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/sysfs.h>
20 #include <asm/pgtable.h>
22 #include <linux/hugetlb.h>
25 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
26 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
27 unsigned long hugepages_treat_as_movable
;
29 static int max_hstate
;
30 unsigned int default_hstate_idx
;
31 struct hstate hstates
[HUGE_MAX_HSTATE
];
33 /* for command line parsing */
34 static struct hstate
* __initdata parsed_hstate
;
35 static unsigned long __initdata default_hstate_max_huge_pages
;
37 #define for_each_hstate(h) \
38 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
41 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
43 static DEFINE_SPINLOCK(hugetlb_lock
);
46 * Region tracking -- allows tracking of reservations and instantiated pages
47 * across the pages in a mapping.
49 * The region data structures are protected by a combination of the mmap_sem
50 * and the hugetlb_instantion_mutex. To access or modify a region the caller
51 * must either hold the mmap_sem for write, or the mmap_sem for read and
52 * the hugetlb_instantiation mutex:
54 * down_write(&mm->mmap_sem);
56 * down_read(&mm->mmap_sem);
57 * mutex_lock(&hugetlb_instantiation_mutex);
60 struct list_head link
;
65 static long region_add(struct list_head
*head
, long f
, long t
)
67 struct file_region
*rg
, *nrg
, *trg
;
69 /* Locate the region we are either in or before. */
70 list_for_each_entry(rg
, head
, link
)
74 /* Round our left edge to the current segment if it encloses us. */
78 /* Check for and consume any regions we now overlap with. */
80 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
81 if (&rg
->link
== head
)
86 /* If this area reaches higher then extend our area to
87 * include it completely. If this is not the first area
88 * which we intend to reuse, free it. */
101 static long region_chg(struct list_head
*head
, long f
, long t
)
103 struct file_region
*rg
, *nrg
;
106 /* Locate the region we are before or in. */
107 list_for_each_entry(rg
, head
, link
)
111 /* If we are below the current region then a new region is required.
112 * Subtle, allocate a new region at the position but make it zero
113 * size such that we can guarantee to record the reservation. */
114 if (&rg
->link
== head
|| t
< rg
->from
) {
115 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
120 INIT_LIST_HEAD(&nrg
->link
);
121 list_add(&nrg
->link
, rg
->link
.prev
);
126 /* Round our left edge to the current segment if it encloses us. */
131 /* Check for and consume any regions we now overlap with. */
132 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
133 if (&rg
->link
== head
)
138 /* We overlap with this area, if it extends futher than
139 * us then we must extend ourselves. Account for its
140 * existing reservation. */
145 chg
-= rg
->to
- rg
->from
;
150 static long region_truncate(struct list_head
*head
, long end
)
152 struct file_region
*rg
, *trg
;
155 /* Locate the region we are either in or before. */
156 list_for_each_entry(rg
, head
, link
)
159 if (&rg
->link
== head
)
162 /* If we are in the middle of a region then adjust it. */
163 if (end
> rg
->from
) {
166 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
169 /* Drop any remaining regions. */
170 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
171 if (&rg
->link
== head
)
173 chg
+= rg
->to
- rg
->from
;
180 static long region_count(struct list_head
*head
, long f
, long t
)
182 struct file_region
*rg
;
185 /* Locate each segment we overlap with, and count that overlap. */
186 list_for_each_entry(rg
, head
, link
) {
195 seg_from
= max(rg
->from
, f
);
196 seg_to
= min(rg
->to
, t
);
198 chg
+= seg_to
- seg_from
;
205 * Convert the address within this vma to the page offset within
206 * the mapping, in pagecache page units; huge pages here.
208 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
209 struct vm_area_struct
*vma
, unsigned long address
)
211 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
212 (vma
->vm_pgoff
>> huge_page_order(h
));
216 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
217 * bits of the reservation map pointer, which are always clear due to
220 #define HPAGE_RESV_OWNER (1UL << 0)
221 #define HPAGE_RESV_UNMAPPED (1UL << 1)
222 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
225 * These helpers are used to track how many pages are reserved for
226 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
227 * is guaranteed to have their future faults succeed.
229 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
230 * the reserve counters are updated with the hugetlb_lock held. It is safe
231 * to reset the VMA at fork() time as it is not in use yet and there is no
232 * chance of the global counters getting corrupted as a result of the values.
234 * The private mapping reservation is represented in a subtly different
235 * manner to a shared mapping. A shared mapping has a region map associated
236 * with the underlying file, this region map represents the backing file
237 * pages which have ever had a reservation assigned which this persists even
238 * after the page is instantiated. A private mapping has a region map
239 * associated with the original mmap which is attached to all VMAs which
240 * reference it, this region map represents those offsets which have consumed
241 * reservation ie. where pages have been instantiated.
