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/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
16 #include <linux/mutex.h>
19 #include <asm/pgtable.h>
21 #include <linux/hugetlb.h>
24 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
25 static unsigned long nr_huge_pages
, free_huge_pages
, resv_huge_pages
;
26 static unsigned long surplus_huge_pages
;
27 static unsigned long nr_overcommit_huge_pages
;
28 unsigned long max_huge_pages
;
29 unsigned long sysctl_overcommit_huge_pages
;
30 static struct list_head hugepage_freelists
[MAX_NUMNODES
];
31 static unsigned int nr_huge_pages_node
[MAX_NUMNODES
];
32 static unsigned int free_huge_pages_node
[MAX_NUMNODES
];
33 static unsigned int surplus_huge_pages_node
[MAX_NUMNODES
];
34 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
35 unsigned long hugepages_treat_as_movable
;
36 static int hugetlb_next_nid
;
39 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
41 static DEFINE_SPINLOCK(hugetlb_lock
);
44 * Region tracking -- allows tracking of reservations and instantiated pages
45 * across the pages in a mapping.
48 struct list_head link
;
53 static long region_add(struct list_head
*head
, long f
, long t
)
55 struct file_region
*rg
, *nrg
, *trg
;
57 /* Locate the region we are either in or before. */
58 list_for_each_entry(rg
, head
, link
)
62 /* Round our left edge to the current segment if it encloses us. */
66 /* Check for and consume any regions we now overlap with. */
68 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
69 if (&rg
->link
== head
)
74 /* If this area reaches higher then extend our area to
75 * include it completely. If this is not the first area
76 * which we intend to reuse, free it. */
89 static long region_chg(struct list_head
*head
, long f
, long t
)
91 struct file_region
*rg
, *nrg
;
94 /* Locate the region we are before or in. */
95 list_for_each_entry(rg
, head
, link
)
99 /* If we are below the current region then a new region is required.
100 * Subtle, allocate a new region at the position but make it zero
101 * size such that we can guarantee to record the reservation. */
102 if (&rg
->link
== head
|| t
< rg
->from
) {
103 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
108 INIT_LIST_HEAD(&nrg
->link
);
109 list_add(&nrg
->link
, rg
->link
.prev
);
114 /* Round our left edge to the current segment if it encloses us. */
119 /* Check for and consume any regions we now overlap with. */
120 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
121 if (&rg
->link
== head
)
126 /* We overlap with this area, if it extends futher than
127 * us then we must extend ourselves. Account for its
128 * existing reservation. */
133 chg
-= rg
->to
- rg
->from
;
138 static long region_truncate(struct list_head
*head
, long end
)
140 struct file_region
*rg
, *trg
;
143 /* Locate the region we are either in or before. */
144 list_for_each_entry(rg
, head
, link
)
147 if (&rg
->link
== head
)
150 /* If we are in the middle of a region then adjust it. */
151 if (end
> rg
->from
) {
154 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
157 /* Drop any remaining regions. */
158 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
159 if (&rg
->link
== head
)
161 chg
+= rg
->to
- rg
->from
;
169 * Convert the address within this vma to the page offset within
170 * the mapping, in base page units.
172 static pgoff_t
vma_page_offset(struct vm_area_struct
*vma
,
173 unsigned long address
)
175 return ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
176 (vma
->vm_pgoff
>> PAGE_SHIFT
);
180 * Convert the address within this vma to the page offset within
181 * the mapping, in pagecache page units; huge pages here.
183 static pgoff_t
vma_pagecache_offset(struct vm_area_struct
*vma
,
184 unsigned long address
)
186 return ((address
- vma
->vm_start
) >> HPAGE_SHIFT
) +
187 (vma
->vm_pgoff
>> (HPAGE_SHIFT
- PAGE_SHIFT
));
190 #define HPAGE_RESV_OWNER (1UL << (BITS_PER_LONG - 1))
191 #define HPAGE_RESV_UNMAPPED (1UL << (BITS_PER_LONG - 2))
192 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
194 * These helpers are used to track how many pages are reserved for
195 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
196 * is guaranteed to have their future faults succeed.
198 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
199 * the reserve counters are updated with the hugetlb_lock held. It is safe
200 * to reset the VMA at fork() time as it is not in use yet and there is no
201 * chance of the global counters getting corrupted as a result of the values.
