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
;
38 #define for_each_hstate(h) \
39 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
42 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
44 static DEFINE_SPINLOCK(hugetlb_lock
);
47 * Region tracking -- allows tracking of reservations and instantiated pages
48 * across the pages in a mapping.
50 * The region data structures are protected by a combination of the mmap_sem
51 * and the hugetlb_instantion_mutex. To access or modify a region the caller
52 * must either hold the mmap_sem for write, or the mmap_sem for read and
53 * the hugetlb_instantiation mutex:
55 * down_write(&mm->mmap_sem);
57 * down_read(&mm->mmap_sem);
58 * mutex_lock(&hugetlb_instantiation_mutex);
61 struct list_head link
;
66 static long region_add(struct list_head
*head
, long f
, long t
)
68 struct file_region
*rg
, *nrg
, *trg
;
70 /* Locate the region we are either in or before. */
71 list_for_each_entry(rg
, head
, link
)
75 /* Round our left edge to the current segment if it encloses us. */
79 /* Check for and consume any regions we now overlap with. */
81 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
82 if (&rg
->link
== head
)
87 /* If this area reaches higher then extend our area to
88 * include it completely. If this is not the first area
89 * which we intend to reuse, free it. */
102 static long region_chg(struct list_head
*head
, long f
, long t
)
104 struct file_region
*rg
, *nrg
;
107 /* Locate the region we are before or in. */
108 list_for_each_entry(rg
, head
, link
)
112 /* If we are below the current region then a new region is required.
113 * Subtle, allocate a new region at the position but make it zero
114 * size such that we can guarantee to record the reservation. */
115 if (&rg
->link
== head
|| t
< rg
->from
) {
116 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
121 INIT_LIST_HEAD(&nrg
->link
);
122 list_add(&nrg
->link
, rg
->link
.prev
);
127 /* Round our left edge to the current segment if it encloses us. */
132 /* Check for and consume any regions we now overlap with. */
133 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
134 if (&rg
->link
== head
)
139 /* We overlap with this area, if it extends futher than
140 * us then we must extend ourselves. Account for its
141 * existing reservation. */
146 chg
-= rg
->to
- rg
->from
;
151 static long region_truncate(struct list_head
*head
, long end
)
153 struct file_region
*rg
, *trg
;
156 /* Locate the region we are either in or before. */
157 list_for_each_entry(rg
, head
, link
)
160 if (&rg
->link
== head
)
163 /* If we are in the middle of a region then adjust it. */
164 if (end
> rg
->from
) {
167 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
170 /* Drop any remaining regions. */
171 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
172 if (&rg
->link
== head
)
174 chg
+= rg
->to
- rg
->from
;
181 static long region_count(struct list_head
*head
, long f
, long t
)
183 struct file_region
*rg
;
186 /* Locate each segment we overlap with, and count that overlap. */
187 list_for_each_entry(rg
, head
, link
) {
196 seg_from
= max(rg
->from
, f
);
197 seg_to
= min(rg
->to
, t
);
199 chg
+= seg_to
- seg_from
;
206 * Convert the address within this vma to the page offset within
207 * the mapping, in pagecache page units; huge pages here.
209 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
210 struct vm_area_struct
*vma
, unsigned long address
)
212 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
213 (vma
->vm_pgoff
>> huge_page_order(h
));
217 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
218 * bits of the reservation map pointer, which are always clear due to
221 #define HPAGE_RESV_OWNER (1UL << 0)
222 #define HPAGE_RESV_UNMAPPED (1UL << 1)
223 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
226 * These helpers are used to track how many pages are reserved for
227 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
228 * is guaranteed to have their future faults succeed.
230 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
231 * the reserve counters are updated with the hugetlb_lock held. It is safe
232 * to reset the VMA at fork() time as it is not in use yet and there is no
233 * chance of the global counters getting corrupted as a result of the values.
235 * The private mapping reservation is represented in a subtly different
236 * manner to a shared mapping. A shared mapping has a region map associated
237 * with the underlying file, this region map represents the backing file
238 * pages which have ever had a reservation assigned which this persists even
239 * after the page is instantiated. A private mapping has a region map
240 * associated with the original mmap which is attached to all VMAs which
241 * reference it, this region map represents those offsets which have consumed
242 * reservation ie. where pages have been instantiated.
244 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
246 return (unsigned long)vma
->vm_private_data
;
249 static void set_vma_private_data(struct vm_area_struct
*vma
,
252 vma
->vm_private_data
= (void *)value
;
257 struct list_head regions
;
260 struct resv_map
*resv_map_alloc(void)
262 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
266 kref_init(&resv_map
->refs
);
267 INIT_LIST_HEAD(&resv_map
->regions
);
272 void resv_map_release(struct kref
*ref
)
274 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
276 /* Clear out any active regions before we release the map. */
277 region_truncate(&resv_map
->regions
, 0);
281 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
283 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
284 if (!(vma
->vm_flags
& VM_SHARED
))
285 return (struct resv_map
*)(get_vma_private_data(vma
) &
290 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
292 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
293 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
295 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
296 HPAGE_RESV_MASK
) | (unsigned long)map
);
299 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
301 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
302 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
304 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
307 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
309 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
311 return (get_vma_private_data(vma
