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 embedded into a resv_map and
139 * protected by a resv_map's lock
142 struct list_head link
;
147 static long region_add(struct resv_map
*resv
, long f
, long t
)
149 struct list_head
*head
= &resv
->regions
;
150 struct file_region
*rg
, *nrg
, *trg
;
152 spin_lock(&resv
->lock
);
153 /* Locate the region we are either in or before. */
154 list_for_each_entry(rg
, head
, link
)
158 /* Round our left edge to the current segment if it encloses us. */
162 /* Check for and consume any regions we now overlap with. */
164 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
165 if (&rg
->link
== head
)
170 /* If this area reaches higher then extend our area to
171 * include it completely. If this is not the first area
172 * which we intend to reuse, free it. */
182 spin_unlock(&resv
->lock
);
186 static long region_chg(struct resv_map
*resv
, long f
, long t
)
188 struct list_head
*head
= &resv
->regions
;
189 struct file_region
*rg
, *nrg
= NULL
;
193 spin_lock(&resv
->lock
);
194 /* Locate the region we are before or in. */
195 list_for_each_entry(rg
, head
, link
)
199 /* If we are below the current region then a new region is required.
200 * Subtle, allocate a new region at the position but make it zero
201 * size such that we can guarantee to record the reservation. */
202 if (&rg
->link
== head
|| t
< rg
->from
) {
204 spin_unlock(&resv
->lock
);
205 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
211 INIT_LIST_HEAD(&nrg
->link
);
215 list_add(&nrg
->link
, rg
->link
.prev
);
220 /* Round our left edge to the current segment if it encloses us. */
225 /* Check for and consume any regions we now overlap with. */
226 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
227 if (&rg
->link
== head
)
232 /* We overlap with this area, if it extends further than
233 * us then we must extend ourselves. Account for its
234 * existing reservation. */
239 chg
-= rg
->to
- rg
->from
;
243 spin_unlock(&resv
->lock
);
244 /* We already know we raced and no longer need the new region */
248 spin_unlock(&resv
->lock
);
252 static long region_truncate(struct resv_map
*resv
, long end
)
254 struct list_head
*head
= &resv
->regions
;
255 struct file_region
*rg
, *trg
;
258 spin_lock(&resv
->lock
);
259 /* Locate the region we are either in or before. */
260 list_for_each_entry(rg
, head
, link
)
263 if (&rg
->link
== head
)
266 /* If we are in the middle of a region then adjust it. */
267 if (end
> rg
->from
) {
270 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
273 /* Drop any remaining regions. */
274 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
275 if (&rg
->link
== head
)
277 chg
+= rg
->to
- rg
->from
;
283 spin_unlock(&resv
->lock
);
287 static long region_count(struct resv_map
*resv
, long f
, long t
)
289 struct list_head
*head
= &resv
->regions
;
290 struct file_region
*rg
;
293 spin_lock(&resv
->lock
);
294 /* Locate each segment we overlap with, and count that overlap. */
295 list_for_each_entry(rg
, head
, link
) {
304 seg_from
= max(rg
->from
, f
);
305 seg_to
= min(rg
->to
, t
);
307 chg
+= seg_to
- seg_from
;
309 spin_unlock(&resv
->lock
);
315 * Convert the address within this vma to the page offset within
316 * the mapping, in pagecache page units; huge pages here.
318 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
319 struct vm_area_struct
*vma
, unsigned long address
)
321 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
322 (vma
->vm_pgoff
>> huge_page_order(h
));
325 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
326 unsigned long address
)
328 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
332 * Return the size of the pages allocated when backing a VMA. In the majority
333 * cases this will be same size as used by the page table entries.
335 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
337 struct hstate
*hstate
;
339 if (!is_vm_hugetlb_page(vma
))
342 hstate
= hstate_vma(vma
);
344 return 1UL << huge_page_shift(hstate
);
346 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
349 * Return the page size being used by the MMU to back a VMA. In the majority
350 * of cases, the page size used by the kernel matches the MMU size. On
351 * architectures where it differs, an architecture-specific version of this
352 * function is required.
354 #ifndef vma_mmu_pagesize
355 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
357 return vma_kernel_pagesize(vma
);
362 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
363 * bits of the reservation map pointer, which are always clear due to
366 #define HPAGE_RESV_OWNER (1UL << 0)
367 #define HPAGE_RESV_UNMAPPED (1UL << 1)
368 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
371 * These helpers are used to track how many pages are reserved for
372 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
373 * is guaranteed to have their future faults succeed.
375 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
376 * the reserve counters are updated with the hugetlb_lock held. It is safe
377 * to reset the VMA at fork() time as it is not in use yet and there is no
378 * chance of the global counters getting corrupted as a result of the values.
380 * The private mapping reservation is represented in a subtly different
381 * manner to a shared mapping. A shared mapping has a region map associated
382 * with the underlying file, this region map represents the backing file
383 * pages which have ever had a reservation assigned which this persists even
384 * after the page is instantiated. A private mapping has a region map
385 * associated with the original mmap which is attached to all VMAs which
386 * reference it, this region map represents those offsets which have consumed
387 * reservation ie. where pages have been instantiated.
389 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
391 return (unsigned long)vma
->vm_private_data
;
394 static void set_vma_private_data(struct vm_area_struct
*vma
,
397 vma
->vm_private_data
= (void *)value
;
400 struct resv_map
*resv_map_alloc(void)
402 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
406 kref_init(&resv_map
->refs
);
407 spin_lock_init(&resv_map
->lock
);
408 INIT_LIST_HEAD(&resv_map
->regions
);
413 void resv_map_release(struct kref
*ref
)
415 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
417 /* Clear out any active regions before we release the map. */
418 region_truncate(resv_map
, 0);
422 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
424 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
425 if (!(vma
->vm_flags
& VM_MAYSHARE
))
426 return (struct resv_map
*)(get_vma_private_data(vma
) &
431 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
434 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
436 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
437 HPAGE_RESV_MASK
) | (unsigned long)map
);
440 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
442 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
443 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
445 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
448 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
450 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
452 return (get_vma_private_data(vma
) & flag
) != 0;
455 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
456 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
458 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
459 if (!(vma
->vm_flags
& VM_MAYSHARE
))
460 vma
->vm_private_data
= (void *)0;
463 /* Returns true if the VMA has associated reserve pages */
464 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
466 if (vma
->vm_flags
& VM_NORESERVE
) {
468 * This address is already reserved by other process(chg == 0),
469 * so, we should decrement reserved count. Without decrementing,
470 * reserve count remains after releasing inode, because this
471 * allocated page will go into page cache and is regarded as
472 * coming from reserved pool in releasing step. Currently, we
473 * don't have any other solution to deal with this situation
474 * properly, so add work-around here.
476 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
482 /* Shared mappings always use reserves */
483 if (vma
->vm_flags
& VM_MAYSHARE
)
487 * Only the process that called mmap() has reserves for
490 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
496 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
498 int nid
= page_to_nid(page
);
499 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
500 h
->free_huge_pages
++;
501 h
->free_huge_pages_node
[nid
]++;
504 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
508 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
509 if (!is_migrate_isolate_page(page
))
512 * if 'non-isolated free hugepage' not found on the list,
513 * the allocation fails.
515 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
517 list_move(&page
->lru
, &h
->hugepage_activelist
);
518 set_page_refcounted(page
);
519 h
->free_huge_pages
--;
520 h
->free_huge_pages_node
[nid
]--;
524 /* Movability of hugepages depends on migration support. */
525 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
527 if (hugepages_treat_as_movable
|| hugepage_migration_support(h
))
528 return GFP_HIGHUSER_MOVABLE
;
533 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
534 struct vm_area_struct
*vma
,
535 unsigned long address
, int avoid_reserve
,
538 struct page
*page
= NULL
;
539 struct mempolicy
*mpol
;
540 nodemask_t
*nodemask
;
541 struct zonelist
*zonelist
;
544 unsigned int cpuset_mems_cookie
;
547 * A child process with MAP_PRIVATE mappings created by their parent
548 * have no page reserves. This check ensures that reservations are
549 * not "stolen". The child may still get SIGKILLed
551 if (!vma_has_reserves(vma
, chg
) &&
552 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
555 /* If reserves cannot be used, ensure enough pages are in the pool */
556 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
560 cpuset_mems_cookie
= read_mems_allowed_begin();
561 zonelist
= huge_zonelist(vma
, address
,
562 htlb_alloc_mask(h
), &mpol
, &nodemask
);
564 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
565 MAX_NR_ZONES
- 1, nodemask
) {
566 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask(h
))) {
567 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
571 if (!vma_has_reserves(vma
, chg
))
574 SetPagePrivate(page
);
575 h
->resv_huge_pages
--;
582 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
590 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
594 VM_BUG_ON(h
->order
>= MAX_ORDER
);
597 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
598 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
599 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
600 1 << PG_referenced
| 1 << PG_dirty
|
601 1 << PG_active
| 1 << PG_reserved
|
602 1 << PG_private
| 1 << PG_writeback
);
604 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
605 set_compound_page_dtor(page
, NULL
);
606 set_page_refcounted(page
);
607 arch_release_hugepage(page
);
608 __free_pages(page
, huge_page_order(h
));
611 struct hstate
*size_to_hstate(unsigned long size
)
616 if (huge_page_size(h
) == size
)
622 static void free_huge_page(struct page
*page
)
625 * Can't pass hstate in here because it is called from the
626 * compound page destructor.
