2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, 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>
24 #include <linux/page-isolation.h>
27 #include <asm/pgtable.h>
31 #include <linux/hugetlb.h>
32 #include <linux/hugetlb_cgroup.h>
33 #include <linux/node.h>
36 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
37 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
38 unsigned long hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 __initdata
LIST_HEAD(huge_boot_pages
);
46 /* for command line parsing */
47 static struct hstate
* __initdata parsed_hstate
;
48 static unsigned long __initdata default_hstate_max_huge_pages
;
49 static unsigned long __initdata default_hstate_size
;
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 DEFINE_SPINLOCK(hugetlb_lock
);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
58 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
60 spin_unlock(&spool
->lock
);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
70 struct hugepage_subpool
*spool
;
72 spool
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
76 spin_lock_init(&spool
->lock
);
78 spool
->max_hpages
= nr_blocks
;
79 spool
->used_hpages
= 0;
84 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
86 spin_lock(&spool
->lock
);
87 BUG_ON(!spool
->count
);
89 unlock_or_release_subpool(spool
);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
100 spin_lock(&spool
->lock
);
101 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
102 spool
->used_hpages
+= delta
;
106 spin_unlock(&spool
->lock
);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
117 spin_lock(&spool
->lock
);
118 spool
->used_hpages
-= delta
;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool
);
124 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
126 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
129 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
131 return subpool_inode(file_inode(vma
->vm_file
));
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link
;
154 static long region_add(struct list_head
*head
, long f
, long t
)
156 struct file_region
*rg
, *nrg
, *trg
;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg
, head
, link
)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
170 if (&rg
->link
== head
)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head
*head
, long f
, long t
)
192 struct file_region
*rg
, *nrg
;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg
, head
, link
)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg
->link
== head
|| t
< rg
->from
) {
204 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
209 INIT_LIST_HEAD(&nrg
->link
);
210 list_add(&nrg
->link
, rg
->link
.prev
);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
222 if (&rg
->link
== head
)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg
-= rg
->to
- rg
->from
;
239 static long region_truncate(struct list_head
*head
, long end
)
241 struct file_region
*rg
, *trg
;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg
, head
, link
)
248 if (&rg
->link
== head
)
251 /* If we are in the middle of a region then adjust it. */
252 if (end
> rg
->from
) {
255 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
260 if (&rg
->link
== head
)
262 chg
+= rg
->to
- rg
->from
;
269 static long region_count(struct list_head
*head
, long f
, long t
)
271 struct file_region
*rg
;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg
, head
, link
) {
284 seg_from
= max(rg
->from
, f
);
285 seg_to
= min(rg
->to
, t
);
287 chg
+= seg_to
- seg_from
;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
298 struct vm_area_struct
*vma
, unsigned long address
)
300 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
301 (vma
->vm_pgoff
>> huge_page_order(h
));
304 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
305 unsigned long address
)
307 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
316 struct hstate
*hstate
;
318 if (!is_vm_hugetlb_page(vma
))
321 hstate
= hstate_vma(vma
);
323 return 1UL << (hstate
->order
+ PAGE_SHIFT
);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
336 return vma_kernel_pagesize(vma
);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
370 return (unsigned long)vma
->vm_private_data
;
373 static void set_vma_private_data(struct vm_area_struct
*vma
,
376 vma
->vm_private_data
= (void *)value
;
381 struct list_head regions
;
384 static struct resv_map
*resv_map_alloc(void)
386 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
390 kref_init(&resv_map
->refs
);
391 INIT_LIST_HEAD(&resv_map
->regions
);
396 static void resv_map_release(struct kref
*ref
)
398 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map
->regions
, 0);
405 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
408 if (!(vma
->vm_flags
& VM_MAYSHARE
))
409 return (struct resv_map
*)(get_vma_private_data(vma
) &
414 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
417 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
419 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
420 HPAGE_RESV_MASK
) | (unsigned long)map
);
423 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
426 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
428 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
431 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
435 return (get_vma_private_data(vma
) & flag
) != 0;
438 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
439 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
441 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
442 if (!(vma
->vm_flags
& VM_MAYSHARE
))
443 vma
->vm_private_data
= (void *)0;
446 /* Returns true if the VMA has associated reserve pages */
447 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
449 if (vma
->vm_flags
& VM_NORESERVE
) {
451 * This address is already reserved by other process(chg == 0),
452 * so, we should decrement reserved count. Without decrementing,
453 * reserve count remains after releasing inode, because this
454 * allocated page will go into page cache and is regarded as
455 * coming from reserved pool in releasing step. Currently, we
456 * don't have any other solution to deal with this situation
457 * properly, so add work-around here.
459 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
465 /* Shared mappings always use reserves */
466 if (vma
->vm_flags
& VM_MAYSHARE
)
470 * Only the process that called mmap() has reserves for
473 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
479 static void copy_gigantic_page(struct page
*dst
, struct page
*src
)
482 struct hstate
*h
= page_hstate(src
);
483 struct page
*dst_base
= dst
;
484 struct page
*src_base
= src
;
486 for (i
= 0; i
< pages_per_huge_page(h
); ) {
488 copy_highpage(dst
, src
);
491 dst
= mem_map_next(dst
, dst_base
, i
);
492 src
= mem_map_next(src
, src_base
, i
);
496 void copy_huge_page(struct page
*dst
, struct page
*src
)
499 struct hstate
*h
= page_hstate(src
);
501 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
502 copy_gigantic_page(dst
, src
);
507 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
509 copy_highpage(dst
+ i
, src
+ i
);
513 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
515 int nid
= page_to_nid(page
);
516 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
517 h
->free_huge_pages
++;
518 h
->free_huge_pages_node
[nid
]++;
521 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
525 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
526 if (!is_migrate_isolate_page(page
))
529 * if 'non-isolated free hugepage' not found on the list,
530 * the allocation fails.
532 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
534 list_move(&page
->lru
, &h
->hugepage_activelist
);
535 set_page_refcounted(page
);
536 h
->free_huge_pages
--;
537 h
->free_huge_pages_node
[nid
]--;
541 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
542 struct vm_area_struct
*vma
,
543 unsigned long address
, int avoid_reserve
,
546 struct page
*page
= NULL
;
547 struct mempolicy
*mpol
;
548 nodemask_t
*nodemask
;
549 struct zonelist
*zonelist
;
552 unsigned int cpuset_mems_cookie
;
555 cpuset_mems_cookie
= get_mems_allowed();
556 zonelist
= huge_zonelist(vma
, address
,
557 htlb_alloc_mask
, &mpol
, &nodemask
);
559 * A child process with MAP_PRIVATE mappings created by their parent
560 * have no page reserves. This check ensures that reservations are
561 * not "stolen". The child may still get SIGKILLed
563 if (!vma_has_reserves(vma
, chg
) &&
564 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
567 /* If reserves cannot be used, ensure enough pages are in the pool */
568 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
571 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
572 MAX_NR_ZONES
- 1, nodemask
) {
573 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
574 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
578 if (!vma_has_reserves(vma
, chg
))
581 SetPagePrivate(page
);
582 h
->resv_huge_pages
--;
589 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
598 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
602 VM_BUG_ON(h
->order
>= MAX_ORDER
);
605 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
606 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
607 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
608 1 << PG_referenced
| 1 << PG_dirty
|
609 1 << PG_active
| 1 << PG_reserved
|
610 1 << PG_private
| 1 << PG_writeback
);
612 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
613 set_compound_page_dtor(page
, NULL
);
614 set_page_refcounted(page
);
615 arch_release_hugepage(page
);
616 __free_pages(page
, huge_page_order(h
));
619 struct hstate
*size_to_hstate(unsigned long size
)
624 if (huge_page_size(h
) == size
)
630 static void free_huge_page(struct page
*page
)
633 * Can't pass hstate in here because it is called from the
634 * compound page destructor.
636 struct hstate
*h
= page_hstate(page
);
637 int nid
= page_to_nid(page
);
638 struct hugepage_subpool
*spool
=
639 (struct hugepage_subpool
*)page_private(page
);
640 bool restore_reserve
;
642 set_page_private(page
, 0);
643 page
->mapping
= NULL
;
644 BUG_ON(page_count(page
));
645 BUG_ON(page_mapcount(page
));
646 restore_reserve
= PagePrivate(page
);
648 spin_lock(&hugetlb_lock
);
649 hugetlb_cgroup_uncharge_page(hstate_index(h
),
650 pages_per_huge_page(h
), page
);
652 h
->resv_huge_pages
++;
654 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
655 /* remove the page from active list */
656 list_del(&page
->lru
);
657 update_and_free_page(h
, page
);
658 h
->surplus_huge_pages
--;
659 h
->surplus_huge_pages_node
[nid
]--;
661 arch_clear_hugepage_flags(page
);
662 enqueue_huge_page(h
, page
);
664 spin_unlock(&hugetlb_lock
);
665 hugepage_subpool_put_pages(spool
, 1);
668 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
670 INIT_LIST_HEAD(&page
->lru
);
671 set_compound_page_dtor(page
, free_huge_page
);
672 spin_lock(&hugetlb_lock
);
673 set_hugetlb_cgroup(page
, NULL
);
675 h
->nr_huge_pages_node
[nid
]++;
676 spin_unlock(&hugetlb_lock
);
677 put_page(page
); /* free it into the hugepage allocator */
680 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
683 int nr_pages
= 1 << order
;
684 struct page
*p
= page
+ 1;
686 /* we rely on prep_new_huge_page to set the destructor */
687 set_compound_order(page
, order
);
689 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
691 set_page_count(p
, 0);
692 p
->first_page
= page
;
697 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
698 * transparent huge pages. See the PageTransHuge() documentation for more
701 int PageHuge(struct page
*page
)
703 compound_page_dtor
*dtor
;
705 if (!PageCompound(page
))
708 page
= compound_head(page
);
709 dtor
= get_compound_page_dtor(page
);
711 return dtor
== free_huge_page
;
713 EXPORT_SYMBOL_GPL(PageHuge
);
716 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
717 * normal or transparent huge pages.