203 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
205 return (unsigned long)vma
->vm_private_data
;
208 static void set_vma_private_data(struct vm_area_struct
*vma
,
211 vma
->vm_private_data
= (void *)value
;
214 static unsigned long vma_resv_huge_pages(struct vm_area_struct
*vma
)
216 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
217 if (!(vma
->vm_flags
& VM_SHARED
))
218 return get_vma_private_data(vma
) & ~HPAGE_RESV_MASK
;
222 static void set_vma_resv_huge_pages(struct vm_area_struct
*vma
,
223 unsigned long reserve
)
225 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
226 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
228 set_vma_private_data(vma
,
229 (get_vma_private_data(vma
) & HPAGE_RESV_MASK
) | reserve
);
232 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
234 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
235 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
237 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
240 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
242 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
244 return (get_vma_private_data(vma
) & flag
) != 0;
247 /* Decrement the reserved pages in the hugepage pool by one */
248 static void decrement_hugepage_resv_vma(struct vm_area_struct
*vma
)
250 if (vma
->vm_flags
& VM_NORESERVE
)
253 if (vma
->vm_flags
& VM_SHARED
) {
254 /* Shared mappings always use reserves */
258 * Only the process that called mmap() has reserves for
261 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
262 unsigned long flags
, reserve
;
264 flags
= (unsigned long)vma
->vm_private_data
&
266 reserve
= (unsigned long)vma
->vm_private_data
- 1;
267 vma
->vm_private_data
= (void *)(reserve
| flags
);
272 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
273 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
275 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
276 if (!(vma
->vm_flags
& VM_SHARED
))
277 vma
->vm_private_data
= (void *)0;
280 /* Returns true if the VMA has associated reserve pages */
281 static int vma_has_private_reserves(struct vm_area_struct
*vma
)
283 if (vma
->vm_flags
& VM_SHARED
)
285 if (!vma_resv_huge_pages(vma
))
290 static void clear_huge_page(struct page
*page
, unsigned long addr
)
295 for (i
= 0; i
< (HPAGE_SIZE
/PAGE_SIZE
); i
++) {
297 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
301 static void copy_huge_page(struct page
*dst
, struct page
*src
,
302 unsigned long addr
, struct vm_area_struct
*vma
)
307 for (i
= 0; i
< HPAGE_SIZE
/PAGE_SIZE
; i
++) {
309 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
313 static void enqueue_huge_page(struct page
*page
)
315 int nid
= page_to_nid(page
);
316 list_add(&page
->lru
, &hugepage_freelists
[nid
]);
318 free_huge_pages_node
[nid
]++;
321 static struct page
*dequeue_huge_page(void)
324 struct page
*page
= NULL
;
326 for (nid
= 0; nid
< MAX_NUMNODES
; ++nid
) {
327 if (!list_empty(&hugepage_freelists
[nid
])) {
328 page
= list_entry(hugepage_freelists
[nid
].next
,
330 list_del(&page
->lru
);
332 free_huge_pages_node
[nid
]--;
339 static struct page
*dequeue_huge_page_vma(struct vm_area_struct
*vma
,
340 unsigned long address
, int avoid_reserve
)
343 struct page
*page
= NULL
;
344 struct mempolicy
*mpol
;
345 nodemask_t
*nodemask
;
346 struct zonelist
*zonelist
= huge_zonelist(vma
, address
,
347 htlb_alloc_mask
, &mpol
, &nodemask
);
352 * A child process with MAP_PRIVATE mappings created by their parent
353 * have no page reserves. This check ensures that reservations are
354 * not "stolen". The child may still get SIGKILLed
356 if (!vma_has_private_reserves(vma
) &&
357 free_huge_pages
- resv_huge_pages
== 0)
360 /* If reserves cannot be used, ensure enough pages are in the pool */
361 if (avoid_reserve
&& free_huge_pages
- resv_huge_pages
== 0)
364 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
365 MAX_NR_ZONES
- 1, nodemask
) {
366 nid
= zone_to_nid(zone
);
367 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
) &&
368 !list_empty(&hugepage_freelists
[nid
])) {
369 page
= list_entry(hugepage_freelists
[nid
].next
,
371 list_del(&page
->lru
);
373 free_huge_pages_node
[nid
]--;
376 decrement_hugepage_resv_vma(vma
);
385 static void update_and_free_page(struct page
*page
)
389 nr_huge_pages_node
[page_to_nid(page
)]--;
390 for (i
= 0; i
< (HPAGE_SIZE
/ PAGE_SIZE
); i
++) {
391 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
392 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
393 1 << PG_private
| 1<< PG_writeback
);
395 set_compound_page_dtor(page
, NULL
);
396 set_page_refcounted(page
);
397 arch_release_hugepage(page
);
398 __free_pages(page
, HUGETLB_PAGE_ORDER
);
401 static void free_huge_page(struct page
*page
)
403 int nid
= page_to_nid(page
);
404 struct address_space
*mapping
;
406 mapping
= (struct address_space
*) page_private(page
);
407 set_page_private(page
, 0);
408 BUG_ON(page_count(page
));
409 INIT_LIST_HEAD(&page
->lru
);
411 spin_lock(&hugetlb_lock
);
412 if (surplus_huge_pages_node
[nid
]) {
413 update_and_free_page(page
);
414 surplus_huge_pages
--;
415 surplus_huge_pages_node
[nid
]--;
417 enqueue_huge_page(page
);
419 spin_unlock(&hugetlb_lock
);
421 hugetlb_put_quota(mapping
, 1);
425 * Increment or decrement surplus_huge_pages. Keep node-specific counters
426 * balanced by operating on them in a round-robin fashion.
427 * Returns 1 if an adjustment was made.
