free_pool_huge_page(h
, &node_states
[N_HIGH_MEMORY
], 1))
1062 * Determine if the huge page at addr within the vma has an associated
1063 * reservation. Where it does not we will need to logically increase
1064 * reservation and actually increase subpool usage before an allocation
1065 * can occur. Where any new reservation would be required the
1066 * reservation change is prepared, but not committed. Once the page
1067 * has been allocated from the subpool and instantiated the change should
1068 * be committed via vma_commit_reservation. No action is required on
1071 static long vma_needs_reservation(struct hstate
*h
,
1072 struct vm_area_struct
*vma
, unsigned long addr
)
1074 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1075 struct inode
*inode
= mapping
->host
;
1077 if (vma
->vm_flags
& VM_MAYSHARE
) {
1078 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1079 return region_chg(&inode
->i_mapping
->private_list
,
1082 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1087 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1088 struct resv_map
*reservations
= vma_resv_map(vma
);
1090 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
1096 static void vma_commit_reservation(struct hstate
*h
,
1097 struct vm_area_struct
*vma
, unsigned long addr
)
1099 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1100 struct inode
*inode
= mapping
->host
;
1102 if (vma
->vm_flags
& VM_MAYSHARE
) {
1103 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1104 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1106 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1107 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1108 struct resv_map
*reservations
= vma_resv_map(vma
);
1110 /* Mark this page used in the map. */
1111 region_add(&reservations
->regions
, idx
, idx
+ 1);
1115 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1116 unsigned long addr
, int avoid_reserve
)
1118 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1119 struct hstate
*h
= hstate_vma(vma
);
1123 struct hugetlb_cgroup
*h_cg
;
1125 idx
= hstate_index(h
);
1127 * Processes that did not create the mapping will have no
1128 * reserves and will not have accounted against subpool
1129 * limit. Check that the subpool limit can be made before
1130 * satisfying the allocation MAP_NORESERVE mappings may also
1131 * need pages and subpool limit allocated allocated if no reserve
1134 chg
= vma_needs_reservation(h
, vma
, addr
);
1136 return ERR_PTR(-ENOMEM
);
1138 if (hugepage_subpool_get_pages(spool
, chg
))
1139 return ERR_PTR(-ENOSPC
);
1141 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1143 hugepage_subpool_put_pages(spool
, chg
);
1144 return ERR_PTR(-ENOSPC
);
1146 spin_lock(&hugetlb_lock
);
1147 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
1149 /* update page cgroup details */
1150 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
),
1152 spin_unlock(&hugetlb_lock
);
1154 spin_unlock(&hugetlb_lock
);
1155 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1157 hugetlb_cgroup_uncharge_cgroup(idx
,
1158 pages_per_huge_page(h
),
1160 hugepage_subpool_put_pages(spool
, chg
);
1161 return ERR_PTR(-ENOSPC
);
1163 spin_lock(&hugetlb_lock
);
1164 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
),
1166 list_move(&page
->lru
, &h
->hugepage_activelist
);
1167 spin_unlock(&hugetlb_lock
);
1170 set_page_private(page
, (unsigned long)spool
);
1172 vma_commit_reservation(h
, vma
, addr
);
1176 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1178 struct huge_bootmem_page
*m
;
1179 int nr_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
1184 addr
= __alloc_bootmem_node_nopanic(
1185 NODE_DATA(hstate_next_node_to_alloc(h
,
1186 &node_states
[N_HIGH_MEMORY
])),
1187 huge_page_size(h
), huge_page_size(h
), 0);
1191 * Use the beginning of the huge page to store the
1192 * huge_bootmem_page struct (until gather_bootmem
1193 * puts them into the mem_map).
1203 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1204 /* Put them into a private list first because mem_map is not up yet */
1205 list_add(&m
->list
, &huge_boot_pages
);
1210 static void prep_compound_huge_page(struct page
*page
, int order
)
1212 if (unlikely(order
> (MAX_ORDER
- 1)))
1213 prep_compound_gigantic_page(page
, order
);
1215 prep_compound_page(page
, order
);
1218 /* Put bootmem huge pages into the standard lists after mem_map is up */
1219 static void __init
gather_bootmem_prealloc(void)
1221 struct huge_bootmem_page
*m
;
1223 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1224 struct hstate
*h
= m
->hstate
;
1227 #ifdef CONFIG_HIGHMEM
1228 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1229 free_bootmem_late((unsigned long)m
,
1230 sizeof(struct huge_bootmem_page
));
1232 page
= virt_to_page(m
);
1234 __ClearPageReserved(page
);
1235 WARN_ON(page_count(page
) != 1);
1236 prep_compound_huge_page(page
, h
->order
);
1237 prep_new_huge_page(h
, page
, page_to_nid(page
));
1239 * If we had gigantic hugepages allocated at boot time, we need
1240 * to restore the 'stolen' pages to totalram_pages in order to
1241 * fix confusing memory reports from free(1) and another
1242 * side-effects, like CommitLimit going negative.
1244 if (h
->order
> (MAX_ORDER
- 1))
1245 totalram_pages
+= 1 << h
->order
;
1249 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1253 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1254 if (h
->order
>= MAX_ORDER
) {
1255 if (!alloc_bootmem_huge_page(h
))
1257 } else if (!alloc_fresh_huge_page(h
,
1258 &node_states
[N_HIGH_MEMORY
]))
1261 h
->max_huge_pages
= i
;
1264 static void __init
hugetlb_init_hstates(void)
1268 for_each_hstate(h
) {
1269 /* oversize hugepages were init'ed in early boot */
1270 if (h
->order
< MAX_ORDER
)
1271 hugetlb_hstate_alloc_pages(h
);
1275 static char * __init
memfmt(char *buf
, unsigned long n
)
1277 if (n
>= (1UL << 30))
1278 sprintf(buf
, "%lu GB", n
>> 30);
1279 else if (n
>= (1UL << 20))
1280 sprintf(buf
, "%lu MB", n
>> 20);
1282 sprintf(buf
, "%lu KB", n
>> 10);
1286 static void __init
report_hugepages(void)
1290 for_each_hstate(h
) {
1292 printk(KERN_INFO
"HugeTLB registered %s page size, "
1293 "pre-allocated %ld pages\n",
1294 memfmt(buf
, huge_page_size(h
)),
1295 h
->free_huge_pages
);
1299 #ifdef CONFIG_HIGHMEM
1300 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1301 nodemask_t
*nodes_allowed
)
1305 if (h
->order
>= MAX_ORDER
)
1308 for_each_node_mask(i
, *nodes_allowed
) {
1309 struct page
*page
, *next
;
1310 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1311 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1312 if (count
>= h
->nr_huge_pages
)
1314 if (PageHighMem(page
))
1316 list_del(&page
->lru
);
1317 update_and_free_page(h
, page
);
1318 h
->free_huge_pages
--;
1319 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1324 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1325 nodemask_t
*nodes_allowed
)
1331 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1332 * balanced by operating on them in a round-robin fashion.
