2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
10 #include <linux/seq_file.h>
11 #include <linux/sysctl.h>
12 #include <linux/highmem.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/nodemask.h>
15 #include <linux/pagemap.h>
16 #include <linux/mempolicy.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
23 #include <asm/pgtable.h>
26 #include <linux/hugetlb.h>
29 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
30 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
31 unsigned long hugepages_treat_as_movable
;
33 static int max_hstate
;
34 unsigned int default_hstate_idx
;
35 struct hstate hstates
[HUGE_MAX_HSTATE
];
37 __initdata
LIST_HEAD(huge_boot_pages
);
39 /* for command line parsing */
40 static struct hstate
* __initdata parsed_hstate
;
41 static unsigned long __initdata default_hstate_max_huge_pages
;
42 static unsigned long __initdata default_hstate_size
;
44 #define for_each_hstate(h) \
45 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
48 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
50 static DEFINE_SPINLOCK(hugetlb_lock
);
53 * Region tracking -- allows tracking of reservations and instantiated pages
54 * across the pages in a mapping.
56 * The region data structures are protected by a combination of the mmap_sem
57 * and the hugetlb_instantion_mutex. To access or modify a region the caller
58 * must either hold the mmap_sem for write, or the mmap_sem for read and
59 * the hugetlb_instantiation mutex:
61 * down_write(&mm->mmap_sem);
63 * down_read(&mm->mmap_sem);
64 * mutex_lock(&hugetlb_instantiation_mutex);
67 struct list_head link
;
72 static long region_add(struct list_head
*head
, long f
, long t
)
74 struct file_region
*rg
, *nrg
, *trg
;
76 /* Locate the region we are either in or before. */
77 list_for_each_entry(rg
, head
, link
)
81 /* Round our left edge to the current segment if it encloses us. */
85 /* Check for and consume any regions we now overlap with. */
87 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
88 if (&rg
->link
== head
)
93 /* If this area reaches higher then extend our area to
94 * include it completely. If this is not the first area
95 * which we intend to reuse, free it. */
108 static long region_chg(struct list_head
*head
, long f
, long t
)
110 struct file_region
*rg
, *nrg
;
113 /* Locate the region we are before or in. */
114 list_for_each_entry(rg
, head
, link
)
118 /* If we are below the current region then a new region is required.
119 * Subtle, allocate a new region at the position but make it zero
120 * size such that we can guarantee to record the reservation. */
121 if (&rg
->link
== head
|| t
< rg
->from
) {
122 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
127 INIT_LIST_HEAD(&nrg
->link
);
128 list_add(&nrg
->link
, rg
->link
.prev
);
133 /* Round our left edge to the current segment if it encloses us. */
138 /* Check for and consume any regions we now overlap with. */
139 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
140 if (&rg
->link
== head
)
145 /* We overlap with this area, if it extends futher than
146 * us then we must extend ourselves. Account for its
147 * existing reservation. */
152 chg
-= rg
->to
- rg
->from
;
157 static long region_truncate(struct list_head
*head
, long end
)
159 struct file_region
*rg
, *trg
;
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg
, head
, link
)
166 if (&rg
->link
== head
)
169 /* If we are in the middle of a region then adjust it. */
170 if (end
> rg
->from
) {
173 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
176 /* Drop any remaining regions. */
177 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
178 if (&rg
->link
== head
)
180 chg
+= rg
->to
- rg
->from
;
187 static long region_count(struct list_head
*head
, long f
, long t
)
189 struct file_region
*rg
;
192 /* Locate each segment we overlap with, and count that overlap. */
193 list_for_each_entry(rg
, head
, link
) {
202 seg_from
= max(rg
->from
, f
);
203 seg_to
= min(rg
->to
, t
);
205 chg
+= seg_to
- seg_from
;
212 * Convert the address within this vma to the page offset within
213 * the mapping, in pagecache page units; huge pages here.
215 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
216 struct vm_area_struct
*vma
, unsigned long address
)
218 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
219 (vma
->vm_pgoff
>> huge_page_order(h
));
223 * Return the size of the pages allocated when backing a VMA. In the majority
224 * cases this will be same size as used by the page table entries.
226 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
228 struct hstate
*hstate
;
230 if (!is_vm_hugetlb_page(vma
))
233 hstate
= hstate_vma(vma
);
235 return 1UL << (hstate
->order
+ PAGE_SHIFT
);
239 * Return the page size being used by the MMU to back a VMA. In the majority
240 * of cases, the page size used by the kernel matches the MMU size. On
241 * architectures where it differs, an architecture-specific version of this
242 * function is required.
244 #ifndef vma_mmu_pagesize
245 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
247 return vma_kernel_pagesize(vma
);
252 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
253 * bits of the reservation map pointer, which are always clear due to
256 #define HPAGE_RESV_OWNER (1UL << 0)
257 #define HPAGE_RESV_UNMAPPED (1UL << 1)
258 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
261 * These helpers are used to track how many pages are reserved for
262 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
263 * is guaranteed to have their future faults succeed.
265 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
266 * the reserve counters are updated with the hugetlb_lock held. It is safe
267 * to reset the VMA at fork() time as it is not in use yet and there is no
268 * chance of the global counters getting corrupted as a result of the values.
270 * The private mapping reservation is represented in a subtly different
271 * manner to a shared mapping. A shared mapping has a region map associated
272 * with the underlying file, this region map represents the backing file
273 * pages which have ever had a reservation assigned which this persists even
274 * after the page is instantiated. A private mapping has a region map
275 * associated with the original mmap which is attached to all VMAs which
276 * reference it, this region map represents those offsets which have consumed
277 * reservation ie. where pages have been instantiated.