429 static int adjust_pool_surplus(int delta
)
435 VM_BUG_ON(delta
!= -1 && delta
!= 1);
437 nid
= next_node(nid
, node_online_map
);
438 if (nid
== MAX_NUMNODES
)
439 nid
= first_node(node_online_map
);
441 /* To shrink on this node, there must be a surplus page */
442 if (delta
< 0 && !surplus_huge_pages_node
[nid
])
444 /* Surplus cannot exceed the total number of pages */
445 if (delta
> 0 && surplus_huge_pages_node
[nid
] >=
446 nr_huge_pages_node
[nid
])
449 surplus_huge_pages
+= delta
;
450 surplus_huge_pages_node
[nid
] += delta
;
453 } while (nid
!= prev_nid
);
459 static struct page
*alloc_fresh_huge_page_node(int nid
)
463 page
= alloc_pages_node(nid
,
464 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
465 __GFP_REPEAT
|__GFP_NOWARN
,
468 if (arch_prepare_hugepage(page
)) {
469 __free_pages(page
, HUGETLB_PAGE_ORDER
);
472 set_compound_page_dtor(page
, free_huge_page
);
473 spin_lock(&hugetlb_lock
);
475 nr_huge_pages_node
[nid
]++;
476 spin_unlock(&hugetlb_lock
);
477 put_page(page
); /* free it into the hugepage allocator */
483 static int alloc_fresh_huge_page(void)
490 start_nid
= hugetlb_next_nid
;
493 page
= alloc_fresh_huge_page_node(hugetlb_next_nid
);
497 * Use a helper variable to find the next node and then
498 * copy it back to hugetlb_next_nid afterwards:
499 * otherwise there's a window in which a racer might
500 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
501 * But we don't need to use a spin_lock here: it really
502 * doesn't matter if occasionally a racer chooses the
503 * same nid as we do. Move nid forward in the mask even
504 * if we just successfully allocated a hugepage so that
505 * the next caller gets hugepages on the next node.
507 next_nid
= next_node(hugetlb_next_nid
, node_online_map
);
508 if (next_nid
== MAX_NUMNODES
)
509 next_nid
= first_node(node_online_map
);
510 hugetlb_next_nid
= next_nid
;
511 } while (!page
&& hugetlb_next_nid
!= start_nid
);
514 count_vm_event(HTLB_BUDDY_PGALLOC
);
516 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
521 static struct page
*alloc_buddy_huge_page(struct vm_area_struct
*vma
,
522 unsigned long address
)
528 * Assume we will successfully allocate the surplus page to
529 * prevent racing processes from causing the surplus to exceed
532 * This however introduces a different race, where a process B
533 * tries to grow the static hugepage pool while alloc_pages() is
534 * called by process A. B will only examine the per-node
535 * counters in determining if surplus huge pages can be
536 * converted to normal huge pages in adjust_pool_surplus(). A
537 * won't be able to increment the per-node counter, until the
538 * lock is dropped by B, but B doesn't drop hugetlb_lock until
539 * no more huge pages can be converted from surplus to normal
540 * state (and doesn't try to convert again). Thus, we have a
541 * case where a surplus huge page exists, the pool is grown, and
542 * the surplus huge page still exists after, even though it
543 * should just have been converted to a normal huge page. This
544 * does not leak memory, though, as the hugepage will be freed
545 * once it is out of use. It also does not allow the counters to
546 * go out of whack in adjust_pool_surplus() as we don't modify
547 * the node values until we've gotten the hugepage and only the
548 * per-node value is checked there.
550 spin_lock(&hugetlb_lock
);
551 if (surplus_huge_pages
>= nr_overcommit_huge_pages
) {
552 spin_unlock(&hugetlb_lock
);
556 surplus_huge_pages
++;
558 spin_unlock(&hugetlb_lock
);
560 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
561 __GFP_REPEAT
|__GFP_NOWARN
,
564 spin_lock(&hugetlb_lock
);
567 * This page is now managed by the hugetlb allocator and has
568 * no users -- drop the buddy allocator's reference.
570 put_page_testzero(page
);
571 VM_BUG_ON(page_count(page
));
572 nid
= page_to_nid(page
);
573 set_compound_page_dtor(page
, free_huge_page
);
575 * We incremented the global counters already
577 nr_huge_pages_node
[nid
]++;
578 surplus_huge_pages_node
[nid
]++;
579 __count_vm_event(HTLB_BUDDY_PGALLOC
);
582 surplus_huge_pages
--;
583 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
585 spin_unlock(&hugetlb_lock
);
591 * Increase the hugetlb pool such that it can accomodate a reservation
594 static int gather_surplus_pages(int delta
)
596 struct list_head surplus_list
;
597 struct page
*page
, *tmp
;
599 int needed
, allocated
;
601 needed
= (resv_huge_pages
+ delta
) - free_huge_pages
;
603 resv_huge_pages
+= delta
;
608 INIT_LIST_HEAD(&surplus_list
);
612 spin_unlock(&hugetlb_lock
);
613 for (i
= 0; i
< needed
; i
++) {
614 page
= alloc_buddy_huge_page(NULL
, 0);
617 * We were not able to allocate enough pages to
618 * satisfy the entire reservation so we free what
619 * we've allocated so far.
621 spin_lock(&hugetlb_lock
);
626 list_add(&page
->lru
, &surplus_list
);
631 * After retaking hugetlb_lock, we need to recalculate 'needed'
632 * because either resv_huge_pages or free_huge_pages may have changed.