) & flag
) != 0;
314 /* Decrement the reserved pages in the hugepage pool by one */
315 static void decrement_hugepage_resv_vma(struct hstate
*h
,
316 struct vm_area_struct
*vma
)
318 if (vma
->vm_flags
& VM_NORESERVE
)
321 if (vma
->vm_flags
& VM_SHARED
) {
322 /* Shared mappings always use reserves */
323 h
->resv_huge_pages
--;
324 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
326 * Only the process that called mmap() has reserves for
329 h
->resv_huge_pages
--;
333 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
334 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
336 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
337 if (!(vma
->vm_flags
& VM_SHARED
))
338 vma
->vm_private_data
= (void *)0;
341 /* Returns true if the VMA has associated reserve pages */
342 static int vma_has_private_reserves(struct vm_area_struct
*vma
)
344 if (vma
->vm_flags
& VM_SHARED
)
346 if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
351 static void clear_huge_page(struct page
*page
,
352 unsigned long addr
, unsigned long sz
)
357 for (i
= 0; i
< sz
/PAGE_SIZE
; i
++) {
359 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
363 static void copy_huge_page(struct page
*dst
, struct page
*src
,
364 unsigned long addr
, struct vm_area_struct
*vma
)
367 struct hstate
*h
= hstate_vma(vma
);
370 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
372 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
376 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
378 int nid
= page_to_nid(page
);
379 list_add(&page
->lru
, &h
->hugepage_freelists
[nid
]);
380 h
->free_huge_pages
++;
381 h
->free_huge_pages_node
[nid
]++;
384 static struct page
*dequeue_huge_page(struct hstate
*h
)
387 struct page
*page
= NULL
;
389 for (nid
= 0; nid
< MAX_NUMNODES
; ++nid
) {
390 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
391 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
393 list_del(&page
->lru
);
394 h
->free_huge_pages
--;
395 h
->free_huge_pages_node
[nid
]--;
402 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
403 struct vm_area_struct
*vma
,
404 unsigned long address
, int avoid_reserve
)
407 struct page
*page
= NULL
;
408 struct mempolicy
*mpol
;
409 nodemask_t
*nodemask
;
410 struct zonelist
*zonelist
= huge_zonelist(vma
, address
,
411 htlb_alloc_mask
, &mpol
, &nodemask
);
416 * A child process with MAP_PRIVATE mappings created by their parent
417 * have no page reserves. This check ensures that reservations are
418 * not "stolen". The child may still get SIGKILLed
420 if (!vma_has_private_reserves(vma
) &&
421 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
424 /* If reserves cannot be used, ensure enough pages are in the pool */
425 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
428 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
429 MAX_NR_ZONES
- 1, nodemask
) {
430 nid
= zone_to_nid(zone
);
431 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
) &&
432 !list_empty(&h
->hugepage_freelists
[nid
])) {
433 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
435 list_del(&page
->lru
);
436 h
->free_huge_pages
--;
437 h
->free_huge_pages_node
[nid
]--;
440 decrement_hugepage_resv_vma(h
, vma
);
449 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
454 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
455 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
456 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
457 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
458 1 << PG_private
| 1<< PG_writeback
);
460 set_compound_page_dtor(page
, NULL
);
461 set_page_refcounted(page
);
462 arch_release_hugepage(page
);
463 __free_pages(page
, huge_page_order(h
));
466 struct hstate
*size_to_hstate(unsigned long size
)
471 if (huge_page_size(h
) == size
)
477 static void free_huge_page(struct page
*page
)
480 * Can't pass hstate in here because it is called from the
481 * compound page destructor.
483 struct hstate
*h
= page_hstate(page
);
484 int nid
= page_to_nid(page
);
485 struct address_space
*mapping
;
487 mapping
= (struct address_space
*) page_private(page
);
488 set_page_private(page
, 0);
489 BUG_ON(page_count(page
));
490 INIT_LIST_HEAD(&page
->lru
);
492 spin_lock(&hugetlb_lock
);
493 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
494 update_and_free_page(h
, page
);
495 h
->surplus_huge_pages
--;
496 h
->surplus_huge_pages_node
[nid
]--;
498 enqueue_huge_page(h
, page
);
500 spin_unlock(&hugetlb_lock
);
502 hugetlb_put_quota(mapping
, 1);
506 * Increment or decrement surplus_huge_pages. Keep node-specific counters
507 * balanced by operating on them in a round-robin fashion.
508 * Returns 1 if an adjustment was made.
510 static int adjust_pool_surplus(struct hstate
*h
, int delta
)
516 VM_BUG_ON(delta
!= -1 && delta
!= 1);
518 nid
= next_node(nid
, node_online_map
);
519 if (nid
== MAX_NUMNODES
)
520 nid
= first_node(node_online_map
);
522 /* To shrink on this node, there must be a surplus page */
523 if (delta
< 0 && !h
->surplus_huge_pages_node
[nid
])
525 /* Surplus cannot exceed the total number of pages */
526 if (delta
> 0 && h
->surplus_huge_pages_node
[nid
] >=
527 h
->nr_huge_pages_node
[nid
])
530 h
->surplus_huge_pages
+= delta
;
531 h
->surplus_huge_pages_node
[nid
] += delta
;
534 } while (nid
!= prev_nid
);
540 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
542 set_compound_page_dtor(page
, free_huge_page
);
543 spin_lock(&hugetlb_lock
);
545 h
->nr_huge_pages_node
[nid
]++;
546 spin_unlock(&hugetlb_lock
);
547 put_page(page
); /* free it into the hugepage allocator */
550 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
554 if (h
->order
>= MAX_ORDER
)
557 page
= alloc_pages_node(nid
,
558 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
559 __GFP_REPEAT
|__GFP_NOWARN
,
562 if (arch_prepare_hugepage(page
)) {
563 __free_pages(page
, HUGETLB_PAGE_ORDER
);
566 prep_new_huge_page(h
, page
, nid
);
573 * Use a helper variable to find the next node and then
574 * copy it back to hugetlb_next_nid afterwards:
575 * otherwise there's a window in which a racer might
576 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
577 * But we don't need to use a spin_lock here: it really
578 * doesn't matter if occasionally a racer chooses the
579 * same nid as we do. Move nid forward in the mask even
580 * if we just successfully allocated a hugepage so that
581 * the next caller gets hugepages on the next node.