279 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
281 return (unsigned long)vma
->vm_private_data
;
284 static void set_vma_private_data(struct vm_area_struct
*vma
,
287 vma
->vm_private_data
= (void *)value
;
292 struct list_head regions
;
295 static struct resv_map
*resv_map_alloc(void)
297 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
301 kref_init(&resv_map
->refs
);
302 INIT_LIST_HEAD(&resv_map
->regions
);
307 static void resv_map_release(struct kref
*ref
)
309 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
311 /* Clear out any active regions before we release the map. */
312 region_truncate(&resv_map
->regions
, 0);
316 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
318 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
319 if (!(vma
->vm_flags
& VM_SHARED
))
320 return (struct resv_map
*)(get_vma_private_data(vma
) &
325 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
327 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
328 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
330 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
331 HPAGE_RESV_MASK
) | (unsigned long)map
);
334 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
336 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
337 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
339 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
342 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
344 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
346 return (get_vma_private_data(vma
) & flag
) != 0;
349 /* Decrement the reserved pages in the hugepage pool by one */
350 static void decrement_hugepage_resv_vma(struct hstate
*h
,
351 struct vm_area_struct
*vma
)
353 if (vma
->vm_flags
& VM_NORESERVE
)
356 if (vma
->vm_flags
& VM_SHARED
) {
357 /* Shared mappings always use reserves */
358 h
->resv_huge_pages
--;
359 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
361 * Only the process that called mmap() has reserves for
364 h
->resv_huge_pages
--;
368 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
369 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
371 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
372 if (!(vma
->vm_flags
& VM_SHARED
))
373 vma
->vm_private_data
= (void *)0;
376 /* Returns true if the VMA has associated reserve pages */
377 static int vma_has_reserves(struct vm_area_struct
*vma
)
379 if (vma
->vm_flags
& VM_SHARED
)
381 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
386 static void clear_gigantic_page(struct page
*page
,
387 unsigned long addr
, unsigned long sz
)
390 struct page
*p
= page
;
393 for (i
= 0; i
< sz
/PAGE_SIZE
; i
++, p
= mem_map_next(p
, page
, i
)) {
395 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
398 static void clear_huge_page(struct page
*page
,
399 unsigned long addr
, unsigned long sz
)
403 if (unlikely(sz
> MAX_ORDER_NR_PAGES
)) {
404 clear_gigantic_page(page
, addr
, sz
);
409 for (i
= 0; i
< sz
/PAGE_SIZE
; i
++) {
411 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
415 static void copy_gigantic_page(struct page
*dst
, struct page
*src
,
416 unsigned long addr
, struct vm_area_struct
*vma
)
419 struct hstate
*h
= hstate_vma(vma
);
420 struct page
*dst_base
= dst
;
421 struct page
*src_base
= src
;
423 for (i
= 0; i
< pages_per_huge_page(h
); ) {
425 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
428 dst
= mem_map_next(dst
, dst_base
, i
);
429 src
= mem_map_next(src
, src_base
, i
);
432 static void copy_huge_page(struct page
*dst
, struct page
*src
,
433 unsigned long addr
, struct vm_area_struct
*vma
)
436 struct hstate
*h
= hstate_vma(vma
);
438 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
439 copy_gigantic_page(dst
, src
, addr
, vma
);
444 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
446 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
450 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
452 int nid
= page_to_nid(page
);
453 list_add(&page
->lru
, &h
->hugepage_freelists
[nid
]);
454 h
->free_huge_pages
++;
455 h
->free_huge_pages_node
[nid
]++;
458 static struct page
*dequeue_huge_page(struct hstate
*h
)
461 struct page
*page
= NULL
;
463 for (nid
= 0; nid
< MAX_NUMNODES
; ++nid
) {
464 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
465 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
467 list_del(&page
->lru
);
468 h
->free_huge_pages
--;
469 h
->free_huge_pages_node
[nid
]--;
476 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
477 struct vm_area_struct
*vma
,
478 unsigned long address
, int avoid_reserve
)
481 struct page
*page
= NULL
;
482 struct mempolicy
*mpol
;
483 nodemask_t
*nodemask
;
484 struct zonelist
*zonelist
= huge_zonelist(vma
, address
,
485 htlb_alloc_mask
, &mpol
, &nodemask
);
490 * A child process with MAP_PRIVATE mappings created by their parent
491 * have no page reserves. This check ensures that reservations are
492 * not "stolen". The child may still get SIGKILLed
494 if (!vma_has_reserves(vma
) &&
495 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
498 /* If reserves cannot be used, ensure enough pages are in the pool */
499 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
502 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
503 MAX_NR_ZONES
- 1, nodemask
) {
504 nid
= zone_to_nid(zone
);
505 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
) &&
506 !list_empty(&h
->hugepage_freelists
[nid
])) {
507 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
509 list_del(&page
->lru
);
510 h
->free_huge_pages
--;
511 h
->free_huge_pages_node
[nid
]--;
514 decrement_hugepage_resv_vma(h
, vma
);
523 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
527 VM_BUG_ON(h
->order
>= MAX_ORDER
);
530 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
531 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
532 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
533 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
534 1 << PG_private
| 1<< PG_writeback
);
536 set_compound_page_dtor(page
, NULL
);
537 set_page_refcounted(page
);
538 arch_release_hugepage(page
);
539 __free_pages(page
, huge_page_order(h
));
542 struct hstate
*size_to_hstate(unsigned long size
)
547 if (huge_page_size(h
) == size
)
553 static void free_huge_page(struct page
*page
)
556 * Can't pass hstate in here because it is called from the
557 * compound page destructor.
559 struct hstate
*h
= page_hstate(page
);
560 int nid
= page_to_nid(page
);
561 struct address_space
*mapping
;
563 mapping
= (struct address_space
*) page_private(page
);
564 set_page_private(page
, 0);
565 BUG_ON(page_count(page
));
566 INIT_LIST_HEAD(&page
->lru
);
568 spin_lock(&hugetlb_lock
);
569 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
570 update_and_free_page(h
, page
);
571 h
->surplus_huge_pages
--;
572 h
->surplus_huge_pages_node
[nid
]--;
574 enqueue_huge_page(h
, page
);
576 spin_unlock(&hugetlb_lock
);
578 hugetlb_put_quota(mapping
, 1);
582 * Increment or decrement surplus_huge_pages. Keep node-specific counters
583 * balanced by operating on them in a round-robin fashion.
584 * Returns 1 if an adjustment was made.
586 static int adjust_pool_surplus(struct hstate
*h
, int delta
)
592 VM_BUG_ON(delta
!= -1 && delta
!= 1);
594 nid
= next_node(nid
, node_online_map
);
595 if (nid
== MAX_NUMNODES
)
596 nid
= first_node(node_online_map
);
598 /* To shrink on this node, there must be a surplus page */
599 if (delta
< 0 && !h
->surplus_huge_pages_node
[nid
])
601 /* Surplus cannot exceed the total number of pages */
602 if (delta
> 0 && h
->surplus_huge_pages_node
[nid
] >=
603 h
->nr_huge_pages_node
[nid
])
606 h
->surplus_huge_pages
+= delta
;
607 h
->surplus_huge_pages_node
[nid
] += delta
;
610 } while (nid
!= prev_nid
);
616 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
618 set_compound_page_dtor(page
, free_huge_page
);
619 spin_lock(&hugetlb_lock
);
621 h
->nr_huge_pages_node
[nid
]++;
622 spin_unlock(&hugetlb_lock
);
623 put_page(page
); /* free it into the hugepage allocator */
626 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
630 if (h
->order
>= MAX_ORDER
)
633 page
= alloc_pages_node(nid
,
634 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
635 __GFP_REPEAT
|__GFP_NOWARN
,
638 if (arch_prepare_hugepage(page
)) {
639 __free_pages(page
, huge_page_order(h
));
642 prep_new_huge_page(h
, page
, nid
);
649 * Use a helper variable to find the next node and then
650 * copy it back to hugetlb_next_nid afterwards:
651 * otherwise there's a window in which a racer might
652 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
653 * But we don't need to use a spin_lock here: it really
654 * doesn't matter if occasionally a racer chooses the
655 * same nid as we do. Move nid forward in the mask even
656 * if we just successfully allocated a hugepage so that
657 * the next caller gets hugepages on the next node.