628 struct hstate
*h
= page_hstate(page
);
629 int nid
= page_to_nid(page
);
630 struct hugepage_subpool
*spool
=
631 (struct hugepage_subpool
*)page_private(page
);
632 bool restore_reserve
;
634 set_page_private(page
, 0);
635 page
->mapping
= NULL
;
636 BUG_ON(page_count(page
));
637 BUG_ON(page_mapcount(page
));
638 restore_reserve
= PagePrivate(page
);
639 ClearPagePrivate(page
);
641 spin_lock(&hugetlb_lock
);
642 hugetlb_cgroup_uncharge_page(hstate_index(h
),
643 pages_per_huge_page(h
), page
);
645 h
->resv_huge_pages
++;
647 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
648 /* remove the page from active list */
649 list_del(&page
->lru
);
650 update_and_free_page(h
, page
);
651 h
->surplus_huge_pages
--;
652 h
->surplus_huge_pages_node
[nid
]--;
654 arch_clear_hugepage_flags(page
);
655 enqueue_huge_page(h
, page
);
657 spin_unlock(&hugetlb_lock
);
658 hugepage_subpool_put_pages(spool
, 1);
661 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
663 INIT_LIST_HEAD(&page
->lru
);
664 set_compound_page_dtor(page
, free_huge_page
);
665 spin_lock(&hugetlb_lock
);
666 set_hugetlb_cgroup(page
, NULL
);
668 h
->nr_huge_pages_node
[nid
]++;
669 spin_unlock(&hugetlb_lock
);
670 put_page(page
); /* free it into the hugepage allocator */
673 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
676 int nr_pages
= 1 << order
;
677 struct page
*p
= page
+ 1;
679 /* we rely on prep_new_huge_page to set the destructor */
680 set_compound_order(page
, order
);
682 __ClearPageReserved(page
);
683 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
686 * For gigantic hugepages allocated through bootmem at
687 * boot, it's safer to be consistent with the not-gigantic
688 * hugepages and clear the PG_reserved bit from all tail pages
689 * too. Otherwse drivers using get_user_pages() to access tail
690 * pages may get the reference counting wrong if they see
691 * PG_reserved set on a tail page (despite the head page not
692 * having PG_reserved set). Enforcing this consistency between
693 * head and tail pages allows drivers to optimize away a check
694 * on the head page when they need know if put_page() is needed
695 * after get_user_pages().
697 __ClearPageReserved(p
);
698 set_page_count(p
, 0);
699 p
->first_page
= page
;
704 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
705 * transparent huge pages. See the PageTransHuge() documentation for more
708 int PageHuge(struct page
*page
)
710 if (!PageCompound(page
))
713 page
= compound_head(page
);
714 return get_compound_page_dtor(page
) == free_huge_page
;
716 EXPORT_SYMBOL_GPL(PageHuge
);
719 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
720 * normal or transparent huge pages.
722 int PageHeadHuge(struct page
*page_head
)
724 if (!PageHead(page_head
))
727 return get_compound_page_dtor(page_head
) == free_huge_page
;
730 pgoff_t
__basepage_index(struct page
*page
)
732 struct page
*page_head
= compound_head(page
);
733 pgoff_t index
= page_index(page_head
);
734 unsigned long compound_idx
;
736 if (!PageHuge(page_head
))
737 return page_index(page
);
739 if (compound_order(page_head
) >= MAX_ORDER
)
740 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
742 compound_idx
= page
- page_head
;
744 return (index
<< compound_order(page_head
)) + compound_idx
;
747 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
751 if (h
->order
>= MAX_ORDER
)
754 page
= alloc_pages_exact_node(nid
,
755 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
756 __GFP_REPEAT
|__GFP_NOWARN
,
759 if (arch_prepare_hugepage(page
)) {
760 __free_pages(page
, huge_page_order(h
));
763 prep_new_huge_page(h
, page
, nid
);
770 * common helper functions for hstate_next_node_to_{alloc|free}.
771 * We may have allocated or freed a huge page based on a different
772 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
773 * be outside of *nodes_allowed. Ensure that we use an allowed
774 * node for alloc or free.
776 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
778 nid
= next_node(nid
, *nodes_allowed
);
779 if (nid
== MAX_NUMNODES
)
780 nid
= first_node(*nodes_allowed
);
781 VM_BUG_ON(nid
>= MAX_NUMNODES
);
786 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
788 if (!node_isset(nid
, *nodes_allowed
))
789 nid
= next_node_allowed(nid
, nodes_allowed
);
794 * returns the previously saved node ["this node"] from which to
795 * allocate a persistent huge page for the pool and advance the
796 * next node from which to allocate, handling wrap at end of node
799 static int hstate_next_node_to_alloc(struct hstate
*h
,
800 nodemask_t
*nodes_allowed
)
804 VM_BUG_ON(!nodes_allowed
);
806 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
807 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
813 * helper for free_pool_huge_page() - return the previously saved
814 * node ["this node"] from which to free a huge page. Advance the
815 * next node id whether or not we find a free huge page to free so
816 * that the next attempt to free addresses the next node.
818 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
822 VM_BUG_ON(!nodes_allowed
);
824 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
825 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
830 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
831 for (nr_nodes = nodes_weight(*mask); \
833 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
836 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
837 for (nr_nodes = nodes_weight(*mask); \
839 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
842 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
848 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
849 page
= alloc_fresh_huge_page_node(h
, node
);
857 count_vm_event(HTLB_BUDDY_PGALLOC
);
859 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
865 * Free huge page from pool from next node to free.
866 * Attempt to keep persistent huge pages more or less
867 * balanced over allowed nodes.
868 * Called with hugetlb_lock locked.
870 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
876 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
878 * If we're returning unused surplus pages, only examine
879 * nodes with surplus pages.
881 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
882 !list_empty(&h
->hugepage_freelists
[node
])) {
884 list_entry(h
->hugepage_freelists
[node
].next
,
886 list_del(&page
->lru
);
887 h
->free_huge_pages
--;
888 h
->free_huge_pages_node
[node
]--;
890 h
->surplus_huge_pages
--;
891 h
->surplus_huge_pages_node
[node
]--;
893 update_and_free_page(h
, page
);
903 * Dissolve a given free hugepage into free buddy pages. This function does
904 * nothing for in-use (including surplus) hugepages.
906 static void dissolve_free_huge_page(struct page
*page
)
908 spin_lock(&hugetlb_lock
);
909 if (PageHuge(page
) && !page_count(page
)) {
910 struct hstate
*h
= page_hstate(page
);
911 int nid
= page_to_nid(page
);
912 list_del(&page
->lru
);
913 h
->free_huge_pages
--;
914 h
->free_huge_pages_node
[nid
]--;
915 update_and_free_page(h
, page
);
917 spin_unlock(&hugetlb_lock
);
921 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
922 * make specified memory blocks removable from the system.
923 * Note that start_pfn should aligned with (minimum) hugepage size.
925 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
927 unsigned int order
= 8 * sizeof(void *);
931 /* Set scan step to minimum hugepage size */
933 if (order
> huge_page_order(h
))
934 order
= huge_page_order(h
);
935 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
936 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
937 dissolve_free_huge_page(pfn_to_page(pfn
));
940 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
945 if (h
->order
>= MAX_ORDER
)
949 * Assume we will successfully allocate the surplus page to
950 * prevent racing processes from causing the surplus to exceed
953 * This however introduces a different race, where a process B
954 * tries to grow the static hugepage pool while alloc_pages() is
955 * called by process A. B will only examine the per-node
956 * counters in determining if surplus huge pages can be
957 * converted to normal huge pages in adjust_pool_surplus(). A
958 * won't be able to increment the per-node counter, until the
959 * lock is dropped by B, but B doesn't drop hugetlb_lock until
960 * no more huge pages can be converted from surplus to normal
961 * state (and doesn't try to convert again). Thus, we have a
962 * case where a surplus huge page exists, the pool is grown, and
963 * the surplus huge page still exists after, even though it
964 * should just have been converted to a normal huge page. This
965 * does not leak memory, though, as the hugepage will be freed
966 * once it is out of use. It also does not allow the counters to
967 * go out of whack in adjust_pool_surplus() as we don't modify
968 * the node values until we've gotten the hugepage and only the
969 * per-node value is checked there.
971 spin_lock(&hugetlb_lock
);
972 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
973 spin_unlock(&hugetlb_lock
);
977 h
->surplus_huge_pages
++;
979 spin_unlock(&hugetlb_lock
);
981 if (nid
== NUMA_NO_NODE
)
982 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
983 __GFP_REPEAT
|__GFP_NOWARN
,
986 page
= alloc_pages_exact_node(nid
,
987 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
988 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
990 if (page
&& arch_prepare_hugepage(page
)) {
991 __free_pages(page
, huge_page_order(h
));
995 spin_lock(&hugetlb_lock
);
997 INIT_LIST_HEAD(&page
->lru
);
998 r_nid
= page_to_nid(page
);
999 set_compound_page_dtor(page
, free_huge_page
);
1000 set_hugetlb_cgroup(page
, NULL
);
1002 * We incremented the global counters already
1004 h
->nr_huge_pages_node
[r_nid
]++;
1005 h
->surplus_huge_pages_node
[r_nid
]++;
1006 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1009 h
->surplus_huge_pages
--;
1010 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1012 spin_unlock(&hugetlb_lock
);
1018 * This allocation function is useful in the context where vma is irrelevant.