583 static int hstate_next_node(struct hstate
*h
)
586 next_nid
= next_node(h
->hugetlb_next_nid
, node_online_map
);
587 if (next_nid
== MAX_NUMNODES
)
588 next_nid
= first_node(node_online_map
);
589 h
->hugetlb_next_nid
= next_nid
;
593 static int alloc_fresh_huge_page(struct hstate
*h
)
600 start_nid
= h
->hugetlb_next_nid
;
603 page
= alloc_fresh_huge_page_node(h
, h
->hugetlb_next_nid
);
606 next_nid
= hstate_next_node(h
);
607 } while (!page
&& h
->hugetlb_next_nid
!= start_nid
);
610 count_vm_event(HTLB_BUDDY_PGALLOC
);
612 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
617 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
618 struct vm_area_struct
*vma
, unsigned long address
)
623 if (h
->order
>= MAX_ORDER
)
627 * Assume we will successfully allocate the surplus page to
628 * prevent racing processes from causing the surplus to exceed
631 * This however introduces a different race, where a process B
632 * tries to grow the static hugepage pool while alloc_pages() is
633 * called by process A. B will only examine the per-node
634 * counters in determining if surplus huge pages can be
635 * converted to normal huge pages in adjust_pool_surplus(). A
636 * won't be able to increment the per-node counter, until the
637 * lock is dropped by B, but B doesn't drop hugetlb_lock until
638 * no more huge pages can be converted from surplus to normal
639 * state (and doesn't try to convert again). Thus, we have a
640 * case where a surplus huge page exists, the pool is grown, and
641 * the surplus huge page still exists after, even though it
642 * should just have been converted to a normal huge page. This
643 * does not leak memory, though, as the hugepage will be freed
644 * once it is out of use. It also does not allow the counters to
645 * go out of whack in adjust_pool_surplus() as we don't modify
646 * the node values until we've gotten the hugepage and only the
647 * per-node value is checked there.
649 spin_lock(&hugetlb_lock
);
650 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
651 spin_unlock(&hugetlb_lock
);
655 h
->surplus_huge_pages
++;
657 spin_unlock(&hugetlb_lock
);
659 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
660 __GFP_REPEAT
|__GFP_NOWARN
,
663 spin_lock(&hugetlb_lock
);
666 * This page is now managed by the hugetlb allocator and has
667 * no users -- drop the buddy allocator's reference.
669 put_page_testzero(page
);
670 VM_BUG_ON(page_count(page
));
671 nid
= page_to_nid(page
);
672 set_compound_page_dtor(page
, free_huge_page
);
674 * We incremented the global counters already
676 h
->nr_huge_pages_node
[nid
]++;
677 h
->surplus_huge_pages_node
[nid
]++;
678 __count_vm_event(HTLB_BUDDY_PGALLOC
);
681 h
->surplus_huge_pages
--;
682 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
684 spin_unlock(&hugetlb_lock
);
690 * Increase the hugetlb pool such that it can accomodate a reservation
693 static int gather_surplus_pages(struct hstate
*h
, int delta
)
695 struct list_head surplus_list
;
696 struct page
*page
, *tmp
;
698 int needed
, allocated
;
700 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
702 h
->resv_huge_pages
+= delta
;
707 INIT_LIST_HEAD(&surplus_list
);
711 spin_unlock(&hugetlb_lock
);
712 for (i
= 0; i
< needed
; i
++) {
713 page
= alloc_buddy_huge_page(h
, NULL
, 0);
716 * We were not able to allocate enough pages to
717 * satisfy the entire reservation so we free what
718 * we've allocated so far.
720 spin_lock(&hugetlb_lock
);
725 list_add(&page
->lru
, &surplus_list
);
730 * After retaking hugetlb_lock, we need to recalculate 'needed'
731 * because either resv_huge_pages or free_huge_pages may have changed.
733 spin_lock(&hugetlb_lock
);
734 needed
= (h
->resv_huge_pages
+ delta
) -
735 (h
->free_huge_pages
+ allocated
);
740 * The surplus_list now contains _at_least_ the number of extra pages
741 * needed to accomodate the reservation. Add the appropriate number
742 * of pages to the hugetlb pool and free the extras back to the buddy
743 * allocator. Commit the entire reservation here to prevent another
744 * process from stealing the pages as they are added to the pool but
745 * before they are reserved.
748 h
->resv_huge_pages
+= delta
;
751 /* Free the needed pages to the hugetlb pool */
752 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
755 list_del(&page
->lru
);
756 enqueue_huge_page(h
, page
);
759 /* Free unnecessary surplus pages to the buddy allocator */
760 if (!list_empty(&surplus_list
)) {
761 spin_unlock(&hugetlb_lock
);
762 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
763 list_del(&page
->lru
);
765 * The page has a reference count of zero already, so
766 * call free_huge_page directly instead of using
767 * put_page. This must be done with hugetlb_lock
768 * unlocked which is safe because free_huge_page takes
769 * hugetlb_lock before deciding how to free the page.
771 free_huge_page(page
);
773 spin_lock(&hugetlb_lock
);
780 * When releasing a hugetlb pool reservation, any surplus pages that were
781 * allocated to satisfy the reservation must be explicitly freed if they were
784 static void return_unused_surplus_pages(struct hstate
*h
,
785 unsigned long unused_resv_pages
)
789 unsigned long nr_pages
;
792 * We want to release as many surplus pages as possible, spread
793 * evenly across all nodes. Iterate across all nodes until we
794 * can no longer free unreserved surplus pages. This occurs when
795 * the nodes with surplus pages have no free pages.
797 unsigned long remaining_iterations
= num_online_nodes();
799 /* Uncommit the reservation */
800 h
->resv_huge_pages
-= unused_resv_pages
;
802 /* Cannot return gigantic pages currently */
803 if (h
->order
>= MAX_ORDER
)
806 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
808 while (remaining_iterations
-- && nr_pages
) {
809 nid
= next_node(nid
, node_online_map
);
810 if (nid
== MAX_NUMNODES
)
811 nid
= first_node(node_online_map
);
813 if (!h
->surplus_huge_pages_node
[nid
])
816 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
817 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
819 list_del(&page
->lru
);
820 update_and_free_page(h
, page
);
821 h
->free_huge_pages
--;
822 h
->free_huge_pages_node
[nid
]--;
823 h
->surplus_huge_pages
--;
824 h
->surplus_huge_pages_node
[nid
]--;
826 remaining_iterations
= num_online_nodes();
832 * Determine if the huge page at addr within the vma has an associated
833 * reservation. Where it does not we will need to logically increase
834 * reservation and actually increase quota before an allocation can occur.
835 * Where any new reservation would be required the reservation change is
836 * prepared, but not committed. Once the page has been quota'd allocated
837 * an instantiated the change should be committed via vma_commit_reservation.
838 * No action is required on failure.
