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>
17 #include <linux/bootmem.h>
18 #include <linux/sysfs.h>
21 #include <asm/pgtable.h>
23 #include <linux/hugetlb.h>
26 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
27 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
28 unsigned long hugepages_treat_as_movable
;
30 static int max_hstate
;
31 unsigned int default_hstate_idx
;
32 struct hstate hstates
[HUGE_MAX_HSTATE
];
34 /* for command line parsing */
35 static struct hstate
* __initdata parsed_hstate
;
36 static unsigned long __initdata default_hstate_max_huge_pages
;
37 static unsigned long __initdata default_hstate_size
;
39 #define for_each_hstate(h) \
40 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
43 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
45 static DEFINE_SPINLOCK(hugetlb_lock
);
48 * Region tracking -- allows tracking of reservations and instantiated pages
49 * across the pages in a mapping.
51 * The region data structures are protected by a combination of the mmap_sem
52 * and the hugetlb_instantion_mutex. To access or modify a region the caller
53 * must either hold the mmap_sem for write, or the mmap_sem for read and
54 * the hugetlb_instantiation mutex:
56 * down_write(&mm->mmap_sem);
58 * down_read(&mm->mmap_sem);
59 * mutex_lock(&hugetlb_instantiation_mutex);
62 struct list_head link
;
67 static long region_add(struct list_head
*head
, long f
, long t
)
69 struct file_region
*rg
, *nrg
, *trg
;
71 /* Locate the region we are either in or before. */
72 list_for_each_entry(rg
, head
, link
)
76 /* Round our left edge to the current segment if it encloses us. */
80 /* Check for and consume any regions we now overlap with. */
82 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
83 if (&rg
->link
== head
)
88 /* If this area reaches higher then extend our area to
89 * include it completely. If this is not the first area
90 * which we intend to reuse, free it. */
103 static long region_chg(struct list_head
*head
, long f
, long t
)
105 struct file_region
*rg
, *nrg
;
108 /* Locate the region we are before or in. */
109 list_for_each_entry(rg
, head
, link
)
113 /* If we are below the current region then a new region is required.
114 * Subtle, allocate a new region at the position but make it zero
115 * size such that we can guarantee to record the reservation. */
116 if (&rg
->link
== head
|| t
< rg
->from
) {
117 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
122 INIT_LIST_HEAD(&nrg
->link
);
123 list_add(&nrg
->link
, rg
->link
.prev
);
128 /* Round our left edge to the current segment if it encloses us. */
133 /* Check for and consume any regions we now overlap with. */
134 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
135 if (&rg
->link
== head
)
140 /* We overlap with this area, if it extends futher than
141 * us then we must extend ourselves. Account for its
142 * existing reservation. */
147 chg
-= rg
->to
- rg
->from
;
152 static long region_truncate(struct list_head
*head
, long end
)
154 struct file_region
*rg
, *trg
;
157 /* Locate the region we are either in or before. */
158 list_for_each_entry(rg
, head
, link
)
161 if (&rg
->link
== head
)
164 /* If we are in the middle of a region then adjust it. */
165 if (end
> rg
->from
) {
168 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
171 /* Drop any remaining regions. */
172 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
173 if (&rg
->link
== head
)
175 chg
+= rg
->to
- rg
->from
;
182 static long region_count(struct list_head
*head
, long f
, long t
)
184 struct file_region
*rg
;
187 /* Locate each segment we overlap with, and count that overlap. */
188 list_for_each_entry(rg
, head
, link
) {
197 seg_from
= max(rg
->from
, f
);
198 seg_to
= min(rg
->to
, t
);
200 chg
+= seg_to
- seg_from
;
207 * Convert the address within this vma to the page offset within
208 * the mapping, in pagecache page units; huge pages here.
210 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
211 struct vm_area_struct
*vma
, unsigned long address
)
213 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
214 (vma
->vm_pgoff
>> huge_page_order(h
));
218 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
219 * bits of the reservation map pointer, which are always clear due to
222 #define HPAGE_RESV_OWNER (1UL << 0)
223 #define HPAGE_RESV_UNMAPPED (1UL << 1)
224 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
227 * These helpers are used to track how many pages are reserved for
228 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
229 * is guaranteed to have their future faults succeed.
231 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
232 * the reserve counters are updated with the hugetlb_lock held. It is safe
233 * to reset the VMA at fork() time as it is not in use yet and there is no
234 * chance of the global counters getting corrupted as a result of the values.
236 * The private mapping reservation is represented in a subtly different
237 * manner to a shared mapping. A shared mapping has a region map associated
238 * with the underlying file, this region map represents the backing file
239 * pages which have ever had a reservation assigned which this persists even
240 * after the page is instantiated. A private mapping has a region map
241 * associated with the original mmap which is attached to all VMAs which
242 * reference it, this region map represents those offsets which have consumed
243 * reservation ie. where pages have been instantiated.
