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 unsigned long hugepages_treat_as_movable
;
39 int hugetlb_max_hstate __read_mostly
;
40 unsigned int default_hstate_idx
;
41 struct hstate hstates
[HUGE_MAX_HSTATE
];
43 __initdata
LIST_HEAD(huge_boot_pages
);
45 /* for command line parsing */
46 static struct hstate
* __initdata parsed_hstate
;
47 static unsigned long __initdata default_hstate_max_huge_pages
;
48 static unsigned long __initdata default_hstate_size
;
51 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
52 * free_huge_pages, and surplus_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_instantiation_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 << huge_page_shift(hstate
);
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 /* Movability of hugepages depends on migration support. */
542 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
544 if (hugepages_treat_as_movable
|| hugepage_migration_support(h
))
545 return GFP_HIGHUSER_MOVABLE
;
550 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
551 struct vm_area_struct
*vma
,
552 unsigned long address
, int avoid_reserve
,
555 struct page
*page
= NULL
;
556 struct mempolicy
*mpol
;
557 nodemask_t
*nodemask
;
558 struct zonelist
*zonelist
;
561 unsigned int cpuset_mems_cookie
;
564 * A child process with MAP_PRIVATE mappings created by their parent
565 * have no page reserves. This check ensures that reservations are
566 * not "stolen". The child may still get SIGKILLed
568 if (!vma_has_reserves(vma
, chg
) &&
569 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
572 /* If reserves cannot be used, ensure enough pages are in the pool */
573 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
577 cpuset_mems_cookie
= get_mems_allowed();
578 zonelist
= huge_zonelist(vma
, address
,
579 htlb_alloc_mask(h
), &mpol
, &nodemask
);
581 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
582 MAX_NR_ZONES
- 1, nodemask
) {
583 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask(h
))) {
584 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
588 if (!vma_has_reserves(vma
, chg
))
591 SetPagePrivate(page
);
592 h
->resv_huge_pages
--;
599 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
607 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
611 VM_BUG_ON(h
->order
>= MAX_ORDER
);
614 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
615 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
616 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
617 1 << PG_referenced
| 1 << PG_dirty
|
618 1 << PG_active
| 1 << PG_reserved
|
619 1 << PG_private
| 1 << PG_writeback
);
621 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
622 set_compound_page_dtor(page
, NULL
);
623 set_page_refcounted(page
);
624 arch_release_hugepage(page
);
625 __free_pages(page
, huge_page_order(h
));
628 struct hstate
*size_to_hstate(unsigned long size
)
633 if (huge_page_size(h
) == size
)
639 static void free_huge_page(struct page
*page
)
642 * Can't pass hstate in here because it is called from the
643 * compound page destructor.
645 struct hstate
*h
= page_hstate(page
);
646 int nid
= page_to_nid(page
);
647 struct hugepage_subpool
*spool
=
648 (struct hugepage_subpool
*)page_private(page
);
649 bool restore_reserve
;
651 set_page_private(page
, 0);
652 page
->mapping
= NULL
;
653 BUG_ON(page_count(page
));
654 BUG_ON(page_mapcount(page
));
655 restore_reserve
= PagePrivate(page
);
656 ClearPagePrivate(page
);
658 spin_lock(&hugetlb_lock
);
659 hugetlb_cgroup_uncharge_page(hstate_index(h
),
660 pages_per_huge_page(h
), page
);
662 h
->resv_huge_pages
++;
664 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
665 /* remove the page from active list */
666 list_del(&page
->lru
);
667 update_and_free_page(h
, page
);
668 h
->surplus_huge_pages
--;
669 h
->surplus_huge_pages_node
[nid
]--;
671 arch_clear_hugepage_flags(page
);
672 enqueue_huge_page(h
, page
);
674 spin_unlock(&hugetlb_lock
);
675 hugepage_subpool_put_pages(spool
, 1);
678 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
680 INIT_LIST_HEAD(&page
->lru
);
681 set_compound_page_dtor(page
, free_huge_page
);
682 spin_lock(&hugetlb_lock
);
683 set_hugetlb_cgroup(page
, NULL
);
685 h
->nr_huge_pages_node
[nid
]++;
686 spin_unlock(&hugetlb_lock
);
687 put_page(page
); /* free it into the hugepage allocator */
690 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
693 int nr_pages
= 1 << order
;
694 struct page
*p
= page
+ 1;
696 /* we rely on prep_new_huge_page to set the destructor */
697 set_compound_order(page
, order
);
699 __ClearPageReserved(page
);
700 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
703 * For gigantic hugepages allocated through bootmem at
704 * boot, it's safer to be consistent with the not-gigantic
705 * hugepages and clear the PG_reserved bit from all tail pages
706 * too. Otherwse drivers using get_user_pages() to access tail
707 * pages may get the reference counting wrong if they see
708 * PG_reserved set on a tail page (despite the head page not
709 * having PG_reserved set). Enforcing this consistency between
710 * head and tail pages allows drivers to optimize away a check
711 * on the head page when they need know if put_page() is needed
712 * after get_user_pages().
714 __ClearPageReserved(p
);
715 set_page_count(p
, 0);
716 p
->first_page
= page
;
721 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
722 * transparent huge pages. See the PageTransHuge() documentation for more
725 int PageHuge(struct page
*page
)
727 compound_page_dtor
*dtor
;
729 if (!PageCompound(page
))
732 page
= compound_head(page
);
733 dtor
= get_compound_page_dtor(page
);
735 return dtor
== free_huge_page
;
737 EXPORT_SYMBOL_GPL(PageHuge
);
739 pgoff_t
__basepage_index(struct page
*page
)
741 struct page
*page_head
= compound_head(page
);
742 pgoff_t index
= page_index(page_head
);
743 unsigned long compound_idx
;
745 if (!PageHuge(page_head
))
746 return page_index(page
);
748 if (compound_order(page_head
) >= MAX_ORDER
)
749 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
751 compound_idx
= page
- page_head
;
753 return (index
<< compound_order(page_head
)) + compound_idx
;
756 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
760 if (h
->order
>= MAX_ORDER
)
763 page
= alloc_pages_exact_node(nid
,
764 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
765 __GFP_REPEAT
|__GFP_NOWARN
,
768 if (arch_prepare_hugepage(page
)) {
769 __free_pages(page
, huge_page_order(h
));
772 prep_new_huge_page(h
, page
, nid
);
779 * common helper functions for hstate_next_node_to_{alloc|free}.
780 * We may have allocated or freed a huge page based on a different
781 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
782 * be outside of *nodes_allowed. Ensure that we use an allowed
783 * node for alloc or free.
785 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
787 nid
= next_node(nid
, *nodes_allowed
);
788 if (nid
== MAX_NUMNODES
)
789 nid
= first_node(*nodes_allowed
);
790 VM_BUG_ON(nid
>= MAX_NUMNODES
);
795 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
797 if (!node_isset(nid
, *nodes_allowed
))
798 nid
= next_node_allowed(nid
, nodes_allowed
);
803 * returns the previously saved node ["this node"] from which to
804 * allocate a persistent huge page for the pool and advance the
805 * next node from which to allocate, handling wrap at end of node
808 static int hstate_next_node_to_alloc(struct hstate
*h
,
809 nodemask_t
*nodes_allowed
)
813 VM_BUG_ON(!nodes_allowed
);
815 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
816 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
822 * helper for free_pool_huge_page() - return the previously saved
823 * node ["this node"] from which to free a huge page. Advance the
824 * next node id whether or not we find a free huge page to free so
825 * that the next attempt to free addresses the next node.
827 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
831 VM_BUG_ON(!nodes_allowed
);
833 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
834 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
839 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
840 for (nr_nodes = nodes_weight(*mask); \
842 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
845 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
846 for (nr_nodes = nodes_weight(*mask); \
848 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
851 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
857 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
858 page
= alloc_fresh_huge_page_node(h
, node
);
866 count_vm_event(HTLB_BUDDY_PGALLOC
);
868 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
874 * Free huge page from pool from next node to free.
875 * Attempt to keep persistent huge pages more or less
876 * balanced over allowed nodes.
877 * Called with hugetlb_lock locked.
879 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
885 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
887 * If we're returning unused surplus pages, only examine
888 * nodes with surplus pages.
890 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
891 !list_empty(&h
->hugepage_freelists
[node
])) {
893 list_entry(h
->hugepage_freelists
[node
].next
,
895 list_del(&page
->lru
);
896 h
->free_huge_pages
--;
897 h
->free_huge_pages_node
[node
]--;
899 h
->surplus_huge_pages
--;
900 h
->surplus_huge_pages_node
[node
]--;
902 update_and_free_page(h
, page
);
912 * Dissolve a given free hugepage into free buddy pages. This function does
913 * nothing for in-use (including surplus) hugepages.
