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 resv_map
*resv
, long f
, long t
)
156 struct list_head
*head
= &resv
->regions
;
157 struct file_region
*rg
, *nrg
, *trg
;
159 /* Locate the region we are either in or before. */
160 list_for_each_entry(rg
, head
, link
)
164 /* Round our left edge to the current segment if it encloses us. */
168 /* Check for and consume any regions we now overlap with. */
170 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
171 if (&rg
->link
== head
)
176 /* If this area reaches higher then extend our area to
177 * include it completely. If this is not the first area
178 * which we intend to reuse, free it. */
191 static long region_chg(struct resv_map
*resv
, long f
, long t
)
193 struct list_head
*head
= &resv
->regions
;
194 struct file_region
*rg
, *nrg
;
197 /* Locate the region we are before or in. */
198 list_for_each_entry(rg
, head
, link
)
202 /* If we are below the current region then a new region is required.
203 * Subtle, allocate a new region at the position but make it zero
204 * size such that we can guarantee to record the reservation. */
205 if (&rg
->link
== head
|| t
< rg
->from
) {
206 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
211 INIT_LIST_HEAD(&nrg
->link
);
212 list_add(&nrg
->link
, rg
->link
.prev
);
217 /* Round our left edge to the current segment if it encloses us. */
222 /* Check for and consume any regions we now overlap with. */
223 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
224 if (&rg
->link
== head
)
229 /* We overlap with this area, if it extends further than
230 * us then we must extend ourselves. Account for its
231 * existing reservation. */
236 chg
-= rg
->to
- rg
->from
;
241 static long region_truncate(struct resv_map
*resv
, long end
)
243 struct list_head
*head
= &resv
->regions
;
244 struct file_region
*rg
, *trg
;
247 /* Locate the region we are either in or before. */
248 list_for_each_entry(rg
, head
, link
)
251 if (&rg
->link
== head
)
254 /* If we are in the middle of a region then adjust it. */
255 if (end
> rg
->from
) {
258 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
261 /* Drop any remaining regions. */
262 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
263 if (&rg
->link
== head
)
265 chg
+= rg
->to
- rg
->from
;
272 static long region_count(struct resv_map
*resv
, long f
, long t
)
274 struct list_head
*head
= &resv
->regions
;
275 struct file_region
*rg
;
278 /* Locate each segment we overlap with, and count that overlap. */
279 list_for_each_entry(rg
, head
, link
) {
288 seg_from
= max(rg
->from
, f
);
289 seg_to
= min(rg
->to
, t
);
291 chg
+= seg_to
- seg_from
;
298 * Convert the address within this vma to the page offset within
299 * the mapping, in pagecache page units; huge pages here.
301 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
302 struct vm_area_struct
*vma
, unsigned long address
)
304 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
305 (vma
->vm_pgoff
>> huge_page_order(h
));
308 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
309 unsigned long address
)
311 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
315 * Return the size of the pages allocated when backing a VMA. In the majority
316 * cases this will be same size as used by the page table entries.
318 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
320 struct hstate
*hstate
;
322 if (!is_vm_hugetlb_page(vma
))
325 hstate
= hstate_vma(vma
);
327 return 1UL << huge_page_shift(hstate
);
329 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
332 * Return the page size being used by the MMU to back a VMA. In the majority
333 * of cases, the page size used by the kernel matches the MMU size. On
334 * architectures where it differs, an architecture-specific version of this
335 * function is required.
337 #ifndef vma_mmu_pagesize
338 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
340 return vma_kernel_pagesize(vma
);
345 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
346 * bits of the reservation map pointer, which are always clear due to
349 #define HPAGE_RESV_OWNER (1UL << 0)
350 #define HPAGE_RESV_UNMAPPED (1UL << 1)
351 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
354 * These helpers are used to track how many pages are reserved for
355 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
356 * is guaranteed to have their future faults succeed.
358 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
359 * the reserve counters are updated with the hugetlb_lock held. It is safe
360 * to reset the VMA at fork() time as it is not in use yet and there is no
361 * chance of the global counters getting corrupted as a result of the values.
363 * The private mapping reservation is represented in a subtly different
364 * manner to a shared mapping. A shared mapping has a region map associated
365 * with the underlying file, this region map represents the backing file
366 * pages which have ever had a reservation assigned which this persists even
367 * after the page is instantiated. A private mapping has a region map
368 * associated with the original mmap which is attached to all VMAs which
369 * reference it, this region map represents those offsets which have consumed
370 * reservation ie. where pages have been instantiated.
372 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
374 return (unsigned long)vma
->vm_private_data
;
377 static void set_vma_private_data(struct vm_area_struct
*vma
,
380 vma
->vm_private_data
= (void *)value
;
383 struct resv_map
*resv_map_alloc(void)
385 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
389 kref_init(&resv_map
->refs
);
390 INIT_LIST_HEAD(&resv_map
->regions
);
395 void resv_map_release(struct kref
*ref
)
397 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
399 /* Clear out any active regions before we release the map. */
400 region_truncate(resv_map
, 0);
404 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
406 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
407 if (!(vma
->vm_flags
& VM_MAYSHARE
))
408 return (struct resv_map
*)(get_vma_private_data(vma
) &
413 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
415 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
416 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
418 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
419 HPAGE_RESV_MASK
) | (unsigned long)map
);
422 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
424 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
425 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
427 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
430 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
432 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
434 return (get_vma_private_data(vma
) & flag
) != 0;
437 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
438 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
440 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
441 if (!(vma
->vm_flags
& VM_MAYSHARE
))
442 vma
->vm_private_data
= (void *)0;
445 /* Returns true if the VMA has associated reserve pages */
446 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
448 if (vma
->vm_flags
& VM_NORESERVE
) {
450 * This address is already reserved by other process(chg == 0),
451 * so, we should decrement reserved count. Without decrementing,
452 * reserve count remains after releasing inode, because this
453 * allocated page will go into page cache and is regarded as
454 * coming from reserved pool in releasing step. Currently, we
455 * don't have any other solution to deal with this situation
456 * properly, so add work-around here.
458 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
464 /* Shared mappings always use reserves */
465 if (vma
->vm_flags
& VM_MAYSHARE
)
469 * Only the process that called mmap() has reserves for
472 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
478 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
480 int nid
= page_to_nid(page
);
481 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
482 h
->free_huge_pages
++;
483 h
->free_huge_pages_node
[nid
]++;
486 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
490 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
491 if (!is_migrate_isolate_page(page
))
494 * if 'non-isolated free hugepage' not found on the list,
495 * the allocation fails.
497 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
499 list_move(&page
->lru
, &h
->hugepage_activelist
);
500 set_page_refcounted(page
);
501 h
->free_huge_pages
--;
502 h
->free_huge_pages_node
[nid
]--;
506 /* Movability of hugepages depends on migration support. */
507 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
509 if (hugepages_treat_as_movable
|| hugepage_migration_support(h
))
510 return GFP_HIGHUSER_MOVABLE
;
515 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
516 struct vm_area_struct
*vma
,
517 unsigned long address
, int avoid_reserve
,
520 struct page
*page
= NULL
;
521 struct mempolicy
*mpol
;
522 nodemask_t
*nodemask
;
523 struct zonelist
*zonelist
;
526 unsigned int cpuset_mems_cookie
;
529 * A child process with MAP_PRIVATE mappings created by their parent
530 * have no page reserves. This check ensures that reservations are
531 * not "stolen". The child may still get SIGKILLed
533 if (!vma_has_reserves(vma
, chg
) &&
534 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
537 /* If reserves cannot be used, ensure enough pages are in the pool */
538 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
542 cpuset_mems_cookie
= read_mems_allowed_begin();
543 zonelist
= huge_zonelist(vma
, address
,
544 htlb_alloc_mask(h
), &mpol
, &nodemask
);
546 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
547 MAX_NR_ZONES
- 1, nodemask
) {
548 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask(h
))) {
549 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
553 if (!vma_has_reserves(vma
, chg
))
556 SetPagePrivate(page
);
557 h
->resv_huge_pages
--;
564 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
572 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
576 VM_BUG_ON(h
->order
>= MAX_ORDER
);
579 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
580 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
581 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
582 1 << PG_referenced
| 1 << PG_dirty
|
583 1 << PG_active
| 1 << PG_reserved
|
584 1 << PG_private
| 1 << PG_writeback
);
586 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
587 set_compound_page_dtor(page
, NULL
);
588 set_page_refcounted(page
);
589 arch_release_hugepage(page
);
590 __free_pages(page
, huge_page_order(h
));
593 struct hstate
*size_to_hstate(unsigned long size
)
598 if (huge_page_size(h
) == size
)
604 static void free_huge_page(struct page
*page
)
607 * Can't pass hstate in here because it is called from the
608 * compound page destructor.