1019 * E.g. soft-offlining uses this function because it only cares physical
1020 * address of error page.
1022 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1024 struct page
*page
= NULL
;
1026 spin_lock(&hugetlb_lock
);
1027 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1028 page
= dequeue_huge_page_node(h
, nid
);
1029 spin_unlock(&hugetlb_lock
);
1032 page
= alloc_buddy_huge_page(h
, nid
);
1038 * Increase the hugetlb pool such that it can accommodate a reservation
1041 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1043 struct list_head surplus_list
;
1044 struct page
*page
, *tmp
;
1046 int needed
, allocated
;
1047 bool alloc_ok
= true;
1049 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1051 h
->resv_huge_pages
+= delta
;
1056 INIT_LIST_HEAD(&surplus_list
);
1060 spin_unlock(&hugetlb_lock
);
1061 for (i
= 0; i
< needed
; i
++) {
1062 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1067 list_add(&page
->lru
, &surplus_list
);
1072 * After retaking hugetlb_lock, we need to recalculate 'needed'
1073 * because either resv_huge_pages or free_huge_pages may have changed.
1075 spin_lock(&hugetlb_lock
);
1076 needed
= (h
->resv_huge_pages
+ delta
) -
1077 (h
->free_huge_pages
+ allocated
);
1082 * We were not able to allocate enough pages to
1083 * satisfy the entire reservation so we free what
1084 * we've allocated so far.
1089 * The surplus_list now contains _at_least_ the number of extra pages
1090 * needed to accommodate the reservation. Add the appropriate number
1091 * of pages to the hugetlb pool and free the extras back to the buddy
1092 * allocator. Commit the entire reservation here to prevent another
1093 * process from stealing the pages as they are added to the pool but
1094 * before they are reserved.
1096 needed
+= allocated
;
1097 h
->resv_huge_pages
+= delta
;
1100 /* Free the needed pages to the hugetlb pool */
1101 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1105 * This page is now managed by the hugetlb allocator and has
1106 * no users -- drop the buddy allocator's reference.
1108 put_page_testzero(page
);
1109 VM_BUG_ON_PAGE(page_count(page
), page
);
1110 enqueue_huge_page(h
, page
);
1113 spin_unlock(&hugetlb_lock
);
1115 /* Free unnecessary surplus pages to the buddy allocator */
1116 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1118 spin_lock(&hugetlb_lock
);
1124 * When releasing a hugetlb pool reservation, any surplus pages that were
1125 * allocated to satisfy the reservation must be explicitly freed if they were
1127 * Called with hugetlb_lock held.
1129 static void return_unused_surplus_pages(struct hstate
*h
,
1130 unsigned long unused_resv_pages
)
1132 unsigned long nr_pages
;
1134 /* Uncommit the reservation */
1135 h
->resv_huge_pages
-= unused_resv_pages
;
1137 /* Cannot return gigantic pages currently */
1138 if (h
->order
>= MAX_ORDER
)
1141 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1144 * We want to release as many surplus pages as possible, spread
1145 * evenly across all nodes with memory. Iterate across these nodes
1146 * until we can no longer free unreserved surplus pages. This occurs
1147 * when the nodes with surplus pages have no free pages.
1148 * free_pool_huge_page() will balance the the freed pages across the
1149 * on-line nodes with memory and will handle the hstate accounting.
1151 while (nr_pages
--) {
1152 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1158 * Determine if the huge page at addr within the vma has an associated
1159 * reservation. Where it does not we will need to logically increase
1160 * reservation and actually increase subpool usage before an allocation
1161 * can occur. Where any new reservation would be required the
1162 * reservation change is prepared, but not committed. Once the page
1163 * has been allocated from the subpool and instantiated the change should
1164 * be committed via vma_commit_reservation. No action is required on
1167 static long vma_needs_reservation(struct hstate
*h
,
1168 struct vm_area_struct
*vma
, unsigned long addr
)
1170 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1171 struct inode
*inode
= mapping
->host
;
1173 if (vma
->vm_flags
& VM_MAYSHARE
) {
1174 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1175 struct resv_map
*resv
= inode
->i_mapping
->private_data
;
1177 return region_chg(resv
, idx
, idx
+ 1);
1179 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1184 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1185 struct resv_map
*resv
= vma_resv_map(vma
);
1187 err
= region_chg(resv
, idx
, idx
+ 1);
1193 static void vma_commit_reservation(struct hstate
*h
,
1194 struct vm_area_struct
*vma
, unsigned long addr
)
1196 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1197 struct inode
*inode
= mapping
->host
;
1199 if (vma
->vm_flags
& VM_MAYSHARE
) {
1200 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1201 struct resv_map
*resv
= inode
->i_mapping
->private_data
;
1203 region_add(resv
, idx
, idx
+ 1);
1205 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1206 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1207 struct resv_map
*resv
= vma_resv_map(vma
);
1209 /* Mark this page used in the map. */
1210 region_add(resv
, idx
, idx
+ 1);
1214 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1215 unsigned long addr
, int avoid_reserve
)
1217 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1218 struct hstate
*h
= hstate_vma(vma
);
1222 struct hugetlb_cgroup
*h_cg
;
1224 idx
= hstate_index(h
);
1226 * Processes that did not create the mapping will have no
1227 * reserves and will not have accounted against subpool
1228 * limit. Check that the subpool limit can be made before
1229 * satisfying the allocation MAP_NORESERVE mappings may also
1230 * need pages and subpool limit allocated allocated if no reserve
1233 chg
= vma_needs_reservation(h
, vma
, addr
);
1235 return ERR_PTR(-ENOMEM
);
1236 if (chg
|| avoid_reserve
)
1237 if (hugepage_subpool_get_pages(spool
, 1))
1238 return ERR_PTR(-ENOSPC
);
1240 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1242 if (chg
|| avoid_reserve
)
1243 hugepage_subpool_put_pages(spool
, 1);
1244 return ERR_PTR(-ENOSPC
);
1246 spin_lock(&hugetlb_lock
);
1247 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1249 spin_unlock(&hugetlb_lock
);
1250 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1252 hugetlb_cgroup_uncharge_cgroup(idx
,
1253 pages_per_huge_page(h
),
1255 if (chg
|| avoid_reserve
)
1256 hugepage_subpool_put_pages(spool
, 1);
1257 return ERR_PTR(-ENOSPC
);
1259 spin_lock(&hugetlb_lock
);
1260 list_move(&page
->lru
, &h
->hugepage_activelist
);
1263 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1264 spin_unlock(&hugetlb_lock
);
1266 set_page_private(page
, (unsigned long)spool
);
1268 vma_commit_reservation(h
, vma
, addr
);
1273 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1274 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1275 * where no ERR_VALUE is expected to be returned.
1277 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1278 unsigned long addr
, int avoid_reserve
)
1280 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1286 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1288 struct huge_bootmem_page
*m
;
1291 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1294 addr
= memblock_virt_alloc_try_nid_nopanic(
1295 huge_page_size(h
), huge_page_size(h
),
1296 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1299 * Use the beginning of the huge page to store the
1300 * huge_bootmem_page struct (until gather_bootmem
1301 * puts them into the mem_map).
1310 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1311 /* Put them into a private list first because mem_map is not up yet */
1312 list_add(&m
->list
, &huge_boot_pages
);
1317 static void prep_compound_huge_page(struct page
*page
, int order
)
1319 if (unlikely(order
> (MAX_ORDER
- 1)))
1320 prep_compound_gigantic_page(page
, order
);
1322 prep_compound_page(page
, order
);
1325 /* Put bootmem huge pages into the standard lists after mem_map is up */
1326 static void __init
gather_bootmem_prealloc(void)
1328 struct huge_bootmem_page
*m
;
1330 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1331 struct hstate
*h
= m
->hstate
;
1334 #ifdef CONFIG_HIGHMEM
1335 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1336 memblock_free_late(__pa(m
),
1337 sizeof(struct huge_bootmem_page
));
1339 page
= virt_to_page(m
);
1341 WARN_ON(page_count(page
) != 1);
1342 prep_compound_huge_page(page
, h
->order
);
1343 WARN_ON(PageReserved(page
));
1344 prep_new_huge_page(h
, page
, page_to_nid(page
));
1346 * If we had gigantic hugepages allocated at boot time, we need
1347 * to restore the 'stolen' pages to totalram_pages in order to
1348 * fix confusing memory reports from free(1) and another
1349 * side-effects, like CommitLimit going negative.