840 static int vma_needs_reservation(struct hstate
*h
,
841 struct vm_area_struct
*vma
, unsigned long addr
)
843 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
844 struct inode
*inode
= mapping
->host
;
846 if (vma
->vm_flags
& VM_SHARED
) {
847 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
848 return region_chg(&inode
->i_mapping
->private_list
,
851 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
856 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
857 struct resv_map
*reservations
= vma_resv_map(vma
);
859 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
865 static void vma_commit_reservation(struct hstate
*h
,
866 struct vm_area_struct
*vma
, unsigned long addr
)
868 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
869 struct inode
*inode
= mapping
->host
;
871 if (vma
->vm_flags
& VM_SHARED
) {
872 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
873 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
875 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
876 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
877 struct resv_map
*reservations
= vma_resv_map(vma
);
879 /* Mark this page used in the map. */
880 region_add(&reservations
->regions
, idx
, idx
+ 1);
884 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
885 unsigned long addr
, int avoid_reserve
)
887 struct hstate
*h
= hstate_vma(vma
);
889 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
890 struct inode
*inode
= mapping
->host
;
894 * Processes that did not create the mapping will have no reserves and
895 * will not have accounted against quota. Check that the quota can be
896 * made before satisfying the allocation
897 * MAP_NORESERVE mappings may also need pages and quota allocated
898 * if no reserve mapping overlaps.
900 chg
= vma_needs_reservation(h
, vma
, addr
);
904 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
905 return ERR_PTR(-ENOSPC
);
907 spin_lock(&hugetlb_lock
);
908 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
909 spin_unlock(&hugetlb_lock
);
912 page
= alloc_buddy_huge_page(h
, vma
, addr
);
914 hugetlb_put_quota(inode
->i_mapping
, chg
);
915 return ERR_PTR(-VM_FAULT_OOM
);
919 set_page_refcounted(page
);
920 set_page_private(page
, (unsigned long) mapping
);
922 vma_commit_reservation(h
, vma
, addr
);
927 static __initdata
LIST_HEAD(huge_boot_pages
);
929 struct huge_bootmem_page
{
930 struct list_head list
;
931 struct hstate
*hstate
;
934 static int __init
alloc_bootmem_huge_page(struct hstate
*h
)
936 struct huge_bootmem_page
*m
;
937 int nr_nodes
= nodes_weight(node_online_map
);
942 addr
= __alloc_bootmem_node_nopanic(
943 NODE_DATA(h
->hugetlb_next_nid
),
944 huge_page_size(h
), huge_page_size(h
), 0);
948 * Use the beginning of the huge page to store the
949 * huge_bootmem_page struct (until gather_bootmem
950 * puts them into the mem_map).
962 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
963 /* Put them into a private list first because mem_map is not up yet */
964 list_add(&m
->list
, &huge_boot_pages
);
969 /* Put bootmem huge pages into the standard lists after mem_map is up */
970 static void __init
gather_bootmem_prealloc(void)
972 struct huge_bootmem_page
*m
;
974 list_for_each_entry(m
, &huge_boot_pages
, list
) {
975 struct page
*page
= virt_to_page(m
);
976 struct hstate
*h
= m
->hstate
;
977 __ClearPageReserved(page
);
978 WARN_ON(page_count(page
) != 1);
979 prep_compound_page(page
, h
->order
);
980 prep_new_huge_page(h
, page
, page_to_nid(page
));
984 static void __init
hugetlb_init_one_hstate(struct hstate
*h
)
988 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
989 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
991 h
->hugetlb_next_nid
= first_node(node_online_map
);
993 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
994 if (h
->order
>= MAX_ORDER
) {
995 if (!alloc_bootmem_huge_page(h
))
997 } else if (!alloc_fresh_huge_page(h
))
1000 h
->max_huge_pages
= h
->free_huge_pages
= h
->nr_huge_pages
= i
;
1003 static void __init
hugetlb_init_hstates(void)
1007 for_each_hstate(h
) {
1008 hugetlb_init_one_hstate(h
);
1012 static void __init
report_hugepages(void)
1016 for_each_hstate(h
) {
1017 printk(KERN_INFO
"Total HugeTLB memory allocated, "
1020 1 << (h
->order
+ PAGE_SHIFT
- 20));
1024 #ifdef CONFIG_SYSCTL
1025 #ifdef CONFIG_HIGHMEM
1026 static void try_to_free_low(struct hstate
*h
, unsigned long count
)
1030 if (h
->order
>= MAX_ORDER
)
1033 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
1034 struct page
*page
, *next
;
1035 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1036 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1037 if (count
>= h
->nr_huge_pages
)
1039 if (PageHighMem(page
))
1041 list_del(&page
->lru
);
1042 update_and_free_page(h
, page
);
1043 h
->free_huge_pages
--;
1044 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1049 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
)
1054 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1055 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
)
1057 unsigned long min_count
, ret
;
1059 if (h
->order
>= MAX_ORDER
)
1060 return h
->max_huge_pages
;
1063 * Increase the pool size
1064 * First take pages out of surplus state. Then make up the
1065 * remaining difference by allocating fresh huge pages.
1067 * We might race with alloc_buddy_huge_page() here and be unable
1068 * to convert a surplus huge page to a normal huge page. That is
1069 * not critical, though, it just means the overall size of the
1070 * pool might be one hugepage larger than it needs to be, but
1071 * within all the constraints specified by the sysctls.
1073 spin_lock(&hugetlb_lock
);
1074 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1075 if (!adjust_pool_surplus(h
, -1))
1079 while (count
> persistent_huge_pages(h
)) {
1081 * If this allocation races such that we no longer need the
1082 * page, free_huge_page will handle it by freeing the page
1083 * and reducing the surplus.
1085 spin_unlock(&hugetlb_lock
);
1086 ret
= alloc_fresh_huge_page(h
);
1087 spin_lock(&hugetlb_lock
);
1094 * Decrease the pool size
1095 * First return free pages to the buddy allocator (being careful
1096 * to keep enough around to satisfy reservations). Then place
1097 * pages into surplus state as needed so the pool will shrink
1098 * to the desired size as pages become free.
1100 * By placing pages into the surplus state independent of the
1101 * overcommit value, we are allowing the surplus pool size to
1102 * exceed overcommit. There are few sane options here. Since
1103 * alloc_buddy_huge_page() is checking the global counter,
1104 * though, we'll note that we're not allowed to exceed surplus
1105 * and won't grow the pool anywhere else. Not until one of the
1106 * sysctls are changed, or the surplus pages go out of use.