1333 * Returns 1 if an adjustment was made.
1335 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1338 int start_nid
, next_nid
;
1341 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1344 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
1346 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
1347 next_nid
= start_nid
;
1353 * To shrink on this node, there must be a surplus page
1355 if (!h
->surplus_huge_pages_node
[nid
]) {
1356 next_nid
= hstate_next_node_to_alloc(h
,
1363 * Surplus cannot exceed the total number of pages
1365 if (h
->surplus_huge_pages_node
[nid
] >=
1366 h
->nr_huge_pages_node
[nid
]) {
1367 next_nid
= hstate_next_node_to_free(h
,
1373 h
->surplus_huge_pages
+= delta
;
1374 h
->surplus_huge_pages_node
[nid
] += delta
;
1377 } while (next_nid
!= start_nid
);
1382 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1383 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1384 nodemask_t
*nodes_allowed
)
1386 unsigned long min_count
, ret
;
1388 if (h
->order
>= MAX_ORDER
)
1389 return h
->max_huge_pages
;
1392 * Increase the pool size
1393 * First take pages out of surplus state. Then make up the
1394 * remaining difference by allocating fresh huge pages.
1396 * We might race with alloc_buddy_huge_page() here and be unable
1397 * to convert a surplus huge page to a normal huge page. That is
1398 * not critical, though, it just means the overall size of the
1399 * pool might be one hugepage larger than it needs to be, but
1400 * within all the constraints specified by the sysctls.
1402 spin_lock(&hugetlb_lock
);
1403 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1404 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1408 while (count
> persistent_huge_pages(h
)) {
1410 * If this allocation races such that we no longer need the
1411 * page, free_huge_page will handle it by freeing the page
1412 * and reducing the surplus.
1414 spin_unlock(&hugetlb_lock
);
1415 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1416 spin_lock(&hugetlb_lock
);
1420 /* Bail for signals. Probably ctrl-c from user */
1421 if (signal_pending(current
))
1426 * Decrease the pool size
1427 * First return free pages to the buddy allocator (being careful
1428 * to keep enough around to satisfy reservations). Then place
1429 * pages into surplus state as needed so the pool will shrink
1430 * to the desired size as pages become free.
1432 * By placing pages into the surplus state independent of the
1433 * overcommit value, we are allowing the surplus pool size to
1434 * exceed overcommit. There are few sane options here. Since
1435 * alloc_buddy_huge_page() is checking the global counter,
1436 * though, we'll note that we're not allowed to exceed surplus
1437 * and won't grow the pool anywhere else. Not until one of the
1438 * sysctls are changed, or the surplus pages go out of use.
1440 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1441 min_count
= max(count
, min_count
);
1442 try_to_free_low(h
, min_count
, nodes_allowed
);
1443 while (min_count
< persistent_huge_pages(h
)) {
1444 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1447 while (count
< persistent_huge_pages(h
)) {
1448 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1452 ret
= persistent_huge_pages(h
);
1453 spin_unlock(&hugetlb_lock
);
1457 #define HSTATE_ATTR_RO(_name) \
1458 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1460 #define HSTATE_ATTR(_name) \
1461 static struct kobj_attribute _name##_attr = \
1462 __ATTR(_name, 0644, _name##_show, _name##_store)
1464 static struct kobject
*hugepages_kobj
;
1465 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1467 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1469 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1473 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1474 if (hstate_kobjs
[i
] == kobj
) {
1476 *nidp
= NUMA_NO_NODE
;
1480 return kobj_to_node_hstate(kobj
, nidp
);
1483 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1484 struct kobj_attribute
*attr
, char *buf
)
1487 unsigned long nr_huge_pages
;
1490 h
= kobj_to_hstate(kobj
, &nid
);
1491 if (nid
== NUMA_NO_NODE
)
1492 nr_huge_pages
= h
->nr_huge_pages
;
1494 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1496 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1499 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1500 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1501 const char *buf
, size_t len
)
1505 unsigned long count
;
1507 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1509 err
= strict_strtoul(buf
, 10, &count
);
1513 h
= kobj_to_hstate(kobj
, &nid
);
1514 if (h
->order
>= MAX_ORDER
) {
1519 if (nid
== NUMA_NO_NODE
) {
1521 * global hstate attribute
1523 if (!(obey_mempolicy
&&
1524 init_nodemask_of_mempolicy(nodes_allowed
))) {
1525 NODEMASK_FREE(nodes_allowed
);
1526 nodes_allowed
= &node_states
[N_HIGH_MEMORY
];
1528 } else if (nodes_allowed
) {
1530 * per node hstate attribute: adjust count to global,
1531 * but restrict alloc/free to the specified node.
1533 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1534 init_nodemask_of_node(nodes_allowed
, nid
);
1536 nodes_allowed
= &node_states
[N_HIGH_MEMORY
];
1538 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1540 if (nodes_allowed
!= &node_states
[N_HIGH_MEMORY
])
1541 NODEMASK_FREE(nodes_allowed
);
1545 NODEMASK_FREE(nodes_allowed
);
1549 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1550 struct kobj_attribute
*attr
, char *buf
)
1552 return nr_hugepages_show_common(kobj
, attr
, buf
);
1555 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1556 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1558 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1560 HSTATE_ATTR(nr_hugepages
);
1565 * hstate attribute for optionally mempolicy-based constraint on persistent
1566 * huge page alloc/free.