659 static int hstate_next_node(struct hstate
*h
)
662 next_nid
= next_node(h
->hugetlb_next_nid
, node_online_map
);
663 if (next_nid
== MAX_NUMNODES
)
664 next_nid
= first_node(node_online_map
);
665 h
->hugetlb_next_nid
= next_nid
;
669 static int alloc_fresh_huge_page(struct hstate
*h
)
676 start_nid
= h
->hugetlb_next_nid
;
679 page
= alloc_fresh_huge_page_node(h
, h
->hugetlb_next_nid
);
682 next_nid
= hstate_next_node(h
);
683 } while (!page
&& h
->hugetlb_next_nid
!= start_nid
);
686 count_vm_event(HTLB_BUDDY_PGALLOC
);
688 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
693 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
694 struct vm_area_struct
*vma
, unsigned long address
)
699 if (h
->order
>= MAX_ORDER
)
703 * Assume we will successfully allocate the surplus page to
704 * prevent racing processes from causing the surplus to exceed
707 * This however introduces a different race, where a process B
708 * tries to grow the static hugepage pool while alloc_pages() is
709 * called by process A. B will only examine the per-node
710 * counters in determining if surplus huge pages can be
711 * converted to normal huge pages in adjust_pool_surplus(). A
712 * won't be able to increment the per-node counter, until the
713 * lock is dropped by B, but B doesn't drop hugetlb_lock until
714 * no more huge pages can be converted from surplus to normal
715 * state (and doesn't try to convert again). Thus, we have a
716 * case where a surplus huge page exists, the pool is grown, and
717 * the surplus huge page still exists after, even though it
718 * should just have been converted to a normal huge page. This
719 * does not leak memory, though, as the hugepage will be freed
720 * once it is out of use. It also does not allow the counters to
721 * go out of whack in adjust_pool_surplus() as we don't modify
722 * the node values until we've gotten the hugepage and only the
723 * per-node value is checked there.
725 spin_lock(&hugetlb_lock
);
726 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
727 spin_unlock(&hugetlb_lock
);
731 h
->surplus_huge_pages
++;
733 spin_unlock(&hugetlb_lock
);
735 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
736 __GFP_REPEAT
|__GFP_NOWARN
,
739 if (page
&& arch_prepare_hugepage(page
)) {
740 __free_pages(page
, huge_page_order(h
));
744 spin_lock(&hugetlb_lock
);
747 * This page is now managed by the hugetlb allocator and has
748 * no users -- drop the buddy allocator's reference.
750 put_page_testzero(page
);
751 VM_BUG_ON(page_count(page
));
752 nid
= page_to_nid(page
);
753 set_compound_page_dtor(page
, free_huge_page
);
755 * We incremented the global counters already
757 h
->nr_huge_pages_node
[nid
]++;
758 h
->surplus_huge_pages_node
[nid
]++;
759 __count_vm_event(HTLB_BUDDY_PGALLOC
);
762 h
->surplus_huge_pages
--;
763 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
765 spin_unlock(&hugetlb_lock
);
771 * Increase the hugetlb pool such that it can accomodate a reservation
774 static int gather_surplus_pages(struct hstate
*h
, int delta
)
776 struct list_head surplus_list
;
777 struct page
*page
, *tmp
;
779 int needed
, allocated
;
781 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
783 h
->resv_huge_pages
+= delta
;
788 INIT_LIST_HEAD(&surplus_list
);
792 spin_unlock(&hugetlb_lock
);
793 for (i
= 0; i
< needed
; i
++) {
794 page
= alloc_buddy_huge_page(h
, NULL
, 0);
797 * We were not able to allocate enough pages to
798 * satisfy the entire reservation so we free what
799 * we've allocated so far.
801 spin_lock(&hugetlb_lock
);
806 list_add(&page
->lru
, &surplus_list
);
811 * After retaking hugetlb_lock, we need to recalculate 'needed'
812 * because either resv_huge_pages or free_huge_pages may have changed.
814 spin_lock(&hugetlb_lock
);
815 needed
= (h
->resv_huge_pages
+ delta
) -
816 (h
->free_huge_pages
+ allocated
);
821 * The surplus_list now contains _at_least_ the number of extra pages
822 * needed to accomodate the reservation. Add the appropriate number
823 * of pages to the hugetlb pool and free the extras back to the buddy
824 * allocator. Commit the entire reservation here to prevent another
825 * process from stealing the pages as they are added to the pool but
826 * before they are reserved.
829 h
->resv_huge_pages
+= delta
;
832 /* Free the needed pages to the hugetlb pool */
833 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
836 list_del(&page
->lru
);
837 enqueue_huge_page(h
, page
);
840 /* Free unnecessary surplus pages to the buddy allocator */
841 if (!list_empty(&surplus_list
)) {
842 spin_unlock(&hugetlb_lock
);
843 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
844 list_del(&page
->lru
);
846 * The page has a reference count of zero already, so
847 * call free_huge_page directly instead of using
848 * put_page. This must be done with hugetlb_lock
849 * unlocked which is safe because free_huge_page takes
850 * hugetlb_lock before deciding how to free the page.
852 free_huge_page(page
);
854 spin_lock(&hugetlb_lock
);
861 * When releasing a hugetlb pool reservation, any surplus pages that were
862 * allocated to satisfy the reservation must be explicitly freed if they were
865 static void return_unused_surplus_pages(struct hstate
*h
,
866 unsigned long unused_resv_pages
)
870 unsigned long nr_pages
;
873 * We want to release as many surplus pages as possible, spread
874 * evenly across all nodes. Iterate across all nodes until we
875 * can no longer free unreserved surplus pages. This occurs when
876 * the nodes with surplus pages have no free pages.
878 unsigned long remaining_iterations
= num_online_nodes();
880 /* Uncommit the reservation */
881 h
->resv_huge_pages
-= unused_resv_pages
;
883 /* Cannot return gigantic pages currently */
884 if (h
->order
>= MAX_ORDER
)
887 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
889 while (remaining_iterations
-- && nr_pages
) {
890 nid
= next_node(nid
, node_online_map
);
891 if (nid
== MAX_NUMNODES
)
892 nid
= first_node(node_online_map
);
894 if (!h
->surplus_huge_pages_node
[nid
])
897 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
898 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
900 list_del(&page
->lru
);
901 update_and_free_page(h
, page
);
902 h
->free_huge_pages
--;
903 h
->free_huge_pages_node
[nid
]--;
904 h
->surplus_huge_pages
--;
905 h
->surplus_huge_pages_node
[nid
]--;
907 remaining_iterations
= num_online_nodes();
913 * Determine if the huge page at addr within the vma has an associated
914 * reservation. Where it does not we will need to logically increase
915 * reservation and actually increase quota before an allocation can occur.