915 static void dissolve_free_huge_page(struct page
*page
)
917 spin_lock(&hugetlb_lock
);
918 if (PageHuge(page
) && !page_count(page
)) {
919 struct hstate
*h
= page_hstate(page
);
920 int nid
= page_to_nid(page
);
921 list_del(&page
->lru
);
922 h
->free_huge_pages
--;
923 h
->free_huge_pages_node
[nid
]--;
924 update_and_free_page(h
, page
);
926 spin_unlock(&hugetlb_lock
);
930 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
931 * make specified memory blocks removable from the system.
932 * Note that start_pfn should aligned with (minimum) hugepage size.
934 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
936 unsigned int order
= 8 * sizeof(void *);
940 /* Set scan step to minimum hugepage size */
942 if (order
> huge_page_order(h
))
943 order
= huge_page_order(h
);
944 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
945 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
946 dissolve_free_huge_page(pfn_to_page(pfn
));
949 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
954 if (h
->order
>= MAX_ORDER
)
958 * Assume we will successfully allocate the surplus page to
959 * prevent racing processes from causing the surplus to exceed
962 * This however introduces a different race, where a process B
963 * tries to grow the static hugepage pool while alloc_pages() is
964 * called by process A. B will only examine the per-node
965 * counters in determining if surplus huge pages can be
966 * converted to normal huge pages in adjust_pool_surplus(). A
967 * won't be able to increment the per-node counter, until the
968 * lock is dropped by B, but B doesn't drop hugetlb_lock until
969 * no more huge pages can be converted from surplus to normal
970 * state (and doesn't try to convert again). Thus, we have a
971 * case where a surplus huge page exists, the pool is grown, and
972 * the surplus huge page still exists after, even though it
973 * should just have been converted to a normal huge page. This
974 * does not leak memory, though, as the hugepage will be freed
975 * once it is out of use. It also does not allow the counters to
976 * go out of whack in adjust_pool_surplus() as we don't modify
977 * the node values until we've gotten the hugepage and only the
978 * per-node value is checked there.
980 spin_lock(&hugetlb_lock
);
981 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
982 spin_unlock(&hugetlb_lock
);
986 h
->surplus_huge_pages
++;
988 spin_unlock(&hugetlb_lock
);
990 if (nid
== NUMA_NO_NODE
)
991 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
992 __GFP_REPEAT
|__GFP_NOWARN
,
995 page
= alloc_pages_exact_node(nid
,
996 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
997 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
999 if (page
&& arch_prepare_hugepage(page
)) {
1000 __free_pages(page
, huge_page_order(h
));
1004 spin_lock(&hugetlb_lock
);
1006 INIT_LIST_HEAD(&page
->lru
);
1007 r_nid
= page_to_nid(page
);
1008 set_compound_page_dtor(page
, free_huge_page
);
1009 set_hugetlb_cgroup(page
, NULL
);
1011 * We incremented the global counters already
1013 h
->nr_huge_pages_node
[r_nid
]++;
1014 h
->surplus_huge_pages_node
[r_nid
]++;
1015 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1018 h
->surplus_huge_pages
--;
1019 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1021 spin_unlock(&hugetlb_lock
);
1027 * This allocation function is useful in the context where vma is irrelevant.
1028 * E.g. soft-offlining uses this function because it only cares physical
1029 * address of error page.
1031 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1033 struct page
*page
= NULL
;
1035 spin_lock(&hugetlb_lock
);
1036 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1037 page
= dequeue_huge_page_node(h
, nid
);
1038 spin_unlock(&hugetlb_lock
);
1041 page
= alloc_buddy_huge_page(h
, nid
);
1047 * Increase the hugetlb pool such that it can accommodate a reservation
1050 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1052 struct list_head surplus_list
;
1053 struct page
*page
, *tmp
;
1055 int needed
, allocated
;
1056 bool alloc_ok
= true;
1058 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1060 h
->resv_huge_pages
+= delta
;
1065 INIT_LIST_HEAD(&surplus_list
);
1069 spin_unlock(&hugetlb_lock
);
1070 for (i
= 0; i
< needed
; i
++) {
1071 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1076 list_add(&page
->lru
, &surplus_list
);
1081 * After retaking hugetlb_lock, we need to recalculate 'needed'
1082 * because either resv_huge_pages or free_huge_pages may have changed.
1084 spin_lock(&hugetlb_lock
);
1085 needed
= (h
->resv_huge_pages
+ delta
) -
1086 (h
->free_huge_pages
+ allocated
);
1091 * We were not able to allocate enough pages to
1092 * satisfy the entire reservation so we free what
1093 * we've allocated so far.
1098 * The surplus_list now contains _at_least_ the number of extra pages
1099 * needed to accommodate the reservation. Add the appropriate number
1100 * of pages to the hugetlb pool and free the extras back to the buddy
1101 * allocator. Commit the entire reservation here to prevent another
1102 * process from stealing the pages as they are added to the pool but
1103 * before they are reserved.
1105 needed
+= allocated
;
1106 h
->resv_huge_pages
+= delta
;
1109 /* Free the needed pages to the hugetlb pool */
1110 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1114 * This page is now managed by the hugetlb allocator and has
1115 * no users -- drop the buddy allocator's reference.
1117 put_page_testzero(page
);
1118 VM_BUG_ON(page_count(page
));
1119 enqueue_huge_page(h
, page
);
1122 spin_unlock(&hugetlb_lock
);
1124 /* Free unnecessary surplus pages to the buddy allocator */
1125 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1127 spin_lock(&hugetlb_lock
);
1133 * When releasing a hugetlb pool reservation, any surplus pages that were
1134 * allocated to satisfy the reservation must be explicitly freed if they were
1136 * Called with hugetlb_lock held.
1138 static void return_unused_surplus_pages(struct hstate
*h
,
1139 unsigned long unused_resv_pages
)
1141 unsigned long nr_pages
;
1143 /* Uncommit the reservation */
1144 h
->resv_huge_pages
-= unused_resv_pages
;
1146 /* Cannot return gigantic pages currently */
1147 if (h
->order
>= MAX_ORDER
)
1150 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1153 * We want to release as many surplus pages as possible, spread
1154 * evenly across all nodes with memory. Iterate across these nodes
1155 * until we can no longer free unreserved surplus pages. This occurs
1156 * when the nodes with surplus pages have no free pages.
1157 * free_pool_huge_page() will balance the the freed pages across the
1158 * on-line nodes with memory and will handle the hstate accounting.
1160 while (nr_pages
--) {
1161 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1167 * Determine if the huge page at addr within the vma has an associated
1168 * reservation. Where it does not we will need to logically increase
1169 * reservation and actually increase subpool usage before an allocation
1170 * can occur. Where any new reservation would be required the
1171 * reservation change is prepared, but not committed. Once the page
1172 * has been allocated from the subpool and instantiated the change should
1173 * be committed via vma_commit_reservation. No action is required on
1176 static long vma_needs_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 return region_chg(&inode
->i_mapping
->private_list
,
1187 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1192 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1193 struct resv_map
*resv
= vma_resv_map(vma
);
1195 err
= region_chg(&resv
->regions
, idx
, idx
+ 1);
1201 static void vma_commit_reservation(struct hstate
*h
,
1202 struct vm_area_struct
*vma
, unsigned long addr
)
1204 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1205 struct inode
*inode
= mapping
->host
;
1207 if (vma
->vm_flags
& VM_MAYSHARE
) {
1208 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1209 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1211 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1212 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1213 struct resv_map
*resv
= vma_resv_map(vma
);
1215 /* Mark this page used in the map. */
1216 region_add(&resv
->regions
, idx
, idx
+ 1);
1220 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1221 unsigned long addr
, int avoid_reserve
)
1223 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1224 struct hstate
*h
= hstate_vma(vma
);
1228 struct hugetlb_cgroup
*h_cg
;
1230 idx
= hstate_index(h
);
1232 * Processes that did not create the mapping will have no
1233 * reserves and will not have accounted against subpool
1234 * limit. Check that the subpool limit can be made before
1235 * satisfying the allocation MAP_NORESERVE mappings may also
1236 * need pages and subpool limit allocated allocated if no reserve
1239 chg
= vma_needs_reservation(h
, vma
, addr
);
1241 return ERR_PTR(-ENOMEM
);
1242 if (chg
|| avoid_reserve
)
1243 if (hugepage_subpool_get_pages(spool
, 1))
1244 return ERR_PTR(-ENOSPC
);
1246 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1248 if (chg
|| avoid_reserve
)
1249 hugepage_subpool_put_pages(spool
, 1);
1250 return ERR_PTR(-ENOSPC
);
1252 spin_lock(&hugetlb_lock
);
1253 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1255 spin_unlock(&hugetlb_lock
);
1256 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1258 hugetlb_cgroup_uncharge_cgroup(idx
,
1259 pages_per_huge_page(h
),
1261 if (chg
|| avoid_reserve
)
1262 hugepage_subpool_put_pages(spool
, 1);
1263 return ERR_PTR(-ENOSPC
);
1265 spin_lock(&hugetlb_lock
);
1266 list_move(&page
->lru
, &h
->hugepage_activelist
);
1269 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1270 spin_unlock(&hugetlb_lock
);
1272 set_page_private(page
, (unsigned long)spool
);
1274 vma_commit_reservation(h
, vma
, addr
);
1279 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1280 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1281 * where no ERR_VALUE is expected to be returned.