719 int PageHeadHuge(struct page
*page_head
)
721 compound_page_dtor
*dtor
;
723 if (!PageHead(page_head
))
726 dtor
= get_compound_page_dtor(page_head
);
728 return dtor
== free_huge_page
;
730 EXPORT_SYMBOL_GPL(PageHeadHuge
);
732 pgoff_t
__basepage_index(struct page
*page
)
734 struct page
*page_head
= compound_head(page
);
735 pgoff_t index
= page_index(page_head
);
736 unsigned long compound_idx
;
738 if (!PageHuge(page_head
))
739 return page_index(page
);
741 if (compound_order(page_head
) >= MAX_ORDER
)
742 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
744 compound_idx
= page
- page_head
;
746 return (index
<< compound_order(page_head
)) + compound_idx
;
749 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
753 if (h
->order
>= MAX_ORDER
)
756 page
= alloc_pages_exact_node(nid
,
757 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
758 __GFP_REPEAT
|__GFP_NOWARN
,
761 if (arch_prepare_hugepage(page
)) {
762 __free_pages(page
, huge_page_order(h
));
765 prep_new_huge_page(h
, page
, nid
);
772 * common helper functions for hstate_next_node_to_{alloc|free}.
773 * We may have allocated or freed a huge page based on a different
774 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
775 * be outside of *nodes_allowed. Ensure that we use an allowed
776 * node for alloc or free.
778 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
780 nid
= next_node(nid
, *nodes_allowed
);
781 if (nid
== MAX_NUMNODES
)
782 nid
= first_node(*nodes_allowed
);
783 VM_BUG_ON(nid
>= MAX_NUMNODES
);
788 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
790 if (!node_isset(nid
, *nodes_allowed
))
791 nid
= next_node_allowed(nid
, nodes_allowed
);
796 * returns the previously saved node ["this node"] from which to
797 * allocate a persistent huge page for the pool and advance the
798 * next node from which to allocate, handling wrap at end of node
801 static int hstate_next_node_to_alloc(struct hstate
*h
,
802 nodemask_t
*nodes_allowed
)
806 VM_BUG_ON(!nodes_allowed
);
808 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
809 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
815 * helper for free_pool_huge_page() - return the previously saved
816 * node ["this node"] from which to free a huge page. Advance the
817 * next node id whether or not we find a free huge page to free so
818 * that the next attempt to free addresses the next node.
820 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
824 VM_BUG_ON(!nodes_allowed
);
826 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
827 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
832 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
833 for (nr_nodes = nodes_weight(*mask); \
835 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
838 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
839 for (nr_nodes = nodes_weight(*mask); \
841 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
844 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
850 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
851 page
= alloc_fresh_huge_page_node(h
, node
);
859 count_vm_event(HTLB_BUDDY_PGALLOC
);
861 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
867 * Free huge page from pool from next node to free.
868 * Attempt to keep persistent huge pages more or less
869 * balanced over allowed nodes.
870 * Called with hugetlb_lock locked.
872 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
878 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
880 * If we're returning unused surplus pages, only examine
881 * nodes with surplus pages.
883 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
884 !list_empty(&h
->hugepage_freelists
[node
])) {
886 list_entry(h
->hugepage_freelists
[node
].next
,
888 list_del(&page
->lru
);
889 h
->free_huge_pages
--;
890 h
->free_huge_pages_node
[node
]--;
892 h
->surplus_huge_pages
--;
893 h
->surplus_huge_pages_node
[node
]--;
895 update_and_free_page(h
, page
);
904 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
909 if (h
->order
>= MAX_ORDER
)
913 * Assume we will successfully allocate the surplus page to
914 * prevent racing processes from causing the surplus to exceed
917 * This however introduces a different race, where a process B
918 * tries to grow the static hugepage pool while alloc_pages() is
919 * called by process A. B will only examine the per-node
920 * counters in determining if surplus huge pages can be
921 * converted to normal huge pages in adjust_pool_surplus(). A
922 * won't be able to increment the per-node counter, until the
923 * lock is dropped by B, but B doesn't drop hugetlb_lock until
924 * no more huge pages can be converted from surplus to normal
925 * state (and doesn't try to convert again). Thus, we have a
926 * case where a surplus huge page exists, the pool is grown, and
927 * the surplus huge page still exists after, even though it
928 * should just have been converted to a normal huge page. This
929 * does not leak memory, though, as the hugepage will be freed
930 * once it is out of use. It also does not allow the counters to
931 * go out of whack in adjust_pool_surplus() as we don't modify
932 * the node values until we've gotten the hugepage and only the
933 * per-node value is checked there.
935 spin_lock(&hugetlb_lock
);
936 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
937 spin_unlock(&hugetlb_lock
);
941 h
->surplus_huge_pages
++;
943 spin_unlock(&hugetlb_lock
);
945 if (nid
== NUMA_NO_NODE
)
946 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
947 __GFP_REPEAT
|__GFP_NOWARN
,
950 page
= alloc_pages_exact_node(nid
,
951 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
952 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
954 if (page
&& arch_prepare_hugepage(page
)) {
955 __free_pages(page
, huge_page_order(h
));
959 spin_lock(&hugetlb_lock
);
961 INIT_LIST_HEAD(&page
->lru
);
962 r_nid
= page_to_nid(page
);
963 set_compound_page_dtor(page
, free_huge_page
);
964 set_hugetlb_cgroup(page
, NULL
);
966 * We incremented the global counters already
968 h
->nr_huge_pages_node
[r_nid
]++;
969 h
->surplus_huge_pages_node
[r_nid
]++;
970 __count_vm_event(HTLB_BUDDY_PGALLOC
);
973 h
->surplus_huge_pages
--;
974 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
976 spin_unlock(&hugetlb_lock
);
982 * This allocation function is useful in the context where vma is irrelevant.
983 * E.g. soft-offlining uses this function because it only cares physical
984 * address of error page.
986 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
988 struct page
*page
= NULL
;
990 spin_lock(&hugetlb_lock
);
991 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
992 page
= dequeue_huge_page_node(h
, nid
);
993 spin_unlock(&hugetlb_lock
);
996 page
= alloc_buddy_huge_page(h
, nid
);
1002 * Increase the hugetlb pool such that it can accommodate a reservation
1005 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1007 struct list_head surplus_list
;
1008 struct page
*page
, *tmp
;
1010 int needed
, allocated
;
1011 bool alloc_ok
= true;
1013 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1015 h
->resv_huge_pages
+= delta
;
1020 INIT_LIST_HEAD(&surplus_list
);
1024 spin_unlock(&hugetlb_lock
);
1025 for (i
= 0; i
< needed
; i
++) {
1026 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1031 list_add(&page
->lru
, &surplus_list
);
1036 * After retaking hugetlb_lock, we need to recalculate 'needed'
1037 * because either resv_huge_pages or free_huge_pages may have changed.
1039 spin_lock(&hugetlb_lock
);
1040 needed
= (h
->resv_huge_pages
+ delta
) -
1041 (h
->free_huge_pages
+ allocated
);
1046 * We were not able to allocate enough pages to
1047 * satisfy the entire reservation so we free what
1048 * we've allocated so far.
1053 * The surplus_list now contains _at_least_ the number of extra pages
1054 * needed to accommodate the reservation. Add the appropriate number
1055 * of pages to the hugetlb pool and free the extras back to the buddy
1056 * allocator. Commit the entire reservation here to prevent another
1057 * process from stealing the pages as they are added to the pool but
1058 * before they are reserved.
1060 needed
+= allocated
;
1061 h
->resv_huge_pages
+= delta
;
1064 /* Free the needed pages to the hugetlb pool */
1065 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1069 * This page is now managed by the hugetlb allocator and has
1070 * no users -- drop the buddy allocator's reference.
1072 put_page_testzero(page
);
1073 VM_BUG_ON(page_count(page
));
1074 enqueue_huge_page(h
, page
);
1077 spin_unlock(&hugetlb_lock
);
1079 /* Free unnecessary surplus pages to the buddy allocator */
1080 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1082 spin_lock(&hugetlb_lock
);
1088 * This routine has two main purposes:
1089 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1090 * in unused_resv_pages. This corresponds to the prior adjustments made
1091 * to the associated reservation map.