1568 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1569 struct kobj_attribute
*attr
, char *buf
)
1571 return nr_hugepages_show_common(kobj
, attr
, buf
);
1574 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1575 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1577 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1579 HSTATE_ATTR(nr_hugepages_mempolicy
);
1583 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1584 struct kobj_attribute
*attr
, char *buf
)
1586 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1587 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1590 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1591 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1594 unsigned long input
;
1595 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1597 if (h
->order
>= MAX_ORDER
)
1600 err
= strict_strtoul(buf
, 10, &input
);
1604 spin_lock(&hugetlb_lock
);
1605 h
->nr_overcommit_huge_pages
= input
;
1606 spin_unlock(&hugetlb_lock
);
1610 HSTATE_ATTR(nr_overcommit_hugepages
);
1612 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1613 struct kobj_attribute
*attr
, char *buf
)
1616 unsigned long free_huge_pages
;
1619 h
= kobj_to_hstate(kobj
, &nid
);
1620 if (nid
== NUMA_NO_NODE
)
1621 free_huge_pages
= h
->free_huge_pages
;
1623 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1625 return sprintf(buf
, "%lu\n", free_huge_pages
);
1627 HSTATE_ATTR_RO(free_hugepages
);
1629 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1630 struct kobj_attribute
*attr
, char *buf
)
1632 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1633 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1635 HSTATE_ATTR_RO(resv_hugepages
);
1637 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1638 struct kobj_attribute
*attr
, char *buf
)
1641 unsigned long surplus_huge_pages
;
1644 h
= kobj_to_hstate(kobj
, &nid
);
1645 if (nid
== NUMA_NO_NODE
)
1646 surplus_huge_pages
= h
->surplus_huge_pages
;
1648 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1650 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1652 HSTATE_ATTR_RO(surplus_hugepages
);
1654 static struct attribute
*hstate_attrs
[] = {
1655 &nr_hugepages_attr
.attr
,
1656 &nr_overcommit_hugepages_attr
.attr
,
1657 &free_hugepages_attr
.attr
,
1658 &resv_hugepages_attr
.attr
,
1659 &surplus_hugepages_attr
.attr
,
1661 &nr_hugepages_mempolicy_attr
.attr
,
1666 static struct attribute_group hstate_attr_group
= {
1667 .attrs
= hstate_attrs
,
1670 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1671 struct kobject
**hstate_kobjs
,
1672 struct attribute_group
*hstate_attr_group
)
1675 int hi
= hstate_index(h
);
1677 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1678 if (!hstate_kobjs
[hi
])
1681 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1683 kobject_put(hstate_kobjs
[hi
]);
1688 static void __init
hugetlb_sysfs_init(void)
1693 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1694 if (!hugepages_kobj
)
1697 for_each_hstate(h
) {
1698 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1699 hstate_kobjs
, &hstate_attr_group
);
1701 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s",
1709 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1710 * with node devices in node_devices[] using a parallel array. The array
1711 * index of a node device or _hstate == node id.
1712 * This is here to avoid any static dependency of the node device driver, in
1713 * the base kernel, on the hugetlb module.
1715 struct node_hstate
{
1716 struct kobject
*hugepages_kobj
;
1717 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1719 struct node_hstate node_hstates
[MAX_NUMNODES
];
1722 * A subset of global hstate attributes for node devices
1724 static struct attribute
*per_node_hstate_attrs
[] = {
1725 &nr_hugepages_attr
.attr
,
1726 &free_hugepages_attr
.attr
,
1727 &surplus_hugepages_attr
.attr
,
1731 static struct attribute_group per_node_hstate_attr_group
= {
1732 .attrs
= per_node_hstate_attrs
,
1736 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1737 * Returns node id via non-NULL nidp.
1739 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1743 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1744 struct node_hstate
*nhs
= &node_hstates
[nid
];
1746 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1747 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1759 * Unregister hstate attributes from a single node device.
1760 * No-op if no hstate attributes attached.
1762 void hugetlb_unregister_node(struct node
*node
)
1765 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1767 if (!nhs
->hugepages_kobj
)
1768 return; /* no hstate attributes */
1770 for_each_hstate(h
) {
1771 int idx
= hstate_index(h
);
1772 if (nhs
->hstate_kobjs
[idx
]) {
1773 kobject_put(nhs
->hstate_kobjs
[idx
]);
1774 nhs
->hstate_kobjs
[idx
] = NULL
;
1778 kobject_put(nhs
->hugepages_kobj
);
1779 nhs
->hugepages_kobj
= NULL
;
1783 * hugetlb module exit: unregister hstate attributes from node devices
1786 static void hugetlb_unregister_all_nodes(void)
1791 * disable node device registrations.
1793 register_hugetlbfs_with_node(NULL
, NULL
);
1796 * remove hstate attributes from any nodes that have them.
1798 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1799 hugetlb_unregister_node(&node_devices
[nid
]);
1803 * Register hstate attributes for a single node device.
1804 * No-op if attributes already registered.
1806 void hugetlb_register_node(struct node
*node
)
1809 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1812 if (nhs
->hugepages_kobj
)
1813 return; /* already allocated */
1815 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1817 if (!nhs
->hugepages_kobj
)
1820 for_each_hstate(h
) {
1821 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1823 &per_node_hstate_attr_group
);
1825 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s"
1827 h
->name
, node
->dev
.id
);
1828 hugetlb_unregister_node(node
);
1835 * hugetlb init time: register hstate attributes for all registered node
1836 * devices of nodes that have memory. All on-line nodes should have
1837 * registered their associated device by this time.
1839 static void hugetlb_register_all_nodes(void)
1843 for_each_node_state(nid
, N_HIGH_MEMORY
) {
1844 struct node
*node
= &node_devices
[nid
];
1845 if (node
->dev
.id
== nid
)
1846 hugetlb_register_node(node
);
1850 * Let the node device driver know we're here so it can
1851 * [un]register hstate attributes on node hotplug.