1108 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1109 min_count
= max(count
, min_count
);
1110 try_to_free_low(h
, min_count
);
1111 while (min_count
< persistent_huge_pages(h
)) {
1112 struct page
*page
= dequeue_huge_page(h
);
1115 update_and_free_page(h
, page
);
1117 while (count
< persistent_huge_pages(h
)) {
1118 if (!adjust_pool_surplus(h
, 1))
1122 ret
= persistent_huge_pages(h
);
1123 spin_unlock(&hugetlb_lock
);
1127 #define HSTATE_ATTR_RO(_name) \
1128 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1130 #define HSTATE_ATTR(_name) \
1131 static struct kobj_attribute _name##_attr = \
1132 __ATTR(_name, 0644, _name##_show, _name##_store)
1134 static struct kobject
*hugepages_kobj
;
1135 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1137 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
)
1140 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1141 if (hstate_kobjs
[i
] == kobj
)
1147 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1148 struct kobj_attribute
*attr
, char *buf
)
1150 struct hstate
*h
= kobj_to_hstate(kobj
);
1151 return sprintf(buf
, "%lu\n", h
->nr_huge_pages
);
1153 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1154 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1157 unsigned long input
;
1158 struct hstate
*h
= kobj_to_hstate(kobj
);
1160 err
= strict_strtoul(buf
, 10, &input
);
1164 h
->max_huge_pages
= set_max_huge_pages(h
, input
);
1168 HSTATE_ATTR(nr_hugepages
);
1170 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1171 struct kobj_attribute
*attr
, char *buf
)
1173 struct hstate
*h
= kobj_to_hstate(kobj
);
1174 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1176 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1177 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1180 unsigned long input
;
1181 struct hstate
*h
= kobj_to_hstate(kobj
);
1183 err
= strict_strtoul(buf
, 10, &input
);
1187 spin_lock(&hugetlb_lock
);
1188 h
->nr_overcommit_huge_pages
= input
;
1189 spin_unlock(&hugetlb_lock
);
1193 HSTATE_ATTR(nr_overcommit_hugepages
);
1195 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1196 struct kobj_attribute
*attr
, char *buf
)
1198 struct hstate
*h
= kobj_to_hstate(kobj
);
1199 return sprintf(buf
, "%lu\n", h
->free_huge_pages
);
1201 HSTATE_ATTR_RO(free_hugepages
);
1203 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1204 struct kobj_attribute
*attr
, char *buf
)
1206 struct hstate
*h
= kobj_to_hstate(kobj
);
1207 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1209 HSTATE_ATTR_RO(resv_hugepages
);
1211 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1212 struct kobj_attribute
*attr
, char *buf
)
1214 struct hstate
*h
= kobj_to_hstate(kobj
);
1215 return sprintf(buf
, "%lu\n", h
->surplus_huge_pages
);
1217 HSTATE_ATTR_RO(surplus_hugepages
);
1219 static struct attribute
*hstate_attrs
[] = {
1220 &nr_hugepages_attr
.attr
,
1221 &nr_overcommit_hugepages_attr
.attr
,
1222 &free_hugepages_attr
.attr
,
1223 &resv_hugepages_attr
.attr
,
1224 &surplus_hugepages_attr
.attr
,
1228 static struct attribute_group hstate_attr_group
= {
1229 .attrs
= hstate_attrs
,
1232 static int __init
hugetlb_sysfs_add_hstate(struct hstate
*h
)
1236 hstate_kobjs
[h
- hstates
] = kobject_create_and_add(h
->name
,
1238 if (!hstate_kobjs
[h
- hstates
])
1241 retval
= sysfs_create_group(hstate_kobjs
[h
- hstates
],
1242 &hstate_attr_group
);
1244 kobject_put(hstate_kobjs
[h
- hstates
]);
1249 static void __init
hugetlb_sysfs_init(void)
1254 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1255 if (!hugepages_kobj
)
1258 for_each_hstate(h
) {
1259 err
= hugetlb_sysfs_add_hstate(h
);
1261 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s",
1266 static void __exit
hugetlb_exit(void)
1270 for_each_hstate(h
) {
1271 kobject_put(hstate_kobjs
[h
- hstates
]);
1274 kobject_put(hugepages_kobj
);
1276 module_exit(hugetlb_exit
);
1278 static int __init
hugetlb_init(void)
1280 BUILD_BUG_ON(HPAGE_SHIFT
== 0);
1282 if (!size_to_hstate(HPAGE_SIZE
)) {
1283 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1284 parsed_hstate
->max_huge_pages
= default_hstate_max_huge_pages
;
1286 default_hstate_idx
= size_to_hstate(HPAGE_SIZE
) - hstates
;
1288 hugetlb_init_hstates();
1290 gather_bootmem_prealloc();
1294 hugetlb_sysfs_init();
1298 module_init(hugetlb_init
);
1300 /* Should be called on processing a hugepagesz=... option */
1301 void __init
hugetlb_add_hstate(unsigned order
)
1304 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1305 printk(KERN_WARNING
"hugepagesz= specified twice, ignoring\n");
1308 BUG_ON(max_hstate
>= HUGE_MAX_HSTATE
);
1310 h
= &hstates
[max_hstate
++];
1312 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1313 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1314 huge_page_size(h
)/1024);
1315 hugetlb_init_one_hstate(h
);
1319 static int __init
hugetlb_setup(char *s
)
1324 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1325 * so this hugepages= parameter goes to the "default hstate".