243 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
245 return (unsigned long)vma
->vm_private_data
;
248 static void set_vma_private_data(struct vm_area_struct
*vma
,
251 vma
->vm_private_data
= (void *)value
;
256 struct list_head regions
;
259 struct resv_map
*resv_map_alloc(void)
261 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
265 kref_init(&resv_map
->refs
);
266 INIT_LIST_HEAD(&resv_map
->regions
);
271 void resv_map_release(struct kref
*ref
)
273 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
275 /* Clear out any active regions before we release the map. */
276 region_truncate(&resv_map
->regions
, 0);
280 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
282 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
283 if (!(vma
->vm_flags
& VM_SHARED
))
284 return (struct resv_map
*)(get_vma_private_data(vma
) &
289 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
291 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
292 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
294 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
295 HPAGE_RESV_MASK
) | (unsigned long)map
);
298 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
300 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
301 VM_BUG_ON(vma
->vm_flags
& VM_SHARED
);
303 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
306 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
308 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
310 return (get_vma_private_data(vma
) & flag
) != 0;
313 /* Decrement the reserved pages in the hugepage pool by one */
314 static void decrement_hugepage_resv_vma(struct hstate
*h
,
315 struct vm_area_struct
*vma
)
317 if (vma
->vm_flags
& VM_NORESERVE
)
320 if (vma
->vm_flags
& VM_SHARED
) {
321 /* Shared mappings always use reserves */
322 h
->resv_huge_pages
--;
323 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
325 * Only the process that called mmap() has reserves for
328 h
->resv_huge_pages
--;
332 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
333 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
335 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
336 if (!(vma
->vm_flags
& VM_SHARED
))
337 vma
->vm_private_data
= (void *)0;
340 /* Returns true if the VMA has associated reserve pages */
341 static int vma_has_private_reserves(struct vm_area_struct
*vma
)
343 if (vma
->vm_flags
& VM_SHARED
)
345 if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
350 static void clear_huge_page(struct page
*page
,
351 unsigned long addr
, unsigned long sz
)
356 for (i
= 0; i
< sz
/PAGE_SIZE
; i
++) {
358 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
362 static void copy_huge_page(struct page
*dst
, struct page
*src
,
363 unsigned long addr
, struct vm_area_struct
*vma
)
366 struct hstate
*h
= hstate_vma(vma
);
369 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
371 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
375 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
377 int nid
= page_to_nid(page
);
378 list_add(&page
->lru
, &h
->hugepage_freelists
[nid
]);
379 h
->free_huge_pages
++;
380 h
->free_huge_pages_node
[nid
]++;
383 static struct page
*dequeue_huge_page(struct hstate
*h
)
386 struct page
*page
= NULL
;
388 for (nid
= 0; nid
< MAX_NUMNODES
; ++nid
) {
389 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
390 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
392 list_del(&page
->lru
);
393 h
->free_huge_pages
--;
394 h
->free_huge_pages_node
[nid
]--;
401 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
402 struct vm_area_struct
*vma
,
403 unsigned long address
, int avoid_reserve
)
406 struct page
*page
= NULL
;
407 struct mempolicy
*mpol
;
408 nodemask_t
*nodemask
;
409 struct zonelist
*zonelist
= huge_zonelist(vma
, address
,
410 htlb_alloc_mask
, &mpol
, &nodemask
);
415 * A child process with MAP_PRIVATE mappings created by their parent
416 * have no page reserves. This check ensures that reservations are
417 * not "stolen". The child may still get SIGKILLed
419 if (!vma_has_private_reserves(vma
) &&
420 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
423 /* If reserves cannot be used, ensure enough pages are in the pool */
424 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
427 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
428 MAX_NR_ZONES
- 1, nodemask
) {
429 nid
= zone_to_nid(zone
);
430 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
) &&
431 !list_empty(&h
->hugepage_freelists
[nid
])) {
432 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
434 list_del(&page
->lru
);
435 h
->free_huge_pages
--;
436 h
->free_huge_pages_node
[nid
]--;
439 decrement_hugepage_resv_vma(h
, vma
);
448 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
453 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
454 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
455 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
456 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
457 1 << PG_private
| 1<< PG_writeback
);
459 set_compound_page_dtor(page
, NULL
);
460 set_page_refcounted(page
);
461 arch_release_hugepage(page
);
462 __free_pages(page
, huge_page_order(h
));
465 struct hstate
*size_to_hstate(unsigned long size
)
470 if (huge_page_size(h
) == size
)
476 static void free_huge_page(struct page
*page
)
479 * Can't pass hstate in here because it is called from the
480 * compound page destructor.
482 struct hstate
*h
= page_hstate(page
);
483 int nid
= page_to_nid(page
);
484 struct address_space
*mapping
;
486 mapping
= (struct address_space
*) page_private(page
);
487 set_page_private(page
, 0);
488 BUG_ON(page_count(page
));
489 INIT_LIST_HEAD(&page
->lru
);
491 spin_lock(&hugetlb_lock
);
492 if (h
->surplus_huge_pages_node
[nid
]) {
493 update_and_free_page(h
, page
);
494 h
->surplus_huge_pages
--;
495 h
->surplus_huge_pages_node
[nid
]--;
497 enqueue_huge_page(h
, page
);
499 spin_unlock(&hugetlb_lock
);
501 hugetlb_put_quota(mapping
, 1);
505 * Increment or decrement surplus_huge_pages. Keep node-specific counters
506 * balanced by operating on them in a round-robin fashion.
507 * Returns 1 if an adjustment was made.
509 static int adjust_pool_surplus(struct hstate
*h
, int delta
)
515 VM_BUG_ON(delta
!= -1 && delta
!= 1);
517 nid
= next_node(nid
, node_online_map
);
518 if (nid
== MAX_NUMNODES
)
519 nid
= first_node(node_online_map
);
521 /* To shrink on this node, there must be a surplus page */
522 if (delta
< 0 && !h
->surplus_huge_pages_node
[nid
])
524 /* Surplus cannot exceed the total number of pages */
525 if (delta
> 0 && h
->surplus_huge_pages_node
[nid
] >=
526 h
->nr_huge_pages_node
[nid
])
529 h
->surplus_huge_pages
+= delta
;
530 h
->surplus_huge_pages_node
[nid
] += delta
;
533 } while (nid
!= prev_nid
);
539 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
541 set_compound_page_dtor(page
, free_huge_page
);
542 spin_lock(&hugetlb_lock
);
544 h
->nr_huge_pages_node
[nid
]++;
545 spin_unlock(&hugetlb_lock
);
546 put_page(page
); /* free it into the hugepage allocator */
549 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
553 page
= alloc_pages_node(nid
,
554 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
555 __GFP_REPEAT
|__GFP_NOWARN
,
558 if (arch_prepare_hugepage(page
)) {
559 __free_pages(page
, HUGETLB_PAGE_ORDER
);
562 prep_new_huge_page(h
, page
, nid
);
569 * Use a helper variable to find the next node and then
570 * copy it back to hugetlb_next_nid afterwards:
571 * otherwise there's a window in which a racer might
572 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
573 * But we don't need to use a spin_lock here: it really
574 * doesn't matter if occasionally a racer chooses the
575 * same nid as we do. Move nid forward in the mask even
576 * if we just successfully allocated a hugepage so that
577 * the next caller gets hugepages on the next node.
