2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
26 #include <asm/pgtable.h>
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
33 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
34 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
35 unsigned long hugepages_treat_as_movable
;
37 static int max_hstate
;
38 unsigned int default_hstate_idx
;
39 struct hstate hstates
[HUGE_MAX_HSTATE
];
41 __initdata
LIST_HEAD(huge_boot_pages
);
43 /* for command line parsing */
44 static struct hstate
* __initdata parsed_hstate
;
45 static unsigned long __initdata default_hstate_max_huge_pages
;
46 static unsigned long __initdata default_hstate_size
;
48 #define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 static DEFINE_SPINLOCK(hugetlb_lock
);
57 * Region tracking -- allows tracking of reservations and instantiated pages
58 * across the pages in a mapping.
60 * The region data structures are protected by a combination of the mmap_sem
61 * and the hugetlb_instantion_mutex. To access or modify a region the caller
62 * must either hold the mmap_sem for write, or the mmap_sem for read and
63 * the hugetlb_instantiation mutex:
65 * down_write(&mm->mmap_sem);
67 * down_read(&mm->mmap_sem);
68 * mutex_lock(&hugetlb_instantiation_mutex);
71 struct list_head link
;
76 static long region_add(struct list_head
*head
, long f
, long t
)
78 struct file_region
*rg
, *nrg
, *trg
;
80 /* Locate the region we are either in or before. */
81 list_for_each_entry(rg
, head
, link
)
85 /* Round our left edge to the current segment if it encloses us. */
89 /* Check for and consume any regions we now overlap with. */
91 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
92 if (&rg
->link
== head
)
97 /* If this area reaches higher then extend our area to
98 * include it completely. If this is not the first area
99 * which we intend to reuse, free it. */
112 static long region_chg(struct list_head
*head
, long f
, long t
)
114 struct file_region
*rg
, *nrg
;
117 /* Locate the region we are before or in. */
118 list_for_each_entry(rg
, head
, link
)
122 /* If we are below the current region then a new region is required.
123 * Subtle, allocate a new region at the position but make it zero
124 * size such that we can guarantee to record the reservation. */
125 if (&rg
->link
== head
|| t
< rg
->from
) {
126 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
131 INIT_LIST_HEAD(&nrg
->link
);
132 list_add(&nrg
->link
, rg
->link
.prev
);
137 /* Round our left edge to the current segment if it encloses us. */
142 /* Check for and consume any regions we now overlap with. */
143 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
144 if (&rg
->link
== head
)
149 /* We overlap with this area, if it extends futher than
150 * us then we must extend ourselves. Account for its
151 * existing reservation. */
156 chg
-= rg
->to
- rg
->from
;
161 static long region_truncate(struct list_head
*head
, long end
)
163 struct file_region
*rg
, *trg
;
166 /* Locate the region we are either in or before. */
167 list_for_each_entry(rg
, head
, link
)
170 if (&rg
->link
== head
)
173 /* If we are in the middle of a region then adjust it. */
174 if (end
> rg
->from
) {
177 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
180 /* Drop any remaining regions. */
181 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
182 if (&rg
->link
== head
)
184 chg
+= rg
->to
- rg
->from
;
191 static long region_count(struct list_head
*head
, long f
, long t
)
193 struct file_region
*rg
;
196 /* Locate each segment we overlap with, and count that overlap. */
197 list_for_each_entry(rg
, head
, link
) {
206 seg_from
= max(rg
->from
, f
);
207 seg_to
= min(rg
->to
, t
);
209 chg
+= seg_to
- seg_from
;
216 * Convert the address within this vma to the page offset within
217 * the mapping, in pagecache page units; huge pages here.
219 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
220 struct vm_area_struct
*vma
, unsigned long address
)
222 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
223 (vma
->vm_pgoff
>> huge_page_order(h
));
226 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
227 unsigned long address
)
229 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
233 * Return the size of the pages allocated when backing a VMA. In the majority
234 * cases this will be same size as used by the page table entries.
236 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
238 struct hstate
*hstate
;
240 if (!is_vm_hugetlb_page(vma
))
243 hstate
= hstate_vma(vma
);
245 return 1UL << (hstate
->order
+ PAGE_SHIFT
);
247 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
250 * Return the page size being used by the MMU to back a VMA. In the majority
251 * of cases, the page size used by the kernel matches the MMU size. On
252 * architectures where it differs, an architecture-specific version of this
253 * function is required.
255 #ifndef vma_mmu_pagesize
256 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
258 return vma_kernel_pagesize(vma
);
263 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
264 * bits of the reservation map pointer, which are always clear due to
267 #define HPAGE_RESV_OWNER (1UL << 0)
268 #define HPAGE_RESV_UNMAPPED (1UL << 1)
269 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
272 * These helpers are used to track how many pages are reserved for
273 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274 * is guaranteed to have their future faults succeed.
276 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277 * the reserve counters are updated with the hugetlb_lock held. It is safe
278 * to reset the VMA at fork() time as it is not in use yet and there is no
279 * chance of the global counters getting corrupted as a result of the values.
281 * The private mapping reservation is represented in a subtly different
282 * manner to a shared mapping. A shared mapping has a region map associated
283 * with the underlying file, this region map represents the backing file
284 * pages which have ever had a reservation assigned which this persists even
285 * after the page is instantiated. A private mapping has a region map
286 * associated with the original mmap which is attached to all VMAs which
287 * reference it, this region map represents those offsets which have consumed
288 * reservation ie. where pages have been instantiated.
290 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
292 return (unsigned long)vma
->vm_private_data
;
295 static void set_vma_private_data(struct vm_area_struct
*vma
,
298 vma
->vm_private_data
= (void *)value
;
303 struct list_head regions
;
306 static struct resv_map
*resv_map_alloc(void)
308 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
312 kref_init(&resv_map
->refs
);
313 INIT_LIST_HEAD(&resv_map
->regions
);
318 static void resv_map_release(struct kref
*ref
)
320 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
322 /* Clear out any active regions before we release the map. */
323 region_truncate(&resv_map
->regions
, 0);
327 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
329 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
330 if (!(vma
->vm_flags
& VM_MAYSHARE
))
331 return (struct resv_map
*)(get_vma_private_data(vma
) &
336 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
338 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
339 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
341 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
342 HPAGE_RESV_MASK
) | (unsigned long)map
);
345 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
347 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
348 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
350 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
353 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
355 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
357 return (get_vma_private_data(vma
) & flag
) != 0;
360 /* Decrement the reserved pages in the hugepage pool by one */
361 static void decrement_hugepage_resv_vma(struct hstate
*h
,
362 struct vm_area_struct
*vma
)
364 if (vma
->vm_flags
& VM_NORESERVE
)
367 if (vma
->vm_flags
& VM_MAYSHARE
) {
368 /* Shared mappings always use reserves */
369 h
->resv_huge_pages
--;
370 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
372 * Only the process that called mmap() has reserves for
375 h
->resv_huge_pages
--;
379 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
380 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
382 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
383 if (!(vma
->vm_flags
& VM_MAYSHARE
))
384 vma
->vm_private_data
= (void *)0;
387 /* Returns true if the VMA has associated reserve pages */
388 static int vma_has_reserves(struct vm_area_struct
*vma
)
390 if (vma
->vm_flags
& VM_MAYSHARE
)
392 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
397 static void clear_gigantic_page(struct page
*page
,
398 unsigned long addr
, unsigned long sz
)
401 struct page
*p
= page
;
404 for (i
= 0; i
< sz
/PAGE_SIZE
; i
++, p
= mem_map_next(p
, page
, i
)) {
406 clear_user_highpage(p
, addr
+ i
* PAGE_SIZE
);
409 static void clear_huge_page(struct page
*page
,
410 unsigned long addr
, unsigned long sz
)
414 if (unlikely(sz
/PAGE_SIZE
> MAX_ORDER_NR_PAGES
)) {
415 clear_gigantic_page(page