1351 if (h
->order
> (MAX_ORDER
- 1))
1352 adjust_managed_page_count(page
, 1 << h
->order
);
1356 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1360 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1361 if (h
->order
>= MAX_ORDER
) {
1362 if (!alloc_bootmem_huge_page(h
))
1364 } else if (!alloc_fresh_huge_page(h
,
1365 &node_states
[N_MEMORY
]))
1368 h
->max_huge_pages
= i
;
1371 static void __init
hugetlb_init_hstates(void)
1375 for_each_hstate(h
) {
1376 /* oversize hugepages were init'ed in early boot */
1377 if (h
->order
< MAX_ORDER
)
1378 hugetlb_hstate_alloc_pages(h
);
1382 static char * __init
memfmt(char *buf
, unsigned long n
)
1384 if (n
>= (1UL << 30))
1385 sprintf(buf
, "%lu GB", n
>> 30);
1386 else if (n
>= (1UL << 20))
1387 sprintf(buf
, "%lu MB", n
>> 20);
1389 sprintf(buf
, "%lu KB", n
>> 10);
1393 static void __init
report_hugepages(void)
1397 for_each_hstate(h
) {
1399 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1400 memfmt(buf
, huge_page_size(h
)),
1401 h
->free_huge_pages
);
1405 #ifdef CONFIG_HIGHMEM
1406 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1407 nodemask_t
*nodes_allowed
)
1411 if (h
->order
>= MAX_ORDER
)
1414 for_each_node_mask(i
, *nodes_allowed
) {
1415 struct page
*page
, *next
;
1416 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1417 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1418 if (count
>= h
->nr_huge_pages
)
1420 if (PageHighMem(page
))
1422 list_del(&page
->lru
);
1423 update_and_free_page(h
, page
);
1424 h
->free_huge_pages
--;
1425 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1430 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1431 nodemask_t
*nodes_allowed
)
1437 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1438 * balanced by operating on them in a round-robin fashion.
1439 * Returns 1 if an adjustment was made.
1441 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1446 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1449 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1450 if (h
->surplus_huge_pages_node
[node
])
1454 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1455 if (h
->surplus_huge_pages_node
[node
] <
1456 h
->nr_huge_pages_node
[node
])
1463 h
->surplus_huge_pages
+= delta
;
1464 h
->surplus_huge_pages_node
[node
] += delta
;
1468 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1469 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1470 nodemask_t
*nodes_allowed
)
1472 unsigned long min_count
, ret
;
1474 if (h
->order
>= MAX_ORDER
)
1475 return h
->max_huge_pages
;
1478 * Increase the pool size
1479 * First take pages out of surplus state. Then make up the
1480 * remaining difference by allocating fresh huge pages.
1482 * We might race with alloc_buddy_huge_page() here and be unable
1483 * to convert a surplus huge page to a normal huge page. That is
1484 * not critical, though, it just means the overall size of the
1485 * pool might be one hugepage larger than it needs to be, but
1486 * within all the constraints specified by the sysctls.
1488 spin_lock(&hugetlb_lock
);
1489 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1490 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1494 while (count
> persistent_huge_pages(h
)) {
1496 * If this allocation races such that we no longer need the
1497 * page, free_huge_page will handle it by freeing the page
1498 * and reducing the surplus.
1500 spin_unlock(&hugetlb_lock
);
1501 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1502 spin_lock(&hugetlb_lock
);
1506 /* Bail for signals. Probably ctrl-c from user */
1507 if (signal_pending(current
))
1512 * Decrease the pool size
1513 * First return free pages to the buddy allocator (being careful
1514 * to keep enough around to satisfy reservations). Then place
1515 * pages into surplus state as needed so the pool will shrink
1516 * to the desired size as pages become free.
1518 * By placing pages into the surplus state independent of the
1519 * overcommit value, we are allowing the surplus pool size to
1520 * exceed overcommit. There are few sane options here. Since
1521 * alloc_buddy_huge_page() is checking the global counter,
1522 * though, we'll note that we're not allowed to exceed surplus
1523 * and won't grow the pool anywhere else. Not until one of the
1524 * sysctls are changed, or the surplus pages go out of use.
1526 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1527 min_count
= max(count
, min_count
);
1528 try_to_free_low(h
, min_count
, nodes_allowed
);
1529 while (min_count
< persistent_huge_pages(h
)) {
1530 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1533 while (count
< persistent_huge_pages(h
)) {
1534 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1538 ret
= persistent_huge_pages(h
);
1539 spin_unlock(&hugetlb_lock
);
1543 #define HSTATE_ATTR_RO(_name) \
1544 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1546 #define HSTATE_ATTR(_name) \
1547 static struct kobj_attribute _name##_attr = \
1548 __ATTR(_name, 0644, _name##_show, _name##_store)
1550 static struct kobject
*hugepages_kobj
;
1551 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1553 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1555 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1559 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1560 if (hstate_kobjs
[i
] == kobj
) {
1562 *nidp
= NUMA_NO_NODE
;
1566 return kobj_to_node_hstate(kobj
, nidp
);
1569 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1570 struct kobj_attribute
*attr
, char *buf
)
1573 unsigned long nr_huge_pages
;
1576 h
= kobj_to_hstate(kobj
, &nid
);
1577 if (nid
== NUMA_NO_NODE
)
1578 nr_huge_pages
= h
->nr_huge_pages
;
1580 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1582 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1585 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1586 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1587 const char *buf
, size_t len
)
1591 unsigned long count
;
1593 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1595 err
= kstrtoul(buf
, 10, &count
);
1599 h
= kobj_to_hstate(kobj
, &nid
);
1600 if (h
->order
>= MAX_ORDER
) {
1605 if (nid
== NUMA_NO_NODE
) {
1607 * global hstate attribute
1609 if (!(obey_mempolicy
&&
1610 init_nodemask_of_mempolicy(nodes_allowed
))) {
1611 NODEMASK_FREE(nodes_allowed
);
1612 nodes_allowed
= &node_states
[N_MEMORY
];
1614 } else if (nodes_allowed
) {
1616 * per node hstate attribute: adjust count to global,
1617 * but restrict alloc/free to the specified node.
1619 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1620 init_nodemask_of_node(nodes_allowed
, nid
);
1622 nodes_allowed
= &node_states
[N_MEMORY
];
1624 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1626 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1627 NODEMASK_FREE(nodes_allowed
);
1631 NODEMASK_FREE(nodes_allowed
);
1635 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1636 struct kobj_attribute
*attr
, char *buf
)
1638 return nr_hugepages_show_common(kobj
, attr
, buf
);
1641 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1642 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1644 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1646 HSTATE_ATTR(nr_hugepages
);
1651 * hstate attribute for optionally mempolicy-based constraint on persistent
1652 * huge page alloc/free.
1654 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1655 struct kobj_attribute
*attr
, char *buf
)
1657 return nr_hugepages_show_common(kobj
, attr
, buf
);
1660 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1661 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1663 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1665 HSTATE_ATTR(nr_hugepages_mempolicy
);
1669 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1670 struct kobj_attribute
*attr
, char *buf
)
1672 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1673 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1676 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1677 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1680 unsigned long input
;
1681 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1683 if (h
->order
>= MAX_ORDER
)
1686 err
= kstrtoul(buf
, 10, &input
);
1690 spin_lock(&hugetlb_lock
);
1691 h
->nr_overcommit_huge_pages
= input
;
1692 spin_unlock(&hugetlb_lock
);
1696 HSTATE_ATTR(nr_overcommit_hugepages
);
1698 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1699 struct kobj_attribute
*attr
, char *buf
)
1702 unsigned long free_huge_pages
;
1705 h
= kobj_to_hstate(kobj
, &nid
);
1706 if (nid
== NUMA_NO_NODE
)
1707 free_huge_pages
= h
->free_huge_pages
;
1709 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1711 return sprintf(buf
, "%lu\n", free_huge_pages
);
1713 HSTATE_ATTR_RO(free_hugepages
);
1715 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1716 struct kobj_attribute
*attr
, char *buf
)
1718 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1719 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1721 HSTATE_ATTR_RO(resv_hugepages
);
1723 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1724 struct kobj_attribute
*attr
, char *buf
)
1727 unsigned long surplus_huge_pages
;
1730 h
= kobj_to_hstate(kobj
, &nid
);
1731 if (nid
== NUMA_NO_NODE
)
1732 surplus_huge_pages
= h
->surplus_huge_pages
;
1734 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1736 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1738 HSTATE_ATTR_RO(surplus_hugepages
);
1740 static struct attribute
*hstate_attrs
[] = {
1741 &nr_hugepages_attr
.attr
,
1742 &nr_overcommit_hugepages_attr
.attr
,
1743 &free_hugepages_attr
.attr
,
1744 &resv_hugepages_attr
.attr
,
1745 &surplus_hugepages_attr
.attr
,
1747 &nr_hugepages_mempolicy_attr
.attr
,
1752 static struct attribute_group hstate_attr_group
= {
1753 .attrs
= hstate_attrs
,
1756 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1757 struct kobject
**hstate_kobjs
,
1758 struct attribute_group
*hstate_attr_group
)
1761 int hi
= hstate_index(h
);
1763 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1764 if (!hstate_kobjs
[hi
])
1767 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1769 kobject_put(hstate_kobjs
[hi
]);
1774 static void __init
hugetlb_sysfs_init(void)
1779 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1780 if (!hugepages_kobj
)
1783 for_each_hstate(h
) {
1784 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1785 hstate_kobjs
, &hstate_attr_group
);
1787 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1794 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1795 * with node devices in node_devices[] using a parallel array. The array
1796 * index of a node device or _hstate == node id.
1797 * This is here to avoid any static dependency of the node device driver, in
1798 * the base kernel, on the hugetlb module.
1800 struct node_hstate
{
1801 struct kobject
*hugepages_kobj
;
1802 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1804 struct node_hstate node_hstates
[MAX_NUMNODES
];
1807 * A subset of global hstate attributes for node devices
1809 static struct attribute
*per_node_hstate_attrs
[] = {
1810 &nr_hugepages_attr
.attr
,
1811 &free_hugepages_attr
.attr
,
1812 &surplus_hugepages_attr
.attr
,
1816 static struct attribute_group per_node_hstate_attr_group
= {
1817 .attrs
= per_node_hstate_attrs
,
1821 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1822 * Returns node id via non-NULL nidp.