1328 mhp
= &default_hstate_max_huge_pages
;
1330 mhp
= &parsed_hstate
->max_huge_pages
;
1332 if (sscanf(s
, "%lu", mhp
) <= 0)
1337 __setup("hugepages=", hugetlb_setup
);
1339 static unsigned int cpuset_mems_nr(unsigned int *array
)
1342 unsigned int nr
= 0;
1344 for_each_node_mask(node
, cpuset_current_mems_allowed
)
1350 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
1351 struct file
*file
, void __user
*buffer
,
1352 size_t *length
, loff_t
*ppos
)
1354 struct hstate
*h
= &default_hstate
;
1358 tmp
= h
->max_huge_pages
;
1361 table
->maxlen
= sizeof(unsigned long);
1362 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1365 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
);
1370 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
1371 struct file
*file
, void __user
*buffer
,
1372 size_t *length
, loff_t
*ppos
)
1374 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
1375 if (hugepages_treat_as_movable
)
1376 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
1378 htlb_alloc_mask
= GFP_HIGHUSER
;
1382 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
1383 struct file
*file
, void __user
*buffer
,
1384 size_t *length
, loff_t
*ppos
)
1386 struct hstate
*h
= &default_hstate
;
1390 tmp
= h
->nr_overcommit_huge_pages
;
1393 table
->maxlen
= sizeof(unsigned long);
1394 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1397 spin_lock(&hugetlb_lock
);
1398 h
->nr_overcommit_huge_pages
= tmp
;
1399 spin_unlock(&hugetlb_lock
);
1405 #endif /* CONFIG_SYSCTL */
1407 int hugetlb_report_meminfo(char *buf
)
1409 struct hstate
*h
= &default_hstate
;
1411 "HugePages_Total: %5lu\n"
1412 "HugePages_Free: %5lu\n"
1413 "HugePages_Rsvd: %5lu\n"
1414 "HugePages_Surp: %5lu\n"
1415 "Hugepagesize: %5lu kB\n",
1419 h
->surplus_huge_pages
,
1420 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
1423 int hugetlb_report_node_meminfo(int nid
, char *buf
)
1425 struct hstate
*h
= &default_hstate
;
1427 "Node %d HugePages_Total: %5u\n"
1428 "Node %d HugePages_Free: %5u\n"
1429 "Node %d HugePages_Surp: %5u\n",
1430 nid
, h
->nr_huge_pages_node
[nid
],
1431 nid
, h
->free_huge_pages_node
[nid
],
1432 nid
, h
->surplus_huge_pages_node
[nid
]);
1435 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1436 unsigned long hugetlb_total_pages(void)
1438 struct hstate
*h
= &default_hstate
;
1439 return h
->nr_huge_pages
* pages_per_huge_page(h
);
1442 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
1446 spin_lock(&hugetlb_lock
);
1448 * When cpuset is configured, it breaks the strict hugetlb page
1449 * reservation as the accounting is done on a global variable. Such
1450 * reservation is completely rubbish in the presence of cpuset because
1451 * the reservation is not checked against page availability for the
1452 * current cpuset. Application can still potentially OOM'ed by kernel
1453 * with lack of free htlb page in cpuset that the task is in.
1454 * Attempt to enforce strict accounting with cpuset is almost
1455 * impossible (or too ugly) because cpuset is too fluid that
1456 * task or memory node can be dynamically moved between cpusets.
1458 * The change of semantics for shared hugetlb mapping with cpuset is
1459 * undesirable. However, in order to preserve some of the semantics,
1460 * we fall back to check against current free page availability as
1461 * a best attempt and hopefully to minimize the impact of changing
1462 * semantics that cpuset has.
1465 if (gather_surplus_pages(h
, delta
) < 0)
1468 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
1469 return_unused_surplus_pages(h
, delta
);
1476 return_unused_surplus_pages(h
, (unsigned long) -delta
);
1479 spin_unlock(&hugetlb_lock
);
1483 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
1485 struct resv_map
*reservations
= vma_resv_map(vma
);
1488 * This new VMA should share its siblings reservation map if present.
1489 * The VMA will only ever have a valid reservation map pointer where
1490 * it is being copied for another still existing VMA. As that VMA
1491 * has a reference to the reservation map it cannot dissappear until
1492 * after this open call completes. It is therefore safe to take a
1493 * new reference here without additional locking.
1496 kref_get(&reservations
->refs
);
1499 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
1501 struct hstate
*h
= hstate_vma(vma
);
1502 struct resv_map
*reservations
= vma_resv_map(vma
);
1503 unsigned long reserve
;
1504 unsigned long start
;
1508 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
1509 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
1511 reserve
= (end
- start
) -
1512 region_count(&reservations
->regions
, start
, end
);
1514 kref_put(&reservations
->refs
, resv_map_release
);
1517 hugetlb_acct_memory(h
, -reserve
);
1522 * We cannot handle pagefaults against hugetlb pages at all. They cause
1523 * handle_mm_fault() to try to instantiate regular-sized pages in the
1524 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1527 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1533 struct vm_operations_struct hugetlb_vm_ops
= {
1534 .fault
= hugetlb_vm_op_fault
,
1535 .open
= hugetlb_vm_op_open
,
1536 .close
= hugetlb_vm_op_close
,
1539 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
1546 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
1548 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
1550 entry
= pte_mkyoung(entry
);
1551 entry
= pte_mkhuge(entry
);
1556 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
1557 unsigned long address
, pte_t
*ptep
)
1561 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
1562 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
1563 update_mmu_cache(vma
, address
, entry
);
1568 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
1569 struct vm_area_struct
*vma
)
1571 pte_t
*src_pte
, *dst_pte
, entry
;
1572 struct page
*ptepage
;
1575 struct hstate
*h
= hstate_vma(vma
);
1576 unsigned long sz
= huge_page_size(h
);
1578 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
1580 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
1581 src_pte
= huge_pte_offset(src
, addr
);
1584 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
1588 /* If the pagetables are shared don't copy or take references */
1589 if (dst_pte
== src_pte
)
1592 spin_lock(&dst
->page_table_lock
);
1593 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
1594 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
1596 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
1597 entry
= huge_ptep_get(src_pte
);
1598 ptepage
= pte_page(entry
);
1600 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
1602 spin_unlock(&src
->page_table_lock
);
1603 spin_unlock(&dst
->page_table_lock
);
1611 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1612 unsigned long end
, struct page
*ref_page
)
1614 struct mm_struct
*mm
= vma
->vm_mm
;
1615 unsigned long address
;
1620 struct hstate
*h
= hstate_vma(vma
);
1621 unsigned long sz
= huge_page_size(h
);
1624 * A page gathering list, protected by per file i_mmap_lock. The
1625 * lock is used to avoid list corruption from multiple unmapping
1626 * of the same page since we are using page->lru.