1092 * 2) Free any unused surplus pages that may have been allocated to satisfy
1093 * the reservation. As many as unused_resv_pages may be freed.
1095 * Called with hugetlb_lock held. However, the lock could be dropped (and
1096 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1097 * we must make sure nobody else can claim pages we are in the process of
1098 * freeing. Do this by ensuring resv_huge_page always is greater than the
1099 * number of huge pages we plan to free when dropping the lock.
1101 static void return_unused_surplus_pages(struct hstate
*h
,
1102 unsigned long unused_resv_pages
)
1104 unsigned long nr_pages
;
1106 /* Cannot return gigantic pages currently */
1107 if (h
->order
>= MAX_ORDER
)
1111 * Part (or even all) of the reservation could have been backed
1112 * by pre-allocated pages. Only free surplus pages.
1114 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1117 * We want to release as many surplus pages as possible, spread
1118 * evenly across all nodes with memory. Iterate across these nodes
1119 * until we can no longer free unreserved surplus pages. This occurs
1120 * when the nodes with surplus pages have no free pages.
1121 * free_pool_huge_page() will balance the the freed pages across the
1122 * on-line nodes with memory and will handle the hstate accounting.
1124 * Note that we decrement resv_huge_pages as we free the pages. If
1125 * we drop the lock, resv_huge_pages will still be sufficiently large
1126 * to cover subsequent pages we may free.
1128 while (nr_pages
--) {
1129 h
->resv_huge_pages
--;
1130 unused_resv_pages
--;
1131 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1133 cond_resched_lock(&hugetlb_lock
);
1137 /* Fully uncommit the reservation */
1138 h
->resv_huge_pages
-= unused_resv_pages
;
1142 * Determine if the huge page at addr within the vma has an associated
1143 * reservation. Where it does not we will need to logically increase
1144 * reservation and actually increase subpool usage before an allocation
1145 * can occur. Where any new reservation would be required the
1146 * reservation change is prepared, but not committed. Once the page
1147 * has been allocated from the subpool and instantiated the change should
1148 * be committed via vma_commit_reservation. No action is required on
1151 static long vma_needs_reservation(struct hstate
*h
,
1152 struct vm_area_struct
*vma
, unsigned long addr
)
1154 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1155 struct inode
*inode
= mapping
->host
;
1157 if (vma
->vm_flags
& VM_MAYSHARE
) {
1158 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1159 return region_chg(&inode
->i_mapping
->private_list
,
1162 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1167 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1168 struct resv_map
*resv
= vma_resv_map(vma
);
1170 err
= region_chg(&resv
->regions
, idx
, idx
+ 1);
1176 static void vma_commit_reservation(struct hstate
*h
,
1177 struct vm_area_struct
*vma
, unsigned long addr
)
1179 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1180 struct inode
*inode
= mapping
->host
;
1182 if (vma
->vm_flags
& VM_MAYSHARE
) {
1183 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1184 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1186 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1187 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1188 struct resv_map
*resv
= vma_resv_map(vma
);
1190 /* Mark this page used in the map. */
1191 region_add(&resv
->regions
, idx
, idx
+ 1);
1195 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1196 unsigned long addr
, int avoid_reserve
)
1198 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1199 struct hstate
*h
= hstate_vma(vma
);
1203 struct hugetlb_cgroup
*h_cg
;
1205 idx
= hstate_index(h
);
1207 * Processes that did not create the mapping will have no
1208 * reserves and will not have accounted against subpool
1209 * limit. Check that the subpool limit can be made before
1210 * satisfying the allocation MAP_NORESERVE mappings may also
1211 * need pages and subpool limit allocated allocated if no reserve
1214 chg
= vma_needs_reservation(h
, vma
, addr
);
1216 return ERR_PTR(-ENOMEM
);
1217 if (chg
|| avoid_reserve
)
1218 if (hugepage_subpool_get_pages(spool
, 1))
1219 return ERR_PTR(-ENOSPC
);
1221 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1223 if (chg
|| avoid_reserve
)
1224 hugepage_subpool_put_pages(spool
, 1);
1225 return ERR_PTR(-ENOSPC
);
1227 spin_lock(&hugetlb_lock
);
1228 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1230 spin_unlock(&hugetlb_lock
);
1231 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1233 hugetlb_cgroup_uncharge_cgroup(idx
,
1234 pages_per_huge_page(h
),
1236 if (chg
|| avoid_reserve
)
1237 hugepage_subpool_put_pages(spool
, 1);
1238 return ERR_PTR(-ENOSPC
);
1240 spin_lock(&hugetlb_lock
);
1241 list_move(&page
->lru
, &h
->hugepage_activelist
);
1244 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1245 spin_unlock(&hugetlb_lock
);
1247 set_page_private(page
, (unsigned long)spool
);
1249 vma_commit_reservation(h
, vma
, addr
);
1253 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1255 struct huge_bootmem_page
*m
;
1258 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1261 addr
= __alloc_bootmem_node_nopanic(NODE_DATA(node
),
1262 huge_page_size(h
), huge_page_size(h
), 0);
1266 * Use the beginning of the huge page to store the
1267 * huge_bootmem_page struct (until gather_bootmem
1268 * puts them into the mem_map).
1277 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1278 /* Put them into a private list first because mem_map is not up yet */
1279 list_add(&m
->list
, &huge_boot_pages
);
1284 static void prep_compound_huge_page(struct page
*page
, int order
)
1286 if (unlikely(order
> (MAX_ORDER
- 1)))
1287 prep_compound_gigantic_page(page
, order
);
1289 prep_compound_page(page
, order
);
1292 /* Put bootmem huge pages into the standard lists after mem_map is up */
1293 static void __init
gather_bootmem_prealloc(void)
1295 struct huge_bootmem_page
*m
;
1297 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1298 struct hstate
*h
= m
->hstate
;
1301 #ifdef CONFIG_HIGHMEM
1302 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1303 free_bootmem_late((unsigned long)m
,
1304 sizeof(struct huge_bootmem_page
));
1306 page
= virt_to_page(m
);
1308 __ClearPageReserved(page
);
1309 WARN_ON(page_count(page
) != 1);
1310 prep_compound_huge_page(page
, h
->order
);
1311 prep_new_huge_page(h
, page
, page_to_nid(page
));
1313 * If we had gigantic hugepages allocated at boot time, we need
1314 * to restore the 'stolen' pages to totalram_pages in order to
1315 * fix confusing memory reports from free(1) and another
1316 * side-effects, like CommitLimit going negative.
1318 if (h
->order
> (MAX_ORDER
- 1))
1319 totalram_pages
+= 1 << h
->order
;
1323 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1327 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1328 if (h
->order
>= MAX_ORDER
) {
1329 if (!alloc_bootmem_huge_page(h
))
1331 } else if (!alloc_fresh_huge_page(h
,
1332 &node_states
[N_MEMORY
]))
1335 h
->max_huge_pages
= i
;
1338 static void __init
hugetlb_init_hstates(void)
1342 for_each_hstate(h
) {
1343 /* oversize hugepages were init'ed in early boot */
1344 if (h
->order
< MAX_ORDER
)
1345 hugetlb_hstate_alloc_pages(h
);
1349 static char * __init
memfmt(char *buf
, unsigned long n
)
1351 if (n
>= (1UL << 30))
1352 sprintf(buf
, "%lu GB", n
>> 30);
1353 else if (n
>= (1UL << 20))
1354 sprintf(buf
, "%lu MB", n
>> 20);
1356 sprintf(buf
, "%lu KB", n
>> 10);
1360 static void __init
report_hugepages(void)
1364 for_each_hstate(h
) {
1366 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1367 memfmt(buf
, huge_page_size(h
)),
1368 h
->free_huge_pages
);
1372 #ifdef CONFIG_HIGHMEM
1373 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1374 nodemask_t
*nodes_allowed
)
1378 if (h
->order
>= MAX_ORDER
)
1381 for_each_node_mask(i
, *nodes_allowed
) {
1382 struct page
*page
, *next
;
1383 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1384 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1385 if (count
>= h
->nr_huge_pages
)
1387 if (PageHighMem(page
))
1389 list_del(&page
->lru
);
1390 update_and_free_page(h
, page
);
1391 h
->free_huge_pages
--;
1392 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1397 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1398 nodemask_t
*nodes_allowed
)
1404 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1405 * balanced by operating on them in a round-robin fashion.
1406 * Returns 1 if an adjustment was made.
1408 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1413 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1416 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1417 if (h
->surplus_huge_pages_node
[node
])
1421 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1422 if (h
->surplus_huge_pages_node
[node
] <
1423 h
->nr_huge_pages_node
[node
])
1430 h
->surplus_huge_pages
+= delta
;
1431 h
->surplus_huge_pages_node
[node
] += delta
;
1435 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1436 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1437 nodemask_t
*nodes_allowed
)
1439 unsigned long min_count
, ret
;
1441 if (h
->order
>= MAX_ORDER
)
1442 return h
->max_huge_pages
;
1445 * Increase the pool size
1446 * First take pages out of surplus state. Then make up the
1447 * remaining difference by allocating fresh huge pages.