634 spin_lock(&hugetlb_lock
);
635 needed
= (resv_huge_pages
+ delta
) - (free_huge_pages
+ allocated
);
640 * The surplus_list now contains _at_least_ the number of extra pages
641 * needed to accomodate the reservation. Add the appropriate number
642 * of pages to the hugetlb pool and free the extras back to the buddy
643 * allocator. Commit the entire reservation here to prevent another
644 * process from stealing the pages as they are added to the pool but
645 * before they are reserved.
648 resv_huge_pages
+= delta
;
651 /* Free the needed pages to the hugetlb pool */
652 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
655 list_del(&page
->lru
);
656 enqueue_huge_page(page
);
659 /* Free unnecessary surplus pages to the buddy allocator */
660 if (!list_empty(&surplus_list
)) {
661 spin_unlock(&hugetlb_lock
);
662 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
663 list_del(&page
->lru
);
665 * The page has a reference count of zero already, so
666 * call free_huge_page directly instead of using
667 * put_page. This must be done with hugetlb_lock
668 * unlocked which is safe because free_huge_page takes
669 * hugetlb_lock before deciding how to free the page.
671 free_huge_page(page
);
673 spin_lock(&hugetlb_lock
);
680 * When releasing a hugetlb pool reservation, any surplus pages that were
681 * allocated to satisfy the reservation must be explicitly freed if they were
684 static void return_unused_surplus_pages(unsigned long unused_resv_pages
)
688 unsigned long nr_pages
;
691 * We want to release as many surplus pages as possible, spread
692 * evenly across all nodes. Iterate across all nodes until we
693 * can no longer free unreserved surplus pages. This occurs when
694 * the nodes with surplus pages have no free pages.
696 unsigned long remaining_iterations
= num_online_nodes();
698 /* Uncommit the reservation */
699 resv_huge_pages
-= unused_resv_pages
;
701 nr_pages
= min(unused_resv_pages
, surplus_huge_pages
);
703 while (remaining_iterations
-- && nr_pages
) {
704 nid
= next_node(nid
, node_online_map
);
705 if (nid
== MAX_NUMNODES
)
706 nid
= first_node(node_online_map
);
708 if (!surplus_huge_pages_node
[nid
])
711 if (!list_empty(&hugepage_freelists
[nid
])) {
712 page
= list_entry(hugepage_freelists
[nid
].next
,
714 list_del(&page
->lru
);
715 update_and_free_page(page
);
717 free_huge_pages_node
[nid
]--;
718 surplus_huge_pages
--;
719 surplus_huge_pages_node
[nid
]--;
721 remaining_iterations
= num_online_nodes();
727 * Determine if the huge page at addr within the vma has an associated
728 * reservation. Where it does not we will need to logically increase
729 * reservation and actually increase quota before an allocation can occur.
730 * Where any new reservation would be required the reservation change is
731 * prepared, but not committed. Once the page has been quota'd allocated
732 * an instantiated the change should be committed via vma_commit_reservation.
733 * No action is required on failure.
735 static int vma_needs_reservation(struct vm_area_struct
*vma
, unsigned long addr
)
737 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
738 struct inode
*inode
= mapping
->host
;
740 if (vma
->vm_flags
& VM_SHARED
) {
741 pgoff_t idx
= vma_pagecache_offset(vma
, addr
);
742 return region_chg(&inode
->i_mapping
->private_list
,
746 if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
752 static void vma_commit_reservation(struct vm_area_struct
*vma
,
755 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
756 struct inode
*inode
= mapping
->host
;
758 if (vma
->vm_flags
& VM_SHARED
) {
759 pgoff_t idx
= vma_pagecache_offset(vma
, addr
);
760 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
764 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
765 unsigned long addr
, int avoid_reserve
)
768 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
769 struct inode
*inode
= mapping
->host
;
773 * Processes that did not create the mapping will have no reserves and
774 * will not have accounted against quota. Check that the quota can be
775 * made before satisfying the allocation
776 * MAP_NORESERVE mappings may also need pages and quota allocated
777 * if no reserve mapping overlaps.