, addr
, sz
);
420 for (i
= 0; i
< sz
/PAGE_SIZE
; i
++) {
422 clear_user_highpage(page
+ i
, addr
+ i
* PAGE_SIZE
);
426 static void copy_gigantic_page(struct page
*dst
, struct page
*src
,
427 unsigned long addr
, struct vm_area_struct
*vma
)
430 struct hstate
*h
= hstate_vma(vma
);
431 struct page
*dst_base
= dst
;
432 struct page
*src_base
= src
;
434 for (i
= 0; i
< pages_per_huge_page(h
); ) {
436 copy_user_highpage(dst
, src
, addr
+ i
*PAGE_SIZE
, vma
);
439 dst
= mem_map_next(dst
, dst_base
, i
);
440 src
= mem_map_next(src
, src_base
, i
);
443 static void copy_huge_page(struct page
*dst
, struct page
*src
,
444 unsigned long addr
, struct vm_area_struct
*vma
)
447 struct hstate
*h
= hstate_vma(vma
);
449 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
450 copy_gigantic_page(dst
, src
, addr
, vma
);
455 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
457 copy_user_highpage(dst
+ i
, src
+ i
, addr
+ i
*PAGE_SIZE
, vma
);
461 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
463 int nid
= page_to_nid(page
);
464 list_add(&page
->lru
, &h
->hugepage_freelists
[nid
]);
465 h
->free_huge_pages
++;
466 h
->free_huge_pages_node
[nid
]++;
469 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
473 if (list_empty(&h
->hugepage_freelists
[nid
]))
475 page
= list_entry(h
->hugepage_freelists
[nid
].next
, struct page
, lru
);
476 list_del(&page
->lru
);
477 h
->free_huge_pages
--;
478 h
->free_huge_pages_node
[nid
]--;
482 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
483 struct vm_area_struct
*vma
,
484 unsigned long address
, int avoid_reserve
)
486 struct page
*page
= NULL
;
487 struct mempolicy
*mpol
;
488 nodemask_t
*nodemask
;
489 struct zonelist
*zonelist
;
494 zonelist
= huge_zonelist(vma
, address
,
495 htlb_alloc_mask
, &mpol
, &nodemask
);
497 * A child process with MAP_PRIVATE mappings created by their parent
498 * have no page reserves. This check ensures that reservations are
499 * not "stolen". The child may still get SIGKILLed
501 if (!vma_has_reserves(vma
) &&
502 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
505 /* If reserves cannot be used, ensure enough pages are in the pool */
506 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
509 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
510 MAX_NR_ZONES
- 1, nodemask
) {
511 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
512 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
515 decrement_hugepage_resv_vma(h
, vma
);
526 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
530 VM_BUG_ON(h
->order
>= MAX_ORDER
);
533 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
534 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
535 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
| 1 << PG_referenced
|
536 1 << PG_dirty
| 1 << PG_active
| 1 << PG_reserved
|
537 1 << PG_private
| 1<< PG_writeback
);
539 set_compound_page_dtor(page
, NULL
);
540 set_page_refcounted(page
);
541 arch_release_hugepage(page
);
542 __free_pages(page
, huge_page_order(h
));
545 struct hstate
*size_to_hstate(unsigned long size
)
550 if (huge_page_size(h
) == size
)
556 static void free_huge_page(struct page
*page
)
559 * Can't pass hstate in here because it is called from the
560 * compound page destructor.
562 struct hstate
*h
= page_hstate(page
);
563 int nid
= page_to_nid(page
);
564 struct address_space
*mapping
;
566 mapping
= (struct address_space
*) page_private(page
);
567 set_page_private(page
, 0);
568 page
->mapping
= NULL
;
569 BUG_ON(page_count(page
));
570 BUG_ON(page_mapcount(page
));
571 INIT_LIST_HEAD(&page
->lru
);
573 spin_lock(&hugetlb_lock
);
574 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
575 update_and_free_page(h
, page
);
576 h
->surplus_huge_pages
--;
577 h
->surplus_huge_pages_node
[nid
]--;
579 enqueue_huge_page(h
, page
);
581 spin_unlock(&hugetlb_lock
);
583 hugetlb_put_quota(mapping
, 1);
586 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
588 set_compound_page_dtor(page
, free_huge_page
);
589 spin_lock(&hugetlb_lock
);
591 h
->nr_huge_pages_node
[nid
]++;
592 spin_unlock(&hugetlb_lock
);
593 put_page(page
); /* free it into the hugepage allocator */
596 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
599 int nr_pages
= 1 << order
;
600 struct page
*p
= page
+ 1;
602 /* we rely on prep_new_huge_page to set the destructor */
603 set_compound_order(page
, order
);
605 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
607 p
->first_page
= page
;
611 int PageHuge(struct page
*page
)
613 compound_page_dtor
*dtor
;
615 if (!PageCompound(page
))
618 page
= compound_head(page
);
619 dtor
= get_compound_page_dtor(page
);
621 return dtor
== free_huge_page
;
624 EXPORT_SYMBOL_GPL(PageHuge
);
626 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
630 if (h
->order
>= MAX_ORDER
)
633 page
= alloc_pages_exact_node(nid
,
634 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
635 __GFP_REPEAT
|__GFP_NOWARN
,
638 if (arch_prepare_hugepage(page
)) {
639 __free_pages(page
, huge_page_order(h
));
642 prep_new_huge_page(h
, page
, nid
);
649 * common helper functions for hstate_next_node_to_{alloc|free}.
650 * We may have allocated or freed a huge page based on a different
651 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
652 * be outside of *nodes_allowed. Ensure that we use an allowed
653 * node for alloc or free.
655 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
657 nid
= next_node(nid
, *nodes_allowed
);
658 if (nid
== MAX_NUMNODES
)
659 nid
= first_node(*nodes_allowed
);
660 VM_BUG_ON(nid
>= MAX_NUMNODES
);
665 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
667 if (!node_isset(nid
, *nodes_allowed
))
668 nid
= next_node_allowed(nid
, nodes_allowed
);
673 * returns the previously saved node ["this node"] from which to
674 * allocate a persistent huge page for the pool and advance the
675 * next node from which to allocate, handling wrap at end of node
678 static int hstate_next_node_to_alloc(struct hstate
*h
,
679 nodemask_t
*nodes_allowed
)
683 VM_BUG_ON(!nodes_allowed
);
685 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
686 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
691 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
698 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
699 next_nid
= start_nid
;
702 page
= alloc_fresh_huge_page_node(h
, next_nid
);
707 next_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
708 } while (next_nid
!= start_nid
);
711 count_vm_event(HTLB_BUDDY_PGALLOC
);
713 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
719 * helper for free_pool_huge_page() - return the previously saved
720 * node ["this node"] from which to free a huge page. Advance the
721 * next node id whether or not we find a free huge page to free so
722 * that the next attempt to free addresses the next node.
724 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
728 VM_BUG_ON(!nodes_allowed
);
730 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
731 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
737 * Free huge page from pool from next node to free.
738 * Attempt to keep persistent huge pages more or less
739 * balanced over allowed nodes.
740 * Called with hugetlb_lock locked.
742 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
749 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
750 next_nid
= start_nid
;
754 * If we're returning unused surplus pages, only examine
755 * nodes with surplus pages.
757 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[next_nid
]) &&
758 !list_empty(&h
->hugepage_freelists
[next_nid
])) {
760 list_entry(h
->hugepage_freelists
[next_nid
].next
,
762 list_del(&page
->lru
);
763 h
->free_huge_pages
--;
764 h
->free_huge_pages_node
[next_nid
]--;
766 h
->surplus_huge_pages
--;
767 h
->surplus_huge_pages_node
[next_nid
]--;
769 update_and_free_page(h
, page
);
773 next_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
774 } while (next_nid
!= start_nid
);
779 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
784 if (h
->order
>= MAX_ORDER
)
788 * Assume we will successfully allocate the surplus page to
789 * prevent racing processes from causing the surplus to exceed
792 * This however introduces a different race, where a process B
793 * tries to grow the static hugepage pool while alloc_pages() is
794 * called by process A. B will only examine the per-node
795 * counters in determining if surplus huge pages can be
796 * converted to normal huge pages in adjust_pool_surplus(). A
797 * won't be able to increment the per-node counter, until the
798 * lock is dropped by B, but B doesn't drop hugetlb_lock until
799 * no more huge pages can be converted from surplus to normal
800 * state (and doesn't try to convert again). Thus, we have a
801 * case where a surplus huge page exists, the pool is grown, and
802 * the surplus huge page still exists after, even though it
803 * should just have been converted to a normal huge page. This
804 * does not leak memory, though, as the hugepage will be freed
805 * once it is out of use. It also does not allow the counters to
806 * go out of whack in adjust_pool_surplus() as we don't modify
807 * the node values until we've gotten the hugepage and only the
808 * per-node value is checked there.