1824 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1828 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1829 struct node_hstate
*nhs
= &node_hstates
[nid
];
1831 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1832 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1844 * Unregister hstate attributes from a single node device.
1845 * No-op if no hstate attributes attached.
1847 static void hugetlb_unregister_node(struct node
*node
)
1850 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1852 if (!nhs
->hugepages_kobj
)
1853 return; /* no hstate attributes */
1855 for_each_hstate(h
) {
1856 int idx
= hstate_index(h
);
1857 if (nhs
->hstate_kobjs
[idx
]) {
1858 kobject_put(nhs
->hstate_kobjs
[idx
]);
1859 nhs
->hstate_kobjs
[idx
] = NULL
;
1863 kobject_put(nhs
->hugepages_kobj
);
1864 nhs
->hugepages_kobj
= NULL
;
1868 * hugetlb module exit: unregister hstate attributes from node devices
1871 static void hugetlb_unregister_all_nodes(void)
1876 * disable node device registrations.
1878 register_hugetlbfs_with_node(NULL
, NULL
);
1881 * remove hstate attributes from any nodes that have them.
1883 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1884 hugetlb_unregister_node(node_devices
[nid
]);
1888 * Register hstate attributes for a single node device.
1889 * No-op if attributes already registered.
1891 static void hugetlb_register_node(struct node
*node
)
1894 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1897 if (nhs
->hugepages_kobj
)
1898 return; /* already allocated */
1900 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1902 if (!nhs
->hugepages_kobj
)
1905 for_each_hstate(h
) {
1906 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1908 &per_node_hstate_attr_group
);
1910 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1911 h
->name
, node
->dev
.id
);
1912 hugetlb_unregister_node(node
);
1919 * hugetlb init time: register hstate attributes for all registered node
1920 * devices of nodes that have memory. All on-line nodes should have
1921 * registered their associated device by this time.
1923 static void hugetlb_register_all_nodes(void)
1927 for_each_node_state(nid
, N_MEMORY
) {
1928 struct node
*node
= node_devices
[nid
];
1929 if (node
->dev
.id
== nid
)
1930 hugetlb_register_node(node
);
1934 * Let the node device driver know we're here so it can
1935 * [un]register hstate attributes on node hotplug.
1937 register_hugetlbfs_with_node(hugetlb_register_node
,
1938 hugetlb_unregister_node
);
1940 #else /* !CONFIG_NUMA */
1942 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1950 static void hugetlb_unregister_all_nodes(void) { }
1952 static void hugetlb_register_all_nodes(void) { }
1956 static void __exit
hugetlb_exit(void)
1960 hugetlb_unregister_all_nodes();
1962 for_each_hstate(h
) {
1963 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1966 kobject_put(hugepages_kobj
);
1968 module_exit(hugetlb_exit
);
1970 static int __init
hugetlb_init(void)
1972 /* Some platform decide whether they support huge pages at boot
1973 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1974 * there is no such support
1976 if (HPAGE_SHIFT
== 0)
1979 if (!size_to_hstate(default_hstate_size
)) {
1980 default_hstate_size
= HPAGE_SIZE
;
1981 if (!size_to_hstate(default_hstate_size
))
1982 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1984 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1985 if (default_hstate_max_huge_pages
)
1986 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1988 hugetlb_init_hstates();
1989 gather_bootmem_prealloc();
1992 hugetlb_sysfs_init();
1993 hugetlb_register_all_nodes();
1994 hugetlb_cgroup_file_init();
1998 module_init(hugetlb_init
);
2000 /* Should be called on processing a hugepagesz=... option */
2001 void __init
hugetlb_add_hstate(unsigned order
)
2006 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2007 pr_warning("hugepagesz= specified twice, ignoring\n");
2010 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2012 h
= &hstates
[hugetlb_max_hstate
++];
2014 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2015 h
->nr_huge_pages
= 0;
2016 h
->free_huge_pages
= 0;
2017 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2018 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2019 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2020 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2021 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2022 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2023 huge_page_size(h
)/1024);
2028 static int __init
hugetlb_nrpages_setup(char *s
)
2031 static unsigned long *last_mhp
;
2034 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2035 * so this hugepages= parameter goes to the "default hstate".
2037 if (!hugetlb_max_hstate
)
2038 mhp
= &default_hstate_max_huge_pages
;
2040 mhp
= &parsed_hstate
->max_huge_pages
;
2042 if (mhp
== last_mhp
) {
2043 pr_warning("hugepages= specified twice without "
2044 "interleaving hugepagesz=, ignoring\n");
2048 if (sscanf(s
, "%lu", mhp
) <= 0)
2052 * Global state is always initialized later in hugetlb_init.
2053 * But we need to allocate >= MAX_ORDER hstates here early to still
2054 * use the bootmem allocator.
2056 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2057 hugetlb_hstate_alloc_pages(parsed_hstate
);
2063 __setup("hugepages=", hugetlb_nrpages_setup
);
2065 static int __init
hugetlb_default_setup(char *s
)
2067 default_hstate_size
= memparse(s
, &s
);
2070 __setup("default_hugepagesz=", hugetlb_default_setup
);
2072 static unsigned int cpuset_mems_nr(unsigned int *array
)
2075 unsigned int nr
= 0;
2077 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2083 #ifdef CONFIG_SYSCTL
2084 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2085 struct ctl_table
*table
, int write
,
2086 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2088 struct hstate
*h
= &default_hstate
;
2092 tmp
= h
->max_huge_pages
;
2094 if (write
&& h
->order
>= MAX_ORDER
)
2098 table
->maxlen
= sizeof(unsigned long);
2099 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2104 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2105 GFP_KERNEL
| __GFP_NORETRY
);
2106 if (!(obey_mempolicy
&&
2107 init_nodemask_of_mempolicy(nodes_allowed
))) {
2108 NODEMASK_FREE(nodes_allowed
);
2109 nodes_allowed
= &node_states
[N_MEMORY
];
2111 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2113 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2114 NODEMASK_FREE(nodes_allowed
);
2120 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2121 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2124 return hugetlb_sysctl_handler_common(false, table
, write
,
2125 buffer
, length
, ppos
);
2129 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2130 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2132 return hugetlb_sysctl_handler_common(true, table
, write
,
2133 buffer
, length
, ppos
);
2135 #endif /* CONFIG_NUMA */
2137 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2138 void __user
*buffer
,
2139 size_t *length
, loff_t
*ppos
)
2141 struct hstate
*h
= &default_hstate
;
2145 tmp
= h
->nr_overcommit_huge_pages
;
2147 if (write
&& h
->order
>= MAX_ORDER
)
2151 table
->maxlen
= sizeof(unsigned long);
2152 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2157 spin_lock(&hugetlb_lock
);
2158 h
->nr_overcommit_huge_pages
= tmp
;
2159 spin_unlock(&hugetlb_lock
);
2165 #endif /* CONFIG_SYSCTL */
2167 void hugetlb_report_meminfo(struct seq_file
*m
)
2169 struct hstate
*h
= &default_hstate
;
2171 "HugePages_Total: %5lu\n"
2172 "HugePages_Free: %5lu\n"
2173 "HugePages_Rsvd: %5lu\n"
2174 "HugePages_Surp: %5lu\n"
2175 "Hugepagesize: %8lu kB\n",
2179 h
->surplus_huge_pages
,
2180 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2183 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2185 struct hstate
*h
= &default_hstate
;
2187 "Node %d HugePages_Total: %5u\n"
2188 "Node %d HugePages_Free: %5u\n"
2189 "Node %d HugePages_Surp: %5u\n",
2190 nid
, h
->nr_huge_pages_node
[nid
],
2191 nid
, h
->free_huge_pages_node
[nid
],
2192 nid
, h
->surplus_huge_pages_node
[nid
]);
2195 void hugetlb_show_meminfo(void)
2200 for_each_node_state(nid
, N_MEMORY
)
2202 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2204 h
->nr_huge_pages_node
[nid
],
2205 h
->free_huge_pages_node
[nid
],
2206 h
->surplus_huge_pages_node
[nid
],
2207 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2210 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2211 unsigned long hugetlb_total_pages(void)
2214 unsigned long nr_total_pages
= 0;
2217 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2218 return nr_total_pages
;
2221 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2225 spin_lock(&hugetlb_lock
);
2227 * When cpuset is configured, it breaks the strict hugetlb page
2228 * reservation as the accounting is done on a global variable. Such
2229 * reservation is completely rubbish in the presence of cpuset because
2230 * the reservation is not checked against page availability for the
2231 * current cpuset. Application can still potentially OOM'ed by kernel
2232 * with lack of free htlb page in cpuset that the task is in.
2233 * Attempt to enforce strict accounting with cpuset is almost
2234 * impossible (or too ugly) because cpuset is too fluid that
2235 * task or memory node can be dynamically moved between cpusets.
2237 * The change of semantics for shared hugetlb mapping with cpuset is
2238 * undesirable. However, in order to preserve some of the semantics,
2239 * we fall back to check against current free page availability as
2240 * a best attempt and hopefully to minimize the impact of changing
2241 * semantics that cpuset has.