579 static int hstate_next_node(struct hstate
*h
)
582 next_nid
= next_node(h
->hugetlb_next_nid
, node_online_map
);
583 if (next_nid
== MAX_NUMNODES
)
584 next_nid
= first_node(node_online_map
);
585 h
->hugetlb_next_nid
= next_nid
;
589 static int alloc_fresh_huge_page(struct hstate
*h
)
596 start_nid
= h
->hugetlb_next_nid
;
599 page
= alloc_fresh_huge_page_node(h
, h
->hugetlb_next_nid
);
602 next_nid
= hstate_next_node(h
);
603 } while (!page
&& h
->hugetlb_next_nid
!= start_nid
);
606 count_vm_event(HTLB_BUDDY_PGALLOC
);
608 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
613 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
614 struct vm_area_struct
*vma
, unsigned long address
)
620 * Assume we will successfully allocate the surplus page to
621 * prevent racing processes from causing the surplus to exceed
624 * This however introduces a different race, where a process B
625 * tries to grow the static hugepage pool while alloc_pages() is
626 * called by process A. B will only examine the per-node
627 * counters in determining if surplus huge pages can be
628 * converted to normal huge pages in adjust_pool_surplus(). A
629 * won't be able to increment the per-node counter, until the
630 * lock is dropped by B, but B doesn't drop hugetlb_lock until
631 * no more huge pages can be converted from surplus to normal
632 * state (and doesn't try to convert again). Thus, we have a
633 * case where a surplus huge page exists, the pool is grown, and
634 * the surplus huge page still exists after, even though it
635 * should just have been converted to a normal huge page. This
636 * does not leak memory, though, as the hugepage will be freed
637 * once it is out of use. It also does not allow the counters to
638 * go out of whack in adjust_pool_surplus() as we don't modify
639 * the node values until we've gotten the hugepage and only the
640 * per-node value is checked there.
642 spin_lock(&hugetlb_lock
);
643 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
644 spin_unlock(&hugetlb_lock
);
648 h
->surplus_huge_pages
++;
650 spin_unlock(&hugetlb_lock
);
652 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
653 __GFP_REPEAT
|__GFP_NOWARN
,
656 spin_lock(&hugetlb_lock
);
659 * This page is now managed by the hugetlb allocator and has
660 * no users -- drop the buddy allocator's reference.
662 put_page_testzero(page
);
663 VM_BUG_ON(page_count(page
));
664 nid
= page_to_nid(page
);
665 set_compound_page_dtor(page
, free_huge_page
);
667 * We incremented the global counters already
669 h
->nr_huge_pages_node
[nid
]++;
670 h
->surplus_huge_pages_node
[nid
]++;
671 __count_vm_event(HTLB_BUDDY_PGALLOC
);
674 h
->surplus_huge_pages
--;
675 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
677 spin_unlock(&hugetlb_lock
);
683 * Increase the hugetlb pool such that it can accomodate a reservation
686 static int gather_surplus_pages(struct hstate
*h
, int delta
)
688 struct list_head surplus_list
;
689 struct page
*page
, *tmp
;
691 int needed
, allocated
;
693 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
695 h
->resv_huge_pages
+= delta
;
700 INIT_LIST_HEAD(&surplus_list
);
704 spin_unlock(&hugetlb_lock
);
705 for (i
= 0; i
< needed
; i
++) {
706 page
= alloc_buddy_huge_page(h
, NULL
, 0);
709 * We were not able to allocate enough pages to
710 * satisfy the entire reservation so we free what
711 * we've allocated so far.
713 spin_lock(&hugetlb_lock
);
718 list_add(&page
->lru
, &surplus_list
);
723 * After retaking hugetlb_lock, we need to recalculate 'needed'
724 * because either resv_huge_pages or free_huge_pages may have changed.
726 spin_lock(&hugetlb_lock
);
727 needed
= (h
->resv_huge_pages
+ delta
) -
728 (h
->free_huge_pages
+ allocated
);
733 * The surplus_list now contains _at_least_ the number of extra pages
734 * needed to accomodate the reservation. Add the appropriate number
735 * of pages to the hugetlb pool and free the extras back to the buddy
736 * allocator. Commit the entire reservation here to prevent another
737 * process from stealing the pages as they are added to the pool but
738 * before they are reserved.
741 h
->resv_huge_pages
+= delta
;
744 /* Free the needed pages to the hugetlb pool */
745 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
748 list_del(&page
->lru
);
749 enqueue_huge_page(h
, page
);
752 /* Free unnecessary surplus pages to the buddy allocator */
753 if (!list_empty(&surplus_list
)) {
754 spin_unlock(&hugetlb_lock
);
755 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
756 list_del(&page
->lru
);
758 * The page has a reference count of zero already, so
759 * call free_huge_page directly instead of using
760 * put_page. This must be done with hugetlb_lock
761 * unlocked which is safe because free_huge_page takes
762 * hugetlb_lock before deciding how to free the page.
764 free_huge_page(page
);
766 spin_lock(&hugetlb_lock
);
773 * When releasing a hugetlb pool reservation, any surplus pages that were
774 * allocated to satisfy the reservation must be explicitly freed if they were
777 static void return_unused_surplus_pages(struct hstate
*h
,
778 unsigned long unused_resv_pages
)
782 unsigned long nr_pages
;
785 * We want to release as many surplus pages as possible, spread
786 * evenly across all nodes. Iterate across all nodes until we
787 * can no longer free unreserved surplus pages. This occurs when
788 * the nodes with surplus pages have no free pages.
790 unsigned long remaining_iterations
= num_online_nodes();
792 /* Uncommit the reservation */
793 h
->resv_huge_pages
-= unused_resv_pages
;
795 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
797 while (remaining_iterations
-- && nr_pages
) {
798 nid
= next_node(nid
, node_online_map
);
799 if (nid
== MAX_NUMNODES
)
800 nid
= first_node(node_online_map
);
802 if (!h
->surplus_huge_pages_node
[nid
])
805 if (!list_empty(&h
->hugepage_freelists
[nid
])) {
806 page
= list_entry(h
->hugepage_freelists
[nid
].next
,
808 list_del(&page
->lru
);
809 update_and_free_page(h
, page
);
810 h
->free_huge_pages
--;
811 h
->free_huge_pages_node
[nid
]--;
812 h
->surplus_huge_pages
--;
813 h
->surplus_huge_pages_node
[nid
]--;
815 remaining_iterations
= num_online_nodes();
821 * Determine if the huge page at addr within the vma has an associated
822 * reservation. Where it does not we will need to logically increase
823 * reservation and actually increase quota before an allocation can occur.