810 spin_lock(&hugetlb_lock
);
811 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
812 spin_unlock(&hugetlb_lock
);
816 h
->surplus_huge_pages
++;
818 spin_unlock(&hugetlb_lock
);
820 if (nid
== NUMA_NO_NODE
)
821 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
822 __GFP_REPEAT
|__GFP_NOWARN
,
825 page
= alloc_pages_exact_node(nid
,
826 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
827 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
829 if (page
&& arch_prepare_hugepage(page
)) {
830 __free_pages(page
, huge_page_order(h
));
834 spin_lock(&hugetlb_lock
);
837 * This page is now managed by the hugetlb allocator and has
838 * no users -- drop the buddy allocator's reference.
840 put_page_testzero(page
);
841 VM_BUG_ON(page_count(page
));
842 r_nid
= page_to_nid(page
);
843 set_compound_page_dtor(page
, free_huge_page
);
845 * We incremented the global counters already
847 h
->nr_huge_pages_node
[r_nid
]++;
848 h
->surplus_huge_pages_node
[r_nid
]++;
849 __count_vm_event(HTLB_BUDDY_PGALLOC
);
852 h
->surplus_huge_pages
--;
853 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
855 spin_unlock(&hugetlb_lock
);
861 * This allocation function is useful in the context where vma is irrelevant.
862 * E.g. soft-offlining uses this function because it only cares physical
863 * address of error page.
865 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
869 spin_lock(&hugetlb_lock
);
870 page
= dequeue_huge_page_node(h
, nid
);
871 spin_unlock(&hugetlb_lock
);
874 page
= alloc_buddy_huge_page(h
, nid
);
880 * Increase the hugetlb pool such that it can accomodate a reservation
883 static int gather_surplus_pages(struct hstate
*h
, int delta
)
885 struct list_head surplus_list
;
886 struct page
*page
, *tmp
;
888 int needed
, allocated
;
890 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
892 h
->resv_huge_pages
+= delta
;
897 INIT_LIST_HEAD(&surplus_list
);
901 spin_unlock(&hugetlb_lock
);
902 for (i
= 0; i
< needed
; i
++) {
903 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
906 * We were not able to allocate enough pages to
907 * satisfy the entire reservation so we free what
908 * we've allocated so far.
910 spin_lock(&hugetlb_lock
);
915 list_add(&page
->lru
, &surplus_list
);
920 * After retaking hugetlb_lock, we need to recalculate 'needed'
921 * because either resv_huge_pages or free_huge_pages may have changed.
923 spin_lock(&hugetlb_lock
);
924 needed
= (h
->resv_huge_pages
+ delta
) -
925 (h
->free_huge_pages
+ allocated
);
930 * The surplus_list now contains _at_least_ the number of extra pages
931 * needed to accomodate the reservation. Add the appropriate number
932 * of pages to the hugetlb pool and free the extras back to the buddy
933 * allocator. Commit the entire reservation here to prevent another
934 * process from stealing the pages as they are added to the pool but
935 * before they are reserved.
938 h
->resv_huge_pages
+= delta
;
941 /* Free the needed pages to the hugetlb pool */
942 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
945 list_del(&page
->lru
);
946 enqueue_huge_page(h
, page
);
949 /* Free unnecessary surplus pages to the buddy allocator */
950 if (!list_empty(&surplus_list
)) {
951 spin_unlock(&hugetlb_lock
);
952 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
953 list_del(&page
->lru
);
955 * The page has a reference count of zero already, so
956 * call free_huge_page directly instead of using
957 * put_page. This must be done with hugetlb_lock
958 * unlocked which is safe because free_huge_page takes
959 * hugetlb_lock before deciding how to free the page.
961 free_huge_page(page
);
963 spin_lock(&hugetlb_lock
);
970 * When releasing a hugetlb pool reservation, any surplus pages that were
971 * allocated to satisfy the reservation must be explicitly freed if they were
973 * Called with hugetlb_lock held.
975 static void return_unused_surplus_pages(struct hstate
*h
,
976 unsigned long unused_resv_pages
)
978 unsigned long nr_pages
;
980 /* Uncommit the reservation */
981 h
->resv_huge_pages
-= unused_resv_pages
;
983 /* Cannot return gigantic pages currently */
984 if (h
->order
>= MAX_ORDER
)
987 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
990 * We want to release as many surplus pages as possible, spread
991 * evenly across all nodes with memory. Iterate across these nodes
992 * until we can no longer free unreserved surplus pages. This occurs
993 * when the nodes with surplus pages have no free pages.
994 * free_pool_huge_page() will balance the the freed pages across the
995 * on-line nodes with memory and will handle the hstate accounting.
998 if (!free_pool_huge_page(h
, &node_states
[N_HIGH_MEMORY
], 1))
1004 * Determine if the huge page at addr within the vma has an associated
1005 * reservation. Where it does not we will need to logically increase
1006 * reservation and actually increase quota before an allocation can occur.
1007 * Where any new reservation would be required the reservation change is
1008 * prepared, but not committed. Once the page has been quota'd allocated
1009 * an instantiated the change should be committed via vma_commit_reservation.
1010 * No action is required on failure.
1012 static long vma_needs_reservation(struct hstate
*h
,
1013 struct vm_area_struct
*vma
, unsigned long addr
)
1015 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1016 struct inode
*inode
= mapping
->host
;
1018 if (vma
->vm_flags
& VM_MAYSHARE
) {
1019 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1020 return region_chg(&inode
->i_mapping
->private_list
,
1023 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1028 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1029 struct resv_map
*reservations
= vma_resv_map(vma
);
1031 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
1037 static void vma_commit_reservation(struct hstate
*h
,
1038 struct vm_area_struct
*vma
, unsigned long addr
)
1040 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1041 struct inode
*inode
= mapping
->host
;
1043 if (vma
->vm_flags
& VM_MAYSHARE
) {
1044 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1045 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1047 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1048 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1049 struct resv_map
*reservations
= vma_resv_map(vma
);
1051 /* Mark this page used in the map. */
1052 region_add(&reservations
->regions
, idx
, idx
+ 1);
1056 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1057 unsigned long addr
, int avoid_reserve
)
1059 struct hstate
*h
= hstate_vma(vma
);
1061 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1062 struct inode
*inode
= mapping
->host
;
1066 * Processes that did not create the mapping will have no reserves and
1067 * will not have accounted against quota. Check that the quota can be
1068 * made before satisfying the allocation
1069 * MAP_NORESERVE mappings may also need pages and quota allocated
1070 * if no reserve mapping overlaps.
1072 chg
= vma_needs_reservation(h
, vma
, addr
);
1074 return ERR_PTR(chg
);
1076 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
1077 return ERR_PTR(-ENOSPC
);
1079 spin_lock(&hugetlb_lock
);
1080 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
1081 spin_unlock(&hugetlb_lock
);
1084 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1086 hugetlb_put_quota(inode
->i_mapping
, chg
);
1087 return ERR_PTR(-VM_FAULT_SIGBUS
);
1091 set_page_refcounted(page
);
1092 set_page_private(page
, (unsigned long) mapping
);
1094 vma_commit_reservation(h
, vma
, addr
);
1099 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1101 struct huge_bootmem_page
*m
;
1102 int nr_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
1107 addr
= __alloc_bootmem_node_nopanic(
1108 NODE_DATA(hstate_next_node_to_alloc(h
,
1109 &node_states
[N_HIGH_MEMORY
])),
1110 huge_page_size(h
), huge_page_size(h
), 0);
1114 * Use the beginning of the huge page to store the
1115 * huge_bootmem_page struct (until gather_bootmem
1116 * puts them into the mem_map).