916 * Where any new reservation would be required the reservation change is
917 * prepared, but not committed. Once the page has been quota'd allocated
918 * an instantiated the change should be committed via vma_commit_reservation.
919 * No action is required on failure.
921 static int vma_needs_reservation(struct hstate
*h
,
922 struct vm_area_struct
*vma
, unsigned long addr
)
924 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
925 struct inode
*inode
= mapping
->host
;
927 if (vma
->vm_flags
& VM_SHARED
) {
928 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
929 return region_chg(&inode
->i_mapping
->private_list
,
932 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
937 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
938 struct resv_map
*reservations
= vma_resv_map(vma
);
940 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
946 static void vma_commit_reservation(struct hstate
*h
,
947 struct vm_area_struct
*vma
, unsigned long addr
)
949 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
950 struct inode
*inode
= mapping
->host
;
952 if (vma
->vm_flags
& VM_SHARED
) {
953 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
954 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
956 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
957 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
958 struct resv_map
*reservations
= vma_resv_map(vma
);
960 /* Mark this page used in the map. */
961 region_add(&reservations
->regions
, idx
, idx
+ 1);
965 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
966 unsigned long addr
, int avoid_reserve
)
968 struct hstate
*h
= hstate_vma(vma
);
970 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
971 struct inode
*inode
= mapping
->host
;
975 * Processes that did not create the mapping will have no reserves and
976 * will not have accounted against quota. Check that the quota can be
977 * made before satisfying the allocation
978 * MAP_NORESERVE mappings may also need pages and quota allocated
979 * if no reserve mapping overlaps.
981 chg
= vma_needs_reservation(h
, vma
, addr
);
985 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
986 return ERR_PTR(-ENOSPC
);
988 spin_lock(&hugetlb_lock
);
989 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
990 spin_unlock(&hugetlb_lock
);
993 page
= alloc_buddy_huge_page(h
, vma
, addr
);
995 hugetlb_put_quota(inode
->i_mapping
, chg
);
996 return ERR_PTR(-VM_FAULT_OOM
);
1000 set_page_refcounted(page
);
1001 set_page_private(page
, (unsigned long) mapping
);
1003 vma_commit_reservation(h
, vma
, addr
);
1008 __attribute__((weak
)) int alloc_bootmem_huge_page(struct hstate
*h
)
1010 struct huge_bootmem_page
*m
;
1011 int nr_nodes
= nodes_weight(node_online_map
);
1016 addr
= __alloc_bootmem_node_nopanic(
1017 NODE_DATA(h
->hugetlb_next_nid
),
1018 huge_page_size(h
), huge_page_size(h
), 0);
1022 * Use the beginning of the huge page to store the
1023 * huge_bootmem_page struct (until gather_bootmem
1024 * puts them into the mem_map).
1030 hstate_next_node(h
);
1036 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1037 /* Put them into a private list first because mem_map is not up yet */
1038 list_add(&m
->list
, &huge_boot_pages
);
1043 static void prep_compound_huge_page(struct page
*page
, int order
)
1045 if (unlikely(order
> (MAX_ORDER
- 1)))
1046 prep_compound_gigantic_page(page
, order
);
1048 prep_compound_page(page
, order
);
1051 /* Put bootmem huge pages into the standard lists after mem_map is up */
1052 static void __init
gather_bootmem_prealloc(void)
1054 struct huge_bootmem_page
*m
;
1056 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1057 struct page
*page
= virt_to_page(m
);
1058 struct hstate
*h
= m
->hstate
;
1059 __ClearPageReserved(page
);
1060 WARN_ON(page_count(page
) != 1);
1061 prep_compound_huge_page(page
, h
->order
);
1062 prep_new_huge_page(h
, page
, page_to_nid(page
));
1066 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1070 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1071 if (h
->order
>= MAX_ORDER
) {
1072 if (!alloc_bootmem_huge_page(h
))
1074 } else if (!alloc_fresh_huge_page(h
))
1077 h
->max_huge_pages
= i
;
1080 static void __init
hugetlb_init_hstates(void)
1084 for_each_hstate(h
) {
1085 /* oversize hugepages were init'ed in early boot */
1086 if (h
->order
< MAX_ORDER
)
1087 hugetlb_hstate_alloc_pages(h
);
1091 static char * __init
memfmt(char *buf
, unsigned long n
)
1093 if (n
>= (1UL << 30))
1094 sprintf(buf
, "%lu GB", n
>> 30);
1095 else if (n
>= (1UL << 20))
1096 sprintf(buf
, "%lu MB", n
>> 20);
1098 sprintf(buf
, "%lu KB", n
>> 10);
1102 static void __init
report_hugepages(void)
1106 for_each_hstate(h
) {
1108 printk(KERN_INFO
"HugeTLB registered %s page size, "
1109 "pre-allocated %ld pages\n",
1110 memfmt(buf
, huge_page_size(h
)),
1111 h
->free_huge_pages
);
1115 #ifdef CONFIG_HIGHMEM
1116 static void try_to_free_low(struct hstate
*h
, unsigned long count
)
1120 if (h
->order
>= MAX_ORDER
)
1123 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
1124 struct page
*page
, *next
;
1125 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1126 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1127 if (count
>= h
->nr_huge_pages
)
1129 if (PageHighMem(page
))
1131 list_del(&page
->lru
);
1132 update_and_free_page(h
, page
);
1133 h
->free_huge_pages
--;
1134 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1139 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
)
1144 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1145 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
)
1147 unsigned long min_count
, ret
;
1149 if (h
->order
>= MAX_ORDER
)
1150 return h
->max_huge_pages
;
1153 * Increase the pool size
1154 * First take pages out of surplus state. Then make up the
1155 * remaining difference by allocating fresh huge pages.
1157 * We might race with alloc_buddy_huge_page() here and be unable
1158 * to convert a surplus huge page to a normal huge page. That is
1159 * not critical, though, it just means the overall size of the
1160 * pool might be one hugepage larger than it needs to be, but
1161 * within all the constraints specified by the sysctls.
1163 spin_lock(&hugetlb_lock
);
1164 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1165 if (!adjust_pool_surplus(h
, -1))
1169 while (count
> persistent_huge_pages(h
)) {
1171 * If this allocation races such that we no longer need the
1172 * page, free_huge_page will handle it by freeing the page
1173 * and reducing the surplus.
1175 spin_unlock(&hugetlb_lock
);
1176 ret
= alloc_fresh_huge_page(h
);
1177 spin_lock(&hugetlb_lock
);
1184 * Decrease the pool size
1185 * First return free pages to the buddy allocator (being careful
1186 * to keep enough around to satisfy reservations). Then place
1187 * pages into surplus state as needed so the pool will shrink
1188 * to the desired size as pages become free.