1283 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1284 unsigned long addr
, int avoid_reserve
)
1286 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1292 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1294 struct huge_bootmem_page
*m
;
1297 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1300 addr
= __alloc_bootmem_node_nopanic(NODE_DATA(node
),
1301 huge_page_size(h
), huge_page_size(h
), 0);
1305 * Use the beginning of the huge page to store the
1306 * huge_bootmem_page struct (until gather_bootmem
1307 * puts them into the mem_map).
1316 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1317 /* Put them into a private list first because mem_map is not up yet */
1318 list_add(&m
->list
, &huge_boot_pages
);
1323 static void prep_compound_huge_page(struct page
*page
, int order
)
1325 if (unlikely(order
> (MAX_ORDER
- 1)))
1326 prep_compound_gigantic_page(page
, order
);
1328 prep_compound_page(page
, order
);
1331 /* Put bootmem huge pages into the standard lists after mem_map is up */
1332 static void __init
gather_bootmem_prealloc(void)
1334 struct huge_bootmem_page
*m
;
1336 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1337 struct hstate
*h
= m
->hstate
;
1340 #ifdef CONFIG_HIGHMEM
1341 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1342 free_bootmem_late((unsigned long)m
,
1343 sizeof(struct huge_bootmem_page
));
1345 page
= virt_to_page(m
);
1347 WARN_ON(page_count(page
) != 1);
1348 prep_compound_huge_page(page
, h
->order
);
1349 WARN_ON(PageReserved(page
));
1350 prep_new_huge_page(h
, page
, page_to_nid(page
));
1352 * If we had gigantic hugepages allocated at boot time, we need
1353 * to restore the 'stolen' pages to totalram_pages in order to
1354 * fix confusing memory reports from free(1) and another
1355 * side-effects, like CommitLimit going negative.
1357 if (h
->order
> (MAX_ORDER
- 1))
1358 adjust_managed_page_count(page
, 1 << h
->order
);
1362 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1366 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1367 if (h
->order
>= MAX_ORDER
) {
1368 if (!alloc_bootmem_huge_page(h
))
1370 } else if (!alloc_fresh_huge_page(h
,
1371 &node_states
[N_MEMORY
]))
1374 h
->max_huge_pages
= i
;
1377 static void __init
hugetlb_init_hstates(void)
1381 for_each_hstate(h
) {
1382 /* oversize hugepages were init'ed in early boot */
1383 if (h
->order
< MAX_ORDER
)
1384 hugetlb_hstate_alloc_pages(h
);
1388 static char * __init
memfmt(char *buf
, unsigned long n
)
1390 if (n
>= (1UL << 30))
1391 sprintf(buf
, "%lu GB", n
>> 30);
1392 else if (n
>= (1UL << 20))
1393 sprintf(buf
, "%lu MB", n
>> 20);
1395 sprintf(buf
, "%lu KB", n
>> 10);
1399 static void __init
report_hugepages(void)
1403 for_each_hstate(h
) {
1405 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1406 memfmt(buf
, huge_page_size(h
)),
1407 h
->free_huge_pages
);
1411 #ifdef CONFIG_HIGHMEM
1412 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1413 nodemask_t
*nodes_allowed
)
1417 if (h
->order
>= MAX_ORDER
)
1420 for_each_node_mask(i
, *nodes_allowed
) {
1421 struct page
*page
, *next
;
1422 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1423 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1424 if (count
>= h
->nr_huge_pages
)
1426 if (PageHighMem(page
))
1428 list_del(&page
->lru
);
1429 update_and_free_page(h
, page
);
1430 h
->free_huge_pages
--;
1431 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1436 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1437 nodemask_t
*nodes_allowed
)
1443 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1444 * balanced by operating on them in a round-robin fashion.
1445 * Returns 1 if an adjustment was made.
1447 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1452 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1455 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1456 if (h
->surplus_huge_pages_node
[node
])
1460 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1461 if (h
->surplus_huge_pages_node
[node
] <
1462 h
->nr_huge_pages_node
[node
])
1469 h
->surplus_huge_pages
+= delta
;
1470 h
->surplus_huge_pages_node
[node
] += delta
;
1474 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1475 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1476 nodemask_t
*nodes_allowed
)
1478 unsigned long min_count
, ret
;
1480 if (h
->order
>= MAX_ORDER
)
1481 return h
->max_huge_pages
;
1484 * Increase the pool size
1485 * First take pages out of surplus state. Then make up the
1486 * remaining difference by allocating fresh huge pages.
1488 * We might race with alloc_buddy_huge_page() here and be unable
1489 * to convert a surplus huge page to a normal huge page. That is
1490 * not critical, though, it just means the overall size of the
1491 * pool might be one hugepage larger than it needs to be, but
1492 * within all the constraints specified by the sysctls.
1494 spin_lock(&hugetlb_lock
);
1495 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1496 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1500 while (count
> persistent_huge_pages(h
)) {
1502 * If this allocation races such that we no longer need the
1503 * page, free_huge_page will handle it by freeing the page
1504 * and reducing the surplus.
1506 spin_unlock(&hugetlb_lock
);
1507 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1508 spin_lock(&hugetlb_lock
);
1512 /* Bail for signals. Probably ctrl-c from user */
1513 if (signal_pending(current
))
1518 * Decrease the pool size
1519 * First return free pages to the buddy allocator (being careful
1520 * to keep enough around to satisfy reservations). Then place
1521 * pages into surplus state as needed so the pool will shrink
1522 * to the desired size as pages become free.
1524 * By placing pages into the surplus state independent of the
1525 * overcommit value, we are allowing the surplus pool size to
1526 * exceed overcommit. There are few sane options here. Since
1527 * alloc_buddy_huge_page() is checking the global counter,
1528 * though, we'll note that we're not allowed to exceed surplus
1529 * and won't grow the pool anywhere else. Not until one of the
1530 * sysctls are changed, or the surplus pages go out of use.
1532 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1533 min_count
= max(count
, min_count
);
1534 try_to_free_low(h
, min_count
, nodes_allowed
);
1535 while (min_count
< persistent_huge_pages(h
)) {
1536 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1539 while (count
< persistent_huge_pages(h
)) {
1540 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1544 ret
= persistent_huge_pages(h
);
1545 spin_unlock(&hugetlb_lock
);
1549 #define HSTATE_ATTR_RO(_name) \
1550 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1552 #define HSTATE_ATTR(_name) \
1553 static struct kobj_attribute _name##_attr = \
1554 __ATTR(_name, 0644, _name##_show, _name##_store)
1556 static struct kobject
*hugepages_kobj
;
1557 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1559 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1561 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1565 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1566 if (hstate_kobjs
[i
] == kobj
) {
1568 *nidp
= NUMA_NO_NODE
;
1572 return kobj_to_node_hstate(kobj
, nidp
);
1575 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1576 struct kobj_attribute
*attr
, char *buf
)
1579 unsigned long nr_huge_pages
;
1582 h
= kobj_to_hstate(kobj
, &nid
);
1583 if (nid
== NUMA_NO_NODE
)
1584 nr_huge_pages
= h
->nr_huge_pages
;
1586 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1588 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1591 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1592 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1593 const char *buf
, size_t len
)
1597 unsigned long count
;
1599 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1601 err
= kstrtoul(buf
, 10, &count
);
1605 h
= kobj_to_hstate(kobj
, &nid
);
1606 if (h
->order
>= MAX_ORDER
) {
1611 if (nid
== NUMA_NO_NODE
) {
1613 * global hstate attribute
1615 if (!(obey_mempolicy
&&
1616 init_nodemask_of_mempolicy(nodes_allowed
))) {
1617 NODEMASK_FREE(nodes_allowed
);
1618 nodes_allowed
= &node_states
[N_MEMORY
];
1620 } else if (nodes_allowed
) {
1622 * per node hstate attribute: adjust count to global,
1623 * but restrict alloc/free to the specified node.