1449 * We might race with alloc_buddy_huge_page() here and be unable
1450 * to convert a surplus huge page to a normal huge page. That is
1451 * not critical, though, it just means the overall size of the
1452 * pool might be one hugepage larger than it needs to be, but
1453 * within all the constraints specified by the sysctls.
1455 spin_lock(&hugetlb_lock
);
1456 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1457 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1461 while (count
> persistent_huge_pages(h
)) {
1463 * If this allocation races such that we no longer need the
1464 * page, free_huge_page will handle it by freeing the page
1465 * and reducing the surplus.
1467 spin_unlock(&hugetlb_lock
);
1468 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1469 spin_lock(&hugetlb_lock
);
1473 /* Bail for signals. Probably ctrl-c from user */
1474 if (signal_pending(current
))
1479 * Decrease the pool size
1480 * First return free pages to the buddy allocator (being careful
1481 * to keep enough around to satisfy reservations). Then place
1482 * pages into surplus state as needed so the pool will shrink
1483 * to the desired size as pages become free.
1485 * By placing pages into the surplus state independent of the
1486 * overcommit value, we are allowing the surplus pool size to
1487 * exceed overcommit. There are few sane options here. Since
1488 * alloc_buddy_huge_page() is checking the global counter,
1489 * though, we'll note that we're not allowed to exceed surplus
1490 * and won't grow the pool anywhere else. Not until one of the
1491 * sysctls are changed, or the surplus pages go out of use.
1493 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1494 min_count
= max(count
, min_count
);
1495 try_to_free_low(h
, min_count
, nodes_allowed
);
1496 while (min_count
< persistent_huge_pages(h
)) {
1497 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1499 cond_resched_lock(&hugetlb_lock
);
1501 while (count
< persistent_huge_pages(h
)) {
1502 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1506 ret
= persistent_huge_pages(h
);
1507 spin_unlock(&hugetlb_lock
);
1511 #define HSTATE_ATTR_RO(_name) \
1512 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1514 #define HSTATE_ATTR(_name) \
1515 static struct kobj_attribute _name##_attr = \
1516 __ATTR(_name, 0644, _name##_show, _name##_store)
1518 static struct kobject
*hugepages_kobj
;
1519 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1521 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1523 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1527 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1528 if (hstate_kobjs
[i
] == kobj
) {
1530 *nidp
= NUMA_NO_NODE
;
1534 return kobj_to_node_hstate(kobj
, nidp
);
1537 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1538 struct kobj_attribute
*attr
, char *buf
)
1541 unsigned long nr_huge_pages
;
1544 h
= kobj_to_hstate(kobj
, &nid
);
1545 if (nid
== NUMA_NO_NODE
)
1546 nr_huge_pages
= h
->nr_huge_pages
;
1548 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1550 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1553 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1554 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1555 const char *buf
, size_t len
)
1559 unsigned long count
;
1561 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1563 err
= strict_strtoul(buf
, 10, &count
);
1567 h
= kobj_to_hstate(kobj
, &nid
);
1568 if (h
->order
>= MAX_ORDER
) {
1573 if (nid
== NUMA_NO_NODE
) {
1575 * global hstate attribute
1577 if (!(obey_mempolicy
&&
1578 init_nodemask_of_mempolicy(nodes_allowed
))) {
1579 NODEMASK_FREE(nodes_allowed
);
1580 nodes_allowed
= &node_states
[N_MEMORY
];
1582 } else if (nodes_allowed
) {
1584 * per node hstate attribute: adjust count to global,
1585 * but restrict alloc/free to the specified node.
1587 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1588 init_nodemask_of_node(nodes_allowed
, nid
);
1590 nodes_allowed
= &node_states
[N_MEMORY
];
1592 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1594 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1595 NODEMASK_FREE(nodes_allowed
);
1599 NODEMASK_FREE(nodes_allowed
);
1603 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1604 struct kobj_attribute
*attr
, char *buf
)
1606 return nr_hugepages_show_common(kobj
, attr
, buf
);
1609 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1610 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1612 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1614 HSTATE_ATTR(nr_hugepages
);
1619 * hstate attribute for optionally mempolicy-based constraint on persistent
1620 * huge page alloc/free.
1622 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1623 struct kobj_attribute
*attr
, char *buf
)
1625 return nr_hugepages_show_common(kobj
, attr
, buf
);
1628 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1629 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1631 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1633 HSTATE_ATTR(nr_hugepages_mempolicy
);
1637 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1638 struct kobj_attribute
*attr
, char *buf
)
1640 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1641 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1644 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1645 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1648 unsigned long input
;
1649 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1651 if (h
->order
>= MAX_ORDER
)
1654 err
= strict_strtoul(buf
, 10, &input
);
1658 spin_lock(&hugetlb_lock
);
1659 h
->nr_overcommit_huge_pages
= input
;
1660 spin_unlock(&hugetlb_lock
);
1664 HSTATE_ATTR(nr_overcommit_hugepages
);
1666 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1667 struct kobj_attribute
*attr
, char *buf
)
1670 unsigned long free_huge_pages
;
1673 h
= kobj_to_hstate(kobj
, &nid
);
1674 if (nid
== NUMA_NO_NODE
)
1675 free_huge_pages
= h
->free_huge_pages
;
1677 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1679 return sprintf(buf
, "%lu\n", free_huge_pages
);
1681 HSTATE_ATTR_RO(free_hugepages
);
1683 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1684 struct kobj_attribute
*attr
, char *buf
)
1686 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1687 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1689 HSTATE_ATTR_RO(resv_hugepages
);
1691 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1692 struct kobj_attribute
*attr
, char *buf
)
1695 unsigned long surplus_huge_pages
;
1698 h
= kobj_to_hstate(kobj
, &nid
);
1699 if (nid
== NUMA_NO_NODE
)
1700 surplus_huge_pages
= h
->surplus_huge_pages
;
1702 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1704 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1706 HSTATE_ATTR_RO(surplus_hugepages
);
1708 static struct attribute
*hstate_attrs
[] = {
1709 &nr_hugepages_attr
.attr
,
1710 &nr_overcommit_hugepages_attr
.attr
,
1711 &free_hugepages_attr
.attr
,
1712 &resv_hugepages_attr
.attr
,
1713 &surplus_hugepages_attr
.attr
,
1715 &nr_hugepages_mempolicy_attr
.attr
,
1720 static struct attribute_group hstate_attr_group
= {
1721 .attrs
= hstate_attrs
,
1724 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1725 struct kobject
**hstate_kobjs
,
1726 struct attribute_group
*hstate_attr_group
)
1729 int hi
= hstate_index(h
);
1731 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1732 if (!hstate_kobjs
[hi
])
1735 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1737 kobject_put(hstate_kobjs
[hi
]);
1742 static void __init
hugetlb_sysfs_init(void)
1747 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1748 if (!hugepages_kobj
)
1751 for_each_hstate(h
) {
1752 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1753 hstate_kobjs
, &hstate_attr_group
);
1755 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1762 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1763 * with node devices in node_devices[] using a parallel array. The array
1764 * index of a node device or _hstate == node id.
1765 * This is here to avoid any static dependency of the node device driver, in
1766 * the base kernel, on the hugetlb module.
1768 struct node_hstate
{
1769 struct kobject
*hugepages_kobj
;
1770 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1772 struct node_hstate node_hstates
[MAX_NUMNODES
];
1775 * A subset of global hstate attributes for node devices
1777 static struct attribute
*per_node_hstate_attrs
[] = {
1778 &nr_hugepages_attr
.attr
,
1779 &free_hugepages_attr
.attr
,
1780 &surplus_hugepages_attr
.attr
,
1784 static struct attribute_group per_node_hstate_attr_group
= {
1785 .attrs
= per_node_hstate_attrs
,
1789 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1790 * Returns node id via non-NULL nidp.
1792 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1796 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1797 struct node_hstate
*nhs
= &node_hstates
[nid
];
1799 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1800 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1812 * Unregister hstate attributes from a single node device.
1813 * No-op if no hstate attributes attached.
1815 static void hugetlb_unregister_node(struct node
*node
)
1818 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1820 if (!nhs
->hugepages_kobj
)
1821 return; /* no hstate attributes */
1823 for_each_hstate(h
) {
1824 int idx
= hstate_index(h
);
1825 if (nhs
->hstate_kobjs
[idx
]) {
1826 kobject_put(nhs
->hstate_kobjs
[idx
]);
1827 nhs
->hstate_kobjs
[idx
] = NULL
;
1831 kobject_put(nhs
->hugepages_kobj
);
1832 nhs
->hugepages_kobj
= NULL
;
1836 * hugetlb module exit: unregister hstate attributes from node devices
1839 static void hugetlb_unregister_all_nodes(void)
1844 * disable node device registrations.
1846 register_hugetlbfs_with_node(NULL
, NULL
);
1849 * remove hstate attributes from any nodes that have them.
1851 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1852 hugetlb_unregister_node(node_devices
[nid
]);
1856 * Register hstate attributes for a single node device.
1857 * No-op if attributes already registered.