245 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
247 return (unsigned long)vma
->vm_private_data
;
250 static void set_vma_private_data(struct vm_area_struct
*vma
,
253 vma
->vm_private_data
= (void *)value
;
258 struct list_head regions
;
261 struct resv_map
*resv_map_alloc(void)
263 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
267 kref_init(&resv_map
->refs
);
268 INIT_LIST_HEAD(&resv_map
->regions
);
273 void resv_map_release(struct kref
*ref
)
275 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
277 /* Clear out any active regions before we release the map. */
278 region_truncate(&resv_map
->regions
, 0);
282 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
284 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
285 if (!(vma
->vm_flags
& VM_SHARED
))
286 return (struct resv_map
*)(get_vma_private_data(vma
) &
291 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
293 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
294 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
296 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
297 HPAGE_RESV_MASK
) | (unsigned long)map
);
300 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
302 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
303 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
305 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
308 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
310 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
312 return (get_vma_private_data(vma
) & flag
) != 0;
315 /* Decrement the reserved pages in the hugepage pool by one */
316 static void decrement_hugepage_resv_vma(struct hstate
*h
,
317 struct vm_area_struct
*vma
)
319 if (vma
->vm_flags
& VM_NORESERVE
)
322 if (vma
->vm_flags
& VM_SHARED
) {
323 /* Shared mappings always use reserves */
324 h
->resv_huge_pages
--;
325 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
327 * Only the process that called mmap() has reserves for
330 h
->resv_huge_pages
--;
334 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
335 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
337 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
338 if (!(vma
->vm_flags
& VM_SHARED
))
339 vma
->vm_private_data
= (void *)0;
342 /* Returns true if the VMA has associated reserve pages */
343 static int vma_has_private_reserves(struct vm_area_struct
*vma
)
345 if (vma
->vm_flags
& VM_SHARED
)
347 if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
352 static void clear_huge_page(struct page
*page
,
353 unsigned long addr
, unsigned long sz
)
358 for (i
= 0; i
< sz
/PAGE_SIZE
; i
++) {
360 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
364 static void copy_huge_page(struct page
*dst
, struct page
*src
,
365 unsigned long addr
, struct vm_area_struct
*vma
)
368 struct hstate
*h
= hstate_vma(vma
);
371 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
373 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
377 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
379 int nid
= page_to_nid(page
);
380 list_add(&page
->lru
, &h
->hugepage_freelists
[nid
]);
381 h
->free_huge_pages
++;
382 h
->free_huge_pages_node
[nid
]++;
385 static struct page
*dequeue_huge_page(struct hstate
*h
)
388 struct page
*page
= NULL
;
390 for (nid
= 0; nid
< MAX_NUMNODES
; ++nid
) {
391 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
392 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
394 list_del(&page
->lru
);
395 h
->free_huge_pages
--;
396 h
->free_huge_pages_node
[nid
]--;
403 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
404 struct vm_area_struct
*vma
,
405 unsigned long address
, int avoid_reserve
)
408 struct page
*page
= NULL
;
409 struct mempolicy
*mpol
;
410 nodemask_t
*nodemask
;
411 struct zonelist
*zonelist
= huge_zonelist(vma
, address
,
412 htlb_alloc_mask
, &mpol
, &nodemask
);
417 * A child process with MAP_PRIVATE mappings created by their parent
418 * have no page reserves. This check ensures that reservations are
419 * not "stolen". The child may still get SIGKILLed
421 if (!vma_has_private_reserves(vma
) &&
422 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
425 /* If reserves cannot be used, ensure enough pages are in the pool */
426 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
429 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
430 MAX_NR_ZONES
- 1, nodemask
) {
431 nid
= zone_to_nid(zone
);
432 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
) &&
433 !list_empty(&h
->hugepage_freelists
[nid
])) {
434 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
436 list_del(&page
->lru
);
437 h
->free_huge_pages
--;
438 h
->free_huge_pages_node
[nid
]--;
441 decrement_hugepage_resv_vma(h
, vma
);
450 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
455 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
456 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
457 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
458 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
459 1 << PG_private
| 1<< PG_writeback
);
461 set_compound_page_dtor(page
, NULL
);
462 set_page_refcounted(page
);
463 arch_release_hugepage(page
);
464 __free_pages(page
, huge_page_order(h
));
467 struct hstate
*size_to_hstate(unsigned long size
)
472 if (huge_page_size(h
) == size
)
478 static void free_huge_page(struct page
*page
)
481 * Can't pass hstate in here because it is called from the
482 * compound page destructor.
484 struct hstate
*h
= page_hstate(page
);
485 int nid
= page_to_nid(page
);
486 struct address_space
*mapping
;
488 mapping
= (struct address_space
*) page_private(page
);
489 set_page_private(page
, 0);
490 BUG_ON(page_count(page
));
491 INIT_LIST_HEAD(&page
->lru
);
493 spin_lock(&hugetlb_lock
);
494 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
495 update_and_free_page(h
, page
);
496 h
->surplus_huge_pages
--;
497 h
->surplus_huge_pages_node
[nid
]--;
499 enqueue_huge_page(h
, page
);
501 spin_unlock(&hugetlb_lock
);
503 hugetlb_put_quota(mapping
, 1);
507 * Increment or decrement surplus_huge_pages. Keep node-specific counters
508 * balanced by operating on them in a round-robin fashion.
509 * Returns 1 if an adjustment was made.
511 static int adjust_pool_surplus(struct hstate
*h
, int delta
)
517 VM_BUG_ON(delta
!= -1 && delta
!= 1);
519 nid
= next_node(nid
, node_online_map
);
520 if (nid
== MAX_NUMNODES
)
521 nid
= first_node(node_online_map
);
523 /* To shrink on this node, there must be a surplus page */
524 if (delta
< 0 && !h
->surplus_huge_pages_node
[nid
])
526 /* Surplus cannot exceed the total number of pages */
527 if (delta
> 0 && h
->surplus_huge_pages_node
[nid
] >=
528 h
->nr_huge_pages_node
[nid
])
531 h
->surplus_huge_pages
+= delta
;
532 h
->surplus_huge_pages_node
[nid
] += delta
;
535 } while (nid
!= prev_nid
);
541 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
543 set_compound_page_dtor(page
, free_huge_page
);
544 spin_lock(&hugetlb_lock
);
546 h
->nr_huge_pages_node
[nid
]++;
547 spin_unlock(&hugetlb_lock
);
548 put_page(page
); /* free it into the hugepage allocator */
551 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
555 if (h
->order
>= MAX_ORDER
)
558 page
= alloc_pages_node(nid
,
559 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
560 __GFP_REPEAT
|__GFP_NOWARN
,
563 if (arch_prepare_hugepage(page
)) {
564 __free_pages(page
, HUGETLB_PAGE_ORDER
);
567 prep_new_huge_page(h
, page
, nid
);
574 * Use a helper variable to find the next node and then
575 * copy it back to hugetlb_next_nid afterwards:
576 * otherwise there's a window in which a racer might
577 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
578 * But we don't need to use a spin_lock here: it really
579 * doesn't matter if occasionally a racer chooses the
580 * same nid as we do. Move nid forward in the mask even
581 * if we just successfully allocated a hugepage so that
582 * the next caller gets hugepages on the next node.