779 chg
= vma_needs_reservation(vma
, addr
);
783 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
784 return ERR_PTR(-ENOSPC
);
786 spin_lock(&hugetlb_lock
);
787 page
= dequeue_huge_page_vma(vma
, addr
, avoid_reserve
);
788 spin_unlock(&hugetlb_lock
);
791 page
= alloc_buddy_huge_page(vma
, addr
);
793 hugetlb_put_quota(inode
->i_mapping
, chg
);
794 return ERR_PTR(-VM_FAULT_OOM
);
798 set_page_refcounted(page
);
799 set_page_private(page
, (unsigned long) mapping
);
801 vma_commit_reservation(vma
, addr
);
806 static int __init
hugetlb_init(void)
810 if (HPAGE_SHIFT
== 0)
813 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
814 INIT_LIST_HEAD(&hugepage_freelists
[i
]);
816 hugetlb_next_nid
= first_node(node_online_map
);
818 for (i
= 0; i
< max_huge_pages
; ++i
) {
819 if (!alloc_fresh_huge_page())
822 max_huge_pages
= free_huge_pages
= nr_huge_pages
= i
;
823 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages
);
826 module_init(hugetlb_init
);
828 static int __init
hugetlb_setup(char *s
)
830 if (sscanf(s
, "%lu", &max_huge_pages
) <= 0)
834 __setup("hugepages=", hugetlb_setup
);
836 static unsigned int cpuset_mems_nr(unsigned int *array
)
841 for_each_node_mask(node
, cpuset_current_mems_allowed
)
848 #ifdef CONFIG_HIGHMEM
849 static void try_to_free_low(unsigned long count
)
853 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
854 struct page
*page
, *next
;
855 list_for_each_entry_safe(page
, next
, &hugepage_freelists
[i
], lru
) {
856 if (count
>= nr_huge_pages
)
858 if (PageHighMem(page
))
860 list_del(&page
->lru
);
861 update_and_free_page(page
);
863 free_huge_pages_node
[page_to_nid(page
)]--;
868 static inline void try_to_free_low(unsigned long count
)
873 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
874 static unsigned long set_max_huge_pages(unsigned long count
)
876 unsigned long min_count
, ret
;
879 * Increase the pool size
880 * First take pages out of surplus state. Then make up the
881 * remaining difference by allocating fresh huge pages.
883 * We might race with alloc_buddy_huge_page() here and be unable
884 * to convert a surplus huge page to a normal huge page. That is
885 * not critical, though, it just means the overall size of the
886 * pool might be one hugepage larger than it needs to be, but
887 * within all the constraints specified by the sysctls.
889 spin_lock(&hugetlb_lock
);
890 while (surplus_huge_pages
&& count
> persistent_huge_pages
) {
891 if (!adjust_pool_surplus(-1))
895 while (count
> persistent_huge_pages
) {
897 * If this allocation races such that we no longer need the
898 * page, free_huge_page will handle it by freeing the page
899 * and reducing the surplus.
901 spin_unlock(&hugetlb_lock
);
902 ret
= alloc_fresh_huge_page();
903 spin_lock(&hugetlb_lock
);
910 * Decrease the pool size
911 * First return free pages to the buddy allocator (being careful
912 * to keep enough around to satisfy reservations). Then place
913 * pages into surplus state as needed so the pool will shrink
914 * to the desired size as pages become free.
916 * By placing pages into the surplus state independent of the
917 * overcommit value, we are allowing the surplus pool size to
918 * exceed overcommit. There are few sane options here. Since
919 * alloc_buddy_huge_page() is checking the global counter,
920 * though, we'll note that we're not allowed to exceed surplus
921 * and won't grow the pool anywhere else. Not until one of the
922 * sysctls are changed, or the surplus pages go out of use.
924 min_count
= resv_huge_pages
+ nr_huge_pages
- free_huge_pages
;
925 min_count
= max(count
, min_count
);
926 try_to_free_low(min_count
);
927 while (min_count
< persistent_huge_pages
) {
928 struct page
*page
= dequeue_huge_page();
931 update_and_free_page(page
);
933 while (count
< persistent_huge_pages
) {
934 if (!adjust_pool_surplus(1))
938 ret
= persistent_huge_pages
;
939 spin_unlock(&hugetlb_lock
);
943 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
944 struct file
*file
, void __user
*buffer
,
945 size_t *length
, loff_t
*ppos
)
947 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
948 max_huge_pages
= set_max_huge_pages(max_huge_pages
);
952 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
953 struct file
*file
, void __user
*buffer
,
954 size_t *length
, loff_t
*ppos
)
956 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
957 if (hugepages_treat_as_movable
)
958 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
960 htlb_alloc_mask
= GFP_HIGHUSER
;
964 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
965 struct file
*file
, void __user
*buffer
,
966 size_t *length
, loff_t
*ppos
)
968 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
969 spin_lock(&hugetlb_lock
);
970 nr_overcommit_huge_pages
= sysctl_overcommit_huge_pages
;
971 spin_unlock(&hugetlb_lock
);
975 #endif /* CONFIG_SYSCTL */
977 int hugetlb_report_meminfo(char *buf
)
980 "HugePages_Total: %5lu\n"
981 "HugePages_Free: %5lu\n"
982 "HugePages_Rsvd: %5lu\n"
983 "HugePages_Surp: %5lu\n"
984 "Hugepagesize: %5lu kB\n",
992 int hugetlb_report_node_meminfo(int nid
, char *buf
)
995 "Node %d HugePages_Total: %5u\n"
996 "Node %d HugePages_Free: %5u\n"
997 "Node %d HugePages_Surp: %5u\n",
998 nid
, nr_huge_pages_node
[nid
],
999 nid
, free_huge_pages_node
[nid
],
1000 nid
, surplus_huge_pages_node
[nid
]);
1003 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1004 unsigned long hugetlb_total_pages(void)
1006 return nr_huge_pages
* (HPAGE_SIZE
/ PAGE_SIZE
);
1009 static int hugetlb_acct_memory(long delta
)
1013 spin_lock(&hugetlb_lock
);
1015 * When cpuset is configured, it breaks the strict hugetlb page
1016 * reservation as the accounting is done on a global variable. Such
1017 * reservation is completely rubbish in the presence of cpuset because
1018 * the reservation is not checked against page availability for the
1019 * current cpuset. Application can still potentially OOM'ed by kernel
1020 * with lack of free htlb page in cpuset that the task is in.