1859 static void hugetlb_register_node(struct node
*node
)
1862 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1865 if (nhs
->hugepages_kobj
)
1866 return; /* already allocated */
1868 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1870 if (!nhs
->hugepages_kobj
)
1873 for_each_hstate(h
) {
1874 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1876 &per_node_hstate_attr_group
);
1878 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1879 h
->name
, node
->dev
.id
);
1880 hugetlb_unregister_node(node
);
1887 * hugetlb init time: register hstate attributes for all registered node
1888 * devices of nodes that have memory. All on-line nodes should have
1889 * registered their associated device by this time.
1891 static void hugetlb_register_all_nodes(void)
1895 for_each_node_state(nid
, N_MEMORY
) {
1896 struct node
*node
= node_devices
[nid
];
1897 if (node
->dev
.id
== nid
)
1898 hugetlb_register_node(node
);
1902 * Let the node device driver know we're here so it can
1903 * [un]register hstate attributes on node hotplug.
1905 register_hugetlbfs_with_node(hugetlb_register_node
,
1906 hugetlb_unregister_node
);
1908 #else /* !CONFIG_NUMA */
1910 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1918 static void hugetlb_unregister_all_nodes(void) { }
1920 static void hugetlb_register_all_nodes(void) { }
1924 static void __exit
hugetlb_exit(void)
1928 hugetlb_unregister_all_nodes();
1930 for_each_hstate(h
) {
1931 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1934 kobject_put(hugepages_kobj
);
1936 module_exit(hugetlb_exit
);
1938 static int __init
hugetlb_init(void)
1940 /* Some platform decide whether they support huge pages at boot
1941 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1942 * there is no such support
1944 if (HPAGE_SHIFT
== 0)
1947 if (!size_to_hstate(default_hstate_size
)) {
1948 default_hstate_size
= HPAGE_SIZE
;
1949 if (!size_to_hstate(default_hstate_size
))
1950 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1952 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1953 if (default_hstate_max_huge_pages
)
1954 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1956 hugetlb_init_hstates();
1957 gather_bootmem_prealloc();
1960 hugetlb_sysfs_init();
1961 hugetlb_register_all_nodes();
1962 hugetlb_cgroup_file_init();
1966 module_init(hugetlb_init
);
1968 /* Should be called on processing a hugepagesz=... option */
1969 void __init
hugetlb_add_hstate(unsigned order
)
1974 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1975 pr_warning("hugepagesz= specified twice, ignoring\n");
1978 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1980 h
= &hstates
[hugetlb_max_hstate
++];
1982 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1983 h
->nr_huge_pages
= 0;
1984 h
->free_huge_pages
= 0;
1985 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1986 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1987 INIT_LIST_HEAD(&h
->hugepage_activelist
);
1988 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
1989 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
1990 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1991 huge_page_size(h
)/1024);
1996 static int __init
hugetlb_nrpages_setup(char *s
)
1999 static unsigned long *last_mhp
;
2002 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2003 * so this hugepages= parameter goes to the "default hstate".
2005 if (!hugetlb_max_hstate
)
2006 mhp
= &default_hstate_max_huge_pages
;
2008 mhp
= &parsed_hstate
->max_huge_pages
;
2010 if (mhp
== last_mhp
) {
2011 pr_warning("hugepages= specified twice without "
2012 "interleaving hugepagesz=, ignoring\n");
2016 if (sscanf(s
, "%lu", mhp
) <= 0)
2020 * Global state is always initialized later in hugetlb_init.
2021 * But we need to allocate >= MAX_ORDER hstates here early to still
2022 * use the bootmem allocator.
2024 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2025 hugetlb_hstate_alloc_pages(parsed_hstate
);
2031 __setup("hugepages=", hugetlb_nrpages_setup
);
2033 static int __init
hugetlb_default_setup(char *s
)
2035 default_hstate_size
= memparse(s
, &s
);
2038 __setup("default_hugepagesz=", hugetlb_default_setup
);
2040 static unsigned int cpuset_mems_nr(unsigned int *array
)
2043 unsigned int nr
= 0;
2045 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2051 #ifdef CONFIG_SYSCTL
2052 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2053 struct ctl_table
*table
, int write
,
2054 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2056 struct hstate
*h
= &default_hstate
;
2060 tmp
= h
->max_huge_pages
;
2062 if (write
&& h
->order
>= MAX_ORDER
)
2066 table
->maxlen
= sizeof(unsigned long);
2067 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2072 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2073 GFP_KERNEL
| __GFP_NORETRY
);
2074 if (!(obey_mempolicy
&&
2075 init_nodemask_of_mempolicy(nodes_allowed
))) {
2076 NODEMASK_FREE(nodes_allowed
);
2077 nodes_allowed
= &node_states
[N_MEMORY
];
2079 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2081 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2082 NODEMASK_FREE(nodes_allowed
);
2088 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2089 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2092 return hugetlb_sysctl_handler_common(false, table
, write
,
2093 buffer
, length
, ppos
);
2097 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2098 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2100 return hugetlb_sysctl_handler_common(true, table
, write
,
2101 buffer
, length
, ppos
);
2103 #endif /* CONFIG_NUMA */
2105 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2106 void __user
*buffer
,
2107 size_t *length
, loff_t
*ppos
)
2109 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2110 if (hugepages_treat_as_movable
)
2111 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2113 htlb_alloc_mask
= GFP_HIGHUSER
;
2117 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2118 void __user
*buffer
,
2119 size_t *length
, loff_t
*ppos
)
2121 struct hstate
*h
= &default_hstate
;
2125 tmp
= h
->nr_overcommit_huge_pages
;
2127 if (write
&& h
->order
>= MAX_ORDER
)
2131 table
->maxlen
= sizeof(unsigned long);
2132 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2137 spin_lock(&hugetlb_lock
);
2138 h
->nr_overcommit_huge_pages
= tmp
;
2139 spin_unlock(&hugetlb_lock
);
2145 #endif /* CONFIG_SYSCTL */
2147 void hugetlb_report_meminfo(struct seq_file
*m
)
2149 struct hstate
*h
= &default_hstate
;
2151 "HugePages_Total: %5lu\n"
2152 "HugePages_Free: %5lu\n"
2153 "HugePages_Rsvd: %5lu\n"
2154 "HugePages_Surp: %5lu\n"
2155 "Hugepagesize: %8lu kB\n",
2159 h
->surplus_huge_pages
,
2160 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2163 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2165 struct hstate
*h
= &default_hstate
;
2167 "Node %d HugePages_Total: %5u\n"
2168 "Node %d HugePages_Free: %5u\n"
2169 "Node %d HugePages_Surp: %5u\n",
2170 nid
, h
->nr_huge_pages_node
[nid
],
2171 nid
, h
->free_huge_pages_node
[nid
],
2172 nid
, h
->surplus_huge_pages_node
[nid
]);
2175 void hugetlb_show_meminfo(void)
2180 for_each_node_state(nid
, N_MEMORY
)
2182 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2184 h
->nr_huge_pages_node
[nid
],
2185 h
->free_huge_pages_node
[nid
],
2186 h
->surplus_huge_pages_node
[nid
],
2187 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2190 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2191 unsigned long hugetlb_total_pages(void)
2194 unsigned long nr_total_pages
= 0;
2197 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2198 return nr_total_pages
;
2201 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2205 spin_lock(&hugetlb_lock
);
2207 * When cpuset is configured, it breaks the strict hugetlb page
2208 * reservation as the accounting is done on a global variable. Such
2209 * reservation is completely rubbish in the presence of cpuset because
2210 * the reservation is not checked against page availability for the
2211 * current cpuset. Application can still potentially OOM'ed by kernel
2212 * with lack of free htlb page in cpuset that the task is in.
2213 * Attempt to enforce strict accounting with cpuset is almost
2214 * impossible (or too ugly) because cpuset is too fluid that
2215 * task or memory node can be dynamically moved between cpusets.
2217 * The change of semantics for shared hugetlb mapping with cpuset is
2218 * undesirable. However, in order to preserve some of the semantics,
2219 * we fall back to check against current free page availability as
2220 * a best attempt and hopefully to minimize the impact of changing
2221 * semantics that cpuset has.
2224 if (gather_surplus_pages(h
, delta
) < 0)
2227 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2228 return_unused_surplus_pages(h
, delta
);
2235 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2238 spin_unlock(&hugetlb_lock
);
2242 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2244 struct resv_map
*resv
= vma_resv_map(vma
);
2247 * This new VMA should share its siblings reservation map if present.
2248 * The VMA will only ever have a valid reservation map pointer where
2249 * it is being copied for another still existing VMA. As that VMA
2250 * has a reference to the reservation map it cannot disappear until
2251 * after this open call completes. It is therefore safe to take a
2252 * new reference here without additional locking.