1628 LIST_HEAD(page_list
);
1630 WARN_ON(!is_vm_hugetlb_page(vma
));
1631 BUG_ON(start
& ~huge_page_mask(h
));
1632 BUG_ON(end
& ~huge_page_mask(h
));
1634 spin_lock(&mm
->page_table_lock
);
1635 for (address
= start
; address
< end
; address
+= sz
) {
1636 ptep
= huge_pte_offset(mm
, address
);
1640 if (huge_pmd_unshare(mm
, &address
, ptep
))
1644 * If a reference page is supplied, it is because a specific
1645 * page is being unmapped, not a range. Ensure the page we
1646 * are about to unmap is the actual page of interest.
1649 pte
= huge_ptep_get(ptep
);
1650 if (huge_pte_none(pte
))
1652 page
= pte_page(pte
);
1653 if (page
!= ref_page
)
1657 * Mark the VMA as having unmapped its page so that
1658 * future faults in this VMA will fail rather than
1659 * looking like data was lost
1661 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
1664 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1665 if (huge_pte_none(pte
))
1668 page
= pte_page(pte
);
1670 set_page_dirty(page
);
1671 list_add(&page
->lru
, &page_list
);
1673 spin_unlock(&mm
->page_table_lock
);
1674 flush_tlb_range(vma
, start
, end
);
1675 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
1676 list_del(&page
->lru
);
1681 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1682 unsigned long end
, struct page
*ref_page
)
1684 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1685 __unmap_hugepage_range(vma
, start
, end
, ref_page
);
1686 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1690 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1691 * mappping it owns the reserve page for. The intention is to unmap the page
1692 * from other VMAs and let the children be SIGKILLed if they are faulting the
1695 int unmap_ref_private(struct mm_struct
*mm
,
1696 struct vm_area_struct
*vma
,
1698 unsigned long address
)
1700 struct vm_area_struct
*iter_vma
;
1701 struct address_space
*mapping
;
1702 struct prio_tree_iter iter
;
1706 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1707 * from page cache lookup which is in HPAGE_SIZE units.
1709 address
= address
& huge_page_mask(hstate_vma(vma
));
1710 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
)
1711 + (vma
->vm_pgoff
>> PAGE_SHIFT
);
1712 mapping
= (struct address_space
*)page_private(page
);
1714 vma_prio_tree_foreach(iter_vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1715 /* Do not unmap the current VMA */
1716 if (iter_vma
== vma
)
1720 * Unmap the page from other VMAs without their own reserves.
1721 * They get marked to be SIGKILLed if they fault in these
1722 * areas. This is because a future no-page fault on this VMA
1723 * could insert a zeroed page instead of the data existing
1724 * from the time of fork. This would look like data corruption
1726 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
1727 unmap_hugepage_range(iter_vma
,
1728 address
, address
+ HPAGE_SIZE
,
1735 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1736 unsigned long address
, pte_t
*ptep
, pte_t pte
,
1737 struct page
*pagecache_page
)
1739 struct hstate
*h
= hstate_vma(vma
);
1740 struct page
*old_page
, *new_page
;
1742 int outside_reserve
= 0;
1744 old_page
= pte_page(pte
);
1747 /* If no-one else is actually using this page, avoid the copy
1748 * and just make the page writable */
1749 avoidcopy
= (page_count(old_page
) == 1);
1751 set_huge_ptep_writable(vma
, address
, ptep
);
1756 * If the process that created a MAP_PRIVATE mapping is about to
1757 * perform a COW due to a shared page count, attempt to satisfy
1758 * the allocation without using the existing reserves. The pagecache
1759 * page is used to determine if the reserve at this address was
1760 * consumed or not. If reserves were used, a partial faulted mapping
1761 * at the time of fork() could consume its reserves on COW instead
1762 * of the full address range.
1764 if (!(vma
->vm_flags
& VM_SHARED
) &&
1765 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
1766 old_page
!= pagecache_page
)
1767 outside_reserve
= 1;
1769 page_cache_get(old_page
);
1770 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
1772 if (IS_ERR(new_page
)) {
1773 page_cache_release(old_page
);
1776 * If a process owning a MAP_PRIVATE mapping fails to COW,
1777 * it is due to references held by a child and an insufficient
1778 * huge page pool. To guarantee the original mappers
1779 * reliability, unmap the page from child processes. The child
1780 * may get SIGKILLed if it later faults.
1782 if (outside_reserve
) {
1783 BUG_ON(huge_pte_none(pte
));
1784 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
1785 BUG_ON(page_count(old_page
) != 1);
1786 BUG_ON(huge_pte_none(pte
));
1787 goto retry_avoidcopy
;
1792 return -PTR_ERR(new_page
);
1795 spin_unlock(&mm
->page_table_lock
);
1796 copy_huge_page(new_page
, old_page
, address
, vma
);
1797 __SetPageUptodate(new_page
);
1798 spin_lock(&mm
->page_table_lock
);
1800 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
1801 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
1803 huge_ptep_clear_flush(vma
, address
, ptep
);
1804 set_huge_pte_at(mm
, address
, ptep
,
1805 make_huge_pte(vma
, new_page
, 1));
1806 /* Make the old page be freed below */
1807 new_page
= old_page
;
1809 page_cache_release(new_page
);
1810 page_cache_release(old_page
);
1814 /* Return the pagecache page at a given address within a VMA */
1815 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
1816 struct vm_area_struct
*vma
, unsigned long address
)
1818 struct address_space
*mapping
;
1821 mapping
= vma
->vm_file
->f_mapping
;
1822 idx
= vma_hugecache_offset(h
, vma
, address
);
1824 return find_lock_page(mapping
, idx
);
1827 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1828 unsigned long address
, pte_t
*ptep
, int write_access
)
1830 struct hstate
*h
= hstate_vma(vma
);
1831 int ret
= VM_FAULT_SIGBUS
;
1835 struct address_space
*mapping
;
1839 * Currently, we are forced to kill the process in the event the
1840 * original mapper has unmapped pages from the child due to a failed
1841 * COW. Warn that such a situation has occured as it may not be obvious
1843 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
1845 "PID %d killed due to inadequate hugepage pool\n",
1850 mapping
= vma
->vm_file
->f_mapping
;
1851 idx
= vma_hugecache_offset(h
, vma
, address
);
1854 * Use page lock to guard against racing truncation
1855 * before we get page_table_lock.