610 struct hstate
*h
= page_hstate(page
);
611 int nid
= page_to_nid(page
);
612 struct hugepage_subpool
*spool
=
613 (struct hugepage_subpool
*)page_private(page
);
614 bool restore_reserve
;
616 set_page_private(page
, 0);
617 page
->mapping
= NULL
;
618 BUG_ON(page_count(page
));
619 BUG_ON(page_mapcount(page
));
620 restore_reserve
= PagePrivate(page
);
621 ClearPagePrivate(page
);
623 spin_lock(&hugetlb_lock
);
624 hugetlb_cgroup_uncharge_page(hstate_index(h
),
625 pages_per_huge_page(h
), page
);
627 h
->resv_huge_pages
++;
629 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
630 /* remove the page from active list */
631 list_del(&page
->lru
);
632 update_and_free_page(h
, page
);
633 h
->surplus_huge_pages
--;
634 h
->surplus_huge_pages_node
[nid
]--;
636 arch_clear_hugepage_flags(page
);
637 enqueue_huge_page(h
, page
);
639 spin_unlock(&hugetlb_lock
);
640 hugepage_subpool_put_pages(spool
, 1);
643 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
645 INIT_LIST_HEAD(&page
->lru
);
646 set_compound_page_dtor(page
, free_huge_page
);
647 spin_lock(&hugetlb_lock
);
648 set_hugetlb_cgroup(page
, NULL
);
650 h
->nr_huge_pages_node
[nid
]++;
651 spin_unlock(&hugetlb_lock
);
652 put_page(page
); /* free it into the hugepage allocator */
655 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
658 int nr_pages
= 1 << order
;
659 struct page
*p
= page
+ 1;
661 /* we rely on prep_new_huge_page to set the destructor */
662 set_compound_order(page
, order
);
664 __ClearPageReserved(page
);
665 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
668 * For gigantic hugepages allocated through bootmem at
669 * boot, it's safer to be consistent with the not-gigantic
670 * hugepages and clear the PG_reserved bit from all tail pages
671 * too. Otherwse drivers using get_user_pages() to access tail
672 * pages may get the reference counting wrong if they see
673 * PG_reserved set on a tail page (despite the head page not
674 * having PG_reserved set). Enforcing this consistency between
675 * head and tail pages allows drivers to optimize away a check
676 * on the head page when they need know if put_page() is needed
677 * after get_user_pages().
679 __ClearPageReserved(p
);
680 set_page_count(p
, 0);
681 p
->first_page
= page
;
686 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
687 * transparent huge pages. See the PageTransHuge() documentation for more
690 int PageHuge(struct page
*page
)
692 if (!PageCompound(page
))
695 page
= compound_head(page
);
696 return get_compound_page_dtor(page
) == free_huge_page
;
698 EXPORT_SYMBOL_GPL(PageHuge
);
701 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
702 * normal or transparent huge pages.
704 int PageHeadHuge(struct page
*page_head
)
706 if (!PageHead(page_head
))
709 return get_compound_page_dtor(page_head
) == free_huge_page
;
712 pgoff_t
__basepage_index(struct page
*page
)
714 struct page
*page_head
= compound_head(page
);
715 pgoff_t index
= page_index(page_head
);
716 unsigned long compound_idx
;
718 if (!PageHuge(page_head
))
719 return page_index(page
);
721 if (compound_order(page_head
) >= MAX_ORDER
)
722 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
724 compound_idx
= page
- page_head
;
726 return (index
<< compound_order(page_head
)) + compound_idx
;
729 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
733 if (h
->order
>= MAX_ORDER
)
736 page
= alloc_pages_exact_node(nid
,
737 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
738 __GFP_REPEAT
|__GFP_NOWARN
,
741 if (arch_prepare_hugepage(page
)) {
742 __free_pages(page
, huge_page_order(h
));
745 prep_new_huge_page(h
, page
, nid
);
752 * common helper functions for hstate_next_node_to_{alloc|free}.
753 * We may have allocated or freed a huge page based on a different
754 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
755 * be outside of *nodes_allowed. Ensure that we use an allowed
756 * node for alloc or free.
758 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
760 nid
= next_node(nid
, *nodes_allowed
);
761 if (nid
== MAX_NUMNODES
)
762 nid
= first_node(*nodes_allowed
);
763 VM_BUG_ON(nid
>= MAX_NUMNODES
);
768 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
770 if (!node_isset(nid
, *nodes_allowed
))
771 nid
= next_node_allowed(nid
, nodes_allowed
);
776 * returns the previously saved node ["this node"] from which to
777 * allocate a persistent huge page for the pool and advance the
778 * next node from which to allocate, handling wrap at end of node
781 static int hstate_next_node_to_alloc(struct hstate
*h
,
782 nodemask_t
*nodes_allowed
)
786 VM_BUG_ON(!nodes_allowed
);
788 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
789 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
795 * helper for free_pool_huge_page() - return the previously saved
796 * node ["this node"] from which to free a huge page. Advance the
797 * next node id whether or not we find a free huge page to free so
798 * that the next attempt to free addresses the next node.
800 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
804 VM_BUG_ON(!nodes_allowed
);
806 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
807 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
812 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
813 for (nr_nodes = nodes_weight(*mask); \
815 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
818 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
819 for (nr_nodes = nodes_weight(*mask); \
821 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
824 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
830 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
831 page
= alloc_fresh_huge_page_node(h
, node
);
839 count_vm_event(HTLB_BUDDY_PGALLOC
);
841 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
847 * Free huge page from pool from next node to free.
848 * Attempt to keep persistent huge pages more or less
849 * balanced over allowed nodes.
850 * Called with hugetlb_lock locked.
852 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
858 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
860 * If we're returning unused surplus pages, only examine
861 * nodes with surplus pages.
863 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
864 !list_empty(&h
->hugepage_freelists
[node
])) {
866 list_entry(h
->hugepage_freelists
[node
].next
,
868 list_del(&page
->lru
);
869 h
->free_huge_pages
--;
870 h
->free_huge_pages_node
[node
]--;
872 h
->surplus_huge_pages
--;
873 h
->surplus_huge_pages_node
[node
]--;
875 update_and_free_page(h
, page
);
885 * Dissolve a given free hugepage into free buddy pages. This function does
886 * nothing for in-use (including surplus) hugepages.
888 static void dissolve_free_huge_page(struct page
*page
)
890 spin_lock(&hugetlb_lock
);
891 if (PageHuge(page
) && !page_count(page
)) {
892 struct hstate
*h
= page_hstate(page
);
893 int nid
= page_to_nid(page
);
894 list_del(&page
->lru
);
895 h
->free_huge_pages
--;
896 h
->free_huge_pages_node
[nid
]--;
897 update_and_free_page(h
, page
);
899 spin_unlock(&hugetlb_lock
);
903 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
904 * make specified memory blocks removable from the system.
905 * Note that start_pfn should aligned with (minimum) hugepage size.
907 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
909 unsigned int order
= 8 * sizeof(void *);
913 /* Set scan step to minimum hugepage size */
915 if (order
> huge_page_order(h
))
916 order
= huge_page_order(h
);
917 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
918 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
919 dissolve_free_huge_page(pfn_to_page(pfn
));
922 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
927 if (h
->order
>= MAX_ORDER
)
931 * Assume we will successfully allocate the surplus page to
932 * prevent racing processes from causing the surplus to exceed
935 * This however introduces a different race, where a process B
936 * tries to grow the static hugepage pool while alloc_pages() is
937 * called by process A. B will only examine the per-node
938 * counters in determining if surplus huge pages can be
939 * converted to normal huge pages in adjust_pool_surplus(). A
940 * won't be able to increment the per-node counter, until the
941 * lock is dropped by B, but B doesn't drop hugetlb_lock until
942 * no more huge pages can be converted from surplus to normal
943 * state (and doesn't try to convert again). Thus, we have a
944 * case where a surplus huge page exists, the pool is grown, and
945 * the surplus huge page still exists after, even though it
946 * should just have been converted to a normal huge page. This
947 * does not leak memory, though, as the hugepage will be freed
948 * once it is out of use. It also does not allow the counters to
949 * go out of whack in adjust_pool_surplus() as we don't modify
950 * the node values until we've gotten the hugepage and only the
951 * per-node value is checked there.
953 spin_lock(&hugetlb_lock
);
954 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
955 spin_unlock(&hugetlb_lock
);
959 h
->surplus_huge_pages
++;
961 spin_unlock(&hugetlb_lock
);
963 if (nid
== NUMA_NO_NODE
)
964 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
965 __GFP_REPEAT
|__GFP_NOWARN
,
968 page
= alloc_pages_exact_node(nid
,
969 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
970 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
972 if (page
&& arch_prepare_hugepage(page
)) {
973 __free_pages(page
, huge_page_order(h
));
977 spin_lock(&hugetlb_lock
);
979 INIT_LIST_HEAD(&page
->lru
);
980 r_nid
= page_to_nid(page
);
981 set_compound_page_dtor(page
, free_huge_page
);
982 set_hugetlb_cgroup(page
, NULL
);
984 * We incremented the global counters already
986 h
->nr_huge_pages_node
[r_nid
]++;
987 h
->surplus_huge_pages_node
[r_nid
]++;
988 __count_vm_event(HTLB_BUDDY_PGALLOC
);
991 h
->surplus_huge_pages
--;
992 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
994 spin_unlock(&hugetlb_lock
);
1000 * This allocation function is useful in the context where vma is irrelevant.
1001 * E.g. soft-offlining uses this function because it only cares physical
1002 * address of error page.