2255 kref_get(&resv
->refs
);
2258 static void resv_map_put(struct vm_area_struct
*vma
)
2260 struct resv_map
*resv
= vma_resv_map(vma
);
2264 kref_put(&resv
->refs
, resv_map_release
);
2267 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2269 struct hstate
*h
= hstate_vma(vma
);
2270 struct resv_map
*resv
= vma_resv_map(vma
);
2271 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2272 unsigned long reserve
;
2273 unsigned long start
;
2277 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2278 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2280 reserve
= (end
- start
) -
2281 region_count(&resv
->regions
, start
, end
);
2286 hugetlb_acct_memory(h
, -reserve
);
2287 hugepage_subpool_put_pages(spool
, reserve
);
2293 * We cannot handle pagefaults against hugetlb pages at all. They cause
2294 * handle_mm_fault() to try to instantiate regular-sized pages in the
2295 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2298 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2304 const struct vm_operations_struct hugetlb_vm_ops
= {
2305 .fault
= hugetlb_vm_op_fault
,
2306 .open
= hugetlb_vm_op_open
,
2307 .close
= hugetlb_vm_op_close
,
2310 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2316 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2317 vma
->vm_page_prot
)));
2319 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2320 vma
->vm_page_prot
));
2322 entry
= pte_mkyoung(entry
);
2323 entry
= pte_mkhuge(entry
);
2324 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2329 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2330 unsigned long address
, pte_t
*ptep
)
2334 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2335 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2336 update_mmu_cache(vma
, address
, ptep
);
2339 static int is_hugetlb_entry_migration(pte_t pte
)
2343 if (huge_pte_none(pte
) || pte_present(pte
))
2345 swp
= pte_to_swp_entry(pte
);
2346 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2352 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2356 if (huge_pte_none(pte
) || pte_present(pte
))
2358 swp
= pte_to_swp_entry(pte
);
2359 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2365 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2366 struct vm_area_struct
*vma
)
2368 pte_t
*src_pte
, *dst_pte
, entry
;
2369 struct page
*ptepage
;
2372 struct hstate
*h
= hstate_vma(vma
);
2373 unsigned long sz
= huge_page_size(h
);
2374 unsigned long mmun_start
; /* For mmu_notifiers */
2375 unsigned long mmun_end
; /* For mmu_notifiers */
2378 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2380 mmun_start
= vma
->vm_start
;
2381 mmun_end
= vma
->vm_end
;
2383 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
2385 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2386 src_pte
= huge_pte_offset(src
, addr
);
2389 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2395 /* If the pagetables are shared don't copy or take references */
2396 if (dst_pte
== src_pte
)
2399 spin_lock(&dst
->page_table_lock
);
2400 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2401 entry
= huge_ptep_get(src_pte
);
2402 if (huge_pte_none(entry
)) { /* skip none entry */
2404 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
2405 is_hugetlb_entry_hwpoisoned(entry
))) {
2406 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
2408 if (is_write_migration_entry(swp_entry
) && cow
) {
2410 * COW mappings require pages in both
2411 * parent and child to be set to read.
2413 make_migration_entry_read(&swp_entry
);
2414 entry
= swp_entry_to_pte(swp_entry
);
2415 set_huge_pte_at(src
, addr
, src_pte
, entry
);
2417 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2420 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2421 entry
= huge_ptep_get(src_pte
);
2422 ptepage
= pte_page(entry
);
2424 page_dup_rmap(ptepage
);
2425 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2427 spin_unlock(&src
->page_table_lock
);
2428 spin_unlock(&dst
->page_table_lock
);
2432 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
2437 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2438 unsigned long start
, unsigned long end
,
2439 struct page
*ref_page
)
2441 int force_flush
= 0;
2442 struct mm_struct
*mm
= vma
->vm_mm
;
2443 unsigned long address
;
2447 struct hstate
*h
= hstate_vma(vma
);
2448 unsigned long sz
= huge_page_size(h
);
2449 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2450 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2452 WARN_ON(!is_vm_hugetlb_page(vma
));
2453 BUG_ON(start
& ~huge_page_mask(h
));
2454 BUG_ON(end
& ~huge_page_mask(h
));
2456 tlb_start_vma(tlb
, vma
);
2457 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2459 spin_lock(&mm
->page_table_lock
);
2460 for (address
= start
; address
< end
; address
+= sz
) {
2461 ptep
= huge_pte_offset(mm
, address
);
2465 if (huge_pmd_unshare(mm
, &address
, ptep
))
2468 pte
= huge_ptep_get(ptep
);
2469 if (huge_pte_none(pte
))
2473 * Migrating hugepage or HWPoisoned hugepage is already
2474 * unmapped and its refcount is dropped, so just clear pte here.
2476 if (unlikely(!pte_present(pte
))) {
2477 huge_pte_clear(mm
, address
, ptep
);
2481 page
= pte_page(pte
);
2483 * If a reference page is supplied, it is because a specific
2484 * page is being unmapped, not a range. Ensure the page we
2485 * are about to unmap is the actual page of interest.
2488 if (page
!= ref_page
)
2492 * Mark the VMA as having unmapped its page so that
2493 * future faults in this VMA will fail rather than
2494 * looking like data was lost
2496 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2499 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2500 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2501 if (huge_pte_dirty(pte
))
2502 set_page_dirty(page
);
2504 page_remove_rmap(page
);
2505 force_flush
= !__tlb_remove_page(tlb
, page
);
2508 /* Bail out after unmapping reference page if supplied */
2512 spin_unlock(&mm
->page_table_lock
);
2514 * mmu_gather ran out of room to batch pages, we break out of
2515 * the PTE lock to avoid doing the potential expensive TLB invalidate
2516 * and page-free while holding it.
2521 if (address
< end
&& !ref_page
)
2524 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2525 tlb_end_vma(tlb
, vma
);
2528 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2529 struct vm_area_struct
*vma
, unsigned long start
,
2530 unsigned long end
, struct page
*ref_page
)
2532 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2535 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2536 * test will fail on a vma being torn down, and not grab a page table
2537 * on its way out. We're lucky that the flag has such an appropriate
2538 * name, and can in fact be safely cleared here. We could clear it
2539 * before the __unmap_hugepage_range above, but all that's necessary
2540 * is to clear it before releasing the i_mmap_mutex. This works
2541 * because in the context this is called, the VMA is about to be
2542 * destroyed and the i_mmap_mutex is held.
2544 vma
->vm_flags
&= ~VM_MAYSHARE
;
2547 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2548 unsigned long end
, struct page
*ref_page
)
2550 struct mm_struct
*mm
;
2551 struct mmu_gather tlb
;
2555 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2556 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2557 tlb_finish_mmu(&tlb
, start
, end
);
2561 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2562 * mappping it owns the reserve page for. The intention is to unmap the page
2563 * from other VMAs and let the children be SIGKILLed if they are faulting the
2566 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2567 struct page
*page
, unsigned long address
)
2569 struct hstate
*h
= hstate_vma(vma
);
2570 struct vm_area_struct
*iter_vma
;
2571 struct address_space
*mapping
;
2575 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2576 * from page cache lookup which is in HPAGE_SIZE units.
2578 address
= address
& huge_page_mask(h
);
2579 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2581 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2584 * Take the mapping lock for the duration of the table walk. As
2585 * this mapping should be shared between all the VMAs,
2586 * __unmap_hugepage_range() is called as the lock is already held
2588 mutex_lock(&mapping
->i_mmap_mutex
);
2589 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2590 /* Do not unmap the current VMA */
2591 if (iter_vma
== vma
)
2595 * Shared VMAs have their own reserves and do not affect
2596 * MAP_PRIVATE accounting but it is possible that a shared
2597 * VMA is using the same page so check and skip such VMAs.
2599 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
2603 * Unmap the page from other VMAs without their own reserves.
2604 * They get marked to be SIGKILLed if they fault in these
2605 * areas. This is because a future no-page fault on this VMA
2606 * could insert a zeroed page instead of the data existing
2607 * from the time of fork. This would look like data corruption
2609 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2610 unmap_hugepage_range(iter_vma
, address
,
2611 address
+ huge_page_size(h
), page
);
2613 mutex_unlock(&mapping
->i_mmap_mutex
);
2619 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2620 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2621 * cannot race with other handlers or page migration.
2622 * Keep the pte_same checks anyway to make transition from the mutex easier.
2624 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2625 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2626 struct page
*pagecache_page
)
2628 struct hstate
*h
= hstate_vma(vma
);
2629 struct page
*old_page
, *new_page
;
2630 int outside_reserve
= 0;
2631 unsigned long mmun_start
; /* For mmu_notifiers */
2632 unsigned long mmun_end
; /* For mmu_notifiers */
2634 old_page
= pte_page(pte
);
2637 /* If no-one else is actually using this page, avoid the copy
2638 * and just make the page writable */
2639 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2640 page_move_anon_rmap(old_page
, vma
, address
);
2641 set_huge_ptep_writable(vma
, address
, ptep
);
2646 * If the process that created a MAP_PRIVATE mapping is about to
2647 * perform a COW due to a shared page count, attempt to satisfy
2648 * the allocation without using the existing reserves. The pagecache
2649 * page is used to determine if the reserve at this address was
2650 * consumed or not. If reserves were used, a partial faulted mapping
2651 * at the time of fork() could consume its reserves on COW instead
2652 * of the full address range.
2654 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2655 old_page
!= pagecache_page
)
2656 outside_reserve
= 1;
2658 page_cache_get(old_page
);
2660 /* Drop page_table_lock as buddy allocator may be called */
2661 spin_unlock(&mm
->page_table_lock
);
2662 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2664 if (IS_ERR(new_page
)) {
2665 long err
= PTR_ERR(new_page
);
2666 page_cache_release(old_page
);
2669 * If a process owning a MAP_PRIVATE mapping fails to COW,
2670 * it is due to references held by a child and an insufficient
2671 * huge page pool. To guarantee the original mappers
2672 * reliability, unmap the page from child processes. The child
2673 * may get SIGKILLed if it later faults.