1853 register_hugetlbfs_with_node(hugetlb_register_node
,
1854 hugetlb_unregister_node
);
1856 #else /* !CONFIG_NUMA */
1858 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1866 static void hugetlb_unregister_all_nodes(void) { }
1868 static void hugetlb_register_all_nodes(void) { }
1872 static void __exit
hugetlb_exit(void)
1876 hugetlb_unregister_all_nodes();
1878 for_each_hstate(h
) {
1879 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1882 kobject_put(hugepages_kobj
);
1884 module_exit(hugetlb_exit
);
1886 static int __init
hugetlb_init(void)
1888 /* Some platform decide whether they support huge pages at boot
1889 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1890 * there is no such support
1892 if (HPAGE_SHIFT
== 0)
1895 if (!size_to_hstate(default_hstate_size
)) {
1896 default_hstate_size
= HPAGE_SIZE
;
1897 if (!size_to_hstate(default_hstate_size
))
1898 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1900 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1901 if (default_hstate_max_huge_pages
)
1902 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1904 hugetlb_init_hstates();
1906 gather_bootmem_prealloc();
1910 hugetlb_sysfs_init();
1912 hugetlb_register_all_nodes();
1916 module_init(hugetlb_init
);
1918 /* Should be called on processing a hugepagesz=... option */
1919 void __init
hugetlb_add_hstate(unsigned order
)
1924 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1925 printk(KERN_WARNING
"hugepagesz= specified twice, ignoring\n");
1928 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1930 h
= &hstates
[hugetlb_max_hstate
++];
1932 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1933 h
->nr_huge_pages
= 0;
1934 h
->free_huge_pages
= 0;
1935 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1936 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1937 INIT_LIST_HEAD(&h
->hugepage_activelist
);
1938 h
->next_nid_to_alloc
= first_node(node_states
[N_HIGH_MEMORY
]);
1939 h
->next_nid_to_free
= first_node(node_states
[N_HIGH_MEMORY
]);
1940 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1941 huge_page_size(h
)/1024);
1943 * Add cgroup control files only if the huge page consists
1944 * of more than two normal pages. This is because we use
1945 * page[2].lru.next for storing cgoup details.
1947 if (order
>= HUGETLB_CGROUP_MIN_ORDER
)
1948 hugetlb_cgroup_file_init(hugetlb_max_hstate
- 1);
1953 static int __init
hugetlb_nrpages_setup(char *s
)
1956 static unsigned long *last_mhp
;
1959 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1960 * so this hugepages= parameter goes to the "default hstate".
1962 if (!hugetlb_max_hstate
)
1963 mhp
= &default_hstate_max_huge_pages
;
1965 mhp
= &parsed_hstate
->max_huge_pages
;
1967 if (mhp
== last_mhp
) {
1968 printk(KERN_WARNING
"hugepages= specified twice without "
1969 "interleaving hugepagesz=, ignoring\n");
1973 if (sscanf(s
, "%lu", mhp
) <= 0)
1977 * Global state is always initialized later in hugetlb_init.
1978 * But we need to allocate >= MAX_ORDER hstates here early to still
1979 * use the bootmem allocator.
1981 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1982 hugetlb_hstate_alloc_pages(parsed_hstate
);
1988 __setup("hugepages=", hugetlb_nrpages_setup
);
1990 static int __init
hugetlb_default_setup(char *s
)
1992 default_hstate_size
= memparse(s
, &s
);
1995 __setup("default_hugepagesz=", hugetlb_default_setup
);
1997 static unsigned int cpuset_mems_nr(unsigned int *array
)
2000 unsigned int nr
= 0;
2002 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2008 #ifdef CONFIG_SYSCTL
2009 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2010 struct ctl_table
*table
, int write
,
2011 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2013 struct hstate
*h
= &default_hstate
;
2017 tmp
= h
->max_huge_pages
;
2019 if (write
&& h
->order
>= MAX_ORDER
)
2023 table
->maxlen
= sizeof(unsigned long);
2024 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2029 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2030 GFP_KERNEL
| __GFP_NORETRY
);
2031 if (!(obey_mempolicy
&&
2032 init_nodemask_of_mempolicy(nodes_allowed
))) {
2033 NODEMASK_FREE(nodes_allowed
);
2034 nodes_allowed
= &node_states
[N_HIGH_MEMORY
];
2036 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2038 if (nodes_allowed
!= &node_states
[N_HIGH_MEMORY
])
2039 NODEMASK_FREE(nodes_allowed
);
2045 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2046 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2049 return hugetlb_sysctl_handler_common(false, table
, write
,
2050 buffer
, length
, ppos
);
2054 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2055 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2057 return hugetlb_sysctl_handler_common(true, table
, write
,
2058 buffer
, length
, ppos
);
2060 #endif /* CONFIG_NUMA */
2062 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2063 void __user
*buffer
,
2064 size_t *length
, loff_t
*ppos
)
2066 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2067 if (hugepages_treat_as_movable
)
2068 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2070 htlb_alloc_mask
= GFP_HIGHUSER
;
2074 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2075 void __user
*buffer
,
2076 size_t *length
, loff_t
*ppos
)
2078 struct hstate
*h
= &default_hstate
;
2082 tmp
= h
->nr_overcommit_huge_pages
;
2084 if (write
&& h
->order
>= MAX_ORDER
)
2088 table
->maxlen
= sizeof(unsigned long);
2089 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2094 spin_lock(&hugetlb_lock
);
2095 h
->nr_overcommit_huge_pages
= tmp
;
2096 spin_unlock(&hugetlb_lock
);
2102 #endif /* CONFIG_SYSCTL */
2104 void hugetlb_report_meminfo(struct seq_file
*m
)
2106 struct hstate
*h
= &default_hstate
;
2108 "HugePages_Total: %5lu\n"
2109 "HugePages_Free: %5lu\n"
2110 "HugePages_Rsvd: %5lu\n"
2111 "HugePages_Surp: %5lu\n"
2112 "Hugepagesize: %8lu kB\n",
2116 h
->surplus_huge_pages
,
2117 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2120 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2122 struct hstate
*h
= &default_hstate
;
2124 "Node %d HugePages_Total: %5u\n"
2125 "Node %d HugePages_Free: %5u\n"
2126 "Node %d HugePages_Surp: %5u\n",
2127 nid
, h
->nr_huge_pages_node
[nid
],
2128 nid
, h
->free_huge_pages_node
[nid
],
2129 nid
, h
->surplus_huge_pages_node
[nid
]);
2132 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2133 unsigned long hugetlb_total_pages(void)
2135 struct hstate
*h
= &default_hstate
;
2136 return h
->nr_huge_pages
* pages_per_huge_page(h
);
2139 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2143 spin_lock(&hugetlb_lock
);
2145 * When cpuset is configured, it breaks the strict hugetlb page
2146 * reservation as the accounting is done on a global variable. Such
2147 * reservation is completely rubbish in the presence of cpuset because
2148 * the reservation is not checked against page availability for the
2149 * current cpuset. Application can still potentially OOM'ed by kernel
2150 * with lack of free htlb page in cpuset that the task is in.