1625 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1626 init_nodemask_of_node(nodes_allowed
, nid
);
1628 nodes_allowed
= &node_states
[N_MEMORY
];
1630 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1632 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1633 NODEMASK_FREE(nodes_allowed
);
1637 NODEMASK_FREE(nodes_allowed
);
1641 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1642 struct kobj_attribute
*attr
, char *buf
)
1644 return nr_hugepages_show_common(kobj
, attr
, buf
);
1647 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1648 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1650 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1652 HSTATE_ATTR(nr_hugepages
);
1657 * hstate attribute for optionally mempolicy-based constraint on persistent
1658 * huge page alloc/free.
1660 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1661 struct kobj_attribute
*attr
, char *buf
)
1663 return nr_hugepages_show_common(kobj
, attr
, buf
);
1666 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1667 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1669 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1671 HSTATE_ATTR(nr_hugepages_mempolicy
);
1675 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1676 struct kobj_attribute
*attr
, char *buf
)
1678 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1679 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1682 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1683 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1686 unsigned long input
;
1687 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1689 if (h
->order
>= MAX_ORDER
)
1692 err
= kstrtoul(buf
, 10, &input
);
1696 spin_lock(&hugetlb_lock
);
1697 h
->nr_overcommit_huge_pages
= input
;
1698 spin_unlock(&hugetlb_lock
);
1702 HSTATE_ATTR(nr_overcommit_hugepages
);
1704 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1705 struct kobj_attribute
*attr
, char *buf
)
1708 unsigned long free_huge_pages
;
1711 h
= kobj_to_hstate(kobj
, &nid
);
1712 if (nid
== NUMA_NO_NODE
)
1713 free_huge_pages
= h
->free_huge_pages
;
1715 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1717 return sprintf(buf
, "%lu\n", free_huge_pages
);
1719 HSTATE_ATTR_RO(free_hugepages
);
1721 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1722 struct kobj_attribute
*attr
, char *buf
)
1724 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1725 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1727 HSTATE_ATTR_RO(resv_hugepages
);
1729 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1730 struct kobj_attribute
*attr
, char *buf
)
1733 unsigned long surplus_huge_pages
;
1736 h
= kobj_to_hstate(kobj
, &nid
);
1737 if (nid
== NUMA_NO_NODE
)
1738 surplus_huge_pages
= h
->surplus_huge_pages
;
1740 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1742 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1744 HSTATE_ATTR_RO(surplus_hugepages
);
1746 static struct attribute
*hstate_attrs
[] = {
1747 &nr_hugepages_attr
.attr
,
1748 &nr_overcommit_hugepages_attr
.attr
,
1749 &free_hugepages_attr
.attr
,
1750 &resv_hugepages_attr
.attr
,
1751 &surplus_hugepages_attr
.attr
,
1753 &nr_hugepages_mempolicy_attr
.attr
,
1758 static struct attribute_group hstate_attr_group
= {
1759 .attrs
= hstate_attrs
,
1762 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1763 struct kobject
**hstate_kobjs
,
1764 struct attribute_group
*hstate_attr_group
)
1767 int hi
= hstate_index(h
);
1769 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1770 if (!hstate_kobjs
[hi
])
1773 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1775 kobject_put(hstate_kobjs
[hi
]);
1780 static void __init
hugetlb_sysfs_init(void)
1785 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1786 if (!hugepages_kobj
)
1789 for_each_hstate(h
) {
1790 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1791 hstate_kobjs
, &hstate_attr_group
);
1793 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1800 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1801 * with node devices in node_devices[] using a parallel array. The array
1802 * index of a node device or _hstate == node id.
1803 * This is here to avoid any static dependency of the node device driver, in
1804 * the base kernel, on the hugetlb module.
1806 struct node_hstate
{
1807 struct kobject
*hugepages_kobj
;
1808 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1810 struct node_hstate node_hstates
[MAX_NUMNODES
];
1813 * A subset of global hstate attributes for node devices
1815 static struct attribute
*per_node_hstate_attrs
[] = {
1816 &nr_hugepages_attr
.attr
,
1817 &free_hugepages_attr
.attr
,
1818 &surplus_hugepages_attr
.attr
,
1822 static struct attribute_group per_node_hstate_attr_group
= {
1823 .attrs
= per_node_hstate_attrs
,
1827 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1828 * Returns node id via non-NULL nidp.
1830 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1834 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1835 struct node_hstate
*nhs
= &node_hstates
[nid
];
1837 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1838 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1850 * Unregister hstate attributes from a single node device.
1851 * No-op if no hstate attributes attached.
1853 static void hugetlb_unregister_node(struct node
*node
)
1856 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1858 if (!nhs
->hugepages_kobj
)
1859 return; /* no hstate attributes */
1861 for_each_hstate(h
) {
1862 int idx
= hstate_index(h
);
1863 if (nhs
->hstate_kobjs
[idx
]) {
1864 kobject_put(nhs
->hstate_kobjs
[idx
]);
1865 nhs
->hstate_kobjs
[idx
] = NULL
;
1869 kobject_put(nhs
->hugepages_kobj
);
1870 nhs
->hugepages_kobj
= NULL
;
1874 * hugetlb module exit: unregister hstate attributes from node devices
1877 static void hugetlb_unregister_all_nodes(void)
1882 * disable node device registrations.
1884 register_hugetlbfs_with_node(NULL
, NULL
);
1887 * remove hstate attributes from any nodes that have them.
1889 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1890 hugetlb_unregister_node(node_devices
[nid
]);
1894 * Register hstate attributes for a single node device.
1895 * No-op if attributes already registered.
1897 static void hugetlb_register_node(struct node
*node
)
1900 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1903 if (nhs
->hugepages_kobj
)
1904 return; /* already allocated */
1906 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1908 if (!nhs
->hugepages_kobj
)
1911 for_each_hstate(h
) {
1912 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1914 &per_node_hstate_attr_group
);
1916 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1917 h
->name
, node
->dev
.id
);
1918 hugetlb_unregister_node(node
);
1925 * hugetlb init time: register hstate attributes for all registered node
1926 * devices of nodes that have memory. All on-line nodes should have
1927 * registered their associated device by this time.
1929 static void hugetlb_register_all_nodes(void)
1933 for_each_node_state(nid
, N_MEMORY
) {
1934 struct node
*node
= node_devices
[nid
];
1935 if (node
->dev
.id
== nid
)
1936 hugetlb_register_node(node
);
1940 * Let the node device driver know we're here so it can
1941 * [un]register hstate attributes on node hotplug.
1943 register_hugetlbfs_with_node(hugetlb_register_node
,
1944 hugetlb_unregister_node
);
1946 #else /* !CONFIG_NUMA */
1948 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1956 static void hugetlb_unregister_all_nodes(void) { }
1958 static void hugetlb_register_all_nodes(void) { }
1962 static void __exit
hugetlb_exit(void)
1966 hugetlb_unregister_all_nodes();
1968 for_each_hstate(h
) {
1969 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1972 kobject_put(hugepages_kobj
);
1974 module_exit(hugetlb_exit
);
1976 static int __init
hugetlb_init(void)
1978 /* Some platform decide whether they support huge pages at boot
1979 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1980 * there is no such support
1982 if (HPAGE_SHIFT
== 0)
1985 if (!size_to_hstate(default_hstate_size
)) {
1986 default_hstate_size
= HPAGE_SIZE
;
1987 if (!size_to_hstate(default_hstate_size
))
1988 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1990 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1991 if (default_hstate_max_huge_pages
)
1992 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1994 hugetlb_init_hstates();
1995 gather_bootmem_prealloc();
1998 hugetlb_sysfs_init();
1999 hugetlb_register_all_nodes();
2000 hugetlb_cgroup_file_init();
2004 module_init(hugetlb_init
);
2006 /* Should be called on processing a hugepagesz=... option */
2007 void __init
hugetlb_add_hstate(unsigned order
)
2012 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2013 pr_warning("hugepagesz= specified twice, ignoring\n");
2016 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2018 h
= &hstates
[hugetlb_max_hstate
++];
2020 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2021 h
->nr_huge_pages
= 0;
2022 h
->free_huge_pages
= 0;
2023 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2024 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2025 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2026 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2027 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2028 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2029 huge_page_size(h
)/1024);
2034 static int __init
hugetlb_nrpages_setup(char *s
)
2037 static unsigned long *last_mhp
;
2040 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2041 * so this hugepages= parameter goes to the "default hstate".