1190 * By placing pages into the surplus state independent of the
1191 * overcommit value, we are allowing the surplus pool size to
1192 * exceed overcommit. There are few sane options here. Since
1193 * alloc_buddy_huge_page() is checking the global counter,
1194 * though, we'll note that we're not allowed to exceed surplus
1195 * and won't grow the pool anywhere else. Not until one of the
1196 * sysctls are changed, or the surplus pages go out of use.
1198 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1199 min_count
= max(count
, min_count
);
1200 try_to_free_low(h
, min_count
);
1201 while (min_count
< persistent_huge_pages(h
)) {
1202 struct page
*page
= dequeue_huge_page(h
);
1205 update_and_free_page(h
, page
);
1207 while (count
< persistent_huge_pages(h
)) {
1208 if (!adjust_pool_surplus(h
, 1))
1212 ret
= persistent_huge_pages(h
);
1213 spin_unlock(&hugetlb_lock
);
1217 #define HSTATE_ATTR_RO(_name) \
1218 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1220 #define HSTATE_ATTR(_name) \
1221 static struct kobj_attribute _name##_attr = \
1222 __ATTR(_name, 0644, _name##_show, _name##_store)
1224 static struct kobject
*hugepages_kobj
;
1225 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1227 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
)
1230 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1231 if (hstate_kobjs
[i
] == kobj
)
1237 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1238 struct kobj_attribute
*attr
, char *buf
)
1240 struct hstate
*h
= kobj_to_hstate(kobj
);
1241 return sprintf(buf
, "%lu\n", h
->nr_huge_pages
);
1243 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1244 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1247 unsigned long input
;
1248 struct hstate
*h
= kobj_to_hstate(kobj
);
1250 err
= strict_strtoul(buf
, 10, &input
);
1254 h
->max_huge_pages
= set_max_huge_pages(h
, input
);
1258 HSTATE_ATTR(nr_hugepages
);
1260 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1261 struct kobj_attribute
*attr
, char *buf
)
1263 struct hstate
*h
= kobj_to_hstate(kobj
);
1264 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1266 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1267 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1270 unsigned long input
;
1271 struct hstate
*h
= kobj_to_hstate(kobj
);
1273 err
= strict_strtoul(buf
, 10, &input
);
1277 spin_lock(&hugetlb_lock
);
1278 h
->nr_overcommit_huge_pages
= input
;
1279 spin_unlock(&hugetlb_lock
);
1283 HSTATE_ATTR(nr_overcommit_hugepages
);
1285 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1286 struct kobj_attribute
*attr
, char *buf
)
1288 struct hstate
*h
= kobj_to_hstate(kobj
);
1289 return sprintf(buf
, "%lu\n", h
->free_huge_pages
);
1291 HSTATE_ATTR_RO(free_hugepages
);
1293 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1294 struct kobj_attribute
*attr
, char *buf
)
1296 struct hstate
*h
= kobj_to_hstate(kobj
);
1297 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1299 HSTATE_ATTR_RO(resv_hugepages
);
1301 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1302 struct kobj_attribute
*attr
, char *buf
)
1304 struct hstate
*h
= kobj_to_hstate(kobj
);
1305 return sprintf(buf
, "%lu\n", h
->surplus_huge_pages
);
1307 HSTATE_ATTR_RO(surplus_hugepages
);
1309 static struct attribute
*hstate_attrs
[] = {
1310 &nr_hugepages_attr
.attr
,
1311 &nr_overcommit_hugepages_attr
.attr
,
1312 &free_hugepages_attr
.attr
,
1313 &resv_hugepages_attr
.attr
,
1314 &surplus_hugepages_attr
.attr
,
1318 static struct attribute_group hstate_attr_group
= {
1319 .attrs
= hstate_attrs
,
1322 static int __init
hugetlb_sysfs_add_hstate(struct hstate
*h
)
1326 hstate_kobjs
[h
- hstates
] = kobject_create_and_add(h
->name
,
1328 if (!hstate_kobjs
[h
- hstates
])
1331 retval
= sysfs_create_group(hstate_kobjs
[h
- hstates
],
1332 &hstate_attr_group
);
1334 kobject_put(hstate_kobjs
[h
- hstates
]);
1339 static void __init
hugetlb_sysfs_init(void)
1344 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1345 if (!hugepages_kobj
)
1348 for_each_hstate(h
) {
1349 err
= hugetlb_sysfs_add_hstate(h
);
1351 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s",
1356 static void __exit
hugetlb_exit(void)
1360 for_each_hstate(h
) {
1361 kobject_put(hstate_kobjs
[h
- hstates
]);
1364 kobject_put(hugepages_kobj
);
1366 module_exit(hugetlb_exit
);
1368 static int __init
hugetlb_init(void)
1370 /* Some platform decide whether they support huge pages at boot
1371 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1372 * there is no such support
1374 if (HPAGE_SHIFT
== 0)
1377 if (!size_to_hstate(default_hstate_size
)) {
1378 default_hstate_size
= HPAGE_SIZE
;
1379 if (!size_to_hstate(default_hstate_size
))
1380 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1382 default_hstate_idx
= size_to_hstate(default_hstate_size
) - hstates
;
1383 if (default_hstate_max_huge_pages
)
1384 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1386 hugetlb_init_hstates();
1388 gather_bootmem_prealloc();
1392 hugetlb_sysfs_init();
1396 module_init(hugetlb_init
);
1398 /* Should be called on processing a hugepagesz=... option */
1399 void __init
hugetlb_add_hstate(unsigned order
)
1404 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1405 printk(KERN_WARNING
"hugepagesz= specified twice, ignoring\n");
1408 BUG_ON(max_hstate
>= HUGE_MAX_HSTATE
);
1410 h
= &hstates
[max_hstate
++];
1412 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1413 h
->nr_huge_pages
= 0;
1414 h
->free_huge_pages
= 0;
1415 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1416 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1417 h
->hugetlb_next_nid
= first_node(node_online_map
);
1418 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1419 huge_page_size(h
)/1024);
1424 static int __init
hugetlb_nrpages_setup(char *s
)
1427 static unsigned long *last_mhp
;
1430 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1431 * so this hugepages= parameter goes to the "default hstate".
1434 mhp
= &default_hstate_max_huge_pages
;
1436 mhp
= &parsed_hstate
->max_huge_pages
;
1438 if (mhp
== last_mhp
) {
1439 printk(KERN_WARNING
"hugepages= specified twice without "
1440 "interleaving hugepagesz=, ignoring\n");
1444 if (sscanf(s
, "%lu", mhp
) <= 0)
1448 * Global state is always initialized later in hugetlb_init.
1449 * But we need to allocate >= MAX_ORDER hstates here early to still
1450 * use the bootmem allocator.