1021 * Attempt to enforce strict accounting with cpuset is almost
1022 * impossible (or too ugly) because cpuset is too fluid that
1023 * task or memory node can be dynamically moved between cpusets.
1025 * The change of semantics for shared hugetlb mapping with cpuset is
1026 * undesirable. However, in order to preserve some of the semantics,
1027 * we fall back to check against current free page availability as
1028 * a best attempt and hopefully to minimize the impact of changing
1029 * semantics that cpuset has.
1032 if (gather_surplus_pages(delta
) < 0)
1035 if (delta
> cpuset_mems_nr(free_huge_pages_node
)) {
1036 return_unused_surplus_pages(delta
);
1043 return_unused_surplus_pages((unsigned long) -delta
);
1046 spin_unlock(&hugetlb_lock
);
1050 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
1052 unsigned long reserve
= vma_resv_huge_pages(vma
);
1054 hugetlb_acct_memory(-reserve
);
1058 * We cannot handle pagefaults against hugetlb pages at all. They cause
1059 * handle_mm_fault() to try to instantiate regular-sized pages in the
1060 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1063 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1069 struct vm_operations_struct hugetlb_vm_ops
= {
1070 .fault
= hugetlb_vm_op_fault
,
1071 .close
= hugetlb_vm_op_close
,
1074 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
1081 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
1083 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
1085 entry
= pte_mkyoung(entry
);
1086 entry
= pte_mkhuge(entry
);
1091 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
1092 unsigned long address
, pte_t
*ptep
)
1096 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
1097 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
1098 update_mmu_cache(vma
, address
, entry
);
1103 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
1104 struct vm_area_struct
*vma
)
1106 pte_t
*src_pte
, *dst_pte
, entry
;
1107 struct page
*ptepage
;
1111 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
1113 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= HPAGE_SIZE
) {
1114 src_pte
= huge_pte_offset(src
, addr
);
1117 dst_pte
= huge_pte_alloc(dst
, addr
);
1121 /* If the pagetables are shared don't copy or take references */
1122 if (dst_pte
== src_pte
)
1125 spin_lock(&dst
->page_table_lock
);
1126 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
1127 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
1129 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
1130 entry
= huge_ptep_get(src_pte
);
1131 ptepage
= pte_page(entry
);
1133 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
1135 spin_unlock(&src
->page_table_lock
);
1136 spin_unlock(&dst
->page_table_lock
);
1144 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1145 unsigned long end
, struct page
*ref_page
)
1147 struct mm_struct
*mm
= vma
->vm_mm
;
1148 unsigned long address
;
1154 * A page gathering list, protected by per file i_mmap_lock. The
1155 * lock is used to avoid list corruption from multiple unmapping
1156 * of the same page since we are using page->lru.
1158 LIST_HEAD(page_list
);
1160 WARN_ON(!is_vm_hugetlb_page(vma
));
1161 BUG_ON(start
& ~HPAGE_MASK
);
1162 BUG_ON(end
& ~HPAGE_MASK
);
1164 spin_lock(&mm
->page_table_lock
);
1165 for (address
= start
; address
< end
; address
+= HPAGE_SIZE
) {
1166 ptep
= huge_pte_offset(mm
, address
);
1170 if (huge_pmd_unshare(mm
, &address
, ptep
))
1174 * If a reference page is supplied, it is because a specific
1175 * page is being unmapped, not a range. Ensure the page we
1176 * are about to unmap is the actual page of interest.
1179 pte
= huge_ptep_get(ptep
);
1180 if (huge_pte_none(pte
))
1182 page
= pte_page(pte
);
1183 if (page
!= ref_page
)
1187 * Mark the VMA as having unmapped its page so that
1188 * future faults in this VMA will fail rather than
1189 * looking like data was lost
1191 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
1194 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1195 if (huge_pte_none(pte
))
1198 page
= pte_page(pte
);
1200 set_page_dirty(page
);
1201 list_add(&page
->lru
, &page_list
);
1203 spin_unlock(&mm
->page_table_lock
);
1204 flush_tlb_range(vma
, start
, end
);
1205 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
1206 list_del(&page
->lru
);
1211 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1212 unsigned long end
, struct page
*ref_page
)
1215 * It is undesirable to test vma->vm_file as it should be non-null
1216 * for valid hugetlb area. However, vm_file will be NULL in the error
1217 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
1218 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
1219 * to clean up. Since no pte has actually been setup, it is safe to
1220 * do nothing in this case.
1223 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1224 __unmap_hugepage_range(vma
, start
, end
, ref_page
);
1225 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1230 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1231 * mappping it owns the reserve page for. The intention is to unmap the page
1232 * from other VMAs and let the children be SIGKILLed if they are faulting the
1235 int unmap_ref_private(struct mm_struct
*mm
,
1236 struct vm_area_struct
*vma
,
1238 unsigned long address
)
1240 struct vm_area_struct
*iter_vma
;
1241 struct address_space
*mapping
;
1242 struct prio_tree_iter iter
;
1246 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1247 * from page cache lookup which is in HPAGE_SIZE units.