2151 * Attempt to enforce strict accounting with cpuset is almost
2152 * impossible (or too ugly) because cpuset is too fluid that
2153 * task or memory node can be dynamically moved between cpusets.
2155 * The change of semantics for shared hugetlb mapping with cpuset is
2156 * undesirable. However, in order to preserve some of the semantics,
2157 * we fall back to check against current free page availability as
2158 * a best attempt and hopefully to minimize the impact of changing
2159 * semantics that cpuset has.
2162 if (gather_surplus_pages(h
, delta
) < 0)
2165 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2166 return_unused_surplus_pages(h
, delta
);
2173 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2176 spin_unlock(&hugetlb_lock
);
2180 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2182 struct resv_map
*reservations
= vma_resv_map(vma
);
2185 * This new VMA should share its siblings reservation map if present.
2186 * The VMA will only ever have a valid reservation map pointer where
2187 * it is being copied for another still existing VMA. As that VMA
2188 * has a reference to the reservation map it cannot disappear until
2189 * after this open call completes. It is therefore safe to take a
2190 * new reference here without additional locking.
2193 kref_get(&reservations
->refs
);
2196 static void resv_map_put(struct vm_area_struct
*vma
)
2198 struct resv_map
*reservations
= vma_resv_map(vma
);
2202 kref_put(&reservations
->refs
, resv_map_release
);
2205 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2207 struct hstate
*h
= hstate_vma(vma
);
2208 struct resv_map
*reservations
= vma_resv_map(vma
);
2209 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2210 unsigned long reserve
;
2211 unsigned long start
;
2215 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2216 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2218 reserve
= (end
- start
) -
2219 region_count(&reservations
->regions
, start
, end
);
2224 hugetlb_acct_memory(h
, -reserve
);
2225 hugepage_subpool_put_pages(spool
, reserve
);
2231 * We cannot handle pagefaults against hugetlb pages at all. They cause
2232 * handle_mm_fault() to try to instantiate regular-sized pages in the
2233 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2236 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2242 const struct vm_operations_struct hugetlb_vm_ops
= {
2243 .fault
= hugetlb_vm_op_fault
,
2244 .open
= hugetlb_vm_op_open
,
2245 .close
= hugetlb_vm_op_close
,
2248 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2255 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
2257 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
2259 entry
= pte_mkyoung(entry
);
2260 entry
= pte_mkhuge(entry
);
2261 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2266 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2267 unsigned long address
, pte_t
*ptep
)
2271 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
2272 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2273 update_mmu_cache(vma
, address
, ptep
);
2277 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2278 struct vm_area_struct
*vma
)
2280 pte_t
*src_pte
, *dst_pte
, entry
;
2281 struct page
*ptepage
;
2284 struct hstate
*h
= hstate_vma(vma
);
2285 unsigned long sz
= huge_page_size(h
);
2287 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2289 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2290 src_pte
= huge_pte_offset(src
, addr
);
2293 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2297 /* If the pagetables are shared don't copy or take references */
2298 if (dst_pte
== src_pte
)
2301 spin_lock(&dst
->page_table_lock
);
2302 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2303 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2305 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2306 entry
= huge_ptep_get(src_pte
);
2307 ptepage
= pte_page(entry
);
2309 page_dup_rmap(ptepage
);
2310 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2312 spin_unlock(&src
->page_table_lock
);
2313 spin_unlock(&dst
->page_table_lock
);
2321 static int is_hugetlb_entry_migration(pte_t pte
)
2325 if (huge_pte_none(pte
) || pte_present(pte
))
2327 swp
= pte_to_swp_entry(pte
);
2328 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2334 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2338 if (huge_pte_none(pte
) || pte_present(pte
))
2340 swp
= pte_to_swp_entry(pte
);
2341 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2347 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2348 unsigned long start
, unsigned long end
,
2349 struct page
*ref_page
)
2351 int force_flush
= 0;
2352 struct mm_struct
*mm
= vma
->vm_mm
;
2353 unsigned long address
;
2357 struct hstate
*h
= hstate_vma(vma
);
2358 unsigned long sz
= huge_page_size(h
);
2360 WARN_ON(!is_vm_hugetlb_page(vma
));
2361 BUG_ON(start
& ~huge_page_mask(h
));
2362 BUG_ON(end
& ~huge_page_mask(h
));
2364 tlb_start_vma(tlb
, vma
);
2365 mmu_notifier_invalidate_range_start(mm
, start
, end
);
2367 spin_lock(&mm
->page_table_lock
);
2368 for (address
= start
; address
< end
; address
+= sz
) {
2369 ptep
= huge_pte_offset(mm
, address
);
2373 if (huge_pmd_unshare(mm
, &address
, ptep
))
2376 pte
= huge_ptep_get(ptep
);
2377 if (huge_pte_none(pte
))
2381 * HWPoisoned hugepage is already unmapped and dropped reference
2383 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
)))
2386 page
= pte_page(pte
);
2388 * If a reference page is supplied, it is because a specific
2389 * page is being unmapped, not a range. Ensure the page we
2390 * are about to unmap is the actual page of interest.
2393 if (page
!= ref_page
)
2397 * Mark the VMA as having unmapped its page so that
2398 * future faults in this VMA will fail rather than
2399 * looking like data was lost
2401 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2404 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2405 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2407 set_page_dirty(page
);
2409 page_remove_rmap(page
);
2410 force_flush
= !__tlb_remove_page(tlb
, page
);
2413 /* Bail out after unmapping reference page if supplied */
2417 spin_unlock(&mm
->page_table_lock
);
2419 * mmu_gather ran out of room to batch pages, we break out of
2420 * the PTE lock to avoid doing the potential expensive TLB invalidate
2421 * and page-free while holding it.