2675 if (outside_reserve
) {
2676 BUG_ON(huge_pte_none(pte
));
2677 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2678 BUG_ON(huge_pte_none(pte
));
2679 spin_lock(&mm
->page_table_lock
);
2680 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2681 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2682 goto retry_avoidcopy
;
2684 * race occurs while re-acquiring page_table_lock, and
2692 /* Caller expects lock to be held */
2693 spin_lock(&mm
->page_table_lock
);
2695 return VM_FAULT_OOM
;
2697 return VM_FAULT_SIGBUS
;
2701 * When the original hugepage is shared one, it does not have
2702 * anon_vma prepared.
2704 if (unlikely(anon_vma_prepare(vma
))) {
2705 page_cache_release(new_page
);
2706 page_cache_release(old_page
);
2707 /* Caller expects lock to be held */
2708 spin_lock(&mm
->page_table_lock
);
2709 return VM_FAULT_OOM
;
2712 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2713 pages_per_huge_page(h
));
2714 __SetPageUptodate(new_page
);
2716 mmun_start
= address
& huge_page_mask(h
);
2717 mmun_end
= mmun_start
+ huge_page_size(h
);
2718 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2720 * Retake the page_table_lock to check for racing updates
2721 * before the page tables are altered
2723 spin_lock(&mm
->page_table_lock
);
2724 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2725 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2726 ClearPagePrivate(new_page
);
2729 huge_ptep_clear_flush(vma
, address
, ptep
);
2730 set_huge_pte_at(mm
, address
, ptep
,
2731 make_huge_pte(vma
, new_page
, 1));
2732 page_remove_rmap(old_page
);
2733 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2734 /* Make the old page be freed below */
2735 new_page
= old_page
;
2737 spin_unlock(&mm
->page_table_lock
);
2738 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2739 page_cache_release(new_page
);
2740 page_cache_release(old_page
);
2742 /* Caller expects lock to be held */
2743 spin_lock(&mm
->page_table_lock
);
2747 /* Return the pagecache page at a given address within a VMA */
2748 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2749 struct vm_area_struct
*vma
, unsigned long address
)
2751 struct address_space
*mapping
;
2754 mapping
= vma
->vm_file
->f_mapping
;
2755 idx
= vma_hugecache_offset(h
, vma
, address
);
2757 return find_lock_page(mapping
, idx
);
2761 * Return whether there is a pagecache page to back given address within VMA.
2762 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2764 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2765 struct vm_area_struct
*vma
, unsigned long address
)
2767 struct address_space
*mapping
;
2771 mapping
= vma
->vm_file
->f_mapping
;
2772 idx
= vma_hugecache_offset(h
, vma
, address
);
2774 page
= find_get_page(mapping
, idx
);
2777 return page
!= NULL
;
2780 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2781 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2783 struct hstate
*h
= hstate_vma(vma
);
2784 int ret
= VM_FAULT_SIGBUS
;
2789 struct address_space
*mapping
;
2793 * Currently, we are forced to kill the process in the event the
2794 * original mapper has unmapped pages from the child due to a failed
2795 * COW. Warn that such a situation has occurred as it may not be obvious
2797 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2798 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2803 mapping
= vma
->vm_file
->f_mapping
;
2804 idx
= vma_hugecache_offset(h
, vma
, address
);
2807 * Use page lock to guard against racing truncation
2808 * before we get page_table_lock.
2811 page
= find_lock_page(mapping
, idx
);
2813 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2816 page
= alloc_huge_page(vma
, address
, 0);
2818 ret
= PTR_ERR(page
);
2822 ret
= VM_FAULT_SIGBUS
;
2825 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2826 __SetPageUptodate(page
);
2828 if (vma
->vm_flags
& VM_MAYSHARE
) {
2830 struct inode
*inode
= mapping
->host
;
2832 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2839 ClearPagePrivate(page
);
2841 spin_lock(&inode
->i_lock
);
2842 inode
->i_blocks
+= blocks_per_huge_page(h
);
2843 spin_unlock(&inode
->i_lock
);
2846 if (unlikely(anon_vma_prepare(vma
))) {
2848 goto backout_unlocked
;
2854 * If memory error occurs between mmap() and fault, some process
2855 * don't have hwpoisoned swap entry for errored virtual address.
2856 * So we need to block hugepage fault by PG_hwpoison bit check.
2858 if (unlikely(PageHWPoison(page
))) {
2859 ret
= VM_FAULT_HWPOISON
|
2860 VM_FAULT_SET_HINDEX(hstate_index(h
));
2861 goto backout_unlocked
;
2866 * If we are going to COW a private mapping later, we examine the
2867 * pending reservations for this page now. This will ensure that
2868 * any allocations necessary to record that reservation occur outside
2871 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2872 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2874 goto backout_unlocked
;
2877 spin_lock(&mm
->page_table_lock
);
2878 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2883 if (!huge_pte_none(huge_ptep_get(ptep
)))
2887 ClearPagePrivate(page
);
2888 hugepage_add_new_anon_rmap(page
, vma
, address
);
2891 page_dup_rmap(page
);
2892 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2893 && (vma
->vm_flags
& VM_SHARED
)));
2894 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2896 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2897 /* Optimization, do the COW without a second fault */
2898 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2901 spin_unlock(&mm
->page_table_lock
);
2907 spin_unlock(&mm
->page_table_lock
);
2914 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2915 unsigned long address
, unsigned int flags
)
2920 struct page
*page
= NULL
;
2921 struct page
*pagecache_page
= NULL
;
2922 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2923 struct hstate
*h
= hstate_vma(vma
);
2925 address
&= huge_page_mask(h
);
2927 ptep
= huge_pte_offset(mm
, address
);
2929 entry
= huge_ptep_get(ptep
);
2930 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2931 migration_entry_wait_huge(mm
, ptep
);
2933 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2934 return VM_FAULT_HWPOISON_LARGE
|
2935 VM_FAULT_SET_HINDEX(hstate_index(h
));
2938 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2940 return VM_FAULT_OOM
;
2943 * Serialize hugepage allocation and instantiation, so that we don't
2944 * get spurious allocation failures if two CPUs race to instantiate
2945 * the same page in the page cache.
2947 mutex_lock(&hugetlb_instantiation_mutex
);
2948 entry
= huge_ptep_get(ptep
);
2949 if (huge_pte_none(entry
)) {
2950 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2957 * If we are going to COW the mapping later, we examine the pending
2958 * reservations for this page now. This will ensure that any
2959 * allocations necessary to record that reservation occur outside the
2960 * spinlock. For private mappings, we also lookup the pagecache
2961 * page now as it is used to determine if a reservation has been
2964 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2965 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2970 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2971 pagecache_page
= hugetlbfs_pagecache_page(h
,
2976 * hugetlb_cow() requires page locks of pte_page(entry) and
2977 * pagecache_page, so here we need take the former one
2978 * when page != pagecache_page or !pagecache_page.
2979 * Note that locking order is always pagecache_page -> page,
2980 * so no worry about deadlock.
2982 page
= pte_page(entry
);
2984 if (page
!= pagecache_page
)
2987 spin_lock(&mm
->page_table_lock
);
2988 /* Check for a racing update before calling hugetlb_cow */
2989 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2990 goto out_page_table_lock
;
2993 if (flags
& FAULT_FLAG_WRITE
) {
2994 if (!huge_pte_write(entry
)) {
2995 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2997 goto out_page_table_lock
;
2999 entry
= huge_pte_mkdirty(entry
);
3001 entry
= pte_mkyoung(entry
);
3002 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3003 flags
& FAULT_FLAG_WRITE
))
3004 update_mmu_cache(vma
, address
, ptep
);
3006 out_page_table_lock
:
3007 spin_unlock(&mm
->page_table_lock
);
3009 if (pagecache_page
) {
3010 unlock_page(pagecache_page
);
3011 put_page(pagecache_page
);
3013 if (page
!= pagecache_page
)
3018 mutex_unlock(&hugetlb_instantiation_mutex
);
3023 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3024 struct page
**pages
, struct vm_area_struct
**vmas
,
3025 unsigned long *position
, unsigned long *nr_pages
,
3026 long i
, unsigned int flags
)
3028 unsigned long pfn_offset
;
3029 unsigned long vaddr
= *position
;
3030 unsigned long remainder
= *nr_pages
;
3031 struct hstate
*h
= hstate_vma(vma
);
3033 spin_lock(&mm
->page_table_lock
);
3034 while (vaddr
< vma
->vm_end
&& remainder
) {
3040 * Some archs (sparc64, sh*) have multiple pte_ts to
3041 * each hugepage. We have to make sure we get the
3042 * first, for the page indexing below to work.
3044 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3045 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3048 * When coredumping, it suits get_dump_page if we just return
3049 * an error where there's an empty slot with no huge pagecache
3050 * to back it. This way, we avoid allocating a hugepage, and
3051 * the sparse dumpfile avoids allocating disk blocks, but its
3052 * huge holes still show up with zeroes where they need to be.