2043 if (!hugetlb_max_hstate
)
2044 mhp
= &default_hstate_max_huge_pages
;
2046 mhp
= &parsed_hstate
->max_huge_pages
;
2048 if (mhp
== last_mhp
) {
2049 pr_warning("hugepages= specified twice without "
2050 "interleaving hugepagesz=, ignoring\n");
2054 if (sscanf(s
, "%lu", mhp
) <= 0)
2058 * Global state is always initialized later in hugetlb_init.
2059 * But we need to allocate >= MAX_ORDER hstates here early to still
2060 * use the bootmem allocator.
2062 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2063 hugetlb_hstate_alloc_pages(parsed_hstate
);
2069 __setup("hugepages=", hugetlb_nrpages_setup
);
2071 static int __init
hugetlb_default_setup(char *s
)
2073 default_hstate_size
= memparse(s
, &s
);
2076 __setup("default_hugepagesz=", hugetlb_default_setup
);
2078 static unsigned int cpuset_mems_nr(unsigned int *array
)
2081 unsigned int nr
= 0;
2083 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2089 #ifdef CONFIG_SYSCTL
2090 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2091 struct ctl_table
*table
, int write
,
2092 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2094 struct hstate
*h
= &default_hstate
;
2098 tmp
= h
->max_huge_pages
;
2100 if (write
&& h
->order
>= MAX_ORDER
)
2104 table
->maxlen
= sizeof(unsigned long);
2105 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2110 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2111 GFP_KERNEL
| __GFP_NORETRY
);
2112 if (!(obey_mempolicy
&&
2113 init_nodemask_of_mempolicy(nodes_allowed
))) {
2114 NODEMASK_FREE(nodes_allowed
);
2115 nodes_allowed
= &node_states
[N_MEMORY
];
2117 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2119 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2120 NODEMASK_FREE(nodes_allowed
);
2126 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2127 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2130 return hugetlb_sysctl_handler_common(false, table
, write
,
2131 buffer
, length
, ppos
);
2135 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2136 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2138 return hugetlb_sysctl_handler_common(true, table
, write
,
2139 buffer
, length
, ppos
);
2141 #endif /* CONFIG_NUMA */
2143 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2144 void __user
*buffer
,
2145 size_t *length
, loff_t
*ppos
)
2147 struct hstate
*h
= &default_hstate
;
2151 tmp
= h
->nr_overcommit_huge_pages
;
2153 if (write
&& h
->order
>= MAX_ORDER
)
2157 table
->maxlen
= sizeof(unsigned long);
2158 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2163 spin_lock(&hugetlb_lock
);
2164 h
->nr_overcommit_huge_pages
= tmp
;
2165 spin_unlock(&hugetlb_lock
);
2171 #endif /* CONFIG_SYSCTL */
2173 void hugetlb_report_meminfo(struct seq_file
*m
)
2175 struct hstate
*h
= &default_hstate
;
2177 "HugePages_Total: %5lu\n"
2178 "HugePages_Free: %5lu\n"
2179 "HugePages_Rsvd: %5lu\n"
2180 "HugePages_Surp: %5lu\n"
2181 "Hugepagesize: %8lu kB\n",
2185 h
->surplus_huge_pages
,
2186 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2189 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2191 struct hstate
*h
= &default_hstate
;
2193 "Node %d HugePages_Total: %5u\n"
2194 "Node %d HugePages_Free: %5u\n"
2195 "Node %d HugePages_Surp: %5u\n",
2196 nid
, h
->nr_huge_pages_node
[nid
],
2197 nid
, h
->free_huge_pages_node
[nid
],
2198 nid
, h
->surplus_huge_pages_node
[nid
]);
2201 void hugetlb_show_meminfo(void)
2206 for_each_node_state(nid
, N_MEMORY
)
2208 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2210 h
->nr_huge_pages_node
[nid
],
2211 h
->free_huge_pages_node
[nid
],
2212 h
->surplus_huge_pages_node
[nid
],
2213 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2216 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2217 unsigned long hugetlb_total_pages(void)
2220 unsigned long nr_total_pages
= 0;
2223 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2224 return nr_total_pages
;
2227 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2231 spin_lock(&hugetlb_lock
);
2233 * When cpuset is configured, it breaks the strict hugetlb page
2234 * reservation as the accounting is done on a global variable. Such
2235 * reservation is completely rubbish in the presence of cpuset because
2236 * the reservation is not checked against page availability for the
2237 * current cpuset. Application can still potentially OOM'ed by kernel
2238 * with lack of free htlb page in cpuset that the task is in.
2239 * Attempt to enforce strict accounting with cpuset is almost
2240 * impossible (or too ugly) because cpuset is too fluid that
2241 * task or memory node can be dynamically moved between cpusets.
2243 * The change of semantics for shared hugetlb mapping with cpuset is
2244 * undesirable. However, in order to preserve some of the semantics,
2245 * we fall back to check against current free page availability as
2246 * a best attempt and hopefully to minimize the impact of changing
2247 * semantics that cpuset has.
2250 if (gather_surplus_pages(h
, delta
) < 0)
2253 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2254 return_unused_surplus_pages(h
, delta
);
2261 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2264 spin_unlock(&hugetlb_lock
);
2268 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2270 struct resv_map
*resv
= vma_resv_map(vma
);
2273 * This new VMA should share its siblings reservation map if present.
2274 * The VMA will only ever have a valid reservation map pointer where
2275 * it is being copied for another still existing VMA. As that VMA
2276 * has a reference to the reservation map it cannot disappear until
2277 * after this open call completes. It is therefore safe to take a
2278 * new reference here without additional locking.
2281 kref_get(&resv
->refs
);
2284 static void resv_map_put(struct vm_area_struct
*vma
)
2286 struct resv_map
*resv
= vma_resv_map(vma
);
2290 kref_put(&resv
->refs
, resv_map_release
);
2293 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2295 struct hstate
*h
= hstate_vma(vma
);
2296 struct resv_map
*resv
= vma_resv_map(vma
);
2297 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2298 unsigned long reserve
;
2299 unsigned long start
;
2303 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2304 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2306 reserve
= (end
- start
) -
2307 region_count(&resv
->regions
, start
, end
);
2312 hugetlb_acct_memory(h
, -reserve
);
2313 hugepage_subpool_put_pages(spool
, reserve
);
2319 * We cannot handle pagefaults against hugetlb pages at all. They cause
2320 * handle_mm_fault() to try to instantiate regular-sized pages in the
2321 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2324 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2330 const struct vm_operations_struct hugetlb_vm_ops
= {
2331 .fault
= hugetlb_vm_op_fault
,
2332 .open
= hugetlb_vm_op_open
,
2333 .close
= hugetlb_vm_op_close
,
2336 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2342 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2343 vma
->vm_page_prot
)));
2345 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2346 vma
->vm_page_prot
));
2348 entry
= pte_mkyoung(entry
);
2349 entry
= pte_mkhuge(entry
);
2350 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2355 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2356 unsigned long address
, pte_t
*ptep
)
2360 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2361 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2362 update_mmu_cache(vma
, address
, ptep
);
2366 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2367 struct vm_area_struct
*vma
)
2369 pte_t
*src_pte
, *dst_pte
, entry
;
2370 struct page
*ptepage
;
2373 struct hstate
*h
= hstate_vma(vma
);
2374 unsigned long sz
= huge_page_size(h
);
2376 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2378 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2379 src_pte
= huge_pte_offset(src
, addr
);
2382 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2386 /* If the pagetables are shared don't copy or take references */
2387 if (dst_pte
== src_pte
)
2390 spin_lock(&dst
->page_table_lock
);
2391 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2392 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2394 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2395 entry
= huge_ptep_get(src_pte
);
2396 ptepage
= pte_page(entry
);
2398 page_dup_rmap(ptepage
);
2399 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2401 spin_unlock(&src
->page_table_lock
);
2402 spin_unlock(&dst
->page_table_lock
);
2410 static int is_hugetlb_entry_migration(pte_t pte
)
2414 if (huge_pte_none(pte
) || pte_present(pte
))
2416 swp
= pte_to_swp_entry(pte
);
2417 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2423 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2427 if (huge_pte_none(pte
) || pte_present(pte
))
2429 swp
= pte_to_swp_entry(pte
);
2430 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2436 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2437 unsigned long start
, unsigned long end
,
2438 struct page
*ref_page
)
2440 int force_flush
= 0;
2441 struct mm_struct
*mm
= vma
->vm_mm
;
2442 unsigned long address
;
2446 struct hstate
*h
= hstate_vma(vma
);
2447 unsigned long sz
= huge_page_size(h
);
2448 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2449 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2451 WARN_ON(!is_vm_hugetlb_page(vma
));
2452 BUG_ON(start
& ~huge_page_mask(h
));
2453 BUG_ON(end
& ~huge_page_mask(h
));
2455 tlb_start_vma(tlb
, vma
);
2456 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2458 spin_lock(&mm
->page_table_lock
);
2459 for (address
= start
; address
< end
; address
+= sz
) {
2460 ptep
= huge_pte_offset(mm
, address
);
2464 if (huge_pmd_unshare(mm
, &address
, ptep
))
2467 pte
= huge_ptep_get(ptep
);
2468 if (huge_pte_none(pte
))
2472 * HWPoisoned hugepage is already unmapped and dropped reference
2474 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2475 huge_pte_clear(mm
, address
, ptep
);
2479 page
= pte_page(pte
);
2481 * If a reference page is supplied, it is because a specific
2482 * page is being unmapped, not a range. Ensure the page we
2483 * are about to unmap is the actual page of interest.