1004 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1006 struct page
*page
= NULL
;
1008 spin_lock(&hugetlb_lock
);
1009 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1010 page
= dequeue_huge_page_node(h
, nid
);
1011 spin_unlock(&hugetlb_lock
);
1014 page
= alloc_buddy_huge_page(h
, nid
);
1020 * Increase the hugetlb pool such that it can accommodate a reservation
1023 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1025 struct list_head surplus_list
;
1026 struct page
*page
, *tmp
;
1028 int needed
, allocated
;
1029 bool alloc_ok
= true;
1031 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1033 h
->resv_huge_pages
+= delta
;
1038 INIT_LIST_HEAD(&surplus_list
);
1042 spin_unlock(&hugetlb_lock
);
1043 for (i
= 0; i
< needed
; i
++) {
1044 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1049 list_add(&page
->lru
, &surplus_list
);
1054 * After retaking hugetlb_lock, we need to recalculate 'needed'
1055 * because either resv_huge_pages or free_huge_pages may have changed.
1057 spin_lock(&hugetlb_lock
);
1058 needed
= (h
->resv_huge_pages
+ delta
) -
1059 (h
->free_huge_pages
+ allocated
);
1064 * We were not able to allocate enough pages to
1065 * satisfy the entire reservation so we free what
1066 * we've allocated so far.
1071 * The surplus_list now contains _at_least_ the number of extra pages
1072 * needed to accommodate the reservation. Add the appropriate number
1073 * of pages to the hugetlb pool and free the extras back to the buddy
1074 * allocator. Commit the entire reservation here to prevent another
1075 * process from stealing the pages as they are added to the pool but
1076 * before they are reserved.
1078 needed
+= allocated
;
1079 h
->resv_huge_pages
+= delta
;
1082 /* Free the needed pages to the hugetlb pool */
1083 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1087 * This page is now managed by the hugetlb allocator and has
1088 * no users -- drop the buddy allocator's reference.
1090 put_page_testzero(page
);
1091 VM_BUG_ON_PAGE(page_count(page
), page
);
1092 enqueue_huge_page(h
, page
);
1095 spin_unlock(&hugetlb_lock
);
1097 /* Free unnecessary surplus pages to the buddy allocator */
1098 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1100 spin_lock(&hugetlb_lock
);
1106 * When releasing a hugetlb pool reservation, any surplus pages that were
1107 * allocated to satisfy the reservation must be explicitly freed if they were
1109 * Called with hugetlb_lock held.
1111 static void return_unused_surplus_pages(struct hstate
*h
,
1112 unsigned long unused_resv_pages
)
1114 unsigned long nr_pages
;
1116 /* Uncommit the reservation */
1117 h
->resv_huge_pages
-= unused_resv_pages
;
1119 /* Cannot return gigantic pages currently */
1120 if (h
->order
>= MAX_ORDER
)
1123 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1126 * We want to release as many surplus pages as possible, spread
1127 * evenly across all nodes with memory. Iterate across these nodes
1128 * until we can no longer free unreserved surplus pages. This occurs
1129 * when the nodes with surplus pages have no free pages.
1130 * free_pool_huge_page() will balance the the freed pages across the
1131 * on-line nodes with memory and will handle the hstate accounting.
1133 while (nr_pages
--) {
1134 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1140 * Determine if the huge page at addr within the vma has an associated
1141 * reservation. Where it does not we will need to logically increase
1142 * reservation and actually increase subpool usage before an allocation
1143 * can occur. Where any new reservation would be required the
1144 * reservation change is prepared, but not committed. Once the page
1145 * has been allocated from the subpool and instantiated the change should
1146 * be committed via vma_commit_reservation. No action is required on
1149 static long vma_needs_reservation(struct hstate
*h
,
1150 struct vm_area_struct
*vma
, unsigned long addr
)
1152 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1153 struct inode
*inode
= mapping
->host
;
1155 if (vma
->vm_flags
& VM_MAYSHARE
) {
1156 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1157 struct resv_map
*resv
= inode
->i_mapping
->private_data
;
1159 return region_chg(resv
, idx
, idx
+ 1);
1161 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1166 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1167 struct resv_map
*resv
= vma_resv_map(vma
);
1169 err
= region_chg(resv
, idx
, idx
+ 1);
1175 static void vma_commit_reservation(struct hstate
*h
,
1176 struct vm_area_struct
*vma
, unsigned long addr
)
1178 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1179 struct inode
*inode
= mapping
->host
;
1181 if (vma
->vm_flags
& VM_MAYSHARE
) {
1182 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1183 struct resv_map
*resv
= inode
->i_mapping
->private_data
;
1185 region_add(resv
, idx
, idx
+ 1);
1187 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1188 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1189 struct resv_map
*resv
= vma_resv_map(vma
);
1191 /* Mark this page used in the map. */
1192 region_add(resv
, idx
, idx
+ 1);
1196 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1197 unsigned long addr
, int avoid_reserve
)
1199 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1200 struct hstate
*h
= hstate_vma(vma
);
1204 struct hugetlb_cgroup
*h_cg
;
1206 idx
= hstate_index(h
);
1208 * Processes that did not create the mapping will have no
1209 * reserves and will not have accounted against subpool
1210 * limit. Check that the subpool limit can be made before
1211 * satisfying the allocation MAP_NORESERVE mappings may also
1212 * need pages and subpool limit allocated allocated if no reserve
1215 chg
= vma_needs_reservation(h
, vma
, addr
);
1217 return ERR_PTR(-ENOMEM
);
1218 if (chg
|| avoid_reserve
)
1219 if (hugepage_subpool_get_pages(spool
, 1))
1220 return ERR_PTR(-ENOSPC
);
1222 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1224 if (chg
|| avoid_reserve
)
1225 hugepage_subpool_put_pages(spool
, 1);
1226 return ERR_PTR(-ENOSPC
);
1228 spin_lock(&hugetlb_lock
);
1229 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1231 spin_unlock(&hugetlb_lock
);
1232 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1234 hugetlb_cgroup_uncharge_cgroup(idx
,
1235 pages_per_huge_page(h
),
1237 if (chg
|| avoid_reserve
)
1238 hugepage_subpool_put_pages(spool
, 1);
1239 return ERR_PTR(-ENOSPC
);
1241 spin_lock(&hugetlb_lock
);
1242 list_move(&page
->lru
, &h
->hugepage_activelist
);
1245 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1246 spin_unlock(&hugetlb_lock
);
1248 set_page_private(page
, (unsigned long)spool
);
1250 vma_commit_reservation(h
, vma
, addr
);
1255 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1256 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1257 * where no ERR_VALUE is expected to be returned.
1259 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1260 unsigned long addr
, int avoid_reserve
)
1262 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1268 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1270 struct huge_bootmem_page
*m
;
1273 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1276 addr
= memblock_virt_alloc_try_nid_nopanic(
1277 huge_page_size(h
), huge_page_size(h
),
1278 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1281 * Use the beginning of the huge page to store the
1282 * huge_bootmem_page struct (until gather_bootmem
1283 * puts them into the mem_map).
1292 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1293 /* Put them into a private list first because mem_map is not up yet */
1294 list_add(&m
->list
, &huge_boot_pages
);
1299 static void prep_compound_huge_page(struct page
*page
, int order
)
1301 if (unlikely(order
> (MAX_ORDER
- 1)))
1302 prep_compound_gigantic_page(page
, order
);
1304 prep_compound_page(page
, order
);
1307 /* Put bootmem huge pages into the standard lists after mem_map is up */
1308 static void __init
gather_bootmem_prealloc(void)
1310 struct huge_bootmem_page
*m
;
1312 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1313 struct hstate
*h
= m
->hstate
;
1316 #ifdef CONFIG_HIGHMEM
1317 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1318 memblock_free_late(__pa(m
),
1319 sizeof(struct huge_bootmem_page
));
1321 page
= virt_to_page(m
);
1323 WARN_ON(page_count(page
) != 1);
1324 prep_compound_huge_page(page
, h
->order
);
1325 WARN_ON(PageReserved(page
));
1326 prep_new_huge_page(h
, page
, page_to_nid(page
));
1328 * If we had gigantic hugepages allocated at boot time, we need
1329 * to restore the 'stolen' pages to totalram_pages in order to
1330 * fix confusing memory reports from free(1) and another
1331 * side-effects, like CommitLimit going negative.
1333 if (h
->order
> (MAX_ORDER
- 1))
1334 adjust_managed_page_count(page
, 1 << h
->order
);
1338 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1342 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1343 if (h
->order
>= MAX_ORDER
) {
1344 if (!alloc_bootmem_huge_page(h
))
1346 } else if (!alloc_fresh_huge_page(h
,
1347 &node_states
[N_MEMORY
]))
1350 h
->max_huge_pages
= i
;
1353 static void __init
hugetlb_init_hstates(void)
1357 for_each_hstate(h
) {
1358 /* oversize hugepages were init'ed in early boot */
1359 if (h
->order
< MAX_ORDER
)
1360 hugetlb_hstate_alloc_pages(h
);
1364 static char * __init
memfmt(char *buf
, unsigned long n
)
1366 if (n
>= (1UL << 30))
1367 sprintf(buf
, "%lu GB", n
>> 30);
1368 else if (n
>= (1UL << 20))
1369 sprintf(buf
, "%lu MB", n
>> 20);
1371 sprintf(buf
, "%lu KB", n
>> 10);
1375 static void __init
report_hugepages(void)
1379 for_each_hstate(h
) {
1381 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1382 memfmt(buf
, huge_page_size(h
)),
1383 h
->free_huge_pages
);
1387 #ifdef CONFIG_HIGHMEM
1388 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1389 nodemask_t
*nodes_allowed
)
1393 if (h
->order
>= MAX_ORDER
)
1396 for_each_node_mask(i
, *nodes_allowed
) {
1397 struct page
*page
, *next
;
1398 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1399 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1400 if (count
>= h
->nr_huge_pages
)
1402 if (PageHighMem(page
))
1404 list_del(&page
->lru
);
1405 update_and_free_page(h
, page
);
1406 h
->free_huge_pages
--;
1407 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1412 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1413 nodemask_t
*nodes_allowed
)
1419 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1420 * balanced by operating on them in a round-robin fashion.