1126 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1127 /* Put them into a private list first because mem_map is not up yet */
1128 list_add(&m
->list
, &huge_boot_pages
);
1133 static void prep_compound_huge_page(struct page
*page
, int order
)
1135 if (unlikely(order
> (MAX_ORDER
- 1)))
1136 prep_compound_gigantic_page(page
, order
);
1138 prep_compound_page(page
, order
);
1141 /* Put bootmem huge pages into the standard lists after mem_map is up */
1142 static void __init
gather_bootmem_prealloc(void)
1144 struct huge_bootmem_page
*m
;
1146 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1147 struct page
*page
= virt_to_page(m
);
1148 struct hstate
*h
= m
->hstate
;
1149 __ClearPageReserved(page
);
1150 WARN_ON(page_count(page
) != 1);
1151 prep_compound_huge_page(page
, h
->order
);
1152 prep_new_huge_page(h
, page
, page_to_nid(page
));
1156 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1160 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1161 if (h
->order
>= MAX_ORDER
) {
1162 if (!alloc_bootmem_huge_page(h
))
1164 } else if (!alloc_fresh_huge_page(h
,
1165 &node_states
[N_HIGH_MEMORY
]))
1168 h
->max_huge_pages
= i
;
1171 static void __init
hugetlb_init_hstates(void)
1175 for_each_hstate(h
) {
1176 /* oversize hugepages were init'ed in early boot */
1177 if (h
->order
< MAX_ORDER
)
1178 hugetlb_hstate_alloc_pages(h
);
1182 static char * __init
memfmt(char *buf
, unsigned long n
)
1184 if (n
>= (1UL << 30))
1185 sprintf(buf
, "%lu GB", n
>> 30);
1186 else if (n
>= (1UL << 20))
1187 sprintf(buf
, "%lu MB", n
>> 20);
1189 sprintf(buf
, "%lu KB", n
>> 10);
1193 static void __init
report_hugepages(void)
1197 for_each_hstate(h
) {
1199 printk(KERN_INFO
"HugeTLB registered %s page size, "
1200 "pre-allocated %ld pages\n",
1201 memfmt(buf
, huge_page_size(h
)),
1202 h
->free_huge_pages
);
1206 #ifdef CONFIG_HIGHMEM
1207 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1208 nodemask_t
*nodes_allowed
)
1212 if (h
->order
>= MAX_ORDER
)
1215 for_each_node_mask(i
, *nodes_allowed
) {
1216 struct page
*page
, *next
;
1217 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1218 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1219 if (count
>= h
->nr_huge_pages
)
1221 if (PageHighMem(page
))
1223 list_del(&page
->lru
);
1224 update_and_free_page(h
, page
);
1225 h
->free_huge_pages
--;
1226 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1231 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1232 nodemask_t
*nodes_allowed
)
1238 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1239 * balanced by operating on them in a round-robin fashion.
1240 * Returns 1 if an adjustment was made.
1242 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1245 int start_nid
, next_nid
;
1248 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1251 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
1253 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
1254 next_nid
= start_nid
;
1260 * To shrink on this node, there must be a surplus page
1262 if (!h
->surplus_huge_pages_node
[nid
]) {
1263 next_nid
= hstate_next_node_to_alloc(h
,
1270 * Surplus cannot exceed the total number of pages
1272 if (h
->surplus_huge_pages_node
[nid
] >=
1273 h
->nr_huge_pages_node
[nid
]) {
1274 next_nid
= hstate_next_node_to_free(h
,
1280 h
->surplus_huge_pages
+= delta
;
1281 h
->surplus_huge_pages_node
[nid
] += delta
;
1284 } while (next_nid
!= start_nid
);
1289 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1290 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1291 nodemask_t
*nodes_allowed
)
1293 unsigned long min_count
, ret
;
1295 if (h
->order
>= MAX_ORDER
)
1296 return h
->max_huge_pages
;
1299 * Increase the pool size
1300 * First take pages out of surplus state. Then make up the
1301 * remaining difference by allocating fresh huge pages.
1303 * We might race with alloc_buddy_huge_page() here and be unable
1304 * to convert a surplus huge page to a normal huge page. That is
1305 * not critical, though, it just means the overall size of the
1306 * pool might be one hugepage larger than it needs to be, but
1307 * within all the constraints specified by the sysctls.
1309 spin_lock(&hugetlb_lock
);
1310 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1311 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1315 while (count
> persistent_huge_pages(h
)) {
1317 * If this allocation races such that we no longer need the
1318 * page, free_huge_page will handle it by freeing the page
1319 * and reducing the surplus.
1321 spin_unlock(&hugetlb_lock
);
1322 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1323 spin_lock(&hugetlb_lock
);
1327 /* Bail for signals. Probably ctrl-c from user */
1328 if (signal_pending(current
))
1333 * Decrease the pool size
1334 * First return free pages to the buddy allocator (being careful
1335 * to keep enough around to satisfy reservations). Then place
1336 * pages into surplus state as needed so the pool will shrink
1337 * to the desired size as pages become free.
1339 * By placing pages into the surplus state independent of the
1340 * overcommit value, we are allowing the surplus pool size to
1341 * exceed overcommit. There are few sane options here. Since
1342 * alloc_buddy_huge_page() is checking the global counter,
1343 * though, we'll note that we're not allowed to exceed surplus
1344 * and won't grow the pool anywhere else. Not until one of the
1345 * sysctls are changed, or the surplus pages go out of use.
1347 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1348 min_count
= max(count
, min_count
);
1349 try_to_free_low(h
, min_count
, nodes_allowed
);
1350 while (min_count
< persistent_huge_pages(h
)) {
1351 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1354 while (count
< persistent_huge_pages(h
)) {
1355 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1359 ret
= persistent_huge_pages(h
);
1360 spin_unlock(&hugetlb_lock
);
1364 #define HSTATE_ATTR_RO(_name) \
1365 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1367 #define HSTATE_ATTR(_name) \
1368 static struct kobj_attribute _name##_attr = \
1369 __ATTR(_name, 0644, _name##_show, _name##_store)
1371 static struct kobject
*hugepages_kobj
;
1372 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1374 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1376 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1380 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1381 if (hstate_kobjs
[i
] == kobj
) {
1383 *nidp
= NUMA_NO_NODE
;
1387 return kobj_to_node_hstate(kobj
, nidp
);
1390 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1391 struct kobj_attribute
*attr
, char *buf
)
1394 unsigned long nr_huge_pages
;
1397 h
= kobj_to_hstate(kobj
, &nid
);
1398 if (nid
== NUMA_NO_NODE
)
1399 nr_huge_pages
= h
->nr_huge_pages
;
1401 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1403 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1405 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1406 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1407 const char *buf
, size_t len
)
1411 unsigned long count
;
1413 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1415 err
= strict_strtoul(buf
, 10, &count
);
1419 h
= kobj_to_hstate(kobj
, &nid
);
1420 if (nid
== NUMA_NO_NODE
) {
1422 * global hstate attribute
1424 if (!(obey_mempolicy
&&
1425 init_nodemask_of_mempolicy(nodes_allowed
))) {
1426 NODEMASK_FREE(nodes_allowed
);
1427 nodes_allowed
= &node_states
[N_HIGH_MEMORY
];
1429 } else if (nodes_allowed
) {
1431 * per node hstate attribute: adjust count to global,
1432 * but restrict alloc/free to the specified node.
1434 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1435 init_nodemask_of_node(nodes_allowed
, nid
);
1437 nodes_allowed
= &node_states
[N_HIGH_MEMORY
];
1439 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1441 if (nodes_allowed
!= &node_states
[N_HIGH_MEMORY
])
1442 NODEMASK_FREE(nodes_allowed
);
1447 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1448 struct kobj_attribute
*attr
, char *buf
)
1450 return nr_hugepages_show_common(kobj
, attr
, buf
);
1453 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1454 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1456 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1458 HSTATE_ATTR(nr_hugepages
);
1463 * hstate attribute for optionally mempolicy-based constraint on persistent
1464 * huge page alloc/free.
1466 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1467 struct kobj_attribute
*attr
, char *buf
)
1469 return nr_hugepages_show_common(kobj
, attr
, buf
);
1472 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1473 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1475 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1477 HSTATE_ATTR(nr_hugepages_mempolicy
);
1481 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1482 struct kobj_attribute
*attr
, char *buf
)
1484 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1485 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1487 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1488 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1491 unsigned long input
;
1492 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1494 err
= strict_strtoul(buf
, 10, &input
);
1498 spin_lock(&hugetlb_lock
);
1499 h
->nr_overcommit_huge_pages
= input
;
1500 spin_unlock(&hugetlb_lock
);
1504 HSTATE_ATTR(nr_overcommit_hugepages
);
1506 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1507 struct kobj_attribute
*attr
, char *buf
)
1510 unsigned long free_huge_pages
;
1513 h
= kobj_to_hstate(kobj
, &nid
);
1514 if (nid
== NUMA_NO_NODE
)
1515 free_huge_pages
= h
->free_huge_pages
;
1517 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1519 return sprintf(buf
, "%lu\n", free_huge_pages
);
1521 HSTATE_ATTR_RO(free_hugepages
);
1523 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1524 struct kobj_attribute
*attr
, char *buf
)
1526 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1527 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1529 HSTATE_ATTR_RO(resv_hugepages
);
1531 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1532 struct kobj_attribute
*attr
, char *buf
)
1535 unsigned long surplus_huge_pages
;
1538 h
= kobj_to_hstate(kobj
, &nid
);
1539 if (nid
== NUMA_NO_NODE
)
1540 surplus_huge_pages
= h
->surplus_huge_pages
;
1542 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1544 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1546 HSTATE_ATTR_RO(surplus_hugepages
);
1548 static struct attribute
*hstate_attrs
[] = {
1549 &nr_hugepages_attr
.attr
,
1550 &nr_overcommit_hugepages_attr
.attr
,
1551 &free_hugepages_attr
.attr
,
1552 &resv_hugepages_attr
.attr
,
1553 &surplus_hugepages_attr
.attr
,
1555 &nr_hugepages_mempolicy_attr
.attr
,
1560 static struct attribute_group hstate_attr_group
= {
1561 .attrs
= hstate_attrs
,
1564 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1565 struct kobject
**hstate_kobjs
,
1566 struct attribute_group
*hstate_attr_group
)
1569 int hi
= h
- hstates
;
1571 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1572 if (!hstate_kobjs
[hi
])
1575 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1577 kobject_put(hstate_kobjs
[hi
]);
1582 static void __init
hugetlb_sysfs_init(void)
1587 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1588 if (!hugepages_kobj
)
1591 for_each_hstate(h
) {
1592 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1593 hstate_kobjs
, &hstate_attr_group
);
1595 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s",
1603 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1604 * with node sysdevs in node_devices[] using a parallel array. The array
1605 * index of a node sysdev or _hstate == node id.