2426 if (address
< end
&& !ref_page
)
2429 mmu_notifier_invalidate_range_end(mm
, start
, end
);
2430 tlb_end_vma(tlb
, vma
);
2433 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2434 struct vm_area_struct
*vma
, unsigned long start
,
2435 unsigned long end
, struct page
*ref_page
)
2437 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2440 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2441 * test will fail on a vma being torn down, and not grab a page table
2442 * on its way out. We're lucky that the flag has such an appropriate
2443 * name, and can in fact be safely cleared here. We could clear it
2444 * before the __unmap_hugepage_range above, but all that's necessary
2445 * is to clear it before releasing the i_mmap_mutex. This works
2446 * because in the context this is called, the VMA is about to be
2447 * destroyed and the i_mmap_mutex is held.
2449 vma
->vm_flags
&= ~VM_MAYSHARE
;
2452 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2453 unsigned long end
, struct page
*ref_page
)
2455 struct mm_struct
*mm
;
2456 struct mmu_gather tlb
;
2460 tlb_gather_mmu(&tlb
, mm
, 0);
2461 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2462 tlb_finish_mmu(&tlb
, start
, end
);
2466 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2467 * mappping it owns the reserve page for. The intention is to unmap the page
2468 * from other VMAs and let the children be SIGKILLed if they are faulting the
2471 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2472 struct page
*page
, unsigned long address
)
2474 struct hstate
*h
= hstate_vma(vma
);
2475 struct vm_area_struct
*iter_vma
;
2476 struct address_space
*mapping
;
2480 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2481 * from page cache lookup which is in HPAGE_SIZE units.
2483 address
= address
& huge_page_mask(h
);
2484 pgoff
= vma_hugecache_offset(h
, vma
, address
);
2485 mapping
= vma
->vm_file
->f_dentry
->d_inode
->i_mapping
;
2488 * Take the mapping lock for the duration of the table walk. As
2489 * this mapping should be shared between all the VMAs,
2490 * __unmap_hugepage_range() is called as the lock is already held
2492 mutex_lock(&mapping
->i_mmap_mutex
);
2493 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2494 /* Do not unmap the current VMA */
2495 if (iter_vma
== vma
)
2499 * Unmap the page from other VMAs without their own reserves.
2500 * They get marked to be SIGKILLed if they fault in these
2501 * areas. This is because a future no-page fault on this VMA
2502 * could insert a zeroed page instead of the data existing
2503 * from the time of fork. This would look like data corruption
2505 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2506 unmap_hugepage_range(iter_vma
, address
,
2507 address
+ huge_page_size(h
), page
);
2509 mutex_unlock(&mapping
->i_mmap_mutex
);
2515 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2516 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2517 * cannot race with other handlers or page migration.
2518 * Keep the pte_same checks anyway to make transition from the mutex easier.
2520 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2521 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2522 struct page
*pagecache_page
)
2524 struct hstate
*h
= hstate_vma(vma
);
2525 struct page
*old_page
, *new_page
;
2527 int outside_reserve
= 0;
2529 old_page
= pte_page(pte
);
2532 /* If no-one else is actually using this page, avoid the copy
2533 * and just make the page writable */
2534 avoidcopy
= (page_mapcount(old_page
) == 1);
2536 if (PageAnon(old_page
))
2537 page_move_anon_rmap(old_page
, vma
, address
);
2538 set_huge_ptep_writable(vma
, address
, ptep
);
2543 * If the process that created a MAP_PRIVATE mapping is about to
2544 * perform a COW due to a shared page count, attempt to satisfy
2545 * the allocation without using the existing reserves. The pagecache
2546 * page is used to determine if the reserve at this address was
2547 * consumed or not. If reserves were used, a partial faulted mapping
2548 * at the time of fork() could consume its reserves on COW instead
2549 * of the full address range.
2551 if (!(vma
->vm_flags
& VM_MAYSHARE
) &&
2552 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2553 old_page
!= pagecache_page
)
2554 outside_reserve
= 1;
2556 page_cache_get(old_page
);
2558 /* Drop page_table_lock as buddy allocator may be called */
2559 spin_unlock(&mm
->page_table_lock
);
2560 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2562 if (IS_ERR(new_page
)) {
2563 long err
= PTR_ERR(new_page
);
2564 page_cache_release(old_page
);
2567 * If a process owning a MAP_PRIVATE mapping fails to COW,
2568 * it is due to references held by a child and an insufficient
2569 * huge page pool. To guarantee the original mappers
2570 * reliability, unmap the page from child processes. The child
2571 * may get SIGKILLed if it later faults.
2573 if (outside_reserve
) {
2574 BUG_ON(huge_pte_none(pte
));
2575 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2576 BUG_ON(huge_pte_none(pte
));
2577 spin_lock(&mm
->page_table_lock
);
2578 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2579 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2580 goto retry_avoidcopy
;
2582 * race occurs while re-acquiring page_table_lock, and
2590 /* Caller expects lock to be held */
2591 spin_lock(&mm
->page_table_lock
);
2593 return VM_FAULT_OOM
;
2595 return VM_FAULT_SIGBUS
;
2599 * When the original hugepage is shared one, it does not have
2600 * anon_vma prepared.
2602 if (unlikely(anon_vma_prepare(vma
))) {
2603 page_cache_release(new_page
);
2604 page_cache_release(old_page
);
2605 /* Caller expects lock to be held */
2606 spin_lock(&mm
->page_table_lock
);
2607 return VM_FAULT_OOM
;
2610 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2611 pages_per_huge_page(h
));
2612 __SetPageUptodate(new_page
);
2615 * Retake the page_table_lock to check for racing updates
2616 * before the page tables are altered
2618 spin_lock(&mm
->page_table_lock
);
2619 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2620 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2622 mmu_notifier_invalidate_range_start(mm
,
2623 address
& huge_page_mask(h
),
2624 (address
& huge_page_mask(h
)) + huge_page_size(h
));
2625 huge_ptep_clear_flush(vma
, address
, ptep
);
2626 set_huge_pte_at(mm
, address
, ptep
,
2627 make_huge_pte(vma
, new_page
, 1));
2628 page_remove_rmap(old_page
);
2629 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2630 /* Make the old page be freed below */
2631 new_page
= old_page
;
2632 mmu_notifier_invalidate_range_end(mm
,
2633 address
& huge_page_mask(h
),
2634 (address
& huge_page_mask(h
)) + huge_page_size(h
));
2636 page_cache_release(new_page
);
2637 page_cache_release(old_page
);
2641 /* Return the pagecache page at a given address within a VMA */
2642 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2643 struct vm_area_struct
*vma
, unsigned long address
)
2645 struct address_space
*mapping
;
2648 mapping
= vma
->vm_file
->f_mapping
;
2649 idx
= vma_hugecache_offset(h
, vma
, address
);
2651 return find_lock_page(mapping
, idx
);
2655 * Return whether there is a pagecache page to back given address within VMA.