1421 * Returns 1 if an adjustment was made.
1423 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1428 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1431 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1432 if (h
->surplus_huge_pages_node
[node
])
1436 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1437 if (h
->surplus_huge_pages_node
[node
] <
1438 h
->nr_huge_pages_node
[node
])
1445 h
->surplus_huge_pages
+= delta
;
1446 h
->surplus_huge_pages_node
[node
] += delta
;
1450 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1451 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1452 nodemask_t
*nodes_allowed
)
1454 unsigned long min_count
, ret
;
1456 if (h
->order
>= MAX_ORDER
)
1457 return h
->max_huge_pages
;
1460 * Increase the pool size
1461 * First take pages out of surplus state. Then make up the
1462 * remaining difference by allocating fresh huge pages.
1464 * We might race with alloc_buddy_huge_page() here and be unable
1465 * to convert a surplus huge page to a normal huge page. That is
1466 * not critical, though, it just means the overall size of the
1467 * pool might be one hugepage larger than it needs to be, but
1468 * within all the constraints specified by the sysctls.
1470 spin_lock(&hugetlb_lock
);
1471 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1472 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1476 while (count
> persistent_huge_pages(h
)) {
1478 * If this allocation races such that we no longer need the
1479 * page, free_huge_page will handle it by freeing the page
1480 * and reducing the surplus.
1482 spin_unlock(&hugetlb_lock
);
1483 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1484 spin_lock(&hugetlb_lock
);
1488 /* Bail for signals. Probably ctrl-c from user */
1489 if (signal_pending(current
))
1494 * Decrease the pool size
1495 * First return free pages to the buddy allocator (being careful
1496 * to keep enough around to satisfy reservations). Then place
1497 * pages into surplus state as needed so the pool will shrink
1498 * to the desired size as pages become free.
1500 * By placing pages into the surplus state independent of the
1501 * overcommit value, we are allowing the surplus pool size to
1502 * exceed overcommit. There are few sane options here. Since
1503 * alloc_buddy_huge_page() is checking the global counter,
1504 * though, we'll note that we're not allowed to exceed surplus
1505 * and won't grow the pool anywhere else. Not until one of the
1506 * sysctls are changed, or the surplus pages go out of use.
1508 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1509 min_count
= max(count
, min_count
);
1510 try_to_free_low(h
, min_count
, nodes_allowed
);
1511 while (min_count
< persistent_huge_pages(h
)) {
1512 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1515 while (count
< persistent_huge_pages(h
)) {
1516 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1520 ret
= persistent_huge_pages(h
);
1521 spin_unlock(&hugetlb_lock
);
1525 #define HSTATE_ATTR_RO(_name) \
1526 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1528 #define HSTATE_ATTR(_name) \
1529 static struct kobj_attribute _name##_attr = \
1530 __ATTR(_name, 0644, _name##_show, _name##_store)
1532 static struct kobject
*hugepages_kobj
;
1533 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1535 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1537 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1541 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1542 if (hstate_kobjs
[i
] == kobj
) {
1544 *nidp
= NUMA_NO_NODE
;
1548 return kobj_to_node_hstate(kobj
, nidp
);
1551 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1552 struct kobj_attribute
*attr
, char *buf
)
1555 unsigned long nr_huge_pages
;
1558 h
= kobj_to_hstate(kobj
, &nid
);
1559 if (nid
== NUMA_NO_NODE
)
1560 nr_huge_pages
= h
->nr_huge_pages
;
1562 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1564 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1567 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1568 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1569 const char *buf
, size_t len
)
1573 unsigned long count
;
1575 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1577 err
= kstrtoul(buf
, 10, &count
);
1581 h
= kobj_to_hstate(kobj
, &nid
);
1582 if (h
->order
>= MAX_ORDER
) {
1587 if (nid
== NUMA_NO_NODE
) {
1589 * global hstate attribute
1591 if (!(obey_mempolicy
&&
1592 init_nodemask_of_mempolicy(nodes_allowed
))) {
1593 NODEMASK_FREE(nodes_allowed
);
1594 nodes_allowed
= &node_states
[N_MEMORY
];
1596 } else if (nodes_allowed
) {
1598 * per node hstate attribute: adjust count to global,
1599 * but restrict alloc/free to the specified node.
1601 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1602 init_nodemask_of_node(nodes_allowed
, nid
);
1604 nodes_allowed
= &node_states
[N_MEMORY
];
1606 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1608 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1609 NODEMASK_FREE(nodes_allowed
);
1613 NODEMASK_FREE(nodes_allowed
);
1617 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1618 struct kobj_attribute
*attr
, char *buf
)
1620 return nr_hugepages_show_common(kobj
, attr
, buf
);
1623 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1624 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1626 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1628 HSTATE_ATTR(nr_hugepages
);
1633 * hstate attribute for optionally mempolicy-based constraint on persistent
1634 * huge page alloc/free.
1636 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1637 struct kobj_attribute
*attr
, char *buf
)
1639 return nr_hugepages_show_common(kobj
, attr
, buf
);
1642 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1643 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1645 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1647 HSTATE_ATTR(nr_hugepages_mempolicy
);
1651 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1652 struct kobj_attribute
*attr
, char *buf
)
1654 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1655 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1658 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1659 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1662 unsigned long input
;
1663 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1665 if (h
->order
>= MAX_ORDER
)
1668 err
= kstrtoul(buf
, 10, &input
);
1672 spin_lock(&hugetlb_lock
);
1673 h
->nr_overcommit_huge_pages
= input
;
1674 spin_unlock(&hugetlb_lock
);
1678 HSTATE_ATTR(nr_overcommit_hugepages
);
1680 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1681 struct kobj_attribute
*attr
, char *buf
)
1684 unsigned long free_huge_pages
;
1687 h
= kobj_to_hstate(kobj
, &nid
);
1688 if (nid
== NUMA_NO_NODE
)
1689 free_huge_pages
= h
->free_huge_pages
;
1691 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1693 return sprintf(buf
, "%lu\n", free_huge_pages
);
1695 HSTATE_ATTR_RO(free_hugepages
);
1697 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1698 struct kobj_attribute
*attr
, char *buf
)
1700 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1701 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1703 HSTATE_ATTR_RO(resv_hugepages
);
1705 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1706 struct kobj_attribute
*attr
, char *buf
)
1709 unsigned long surplus_huge_pages
;
1712 h
= kobj_to_hstate(kobj
, &nid
);
1713 if (nid
== NUMA_NO_NODE
)
1714 surplus_huge_pages
= h
->surplus_huge_pages
;
1716 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1718 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1720 HSTATE_ATTR_RO(surplus_hugepages
);
1722 static struct attribute
*hstate_attrs
[] = {
1723 &nr_hugepages_attr
.attr
,
1724 &nr_overcommit_hugepages_attr
.attr
,
1725 &free_hugepages_attr
.attr
,
1726 &resv_hugepages_attr
.attr
,
1727 &surplus_hugepages_attr
.attr
,
1729 &nr_hugepages_mempolicy_attr
.attr
,
1734 static struct attribute_group hstate_attr_group
= {
1735 .attrs
= hstate_attrs
,
1738 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1739 struct kobject
**hstate_kobjs
,
1740 struct attribute_group
*hstate_attr_group
)
1743 int hi
= hstate_index(h
);
1745 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1746 if (!hstate_kobjs
[hi
])
1749 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1751 kobject_put(hstate_kobjs
[hi
]);
1756 static void __init
hugetlb_sysfs_init(void)
1761 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1762 if (!hugepages_kobj
)
1765 for_each_hstate(h
) {
1766 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1767 hstate_kobjs
, &hstate_attr_group
);
1769 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1776 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1777 * with node devices in node_devices[] using a parallel array. The array
1778 * index of a node device or _hstate == node id.
1779 * This is here to avoid any static dependency of the node device driver, in
1780 * the base kernel, on the hugetlb module.
1782 struct node_hstate
{
1783 struct kobject
*hugepages_kobj
;
1784 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1786 struct node_hstate node_hstates
[MAX_NUMNODES
];
1789 * A subset of global hstate attributes for node devices
1791 static struct attribute
*per_node_hstate_attrs
[] = {
1792 &nr_hugepages_attr
.attr
,
1793 &free_hugepages_attr
.attr
,
1794 &surplus_hugepages_attr
.attr
,
1798 static struct attribute_group per_node_hstate_attr_group
= {
1799 .attrs
= per_node_hstate_attrs
,
1803 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1804 * Returns node id via non-NULL nidp.
1806 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1810 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1811 struct node_hstate
*nhs
= &node_hstates
[nid
];
1813 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1814 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1826 * Unregister hstate attributes from a single node device.