2244 if (gather_surplus_pages(h
, delta
) < 0)
2247 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2248 return_unused_surplus_pages(h
, delta
);
2255 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2258 spin_unlock(&hugetlb_lock
);
2262 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2264 struct resv_map
*resv
= vma_resv_map(vma
);
2267 * This new VMA should share its siblings reservation map if present.
2268 * The VMA will only ever have a valid reservation map pointer where
2269 * it is being copied for another still existing VMA. As that VMA
2270 * has a reference to the reservation map it cannot disappear until
2271 * after this open call completes. It is therefore safe to take a
2272 * new reference here without additional locking.
2275 kref_get(&resv
->refs
);
2278 static void resv_map_put(struct vm_area_struct
*vma
)
2280 struct resv_map
*resv
= vma_resv_map(vma
);
2284 kref_put(&resv
->refs
, resv_map_release
);
2287 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2289 struct hstate
*h
= hstate_vma(vma
);
2290 struct resv_map
*resv
= vma_resv_map(vma
);
2291 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2292 unsigned long reserve
;
2293 unsigned long start
;
2297 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2298 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2300 reserve
= (end
- start
) -
2301 region_count(resv
, start
, end
);
2306 hugetlb_acct_memory(h
, -reserve
);
2307 hugepage_subpool_put_pages(spool
, reserve
);
2313 * We cannot handle pagefaults against hugetlb pages at all. They cause
2314 * handle_mm_fault() to try to instantiate regular-sized pages in the
2315 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2318 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2324 const struct vm_operations_struct hugetlb_vm_ops
= {
2325 .fault
= hugetlb_vm_op_fault
,
2326 .open
= hugetlb_vm_op_open
,
2327 .close
= hugetlb_vm_op_close
,
2330 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2336 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2337 vma
->vm_page_prot
)));
2339 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2340 vma
->vm_page_prot
));
2342 entry
= pte_mkyoung(entry
);
2343 entry
= pte_mkhuge(entry
);
2344 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2349 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2350 unsigned long address
, pte_t
*ptep
)
2354 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2355 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2356 update_mmu_cache(vma
, address
, ptep
);
2360 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2361 struct vm_area_struct
*vma
)
2363 pte_t
*src_pte
, *dst_pte
, entry
;
2364 struct page
*ptepage
;
2367 struct hstate
*h
= hstate_vma(vma
);
2368 unsigned long sz
= huge_page_size(h
);
2369 unsigned long mmun_start
; /* For mmu_notifiers */
2370 unsigned long mmun_end
; /* For mmu_notifiers */
2373 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2375 mmun_start
= vma
->vm_start
;
2376 mmun_end
= vma
->vm_end
;
2378 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
2380 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2381 spinlock_t
*src_ptl
, *dst_ptl
;
2382 src_pte
= huge_pte_offset(src
, addr
);
2385 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2391 /* If the pagetables are shared don't copy or take references */
2392 if (dst_pte
== src_pte
)
2395 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
2396 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
2397 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
2398 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2400 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2401 entry
= huge_ptep_get(src_pte
);
2402 ptepage
= pte_page(entry
);
2404 page_dup_rmap(ptepage
);
2405 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2407 spin_unlock(src_ptl
);
2408 spin_unlock(dst_ptl
);
2412 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
2417 static int is_hugetlb_entry_migration(pte_t pte
)
2421 if (huge_pte_none(pte
) || pte_present(pte
))
2423 swp
= pte_to_swp_entry(pte
);
2424 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2430 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2434 if (huge_pte_none(pte
) || pte_present(pte
))
2436 swp
= pte_to_swp_entry(pte
);
2437 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2443 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2444 unsigned long start
, unsigned long end
,
2445 struct page
*ref_page
)
2447 int force_flush
= 0;
2448 struct mm_struct
*mm
= vma
->vm_mm
;
2449 unsigned long address
;
2454 struct hstate
*h
= hstate_vma(vma
);
2455 unsigned long sz
= huge_page_size(h
);
2456 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2457 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2459 WARN_ON(!is_vm_hugetlb_page(vma
));
2460 BUG_ON(start
& ~huge_page_mask(h
));
2461 BUG_ON(end
& ~huge_page_mask(h
));
2463 tlb_start_vma(tlb
, vma
);
2464 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2466 for (address
= start
; address
< end
; address
+= sz
) {
2467 ptep
= huge_pte_offset(mm
, address
);
2471 ptl
= huge_pte_lock(h
, mm
, ptep
);
2472 if (huge_pmd_unshare(mm
, &address
, ptep
))
2475 pte
= huge_ptep_get(ptep
);
2476 if (huge_pte_none(pte
))
2480 * HWPoisoned hugepage is already unmapped and dropped reference
2482 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2483 huge_pte_clear(mm
, address
, ptep
);
2487 page
= pte_page(pte
);
2489 * If a reference page is supplied, it is because a specific
2490 * page is being unmapped, not a range. Ensure the page we
2491 * are about to unmap is the actual page of interest.
2494 if (page
!= ref_page
)
2498 * Mark the VMA as having unmapped its page so that
2499 * future faults in this VMA will fail rather than
2500 * looking like data was lost
2502 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2505 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2506 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2507 if (huge_pte_dirty(pte
))
2508 set_page_dirty(page
);
2510 page_remove_rmap(page
);
2511 force_flush
= !__tlb_remove_page(tlb
, page
);
2516 /* Bail out after unmapping reference page if supplied */
2525 * mmu_gather ran out of room to batch pages, we break out of
2526 * the PTE lock to avoid doing the potential expensive TLB invalidate
2527 * and page-free while holding it.
2532 if (address
< end
&& !ref_page
)
2535 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2536 tlb_end_vma(tlb
, vma
);
2539 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2540 struct vm_area_struct
*vma
, unsigned long start
,
2541 unsigned long end
, struct page
*ref_page
)
2543 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2546 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2547 * test will fail on a vma being torn down, and not grab a page table
2548 * on its way out. We're lucky that the flag has such an appropriate
2549 * name, and can in fact be safely cleared here. We could clear it
2550 * before the __unmap_hugepage_range above, but all that's necessary
2551 * is to clear it before releasing the i_mmap_mutex. This works
2552 * because in the context this is called, the VMA is about to be
2553 * destroyed and the i_mmap_mutex is held.
2555 vma
->vm_flags
&= ~VM_MAYSHARE
;
2558 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2559 unsigned long end
, struct page
*ref_page
)
2561 struct mm_struct
*mm
;
2562 struct mmu_gather tlb
;
2566 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2567 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2568 tlb_finish_mmu(&tlb
, start
, end
);
2572 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2573 * mappping it owns the reserve page for. The intention is to unmap the page
2574 * from other VMAs and let the children be SIGKILLed if they are faulting the
2577 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2578 struct page
*page
, unsigned long address
)
2580 struct hstate
*h
= hstate_vma(vma
);
2581 struct vm_area_struct
*iter_vma
;
2582 struct address_space
*mapping
;
2586 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2587 * from page cache lookup which is in HPAGE_SIZE units.
2589 address
= address
& huge_page_mask(h
);
2590 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2592 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2595 * Take the mapping lock for the duration of the table walk. As
2596 * this mapping should be shared between all the VMAs,
2597 * __unmap_hugepage_range() is called as the lock is already held
2599 mutex_lock(&mapping
->i_mmap_mutex
);
2600 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2601 /* Do not unmap the current VMA */
2602 if (iter_vma
== vma
)
2606 * Unmap the page from other VMAs without their own reserves.
2607 * They get marked to be SIGKILLed if they fault in these
2608 * areas. This is because a future no-page fault on this VMA
2609 * could insert a zeroed page instead of the data existing
2610 * from the time of fork. This would look like data corruption
2612 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2613 unmap_hugepage_range(iter_vma
, address
,
2614 address
+ huge_page_size(h
), page
);
2616 mutex_unlock(&mapping
->i_mmap_mutex
);
2622 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2623 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2624 * cannot race with other handlers or page migration.
2625 * Keep the pte_same checks anyway to make transition from the mutex easier.
2627 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2628 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2629 struct page
*pagecache_page
, spinlock_t
*ptl
)
2631 struct hstate
*h
= hstate_vma(vma
);
2632 struct page
*old_page
, *new_page
;
2633 int outside_reserve
= 0;
2634 unsigned long mmun_start
; /* For mmu_notifiers */
2635 unsigned long mmun_end
; /* For mmu_notifiers */
2637 old_page
= pte_page(pte
);
2640 /* If no-one else is actually using this page, avoid the copy
2641 * and just make the page writable */
2642 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2643 page_move_anon_rmap(old_page
, vma
, address
);
2644 set_huge_ptep_writable(vma
, address
, ptep
);
2649 * If the process that created a MAP_PRIVATE mapping is about to
2650 * perform a COW due to a shared page count, attempt to satisfy
2651 * the allocation without using the existing reserves. The pagecache
2652 * page is used to determine if the reserve at this address was
2653 * consumed or not. If reserves were used, a partial faulted mapping
2654 * at the time of fork() could consume its reserves on COW instead
2655 * of the full address range.
2657 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2658 old_page
!= pagecache_page
)
2659 outside_reserve
= 1;
2661 page_cache_get(old_page
);
2663 /* Drop page table lock as buddy allocator may be called */
2665 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2667 if (IS_ERR(new_page
)) {
2668 long err
= PTR_ERR(new_page
);
2669 page_cache_release(old_page
);
2672 * If a process owning a MAP_PRIVATE mapping fails to COW,
2673 * it is due to references held by a child and an insufficient
2674 * huge page pool. To guarantee the original mappers
2675 * reliability, unmap the page from child processes. The child
2676 * may get SIGKILLed if it later faults.