824 * Where any new reservation would be required the reservation change is
825 * prepared, but not committed. Once the page has been quota'd allocated
826 * an instantiated the change should be committed via vma_commit_reservation.
827 * No action is required on failure.
829 static int vma_needs_reservation(struct hstate
*h
,
830 struct vm_area_struct
*vma
, unsigned long addr
)
832 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
833 struct inode
*inode
= mapping
->host
;
835 if (vma
->vm_flags
& VM_SHARED
) {
836 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
837 return region_chg(&inode
->i_mapping
->private_list
,
840 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
845 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
846 struct resv_map
*reservations
= vma_resv_map(vma
);
848 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
854 static void vma_commit_reservation(struct hstate
*h
,
855 struct vm_area_struct
*vma
, unsigned long addr
)
857 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
858 struct inode
*inode
= mapping
->host
;
860 if (vma
->vm_flags
& VM_SHARED
) {
861 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
862 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
864 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
865 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
866 struct resv_map
*reservations
= vma_resv_map(vma
);
868 /* Mark this page used in the map. */
869 region_add(&reservations
->regions
, idx
, idx
+ 1);
873 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
874 unsigned long addr
, int avoid_reserve
)
876 struct hstate
*h
= hstate_vma(vma
);
878 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
879 struct inode
*inode
= mapping
->host
;
883 * Processes that did not create the mapping will have no reserves and
884 * will not have accounted against quota. Check that the quota can be
885 * made before satisfying the allocation
886 * MAP_NORESERVE mappings may also need pages and quota allocated
887 * if no reserve mapping overlaps.
889 chg
= vma_needs_reservation(h
, vma
, addr
);
893 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
894 return ERR_PTR(-ENOSPC
);
896 spin_lock(&hugetlb_lock
);
897 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
898 spin_unlock(&hugetlb_lock
);
901 page
= alloc_buddy_huge_page(h
, vma
, addr
);
903 hugetlb_put_quota(inode
->i_mapping
, chg
);
904 return ERR_PTR(-VM_FAULT_OOM
);
908 set_page_refcounted(page
);
909 set_page_private(page
, (unsigned long) mapping
);
911 vma_commit_reservation(h
, vma
, addr
);
916 static void __init
hugetlb_init_one_hstate(struct hstate
*h
)
920 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
921 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
923 h
->hugetlb_next_nid
= first_node(node_online_map
);
925 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
926 if (!alloc_fresh_huge_page(h
))
929 h
->max_huge_pages
= h
->free_huge_pages
= h
->nr_huge_pages
= i
;
932 static void __init
hugetlb_init_hstates(void)
937 hugetlb_init_one_hstate(h
);
941 static void __init
report_hugepages(void)
946 printk(KERN_INFO
"Total HugeTLB memory allocated, "
949 1 << (h
->order
+ PAGE_SHIFT
- 20));
954 #ifdef CONFIG_HIGHMEM
955 static void try_to_free_low(struct hstate
*h
, unsigned long count
)
959 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
960 struct page
*page
, *next
;
961 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
962 list_for_each_entry_safe(page
, next
, freel
, lru
) {
963 if (count
>= h
->nr_huge_pages
)
965 if (PageHighMem(page
))
967 list_del(&page
->lru
);
968 update_and_free_page(h
, page
);
969 h
->free_huge_pages
--;
970 h
->free_huge_pages_node
[page_to_nid(page
)]--;
975 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
)
980 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
981 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
)
983 unsigned long min_count
, ret
;
986 * Increase the pool size
987 * First take pages out of surplus state. Then make up the
988 * remaining difference by allocating fresh huge pages.
990 * We might race with alloc_buddy_huge_page() here and be unable
991 * to convert a surplus huge page to a normal huge page. That is
992 * not critical, though, it just means the overall size of the
993 * pool might be one hugepage larger than it needs to be, but
994 * within all the constraints specified by the sysctls.
996 spin_lock(&hugetlb_lock
);
997 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
998 if (!adjust_pool_surplus(h
, -1))
1002 while (count
> persistent_huge_pages(h
)) {
1004 * If this allocation races such that we no longer need the
1005 * page, free_huge_page will handle it by freeing the page
1006 * and reducing the surplus.
1008 spin_unlock(&hugetlb_lock
);
1009 ret
= alloc_fresh_huge_page(h
);
1010 spin_lock(&hugetlb_lock
);
1017 * Decrease the pool size
1018 * First return free pages to the buddy allocator (being careful
1019 * to keep enough around to satisfy reservations). Then place
1020 * pages into surplus state as needed so the pool will shrink
1021 * to the desired size as pages become free.
1023 * By placing pages into the surplus state independent of the
1024 * overcommit value, we are allowing the surplus pool size to
1025 * exceed overcommit. There are few sane options here. Since
1026 * alloc_buddy_huge_page() is checking the global counter,
1027 * though, we'll note that we're not allowed to exceed surplus
1028 * and won't grow the pool anywhere else. Not until one of the
1029 * sysctls are changed, or the surplus pages go out of use.