1606 * This is here to avoid any static dependency of the node sysdev driver, in
1607 * the base kernel, on the hugetlb module.
1609 struct node_hstate
{
1610 struct kobject
*hugepages_kobj
;
1611 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1613 struct node_hstate node_hstates
[MAX_NUMNODES
];
1616 * A subset of global hstate attributes for node sysdevs
1618 static struct attribute
*per_node_hstate_attrs
[] = {
1619 &nr_hugepages_attr
.attr
,
1620 &free_hugepages_attr
.attr
,
1621 &surplus_hugepages_attr
.attr
,
1625 static struct attribute_group per_node_hstate_attr_group
= {
1626 .attrs
= per_node_hstate_attrs
,
1630 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1631 * Returns node id via non-NULL nidp.
1633 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1637 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1638 struct node_hstate
*nhs
= &node_hstates
[nid
];
1640 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1641 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1653 * Unregister hstate attributes from a single node sysdev.
1654 * No-op if no hstate attributes attached.
1656 void hugetlb_unregister_node(struct node
*node
)
1659 struct node_hstate
*nhs
= &node_hstates
[node
->sysdev
.id
];
1661 if (!nhs
->hugepages_kobj
)
1662 return; /* no hstate attributes */
1665 if (nhs
->hstate_kobjs
[h
- hstates
]) {
1666 kobject_put(nhs
->hstate_kobjs
[h
- hstates
]);
1667 nhs
->hstate_kobjs
[h
- hstates
] = NULL
;
1670 kobject_put(nhs
->hugepages_kobj
);
1671 nhs
->hugepages_kobj
= NULL
;
1675 * hugetlb module exit: unregister hstate attributes from node sysdevs
1678 static void hugetlb_unregister_all_nodes(void)
1683 * disable node sysdev registrations.
1685 register_hugetlbfs_with_node(NULL
, NULL
);
1688 * remove hstate attributes from any nodes that have them.
1690 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1691 hugetlb_unregister_node(&node_devices
[nid
]);
1695 * Register hstate attributes for a single node sysdev.
1696 * No-op if attributes already registered.
1698 void hugetlb_register_node(struct node
*node
)
1701 struct node_hstate
*nhs
= &node_hstates
[node
->sysdev
.id
];
1704 if (nhs
->hugepages_kobj
)
1705 return; /* already allocated */
1707 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1708 &node
->sysdev
.kobj
);
1709 if (!nhs
->hugepages_kobj
)
1712 for_each_hstate(h
) {
1713 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1715 &per_node_hstate_attr_group
);
1717 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s"
1719 h
->name
, node
->sysdev
.id
);
1720 hugetlb_unregister_node(node
);
1727 * hugetlb init time: register hstate attributes for all registered node
1728 * sysdevs of nodes that have memory. All on-line nodes should have
1729 * registered their associated sysdev by this time.
1731 static void hugetlb_register_all_nodes(void)
1735 for_each_node_state(nid
, N_HIGH_MEMORY
) {
1736 struct node
*node
= &node_devices
[nid
];
1737 if (node
->sysdev
.id
== nid
)
1738 hugetlb_register_node(node
);
1742 * Let the node sysdev driver know we're here so it can
1743 * [un]register hstate attributes on node hotplug.
1745 register_hugetlbfs_with_node(hugetlb_register_node
,
1746 hugetlb_unregister_node
);
1748 #else /* !CONFIG_NUMA */
1750 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1758 static void hugetlb_unregister_all_nodes(void) { }
1760 static void hugetlb_register_all_nodes(void) { }
1764 static void __exit
hugetlb_exit(void)
1768 hugetlb_unregister_all_nodes();
1770 for_each_hstate(h
) {
1771 kobject_put(hstate_kobjs
[h
- hstates
]);
1774 kobject_put(hugepages_kobj
);
1776 module_exit(hugetlb_exit
);
1778 static int __init
hugetlb_init(void)
1780 /* Some platform decide whether they support huge pages at boot
1781 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1782 * there is no such support
1784 if (HPAGE_SHIFT
== 0)
1787 if (!size_to_hstate(default_hstate_size
)) {
1788 default_hstate_size
= HPAGE_SIZE
;
1789 if (!size_to_hstate(default_hstate_size
))
1790 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1792 default_hstate_idx
= size_to_hstate(default_hstate_size
) - hstates
;
1793 if (default_hstate_max_huge_pages
)
1794 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1796 hugetlb_init_hstates();
1798 gather_bootmem_prealloc();
1802 hugetlb_sysfs_init();
1804 hugetlb_register_all_nodes();
1808 module_init(hugetlb_init
);
1810 /* Should be called on processing a hugepagesz=... option */
1811 void __init
hugetlb_add_hstate(unsigned order
)
1816 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1817 printk(KERN_WARNING
"hugepagesz= specified twice, ignoring\n");
1820 BUG_ON(max_hstate
>= HUGE_MAX_HSTATE
);
1822 h
= &hstates
[max_hstate
++];
1824 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1825 h
->nr_huge_pages
= 0;
1826 h
->free_huge_pages
= 0;
1827 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1828 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1829 h
->next_nid_to_alloc
= first_node(node_states
[N_HIGH_MEMORY
]);
1830 h
->next_nid_to_free
= first_node(node_states
[N_HIGH_MEMORY
]);
1831 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1832 huge_page_size(h
)/1024);
1837 static int __init
hugetlb_nrpages_setup(char *s
)
1840 static unsigned long *last_mhp
;
1843 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1844 * so this hugepages= parameter goes to the "default hstate".
1847 mhp
= &default_hstate_max_huge_pages
;
1849 mhp
= &parsed_hstate
->max_huge_pages
;
1851 if (mhp
== last_mhp
) {
1852 printk(KERN_WARNING
"hugepages= specified twice without "
1853 "interleaving hugepagesz=, ignoring\n");
1857 if (sscanf(s
, "%lu", mhp
) <= 0)
1861 * Global state is always initialized later in hugetlb_init.
1862 * But we need to allocate >= MAX_ORDER hstates here early to still
1863 * use the bootmem allocator.