1827 * No-op if no hstate attributes attached.
1829 static void hugetlb_unregister_node(struct node
*node
)
1832 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1834 if (!nhs
->hugepages_kobj
)
1835 return; /* no hstate attributes */
1837 for_each_hstate(h
) {
1838 int idx
= hstate_index(h
);
1839 if (nhs
->hstate_kobjs
[idx
]) {
1840 kobject_put(nhs
->hstate_kobjs
[idx
]);
1841 nhs
->hstate_kobjs
[idx
] = NULL
;
1845 kobject_put(nhs
->hugepages_kobj
);
1846 nhs
->hugepages_kobj
= NULL
;
1850 * hugetlb module exit: unregister hstate attributes from node devices
1853 static void hugetlb_unregister_all_nodes(void)
1858 * disable node device registrations.
1860 register_hugetlbfs_with_node(NULL
, NULL
);
1863 * remove hstate attributes from any nodes that have them.
1865 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1866 hugetlb_unregister_node(node_devices
[nid
]);
1870 * Register hstate attributes for a single node device.
1871 * No-op if attributes already registered.
1873 static void hugetlb_register_node(struct node
*node
)
1876 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1879 if (nhs
->hugepages_kobj
)
1880 return; /* already allocated */
1882 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1884 if (!nhs
->hugepages_kobj
)
1887 for_each_hstate(h
) {
1888 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1890 &per_node_hstate_attr_group
);
1892 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1893 h
->name
, node
->dev
.id
);
1894 hugetlb_unregister_node(node
);
1901 * hugetlb init time: register hstate attributes for all registered node
1902 * devices of nodes that have memory. All on-line nodes should have
1903 * registered their associated device by this time.
1905 static void hugetlb_register_all_nodes(void)
1909 for_each_node_state(nid
, N_MEMORY
) {
1910 struct node
*node
= node_devices
[nid
];
1911 if (node
->dev
.id
== nid
)
1912 hugetlb_register_node(node
);
1916 * Let the node device driver know we're here so it can
1917 * [un]register hstate attributes on node hotplug.
1919 register_hugetlbfs_with_node(hugetlb_register_node
,
1920 hugetlb_unregister_node
);
1922 #else /* !CONFIG_NUMA */
1924 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1932 static void hugetlb_unregister_all_nodes(void) { }
1934 static void hugetlb_register_all_nodes(void) { }
1938 static void __exit
hugetlb_exit(void)
1942 hugetlb_unregister_all_nodes();
1944 for_each_hstate(h
) {
1945 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1948 kobject_put(hugepages_kobj
);
1950 module_exit(hugetlb_exit
);
1952 static int __init
hugetlb_init(void)
1954 /* Some platform decide whether they support huge pages at boot
1955 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1956 * there is no such support
1958 if (HPAGE_SHIFT
== 0)
1961 if (!size_to_hstate(default_hstate_size
)) {
1962 default_hstate_size
= HPAGE_SIZE
;
1963 if (!size_to_hstate(default_hstate_size
))
1964 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1966 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1967 if (default_hstate_max_huge_pages
)
1968 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1970 hugetlb_init_hstates();
1971 gather_bootmem_prealloc();
1974 hugetlb_sysfs_init();
1975 hugetlb_register_all_nodes();
1976 hugetlb_cgroup_file_init();
1980 module_init(hugetlb_init
);
1982 /* Should be called on processing a hugepagesz=... option */
1983 void __init
hugetlb_add_hstate(unsigned order
)
1988 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1989 pr_warning("hugepagesz= specified twice, ignoring\n");
1992 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1994 h
= &hstates
[hugetlb_max_hstate
++];
1996 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1997 h
->nr_huge_pages
= 0;
1998 h
->free_huge_pages
= 0;
1999 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2000 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2001 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2002 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2003 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2004 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2005 huge_page_size(h
)/1024);
2010 static int __init
hugetlb_nrpages_setup(char *s
)
2013 static unsigned long *last_mhp
;
2016 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2017 * so this hugepages= parameter goes to the "default hstate".
2019 if (!hugetlb_max_hstate
)
2020 mhp
= &default_hstate_max_huge_pages
;
2022 mhp
= &parsed_hstate
->max_huge_pages
;
2024 if (mhp
== last_mhp
) {
2025 pr_warning("hugepages= specified twice without "
2026 "interleaving hugepagesz=, ignoring\n");
2030 if (sscanf(s
, "%lu", mhp
) <= 0)
2034 * Global state is always initialized later in hugetlb_init.
2035 * But we need to allocate >= MAX_ORDER hstates here early to still
2036 * use the bootmem allocator.
2038 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2039 hugetlb_hstate_alloc_pages(parsed_hstate
);
2045 __setup("hugepages=", hugetlb_nrpages_setup
);
2047 static int __init
hugetlb_default_setup(char *s
)
2049 default_hstate_size
= memparse(s
, &s
);
2052 __setup("default_hugepagesz=", hugetlb_default_setup
);
2054 static unsigned int cpuset_mems_nr(unsigned int *array
)
2057 unsigned int nr
= 0;
2059 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2065 #ifdef CONFIG_SYSCTL
2066 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2067 struct ctl_table
*table
, int write
,
2068 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2070 struct hstate
*h
= &default_hstate
;
2074 tmp
= h
->max_huge_pages
;
2076 if (write
&& h
->order
>= MAX_ORDER
)
2080 table
->maxlen
= sizeof(unsigned long);
2081 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2086 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2087 GFP_KERNEL
| __GFP_NORETRY
);
2088 if (!(obey_mempolicy
&&
2089 init_nodemask_of_mempolicy(nodes_allowed
))) {
2090 NODEMASK_FREE(nodes_allowed
);
2091 nodes_allowed
= &node_states
[N_MEMORY
];
2093 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2095 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2096 NODEMASK_FREE(nodes_allowed
);
2102 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2103 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2106 return hugetlb_sysctl_handler_common(false, table
, write
,
2107 buffer
, length
, ppos
);
2111 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2112 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2114 return hugetlb_sysctl_handler_common(true, table
, write
,
2115 buffer
, length
, ppos
);
2117 #endif /* CONFIG_NUMA */
2119 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2120 void __user
*buffer
,
2121 size_t *length
, loff_t
*ppos
)
2123 struct hstate
*h
= &default_hstate
;
2127 tmp
= h
->nr_overcommit_huge_pages
;
2129 if (write
&& h
->order
>= MAX_ORDER
)
2133 table
->maxlen
= sizeof(unsigned long);
2134 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2139 spin_lock(&hugetlb_lock
);
2140 h
->nr_overcommit_huge_pages
= tmp
;
2141 spin_unlock(&hugetlb_lock
);
2147 #endif /* CONFIG_SYSCTL */
2149 void hugetlb_report_meminfo(struct seq_file
*m
)
2151 struct hstate
*h
= &default_hstate
;
2153 "HugePages_Total: %5lu\n"
2154 "HugePages_Free: %5lu\n"
2155 "HugePages_Rsvd: %5lu\n"
2156 "HugePages_Surp: %5lu\n"
2157 "Hugepagesize: %8lu kB\n",
2161 h
->surplus_huge_pages
,
2162 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2165 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2167 struct hstate
*h
= &default_hstate
;
2169 "Node %d HugePages_Total: %5u\n"
2170 "Node %d HugePages_Free: %5u\n"
2171 "Node %d HugePages_Surp: %5u\n",
2172 nid
, h
->nr_huge_pages_node
[nid
],
2173 nid
, h
->free_huge_pages_node
[nid
],
2174 nid
, h
->surplus_huge_pages_node
[nid
]);
2177 void hugetlb_show_meminfo(void)
2182 for_each_node_state(nid
, N_MEMORY
)
2184 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2186 h
->nr_huge_pages_node
[nid
],
2187 h
->free_huge_pages_node
[nid
],
2188 h
->surplus_huge_pages_node
[nid
],
2189 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2192 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2193 unsigned long hugetlb_total_pages(void)
2196 unsigned long nr_total_pages
= 0;
2199 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2200 return nr_total_pages
;
2203 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2207 spin_lock(&hugetlb_lock
);
2209 * When cpuset is configured, it breaks the strict hugetlb page
2210 * reservation as the accounting is done on a global variable. Such
2211 * reservation is completely rubbish in the presence of cpuset because
2212 * the reservation is not checked against page availability for the
2213 * current cpuset. Application can still potentially OOM'ed by kernel
2214 * with lack of free htlb page in cpuset that the task is in.
2215 * Attempt to enforce strict accounting with cpuset is almost
2216 * impossible (or too ugly) because cpuset is too fluid that
2217 * task or memory node can be dynamically moved between cpusets.
2219 * The change of semantics for shared hugetlb mapping with cpuset is
2220 * undesirable. However, in order to preserve some of the semantics,
2221 * we fall back to check against current free page availability as
2222 * a best attempt and hopefully to minimize the impact of changing
2223 * semantics that cpuset has.
2226 if (gather_surplus_pages(h
, delta
) < 0)
2229 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2230 return_unused_surplus_pages(h
, delta
);
2237 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2240 spin_unlock(&hugetlb_lock
);
2244 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2246 struct resv_map
*resv
= vma_resv_map(vma
);
2249 * This new VMA should share its siblings reservation map if present.