2656 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2658 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2659 struct vm_area_struct
*vma
, unsigned long address
)
2661 struct address_space
*mapping
;
2665 mapping
= vma
->vm_file
->f_mapping
;
2666 idx
= vma_hugecache_offset(h
, vma
, address
);
2668 page
= find_get_page(mapping
, idx
);
2671 return page
!= NULL
;
2674 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2675 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2677 struct hstate
*h
= hstate_vma(vma
);
2678 int ret
= VM_FAULT_SIGBUS
;
2683 struct address_space
*mapping
;
2687 * Currently, we are forced to kill the process in the event the
2688 * original mapper has unmapped pages from the child due to a failed
2689 * COW. Warn that such a situation has occurred as it may not be obvious
2691 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2693 "PID %d killed due to inadequate hugepage pool\n",
2698 mapping
= vma
->vm_file
->f_mapping
;
2699 idx
= vma_hugecache_offset(h
, vma
, address
);
2702 * Use page lock to guard against racing truncation
2703 * before we get page_table_lock.
2706 page
= find_lock_page(mapping
, idx
);
2708 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2711 page
= alloc_huge_page(vma
, address
, 0);
2713 ret
= PTR_ERR(page
);
2717 ret
= VM_FAULT_SIGBUS
;
2720 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2721 __SetPageUptodate(page
);
2723 if (vma
->vm_flags
& VM_MAYSHARE
) {
2725 struct inode
*inode
= mapping
->host
;
2727 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2735 spin_lock(&inode
->i_lock
);
2736 inode
->i_blocks
+= blocks_per_huge_page(h
);
2737 spin_unlock(&inode
->i_lock
);
2740 if (unlikely(anon_vma_prepare(vma
))) {
2742 goto backout_unlocked
;
2748 * If memory error occurs between mmap() and fault, some process
2749 * don't have hwpoisoned swap entry for errored virtual address.
2750 * So we need to block hugepage fault by PG_hwpoison bit check.
2752 if (unlikely(PageHWPoison(page
))) {
2753 ret
= VM_FAULT_HWPOISON
|
2754 VM_FAULT_SET_HINDEX(hstate_index(h
));
2755 goto backout_unlocked
;
2760 * If we are going to COW a private mapping later, we examine the
2761 * pending reservations for this page now. This will ensure that
2762 * any allocations necessary to record that reservation occur outside
2765 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2766 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2768 goto backout_unlocked
;
2771 spin_lock(&mm
->page_table_lock
);
2772 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2777 if (!huge_pte_none(huge_ptep_get(ptep
)))
2781 hugepage_add_new_anon_rmap(page
, vma
, address
);
2783 page_dup_rmap(page
);
2784 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2785 && (vma
->vm_flags
& VM_SHARED
)));
2786 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2788 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2789 /* Optimization, do the COW without a second fault */
2790 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2793 spin_unlock(&mm
->page_table_lock
);
2799 spin_unlock(&mm
->page_table_lock
);
2806 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2807 unsigned long address
, unsigned int flags
)
2812 struct page
*page
= NULL
;
2813 struct page
*pagecache_page
= NULL
;
2814 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2815 struct hstate
*h
= hstate_vma(vma
);
2817 address
&= huge_page_mask(h
);
2819 ptep
= huge_pte_offset(mm
, address
);
2821 entry
= huge_ptep_get(ptep
);
2822 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2823 migration_entry_wait(mm
, (pmd_t
*)ptep
, address
);
2825 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2826 return VM_FAULT_HWPOISON_LARGE
|
2827 VM_FAULT_SET_HINDEX(hstate_index(h
));
2830 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2832 return VM_FAULT_OOM
;
2835 * Serialize hugepage allocation and instantiation, so that we don't
2836 * get spurious allocation failures if two CPUs race to instantiate
2837 * the same page in the page cache.
2839 mutex_lock(&hugetlb_instantiation_mutex
);
2840 entry
= huge_ptep_get(ptep
);
2841 if (huge_pte_none(entry
)) {
2842 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2849 * If we are going to COW the mapping later, we examine the pending
2850 * reservations for this page now. This will ensure that any
2851 * allocations necessary to record that reservation occur outside the
2852 * spinlock. For private mappings, we also lookup the pagecache
2853 * page now as it is used to determine if a reservation has been
2856 if ((flags
& FAULT_FLAG_WRITE
) && !pte_write(entry
)) {
2857 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2862 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2863 pagecache_page
= hugetlbfs_pagecache_page(h
,
2868 * hugetlb_cow() requires page locks of pte_page(entry) and
2869 * pagecache_page, so here we need take the former one
2870 * when page != pagecache_page or !pagecache_page.
2871 * Note that locking order is always pagecache_page -> page,
2872 * so no worry about deadlock.