584 static int hstate_next_node(struct hstate
*h
)
587 next_nid
= next_node(h
->hugetlb_next_nid
, node_online_map
);
588 if (next_nid
== MAX_NUMNODES
)
589 next_nid
= first_node(node_online_map
);
590 h
->hugetlb_next_nid
= next_nid
;
594 static int alloc_fresh_huge_page(struct hstate
*h
)
601 start_nid
= h
->hugetlb_next_nid
;
604 page
= alloc_fresh_huge_page_node(h
, h
->hugetlb_next_nid
);
607 next_nid
= hstate_next_node(h
);
608 } while (!page
&& h
->hugetlb_next_nid
!= start_nid
);
611 count_vm_event(HTLB_BUDDY_PGALLOC
);
613 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
618 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
619 struct vm_area_struct
*vma
, unsigned long address
)
624 if (h
->order
>= MAX_ORDER
)
628 * Assume we will successfully allocate the surplus page to
629 * prevent racing processes from causing the surplus to exceed
632 * This however introduces a different race, where a process B
633 * tries to grow the static hugepage pool while alloc_pages() is
634 * called by process A. B will only examine the per-node
635 * counters in determining if surplus huge pages can be
636 * converted to normal huge pages in adjust_pool_surplus(). A
637 * won't be able to increment the per-node counter, until the
638 * lock is dropped by B, but B doesn't drop hugetlb_lock until
639 * no more huge pages can be converted from surplus to normal
640 * state (and doesn't try to convert again). Thus, we have a
641 * case where a surplus huge page exists, the pool is grown, and
642 * the surplus huge page still exists after, even though it
643 * should just have been converted to a normal huge page. This
644 * does not leak memory, though, as the hugepage will be freed
645 * once it is out of use. It also does not allow the counters to
646 * go out of whack in adjust_pool_surplus() as we don't modify
647 * the node values until we've gotten the hugepage and only the
648 * per-node value is checked there.
650 spin_lock(&hugetlb_lock
);
651 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
652 spin_unlock(&hugetlb_lock
);
656 h
->surplus_huge_pages
++;
658 spin_unlock(&hugetlb_lock
);
660 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
661 __GFP_REPEAT
|__GFP_NOWARN
,
664 spin_lock(&hugetlb_lock
);
667 * This page is now managed by the hugetlb allocator and has
668 * no users -- drop the buddy allocator's reference.
670 put_page_testzero(page
);
671 VM_BUG_ON(page_count(page
));
672 nid
= page_to_nid(page
);
673 set_compound_page_dtor(page
, free_huge_page
);
675 * We incremented the global counters already
677 h
->nr_huge_pages_node
[nid
]++;
678 h
->surplus_huge_pages_node
[nid
]++;
679 __count_vm_event(HTLB_BUDDY_PGALLOC
);
682 h
->surplus_huge_pages
--;
683 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
685 spin_unlock(&hugetlb_lock
);
691 * Increase the hugetlb pool such that it can accomodate a reservation
694 static int gather_surplus_pages(struct hstate
*h
, int delta
)
696 struct list_head surplus_list
;
697 struct page
*page
, *tmp
;
699 int needed
, allocated
;
701 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
703 h
->resv_huge_pages
+= delta
;
708 INIT_LIST_HEAD(&surplus_list
);
712 spin_unlock(&hugetlb_lock
);
713 for (i
= 0; i
< needed
; i
++) {
714 page
= alloc_buddy_huge_page(h
, NULL
, 0);
717 * We were not able to allocate enough pages to
718 * satisfy the entire reservation so we free what
719 * we've allocated so far.
721 spin_lock(&hugetlb_lock
);
726 list_add(&page
->lru
, &surplus_list
);
731 * After retaking hugetlb_lock, we need to recalculate 'needed'
732 * because either resv_huge_pages or free_huge_pages may have changed.
734 spin_lock(&hugetlb_lock
);
735 needed
= (h
->resv_huge_pages
+ delta
) -
736 (h
->free_huge_pages
+ allocated
);
741 * The surplus_list now contains _at_least_ the number of extra pages
742 * needed to accomodate the reservation. Add the appropriate number
743 * of pages to the hugetlb pool and free the extras back to the buddy
744 * allocator. Commit the entire reservation here to prevent another
745 * process from stealing the pages as they are added to the pool but
746 * before they are reserved.
749 h
->resv_huge_pages
+= delta
;
752 /* Free the needed pages to the hugetlb pool */
753 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
756 list_del(&page
->lru
);
757 enqueue_huge_page(h
, page
);
760 /* Free unnecessary surplus pages to the buddy allocator */
761 if (!list_empty(&surplus_list
)) {
762 spin_unlock(&hugetlb_lock
);
763 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
764 list_del(&page
->lru
);
766 * The page has a reference count of zero already, so
767 * call free_huge_page directly instead of using
768 * put_page. This must be done with hugetlb_lock
769 * unlocked which is safe because free_huge_page takes
770 * hugetlb_lock before deciding how to free the page.
772 free_huge_page(page
);
774 spin_lock(&hugetlb_lock
);
781 * When releasing a hugetlb pool reservation, any surplus pages that were
782 * allocated to satisfy the reservation must be explicitly freed if they were
785 static void return_unused_surplus_pages(struct hstate
*h
,
786 unsigned long unused_resv_pages
)
790 unsigned long nr_pages
;
793 * We want to release as many surplus pages as possible, spread
794 * evenly across all nodes. Iterate across all nodes until we
795 * can no longer free unreserved surplus pages. This occurs when
796 * the nodes with surplus pages have no free pages.
798 unsigned long remaining_iterations
= num_online_nodes();
800 /* Uncommit the reservation */
801 h
->resv_huge_pages
-= unused_resv_pages
;
803 /* Cannot return gigantic pages currently */
804 if (h
->order
>= MAX_ORDER
)
807 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
809 while (remaining_iterations
-- && nr_pages
) {
810 nid
= next_node(nid
, node_online_map
);
811 if (nid
== MAX_NUMNODES
)
812 nid
= first_node(node_online_map
);
814 if (!h
->surplus_huge_pages_node
[nid
])
817 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
818 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
820 list_del(&page
->lru
);
821 update_and_free_page(h
, page
);
822 h
->free_huge_pages
--;
823 h
->free_huge_pages_node
[nid
]--;
824 h
->surplus_huge_pages
--;
825 h
->surplus_huge_pages_node
[nid
]--;
827 remaining_iterations
= num_online_nodes();
833 * Determine if the huge page at addr within the vma has an associated
834 * reservation. Where it does not we will need to logically increase
835 * reservation and actually increase quota before an allocation can occur.
836 * Where any new reservation would be required the reservation change is
837 * prepared, but not committed. Once the page has been quota'd allocated
838 * an instantiated the change should be committed via vma_commit_reservation.
839 * No action is required on failure.