2250 * The VMA will only ever have a valid reservation map pointer where
2251 * it is being copied for another still existing VMA. As that VMA
2252 * has a reference to the reservation map it cannot disappear until
2253 * after this open call completes. It is therefore safe to take a
2254 * new reference here without additional locking.
2257 kref_get(&resv
->refs
);
2260 static void resv_map_put(struct vm_area_struct
*vma
)
2262 struct resv_map
*resv
= vma_resv_map(vma
);
2266 kref_put(&resv
->refs
, resv_map_release
);
2269 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2271 struct hstate
*h
= hstate_vma(vma
);
2272 struct resv_map
*resv
= vma_resv_map(vma
);
2273 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2274 unsigned long reserve
;
2275 unsigned long start
;
2279 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2280 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2282 reserve
= (end
- start
) -
2283 region_count(resv
, start
, end
);
2288 hugetlb_acct_memory(h
, -reserve
);
2289 hugepage_subpool_put_pages(spool
, reserve
);
2295 * We cannot handle pagefaults against hugetlb pages at all. They cause
2296 * handle_mm_fault() to try to instantiate regular-sized pages in the
2297 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2300 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2306 const struct vm_operations_struct hugetlb_vm_ops
= {
2307 .fault
= hugetlb_vm_op_fault
,
2308 .open
= hugetlb_vm_op_open
,
2309 .close
= hugetlb_vm_op_close
,
2312 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2318 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2319 vma
->vm_page_prot
)));
2321 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2322 vma
->vm_page_prot
));
2324 entry
= pte_mkyoung(entry
);
2325 entry
= pte_mkhuge(entry
);
2326 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2331 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2332 unsigned long address
, pte_t
*ptep
)
2336 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2337 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2338 update_mmu_cache(vma
, address
, ptep
);
2342 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2343 struct vm_area_struct
*vma
)
2345 pte_t
*src_pte
, *dst_pte
, entry
;
2346 struct page
*ptepage
;
2349 struct hstate
*h
= hstate_vma(vma
);
2350 unsigned long sz
= huge_page_size(h
);
2351 unsigned long mmun_start
; /* For mmu_notifiers */
2352 unsigned long mmun_end
; /* For mmu_notifiers */
2355 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2357 mmun_start
= vma
->vm_start
;
2358 mmun_end
= vma
->vm_end
;
2360 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
2362 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2363 spinlock_t
*src_ptl
, *dst_ptl
;
2364 src_pte
= huge_pte_offset(src
, addr
);
2367 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2373 /* If the pagetables are shared don't copy or take references */
2374 if (dst_pte
== src_pte
)
2377 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
2378 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
2379 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
2380 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2382 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2383 entry
= huge_ptep_get(src_pte
);
2384 ptepage
= pte_page(entry
);
2386 page_dup_rmap(ptepage
);
2387 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2389 spin_unlock(src_ptl
);
2390 spin_unlock(dst_ptl
);
2394 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
2399 static int is_hugetlb_entry_migration(pte_t pte
)
2403 if (huge_pte_none(pte
) || pte_present(pte
))
2405 swp
= pte_to_swp_entry(pte
);
2406 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2412 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2416 if (huge_pte_none(pte
) || pte_present(pte
))
2418 swp
= pte_to_swp_entry(pte
);
2419 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2425 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2426 unsigned long start
, unsigned long end
,
2427 struct page
*ref_page
)
2429 int force_flush
= 0;
2430 struct mm_struct
*mm
= vma
->vm_mm
;
2431 unsigned long address
;
2436 struct hstate
*h
= hstate_vma(vma
);
2437 unsigned long sz
= huge_page_size(h
);
2438 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2439 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2441 WARN_ON(!is_vm_hugetlb_page(vma
));
2442 BUG_ON(start
& ~huge_page_mask(h
));
2443 BUG_ON(end
& ~huge_page_mask(h
));
2445 tlb_start_vma(tlb
, vma
);
2446 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2448 for (address
= start
; address
< end
; address
+= sz
) {
2449 ptep
= huge_pte_offset(mm
, address
);
2453 ptl
= huge_pte_lock(h
, mm
, ptep
);
2454 if (huge_pmd_unshare(mm
, &address
, ptep
))
2457 pte
= huge_ptep_get(ptep
);
2458 if (huge_pte_none(pte
))
2462 * HWPoisoned hugepage is already unmapped and dropped reference
2464 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2465 huge_pte_clear(mm
, address
, ptep
);
2469 page
= pte_page(pte
);
2471 * If a reference page is supplied, it is because a specific
2472 * page is being unmapped, not a range. Ensure the page we
2473 * are about to unmap is the actual page of interest.
2476 if (page
!= ref_page
)
2480 * Mark the VMA as having unmapped its page so that
2481 * future faults in this VMA will fail rather than
2482 * looking like data was lost
2484 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2487 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2488 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2489 if (huge_pte_dirty(pte
))
2490 set_page_dirty(page
);
2492 page_remove_rmap(page
);
2493 force_flush
= !__tlb_remove_page(tlb
, page
);
2498 /* Bail out after unmapping reference page if supplied */
2507 * mmu_gather ran out of room to batch pages, we break out of
2508 * the PTE lock to avoid doing the potential expensive TLB invalidate
2509 * and page-free while holding it.
2514 if (address
< end
&& !ref_page
)
2517 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2518 tlb_end_vma(tlb
, vma
);
2521 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2522 struct vm_area_struct
*vma
, unsigned long start
,
2523 unsigned long end
, struct page
*ref_page
)
2525 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2528 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2529 * test will fail on a vma being torn down, and not grab a page table
2530 * on its way out. We're lucky that the flag has such an appropriate
2531 * name, and can in fact be safely cleared here. We could clear it
2532 * before the __unmap_hugepage_range above, but all that's necessary
2533 * is to clear it before releasing the i_mmap_mutex. This works
2534 * because in the context this is called, the VMA is about to be
2535 * destroyed and the i_mmap_mutex is held.
2537 vma
->vm_flags
&= ~VM_MAYSHARE
;
2540 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2541 unsigned long end
, struct page
*ref_page
)
2543 struct mm_struct
*mm
;
2544 struct mmu_gather tlb
;
2548 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2549 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2550 tlb_finish_mmu(&tlb
, start
, end
);
2554 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2555 * mappping it owns the reserve page for. The intention is to unmap the page
2556 * from other VMAs and let the children be SIGKILLed if they are faulting the
2559 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2560 struct page
*page
, unsigned long address
)
2562 struct hstate
*h
= hstate_vma(vma
);
2563 struct vm_area_struct
*iter_vma
;
2564 struct address_space
*mapping
;
2568 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2569 * from page cache lookup which is in HPAGE_SIZE units.
2571 address
= address
& huge_page_mask(h
);
2572 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2574 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2577 * Take the mapping lock for the duration of the table walk. As
2578 * this mapping should be shared between all the VMAs,
2579 * __unmap_hugepage_range() is called as the lock is already held
2581 mutex_lock(&mapping
->i_mmap_mutex
);
2582 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2583 /* Do not unmap the current VMA */
2584 if (iter_vma
== vma
)
2588 * Unmap the page from other VMAs without their own reserves.
2589 * They get marked to be SIGKILLed if they fault in these
2590 * areas. This is because a future no-page fault on this VMA
2591 * could insert a zeroed page instead of the data existing
2592 * from the time of fork. This would look like data corruption
2594 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2595 unmap_hugepage_range(iter_vma
, address
,
2596 address
+ huge_page_size(h
), page
);
2598 mutex_unlock(&mapping
->i_mmap_mutex
);
2604 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2605 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2606 * cannot race with other handlers or page migration.
2607 * Keep the pte_same checks anyway to make transition from the mutex easier.
2609 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2610 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2611 struct page
*pagecache_page
, spinlock_t
*ptl
)
2613 struct hstate
*h
= hstate_vma(vma
);
2614 struct page
*old_page
, *new_page
;
2615 int outside_reserve
= 0;
2616 unsigned long mmun_start
; /* For mmu_notifiers */
2617 unsigned long mmun_end
; /* For mmu_notifiers */
2619 old_page
= pte_page(pte
);
2622 /* If no-one else is actually using this page, avoid the copy
2623 * and just make the page writable */
2624 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2625 page_move_anon_rmap(old_page
, vma
, address
);
2626 set_huge_ptep_writable(vma
, address
, ptep
);
2631 * If the process that created a MAP_PRIVATE mapping is about to
2632 * perform a COW due to a shared page count, attempt to satisfy
2633 * the allocation without using the existing reserves. The pagecache
2634 * page is used to determine if the reserve at this address was
2635 * consumed or not. If reserves were used, a partial faulted mapping
2636 * at the time of fork() could consume its reserves on COW instead
2637 * of the full address range.
2639 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2640 old_page
!= pagecache_page
)
2641 outside_reserve
= 1;
2643 page_cache_get(old_page
);
2645 /* Drop page table lock as buddy allocator may be called */
2647 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2649 if (IS_ERR(new_page
)) {
2650 long err
= PTR_ERR(new_page
);
2651 page_cache_release(old_page
);
2654 * If a process owning a MAP_PRIVATE mapping fails to COW,
2655 * it is due to references held by a child and an insufficient
2656 * huge page pool. To guarantee the original mappers
2657 * reliability, unmap the page from child processes. The child
2658 * may get SIGKILLed if it later faults.