2874 page
= pte_page(entry
);
2876 if (page
!= pagecache_page
)
2879 spin_lock(&mm
->page_table_lock
);
2880 /* Check for a racing update before calling hugetlb_cow */
2881 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2882 goto out_page_table_lock
;
2885 if (flags
& FAULT_FLAG_WRITE
) {
2886 if (!pte_write(entry
)) {
2887 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2889 goto out_page_table_lock
;
2891 entry
= pte_mkdirty(entry
);
2893 entry
= pte_mkyoung(entry
);
2894 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2895 flags
& FAULT_FLAG_WRITE
))
2896 update_mmu_cache(vma
, address
, ptep
);
2898 out_page_table_lock
:
2899 spin_unlock(&mm
->page_table_lock
);
2901 if (pagecache_page
) {
2902 unlock_page(pagecache_page
);
2903 put_page(pagecache_page
);
2905 if (page
!= pagecache_page
)
2910 mutex_unlock(&hugetlb_instantiation_mutex
);
2915 /* Can be overriden by architectures */
2916 __attribute__((weak
)) struct page
*
2917 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
2918 pud_t
*pud
, int write
)
2924 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2925 struct page
**pages
, struct vm_area_struct
**vmas
,
2926 unsigned long *position
, int *length
, int i
,
2929 unsigned long pfn_offset
;
2930 unsigned long vaddr
= *position
;
2931 int remainder
= *length
;
2932 struct hstate
*h
= hstate_vma(vma
);
2934 spin_lock(&mm
->page_table_lock
);
2935 while (vaddr
< vma
->vm_end
&& remainder
) {
2941 * Some archs (sparc64, sh*) have multiple pte_ts to
2942 * each hugepage. We have to make sure we get the
2943 * first, for the page indexing below to work.
2945 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2946 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
2949 * When coredumping, it suits get_dump_page if we just return
2950 * an error where there's an empty slot with no huge pagecache
2951 * to back it. This way, we avoid allocating a hugepage, and
2952 * the sparse dumpfile avoids allocating disk blocks, but its
2953 * huge holes still show up with zeroes where they need to be.
2955 if (absent
&& (flags
& FOLL_DUMP
) &&
2956 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
2962 ((flags
& FOLL_WRITE
) && !pte_write(huge_ptep_get(pte
)))) {
2965 spin_unlock(&mm
->page_table_lock
);
2966 ret
= hugetlb_fault(mm
, vma
, vaddr
,
2967 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
2968 spin_lock(&mm
->page_table_lock
);
2969 if (!(ret
& VM_FAULT_ERROR
))
2976 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2977 page
= pte_page(huge_ptep_get(pte
));
2980 pages
[i
] = mem_map_offset(page
, pfn_offset
);
2991 if (vaddr
< vma
->vm_end
&& remainder
&&
2992 pfn_offset
< pages_per_huge_page(h
)) {
2994 * We use pfn_offset to avoid touching the pageframes
2995 * of this compound page.
3000 spin_unlock(&mm
->page_table_lock
);
3001 *length
= remainder
;
3004 return i
? i
: -EFAULT
;
3007 void hugetlb_change_protection(struct vm_area_struct
*vma
,
3008 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3010 struct mm_struct
*mm
= vma
->vm_mm
;
3011 unsigned long start
= address
;
3014 struct hstate
*h
= hstate_vma(vma
);
3016 BUG_ON(address
>= end
);
3017 flush_cache_range(vma
, address
, end
);
3019 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3020 spin_lock(&mm
->page_table_lock
);
3021 for (; address
< end
; address
+= huge_page_size(h
)) {
3022 ptep
= huge_pte_offset(mm
, address
);
3025 if (huge_pmd_unshare(mm
, &address
, ptep
))
3027 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3028 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3029 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
3030 set_huge_pte_at(mm
, address
, ptep
, pte
);
3033 spin_unlock(&mm
->page_table_lock
);
3035 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3036 * may have cleared our pud entry and done put_page on the page table:
3037 * once we release i_mmap_mutex, another task can do the final put_page
3038 * and that page table be reused and filled with junk.
3040 flush_tlb_range(vma
, start
, end
);
3041 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3044 int hugetlb_reserve_pages(struct inode
*inode
,
3046 struct vm_area_struct
*vma
,
3047 vm_flags_t vm_flags
)
3050 struct hstate
*h
= hstate_inode(inode
);
3051 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3054 * Only apply hugepage reservation if asked. At fault time, an
3055 * attempt will be made for VM_NORESERVE to allocate a page
3056 * without using reserves
3058 if (vm_flags
& VM_NORESERVE
)
3062 * Shared mappings base their reservation on the number of pages that
3063 * are already allocated on behalf of the file. Private mappings need
3064 * to reserve the full area even if read-only as mprotect() may be
3065 * called to make the mapping read-write. Assume !vma is a shm mapping
3067 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3068 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3070 struct resv_map
*resv_map
= resv_map_alloc();
3076 set_vma_resv_map(vma
, resv_map
);
3077 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3085 /* There must be enough pages in the subpool for the mapping */
3086 if (hugepage_subpool_get_pages(spool
, chg
)) {
3092 * Check enough hugepages are available for the reservation.
3093 * Hand the pages back to the subpool if there are not
3095 ret
= hugetlb_acct_memory(h
, chg
);
3097 hugepage_subpool_put_pages(spool
, chg
);
3102 * Account for the reservations made. Shared mappings record regions
3103 * that have reservations as they are shared by multiple VMAs.
3104 * When the last VMA disappears, the region map says how much
3105 * the reservation was and the page cache tells how much of
3106 * the reservation was consumed. Private mappings are per-VMA and
3107 * only the consumed reservations are tracked. When the VMA
3108 * disappears, the original reservation is the VMA size and the
3109 * consumed reservations are stored in the map. Hence, nothing
3110 * else has to be done for private mappings here
3112 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3113 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3121 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3123 struct hstate
*h
= hstate_inode(inode
);
3124 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3125 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3127 spin_lock(&inode
->i_lock
);
3128 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3129 spin_unlock(&inode
->i_lock
);
3131 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3132 hugetlb_acct_memory(h
, -(chg
- freed
));
3135 #ifdef CONFIG_MEMORY_FAILURE
3137 /* Should be called in hugetlb_lock */
3138 static int is_hugepage_on_freelist(struct page
*hpage
)
3142 struct hstate
*h
= page_hstate(hpage
);
3143 int nid
= page_to_nid(hpage
);
3145 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3152 * This function is called from memory failure code.
3153 * Assume the caller holds page lock of the head page.
3155 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3157 struct hstate
*h
= page_hstate(hpage
);
3158 int nid
= page_to_nid(hpage
);
3161 spin_lock(&hugetlb_lock
);
3162 if (is_hugepage_on_freelist(hpage
)) {
3163 list_del(&hpage
->lru
);
3164 set_page_refcounted(hpage
);
3165 h
->free_huge_pages
--;
3166 h
->free_huge_pages_node
[nid
]--;
3169 spin_unlock(&hugetlb_lock
);