1452 if (max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1453 hugetlb_hstate_alloc_pages(parsed_hstate
);
1459 __setup("hugepages=", hugetlb_nrpages_setup
);
1461 static int __init
hugetlb_default_setup(char *s
)
1463 default_hstate_size
= memparse(s
, &s
);
1466 __setup("default_hugepagesz=", hugetlb_default_setup
);
1468 static unsigned int cpuset_mems_nr(unsigned int *array
)
1471 unsigned int nr
= 0;
1473 for_each_node_mask(node
, cpuset_current_mems_allowed
)
1479 #ifdef CONFIG_SYSCTL
1480 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
1481 struct file
*file
, void __user
*buffer
,
1482 size_t *length
, loff_t
*ppos
)
1484 struct hstate
*h
= &default_hstate
;
1488 tmp
= h
->max_huge_pages
;
1491 table
->maxlen
= sizeof(unsigned long);
1492 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1495 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
);
1500 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
1501 struct file
*file
, void __user
*buffer
,
1502 size_t *length
, loff_t
*ppos
)
1504 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
1505 if (hugepages_treat_as_movable
)
1506 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
1508 htlb_alloc_mask
= GFP_HIGHUSER
;
1512 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
1513 struct file
*file
, void __user
*buffer
,
1514 size_t *length
, loff_t
*ppos
)
1516 struct hstate
*h
= &default_hstate
;
1520 tmp
= h
->nr_overcommit_huge_pages
;
1523 table
->maxlen
= sizeof(unsigned long);
1524 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1527 spin_lock(&hugetlb_lock
);
1528 h
->nr_overcommit_huge_pages
= tmp
;
1529 spin_unlock(&hugetlb_lock
);
1535 #endif /* CONFIG_SYSCTL */
1537 void hugetlb_report_meminfo(struct seq_file
*m
)
1539 struct hstate
*h
= &default_hstate
;
1541 "HugePages_Total: %5lu\n"
1542 "HugePages_Free: %5lu\n"
1543 "HugePages_Rsvd: %5lu\n"
1544 "HugePages_Surp: %5lu\n"
1545 "Hugepagesize: %8lu kB\n",
1549 h
->surplus_huge_pages
,
1550 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
1553 int hugetlb_report_node_meminfo(int nid
, char *buf
)
1555 struct hstate
*h
= &default_hstate
;
1557 "Node %d HugePages_Total: %5u\n"
1558 "Node %d HugePages_Free: %5u\n"
1559 "Node %d HugePages_Surp: %5u\n",
1560 nid
, h
->nr_huge_pages_node
[nid
],
1561 nid
, h
->free_huge_pages_node
[nid
],
1562 nid
, h
->surplus_huge_pages_node
[nid
]);
1565 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1566 unsigned long hugetlb_total_pages(void)
1568 struct hstate
*h
= &default_hstate
;
1569 return h
->nr_huge_pages
* pages_per_huge_page(h
);
1572 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
1576 spin_lock(&hugetlb_lock
);
1578 * When cpuset is configured, it breaks the strict hugetlb page
1579 * reservation as the accounting is done on a global variable. Such
1580 * reservation is completely rubbish in the presence of cpuset because
1581 * the reservation is not checked against page availability for the
1582 * current cpuset. Application can still potentially OOM'ed by kernel
1583 * with lack of free htlb page in cpuset that the task is in.
1584 * Attempt to enforce strict accounting with cpuset is almost
1585 * impossible (or too ugly) because cpuset is too fluid that
1586 * task or memory node can be dynamically moved between cpusets.
1588 * The change of semantics for shared hugetlb mapping with cpuset is
1589 * undesirable. However, in order to preserve some of the semantics,
1590 * we fall back to check against current free page availability as
1591 * a best attempt and hopefully to minimize the impact of changing
1592 * semantics that cpuset has.
1595 if (gather_surplus_pages(h
, delta
) < 0)
1598 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
1599 return_unused_surplus_pages(h
, delta
);
1606 return_unused_surplus_pages(h
, (unsigned long) -delta
);
1609 spin_unlock(&hugetlb_lock
);
1613 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
1615 struct resv_map
*reservations
= vma_resv_map(vma
);
1618 * This new VMA should share its siblings reservation map if present.
1619 * The VMA will only ever have a valid reservation map pointer where
1620 * it is being copied for another still existing VMA. As that VMA
1621 * has a reference to the reservation map it cannot dissappear until
1622 * after this open call completes. It is therefore safe to take a
1623 * new reference here without additional locking.
1626 kref_get(&reservations
->refs
);
1629 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
1631 struct hstate
*h
= hstate_vma(vma
);
1632 struct resv_map
*reservations
= vma_resv_map(vma
);
1633 unsigned long reserve
;
1634 unsigned long start
;
1638 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
1639 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
1641 reserve
= (end
- start
) -
1642 region_count(&reservations
->regions
, start
, end
);
1644 kref_put(&reservations
->refs
, resv_map_release
);
1647 hugetlb_acct_memory(h
, -reserve
);
1648 hugetlb_put_quota(vma
->vm_file
->f_mapping
, reserve
);
1654 * We cannot handle pagefaults against hugetlb pages at all. They cause
1655 * handle_mm_fault() to try to instantiate regular-sized pages in the
1656 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1659 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1665 struct vm_operations_struct hugetlb_vm_ops
= {
1666 .fault
= hugetlb_vm_op_fault
,
1667 .open
= hugetlb_vm_op_open
,
1668 .close
= hugetlb_vm_op_close
,
1671 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
1678 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
1680 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
1682 entry
= pte_mkyoung(entry
);
1683 entry
= pte_mkhuge(entry
);
1688 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
1689 unsigned long address
, pte_t
*ptep
)
1693 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
1694 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
1695 update_mmu_cache(vma
, address
, entry
);
1700 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
1701 struct vm_area_struct
*vma
)
1703 pte_t
*src_pte
, *dst_pte
, entry
;
1704 struct page
*ptepage
;
1707 struct hstate
*h
= hstate_vma(vma
);
1708 unsigned long sz
= huge_page_size(h
);
1710 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
1712 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
1713 src_pte
= huge_pte_offset(src
, addr
);
1716 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
1720 /* If the pagetables are shared don't copy or take references */
1721 if (dst_pte
== src_pte
)
1724 spin_lock(&dst
->page_table_lock
);
1725 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
1726 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
1728 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
1729 entry
= huge_ptep_get(src_pte
);
1730 ptepage
= pte_page(entry
);
1732 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
1734 spin_unlock(&src
->page_table_lock
);
1735 spin_unlock(&dst
->page_table_lock
);
1743 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1744 unsigned long end
, struct page
*ref_page
)
1746 struct mm_struct
*mm
= vma
->vm_mm
;
1747 unsigned long address
;
1752 struct hstate
*h
= hstate_vma(vma
);
1753 unsigned long sz
= huge_page_size(h
);
1756 * A page gathering list, protected by per file i_mmap_lock. The
1757 * lock is used to avoid list corruption from multiple unmapping
1758 * of the same page since we are using page->lru.