1858 page
= find_lock_page(mapping
, idx
);
1860 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
1863 page
= alloc_huge_page(vma
, address
, 0);
1865 ret
= -PTR_ERR(page
);
1868 clear_huge_page(page
, address
, huge_page_size(h
));
1869 __SetPageUptodate(page
);
1871 if (vma
->vm_flags
& VM_SHARED
) {
1873 struct inode
*inode
= mapping
->host
;
1875 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
1883 spin_lock(&inode
->i_lock
);
1884 inode
->i_blocks
+= blocks_per_huge_page(h
);
1885 spin_unlock(&inode
->i_lock
);
1890 spin_lock(&mm
->page_table_lock
);
1891 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
1896 if (!huge_pte_none(huge_ptep_get(ptep
)))
1899 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
1900 && (vma
->vm_flags
& VM_SHARED
)));
1901 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
1903 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1904 /* Optimization, do the COW without a second fault */
1905 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
1908 spin_unlock(&mm
->page_table_lock
);
1914 spin_unlock(&mm
->page_table_lock
);
1920 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1921 unsigned long address
, int write_access
)
1926 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
1927 struct hstate
*h
= hstate_vma(vma
);
1929 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
1931 return VM_FAULT_OOM
;
1934 * Serialize hugepage allocation and instantiation, so that we don't
1935 * get spurious allocation failures if two CPUs race to instantiate
1936 * the same page in the page cache.
1938 mutex_lock(&hugetlb_instantiation_mutex
);
1939 entry
= huge_ptep_get(ptep
);
1940 if (huge_pte_none(entry
)) {
1941 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
1942 mutex_unlock(&hugetlb_instantiation_mutex
);
1948 spin_lock(&mm
->page_table_lock
);
1949 /* Check for a racing update before calling hugetlb_cow */
1950 if (likely(pte_same(entry
, huge_ptep_get(ptep
))))
1951 if (write_access
&& !pte_write(entry
)) {
1953 page
= hugetlbfs_pagecache_page(h
, vma
, address
);
1954 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
, page
);
1960 spin_unlock(&mm
->page_table_lock
);
1961 mutex_unlock(&hugetlb_instantiation_mutex
);
1966 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1967 struct page
**pages
, struct vm_area_struct
**vmas
,
1968 unsigned long *position
, int *length
, int i
,
1971 unsigned long pfn_offset
;
1972 unsigned long vaddr
= *position
;
1973 int remainder
= *length
;
1974 struct hstate
*h
= hstate_vma(vma
);
1976 spin_lock(&mm
->page_table_lock
);
1977 while (vaddr
< vma
->vm_end
&& remainder
) {
1982 * Some archs (sparc64, sh*) have multiple pte_ts to
1983 * each hugepage. We have to make * sure we get the
1984 * first, for the page indexing below to work.
1986 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
1988 if (!pte
|| huge_pte_none(huge_ptep_get(pte
)) ||
1989 (write
&& !pte_write(huge_ptep_get(pte
)))) {
1992 spin_unlock(&mm
->page_table_lock
);
1993 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
1994 spin_lock(&mm
->page_table_lock
);
1995 if (!(ret
& VM_FAULT_ERROR
))
2004 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2005 page
= pte_page(huge_ptep_get(pte
));
2009 pages
[i
] = page
+ pfn_offset
;
2019 if (vaddr
< vma
->vm_end
&& remainder
&&
2020 pfn_offset
< pages_per_huge_page(h
)) {
2022 * We use pfn_offset to avoid touching the pageframes
2023 * of this compound page.
2028 spin_unlock(&mm
->page_table_lock
);
2029 *length
= remainder
;
2035 void hugetlb_change_protection(struct vm_area_struct
*vma
,
2036 unsigned long address
, unsigned long end
, pgprot_t newprot
)
2038 struct mm_struct
*mm
= vma
->vm_mm
;
2039 unsigned long start
= address
;
2042 struct hstate
*h
= hstate_vma(vma
);
2044 BUG_ON(address
>= end
);
2045 flush_cache_range(vma
, address
, end
);
2047 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2048 spin_lock(&mm
->page_table_lock
);
2049 for (; address
< end
; address
+= huge_page_size(h
)) {
2050 ptep
= huge_pte_offset(mm
, address
);
2053 if (huge_pmd_unshare(mm
, &address
, ptep
))
2055 if (!huge_pte_none(huge_ptep_get(ptep
))) {
2056 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2057 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
2058 set_huge_pte_at(mm
, address
, ptep
, pte
);
2061 spin_unlock(&mm
->page_table_lock
);
2062 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2064 flush_tlb_range(vma
, start
, end
);
2067 int hugetlb_reserve_pages(struct inode
*inode
,
2069 struct vm_area_struct
*vma
)
2072 struct hstate
*h
= hstate_inode(inode
);
2074 if (vma
&& vma
->vm_flags
& VM_NORESERVE
)
2078 * Shared mappings base their reservation on the number of pages that
2079 * are already allocated on behalf of the file. Private mappings need
2080 * to reserve the full area even if read-only as mprotect() may be
2081 * called to make the mapping read-write. Assume !vma is a shm mapping
2083 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2084 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
2086 struct resv_map
*resv_map
= resv_map_alloc();
2092 set_vma_resv_map(vma
, resv_map
);
2093 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
2099 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
2101 ret
= hugetlb_acct_memory(h
, chg
);
2103 hugetlb_put_quota(inode
->i_mapping
, chg
);
2106 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2107 region_add(&inode
->i_mapping
->private_list
, from
, to
);
2111 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
2113 struct hstate
*h
= hstate_inode(inode
);
2114 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
2116 spin_lock(&inode
->i_lock
);
2117 inode
->i_blocks
-= blocks_per_huge_page(h
);
2118 spin_unlock(&inode
->i_lock
);
2120 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
2121 hugetlb_acct_memory(h
, -(chg
- freed
));