2486 if (page
!= ref_page
)
2490 * Mark the VMA as having unmapped its page so that
2491 * future faults in this VMA will fail rather than
2492 * looking like data was lost
2494 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2497 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2498 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2499 if (huge_pte_dirty(pte
))
2500 set_page_dirty(page
);
2502 page_remove_rmap(page
);
2503 force_flush
= !__tlb_remove_page(tlb
, page
);
2506 /* Bail out after unmapping reference page if supplied */
2510 spin_unlock(&mm
->page_table_lock
);
2512 * mmu_gather ran out of room to batch pages, we break out of
2513 * the PTE lock to avoid doing the potential expensive TLB invalidate
2514 * and page-free while holding it.
2519 if (address
< end
&& !ref_page
)
2522 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2523 tlb_end_vma(tlb
, vma
);
2526 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2527 struct vm_area_struct
*vma
, unsigned long start
,
2528 unsigned long end
, struct page
*ref_page
)
2530 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2533 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2534 * test will fail on a vma being torn down, and not grab a page table
2535 * on its way out. We're lucky that the flag has such an appropriate
2536 * name, and can in fact be safely cleared here. We could clear it
2537 * before the __unmap_hugepage_range above, but all that's necessary
2538 * is to clear it before releasing the i_mmap_mutex. This works
2539 * because in the context this is called, the VMA is about to be
2540 * destroyed and the i_mmap_mutex is held.
2542 vma
->vm_flags
&= ~VM_MAYSHARE
;
2545 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2546 unsigned long end
, struct page
*ref_page
)
2548 struct mm_struct
*mm
;
2549 struct mmu_gather tlb
;
2553 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2554 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2555 tlb_finish_mmu(&tlb
, start
, end
);
2559 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2560 * mappping it owns the reserve page for. The intention is to unmap the page
2561 * from other VMAs and let the children be SIGKILLed if they are faulting the
2564 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2565 struct page
*page
, unsigned long address
)
2567 struct hstate
*h
= hstate_vma(vma
);
2568 struct vm_area_struct
*iter_vma
;
2569 struct address_space
*mapping
;
2573 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2574 * from page cache lookup which is in HPAGE_SIZE units.
2576 address
= address
& huge_page_mask(h
);
2577 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2579 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2582 * Take the mapping lock for the duration of the table walk. As
2583 * this mapping should be shared between all the VMAs,
2584 * __unmap_hugepage_range() is called as the lock is already held
2586 mutex_lock(&mapping
->i_mmap_mutex
);
2587 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2588 /* Do not unmap the current VMA */
2589 if (iter_vma
== vma
)
2593 * Unmap the page from other VMAs without their own reserves.
2594 * They get marked to be SIGKILLed if they fault in these
2595 * areas. This is because a future no-page fault on this VMA
2596 * could insert a zeroed page instead of the data existing
2597 * from the time of fork. This would look like data corruption
2599 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2600 unmap_hugepage_range(iter_vma
, address
,
2601 address
+ huge_page_size(h
), page
);
2603 mutex_unlock(&mapping
->i_mmap_mutex
);
2609 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2610 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2611 * cannot race with other handlers or page migration.
2612 * Keep the pte_same checks anyway to make transition from the mutex easier.
2614 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2615 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2616 struct page
*pagecache_page
)
2618 struct hstate
*h
= hstate_vma(vma
);
2619 struct page
*old_page
, *new_page
;
2620 int outside_reserve
= 0;
2621 unsigned long mmun_start
; /* For mmu_notifiers */
2622 unsigned long mmun_end
; /* For mmu_notifiers */
2624 old_page
= pte_page(pte
);
2627 /* If no-one else is actually using this page, avoid the copy
2628 * and just make the page writable */
2629 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2630 page_move_anon_rmap(old_page
, vma
, address
);
2631 set_huge_ptep_writable(vma
, address
, ptep
);
2636 * If the process that created a MAP_PRIVATE mapping is about to
2637 * perform a COW due to a shared page count, attempt to satisfy
2638 * the allocation without using the existing reserves. The pagecache
2639 * page is used to determine if the reserve at this address was
2640 * consumed or not. If reserves were used, a partial faulted mapping
2641 * at the time of fork() could consume its reserves on COW instead
2642 * of the full address range.
2644 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2645 old_page
!= pagecache_page
)
2646 outside_reserve
= 1;
2648 page_cache_get(old_page
);
2650 /* Drop page_table_lock as buddy allocator may be called */
2651 spin_unlock(&mm
->page_table_lock
);
2652 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2654 if (IS_ERR(new_page
)) {
2655 long err
= PTR_ERR(new_page
);
2656 page_cache_release(old_page
);
2659 * If a process owning a MAP_PRIVATE mapping fails to COW,
2660 * it is due to references held by a child and an insufficient
2661 * huge page pool. To guarantee the original mappers
2662 * reliability, unmap the page from child processes. The child
2663 * may get SIGKILLed if it later faults.
2665 if (outside_reserve
) {
2666 BUG_ON(huge_pte_none(pte
));
2667 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2668 BUG_ON(huge_pte_none(pte
));
2669 spin_lock(&mm
->page_table_lock
);
2670 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2671 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2672 goto retry_avoidcopy
;
2674 * race occurs while re-acquiring page_table_lock, and
2682 /* Caller expects lock to be held */
2683 spin_lock(&mm
->page_table_lock
);
2685 return VM_FAULT_OOM
;
2687 return VM_FAULT_SIGBUS
;
2691 * When the original hugepage is shared one, it does not have
2692 * anon_vma prepared.
2694 if (unlikely(anon_vma_prepare(vma
))) {
2695 page_cache_release(new_page
);
2696 page_cache_release(old_page
);
2697 /* Caller expects lock to be held */
2698 spin_lock(&mm
->page_table_lock
);
2699 return VM_FAULT_OOM
;
2702 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2703 pages_per_huge_page(h
));
2704 __SetPageUptodate(new_page
);
2706 mmun_start
= address
& huge_page_mask(h
);
2707 mmun_end
= mmun_start
+ huge_page_size(h
);
2708 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2710 * Retake the page_table_lock to check for racing updates
2711 * before the page tables are altered
2713 spin_lock(&mm
->page_table_lock
);
2714 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2715 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2716 ClearPagePrivate(new_page
);
2719 huge_ptep_clear_flush(vma
, address
, ptep
);
2720 set_huge_pte_at(mm
, address
, ptep
,
2721 make_huge_pte(vma
, new_page
, 1));
2722 page_remove_rmap(old_page
);
2723 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2724 /* Make the old page be freed below */
2725 new_page
= old_page
;
2727 spin_unlock(&mm
->page_table_lock
);
2728 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2729 page_cache_release(new_page
);
2730 page_cache_release(old_page
);
2732 /* Caller expects lock to be held */
2733 spin_lock(&mm
->page_table_lock
);
2737 /* Return the pagecache page at a given address within a VMA */
2738 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2739 struct vm_area_struct
*vma
, unsigned long address
)
2741 struct address_space
*mapping
;
2744 mapping
= vma
->vm_file
->f_mapping
;
2745 idx
= vma_hugecache_offset(h
, vma
, address
);
2747 return find_lock_page(mapping
, idx
);
2751 * Return whether there is a pagecache page to back given address within VMA.