1249 address
= address
& huge_page_mask(hstate_vma(vma
));
1250 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
)
1251 + (vma
->vm_pgoff
>> PAGE_SHIFT
);
1252 mapping
= (struct address_space
*)page_private(page
);
1254 vma_prio_tree_foreach(iter_vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1255 /* Do not unmap the current VMA */
1256 if (iter_vma
== vma
)
1260 * Unmap the page from other VMAs without their own reserves.
1261 * They get marked to be SIGKILLed if they fault in these
1262 * areas. This is because a future no-page fault on this VMA
1263 * could insert a zeroed page instead of the data existing
1264 * from the time of fork. This would look like data corruption
1266 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
1267 unmap_hugepage_range(iter_vma
,
1268 address
, address
+ HPAGE_SIZE
,
1275 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1276 unsigned long address
, pte_t
*ptep
, pte_t pte
,
1277 struct page
*pagecache_page
)
1279 struct page
*old_page
, *new_page
;
1281 int outside_reserve
= 0;
1283 old_page
= pte_page(pte
);
1286 /* If no-one else is actually using this page, avoid the copy
1287 * and just make the page writable */
1288 avoidcopy
= (page_count(old_page
) == 1);
1290 set_huge_ptep_writable(vma
, address
, ptep
);
1295 * If the process that created a MAP_PRIVATE mapping is about to
1296 * perform a COW due to a shared page count, attempt to satisfy
1297 * the allocation without using the existing reserves. The pagecache
1298 * page is used to determine if the reserve at this address was
1299 * consumed or not. If reserves were used, a partial faulted mapping
1300 * at the time of fork() could consume its reserves on COW instead
1301 * of the full address range.
1303 if (!(vma
->vm_flags
& VM_SHARED
) &&
1304 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
1305 old_page
!= pagecache_page
)
1306 outside_reserve
= 1;
1308 page_cache_get(old_page
);
1309 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
1311 if (IS_ERR(new_page
)) {
1312 page_cache_release(old_page
);
1315 * If a process owning a MAP_PRIVATE mapping fails to COW,
1316 * it is due to references held by a child and an insufficient
1317 * huge page pool. To guarantee the original mappers
1318 * reliability, unmap the page from child processes. The child
1319 * may get SIGKILLed if it later faults.
1321 if (outside_reserve
) {
1322 BUG_ON(huge_pte_none(pte
));
1323 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
1324 BUG_ON(page_count(old_page
) != 1);
1325 BUG_ON(huge_pte_none(pte
));
1326 goto retry_avoidcopy
;
1331 return -PTR_ERR(new_page
);
1334 spin_unlock(&mm
->page_table_lock
);
1335 copy_huge_page(new_page
, old_page
, address
, vma
);
1336 __SetPageUptodate(new_page
);
1337 spin_lock(&mm
->page_table_lock
);
1339 ptep
= huge_pte_offset(mm
, address
& HPAGE_MASK
);
1340 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
1342 huge_ptep_clear_flush(vma
, address
, ptep
);
1343 set_huge_pte_at(mm
, address
, ptep
,
1344 make_huge_pte(vma
, new_page
, 1));
1345 /* Make the old page be freed below */
1346 new_page
= old_page
;
1348 page_cache_release(new_page
);
1349 page_cache_release(old_page
);
1353 /* Return the pagecache page at a given address within a VMA */
1354 static struct page
*hugetlbfs_pagecache_page(struct vm_area_struct
*vma
,
1355 unsigned long address
)
1357 struct address_space
*mapping
;
1360 mapping
= vma
->vm_file
->f_mapping
;
1361 idx
= vma_pagecache_offset(vma
, address
);
1363 return find_lock_page(mapping
, idx
);
1366 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1367 unsigned long address
, pte_t
*ptep
, int write_access
)
1369 int ret
= VM_FAULT_SIGBUS
;
1373 struct address_space
*mapping
;
1377 * Currently, we are forced to kill the process in the event the
1378 * original mapper has unmapped pages from the child due to a failed
1379 * COW. Warn that such a situation has occured as it may not be obvious
1381 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
1383 "PID %d killed due to inadequate hugepage pool\n",
1388 mapping
= vma
->vm_file
->f_mapping
;
1389 idx
= vma_pagecache_offset(vma
, address
);
1392 * Use page lock to guard against racing truncation
1393 * before we get page_table_lock.