3054 if (absent
&& (flags
& FOLL_DUMP
) &&
3055 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3061 * We need call hugetlb_fault for both hugepages under migration
3062 * (in which case hugetlb_fault waits for the migration,) and
3063 * hwpoisoned hugepages (in which case we need to prevent the
3064 * caller from accessing to them.) In order to do this, we use
3065 * here is_swap_pte instead of is_hugetlb_entry_migration and
3066 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3067 * both cases, and because we can't follow correct pages
3068 * directly from any kind of swap entries.
3070 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3071 ((flags
& FOLL_WRITE
) &&
3072 !huge_pte_write(huge_ptep_get(pte
)))) {
3075 spin_unlock(&mm
->page_table_lock
);
3076 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3077 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3078 spin_lock(&mm
->page_table_lock
);
3079 if (!(ret
& VM_FAULT_ERROR
))
3086 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3087 page
= pte_page(huge_ptep_get(pte
));
3090 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3101 if (vaddr
< vma
->vm_end
&& remainder
&&
3102 pfn_offset
< pages_per_huge_page(h
)) {
3104 * We use pfn_offset to avoid touching the pageframes
3105 * of this compound page.
3110 spin_unlock(&mm
->page_table_lock
);
3111 *nr_pages
= remainder
;
3114 return i
? i
: -EFAULT
;
3117 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3118 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3120 struct mm_struct
*mm
= vma
->vm_mm
;
3121 unsigned long start
= address
;
3124 struct hstate
*h
= hstate_vma(vma
);
3125 unsigned long pages
= 0;
3127 BUG_ON(address
>= end
);
3128 flush_cache_range(vma
, address
, end
);
3130 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3131 spin_lock(&mm
->page_table_lock
);
3132 for (; address
< end
; address
+= huge_page_size(h
)) {
3133 ptep
= huge_pte_offset(mm
, address
);
3136 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3140 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3141 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3142 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3143 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3144 set_huge_pte_at(mm
, address
, ptep
, pte
);
3148 spin_unlock(&mm
->page_table_lock
);
3150 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3151 * may have cleared our pud entry and done put_page on the page table:
3152 * once we release i_mmap_mutex, another task can do the final put_page
3153 * and that page table be reused and filled with junk.
3155 flush_tlb_range(vma
, start
, end
);
3156 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3158 return pages
<< h
->order
;
3161 int hugetlb_reserve_pages(struct inode
*inode
,
3163 struct vm_area_struct
*vma
,
3164 vm_flags_t vm_flags
)
3167 struct hstate
*h
= hstate_inode(inode
);
3168 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3171 * Only apply hugepage reservation if asked. At fault time, an
3172 * attempt will be made for VM_NORESERVE to allocate a page
3173 * without using reserves
3175 if (vm_flags
& VM_NORESERVE
)
3179 * Shared mappings base their reservation on the number of pages that
3180 * are already allocated on behalf of the file. Private mappings need
3181 * to reserve the full area even if read-only as mprotect() may be
3182 * called to make the mapping read-write. Assume !vma is a shm mapping
3184 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3185 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3187 struct resv_map
*resv_map
= resv_map_alloc();
3193 set_vma_resv_map(vma
, resv_map
);
3194 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3202 /* There must be enough pages in the subpool for the mapping */
3203 if (hugepage_subpool_get_pages(spool
, chg
)) {
3209 * Check enough hugepages are available for the reservation.
3210 * Hand the pages back to the subpool if there are not
3212 ret
= hugetlb_acct_memory(h
, chg
);
3214 hugepage_subpool_put_pages(spool
, chg
);
3219 * Account for the reservations made. Shared mappings record regions
3220 * that have reservations as they are shared by multiple VMAs.
3221 * When the last VMA disappears, the region map says how much
3222 * the reservation was and the page cache tells how much of
3223 * the reservation was consumed. Private mappings are per-VMA and
3224 * only the consumed reservations are tracked. When the VMA
3225 * disappears, the original reservation is the VMA size and the
3226 * consumed reservations are stored in the map. Hence, nothing
3227 * else has to be done for private mappings here
3229 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3230 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3238 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3240 struct hstate
*h
= hstate_inode(inode
);
3241 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3242 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3244 spin_lock(&inode
->i_lock
);
3245 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3246 spin_unlock(&inode
->i_lock
);
3248 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3249 hugetlb_acct_memory(h
, -(chg
- freed
));
3252 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3253 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3254 struct vm_area_struct
*vma
,
3255 unsigned long addr
, pgoff_t idx
)
3257 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3259 unsigned long sbase
= saddr
& PUD_MASK
;
3260 unsigned long s_end
= sbase
+ PUD_SIZE
;
3262 /* Allow segments to share if only one is marked locked */
3263 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3264 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3267 * match the virtual addresses, permission and the alignment of the
3270 if (pmd_index(addr
) != pmd_index(saddr
) ||
3271 vm_flags
!= svm_flags
||
3272 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3278 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3280 unsigned long base
= addr
& PUD_MASK
;
3281 unsigned long end
= base
+ PUD_SIZE
;
3284 * check on proper vm_flags and page table alignment
3286 if (vma
->vm_flags
& VM_MAYSHARE
&&
3287 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3293 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3294 * and returns the corresponding pte. While this is not necessary for the
3295 * !shared pmd case because we can allocate the pmd later as well, it makes the
3296 * code much cleaner. pmd allocation is essential for the shared case because
3297 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3298 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3299 * bad pmd for sharing.
3301 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3303 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3304 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3305 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3307 struct vm_area_struct
*svma
;
3308 unsigned long saddr
;
3312 if (!vma_shareable(vma
, addr
))
3313 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3315 mutex_lock(&mapping
->i_mmap_mutex
);
3316 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3320 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3322 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3324 get_page(virt_to_page(spte
));
3333 spin_lock(&mm
->page_table_lock
);
3335 pud_populate(mm
, pud
,
3336 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3338 put_page(virt_to_page(spte
));
3339 spin_unlock(&mm
->page_table_lock
);
3341 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3342 mutex_unlock(&mapping
->i_mmap_mutex
);
3347 * unmap huge page backed by shared pte.
3349 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3350 * indicated by page_count > 1, unmap is achieved by clearing pud and
3351 * decrementing the ref count. If count == 1, the pte page is not shared.
3353 * called with vma->vm_mm->page_table_lock held.
3355 * returns: 1 successfully unmapped a shared pte page
3356 * 0 the underlying pte page is not shared, or it is the last user
3358 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3360 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3361 pud_t
*pud
= pud_offset(pgd
, *addr
);
3363 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3364 if (page_count(virt_to_page(ptep
)) == 1)
3368 put_page(virt_to_page(ptep
));
3369 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3372 #define want_pmd_share() (1)
3373 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3374 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3378 #define want_pmd_share() (0)
3379 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3381 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3382 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3383 unsigned long addr
, unsigned long sz
)
3389 pgd
= pgd_offset(mm
, addr
);
3390 pud
= pud_alloc(mm
, pgd
, addr
);
3392 if (sz
== PUD_SIZE
) {
3395 BUG_ON(sz
!= PMD_SIZE
);
3396 if (want_pmd_share() && pud_none(*pud
))
3397 pte
= huge_pmd_share(mm
, addr
, pud
);
3399 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3402 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3407 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3413 pgd
= pgd_offset(mm
, addr
);
3414 if (pgd_present(*pgd
)) {
3415 pud
= pud_offset(pgd
, addr
);
3416 if (pud_present(*pud
)) {
3418 return (pte_t
*)pud
;
3419 pmd
= pmd_offset(pud
, addr
);
3422 return (pte_t
*) pmd
;
3426 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3427 pmd_t
*pmd
, int write
)
3431 page
= pte_page(*(pte_t
*)pmd
);
3433 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3438 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3439 pud_t
*pud
, int write
)
3443 page
= pte_page(*(pte_t
*)pud
);
3445 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3449 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3451 /* Can be overriden by architectures */
3452 __attribute__((weak
)) struct page
*
3453 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3454 pud_t
*pud
, int write
)
3460 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3462 #ifdef CONFIG_MEMORY_FAILURE
3464 /* Should be called in hugetlb_lock */
3465 static int is_hugepage_on_freelist(struct page
*hpage
)
3469 struct hstate
*h
= page_hstate(hpage
);
3470 int nid
= page_to_nid(hpage
);
3472 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3479 * This function is called from memory failure code.
3480 * Assume the caller holds page lock of the head page.
3482 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3484 struct hstate
*h
= page_hstate(hpage
);
3485 int nid
= page_to_nid(hpage
);
3488 spin_lock(&hugetlb_lock
);
3489 if (is_hugepage_on_freelist(hpage
)) {
3491 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3492 * but dangling hpage->lru can trigger list-debug warnings
3493 * (this happens when we call unpoison_memory() on it),
3494 * so let it point to itself with list_del_init().
3496 list_del_init(&hpage
->lru
);
3497 set_page_refcounted(hpage
);
3498 h
->free_huge_pages
--;
3499 h
->free_huge_pages_node
[nid
]--;
3502 spin_unlock(&hugetlb_lock
);