841 static int vma_needs_reservation(struct hstate
*h
,
842 struct vm_area_struct
*vma
, unsigned long addr
)
844 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
845 struct inode
*inode
= mapping
->host
;
847 if (vma
->vm_flags
& VM_SHARED
) {
848 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
849 return region_chg(&inode
->i_mapping
->private_list
,
852 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
857 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
858 struct resv_map
*reservations
= vma_resv_map(vma
);
860 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
866 static void vma_commit_reservation(struct hstate
*h
,
867 struct vm_area_struct
*vma
, unsigned long addr
)
869 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
870 struct inode
*inode
= mapping
->host
;
872 if (vma
->vm_flags
& VM_SHARED
) {
873 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
874 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
876 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
877 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
878 struct resv_map
*reservations
= vma_resv_map(vma
);
880 /* Mark this page used in the map. */
881 region_add(&reservations
->regions
, idx
, idx
+ 1);
885 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
886 unsigned long addr
, int avoid_reserve
)
888 struct hstate
*h
= hstate_vma(vma
);
890 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
891 struct inode
*inode
= mapping
->host
;
895 * Processes that did not create the mapping will have no reserves and
896 * will not have accounted against quota. Check that the quota can be
897 * made before satisfying the allocation
898 * MAP_NORESERVE mappings may also need pages and quota allocated
899 * if no reserve mapping overlaps.
901 chg
= vma_needs_reservation(h
, vma
, addr
);
905 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
906 return ERR_PTR(-ENOSPC
);
908 spin_lock(&hugetlb_lock
);
909 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
910 spin_unlock(&hugetlb_lock
);
913 page
= alloc_buddy_huge_page(h
, vma
, addr
);
915 hugetlb_put_quota(inode
->i_mapping
, chg
);
916 return ERR_PTR(-VM_FAULT_OOM
);
920 set_page_refcounted(page
);
921 set_page_private(page
, (unsigned long) mapping
);
923 vma_commit_reservation(h
, vma
, addr
);
928 static __initdata
LIST_HEAD(huge_boot_pages
);
930 struct huge_bootmem_page
{
931 struct list_head list
;
932 struct hstate
*hstate
;
935 static int __init
alloc_bootmem_huge_page(struct hstate
*h
)
937 struct huge_bootmem_page
*m
;
938 int nr_nodes
= nodes_weight(node_online_map
);
943 addr
= __alloc_bootmem_node_nopanic(
944 NODE_DATA(h
->hugetlb_next_nid
),
945 huge_page_size(h
), huge_page_size(h
), 0);
949 * Use the beginning of the huge page to store the
950 * huge_bootmem_page struct (until gather_bootmem
951 * puts them into the mem_map).
963 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
964 /* Put them into a private list first because mem_map is not up yet */
965 list_add(&m
->list
, &huge_boot_pages
);
970 /* Put bootmem huge pages into the standard lists after mem_map is up */
971 static void __init
gather_bootmem_prealloc(void)
973 struct huge_bootmem_page
*m
;
975 list_for_each_entry(m
, &huge_boot_pages
, list
) {
976 struct page
*page
= virt_to_page(m
);
977 struct hstate
*h
= m
->hstate
;
978 __ClearPageReserved(page
);
979 WARN_ON(page_count(page
) != 1);
980 prep_compound_page(page
, h
->order
);
981 prep_new_huge_page(h
, page
, page_to_nid(page
));
985 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
989 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
990 if (h
->order
>= MAX_ORDER
) {
991 if (!alloc_bootmem_huge_page(h
))
993 } else if (!alloc_fresh_huge_page(h
))
996 h
->max_huge_pages
= i
;
999 static void __init
hugetlb_init_hstates(void)
1003 for_each_hstate(h
) {
1004 /* oversize hugepages were init'ed in early boot */
1005 if (h
->order
< MAX_ORDER
)
1006 hugetlb_hstate_alloc_pages(h
);
1010 static char * __init
memfmt(char *buf
, unsigned long n
)
1012 if (n
>= (1UL << 30))
1013 sprintf(buf
, "%lu GB", n
>> 30);
1014 else if (n
>= (1UL << 20))
1015 sprintf(buf
, "%lu MB", n
>> 20);
1017 sprintf(buf
, "%lu KB", n
>> 10);
1021 static void __init
report_hugepages(void)
1025 for_each_hstate(h
) {
1027 printk(KERN_INFO
"HugeTLB registered %s page size, "
1028 "pre-allocated %ld pages\n",
1029 memfmt(buf
, huge_page_size(h
)),
1030 h
->free_huge_pages
);
1034 #ifdef CONFIG_SYSCTL
1035 #ifdef CONFIG_HIGHMEM
1036 static void try_to_free_low(struct hstate
*h
, unsigned long count
)
1040 if (h
->order
>= MAX_ORDER
)
1043 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
1044 struct page
*page
, *next
;
1045 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1046 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1047 if (count
>= h
->nr_huge_pages
)
1049 if (PageHighMem(page
))
1051 list_del(&page
->lru
);
1052 update_and_free_page(h
, page
);
1053 h
->free_huge_pages
--;
1054 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1059 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
)
1064 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1065 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
)
1067 unsigned long min_count
, ret
;
1069 if (h
->order
>= MAX_ORDER
)
1070 return h
->max_huge_pages
;
1073 * Increase the pool size
1074 * First take pages out of surplus state. Then make up the
1075 * remaining difference by allocating fresh huge pages.
1077 * We might race with alloc_buddy_huge_page() here and be unable
1078 * to convert a surplus huge page to a normal huge page. That is
1079 * not critical, though, it just means the overall size of the
1080 * pool might be one hugepage larger than it needs to be, but
1081 * within all the constraints specified by the sysctls.
1083 spin_lock(&hugetlb_lock
);
1084 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1085 if (!adjust_pool_surplus(h
, -1))
1089 while (count
> persistent_huge_pages(h
)) {
1091 * If this allocation races such that we no longer need the
1092 * page, free_huge_page will handle it by freeing the page
1093 * and reducing the surplus.
1095 spin_unlock(&hugetlb_lock
);
1096 ret
= alloc_fresh_huge_page(h
);
1097 spin_lock(&hugetlb_lock
);
1104 * Decrease the pool size
1105 * First return free pages to the buddy allocator (being careful
1106 * to keep enough around to satisfy reservations). Then place
1107 * pages into surplus state as needed so the pool will shrink
1108 * to the desired size as pages become free.
1110 * By placing pages into the surplus state independent of the
1111 * overcommit value, we are allowing the surplus pool size to
1112 * exceed overcommit. There are few sane options here. Since
1113 * alloc_buddy_huge_page() is checking the global counter,
1114 * though, we'll note that we're not allowed to exceed surplus
1115 * and won't grow the pool anywhere else. Not until one of the
1116 * sysctls are changed, or the surplus pages go out of use.