2678 if (outside_reserve
) {
2679 BUG_ON(huge_pte_none(pte
));
2680 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2681 BUG_ON(huge_pte_none(pte
));
2683 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2684 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2685 goto retry_avoidcopy
;
2687 * race occurs while re-acquiring page table
2688 * lock, and our job is done.
2695 /* Caller expects lock to be held */
2698 return VM_FAULT_OOM
;
2700 return VM_FAULT_SIGBUS
;
2704 * When the original hugepage is shared one, it does not have
2705 * anon_vma prepared.
2707 if (unlikely(anon_vma_prepare(vma
))) {
2708 page_cache_release(new_page
);
2709 page_cache_release(old_page
);
2710 /* Caller expects lock to be held */
2712 return VM_FAULT_OOM
;
2715 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2716 pages_per_huge_page(h
));
2717 __SetPageUptodate(new_page
);
2719 mmun_start
= address
& huge_page_mask(h
);
2720 mmun_end
= mmun_start
+ huge_page_size(h
);
2721 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2723 * Retake the page table lock to check for racing updates
2724 * before the page tables are altered
2727 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2728 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2729 ClearPagePrivate(new_page
);
2732 huge_ptep_clear_flush(vma
, address
, ptep
);
2733 set_huge_pte_at(mm
, address
, ptep
,
2734 make_huge_pte(vma
, new_page
, 1));
2735 page_remove_rmap(old_page
);
2736 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2737 /* Make the old page be freed below */
2738 new_page
= old_page
;
2741 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2742 page_cache_release(new_page
);
2743 page_cache_release(old_page
);
2745 /* Caller expects lock to be held */
2750 /* Return the pagecache page at a given address within a VMA */
2751 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2752 struct vm_area_struct
*vma
, unsigned long address
)
2754 struct address_space
*mapping
;
2757 mapping
= vma
->vm_file
->f_mapping
;
2758 idx
= vma_hugecache_offset(h
, vma
, address
);
2760 return find_lock_page(mapping
, idx
);
2764 * Return whether there is a pagecache page to back given address within VMA.
2765 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2767 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2768 struct vm_area_struct
*vma
, unsigned long address
)
2770 struct address_space
*mapping
;
2774 mapping
= vma
->vm_file
->f_mapping
;
2775 idx
= vma_hugecache_offset(h
, vma
, address
);
2777 page
= find_get_page(mapping
, idx
);
2780 return page
!= NULL
;
2783 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2784 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2786 struct hstate
*h
= hstate_vma(vma
);
2787 int ret
= VM_FAULT_SIGBUS
;
2792 struct address_space
*mapping
;
2797 * Currently, we are forced to kill the process in the event the
2798 * original mapper has unmapped pages from the child due to a failed
2799 * COW. Warn that such a situation has occurred as it may not be obvious
2801 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2802 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2807 mapping
= vma
->vm_file
->f_mapping
;
2808 idx
= vma_hugecache_offset(h
, vma
, address
);
2811 * Use page lock to guard against racing truncation
2812 * before we get page_table_lock.
2815 page
= find_lock_page(mapping
, idx
);
2817 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2820 page
= alloc_huge_page(vma
, address
, 0);
2822 ret
= PTR_ERR(page
);
2826 ret
= VM_FAULT_SIGBUS
;
2829 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2830 __SetPageUptodate(page
);
2832 if (vma
->vm_flags
& VM_MAYSHARE
) {
2834 struct inode
*inode
= mapping
->host
;
2836 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2843 ClearPagePrivate(page
);
2845 spin_lock(&inode
->i_lock
);
2846 inode
->i_blocks
+= blocks_per_huge_page(h
);
2847 spin_unlock(&inode
->i_lock
);
2850 if (unlikely(anon_vma_prepare(vma
))) {
2852 goto backout_unlocked
;
2858 * If memory error occurs between mmap() and fault, some process
2859 * don't have hwpoisoned swap entry for errored virtual address.
2860 * So we need to block hugepage fault by PG_hwpoison bit check.
2862 if (unlikely(PageHWPoison(page
))) {
2863 ret
= VM_FAULT_HWPOISON
|
2864 VM_FAULT_SET_HINDEX(hstate_index(h
));
2865 goto backout_unlocked
;
2870 * If we are going to COW a private mapping later, we examine the
2871 * pending reservations for this page now. This will ensure that
2872 * any allocations necessary to record that reservation occur outside
2875 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2876 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2878 goto backout_unlocked
;
2881 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
2883 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2888 if (!huge_pte_none(huge_ptep_get(ptep
)))
2892 ClearPagePrivate(page
);
2893 hugepage_add_new_anon_rmap(page
, vma
, address
);
2896 page_dup_rmap(page
);
2897 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2898 && (vma
->vm_flags
& VM_SHARED
)));
2899 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2901 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2902 /* Optimization, do the COW without a second fault */
2903 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
2919 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2920 unsigned long address
, unsigned int flags
)
2926 struct page
*page
= NULL
;
2927 struct page
*pagecache_page
= NULL
;
2928 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2929 struct hstate
*h
= hstate_vma(vma
);
2931 address
&= huge_page_mask(h
);
2933 ptep
= huge_pte_offset(mm
, address
);
2935 entry
= huge_ptep_get(ptep
);
2936 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2937 migration_entry_wait_huge(vma
, mm
, ptep
);
2939 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2940 return VM_FAULT_HWPOISON_LARGE
|
2941 VM_FAULT_SET_HINDEX(hstate_index(h
));
2944 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2946 return VM_FAULT_OOM
;
2949 * Serialize hugepage allocation and instantiation, so that we don't
2950 * get spurious allocation failures if two CPUs race to instantiate
2951 * the same page in the page cache.
2953 mutex_lock(&hugetlb_instantiation_mutex
);
2954 entry
= huge_ptep_get(ptep
);
2955 if (huge_pte_none(entry
)) {
2956 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2963 * If we are going to COW the mapping later, we examine the pending
2964 * reservations for this page now. This will ensure that any
2965 * allocations necessary to record that reservation occur outside the
2966 * spinlock. For private mappings, we also lookup the pagecache
2967 * page now as it is used to determine if a reservation has been
2970 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2971 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2976 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2977 pagecache_page
= hugetlbfs_pagecache_page(h
,
2982 * hugetlb_cow() requires page locks of pte_page(entry) and
2983 * pagecache_page, so here we need take the former one
2984 * when page != pagecache_page or !pagecache_page.
2985 * Note that locking order is always pagecache_page -> page,
2986 * so no worry about deadlock.
2988 page
= pte_page(entry
);
2990 if (page
!= pagecache_page
)
2993 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
2995 /* Check for a racing update before calling hugetlb_cow */
2996 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3000 if (flags
& FAULT_FLAG_WRITE
) {
3001 if (!huge_pte_write(entry
)) {
3002 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3003 pagecache_page
, ptl
);
3006 entry
= huge_pte_mkdirty(entry
);
3008 entry
= pte_mkyoung(entry
);
3009 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3010 flags
& FAULT_FLAG_WRITE
))
3011 update_mmu_cache(vma
, address
, ptep
);
3016 if (pagecache_page
) {
3017 unlock_page(pagecache_page
);
3018 put_page(pagecache_page
);
3020 if (page
!= pagecache_page
)
3025 mutex_unlock(&hugetlb_instantiation_mutex
);
3030 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3031 struct page
**pages
, struct vm_area_struct
**vmas
,
3032 unsigned long *position
, unsigned long *nr_pages
,
3033 long i
, unsigned int flags
)
3035 unsigned long pfn_offset
;
3036 unsigned long vaddr
= *position
;
3037 unsigned long remainder
= *nr_pages
;
3038 struct hstate
*h
= hstate_vma(vma
);
3040 while (vaddr
< vma
->vm_end
&& remainder
) {
3042 spinlock_t
*ptl
= NULL
;
3047 * Some archs (sparc64, sh*) have multiple pte_ts to
3048 * each hugepage. We have to make sure we get the
3049 * first, for the page indexing below to work.
3051 * Note that page table lock is not held when pte is null.
3053 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3055 ptl
= huge_pte_lock(h
, mm
, pte
);
3056 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3059 * When coredumping, it suits get_dump_page if we just return
3060 * an error where there's an empty slot with no huge pagecache
3061 * to back it. This way, we avoid allocating a hugepage, and
3062 * the sparse dumpfile avoids allocating disk blocks, but its
3063 * huge holes still show up with zeroes where they need to be.
3065 if (absent
&& (flags
& FOLL_DUMP
) &&
3066 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3074 * We need call hugetlb_fault for both hugepages under migration
3075 * (in which case hugetlb_fault waits for the migration,) and
3076 * hwpoisoned hugepages (in which case we need to prevent the
3077 * caller from accessing to them.) In order to do this, we use
3078 * here is_swap_pte instead of is_hugetlb_entry_migration and
3079 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3080 * both cases, and because we can't follow correct pages
3081 * directly from any kind of swap entries.