1865 if (max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1866 hugetlb_hstate_alloc_pages(parsed_hstate
);
1872 __setup("hugepages=", hugetlb_nrpages_setup
);
1874 static int __init
hugetlb_default_setup(char *s
)
1876 default_hstate_size
= memparse(s
, &s
);
1879 __setup("default_hugepagesz=", hugetlb_default_setup
);
1881 static unsigned int cpuset_mems_nr(unsigned int *array
)
1884 unsigned int nr
= 0;
1886 for_each_node_mask(node
, cpuset_current_mems_allowed
)
1892 #ifdef CONFIG_SYSCTL
1893 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
1894 struct ctl_table
*table
, int write
,
1895 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1897 struct hstate
*h
= &default_hstate
;
1901 tmp
= h
->max_huge_pages
;
1904 table
->maxlen
= sizeof(unsigned long);
1905 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
1908 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
1909 GFP_KERNEL
| __GFP_NORETRY
);
1910 if (!(obey_mempolicy
&&
1911 init_nodemask_of_mempolicy(nodes_allowed
))) {
1912 NODEMASK_FREE(nodes_allowed
);
1913 nodes_allowed
= &node_states
[N_HIGH_MEMORY
];
1915 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
1917 if (nodes_allowed
!= &node_states
[N_HIGH_MEMORY
])
1918 NODEMASK_FREE(nodes_allowed
);
1924 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
1925 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1928 return hugetlb_sysctl_handler_common(false, table
, write
,
1929 buffer
, length
, ppos
);
1933 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
1934 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1936 return hugetlb_sysctl_handler_common(true, table
, write
,
1937 buffer
, length
, ppos
);
1939 #endif /* CONFIG_NUMA */
1941 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
1942 void __user
*buffer
,
1943 size_t *length
, loff_t
*ppos
)
1945 proc_dointvec(table
, write
, buffer
, length
, ppos
);
1946 if (hugepages_treat_as_movable
)
1947 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
1949 htlb_alloc_mask
= GFP_HIGHUSER
;
1953 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
1954 void __user
*buffer
,
1955 size_t *length
, loff_t
*ppos
)
1957 struct hstate
*h
= &default_hstate
;
1961 tmp
= h
->nr_overcommit_huge_pages
;
1964 table
->maxlen
= sizeof(unsigned long);
1965 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
1968 spin_lock(&hugetlb_lock
);
1969 h
->nr_overcommit_huge_pages
= tmp
;
1970 spin_unlock(&hugetlb_lock
);
1976 #endif /* CONFIG_SYSCTL */
1978 void hugetlb_report_meminfo(struct seq_file
*m
)
1980 struct hstate
*h
= &default_hstate
;
1982 "HugePages_Total: %5lu\n"
1983 "HugePages_Free: %5lu\n"
1984 "HugePages_Rsvd: %5lu\n"
1985 "HugePages_Surp: %5lu\n"
1986 "Hugepagesize: %8lu kB\n",
1990 h
->surplus_huge_pages
,
1991 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
1994 int hugetlb_report_node_meminfo(int nid
, char *buf
)
1996 struct hstate
*h
= &default_hstate
;
1998 "Node %d HugePages_Total: %5u\n"
1999 "Node %d HugePages_Free: %5u\n"
2000 "Node %d HugePages_Surp: %5u\n",
2001 nid
, h
->nr_huge_pages_node
[nid
],
2002 nid
, h
->free_huge_pages_node
[nid
],
2003 nid
, h
->surplus_huge_pages_node
[nid
]);
2006 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2007 unsigned long hugetlb_total_pages(void)
2009 struct hstate
*h
= &default_hstate
;
2010 return h
->nr_huge_pages
* pages_per_huge_page(h
);
2013 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2017 spin_lock(&hugetlb_lock
);
2019 * When cpuset is configured, it breaks the strict hugetlb page
2020 * reservation as the accounting is done on a global variable. Such
2021 * reservation is completely rubbish in the presence of cpuset because
2022 * the reservation is not checked against page availability for the
2023 * current cpuset. Application can still potentially OOM'ed by kernel
2024 * with lack of free htlb page in cpuset that the task is in.
2025 * Attempt to enforce strict accounting with cpuset is almost
2026 * impossible (or too ugly) because cpuset is too fluid that
2027 * task or memory node can be dynamically moved between cpusets.
2029 * The change of semantics for shared hugetlb mapping with cpuset is
2030 * undesirable. However, in order to preserve some of the semantics,
2031 * we fall back to check against current free page availability as
2032 * a best attempt and hopefully to minimize the impact of changing
2033 * semantics that cpuset has.
2036 if (gather_surplus_pages(h
, delta
) < 0)
2039 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2040 return_unused_surplus_pages(h
, delta
);
2047 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2050 spin_unlock(&hugetlb_lock
);
2054 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2056 struct resv_map
*reservations
= vma_resv_map(vma
);
2059 * This new VMA should share its siblings reservation map if present.
2060 * The VMA will only ever have a valid reservation map pointer where
2061 * it is being copied for another still existing VMA. As that VMA
2062 * has a reference to the reservation map it cannot dissappear until
2063 * after this open call completes. It is therefore safe to take a
2064 * new reference here without additional locking.
2067 kref_get(&reservations
->refs
);
2070 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2072 struct hstate
*h
= hstate_vma(vma
);
2073 struct resv_map
*reservations
= vma_resv_map(vma
);
2074 unsigned long reserve
;
2075 unsigned long start
;
2079 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2080 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2082 reserve
= (end
- start
) -
2083 region_count(&reservations
->regions
, start
, end
);
2085 kref_put(&reservations
->refs
, resv_map_release
);
2088 hugetlb_acct_memory(h
, -reserve
);
2089 hugetlb_put_quota(vma
->vm_file
->f_mapping
, reserve
);
2095 * We cannot handle pagefaults against hugetlb pages at all. They cause
2096 * handle_mm_fault() to try to instantiate regular-sized pages in the
2097 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2100 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2106 const struct vm_operations_struct hugetlb_vm_ops
= {
2107 .fault
= hugetlb_vm_op_fault
,
2108 .open
= hugetlb_vm_op_open
,
2109 .close
= hugetlb_vm_op_close
,
2112 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2119 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
2121 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
2123 entry
= pte_mkyoung(entry
);
2124 entry
= pte_mkhuge(entry
);
2129 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2130 unsigned long address
, pte_t
*ptep
)
2134 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
2135 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1)) {
2136 update_mmu_cache(vma
, address
, ptep
);
2141 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2142 struct vm_area_struct
*vma
)
2144 pte_t
*src_pte
, *dst_pte
, entry
;
2145 struct page
*ptepage
;
2148 struct hstate
*h
= hstate_vma(vma
);
2149 unsigned long sz
= huge_page_size(h
);
2151 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2153 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2154 src_pte
= huge_pte_offset(src
, addr
);
2157 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2161 /* If the pagetables are shared don't copy or take references */
2162 if (dst_pte
== src_pte
)
2165 spin_lock(&dst
->page_table_lock
);
2166 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2167 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2169 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2170 entry
= huge_ptep_get(src_pte
);
2171 ptepage
= pte_page(entry
);
2173 page_dup_rmap(ptepage
);
2174 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2176 spin_unlock(&src
->page_table_lock
);
2177 spin_unlock(&dst
->page_table_lock
);
2185 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2189 if (huge_pte_none(pte
) || pte_present(pte
))
2191 swp
= pte_to_swp_entry(pte
);
2192 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
)) {
2198 void __unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2199 unsigned long end
, struct page
*ref_page
)
2201 struct mm_struct
*mm
= vma
->vm_mm
;
2202 unsigned long address
;
2207 struct hstate
*h
= hstate_vma(vma
);
2208 unsigned long sz
= huge_page_size(h
);
2211 * A page gathering list, protected by per file i_mmap_lock. The
2212 * lock is used to avoid list corruption from multiple unmapping
2213 * of the same page since we are using page->lru.
2215 LIST_HEAD(page_list
);
2217 WARN_ON(!is_vm_hugetlb_page(vma
));
2218 BUG_ON(start
& ~huge_page_mask(h
));
2219 BUG_ON(end
& ~huge_page_mask(h
));
2221 mmu_notifier_invalidate_range_start(mm
, start
, end
);
2222 spin_lock(&mm
->page_table_lock
);
2223 for (address
= start
; address
< end
; address
+= sz
) {
2224 ptep
= huge_pte_offset(mm
, address
);
2228 if (huge_pmd_unshare(mm
, &address
, ptep
))
2232 * If a reference page is supplied, it is because a specific
2233 * page is being unmapped, not a range. Ensure the page we
2234 * are about to unmap is the actual page of interest.
2237 pte
= huge_ptep_get(ptep
);
2238 if (huge_pte_none(pte
))
2240 page
= pte_page(pte
);
2241 if (page
!= ref_page
)
2245 * Mark the VMA as having unmapped its page so that
2246 * future faults in this VMA will fail rather than
2247 * looking like data was lost
2249 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2252 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2253 if (huge_pte_none(pte
))
2257 * HWPoisoned hugepage is already unmapped and dropped reference
2259 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
)))
2262 page
= pte_page(pte
);
2264 set_page_dirty(page
);
2265 list_add(&page
->lru
, &page_list
);
2267 spin_unlock(&mm
->page_table_lock
);
2268 flush_tlb_range(vma
, start
, end
);
2269 mmu_notifier_invalidate_range_end(mm
, start
, end
);
2270 list_for_each_entry_safe(page
, tmp
, &page_list
, lru
) {
2271 page_remove_rmap(page
);
2272 list_del(&page
->lru
);
2277 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2278 unsigned long end
, struct page
*ref_page
)
2280 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2281 __unmap_hugepage_range(vma
, start
, end
, ref_page
);
2282 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2286 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2287 * mappping it owns the reserve page for. The intention is to unmap the page
2288 * from other VMAs and let the children be SIGKILLed if they are faulting the
2291 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2292 struct page
*page
, unsigned long address
)
2294 struct hstate
*h
= hstate_vma(vma
);
2295 struct vm_area_struct
*iter_vma
;
2296 struct address_space
*mapping
;
2297 struct prio_tree_iter iter
;
2301 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2302 * from page cache lookup which is in HPAGE_SIZE units.