1118 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1119 min_count
= max(count
, min_count
);
1120 try_to_free_low(h
, min_count
);
1121 while (min_count
< persistent_huge_pages(h
)) {
1122 struct page
*page
= dequeue_huge_page(h
);
1125 update_and_free_page(h
, page
);
1127 while (count
< persistent_huge_pages(h
)) {
1128 if (!adjust_pool_surplus(h
, 1))
1132 ret
= persistent_huge_pages(h
);
1133 spin_unlock(&hugetlb_lock
);
1137 #define HSTATE_ATTR_RO(_name) \
1138 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1140 #define HSTATE_ATTR(_name) \
1141 static struct kobj_attribute _name##_attr = \
1142 __ATTR(_name, 0644, _name##_show, _name##_store)
1144 static struct kobject
*hugepages_kobj
;
1145 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1147 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
)
1150 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1151 if (hstate_kobjs
[i
] == kobj
)
1157 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1158 struct kobj_attribute
*attr
, char *buf
)
1160 struct hstate
*h
= kobj_to_hstate(kobj
);
1161 return sprintf(buf
, "%lu\n", h
->nr_huge_pages
);
1163 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1164 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1167 unsigned long input
;
1168 struct hstate
*h
= kobj_to_hstate(kobj
);
1170 err
= strict_strtoul(buf
, 10, &input
);
1174 h
->max_huge_pages
= set_max_huge_pages(h
, input
);
1178 HSTATE_ATTR(nr_hugepages
);
1180 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1181 struct kobj_attribute
*attr
, char *buf
)
1183 struct hstate
*h
= kobj_to_hstate(kobj
);
1184 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1186 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1187 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1190 unsigned long input
;
1191 struct hstate
*h
= kobj_to_hstate(kobj
);
1193 err
= strict_strtoul(buf
, 10, &input
);
1197 spin_lock(&hugetlb_lock
);
1198 h
->nr_overcommit_huge_pages
= input
;
1199 spin_unlock(&hugetlb_lock
);
1203 HSTATE_ATTR(nr_overcommit_hugepages
);
1205 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1206 struct kobj_attribute
*attr
, char *buf
)
1208 struct hstate
*h
= kobj_to_hstate(kobj
);
1209 return sprintf(buf
, "%lu\n", h
->free_huge_pages
);
1211 HSTATE_ATTR_RO(free_hugepages
);
1213 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1214 struct kobj_attribute
*attr
, char *buf
)
1216 struct hstate
*h
= kobj_to_hstate(kobj
);
1217 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1219 HSTATE_ATTR_RO(resv_hugepages
);
1221 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1222 struct kobj_attribute
*attr
, char *buf
)
1224 struct hstate
*h
= kobj_to_hstate(kobj
);
1225 return sprintf(buf
, "%lu\n", h
->surplus_huge_pages
);
1227 HSTATE_ATTR_RO(surplus_hugepages
);
1229 static struct attribute
*hstate_attrs
[] = {
1230 &nr_hugepages_attr
.attr
,
1231 &nr_overcommit_hugepages_attr
.attr
,
1232 &free_hugepages_attr
.attr
,
1233 &resv_hugepages_attr
.attr
,
1234 &surplus_hugepages_attr
.attr
,
1238 static struct attribute_group hstate_attr_group
= {
1239 .attrs
= hstate_attrs
,
1242 static int __init
hugetlb_sysfs_add_hstate(struct hstate
*h
)
1246 hstate_kobjs
[h
- hstates
] = kobject_create_and_add(h
->name
,
1248 if (!hstate_kobjs
[h
- hstates
])
1251 retval
= sysfs_create_group(hstate_kobjs
[h
- hstates
],
1252 &hstate_attr_group
);
1254 kobject_put(hstate_kobjs
[h
- hstates
]);
1259 static void __init
hugetlb_sysfs_init(void)
1264 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1265 if (!hugepages_kobj
)
1268 for_each_hstate(h
) {
1269 err
= hugetlb_sysfs_add_hstate(h
);
1271 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s",
1276 static void __exit
hugetlb_exit(void)
1280 for_each_hstate(h
) {
1281 kobject_put(hstate_kobjs
[h
- hstates
]);
1284 kobject_put(hugepages_kobj
);
1286 module_exit(hugetlb_exit
);
1288 static int __init
hugetlb_init(void)
1290 BUILD_BUG_ON(HPAGE_SHIFT
== 0);
1292 if (!size_to_hstate(default_hstate_size
)) {
1293 default_hstate_size
= HPAGE_SIZE
;
1294 if (!size_to_hstate(default_hstate_size
))
1295 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1297 default_hstate_idx
= size_to_hstate(default_hstate_size
) - hstates
;
1298 if (default_hstate_max_huge_pages
)
1299 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1301 hugetlb_init_hstates();
1303 gather_bootmem_prealloc();
1307 hugetlb_sysfs_init();
1311 module_init(hugetlb_init
);
1313 /* Should be called on processing a hugepagesz=... option */
1314 void __init
hugetlb_add_hstate(unsigned order
)
1319 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1320 printk(KERN_WARNING
"hugepagesz= specified twice, ignoring\n");
1323 BUG_ON(max_hstate
>= HUGE_MAX_HSTATE
);
1325 h
= &hstates
[max_hstate
++];
1327 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1328 h
->nr_huge_pages
= 0;
1329 h
->free_huge_pages
= 0;
1330 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1331 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1332 h
->hugetlb_next_nid
= first_node(node_online_map
);
1333 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1334 huge_page_size(h
)/1024);
1339 static int __init
hugetlb_nrpages_setup(char *s
)
1342 static unsigned long *last_mhp
;
1345 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1346 * so this hugepages= parameter goes to the "default hstate".
1349 mhp
= &default_hstate_max_huge_pages
;
1351 mhp
= &parsed_hstate
->max_huge_pages
;
1353 if (mhp
== last_mhp
) {
1354 printk(KERN_WARNING
"hugepages= specified twice without "
1355 "interleaving hugepagesz=, ignoring\n");
1359 if (sscanf(s
, "%lu", mhp
) <= 0)
1363 * Global state is always initialized later in hugetlb_init.
1364 * But we need to allocate >= MAX_ORDER hstates here early to still
1365 * use the bootmem allocator.