2752 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2754 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2755 struct vm_area_struct
*vma
, unsigned long address
)
2757 struct address_space
*mapping
;
2761 mapping
= vma
->vm_file
->f_mapping
;
2762 idx
= vma_hugecache_offset(h
, vma
, address
);
2764 page
= find_get_page(mapping
, idx
);
2767 return page
!= NULL
;
2770 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2771 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2773 struct hstate
*h
= hstate_vma(vma
);
2774 int ret
= VM_FAULT_SIGBUS
;
2779 struct address_space
*mapping
;
2783 * Currently, we are forced to kill the process in the event the
2784 * original mapper has unmapped pages from the child due to a failed
2785 * COW. Warn that such a situation has occurred as it may not be obvious
2787 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2788 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2793 mapping
= vma
->vm_file
->f_mapping
;
2794 idx
= vma_hugecache_offset(h
, vma
, address
);
2797 * Use page lock to guard against racing truncation
2798 * before we get page_table_lock.
2801 page
= find_lock_page(mapping
, idx
);
2803 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2806 page
= alloc_huge_page(vma
, address
, 0);
2808 ret
= PTR_ERR(page
);
2812 ret
= VM_FAULT_SIGBUS
;
2815 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2816 __SetPageUptodate(page
);
2818 if (vma
->vm_flags
& VM_MAYSHARE
) {
2820 struct inode
*inode
= mapping
->host
;
2822 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2829 ClearPagePrivate(page
);
2831 spin_lock(&inode
->i_lock
);
2832 inode
->i_blocks
+= blocks_per_huge_page(h
);
2833 spin_unlock(&inode
->i_lock
);
2836 if (unlikely(anon_vma_prepare(vma
))) {
2838 goto backout_unlocked
;
2844 * If memory error occurs between mmap() and fault, some process
2845 * don't have hwpoisoned swap entry for errored virtual address.
2846 * So we need to block hugepage fault by PG_hwpoison bit check.
2848 if (unlikely(PageHWPoison(page
))) {
2849 ret
= VM_FAULT_HWPOISON
|
2850 VM_FAULT_SET_HINDEX(hstate_index(h
));
2851 goto backout_unlocked
;
2856 * If we are going to COW a private mapping later, we examine the
2857 * pending reservations for this page now. This will ensure that
2858 * any allocations necessary to record that reservation occur outside
2861 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2862 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2864 goto backout_unlocked
;
2867 spin_lock(&mm
->page_table_lock
);
2868 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2873 if (!huge_pte_none(huge_ptep_get(ptep
)))
2877 ClearPagePrivate(page
);
2878 hugepage_add_new_anon_rmap(page
, vma
, address
);
2881 page_dup_rmap(page
);
2882 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2883 && (vma
->vm_flags
& VM_SHARED
)));
2884 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2886 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2887 /* Optimization, do the COW without a second fault */
2888 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2891 spin_unlock(&mm
->page_table_lock
);
2897 spin_unlock(&mm
->page_table_lock
);
2904 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2905 unsigned long address
, unsigned int flags
)
2910 struct page
*page
= NULL
;
2911 struct page
*pagecache_page
= NULL
;
2912 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2913 struct hstate
*h
= hstate_vma(vma
);
2915 address
&= huge_page_mask(h
);
2917 ptep
= huge_pte_offset(mm
, address
);
2919 entry
= huge_ptep_get(ptep
);
2920 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2921 migration_entry_wait_huge(mm
, ptep
);
2923 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2924 return VM_FAULT_HWPOISON_LARGE
|
2925 VM_FAULT_SET_HINDEX(hstate_index(h
));
2928 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2930 return VM_FAULT_OOM
;
2933 * Serialize hugepage allocation and instantiation, so that we don't
2934 * get spurious allocation failures if two CPUs race to instantiate
2935 * the same page in the page cache.
2937 mutex_lock(&hugetlb_instantiation_mutex
);
2938 entry
= huge_ptep_get(ptep
);
2939 if (huge_pte_none(entry
)) {
2940 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2947 * If we are going to COW the mapping later, we examine the pending
2948 * reservations for this page now. This will ensure that any
2949 * allocations necessary to record that reservation occur outside the
2950 * spinlock. For private mappings, we also lookup the pagecache
2951 * page now as it is used to determine if a reservation has been
2954 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2955 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2960 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2961 pagecache_page
= hugetlbfs_pagecache_page(h
,
2966 * hugetlb_cow() requires page locks of pte_page(entry) and
2967 * pagecache_page, so here we need take the former one
2968 * when page != pagecache_page or !pagecache_page.
2969 * Note that locking order is always pagecache_page -> page,
2970 * so no worry about deadlock.
2972 page
= pte_page(entry
);
2974 if (page
!= pagecache_page
)
2977 spin_lock(&mm
->page_table_lock
);
2978 /* Check for a racing update before calling hugetlb_cow */
2979 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2980 goto out_page_table_lock
;
2983 if (flags
& FAULT_FLAG_WRITE
) {
2984 if (!huge_pte_write(entry
)) {
2985 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2987 goto out_page_table_lock
;
2989 entry
= huge_pte_mkdirty(entry
);
2991 entry
= pte_mkyoung(entry
);
2992 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2993 flags
& FAULT_FLAG_WRITE
))
2994 update_mmu_cache(vma
, address
, ptep
);
2996 out_page_table_lock
:
2997 spin_unlock(&mm
->page_table_lock
);
2999 if (pagecache_page
) {
3000 unlock_page(pagecache_page
);
3001 put_page(pagecache_page
);
3003 if (page
!= pagecache_page
)
3008 mutex_unlock(&hugetlb_instantiation_mutex
);
3013 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3014 struct page
**pages
, struct vm_area_struct
**vmas
,
3015 unsigned long *position
, unsigned long *nr_pages
,
3016 long i
, unsigned int flags
)
3018 unsigned long pfn_offset
;
3019 unsigned long vaddr
= *position
;
3020 unsigned long remainder
= *nr_pages
;
3021 struct hstate
*h
= hstate_vma(vma
);
3023 spin_lock(&mm
->page_table_lock
);
3024 while (vaddr
< vma
->vm_end
&& remainder
) {
3030 * Some archs (sparc64, sh*) have multiple pte_ts to
3031 * each hugepage. We have to make sure we get the
3032 * first, for the page indexing below to work.
3034 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3035 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3038 * When coredumping, it suits get_dump_page if we just return
3039 * an error where there's an empty slot with no huge pagecache
3040 * to back it. This way, we avoid allocating a hugepage, and
3041 * the sparse dumpfile avoids allocating disk blocks, but its
3042 * huge holes still show up with zeroes where they need to be.
3044 if (absent
&& (flags
& FOLL_DUMP
) &&
3045 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3051 * We need call hugetlb_fault for both hugepages under migration
3052 * (in which case hugetlb_fault waits for the migration,) and
3053 * hwpoisoned hugepages (in which case we need to prevent the
3054 * caller from accessing to them.) In order to do this, we use
3055 * here is_swap_pte instead of is_hugetlb_entry_migration and
3056 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3057 * both cases, and because we can't follow correct pages
3058 * directly from any kind of swap entries.
3060 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3061 ((flags
& FOLL_WRITE
) &&
3062 !huge_pte_write(huge_ptep_get(pte
)))) {
3065 spin_unlock(&mm
->page_table_lock
);
3066 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3067 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3068 spin_lock(&mm
->page_table_lock
);
3069 if (!(ret
& VM_FAULT_ERROR
))
3076 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3077 page
= pte_page(huge_ptep_get(pte
));
3080 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3091 if (vaddr
< vma
->vm_end
&& remainder
&&
3092 pfn_offset
< pages_per_huge_page(h
)) {
3094 * We use pfn_offset to avoid touching the pageframes
3095 * of this compound page.
3100 spin_unlock(&mm
->page_table_lock
);
3101 *nr_pages
= remainder
;
3104 return i
? i
: -EFAULT
;
3107 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3108 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3110 struct mm_struct
*mm
= vma
->vm_mm
;
3111 unsigned long start
= address
;
3114 struct hstate
*h
= hstate_vma(vma
);
3115 unsigned long pages
= 0;
3117 BUG_ON(address
>= end
);
3118 flush_cache_range(vma
, address
, end
);
3120 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3121 spin_lock(&mm
->page_table_lock
);
3122 for (; address
< end
; address
+= huge_page_size(h
)) {
3123 ptep
= huge_pte_offset(mm
, address
);
3126 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3130 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3131 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3132 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3133 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3134 set_huge_pte_at(mm
, address
, ptep
, pte
);
3138 spin_unlock(&mm
->page_table_lock
);
3140 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3141 * may have cleared our pud entry and done put_page on the page table:
3142 * once we release i_mmap_mutex, another task can do the final put_page
3143 * and that page table be reused and filled with junk.