2304 address
= address
& huge_page_mask(h
);
2305 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
)
2306 + (vma
->vm_pgoff
>> PAGE_SHIFT
);
2307 mapping
= (struct address_space
*)page_private(page
);
2310 * Take the mapping lock for the duration of the table walk. As
2311 * this mapping should be shared between all the VMAs,
2312 * __unmap_hugepage_range() is called as the lock is already held
2314 spin_lock(&mapping
->i_mmap_lock
);
2315 vma_prio_tree_foreach(iter_vma
, &iter
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2316 /* Do not unmap the current VMA */
2317 if (iter_vma
== vma
)
2321 * Unmap the page from other VMAs without their own reserves.
2322 * They get marked to be SIGKILLed if they fault in these
2323 * areas. This is because a future no-page fault on this VMA
2324 * could insert a zeroed page instead of the data existing
2325 * from the time of fork. This would look like data corruption
2327 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2328 __unmap_hugepage_range(iter_vma
,
2329 address
, address
+ huge_page_size(h
),
2332 spin_unlock(&mapping
->i_mmap_lock
);
2338 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2340 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2341 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2342 struct page
*pagecache_page
)
2344 struct hstate
*h
= hstate_vma(vma
);
2345 struct page
*old_page
, *new_page
;
2347 int outside_reserve
= 0;
2349 old_page
= pte_page(pte
);
2352 /* If no-one else is actually using this page, avoid the copy
2353 * and just make the page writable */
2354 avoidcopy
= (page_mapcount(old_page
) == 1);
2356 if (PageAnon(old_page
))
2357 page_move_anon_rmap(old_page
, vma
, address
);
2358 set_huge_ptep_writable(vma
, address
, ptep
);
2363 * If the process that created a MAP_PRIVATE mapping is about to
2364 * perform a COW due to a shared page count, attempt to satisfy
2365 * the allocation without using the existing reserves. The pagecache
2366 * page is used to determine if the reserve at this address was
2367 * consumed or not. If reserves were used, a partial faulted mapping
2368 * at the time of fork() could consume its reserves on COW instead
2369 * of the full address range.
2371 if (!(vma
->vm_flags
& VM_MAYSHARE
) &&
2372 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2373 old_page
!= pagecache_page
)
2374 outside_reserve
= 1;
2376 page_cache_get(old_page
);
2378 /* Drop page_table_lock as buddy allocator may be called */
2379 spin_unlock(&mm
->page_table_lock
);
2380 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2382 if (IS_ERR(new_page
)) {
2383 page_cache_release(old_page
);
2386 * If a process owning a MAP_PRIVATE mapping fails to COW,
2387 * it is due to references held by a child and an insufficient
2388 * huge page pool. To guarantee the original mappers
2389 * reliability, unmap the page from child processes. The child
2390 * may get SIGKILLed if it later faults.
2392 if (outside_reserve
) {
2393 BUG_ON(huge_pte_none(pte
));
2394 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2395 BUG_ON(page_count(old_page
) != 1);
2396 BUG_ON(huge_pte_none(pte
));
2397 spin_lock(&mm
->page_table_lock
);
2398 goto retry_avoidcopy
;
2403 /* Caller expects lock to be held */
2404 spin_lock(&mm
->page_table_lock
);
2405 return -PTR_ERR(new_page
);
2409 * When the original hugepage is shared one, it does not have
2410 * anon_vma prepared.
2412 if (unlikely(anon_vma_prepare(vma
)))
2413 return VM_FAULT_OOM
;
2415 copy_huge_page(new_page
, old_page
, address
, vma
);
2416 __SetPageUptodate(new_page
);
2419 * Retake the page_table_lock to check for racing updates
2420 * before the page tables are altered
2422 spin_lock(&mm
->page_table_lock
);
2423 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2424 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2426 mmu_notifier_invalidate_range_start(mm
,
2427 address
& huge_page_mask(h
),
2428 (address
& huge_page_mask(h
)) + huge_page_size(h
));
2429 huge_ptep_clear_flush(vma
, address
, ptep
);
2430 set_huge_pte_at(mm
, address
, ptep
,
2431 make_huge_pte(vma
, new_page
, 1));
2432 page_remove_rmap(old_page
);
2433 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2434 /* Make the old page be freed below */
2435 new_page
= old_page
;
2436 mmu_notifier_invalidate_range_end(mm
,
2437 address
& huge_page_mask(h
),
2438 (address
& huge_page_mask(h
)) + huge_page_size(h
));
2440 page_cache_release(new_page
);
2441 page_cache_release(old_page
);
2445 /* Return the pagecache page at a given address within a VMA */
2446 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2447 struct vm_area_struct
*vma
, unsigned long address
)
2449 struct address_space
*mapping
;
2452 mapping
= vma
->vm_file
->f_mapping
;
2453 idx
= vma_hugecache_offset(h
, vma
, address
);
2455 return find_lock_page(mapping
, idx
);
2459 * Return whether there is a pagecache page to back given address within VMA.
2460 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2462 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2463 struct vm_area_struct
*vma
, unsigned long address
)
2465 struct address_space
*mapping
;
2469 mapping
= vma
->vm_file
->f_mapping
;
2470 idx
= vma_hugecache_offset(h
, vma
, address
);
2472 page
= find_get_page(mapping
, idx
);
2475 return page
!= NULL
;
2478 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2479 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2481 struct hstate
*h
= hstate_vma(vma
);
2482 int ret
= VM_FAULT_SIGBUS
;
2486 struct address_space
*mapping
;
2490 * Currently, we are forced to kill the process in the event the
2491 * original mapper has unmapped pages from the child due to a failed
2492 * COW. Warn that such a situation has occured as it may not be obvious
2494 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2496 "PID %d killed due to inadequate hugepage pool\n",
2501 mapping
= vma
->vm_file
->f_mapping
;
2502 idx
= vma_hugecache_offset(h
, vma
, address
);
2505 * Use page lock to guard against racing truncation
2506 * before we get page_table_lock.
2509 page
= find_lock_page(mapping
, idx
);
2511 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2514 page
= alloc_huge_page(vma
, address
, 0);
2516 ret
= -PTR_ERR(page
);
2519 clear_huge_page(page
, address
, huge_page_size(h
));
2520 __SetPageUptodate(page
);
2522 if (vma
->vm_flags
& VM_MAYSHARE
) {
2524 struct inode
*inode
= mapping
->host
;
2526 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2534 spin_lock(&inode
->i_lock
);
2535 inode
->i_blocks
+= blocks_per_huge_page(h
);
2536 spin_unlock(&inode
->i_lock
);
2537 page_dup_rmap(page
);
2540 if (unlikely(anon_vma_prepare(vma
))) {
2542 goto backout_unlocked
;
2544 hugepage_add_new_anon_rmap(page
, vma
, address
);
2548 * If memory error occurs between mmap() and fault, some process
2549 * don't have hwpoisoned swap entry for errored virtual address.
2550 * So we need to block hugepage fault by PG_hwpoison bit check.