1367 if (max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1368 hugetlb_hstate_alloc_pages(parsed_hstate
);
1374 __setup("hugepages=", hugetlb_nrpages_setup
);
1376 static int __init
hugetlb_default_setup(char *s
)
1378 default_hstate_size
= memparse(s
, &s
);
1381 __setup("default_hugepagesz=", hugetlb_default_setup
);
1383 static unsigned int cpuset_mems_nr(unsigned int *array
)
1386 unsigned int nr
= 0;
1388 for_each_node_mask(node
, cpuset_current_mems_allowed
)
1394 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
1395 struct file
*file
, void __user
*buffer
,
1396 size_t *length
, loff_t
*ppos
)
1398 struct hstate
*h
= &default_hstate
;
1402 tmp
= h
->max_huge_pages
;
1405 table
->maxlen
= sizeof(unsigned long);
1406 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1409 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
);
1414 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
1415 struct file
*file
, void __user
*buffer
,
1416 size_t *length
, loff_t
*ppos
)
1418 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
1419 if (hugepages_treat_as_movable
)
1420 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
1422 htlb_alloc_mask
= GFP_HIGHUSER
;
1426 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
1427 struct file
*file
, void __user
*buffer
,
1428 size_t *length
, loff_t
*ppos
)
1430 struct hstate
*h
= &default_hstate
;
1434 tmp
= h
->nr_overcommit_huge_pages
;
1437 table
->maxlen
= sizeof(unsigned long);
1438 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1441 spin_lock(&hugetlb_lock
);
1442 h
->nr_overcommit_huge_pages
= tmp
;
1443 spin_unlock(&hugetlb_lock
);
1449 #endif /* CONFIG_SYSCTL */
1451 int hugetlb_report_meminfo(char *buf
)
1453 struct hstate
*h
= &default_hstate
;
1455 "HugePages_Total: %5lu\n"
1456 "HugePages_Free: %5lu\n"
1457 "HugePages_Rsvd: %5lu\n"
1458 "HugePages_Surp: %5lu\n"
1459 "Hugepagesize: %5lu kB\n",
1463 h
->surplus_huge_pages
,
1464 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
1467 int hugetlb_report_node_meminfo(int nid
, char *buf
)
1469 struct hstate
*h
= &default_hstate
;
1471 "Node %d HugePages_Total: %5u\n"
1472 "Node %d HugePages_Free: %5u\n"
1473 "Node %d HugePages_Surp: %5u\n",
1474 nid
, h
->nr_huge_pages_node
[nid
],
1475 nid
, h
->free_huge_pages_node
[nid
],
1476 nid
, h
->surplus_huge_pages_node
[nid
]);
1479 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1480 unsigned long hugetlb_total_pages(void)
1482 struct hstate
*h
= &default_hstate
;
1483 return h
->nr_huge_pages
* pages_per_huge_page(h
);
1486 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
1490 spin_lock(&hugetlb_lock
);
1492 * When cpuset is configured, it breaks the strict hugetlb page
1493 * reservation as the accounting is done on a global variable. Such
1494 * reservation is completely rubbish in the presence of cpuset because
1495 * the reservation is not checked against page availability for the
1496 * current cpuset. Application can still potentially OOM'ed by kernel
1497 * with lack of free htlb page in cpuset that the task is in.
1498 * Attempt to enforce strict accounting with cpuset is almost
1499 * impossible (or too ugly) because cpuset is too fluid that
1500 * task or memory node can be dynamically moved between cpusets.
1502 * The change of semantics for shared hugetlb mapping with cpuset is
1503 * undesirable. However, in order to preserve some of the semantics,
1504 * we fall back to check against current free page availability as
1505 * a best attempt and hopefully to minimize the impact of changing
1506 * semantics that cpuset has.
1509 if (gather_surplus_pages(h
, delta
) < 0)
1512 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
1513 return_unused_surplus_pages(h
, delta
);
1520 return_unused_surplus_pages(h
, (unsigned long) -delta
);
1523 spin_unlock(&hugetlb_lock
);
1527 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
1529 struct resv_map
*reservations
= vma_resv_map(vma
);
1532 * This new VMA should share its siblings reservation map if present.
1533 * The VMA will only ever have a valid reservation map pointer where
1534 * it is being copied for another still existing VMA. As that VMA
1535 * has a reference to the reservation map it cannot dissappear until
1536 * after this open call completes. It is therefore safe to take a
1537 * new reference here without additional locking.
1540 kref_get(&reservations
->refs
);
1543 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
1545 struct hstate
*h
= hstate_vma(vma
);
1546 struct resv_map
*reservations
= vma_resv_map(vma
);
1547 unsigned long reserve
;
1548 unsigned long start
;
1552 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
1553 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
1555 reserve
= (end
- start
) -
1556 region_count(&reservations
->regions
, start
, end
);
1558 kref_put(&reservations
->refs
, resv_map_release
);
1561 hugetlb_acct_memory(h
, -reserve
);
1566 * We cannot handle pagefaults against hugetlb pages at all. They cause
1567 * handle_mm_fault() to try to instantiate regular-sized pages in the
1568 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1571 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1577 struct vm_operations_struct hugetlb_vm_ops
= {
1578 .fault
= hugetlb_vm_op_fault
,
1579 .open
= hugetlb_vm_op_open
,
1580 .close
= hugetlb_vm_op_close
,
1583 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
1590 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
1592 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
1594 entry
= pte_mkyoung(entry
);
1595 entry
= pte_mkhuge(entry
);
1600 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
1601 unsigned long address
, pte_t
*ptep
)
1605 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
1606 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
1607 update_mmu_cache(vma
, address
, entry
);
1612 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
1613 struct vm_area_struct
*vma
)
1615 pte_t
*src_pte
, *dst_pte
, entry
;
1616 struct page
*ptepage
;
1619 struct hstate
*h
= hstate_vma(vma
);
1620 unsigned long sz
= huge_page_size(h
);
1622 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
1624 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
1625 src_pte
= huge_pte_offset(src
, addr
);
1628 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
1632 /* If the pagetables are shared don't copy or take references */
1633 if (dst_pte
== src_pte
)
1636 spin_lock(&dst
->page_table_lock
);
1637 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
1638 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
1640 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
1641 entry
= huge_ptep_get(src_pte
);
1642 ptepage
= pte_page(entry
);
1644 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
1646 spin_unlock(&src
->page_table_lock
);
1647 spin_unlock(&dst
->page_table_lock
);
1655 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1656 unsigned long end
, struct page
*ref_page
)
1658 struct mm_struct
*mm
= vma
->vm_mm
;
1659 unsigned long address
;
1664 struct hstate
*h
= hstate_vma(vma
);
1665 unsigned long sz
= huge_page_size(h
);
1668 * A page gathering list, protected by per file i_mmap_lock. The
1669 * lock is used to avoid list corruption from multiple unmapping
1670 * of the same page since we are using page->lru.