1760 LIST_HEAD(page_list
);
1762 WARN_ON(!is_vm_hugetlb_page(vma
));
1763 BUG_ON(start
& ~huge_page_mask(h
));
1764 BUG_ON(end
& ~huge_page_mask(h
));
1766 mmu_notifier_invalidate_range_start(mm
, start
, end
);
1767 spin_lock(&mm
->page_table_lock
);
1768 for (address
= start
; address
< end
; address
+= sz
) {
1769 ptep
= huge_pte_offset(mm
, address
);
1773 if (huge_pmd_unshare(mm
, &address
, ptep
))
1777 * If a reference page is supplied, it is because a specific
1778 * page is being unmapped, not a range. Ensure the page we
1779 * are about to unmap is the actual page of interest.
1782 pte
= huge_ptep_get(ptep
);
1783 if (huge_pte_none(pte
))
1785 page
= pte_page(pte
);
1786 if (page
!= ref_page
)
1790 * Mark the VMA as having unmapped its page so that
1791 * future faults in this VMA will fail rather than
1792 * looking like data was lost
1794 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
1797 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1798 if (huge_pte_none(pte
))
1801 page
= pte_page(pte
);
1803 set_page_dirty(page
);
1804 list_add(&page
->lru
, &page_list
);
1806 spin_unlock(&mm
->page_table_lock
);
1807 flush_tlb_range(vma
, start
, end
);
1808 mmu_notifier_invalidate_range_end(mm
, start
, end
);
1809 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
1810 list_del(&page
->lru
);
1815 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1816 unsigned long end
, struct page
*ref_page
)
1818 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1819 __unmap_hugepage_range(vma
, start
, end
, ref_page
);
1820 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1824 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1825 * mappping it owns the reserve page for. The intention is to unmap the page
1826 * from other VMAs and let the children be SIGKILLed if they are faulting the
1829 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1830 struct page
*page
, unsigned long address
)
1832 struct hstate
*h
= hstate_vma(vma
);
1833 struct vm_area_struct
*iter_vma
;
1834 struct address_space
*mapping
;
1835 struct prio_tree_iter iter
;
1839 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1840 * from page cache lookup which is in HPAGE_SIZE units.
1842 address
= address
& huge_page_mask(h
);
1843 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
)
1844 + (vma
->vm_pgoff
>> PAGE_SHIFT
);
1845 mapping
= (struct address_space
*)page_private(page
);
1847 vma_prio_tree_foreach(iter_vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1848 /* Do not unmap the current VMA */
1849 if (iter_vma
== vma
)
1853 * Unmap the page from other VMAs without their own reserves.
1854 * They get marked to be SIGKILLed if they fault in these
1855 * areas. This is because a future no-page fault on this VMA
1856 * could insert a zeroed page instead of the data existing
1857 * from the time of fork. This would look like data corruption
1859 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
1860 unmap_hugepage_range(iter_vma
,
1861 address
, address
+ huge_page_size(h
),
1868 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1869 unsigned long address
, pte_t
*ptep
, pte_t pte
,
1870 struct page
*pagecache_page
)
1872 struct hstate
*h
= hstate_vma(vma
);
1873 struct page
*old_page
, *new_page
;
1875 int outside_reserve
= 0;
1877 old_page
= pte_page(pte
);
1880 /* If no-one else is actually using this page, avoid the copy
1881 * and just make the page writable */
1882 avoidcopy
= (page_count(old_page
) == 1);
1884 set_huge_ptep_writable(vma
, address
, ptep
);
1889 * If the process that created a MAP_PRIVATE mapping is about to
1890 * perform a COW due to a shared page count, attempt to satisfy
1891 * the allocation without using the existing reserves. The pagecache
1892 * page is used to determine if the reserve at this address was
1893 * consumed or not. If reserves were used, a partial faulted mapping
1894 * at the time of fork() could consume its reserves on COW instead
1895 * of the full address range.
1897 if (!(vma
->vm_flags
& VM_SHARED
) &&
1898 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
1899 old_page
!= pagecache_page
)
1900 outside_reserve
= 1;
1902 page_cache_get(old_page
);
1903 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
1905 if (IS_ERR(new_page
)) {
1906 page_cache_release(old_page
);
1909 * If a process owning a MAP_PRIVATE mapping fails to COW,
1910 * it is due to references held by a child and an insufficient
1911 * huge page pool. To guarantee the original mappers
1912 * reliability, unmap the page from child processes. The child
1913 * may get SIGKILLed if it later faults.
1915 if (outside_reserve
) {
1916 BUG_ON(huge_pte_none(pte
));
1917 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
1918 BUG_ON(page_count(old_page
) != 1);
1919 BUG_ON(huge_pte_none(pte
));
1920 goto retry_avoidcopy
;
1925 return -PTR_ERR(new_page
);
1928 spin_unlock(&mm
->page_table_lock
);
1929 copy_huge_page(new_page
, old_page
, address
, vma
);
1930 __SetPageUptodate(new_page
);
1931 spin_lock(&mm
->page_table_lock
);
1933 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
1934 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
1936 huge_ptep_clear_flush(vma
, address
, ptep
);
1937 set_huge_pte_at(mm
, address
, ptep
,
1938 make_huge_pte(vma
, new_page
, 1));
1939 /* Make the old page be freed below */
1940 new_page
= old_page
;
1942 page_cache_release(new_page
);
1943 page_cache_release(old_page
);
1947 /* Return the pagecache page at a given address within a VMA */
1948 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
1949 struct vm_area_struct
*vma
, unsigned long address
)
1951 struct address_space
*mapping
;
1954 mapping
= vma
->vm_file
->f_mapping
;
1955 idx
= vma_hugecache_offset(h
, vma
, address
);
1957 return find_lock_page(mapping
, idx
);
1960 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1961 unsigned long address
, pte_t
*ptep
, int write_access
)
1963 struct hstate
*h
= hstate_vma(vma
);
1964 int ret
= VM_FAULT_SIGBUS
;
1968 struct address_space
*mapping
;
1972 * Currently, we are forced to kill the process in the event the
1973 * original mapper has unmapped pages from the child due to a failed
1974 * COW. Warn that such a situation has occured as it may not be obvious
1976 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
1978 "PID %d killed due to inadequate hugepage pool\n",
1983 mapping
= vma
->vm_file
->f_mapping
;
1984 idx
= vma_hugecache_offset(h
, vma
, address
);
1987 * Use page lock to guard against racing truncation
1988 * before we get page_table_lock.