3145 flush_tlb_range(vma
, start
, end
);
3146 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3148 return pages
<< h
->order
;
3151 int hugetlb_reserve_pages(struct inode
*inode
,
3153 struct vm_area_struct
*vma
,
3154 vm_flags_t vm_flags
)
3157 struct hstate
*h
= hstate_inode(inode
);
3158 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3161 * Only apply hugepage reservation if asked. At fault time, an
3162 * attempt will be made for VM_NORESERVE to allocate a page
3163 * without using reserves
3165 if (vm_flags
& VM_NORESERVE
)
3169 * Shared mappings base their reservation on the number of pages that
3170 * are already allocated on behalf of the file. Private mappings need
3171 * to reserve the full area even if read-only as mprotect() may be
3172 * called to make the mapping read-write. Assume !vma is a shm mapping
3174 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3175 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3177 struct resv_map
*resv_map
= resv_map_alloc();
3183 set_vma_resv_map(vma
, resv_map
);
3184 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3192 /* There must be enough pages in the subpool for the mapping */
3193 if (hugepage_subpool_get_pages(spool
, chg
)) {
3199 * Check enough hugepages are available for the reservation.
3200 * Hand the pages back to the subpool if there are not
3202 ret
= hugetlb_acct_memory(h
, chg
);
3204 hugepage_subpool_put_pages(spool
, chg
);
3209 * Account for the reservations made. Shared mappings record regions
3210 * that have reservations as they are shared by multiple VMAs.
3211 * When the last VMA disappears, the region map says how much
3212 * the reservation was and the page cache tells how much of
3213 * the reservation was consumed. Private mappings are per-VMA and
3214 * only the consumed reservations are tracked. When the VMA
3215 * disappears, the original reservation is the VMA size and the
3216 * consumed reservations are stored in the map. Hence, nothing
3217 * else has to be done for private mappings here
3219 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3220 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3228 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3230 struct hstate
*h
= hstate_inode(inode
);
3231 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3232 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3234 spin_lock(&inode
->i_lock
);
3235 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3236 spin_unlock(&inode
->i_lock
);
3238 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3239 hugetlb_acct_memory(h
, -(chg
- freed
));
3242 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3243 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3244 struct vm_area_struct
*vma
,
3245 unsigned long addr
, pgoff_t idx
)
3247 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3249 unsigned long sbase
= saddr
& PUD_MASK
;
3250 unsigned long s_end
= sbase
+ PUD_SIZE
;
3252 /* Allow segments to share if only one is marked locked */
3253 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3254 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3257 * match the virtual addresses, permission and the alignment of the
3260 if (pmd_index(addr
) != pmd_index(saddr
) ||
3261 vm_flags
!= svm_flags
||
3262 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3268 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3270 unsigned long base
= addr
& PUD_MASK
;
3271 unsigned long end
= base
+ PUD_SIZE
;
3274 * check on proper vm_flags and page table alignment
3276 if (vma
->vm_flags
& VM_MAYSHARE
&&
3277 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3283 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3284 * and returns the corresponding pte. While this is not necessary for the
3285 * !shared pmd case because we can allocate the pmd later as well, it makes the
3286 * code much cleaner. pmd allocation is essential for the shared case because
3287 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3288 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3289 * bad pmd for sharing.
3291 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3293 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3294 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3295 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3297 struct vm_area_struct
*svma
;
3298 unsigned long saddr
;
3302 if (!vma_shareable(vma
, addr
))
3303 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3305 mutex_lock(&mapping
->i_mmap_mutex
);
3306 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3310 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3312 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3314 get_page(virt_to_page(spte
));
3323 spin_lock(&mm
->page_table_lock
);
3325 pud_populate(mm
, pud
,
3326 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3328 put_page(virt_to_page(spte
));
3329 spin_unlock(&mm
->page_table_lock
);
3331 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3332 mutex_unlock(&mapping
->i_mmap_mutex
);
3337 * unmap huge page backed by shared pte.
3339 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3340 * indicated by page_count > 1, unmap is achieved by clearing pud and
3341 * decrementing the ref count. If count == 1, the pte page is not shared.
3343 * called with vma->vm_mm->page_table_lock held.
3345 * returns: 1 successfully unmapped a shared pte page
3346 * 0 the underlying pte page is not shared, or it is the last user
3348 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3350 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3351 pud_t
*pud
= pud_offset(pgd
, *addr
);
3353 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3354 if (page_count(virt_to_page(ptep
)) == 1)
3358 put_page(virt_to_page(ptep
));
3359 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3362 #define want_pmd_share() (1)
3363 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3364 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3368 #define want_pmd_share() (0)
3369 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3371 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3372 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3373 unsigned long addr
, unsigned long sz
)
3379 pgd
= pgd_offset(mm
, addr
);
3380 pud
= pud_alloc(mm
, pgd
, addr
);
3382 if (sz
== PUD_SIZE
) {
3385 BUG_ON(sz
!= PMD_SIZE
);
3386 if (want_pmd_share() && pud_none(*pud
))
3387 pte
= huge_pmd_share(mm
, addr
, pud
);
3389 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3392 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3397 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3403 pgd
= pgd_offset(mm
, addr
);
3404 if (pgd_present(*pgd
)) {
3405 pud
= pud_offset(pgd
, addr
);
3406 if (pud_present(*pud
)) {
3408 return (pte_t
*)pud
;
3409 pmd
= pmd_offset(pud
, addr
);
3412 return (pte_t
*) pmd
;
3416 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3417 pmd_t
*pmd
, int write
)
3421 page
= pte_page(*(pte_t
*)pmd
);
3423 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3428 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3429 pud_t
*pud
, int write
)
3433 page
= pte_page(*(pte_t
*)pud
);
3435 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3439 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3441 /* Can be overriden by architectures */
3442 __attribute__((weak
)) struct page
*
3443 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3444 pud_t
*pud
, int write
)
3450 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3452 #ifdef CONFIG_MEMORY_FAILURE
3454 /* Should be called in hugetlb_lock */
3455 static int is_hugepage_on_freelist(struct page
*hpage
)
3459 struct hstate
*h
= page_hstate(hpage
);
3460 int nid
= page_to_nid(hpage
);
3462 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3469 * This function is called from memory failure code.
3470 * Assume the caller holds page lock of the head page.
3472 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3474 struct hstate
*h
= page_hstate(hpage
);
3475 int nid
= page_to_nid(hpage
);
3478 spin_lock(&hugetlb_lock
);
3479 if (is_hugepage_on_freelist(hpage
)) {
3481 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3482 * but dangling hpage->lru can trigger list-debug warnings
3483 * (this happens when we call unpoison_memory() on it),
3484 * so let it point to itself with list_del_init().
3486 list_del_init(&hpage
->lru
);
3487 set_page_refcounted(hpage
);
3488 h
->free_huge_pages
--;
3489 h
->free_huge_pages_node
[nid
]--;
3492 spin_unlock(&hugetlb_lock
);
3497 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3499 VM_BUG_ON(!PageHead(page
));
3500 if (!get_page_unless_zero(page
))
3502 spin_lock(&hugetlb_lock
);
3503 list_move_tail(&page
->lru
, list
);
3504 spin_unlock(&hugetlb_lock
);
3508 void putback_active_hugepage(struct page
*page
)
3510 VM_BUG_ON(!PageHead(page
));
3511 spin_lock(&hugetlb_lock
);
3512 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3513 spin_unlock(&hugetlb_lock
);
3517 bool is_hugepage_active(struct page
*page
)
3519 VM_BUG_ON(!PageHuge(page
));
3521 * This function can be called for a tail page because the caller,
3522 * scan_movable_pages, scans through a given pfn-range which typically
3523 * covers one memory block. In systems using gigantic hugepage (1GB
3524 * for x86_64,) a hugepage is larger than a memory block, and we don't
3525 * support migrating such large hugepages for now, so return false
3526 * when called for tail pages.
3531 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3532 * so we should return false for them.
3534 if (unlikely(PageHWPoison(page
)))
3536 return page_count(page
) > 0;