1031 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1032 min_count
= max(count
, min_count
);
1033 try_to_free_low(h
, min_count
);
1034 while (min_count
< persistent_huge_pages(h
)) {
1035 struct page
*page
= dequeue_huge_page(h
);
1038 update_and_free_page(h
, page
);
1040 while (count
< persistent_huge_pages(h
)) {
1041 if (!adjust_pool_surplus(h
, 1))
1045 ret
= persistent_huge_pages(h
);
1046 spin_unlock(&hugetlb_lock
);
1050 #define HSTATE_ATTR_RO(_name) \
1051 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1053 #define HSTATE_ATTR(_name) \
1054 static struct kobj_attribute _name##_attr = \
1055 __ATTR(_name, 0644, _name##_show, _name##_store)
1057 static struct kobject
*hugepages_kobj
;
1058 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1060 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
)
1063 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1064 if (hstate_kobjs
[i
] == kobj
)
1070 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1071 struct kobj_attribute
*attr
, char *buf
)
1073 struct hstate
*h
= kobj_to_hstate(kobj
);
1074 return sprintf(buf
, "%lu\n", h
->nr_huge_pages
);
1076 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1077 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1080 unsigned long input
;
1081 struct hstate
*h
= kobj_to_hstate(kobj
);
1083 err
= strict_strtoul(buf
, 10, &input
);
1087 h
->max_huge_pages
= set_max_huge_pages(h
, input
);
1091 HSTATE_ATTR(nr_hugepages
);
1093 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1094 struct kobj_attribute
*attr
, char *buf
)
1096 struct hstate
*h
= kobj_to_hstate(kobj
);
1097 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1099 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1100 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1103 unsigned long input
;
1104 struct hstate
*h
= kobj_to_hstate(kobj
);
1106 err
= strict_strtoul(buf
, 10, &input
);
1110 spin_lock(&hugetlb_lock
);
1111 h
->nr_overcommit_huge_pages
= input
;
1112 spin_unlock(&hugetlb_lock
);
1116 HSTATE_ATTR(nr_overcommit_hugepages
);
1118 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1119 struct kobj_attribute
*attr
, char *buf
)
1121 struct hstate
*h
= kobj_to_hstate(kobj
);
1122 return sprintf(buf
, "%lu\n", h
->free_huge_pages
);
1124 HSTATE_ATTR_RO(free_hugepages
);
1126 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1127 struct kobj_attribute
*attr
, char *buf
)
1129 struct hstate
*h
= kobj_to_hstate(kobj
);
1130 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1132 HSTATE_ATTR_RO(resv_hugepages
);
1134 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1135 struct kobj_attribute
*attr
, char *buf
)
1137 struct hstate
*h
= kobj_to_hstate(kobj
);
1138 return sprintf(buf
, "%lu\n", h
->surplus_huge_pages
);
1140 HSTATE_ATTR_RO(surplus_hugepages
);
1142 static struct attribute
*hstate_attrs
[] = {
1143 &nr_hugepages_attr
.attr
,
1144 &nr_overcommit_hugepages_attr
.attr
,
1145 &free_hugepages_attr
.attr
,
1146 &resv_hugepages_attr
.attr
,
1147 &surplus_hugepages_attr
.attr
,
1151 static struct attribute_group hstate_attr_group
= {
1152 .attrs
= hstate_attrs
,
1155 static int __init
hugetlb_sysfs_add_hstate(struct hstate
*h
)
1159 hstate_kobjs
[h
- hstates
] = kobject_create_and_add(h
->name
,
1161 if (!hstate_kobjs
[h
- hstates
])
1164 retval
= sysfs_create_group(hstate_kobjs
[h
- hstates
],
1165 &hstate_attr_group
);
1167 kobject_put(hstate_kobjs
[h
- hstates
]);
1172 static void __init
hugetlb_sysfs_init(void)
1177 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1178 if (!hugepages_kobj
)
1181 for_each_hstate(h
) {
1182 err
= hugetlb_sysfs_add_hstate(h
);
1184 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s",
1189 static void __exit
hugetlb_exit(void)
1193 for_each_hstate(h
) {
1194 kobject_put(hstate_kobjs
[h
- hstates
]);
1197 kobject_put(hugepages_kobj
);
1199 module_exit(hugetlb_exit
);
1201 static int __init
hugetlb_init(void)
1203 BUILD_BUG_ON(HPAGE_SHIFT
== 0);
1205 if (!size_to_hstate(HPAGE_SIZE
)) {
1206 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1207 parsed_hstate
->max_huge_pages
= default_hstate_max_huge_pages
;
1209 default_hstate_idx
= size_to_hstate(HPAGE_SIZE
) - hstates
;
1211 hugetlb_init_hstates();
1215 hugetlb_sysfs_init();
1219 module_init(hugetlb_init
);
1221 /* Should be called on processing a hugepagesz=... option */
1222 void __init
hugetlb_add_hstate(unsigned order
)
1225 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1226 printk(KERN_WARNING
"hugepagesz= specified twice, ignoring\n");
1229 BUG_ON(max_hstate
>= HUGE_MAX_HSTATE
);
1231 h
= &hstates
[max_hstate
++];
1233 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1234 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1235 huge_page_size(h
)/1024);
1236 hugetlb_init_one_hstate(h
);
1240 static int __init
hugetlb_setup(char *s
)
1245 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1246 * so this hugepages= parameter goes to the "default hstate".