2660 if (outside_reserve
) {
2661 BUG_ON(huge_pte_none(pte
));
2662 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2663 BUG_ON(huge_pte_none(pte
));
2665 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2666 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2667 goto retry_avoidcopy
;
2669 * race occurs while re-acquiring page table
2670 * lock, and our job is done.
2677 /* Caller expects lock to be held */
2680 return VM_FAULT_OOM
;
2682 return VM_FAULT_SIGBUS
;
2686 * When the original hugepage is shared one, it does not have
2687 * anon_vma prepared.
2689 if (unlikely(anon_vma_prepare(vma
))) {
2690 page_cache_release(new_page
);
2691 page_cache_release(old_page
);
2692 /* Caller expects lock to be held */
2694 return VM_FAULT_OOM
;
2697 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2698 pages_per_huge_page(h
));
2699 __SetPageUptodate(new_page
);
2701 mmun_start
= address
& huge_page_mask(h
);
2702 mmun_end
= mmun_start
+ huge_page_size(h
);
2703 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2705 * Retake the page table lock to check for racing updates
2706 * before the page tables are altered
2709 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2710 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2711 ClearPagePrivate(new_page
);
2714 huge_ptep_clear_flush(vma
, address
, ptep
);
2715 set_huge_pte_at(mm
, address
, ptep
,
2716 make_huge_pte(vma
, new_page
, 1));
2717 page_remove_rmap(old_page
);
2718 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2719 /* Make the old page be freed below */
2720 new_page
= old_page
;
2723 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2724 page_cache_release(new_page
);
2725 page_cache_release(old_page
);
2727 /* Caller expects lock to be held */
2732 /* Return the pagecache page at a given address within a VMA */
2733 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2734 struct vm_area_struct
*vma
, unsigned long address
)
2736 struct address_space
*mapping
;
2739 mapping
= vma
->vm_file
->f_mapping
;
2740 idx
= vma_hugecache_offset(h
, vma
, address
);
2742 return find_lock_page(mapping
, idx
);
2746 * Return whether there is a pagecache page to back given address within VMA.
2747 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2749 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2750 struct vm_area_struct
*vma
, unsigned long address
)
2752 struct address_space
*mapping
;
2756 mapping
= vma
->vm_file
->f_mapping
;
2757 idx
= vma_hugecache_offset(h
, vma
, address
);
2759 page
= find_get_page(mapping
, idx
);
2762 return page
!= NULL
;
2765 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2766 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2768 struct hstate
*h
= hstate_vma(vma
);
2769 int ret
= VM_FAULT_SIGBUS
;
2774 struct address_space
*mapping
;
2779 * Currently, we are forced to kill the process in the event the
2780 * original mapper has unmapped pages from the child due to a failed
2781 * COW. Warn that such a situation has occurred as it may not be obvious
2783 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2784 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2789 mapping
= vma
->vm_file
->f_mapping
;
2790 idx
= vma_hugecache_offset(h
, vma
, address
);
2793 * Use page lock to guard against racing truncation
2794 * before we get page_table_lock.
2797 page
= find_lock_page(mapping
, idx
);
2799 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2802 page
= alloc_huge_page(vma
, address
, 0);
2804 ret
= PTR_ERR(page
);
2808 ret
= VM_FAULT_SIGBUS
;
2811 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2812 __SetPageUptodate(page
);
2814 if (vma
->vm_flags
& VM_MAYSHARE
) {
2816 struct inode
*inode
= mapping
->host
;
2818 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2825 ClearPagePrivate(page
);
2827 spin_lock(&inode
->i_lock
);
2828 inode
->i_blocks
+= blocks_per_huge_page(h
);
2829 spin_unlock(&inode
->i_lock
);
2832 if (unlikely(anon_vma_prepare(vma
))) {
2834 goto backout_unlocked
;
2840 * If memory error occurs between mmap() and fault, some process
2841 * don't have hwpoisoned swap entry for errored virtual address.
2842 * So we need to block hugepage fault by PG_hwpoison bit check.
2844 if (unlikely(PageHWPoison(page
))) {
2845 ret
= VM_FAULT_HWPOISON
|
2846 VM_FAULT_SET_HINDEX(hstate_index(h
));
2847 goto backout_unlocked
;
2852 * If we are going to COW a private mapping later, we examine the
2853 * pending reservations for this page now. This will ensure that
2854 * any allocations necessary to record that reservation occur outside
2857 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2858 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2860 goto backout_unlocked
;
2863 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
2865 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2870 if (!huge_pte_none(huge_ptep_get(ptep
)))
2874 ClearPagePrivate(page
);
2875 hugepage_add_new_anon_rmap(page
, vma
, address
);
2878 page_dup_rmap(page
);
2879 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2880 && (vma
->vm_flags
& VM_SHARED
)));
2881 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2883 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2884 /* Optimization, do the COW without a second fault */
2885 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
2901 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2902 unsigned long address
, unsigned int flags
)
2908 struct page
*page
= NULL
;
2909 struct page
*pagecache_page
= NULL
;
2910 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2911 struct hstate
*h
= hstate_vma(vma
);
2913 address
&= huge_page_mask(h
);
2915 ptep
= huge_pte_offset(mm
, address
);
2917 entry
= huge_ptep_get(ptep
);
2918 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2919 migration_entry_wait_huge(vma
, mm
, ptep
);
2921 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2922 return VM_FAULT_HWPOISON_LARGE
|
2923 VM_FAULT_SET_HINDEX(hstate_index(h
));
2926 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2928 return VM_FAULT_OOM
;
2931 * Serialize hugepage allocation and instantiation, so that we don't
2932 * get spurious allocation failures if two CPUs race to instantiate
2933 * the same page in the page cache.
2935 mutex_lock(&hugetlb_instantiation_mutex
);
2936 entry
= huge_ptep_get(ptep
);
2937 if (huge_pte_none(entry
)) {
2938 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2945 * If we are going to COW the mapping later, we examine the pending
2946 * reservations for this page now. This will ensure that any
2947 * allocations necessary to record that reservation occur outside the
2948 * spinlock. For private mappings, we also lookup the pagecache
2949 * page now as it is used to determine if a reservation has been
2952 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2953 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2958 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2959 pagecache_page
= hugetlbfs_pagecache_page(h
,
2964 * hugetlb_cow() requires page locks of pte_page(entry) and
2965 * pagecache_page, so here we need take the former one
2966 * when page != pagecache_page or !pagecache_page.
2967 * Note that locking order is always pagecache_page -> page,
2968 * so no worry about deadlock.
2970 page
= pte_page(entry
);
2972 if (page
!= pagecache_page
)
2975 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
2977 /* Check for a racing update before calling hugetlb_cow */
2978 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2982 if (flags
& FAULT_FLAG_WRITE
) {
2983 if (!huge_pte_write(entry
)) {
2984 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2985 pagecache_page
, ptl
);
2988 entry
= huge_pte_mkdirty(entry
);
2990 entry
= pte_mkyoung(entry
);
2991 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2992 flags
& FAULT_FLAG_WRITE
))
2993 update_mmu_cache(vma
, address
, ptep
);
2998 if (pagecache_page
) {
2999 unlock_page(pagecache_page
);
3000 put_page(pagecache_page
);
3002 if (page
!= pagecache_page
)
3007 mutex_unlock(&hugetlb_instantiation_mutex
);
3012 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3013 struct page
**pages
, struct vm_area_struct
**vmas
,
3014 unsigned long *position
, unsigned long *nr_pages
,
3015 long i
, unsigned int flags
)
3017 unsigned long pfn_offset
;
3018 unsigned long vaddr
= *position
;
3019 unsigned long remainder
= *nr_pages
;
3020 struct hstate
*h
= hstate_vma(vma
);
3022 while (vaddr
< vma
->vm_end
&& remainder
) {
3024 spinlock_t
*ptl
= NULL
;
3029 * Some archs (sparc64, sh*) have multiple pte_ts to
3030 * each hugepage. We have to make sure we get the
3031 * first, for the page indexing below to work.
3033 * Note that page table lock is not held when pte is null.
3035 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3037 ptl
= huge_pte_lock(h
, mm
, pte
);
3038 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3041 * When coredumping, it suits get_dump_page if we just return
3042 * an error where there's an empty slot with no huge pagecache
3043 * to back it. This way, we avoid allocating a hugepage, and
3044 * the sparse dumpfile avoids allocating disk blocks, but its
3045 * huge holes still show up with zeroes where they need to be.
3047 if (absent
&& (flags
& FOLL_DUMP
) &&
3048 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3056 * We need call hugetlb_fault for both hugepages under migration
3057 * (in which case hugetlb_fault waits for the migration,) and
3058 * hwpoisoned hugepages (in which case we need to prevent the
3059 * caller from accessing to them.) In order to do this, we use
3060 * here is_swap_pte instead of is_hugetlb_entry_migration and
3061 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3062 * both cases, and because we can't follow correct pages
3063 * directly from any kind of swap entries.