2552 if (unlikely(PageHWPoison(page
))) {
2553 ret
= VM_FAULT_HWPOISON
;
2554 goto backout_unlocked
;
2556 page_dup_rmap(page
);
2560 * If we are going to COW a private mapping later, we examine the
2561 * pending reservations for this page now. This will ensure that
2562 * any allocations necessary to record that reservation occur outside
2565 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2566 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2568 goto backout_unlocked
;
2571 spin_lock(&mm
->page_table_lock
);
2572 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2577 if (!huge_pte_none(huge_ptep_get(ptep
)))
2580 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2581 && (vma
->vm_flags
& VM_SHARED
)));
2582 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2584 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2585 /* Optimization, do the COW without a second fault */
2586 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2589 spin_unlock(&mm
->page_table_lock
);
2595 spin_unlock(&mm
->page_table_lock
);
2602 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2603 unsigned long address
, unsigned int flags
)
2608 struct page
*page
= NULL
;
2609 struct page
*pagecache_page
= NULL
;
2610 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2611 struct hstate
*h
= hstate_vma(vma
);
2613 ptep
= huge_pte_offset(mm
, address
);
2615 entry
= huge_ptep_get(ptep
);
2616 if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2617 return VM_FAULT_HWPOISON
;
2620 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2622 return VM_FAULT_OOM
;
2625 * Serialize hugepage allocation and instantiation, so that we don't
2626 * get spurious allocation failures if two CPUs race to instantiate
2627 * the same page in the page cache.
2629 mutex_lock(&hugetlb_instantiation_mutex
);
2630 entry
= huge_ptep_get(ptep
);
2631 if (huge_pte_none(entry
)) {
2632 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2639 * If we are going to COW the mapping later, we examine the pending
2640 * reservations for this page now. This will ensure that any
2641 * allocations necessary to record that reservation occur outside the
2642 * spinlock. For private mappings, we also lookup the pagecache
2643 * page now as it is used to determine if a reservation has been
2646 if ((flags
& FAULT_FLAG_WRITE
) && !pte_write(entry
)) {
2647 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2652 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2653 pagecache_page
= hugetlbfs_pagecache_page(h
,
2658 * hugetlb_cow() requires page locks of pte_page(entry) and
2659 * pagecache_page, so here we need take the former one
2660 * when page != pagecache_page or !pagecache_page.
2661 * Note that locking order is always pagecache_page -> page,
2662 * so no worry about deadlock.
2664 page
= pte_page(entry
);
2665 if (page
!= pagecache_page
)
2668 spin_lock(&mm
->page_table_lock
);
2669 /* Check for a racing update before calling hugetlb_cow */
2670 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2671 goto out_page_table_lock
;
2674 if (flags
& FAULT_FLAG_WRITE
) {
2675 if (!pte_write(entry
)) {
2676 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2678 goto out_page_table_lock
;
2680 entry
= pte_mkdirty(entry
);
2682 entry
= pte_mkyoung(entry
);
2683 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2684 flags
& FAULT_FLAG_WRITE
))
2685 update_mmu_cache(vma
, address
, ptep
);
2687 out_page_table_lock
:
2688 spin_unlock(&mm
->page_table_lock
);
2690 if (pagecache_page
) {
2691 unlock_page(pagecache_page
);
2692 put_page(pagecache_page
);
2697 mutex_unlock(&hugetlb_instantiation_mutex
);
2702 /* Can be overriden by architectures */
2703 __attribute__((weak
)) struct page
*
2704 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
2705 pud_t
*pud
, int write
)
2711 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2712 struct page
**pages
, struct vm_area_struct
**vmas
,
2713 unsigned long *position
, int *length
, int i
,
2716 unsigned long pfn_offset
;
2717 unsigned long vaddr
= *position
;
2718 int remainder
= *length
;
2719 struct hstate
*h
= hstate_vma(vma
);
2721 spin_lock(&mm
->page_table_lock
);
2722 while (vaddr
< vma
->vm_end
&& remainder
) {
2728 * Some archs (sparc64, sh*) have multiple pte_ts to
2729 * each hugepage. We have to make sure we get the
2730 * first, for the page indexing below to work.
2732 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2733 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
2736 * When coredumping, it suits get_dump_page if we just return
2737 * an error where there's an empty slot with no huge pagecache
2738 * to back it. This way, we avoid allocating a hugepage, and
2739 * the sparse dumpfile avoids allocating disk blocks, but its
2740 * huge holes still show up with zeroes where they need to be.
2742 if (absent
&& (flags
& FOLL_DUMP
) &&
2743 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
2749 ((flags
& FOLL_WRITE
) && !pte_write(huge_ptep_get(pte
)))) {
2752 spin_unlock(&mm
->page_table_lock
);
2753 ret
= hugetlb_fault(mm
, vma
, vaddr
,
2754 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
2755 spin_lock(&mm
->page_table_lock
);
2756 if (!(ret
& VM_FAULT_ERROR
))
2763 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2764 page
= pte_page(huge_ptep_get(pte
));
2767 pages
[i
] = mem_map_offset(page
, pfn_offset
);
2778 if (vaddr
< vma
->vm_end
&& remainder
&&
2779 pfn_offset
< pages_per_huge_page(h
)) {
2781 * We use pfn_offset to avoid touching the pageframes
2782 * of this compound page.
2787 spin_unlock(&mm
->page_table_lock
);
2788 *length
= remainder
;
2791 return i
? i
: -EFAULT
;
2794 void hugetlb_change_protection(struct vm_area_struct
*vma
,
2795 unsigned long address
, unsigned long end
, pgprot_t newprot
)
2797 struct mm_struct
*mm
= vma
->vm_mm
;
2798 unsigned long start
= address
;
2801 struct hstate
*h
= hstate_vma(vma
);
2803 BUG_ON(address
>= end
);
2804 flush_cache_range(vma
, address
, end
);
2806 spin_lock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2807 spin_lock(&mm
->page_table_lock
);
2808 for (; address
< end
; address
+= huge_page_size(h
)) {
2809 ptep
= huge_pte_offset(mm
, address
);
2812 if (huge_pmd_unshare(mm
, &address
, ptep
))
2814 if (!huge_pte_none(huge_ptep_get(ptep
))) {
2815 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2816 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
2817 set_huge_pte_at(mm
, address
, ptep
, pte
);
2820 spin_unlock(&mm
->page_table_lock
);
2821 spin_unlock(&vma
->vm_file
->f_mapping
->i_mmap_lock
);
2823 flush_tlb_range(vma
, start
, end
);
2826 int hugetlb_reserve_pages(struct inode
*inode
,
2828 struct vm_area_struct
*vma
,
2832 struct hstate
*h
= hstate_inode(inode
);
2835 * Only apply hugepage reservation if asked. At fault time, an
2836 * attempt will be made for VM_NORESERVE to allocate a page
2837 * and filesystem quota without using reserves
2839 if (acctflag
& VM_NORESERVE
)
2843 * Shared mappings base their reservation on the number of pages that
2844 * are already allocated on behalf of the file. Private mappings need
2845 * to reserve the full area even if read-only as mprotect() may be
2846 * called to make the mapping read-write. Assume !vma is a shm mapping
2848 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
2849 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
2851 struct resv_map
*resv_map
= resv_map_alloc();
2857 set_vma_resv_map(vma
, resv_map
);
2858 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
2864 /* There must be enough filesystem quota for the mapping */
2865 if (hugetlb_get_quota(inode
->i_mapping
, chg
))
2869 * Check enough hugepages are available for the reservation.
2870 * Hand back the quota if there are not
2872 ret
= hugetlb_acct_memory(h
, chg
);
2874 hugetlb_put_quota(inode
->i_mapping
, chg
);
2879 * Account for the reservations made. Shared mappings record regions
2880 * that have reservations as they are shared by multiple VMAs.
2881 * When the last VMA disappears, the region map says how much
2882 * the reservation was and the page cache tells how much of
2883 * the reservation was consumed. Private mappings are per-VMA and
2884 * only the consumed reservations are tracked. When the VMA
2885 * disappears, the original reservation is the VMA size and the
2886 * consumed reservations are stored in the map. Hence, nothing
2887 * else has to be done for private mappings here
2889 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
2890 region_add(&inode
->i_mapping
->private_list
, from
, to
);
2894 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
2896 struct hstate
*h
= hstate_inode(inode
);
2897 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
2899 spin_lock(&inode
->i_lock
);
2900 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
2901 spin_unlock(&inode
->i_lock
);
2903 hugetlb_put_quota(inode
->i_mapping
, (chg
- freed
));
2904 hugetlb_acct_memory(h
, -(chg
- freed
));
2908 * This function is called from memory failure code.
2909 * Assume the caller holds page lock of the head page.
2911 void __isolate_hwpoisoned_huge_page(struct page
*hpage
)
2913 struct hstate
*h
= page_hstate(hpage
);
2914 int nid
= page_to_nid(hpage
);
2916 spin_lock(&hugetlb_lock
);
2917 list_del(&hpage
->lru
);
2918 h
->free_huge_pages
--;
2919 h
->free_huge_pages_node
[nid
]--;
2920 spin_unlock(&hugetlb_lock
);