1249 mhp
= &default_hstate_max_huge_pages
;
1251 mhp
= &parsed_hstate
->max_huge_pages
;
1253 if (sscanf(s
, "%lu", mhp
) <= 0)
1258 __setup("hugepages=", hugetlb_setup
);
1260 static unsigned int cpuset_mems_nr(unsigned int *array
)
1263 unsigned int nr
= 0;
1265 for_each_node_mask(node
, cpuset_current_mems_allowed
)
1271 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
1272 struct file
*file
, void __user
*buffer
,
1273 size_t *length
, loff_t
*ppos
)
1275 struct hstate
*h
= &default_hstate
;
1279 tmp
= h
->max_huge_pages
;
1282 table
->maxlen
= sizeof(unsigned long);
1283 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1286 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
);
1291 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
1292 struct file
*file
, void __user
*buffer
,
1293 size_t *length
, loff_t
*ppos
)
1295 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
1296 if (hugepages_treat_as_movable
)
1297 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
1299 htlb_alloc_mask
= GFP_HIGHUSER
;
1303 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
1304 struct file
*file
, void __user
*buffer
,
1305 size_t *length
, loff_t
*ppos
)
1307 struct hstate
*h
= &default_hstate
;
1311 tmp
= h
->nr_overcommit_huge_pages
;
1314 table
->maxlen
= sizeof(unsigned long);
1315 proc_doulongvec_minmax(table
, write
, file
, buffer
, length
, ppos
);
1318 spin_lock(&hugetlb_lock
);
1319 h
->nr_overcommit_huge_pages
= tmp
;
1320 spin_unlock(&hugetlb_lock
);
1326 #endif /* CONFIG_SYSCTL */
1328 int hugetlb_report_meminfo(char *buf
)
1330 struct hstate
*h
= &default_hstate
;
1332 "HugePages_Total: %5lu\n"
1333 "HugePages_Free: %5lu\n"
1334 "HugePages_Rsvd: %5lu\n"
1335 "HugePages_Surp: %5lu\n"
1336 "Hugepagesize: %5lu kB\n",
1340 h
->surplus_huge_pages
,
1341 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
1344 int hugetlb_report_node_meminfo(int nid
, char *buf
)
1346 struct hstate
*h
= &default_hstate
;
1348 "Node %d HugePages_Total: %5u\n"
1349 "Node %d HugePages_Free: %5u\n"
1350 "Node %d HugePages_Surp: %5u\n",
1351 nid
, h
->nr_huge_pages_node
[nid
],
1352 nid
, h
->free_huge_pages_node
[nid
],
1353 nid
, h
->surplus_huge_pages_node
[nid
]);
1356 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1357 unsigned long hugetlb_total_pages(void)
1359 struct hstate
*h
= &default_hstate
;
1360 return h
->nr_huge_pages
* pages_per_huge_page(h
);
1363 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
1367 spin_lock(&hugetlb_lock
);
1369 * When cpuset is configured, it breaks the strict hugetlb page
1370 * reservation as the accounting is done on a global variable. Such
1371 * reservation is completely rubbish in the presence of cpuset because
1372 * the reservation is not checked against page availability for the
1373 * current cpuset. Application can still potentially OOM'ed by kernel
1374 * with lack of free htlb page in cpuset that the task is in.
1375 * Attempt to enforce strict accounting with cpuset is almost
1376 * impossible (or too ugly) because cpuset is too fluid that
1377 * task or memory node can be dynamically moved between cpusets.
1379 * The change of semantics for shared hugetlb mapping with cpuset is
1380 * undesirable. However, in order to preserve some of the semantics,
1381 * we fall back to check against current free page availability as
1382 * a best attempt and hopefully to minimize the impact of changing
1383 * semantics that cpuset has.
1386 if (gather_surplus_pages(h
, delta
) < 0)
1389 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
1390 return_unused_surplus_pages(h
, delta
);
1397 return_unused_surplus_pages(h
, (unsigned long) -delta
);
1400 spin_unlock(&hugetlb_lock
);
1404 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
1406 struct resv_map
*reservations
= vma_resv_map(vma
);
1409 * This new VMA should share its siblings reservation map if present.
1410 * The VMA will only ever have a valid reservation map pointer where
1411 * it is being copied for another still existing VMA. As that VMA
1412 * has a reference to the reservation map it cannot dissappear until
1413 * after this open call completes. It is therefore safe to take a
1414 * new reference here without additional locking.
1417 kref_get(&reservations
->refs
);
1420 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
1422 struct hstate
*h
= hstate_vma(vma
);
1423 struct resv_map
*reservations
= vma_resv_map(vma
);
1424 unsigned long reserve
;
1425 unsigned long start
;
1429 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
1430 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
1432 reserve
= (end
- start
) -
1433 region_count(&reservations
->regions
, start
, end
);
1435 kref_put(&reservations
->refs
, resv_map_release
);
1438 hugetlb_acct_memory(h
, -reserve
);
1443 * We cannot handle pagefaults against hugetlb pages at all. They cause
1444 * handle_mm_fault() to try to instantiate regular-sized pages in the
1445 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1448 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
1454 struct vm_operations_struct hugetlb_vm_ops
= {
1455 .fault
= hugetlb_vm_op_fault
,
1456 .open
= hugetlb_vm_op_open
,
1457 .close
= hugetlb_vm_op_close
,
1460 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
1467 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
1469 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
1471 entry
= pte_mkyoung(entry
);
1472 entry
= pte_mkhuge(entry
);
1477 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
1478 unsigned long address
, pte_t
*ptep
)
1482 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
1483 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
1484 update_mmu_cache(vma
, address
, entry
);
1489 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
1490 struct vm_area_struct
*vma
)
1492 pte_t
*src_pte
, *dst_pte
, entry
;
1493 struct page
*ptepage
;
1496 struct hstate
*h
= hstate_vma(vma
);
1497 unsigned long sz
= huge_page_size(h
);
1499 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
1501 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
1502 src_pte
= huge_pte_offset(src
, addr
);
1505 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
1509 /* If the pagetables are shared don't copy or take references */
1510 if (dst_pte
== src_pte
)
1513 spin_lock(&dst
->page_table_lock
);
1514 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
1515 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
1517 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
1518 entry
= huge_ptep_get(src_pte
);
1519 ptepage
= pte_page(entry
);
1521 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
1523 spin_unlock(&src
->page_table_lock
);
1524 spin_unlock(&dst
->page_table_lock
);
1532 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1533 unsigned long end
, struct page
*ref_page
)
1535 struct mm_struct
*mm
= vma
->vm_mm
;
1536 unsigned long address
;
1541 struct hstate
*h
= hstate_vma(vma
);
1542 unsigned long sz
= huge_page_size(h
);
1545 * A page gathering list, protected by per file i_mmap_lock. The
1546 * lock is used to avoid list corruption from multiple unmapping
1547 * of the same page since we are using page->lru.
1549 LIST_HEAD(page_list
);
1551 WARN_ON(!is_vm_hugetlb_page(vma
));
1552 BUG_ON(start
& ~huge_page_mask(h
));
1553 BUG_ON(end
& ~huge_page_mask(h
));
1555 spin_lock(&mm
->page_table_lock
);
1556 for (address
= start
; address
< end
; address
+= sz
) {
1557 ptep
= huge_pte_offset(mm
, address
);
1561 if (huge_pmd_unshare(mm
, &address
, ptep
))
1565 * If a reference page is supplied, it is because a specific
1566 * page is being unmapped, not a range. Ensure the page we
1567 * are about to unmap is the actual page of interest.