3065 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3066 ((flags
& FOLL_WRITE
) &&
3067 !huge_pte_write(huge_ptep_get(pte
)))) {
3072 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3073 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3074 if (!(ret
& VM_FAULT_ERROR
))
3081 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3082 page
= pte_page(huge_ptep_get(pte
));
3085 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3086 get_page_foll(pages
[i
]);
3096 if (vaddr
< vma
->vm_end
&& remainder
&&
3097 pfn_offset
< pages_per_huge_page(h
)) {
3099 * We use pfn_offset to avoid touching the pageframes
3100 * of this compound page.
3106 *nr_pages
= remainder
;
3109 return i
? i
: -EFAULT
;
3112 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3113 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3115 struct mm_struct
*mm
= vma
->vm_mm
;
3116 unsigned long start
= address
;
3119 struct hstate
*h
= hstate_vma(vma
);
3120 unsigned long pages
= 0;
3122 BUG_ON(address
>= end
);
3123 flush_cache_range(vma
, address
, end
);
3125 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3126 for (; address
< end
; address
+= huge_page_size(h
)) {
3128 ptep
= huge_pte_offset(mm
, address
);
3131 ptl
= huge_pte_lock(h
, mm
, ptep
);
3132 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3137 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3138 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3139 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3140 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3141 set_huge_pte_at(mm
, address
, ptep
, pte
);
3147 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3148 * may have cleared our pud entry and done put_page on the page table:
3149 * once we release i_mmap_mutex, another task can do the final put_page
3150 * and that page table be reused and filled with junk.
3152 flush_tlb_range(vma
, start
, end
);
3153 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3155 return pages
<< h
->order
;
3158 int hugetlb_reserve_pages(struct inode
*inode
,
3160 struct vm_area_struct
*vma
,
3161 vm_flags_t vm_flags
)
3164 struct hstate
*h
= hstate_inode(inode
);
3165 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3166 struct resv_map
*resv_map
;
3169 * Only apply hugepage reservation if asked. At fault time, an
3170 * attempt will be made for VM_NORESERVE to allocate a page
3171 * without using reserves
3173 if (vm_flags
& VM_NORESERVE
)
3177 * Shared mappings base their reservation on the number of pages that
3178 * are already allocated on behalf of the file. Private mappings need
3179 * to reserve the full area even if read-only as mprotect() may be
3180 * called to make the mapping read-write. Assume !vma is a shm mapping
3182 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
3183 resv_map
= inode
->i_mapping
->private_data
;
3185 chg
= region_chg(resv_map
, from
, to
);
3188 resv_map
= resv_map_alloc();
3194 set_vma_resv_map(vma
, resv_map
);
3195 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3203 /* There must be enough pages in the subpool for the mapping */
3204 if (hugepage_subpool_get_pages(spool
, chg
)) {
3210 * Check enough hugepages are available for the reservation.
3211 * Hand the pages back to the subpool if there are not
3213 ret
= hugetlb_acct_memory(h
, chg
);
3215 hugepage_subpool_put_pages(spool
, chg
);
3220 * Account for the reservations made. Shared mappings record regions
3221 * that have reservations as they are shared by multiple VMAs.
3222 * When the last VMA disappears, the region map says how much
3223 * the reservation was and the page cache tells how much of
3224 * the reservation was consumed. Private mappings are per-VMA and
3225 * only the consumed reservations are tracked. When the VMA
3226 * disappears, the original reservation is the VMA size and the
3227 * consumed reservations are stored in the map. Hence, nothing
3228 * else has to be done for private mappings here
3230 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3231 region_add(resv_map
, from
, to
);
3239 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3241 struct hstate
*h
= hstate_inode(inode
);
3242 struct resv_map
*resv_map
= inode
->i_mapping
->private_data
;
3244 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3247 chg
= region_truncate(resv_map
, offset
);
3248 spin_lock(&inode
->i_lock
);
3249 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3250 spin_unlock(&inode
->i_lock
);
3252 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3253 hugetlb_acct_memory(h
, -(chg
- freed
));
3256 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3257 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3258 struct vm_area_struct
*vma
,
3259 unsigned long addr
, pgoff_t idx
)
3261 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3263 unsigned long sbase
= saddr
& PUD_MASK
;
3264 unsigned long s_end
= sbase
+ PUD_SIZE
;
3266 /* Allow segments to share if only one is marked locked */
3267 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3268 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3271 * match the virtual addresses, permission and the alignment of the
3274 if (pmd_index(addr
) != pmd_index(saddr
) ||
3275 vm_flags
!= svm_flags
||
3276 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3282 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3284 unsigned long base
= addr
& PUD_MASK
;
3285 unsigned long end
= base
+ PUD_SIZE
;
3288 * check on proper vm_flags and page table alignment
3290 if (vma
->vm_flags
& VM_MAYSHARE
&&
3291 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3297 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3298 * and returns the corresponding pte. While this is not necessary for the
3299 * !shared pmd case because we can allocate the pmd later as well, it makes the
3300 * code much cleaner. pmd allocation is essential for the shared case because
3301 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3302 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3303 * bad pmd for sharing.
3305 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3307 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3308 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3309 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3311 struct vm_area_struct
*svma
;
3312 unsigned long saddr
;
3317 if (!vma_shareable(vma
, addr
))
3318 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3320 mutex_lock(&mapping
->i_mmap_mutex
);
3321 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3325 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3327 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3329 get_page(virt_to_page(spte
));
3338 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
3341 pud_populate(mm
, pud
,
3342 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3344 put_page(virt_to_page(spte
));
3347 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3348 mutex_unlock(&mapping
->i_mmap_mutex
);
3353 * unmap huge page backed by shared pte.
3355 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3356 * indicated by page_count > 1, unmap is achieved by clearing pud and
3357 * decrementing the ref count. If count == 1, the pte page is not shared.
3359 * called with page table lock held.
3361 * returns: 1 successfully unmapped a shared pte page
3362 * 0 the underlying pte page is not shared, or it is the last user
3364 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3366 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3367 pud_t
*pud
= pud_offset(pgd
, *addr
);
3369 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3370 if (page_count(virt_to_page(ptep
)) == 1)
3374 put_page(virt_to_page(ptep
));
3375 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3378 #define want_pmd_share() (1)
3379 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3380 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3384 #define want_pmd_share() (0)
3385 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3387 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3388 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3389 unsigned long addr
, unsigned long sz
)
3395 pgd
= pgd_offset(mm
, addr
);
3396 pud
= pud_alloc(mm
, pgd
, addr
);
3398 if (sz
== PUD_SIZE
) {
3401 BUG_ON(sz
!= PMD_SIZE
);
3402 if (want_pmd_share() && pud_none(*pud
))
3403 pte
= huge_pmd_share(mm
, addr
, pud
);
3405 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3408 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3413 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3419 pgd
= pgd_offset(mm
, addr
);
3420 if (pgd_present(*pgd
)) {
3421 pud
= pud_offset(pgd
, addr
);
3422 if (pud_present(*pud
)) {
3424 return (pte_t
*)pud
;
3425 pmd
= pmd_offset(pud
, addr
);
3428 return (pte_t
*) pmd
;
3432 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3433 pmd_t
*pmd
, int write
)
3437 page
= pte_page(*(pte_t
*)pmd
);
3439 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3444 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3445 pud_t
*pud
, int write
)
3449 page
= pte_page(*(pte_t
*)pud
);
3451 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3455 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3457 /* Can be overriden by architectures */
3458 __attribute__((weak
)) struct page
*
3459 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3460 pud_t
*pud
, int write
)
3466 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3468 #ifdef CONFIG_MEMORY_FAILURE
3470 /* Should be called in hugetlb_lock */
3471 static int is_hugepage_on_freelist(struct page
*hpage
)
3475 struct hstate
*h
= page_hstate(hpage
);
3476 int nid
= page_to_nid(hpage
);
3478 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3485 * This function is called from memory failure code.
3486 * Assume the caller holds page lock of the head page.
3488 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3490 struct hstate
*h
= page_hstate(hpage
);
3491 int nid
= page_to_nid(hpage
);
3494 spin_lock(&hugetlb_lock
);
3495 if (is_hugepage_on_freelist(hpage
)) {
3497 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3498 * but dangling hpage->lru can trigger list-debug warnings
3499 * (this happens when we call unpoison_memory() on it),
3500 * so let it point to itself with list_del_init().
3502 list_del_init(&hpage
->lru
);
3503 set_page_refcounted(hpage
);
3504 h
->free_huge_pages
--;
3505 h
->free_huge_pages_node
[nid
]--;
3508 spin_unlock(&hugetlb_lock
);
3513 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3515 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3516 if (!get_page_unless_zero(page
))
3518 spin_lock(&hugetlb_lock
);
3519 list_move_tail(&page
->lru
, list
);
3520 spin_unlock(&hugetlb_lock
);
3524 void putback_active_hugepage(struct page
*page
)
3526 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3527 spin_lock(&hugetlb_lock
);
3528 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3529 spin_unlock(&hugetlb_lock
);
3533 bool is_hugepage_active(struct page
*page
)
3535 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
3537 * This function can be called for a tail page because the caller,
3538 * scan_movable_pages, scans through a given pfn-range which typically
3539 * covers one memory block. In systems using gigantic hugepage (1GB
3540 * for x86_64,) a hugepage is larger than a memory block, and we don't
3541 * support migrating such large hugepages for now, so return false
3542 * when called for tail pages.
3547 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3548 * so we should return false for them.
3550 if (unlikely(PageHWPoison(page
)))
3552 return page_count(page
) > 0;