3083 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3084 ((flags
& FOLL_WRITE
) &&
3085 !huge_pte_write(huge_ptep_get(pte
)))) {
3090 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3091 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3092 if (!(ret
& VM_FAULT_ERROR
))
3099 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3100 page
= pte_page(huge_ptep_get(pte
));
3103 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3104 get_page_foll(pages
[i
]);
3114 if (vaddr
< vma
->vm_end
&& remainder
&&
3115 pfn_offset
< pages_per_huge_page(h
)) {
3117 * We use pfn_offset to avoid touching the pageframes
3118 * of this compound page.
3124 *nr_pages
= remainder
;
3127 return i
? i
: -EFAULT
;
3130 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3131 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3133 struct mm_struct
*mm
= vma
->vm_mm
;
3134 unsigned long start
= address
;
3137 struct hstate
*h
= hstate_vma(vma
);
3138 unsigned long pages
= 0;
3140 BUG_ON(address
>= end
);
3141 flush_cache_range(vma
, address
, end
);
3143 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3144 for (; address
< end
; address
+= huge_page_size(h
)) {
3146 ptep
= huge_pte_offset(mm
, address
);
3149 ptl
= huge_pte_lock(h
, mm
, ptep
);
3150 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3155 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3156 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3157 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3158 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3159 set_huge_pte_at(mm
, address
, ptep
, pte
);
3165 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3166 * may have cleared our pud entry and done put_page on the page table:
3167 * once we release i_mmap_mutex, another task can do the final put_page
3168 * and that page table be reused and filled with junk.
3170 flush_tlb_range(vma
, start
, end
);
3171 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3173 return pages
<< h
->order
;
3176 int hugetlb_reserve_pages(struct inode
*inode
,
3178 struct vm_area_struct
*vma
,
3179 vm_flags_t vm_flags
)
3182 struct hstate
*h
= hstate_inode(inode
);
3183 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3184 struct resv_map
*resv_map
;
3187 * Only apply hugepage reservation if asked. At fault time, an
3188 * attempt will be made for VM_NORESERVE to allocate a page
3189 * without using reserves
3191 if (vm_flags
& VM_NORESERVE
)
3195 * Shared mappings base their reservation on the number of pages that
3196 * are already allocated on behalf of the file. Private mappings need
3197 * to reserve the full area even if read-only as mprotect() may be
3198 * called to make the mapping read-write. Assume !vma is a shm mapping
3200 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
3201 resv_map
= inode
->i_mapping
->private_data
;
3203 chg
= region_chg(resv_map
, from
, to
);
3206 resv_map
= resv_map_alloc();
3212 set_vma_resv_map(vma
, resv_map
);
3213 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3221 /* There must be enough pages in the subpool for the mapping */
3222 if (hugepage_subpool_get_pages(spool
, chg
)) {
3228 * Check enough hugepages are available for the reservation.
3229 * Hand the pages back to the subpool if there are not
3231 ret
= hugetlb_acct_memory(h
, chg
);
3233 hugepage_subpool_put_pages(spool
, chg
);
3238 * Account for the reservations made. Shared mappings record regions
3239 * that have reservations as they are shared by multiple VMAs.
3240 * When the last VMA disappears, the region map says how much
3241 * the reservation was and the page cache tells how much of
3242 * the reservation was consumed. Private mappings are per-VMA and
3243 * only the consumed reservations are tracked. When the VMA
3244 * disappears, the original reservation is the VMA size and the
3245 * consumed reservations are stored in the map. Hence, nothing
3246 * else has to be done for private mappings here
3248 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3249 region_add(resv_map
, from
, to
);
3257 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3259 struct hstate
*h
= hstate_inode(inode
);
3260 struct resv_map
*resv_map
= inode
->i_mapping
->private_data
;
3262 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3265 chg
= region_truncate(resv_map
, offset
);
3266 spin_lock(&inode
->i_lock
);
3267 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3268 spin_unlock(&inode
->i_lock
);
3270 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3271 hugetlb_acct_memory(h
, -(chg
- freed
));
3274 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3275 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3276 struct vm_area_struct
*vma
,
3277 unsigned long addr
, pgoff_t idx
)
3279 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3281 unsigned long sbase
= saddr
& PUD_MASK
;
3282 unsigned long s_end
= sbase
+ PUD_SIZE
;
3284 /* Allow segments to share if only one is marked locked */
3285 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3286 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3289 * match the virtual addresses, permission and the alignment of the
3292 if (pmd_index(addr
) != pmd_index(saddr
) ||
3293 vm_flags
!= svm_flags
||
3294 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3300 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3302 unsigned long base
= addr
& PUD_MASK
;
3303 unsigned long end
= base
+ PUD_SIZE
;
3306 * check on proper vm_flags and page table alignment
3308 if (vma
->vm_flags
& VM_MAYSHARE
&&
3309 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3315 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3316 * and returns the corresponding pte. While this is not necessary for the
3317 * !shared pmd case because we can allocate the pmd later as well, it makes the
3318 * code much cleaner. pmd allocation is essential for the shared case because
3319 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3320 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3321 * bad pmd for sharing.
3323 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3325 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3326 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3327 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3329 struct vm_area_struct
*svma
;
3330 unsigned long saddr
;
3335 if (!vma_shareable(vma
, addr
))
3336 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3338 mutex_lock(&mapping
->i_mmap_mutex
);
3339 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3343 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3345 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3347 get_page(virt_to_page(spte
));
3356 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
3359 pud_populate(mm
, pud
,
3360 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3362 put_page(virt_to_page(spte
));
3365 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3366 mutex_unlock(&mapping
->i_mmap_mutex
);
3371 * unmap huge page backed by shared pte.
3373 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3374 * indicated by page_count > 1, unmap is achieved by clearing pud and
3375 * decrementing the ref count. If count == 1, the pte page is not shared.
3377 * called with page table lock held.
3379 * returns: 1 successfully unmapped a shared pte page
3380 * 0 the underlying pte page is not shared, or it is the last user
3382 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3384 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3385 pud_t
*pud
= pud_offset(pgd
, *addr
);
3387 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3388 if (page_count(virt_to_page(ptep
)) == 1)
3392 put_page(virt_to_page(ptep
));
3393 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3396 #define want_pmd_share() (1)
3397 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3398 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3402 #define want_pmd_share() (0)
3403 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3405 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3406 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3407 unsigned long addr
, unsigned long sz
)
3413 pgd
= pgd_offset(mm
, addr
);
3414 pud
= pud_alloc(mm
, pgd
, addr
);
3416 if (sz
== PUD_SIZE
) {
3419 BUG_ON(sz
!= PMD_SIZE
);
3420 if (want_pmd_share() && pud_none(*pud
))
3421 pte
= huge_pmd_share(mm
, addr
, pud
);
3423 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3426 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3431 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3437 pgd
= pgd_offset(mm
, addr
);
3438 if (pgd_present(*pgd
)) {
3439 pud
= pud_offset(pgd
, addr
);
3440 if (pud_present(*pud
)) {
3442 return (pte_t
*)pud
;
3443 pmd
= pmd_offset(pud
, addr
);
3446 return (pte_t
*) pmd
;
3450 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3451 pmd_t
*pmd
, int write
)
3455 page
= pte_page(*(pte_t
*)pmd
);
3457 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3462 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3463 pud_t
*pud
, int write
)
3467 page
= pte_page(*(pte_t
*)pud
);
3469 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3473 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3475 /* Can be overriden by architectures */
3476 __attribute__((weak
)) struct page
*
3477 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3478 pud_t
*pud
, int write
)
3484 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3486 #ifdef CONFIG_MEMORY_FAILURE
3488 /* Should be called in hugetlb_lock */
3489 static int is_hugepage_on_freelist(struct page
*hpage
)
3493 struct hstate
*h
= page_hstate(hpage
);
3494 int nid
= page_to_nid(hpage
);
3496 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3503 * This function is called from memory failure code.
3504 * Assume the caller holds page lock of the head page.
3506 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3508 struct hstate
*h
= page_hstate(hpage
);
3509 int nid
= page_to_nid(hpage
);
3512 spin_lock(&hugetlb_lock
);
3513 if (is_hugepage_on_freelist(hpage
)) {
3515 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3516 * but dangling hpage->lru can trigger list-debug warnings
3517 * (this happens when we call unpoison_memory() on it),
3518 * so let it point to itself with list_del_init().
3520 list_del_init(&hpage
->lru
);
3521 set_page_refcounted(hpage
);
3522 h
->free_huge_pages
--;
3523 h
->free_huge_pages_node
[nid
]--;
3526 spin_unlock(&hugetlb_lock
);
3531 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3533 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3534 if (!get_page_unless_zero(page
))
3536 spin_lock(&hugetlb_lock
);
3537 list_move_tail(&page
->lru
, list
);
3538 spin_unlock(&hugetlb_lock
);
3542 void putback_active_hugepage(struct page
*page
)
3544 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3545 spin_lock(&hugetlb_lock
);
3546 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3547 spin_unlock(&hugetlb_lock
);
3551 bool is_hugepage_active(struct page
*page
)
3553 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
3555 * This function can be called for a tail page because the caller,
3556 * scan_movable_pages, scans through a given pfn-range which typically
3557 * covers one memory block. In systems using gigantic hugepage (1GB
3558 * for x86_64,) a hugepage is larger than a memory block, and we don't
3559 * support migrating such large hugepages for now, so return false
3560 * when called for tail pages.
3565 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3566 * so we should return false for them.
3568 if (unlikely(PageHWPoison(page
)))
3570 return page_count(page
) > 0;