1991 page
= find_lock_page(mapping
, idx
);
1993 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
1996 page
= alloc_huge_page(vma
, address
, 0);
1998 ret
= -PTR_ERR(page
);
2001 clear_huge_page(page
, address
, huge_page_size(h
));
2002 __SetPageUptodate(page
);
2004 if (vma
->vm_flags
& VM_SHARED
) {
2006 struct inode
*inode
= mapping
->host
;
2008 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2016 spin_lock(&inode
->i_lock
);
2017 inode
->i_blocks
+= blocks_per_huge_page(h
);
2018 spin_unlock(&inode
->i_lock
);
2024 * If we are going to COW a private mapping later, we examine the
2025 * pending reservations for this page now. This will ensure that
2026 * any allocations necessary to record that reservation occur outside
2029 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
))
2030 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2032 goto backout_unlocked
;
2035 spin_lock(&mm
->page_table_lock
);
2036 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2041 if (!huge_pte_none(huge_ptep_get(ptep
)))
2044 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2045 && (vma
->vm_flags
& VM_SHARED
)));
2046 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2048 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
2049 /* Optimization, do the COW without a second fault */
2050 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2053 spin_unlock(&mm
->page_table_lock
);
2059 spin_unlock(&mm
->page_table_lock
);
2066 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2067 unsigned long address
, int write_access
)
2072 struct page
*pagecache_page
= NULL
;
2073 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2074 struct hstate
*h
= hstate_vma(vma
);
2076 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2078 return VM_FAULT_OOM
;
2081 * Serialize hugepage allocation and instantiation, so that we don't
2082 * get spurious allocation failures if two CPUs race to instantiate
2083 * the same page in the page cache.
2085 mutex_lock(&hugetlb_instantiation_mutex
);
2086 entry
= huge_ptep_get(ptep
);
2087 if (huge_pte_none(entry
)) {
2088 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
2095 * If we are going to COW the mapping later, we examine the pending
2096 * reservations for this page now. This will ensure that any
2097 * allocations necessary to record that reservation occur outside the
2098 * spinlock. For private mappings, we also lookup the pagecache
2099 * page now as it is used to determine if a reservation has been
2102 if (write_access
&& !pte_write(entry
)) {
2103 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2108 if (!(vma
->vm_flags
& VM_SHARED
))
2109 pagecache_page
= hugetlbfs_pagecache_page(h
,
2113 spin_lock(&mm
->page_table_lock
);
2114 /* Check for a racing update before calling hugetlb_cow */
2115 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2116 goto out_page_table_lock
;
2120 if (!pte_write(entry
)) {
2121 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2123 goto out_page_table_lock
;
2125 entry
= pte_mkdirty(entry
);
2127 entry
= pte_mkyoung(entry
);
2128 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, write_access
))
2129 update_mmu_cache(vma
, address
, entry
);
2131 out_page_table_lock
:
2132 spin_unlock(&mm
->page_table_lock
);
2134 if (pagecache_page
) {
2135 unlock_page(pagecache_page
);
2136 put_page(pagecache_page
);
2140 mutex_unlock(&hugetlb_instantiation_mutex
);
2145 /* Can be overriden by architectures */
2146 __attribute__((weak
)) struct page
*
2147 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
2148 pud_t
*pud
, int write
)
2154 static int huge_zeropage_ok(pte_t
*ptep
, int write
, int shared
)
2156 if (!ptep
|| write
|| shared
)
2159 return huge_pte_none(huge_ptep_get(ptep
));
2162 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2163 struct page
**pages
, struct vm_area_struct
**vmas
,
2164 unsigned long *position
, int *length
, int i
,
2167 unsigned long pfn_offset
;
2168 unsigned long vaddr
= *position
;
2169 int remainder
= *length
;
2170 struct hstate
*h
= hstate_vma(vma
);
2171 int zeropage_ok
= 0;
2172 int shared
= vma
->vm_flags
& VM_SHARED
;
2174 spin_lock(&mm
->page_table_lock
);
2175 while (vaddr
< vma
->vm_end
&& remainder
) {
2180 * Some archs (sparc64, sh*) have multiple pte_ts to
2181 * each hugepage. We have to make * sure we get the
2182 * first, for the page indexing below to work.
2184 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2185 if (huge_zeropage_ok(pte
, write
, shared
))
2189 (huge_pte_none(huge_ptep_get(pte
)) && !zeropage_ok
) ||
2190 (write
&& !pte_write(huge_ptep_get(pte
)))) {
2193 spin_unlock(&mm
->page_table_lock
);
2194 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
2195 spin_lock(&mm
->page_table_lock
);
2196 if (!(ret
& VM_FAULT_ERROR
))
2205 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2206 page
= pte_page(huge_ptep_get(pte
));
2210 pages
[i
] = ZERO_PAGE(0);
2212 pages
[i
] = mem_map_offset(page
, pfn_offset
);
2223 if (vaddr
< vma
->vm_end
&& remainder
&&
2224 pfn_offset
< pages_per_huge_page(h
)) {
2226 * We use pfn_offset to avoid touching the pageframes
2227 * of this compound page.
2232 spin_unlock(&mm
->page_table_lock
);
2233 *length
= remainder
;
2239 void hugetlb_change_protection(struct vm_area_struct
*vma
,
2240 unsigned long address
, unsigned long end
, pgprot_t newprot
)
2242 struct mm_struct
*mm
= vma
->vm_mm
;
2243 unsigned long start
= address
;
2246 struct hstate
*h
= hstate_vma(vma
);
2248 BUG_ON(address
>= end
);
2249 flush_cache_range(vma
, address
, end
);
2251 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2252 spin_lock(&mm
->page_table_lock
);
2253 for (; address
< end
; address
+= huge_page_size(h
)) {
2254 ptep
= huge_pte_offset(mm
, address
);
2257 if (huge_pmd_unshare(mm
, &address
, ptep
))
2259 if (!huge_pte_none(huge_ptep_get(ptep
))) {
2260 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2261 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
2262 set_huge_pte_at(mm
, address
, ptep
, pte
);
2265 spin_unlock(&mm
->page_table_lock
);
2266 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2268 flush_tlb_range(vma
, start
, end
);
2271 int hugetlb_reserve_pages(struct inode
*inode
,
2273 struct vm_area_struct
*vma
)
2276 struct hstate
*h
= hstate_inode(inode
);
2278 if (vma
&& vma
->vm_flags
& VM_NORESERVE
)
2282 * Shared mappings base their reservation on the number of pages that
2283 * are already allocated on behalf of the file. Private mappings need
2284 * to reserve the full area even if read-only as mprotect() may be
2285 * called to make the mapping read-write. Assume !vma is a shm mapping
2287 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2288 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
2290 struct resv_map
*resv_map
= resv_map_alloc();
2296 set_vma_resv_map(vma
, resv_map
);
2297 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
2303 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
2305 ret
= hugetlb_acct_memory(h
, chg
);
2307 hugetlb_put_quota(inode
->i_mapping
, chg
);
2310 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2311 region_add(&inode
->i_mapping
->private_list
, from
, to
);
2315 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
2317 struct hstate
*h
= hstate_inode(inode
);
2318 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
2320 spin_lock(&inode
->i_lock
);
2321 inode
->i_blocks
-= blocks_per_huge_page(h
);
2322 spin_unlock(&inode
->i_lock
);
2324 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
2325 hugetlb_acct_memory(h
, -(chg
- freed
));