1672 LIST_HEAD(page_list
);
1674 WARN_ON(!is_vm_hugetlb_page(vma
));
1675 BUG_ON(start
& ~huge_page_mask(h
));
1676 BUG_ON(end
& ~huge_page_mask(h
));
1678 spin_lock(&mm
->page_table_lock
);
1679 for (address
= start
; address
< end
; address
+= sz
) {
1680 ptep
= huge_pte_offset(mm
, address
);
1684 if (huge_pmd_unshare(mm
, &address
, ptep
))
1688 * If a reference page is supplied, it is because a specific
1689 * page is being unmapped, not a range. Ensure the page we
1690 * are about to unmap is the actual page of interest.
1693 pte
= huge_ptep_get(ptep
);
1694 if (huge_pte_none(pte
))
1696 page
= pte_page(pte
);
1697 if (page
!= ref_page
)
1701 * Mark the VMA as having unmapped its page so that
1702 * future faults in this VMA will fail rather than
1703 * looking like data was lost
1705 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
1708 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1709 if (huge_pte_none(pte
))
1712 page
= pte_page(pte
);
1714 set_page_dirty(page
);
1715 list_add(&page
->lru
, &page_list
);
1717 spin_unlock(&mm
->page_table_lock
);
1718 flush_tlb_range(vma
, start
, end
);
1719 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
1720 list_del(&page
->lru
);
1725 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1726 unsigned long end
, struct page
*ref_page
)
1728 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1729 __unmap_hugepage_range(vma
, start
, end
, ref_page
);
1730 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1734 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1735 * mappping it owns the reserve page for. The intention is to unmap the page
1736 * from other VMAs and let the children be SIGKILLed if they are faulting the
1739 int unmap_ref_private(struct mm_struct
*mm
,
1740 struct vm_area_struct
*vma
,
1742 unsigned long address
)
1744 struct vm_area_struct
*iter_vma
;
1745 struct address_space
*mapping
;
1746 struct prio_tree_iter iter
;
1750 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1751 * from page cache lookup which is in HPAGE_SIZE units.
1753 address
= address
& huge_page_mask(hstate_vma(vma
));
1754 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
)
1755 + (vma
->vm_pgoff
>> PAGE_SHIFT
);
1756 mapping
= (struct address_space
*)page_private(page
);
1758 vma_prio_tree_foreach(iter_vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1759 /* Do not unmap the current VMA */
1760 if (iter_vma
== vma
)
1764 * Unmap the page from other VMAs without their own reserves.
1765 * They get marked to be SIGKILLed if they fault in these
1766 * areas. This is because a future no-page fault on this VMA
1767 * could insert a zeroed page instead of the data existing
1768 * from the time of fork. This would look like data corruption
1770 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
1771 unmap_hugepage_range(iter_vma
,
1772 address
, address
+ HPAGE_SIZE
,
1779 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1780 unsigned long address
, pte_t
*ptep
, pte_t pte
,
1781 struct page
*pagecache_page
)
1783 struct hstate
*h
= hstate_vma(vma
);
1784 struct page
*old_page
, *new_page
;
1786 int outside_reserve
= 0;
1788 old_page
= pte_page(pte
);
1791 /* If no-one else is actually using this page, avoid the copy
1792 * and just make the page writable */
1793 avoidcopy
= (page_count(old_page
) == 1);
1795 set_huge_ptep_writable(vma
, address
, ptep
);
1800 * If the process that created a MAP_PRIVATE mapping is about to
1801 * perform a COW due to a shared page count, attempt to satisfy
1802 * the allocation without using the existing reserves. The pagecache
1803 * page is used to determine if the reserve at this address was
1804 * consumed or not. If reserves were used, a partial faulted mapping
1805 * at the time of fork() could consume its reserves on COW instead
1806 * of the full address range.
1808 if (!(vma
->vm_flags
& VM_SHARED
) &&
1809 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
1810 old_page
!= pagecache_page
)
1811 outside_reserve
= 1;
1813 page_cache_get(old_page
);
1814 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
1816 if (IS_ERR(new_page
)) {
1817 page_cache_release(old_page
);
1820 * If a process owning a MAP_PRIVATE mapping fails to COW,
1821 * it is due to references held by a child and an insufficient
1822 * huge page pool. To guarantee the original mappers
1823 * reliability, unmap the page from child processes. The child
1824 * may get SIGKILLed if it later faults.
1826 if (outside_reserve
) {
1827 BUG_ON(huge_pte_none(pte
));
1828 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
1829 BUG_ON(page_count(old_page
) != 1);
1830 BUG_ON(huge_pte_none(pte
));
1831 goto retry_avoidcopy
;
1836 return -PTR_ERR(new_page
);
1839 spin_unlock(&mm
->page_table_lock
);
1840 copy_huge_page(new_page
, old_page
, address
, vma
);
1841 __SetPageUptodate(new_page
);
1842 spin_lock(&mm
->page_table_lock
);
1844 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
1845 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
1847 huge_ptep_clear_flush(vma
, address
, ptep
);
1848 set_huge_pte_at(mm
, address
, ptep
,
1849 make_huge_pte(vma
, new_page
, 1));
1850 /* Make the old page be freed below */
1851 new_page
= old_page
;
1853 page_cache_release(new_page
);
1854 page_cache_release(old_page
);
1858 /* Return the pagecache page at a given address within a VMA */
1859 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
1860 struct vm_area_struct
*vma
, unsigned long address
)
1862 struct address_space
*mapping
;
1865 mapping
= vma
->vm_file
->f_mapping
;
1866 idx
= vma_hugecache_offset(h
, vma
, address
);
1868 return find_lock_page(mapping
, idx
);
1871 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1872 unsigned long address
, pte_t
*ptep
, int write_access
)
1874 struct hstate
*h
= hstate_vma(vma
);
1875 int ret
= VM_FAULT_SIGBUS
;
1879 struct address_space
*mapping
;
1883 * Currently, we are forced to kill the process in the event the
1884 * original mapper has unmapped pages from the child due to a failed
1885 * COW. Warn that such a situation has occured as it may not be obvious
1887 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
1889 "PID %d killed due to inadequate hugepage pool\n",
1894 mapping
= vma
->vm_file
->f_mapping
;
1895 idx
= vma_hugecache_offset(h
, vma
, address
);
1898 * Use page lock to guard against racing truncation
1899 * before we get page_table_lock.