1396 page
= find_lock_page(mapping
, idx
);
1398 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
1401 page
= alloc_huge_page(vma
, address
, 0);
1403 ret
= -PTR_ERR(page
);
1406 clear_huge_page(page
, address
);
1407 __SetPageUptodate(page
);
1409 if (vma
->vm_flags
& VM_SHARED
) {
1411 struct inode
*inode
= mapping
->host
;
1413 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
1421 spin_lock(&inode
->i_lock
);
1422 inode
->i_blocks
+= BLOCKS_PER_HUGEPAGE
;
1423 spin_unlock(&inode
->i_lock
);
1428 spin_lock(&mm
->page_table_lock
);
1429 size
= i_size_read(mapping
->host
) >> HPAGE_SHIFT
;
1434 if (!huge_pte_none(huge_ptep_get(ptep
)))
1437 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
1438 && (vma
->vm_flags
& VM_SHARED
)));
1439 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
1441 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1442 /* Optimization, do the COW without a second fault */
1443 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
1446 spin_unlock(&mm
->page_table_lock
);
1452 spin_unlock(&mm
->page_table_lock
);
1458 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1459 unsigned long address
, int write_access
)
1464 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
1466 ptep
= huge_pte_alloc(mm
, address
);
1468 return VM_FAULT_OOM
;
1471 * Serialize hugepage allocation and instantiation, so that we don't
1472 * get spurious allocation failures if two CPUs race to instantiate
1473 * the same page in the page cache.
1475 mutex_lock(&hugetlb_instantiation_mutex
);
1476 entry
= huge_ptep_get(ptep
);
1477 if (huge_pte_none(entry
)) {
1478 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
1479 mutex_unlock(&hugetlb_instantiation_mutex
);
1485 spin_lock(&mm
->page_table_lock
);
1486 /* Check for a racing update before calling hugetlb_cow */
1487 if (likely(pte_same(entry
, huge_ptep_get(ptep
))))
1488 if (write_access
&& !pte_write(entry
)) {
1490 page
= hugetlbfs_pagecache_page(vma
, address
);
1491 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
, page
);
1497 spin_unlock(&mm
->page_table_lock
);
1498 mutex_unlock(&hugetlb_instantiation_mutex
);
1503 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1504 struct page
**pages
, struct vm_area_struct
**vmas
,
1505 unsigned long *position
, int *length
, int i
,
1508 unsigned long pfn_offset
;
1509 unsigned long vaddr
= *position
;
1510 int remainder
= *length
;
1512 spin_lock(&mm
->page_table_lock
);
1513 while (vaddr
< vma
->vm_end
&& remainder
) {
1518 * Some archs (sparc64, sh*) have multiple pte_ts to
1519 * each hugepage. We have to make * sure we get the
1520 * first, for the page indexing below to work.
1522 pte
= huge_pte_offset(mm
, vaddr
& HPAGE_MASK
);
1524 if (!pte
|| huge_pte_none(huge_ptep_get(pte
)) ||
1525 (write
&& !pte_write(huge_ptep_get(pte
)))) {
1528 spin_unlock(&mm
->page_table_lock
);
1529 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
1530 spin_lock(&mm
->page_table_lock
);
1531 if (!(ret
& VM_FAULT_ERROR
))
1540 pfn_offset
= (vaddr
& ~HPAGE_MASK
) >> PAGE_SHIFT
;
1541 page
= pte_page(huge_ptep_get(pte
));
1545 pages
[i
] = page
+ pfn_offset
;
1555 if (vaddr
< vma
->vm_end
&& remainder
&&
1556 pfn_offset
< HPAGE_SIZE
/PAGE_SIZE
) {
1558 * We use pfn_offset to avoid touching the pageframes
1559 * of this compound page.
1564 spin_unlock(&mm
->page_table_lock
);
1565 *length
= remainder
;
1571 void hugetlb_change_protection(struct vm_area_struct
*vma
,
1572 unsigned long address
, unsigned long end
, pgprot_t newprot
)
1574 struct mm_struct
*mm
= vma
->vm_mm
;
1575 unsigned long start
= address
;
1579 BUG_ON(address
>= end
);
1580 flush_cache_range(vma
, address
, end
);
1582 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1583 spin_lock(&mm
->page_table_lock
);
1584 for (; address
< end
; address
+= HPAGE_SIZE
) {
1585 ptep
= huge_pte_offset(mm
, address
);
1588 if (huge_pmd_unshare(mm
, &address
, ptep
))
1590 if (!huge_pte_none(huge_ptep_get(ptep
))) {
1591 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1592 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
1593 set_huge_pte_at(mm
, address
, ptep
, pte
);
1596 spin_unlock(&mm
->page_table_lock
);
1597 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1599 flush_tlb_range(vma
, start
, end
);
1602 int hugetlb_reserve_pages(struct inode
*inode
,
1604 struct vm_area_struct
*vma
)
1608 if (vma
&& vma
->vm_flags
& VM_NORESERVE
)
1612 * Shared mappings base their reservation on the number of pages that
1613 * are already allocated on behalf of the file. Private mappings need
1614 * to reserve the full area even if read-only as mprotect() may be
1615 * called to make the mapping read-write. Assume !vma is a shm mapping
1617 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
1618 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
1621 set_vma_resv_huge_pages(vma
, chg
);
1622 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
1628 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
1630 ret
= hugetlb_acct_memory(chg
);
1632 hugetlb_put_quota(inode
->i_mapping
, chg
);
1635 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
1636 region_add(&inode
->i_mapping
->private_list
, from
, to
);
1640 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
1642 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
1644 spin_lock(&inode
->i_lock
);
1645 inode
->i_blocks
-= BLOCKS_PER_HUGEPAGE
* freed
;
1646 spin_unlock(&inode
->i_lock
);
1648 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
1649 hugetlb_acct_memory(-(chg
- freed
));