1570 pte
= huge_ptep_get(ptep
);
1571 if (huge_pte_none(pte
))
1573 page
= pte_page(pte
);
1574 if (page
!= ref_page
)
1578 * Mark the VMA as having unmapped its page so that
1579 * future faults in this VMA will fail rather than
1580 * looking like data was lost
1582 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
1585 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1586 if (huge_pte_none(pte
))
1589 page
= pte_page(pte
);
1591 set_page_dirty(page
);
1592 list_add(&page
->lru
, &page_list
);
1594 spin_unlock(&mm
->page_table_lock
);
1595 flush_tlb_range(vma
, start
, end
);
1596 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
1597 list_del(&page
->lru
);
1602 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
1603 unsigned long end
, struct page
*ref_page
)
1605 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1606 __unmap_hugepage_range(vma
, start
, end
, ref_page
);
1607 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1611 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1612 * mappping it owns the reserve page for. The intention is to unmap the page
1613 * from other VMAs and let the children be SIGKILLed if they are faulting the
1616 int unmap_ref_private(struct mm_struct
*mm
,
1617 struct vm_area_struct
*vma
,
1619 unsigned long address
)
1621 struct vm_area_struct
*iter_vma
;
1622 struct address_space
*mapping
;
1623 struct prio_tree_iter iter
;
1627 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1628 * from page cache lookup which is in HPAGE_SIZE units.
1630 address
= address
& huge_page_mask(hstate_vma(vma
));
1631 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
)
1632 + (vma
->vm_pgoff
>> PAGE_SHIFT
);
1633 mapping
= (struct address_space
*)page_private(page
);
1635 vma_prio_tree_foreach(iter_vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
1636 /* Do not unmap the current VMA */
1637 if (iter_vma
== vma
)
1641 * Unmap the page from other VMAs without their own reserves.
1642 * They get marked to be SIGKILLed if they fault in these
1643 * areas. This is because a future no-page fault on this VMA
1644 * could insert a zeroed page instead of the data existing
1645 * from the time of fork. This would look like data corruption
1647 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
1648 unmap_hugepage_range(iter_vma
,
1649 address
, address
+ HPAGE_SIZE
,
1656 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1657 unsigned long address
, pte_t
*ptep
, pte_t pte
,
1658 struct page
*pagecache_page
)
1660 struct hstate
*h
= hstate_vma(vma
);
1661 struct page
*old_page
, *new_page
;
1663 int outside_reserve
= 0;
1665 old_page
= pte_page(pte
);
1668 /* If no-one else is actually using this page, avoid the copy
1669 * and just make the page writable */
1670 avoidcopy
= (page_count(old_page
) == 1);
1672 set_huge_ptep_writable(vma
, address
, ptep
);
1677 * If the process that created a MAP_PRIVATE mapping is about to
1678 * perform a COW due to a shared page count, attempt to satisfy
1679 * the allocation without using the existing reserves. The pagecache
1680 * page is used to determine if the reserve at this address was
1681 * consumed or not. If reserves were used, a partial faulted mapping
1682 * at the time of fork() could consume its reserves on COW instead
1683 * of the full address range.
1685 if (!(vma
->vm_flags
& VM_SHARED
) &&
1686 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
1687 old_page
!= pagecache_page
)
1688 outside_reserve
= 1;
1690 page_cache_get(old_page
);
1691 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
1693 if (IS_ERR(new_page
)) {
1694 page_cache_release(old_page
);
1697 * If a process owning a MAP_PRIVATE mapping fails to COW,
1698 * it is due to references held by a child and an insufficient
1699 * huge page pool. To guarantee the original mappers
1700 * reliability, unmap the page from child processes. The child
1701 * may get SIGKILLed if it later faults.
1703 if (outside_reserve
) {
1704 BUG_ON(huge_pte_none(pte
));
1705 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
1706 BUG_ON(page_count(old_page
) != 1);
1707 BUG_ON(huge_pte_none(pte
));
1708 goto retry_avoidcopy
;
1713 return -PTR_ERR(new_page
);
1716 spin_unlock(&mm
->page_table_lock
);
1717 copy_huge_page(new_page
, old_page
, address
, vma
);
1718 __SetPageUptodate(new_page
);
1719 spin_lock(&mm
->page_table_lock
);
1721 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
1722 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
1724 huge_ptep_clear_flush(vma
, address
, ptep
);
1725 set_huge_pte_at(mm
, address
, ptep
,
1726 make_huge_pte(vma
, new_page
, 1));
1727 /* Make the old page be freed below */
1728 new_page
= old_page
;
1730 page_cache_release(new_page
);
1731 page_cache_release(old_page
);
1735 /* Return the pagecache page at a given address within a VMA */
1736 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
1737 struct vm_area_struct
*vma
, unsigned long address
)
1739 struct address_space
*mapping
;
1742 mapping
= vma
->vm_file
->f_mapping
;
1743 idx
= vma_hugecache_offset(h
, vma
, address
);
1745 return find_lock_page(mapping
, idx
);
1748 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1749 unsigned long address
, pte_t
*ptep
, int write_access
)
1751 struct hstate
*h
= hstate_vma(vma
);
1752 int ret
= VM_FAULT_SIGBUS
;
1756 struct address_space
*mapping
;
1760 * Currently, we are forced to kill the process in the event the
1761 * original mapper has unmapped pages from the child due to a failed
1762 * COW. Warn that such a situation has occured as it may not be obvious
1764 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
1766 "PID %d killed due to inadequate hugepage pool\n",
1771 mapping
= vma
->vm_file
->f_mapping
;
1772 idx
= vma_hugecache_offset(h
, vma
, address
);
1775 * Use page lock to guard against racing truncation
1776 * before we get page_table_lock.