1902 page
= find_lock_page(mapping
, idx
);
1904 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
1907 page
= alloc_huge_page(vma
, address
, 0);
1909 ret
= -PTR_ERR(page
);
1912 clear_huge_page(page
, address
, huge_page_size(h
));
1913 __SetPageUptodate(page
);
1915 if (vma
->vm_flags
& VM_SHARED
) {
1917 struct inode
*inode
= mapping
->host
;
1919 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
1927 spin_lock(&inode
->i_lock
);
1928 inode
->i_blocks
+= blocks_per_huge_page(h
);
1929 spin_unlock(&inode
->i_lock
);
1934 spin_lock(&mm
->page_table_lock
);
1935 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
1940 if (!huge_pte_none(huge_ptep_get(ptep
)))
1943 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
1944 && (vma
->vm_flags
& VM_SHARED
)));
1945 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
1947 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1948 /* Optimization, do the COW without a second fault */
1949 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
1952 spin_unlock(&mm
->page_table_lock
);
1958 spin_unlock(&mm
->page_table_lock
);
1964 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1965 unsigned long address
, int write_access
)
1970 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
1971 struct hstate
*h
= hstate_vma(vma
);
1973 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
1975 return VM_FAULT_OOM
;
1978 * Serialize hugepage allocation and instantiation, so that we don't
1979 * get spurious allocation failures if two CPUs race to instantiate
1980 * the same page in the page cache.
1982 mutex_lock(&hugetlb_instantiation_mutex
);
1983 entry
= huge_ptep_get(ptep
);
1984 if (huge_pte_none(entry
)) {
1985 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
1986 mutex_unlock(&hugetlb_instantiation_mutex
);
1992 spin_lock(&mm
->page_table_lock
);
1993 /* Check for a racing update before calling hugetlb_cow */
1994 if (likely(pte_same(entry
, huge_ptep_get(ptep
))))
1995 if (write_access
&& !pte_write(entry
)) {
1997 page
= hugetlbfs_pagecache_page(h
, vma
, address
);
1998 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
, page
);
2004 spin_unlock(&mm
->page_table_lock
);
2005 mutex_unlock(&hugetlb_instantiation_mutex
);
2010 /* Can be overriden by architectures */
2011 __attribute__((weak
)) struct page
*
2012 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
2013 pud_t
*pud
, int write
)
2019 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2020 struct page
**pages
, struct vm_area_struct
**vmas
,
2021 unsigned long *position
, int *length
, int i
,
2024 unsigned long pfn_offset
;
2025 unsigned long vaddr
= *position
;
2026 int remainder
= *length
;
2027 struct hstate
*h
= hstate_vma(vma
);
2029 spin_lock(&mm
->page_table_lock
);
2030 while (vaddr
< vma
->vm_end
&& remainder
) {
2035 * Some archs (sparc64, sh*) have multiple pte_ts to
2036 * each hugepage. We have to make * sure we get the
2037 * first, for the page indexing below to work.
2039 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2041 if (!pte
|| huge_pte_none(huge_ptep_get(pte
)) ||
2042 (write
&& !pte_write(huge_ptep_get(pte
)))) {
2045 spin_unlock(&mm
->page_table_lock
);
2046 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
2047 spin_lock(&mm
->page_table_lock
);
2048 if (!(ret
& VM_FAULT_ERROR
))
2057 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2058 page
= pte_page(huge_ptep_get(pte
));
2062 pages
[i
] = page
+ pfn_offset
;
2072 if (vaddr
< vma
->vm_end
&& remainder
&&
2073 pfn_offset
< pages_per_huge_page(h
)) {
2075 * We use pfn_offset to avoid touching the pageframes
2076 * of this compound page.
2081 spin_unlock(&mm
->page_table_lock
);
2082 *length
= remainder
;
2088 void hugetlb_change_protection(struct vm_area_struct
*vma
,
2089 unsigned long address
, unsigned long end
, pgprot_t newprot
)
2091 struct mm_struct
*mm
= vma
->vm_mm
;
2092 unsigned long start
= address
;
2095 struct hstate
*h
= hstate_vma(vma
);
2097 BUG_ON(address
>= end
);
2098 flush_cache_range(vma
, address
, end
);
2100 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2101 spin_lock(&mm
->page_table_lock
);
2102 for (; address
< end
; address
+= huge_page_size(h
)) {
2103 ptep
= huge_pte_offset(mm
, address
);
2106 if (huge_pmd_unshare(mm
, &address
, ptep
))
2108 if (!huge_pte_none(huge_ptep_get(ptep
))) {
2109 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2110 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
2111 set_huge_pte_at(mm
, address
, ptep
, pte
);
2114 spin_unlock(&mm
->page_table_lock
);
2115 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2117 flush_tlb_range(vma
, start
, end
);
2120 int hugetlb_reserve_pages(struct inode
*inode
,
2122 struct vm_area_struct
*vma
)
2125 struct hstate
*h
= hstate_inode(inode
);
2127 if (vma
&& vma
->vm_flags
& VM_NORESERVE
)
2131 * Shared mappings base their reservation on the number of pages that
2132 * are already allocated on behalf of the file. Private mappings need
2133 * to reserve the full area even if read-only as mprotect() may be
2134 * called to make the mapping read-write. Assume !vma is a shm mapping
2136 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2137 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
2139 struct resv_map
*resv_map
= resv_map_alloc();
2145 set_vma_resv_map(vma
, resv_map
);
2146 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
2152 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
2154 ret
= hugetlb_acct_memory(h
, chg
);
2156 hugetlb_put_quota(inode
->i_mapping
, chg
);
2159 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2160 region_add(&inode
->i_mapping
->private_list
, from
, to
);
2164 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
2166 struct hstate
*h
= hstate_inode(inode
);
2167 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
2169 spin_lock(&inode
->i_lock
);
2170 inode
->i_blocks
-= blocks_per_huge_page(h
);
2171 spin_unlock(&inode
->i_lock
);
2173 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
2174 hugetlb_acct_memory(h
, -(chg
- freed
));