1779 page
= find_lock_page(mapping
, idx
);
1781 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
1784 page
= alloc_huge_page(vma
, address
, 0);
1786 ret
= -PTR_ERR(page
);
1789 clear_huge_page(page
, address
, huge_page_size(h
));
1790 __SetPageUptodate(page
);
1792 if (vma
->vm_flags
& VM_SHARED
) {
1794 struct inode
*inode
= mapping
->host
;
1796 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
1804 spin_lock(&inode
->i_lock
);
1805 inode
->i_blocks
+= blocks_per_huge_page(h
);
1806 spin_unlock(&inode
->i_lock
);
1811 spin_lock(&mm
->page_table_lock
);
1812 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
1817 if (!huge_pte_none(huge_ptep_get(ptep
)))
1820 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
1821 && (vma
->vm_flags
& VM_SHARED
)));
1822 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
1824 if (write_access
&& !(vma
->vm_flags
& VM_SHARED
)) {
1825 /* Optimization, do the COW without a second fault */
1826 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
1829 spin_unlock(&mm
->page_table_lock
);
1835 spin_unlock(&mm
->page_table_lock
);
1841 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1842 unsigned long address
, int write_access
)
1847 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
1848 struct hstate
*h
= hstate_vma(vma
);
1850 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
1852 return VM_FAULT_OOM
;
1855 * Serialize hugepage allocation and instantiation, so that we don't
1856 * get spurious allocation failures if two CPUs race to instantiate
1857 * the same page in the page cache.
1859 mutex_lock(&hugetlb_instantiation_mutex
);
1860 entry
= huge_ptep_get(ptep
);
1861 if (huge_pte_none(entry
)) {
1862 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, write_access
);
1863 mutex_unlock(&hugetlb_instantiation_mutex
);
1869 spin_lock(&mm
->page_table_lock
);
1870 /* Check for a racing update before calling hugetlb_cow */
1871 if (likely(pte_same(entry
, huge_ptep_get(ptep
))))
1872 if (write_access
&& !pte_write(entry
)) {
1874 page
= hugetlbfs_pagecache_page(h
, vma
, address
);
1875 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
, page
);
1881 spin_unlock(&mm
->page_table_lock
);
1882 mutex_unlock(&hugetlb_instantiation_mutex
);
1887 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
1888 struct page
**pages
, struct vm_area_struct
**vmas
,
1889 unsigned long *position
, int *length
, int i
,
1892 unsigned long pfn_offset
;
1893 unsigned long vaddr
= *position
;
1894 int remainder
= *length
;
1895 struct hstate
*h
= hstate_vma(vma
);
1897 spin_lock(&mm
->page_table_lock
);
1898 while (vaddr
< vma
->vm_end
&& remainder
) {
1903 * Some archs (sparc64, sh*) have multiple pte_ts to
1904 * each hugepage. We have to make * sure we get the
1905 * first, for the page indexing below to work.
1907 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
1909 if (!pte
|| huge_pte_none(huge_ptep_get(pte
)) ||
1910 (write
&& !pte_write(huge_ptep_get(pte
)))) {
1913 spin_unlock(&mm
->page_table_lock
);
1914 ret
= hugetlb_fault(mm
, vma
, vaddr
, write
);
1915 spin_lock(&mm
->page_table_lock
);
1916 if (!(ret
& VM_FAULT_ERROR
))
1925 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
1926 page
= pte_page(huge_ptep_get(pte
));
1930 pages
[i
] = page
+ pfn_offset
;
1940 if (vaddr
< vma
->vm_end
&& remainder
&&
1941 pfn_offset
< pages_per_huge_page(h
)) {
1943 * We use pfn_offset to avoid touching the pageframes
1944 * of this compound page.
1949 spin_unlock(&mm
->page_table_lock
);
1950 *length
= remainder
;
1956 void hugetlb_change_protection(struct vm_area_struct
*vma
,
1957 unsigned long address
, unsigned long end
, pgprot_t newprot
)
1959 struct mm_struct
*mm
= vma
->vm_mm
;
1960 unsigned long start
= address
;
1963 struct hstate
*h
= hstate_vma(vma
);
1965 BUG_ON(address
>= end
);
1966 flush_cache_range(vma
, address
, end
);
1968 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1969 spin_lock(&mm
->page_table_lock
);
1970 for (; address
< end
; address
+= huge_page_size(h
)) {
1971 ptep
= huge_pte_offset(mm
, address
);
1974 if (huge_pmd_unshare(mm
, &address
, ptep
))
1976 if (!huge_pte_none(huge_ptep_get(ptep
))) {
1977 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
1978 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
1979 set_huge_pte_at(mm
, address
, ptep
, pte
);
1982 spin_unlock(&mm
->page_table_lock
);
1983 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
1985 flush_tlb_range(vma
, start
, end
);
1988 int hugetlb_reserve_pages(struct inode
*inode
,
1990 struct vm_area_struct
*vma
)
1993 struct hstate
*h
= hstate_inode(inode
);
1995 if (vma
&& vma
->vm_flags
& VM_NORESERVE
)
1999 * Shared mappings base their reservation on the number of pages that
2000 * are already allocated on behalf of the file. Private mappings need
2001 * to reserve the full area even if read-only as mprotect() may be
2002 * called to make the mapping read-write. Assume !vma is a shm mapping
2004 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2005 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
2007 struct resv_map
*resv_map
= resv_map_alloc();
2013 set_vma_resv_map(vma
, resv_map
);
2014 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
2020 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
2022 ret
= hugetlb_acct_memory(h
, chg
);
2024 hugetlb_put_quota(inode
->i_mapping
, chg
);
2027 if (!vma
|| vma
->vm_flags
& VM_SHARED
)
2028 region_add(&inode
->i_mapping
->private_list
, from
, to
);
2032 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
2034 struct hstate
*h
= hstate_inode(inode
);
2035 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
2037 spin_lock(&inode
->i_lock
);
2038 inode
->i_blocks
-= blocks_per_huge_page(h
);
2039 spin_unlock(&inode
->i_lock
);
2041 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
2042 hugetlb_acct_memory(h
, -(chg
- freed
));