2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
27 #include <asm/pgtable.h>
31 #include <linux/hugetlb.h>
32 #include <linux/hugetlb_cgroup.h>
33 #include <linux/node.h>
36 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
37 unsigned long hugepages_treat_as_movable
;
39 int hugetlb_max_hstate __read_mostly
;
40 unsigned int default_hstate_idx
;
41 struct hstate hstates
[HUGE_MAX_HSTATE
];
43 __initdata
LIST_HEAD(huge_boot_pages
);
45 /* for command line parsing */
46 static struct hstate
* __initdata parsed_hstate
;
47 static unsigned long __initdata default_hstate_max_huge_pages
;
48 static unsigned long __initdata default_hstate_size
;
51 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
52 * free_huge_pages, and surplus_huge_pages.
54 DEFINE_SPINLOCK(hugetlb_lock
);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
58 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
60 spin_unlock(&spool
->lock
);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
70 struct hugepage_subpool
*spool
;
72 spool
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
76 spin_lock_init(&spool
->lock
);
78 spool
->max_hpages
= nr_blocks
;
79 spool
->used_hpages
= 0;
84 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
86 spin_lock(&spool
->lock
);
87 BUG_ON(!spool
->count
);
89 unlock_or_release_subpool(spool
);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
100 spin_lock(&spool
->lock
);
101 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
102 spool
->used_hpages
+= delta
;
106 spin_unlock(&spool
->lock
);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
117 spin_lock(&spool
->lock
);
118 spool
->used_hpages
-= delta
;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool
);
124 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
126 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
129 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
131 return subpool_inode(file_inode(vma
->vm_file
));
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation_mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link
;
154 static long region_add(struct list_head
*head
, long f
, long t
)
156 struct file_region
*rg
, *nrg
, *trg
;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg
, head
, link
)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
170 if (&rg
->link
== head
)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head
*head
, long f
, long t
)
192 struct file_region
*rg
, *nrg
;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg
, head
, link
)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg
->link
== head
|| t
< rg
->from
) {
204 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
209 INIT_LIST_HEAD(&nrg
->link
);
210 list_add(&nrg
->link
, rg
->link
.prev
);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
222 if (&rg
->link
== head
)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg
-= rg
->to
- rg
->from
;
239 static long region_truncate(struct list_head
*head
, long end
)
241 struct file_region
*rg
, *trg
;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg
, head
, link
)
248 if (&rg
->link
== head
)
251 /* If we are in the middle of a region then adjust it. */
252 if (end
> rg
->from
) {
255 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
260 if (&rg
->link
== head
)
262 chg
+= rg
->to
- rg
->from
;
269 static long region_count(struct list_head
*head
, long f
, long t
)
271 struct file_region
*rg
;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg
, head
, link
) {
284 seg_from
= max(rg
->from
, f
);
285 seg_to
= min(rg
->to
, t
);
287 chg
+= seg_to
- seg_from
;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
298 struct vm_area_struct
*vma
, unsigned long address
)
300 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
301 (vma
->vm_pgoff
>> huge_page_order(h
));
304 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
305 unsigned long address
)
307 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
316 struct hstate
*hstate
;
318 if (!is_vm_hugetlb_page(vma
))
321 hstate
= hstate_vma(vma
);
323 return 1UL << huge_page_shift(hstate
);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
336 return vma_kernel_pagesize(vma
);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
370 return (unsigned long)vma
->vm_private_data
;
373 static void set_vma_private_data(struct vm_area_struct
*vma
,
376 vma
->vm_private_data
= (void *)value
;
381 struct list_head regions
;
384 static struct resv_map
*resv_map_alloc(void)
386 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
390 kref_init(&resv_map
->refs
);
391 INIT_LIST_HEAD(&resv_map
->regions
);
396 static void resv_map_release(struct kref
*ref
)
398 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map
->regions
, 0);
405 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
408 if (!(vma
->vm_flags
& VM_MAYSHARE
))
409 return (struct resv_map
*)(get_vma_private_data(vma
) &
414 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
417 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
419 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
420 HPAGE_RESV_MASK
) | (unsigned long)map
);
423 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
426 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
428 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
431 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
435 return (get_vma_private_data(vma
) & flag
) != 0;
438 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
439 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
441 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
442 if (!(vma
->vm_flags
& VM_MAYSHARE
))
443 vma
->vm_private_data
= (void *)0;
446 /* Returns true if the VMA has associated reserve pages */
447 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
449 if (vma
->vm_flags
& VM_NORESERVE
) {
451 * This address is already reserved by other process(chg == 0),
452 * so, we should decrement reserved count. Without decrementing,
453 * reserve count remains after releasing inode, because this
454 * allocated page will go into page cache and is regarded as
455 * coming from reserved pool in releasing step. Currently, we
456 * don't have any other solution to deal with this situation
457 * properly, so add work-around here.
459 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
465 /* Shared mappings always use reserves */
466 if (vma
->vm_flags
& VM_MAYSHARE
)
470 * Only the process that called mmap() has reserves for
473 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
479 static void copy_gigantic_page(struct page
*dst
, struct page
*src
)
482 struct hstate
*h
= page_hstate(src
);
483 struct page
*dst_base
= dst
;
484 struct page
*src_base
= src
;
486 for (i
= 0; i
< pages_per_huge_page(h
); ) {
488 copy_highpage(dst
, src
);
491 dst
= mem_map_next(dst
, dst_base
, i
);
492 src
= mem_map_next(src
, src_base
, i
);
496 void copy_huge_page(struct page
*dst
, struct page
*src
)
499 struct hstate
*h
= page_hstate(src
);
501 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
502 copy_gigantic_page(dst
, src
);
507 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
509 copy_highpage(dst
+ i
, src
+ i
);
513 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
515 int nid
= page_to_nid(page
);
516 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
517 h
->free_huge_pages
++;
518 h
->free_huge_pages_node
[nid
]++;
521 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
525 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
526 if (!is_migrate_isolate_page(page
))
529 * if 'non-isolated free hugepage' not found on the list,
530 * the allocation fails.
532 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
534 list_move(&page
->lru
, &h
->hugepage_activelist
);
535 set_page_refcounted(page
);
536 h
->free_huge_pages
--;
537 h
->free_huge_pages_node
[nid
]--;
541 /* Movability of hugepages depends on migration support. */
542 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
544 if (hugepages_treat_as_movable
|| hugepage_migration_support(h
))
545 return GFP_HIGHUSER_MOVABLE
;
550 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
551 struct vm_area_struct
*vma
,
552 unsigned long address
, int avoid_reserve
,
555 struct page
*page
= NULL
;
556 struct mempolicy
*mpol
;
557 nodemask_t
*nodemask
;
558 struct zonelist
*zonelist
;
561 unsigned int cpuset_mems_cookie
;
564 * A child process with MAP_PRIVATE mappings created by their parent
565 * have no page reserves. This check ensures that reservations are
566 * not "stolen". The child may still get SIGKILLed
568 if (!vma_has_reserves(vma
, chg
) &&
569 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
572 /* If reserves cannot be used, ensure enough pages are in the pool */
573 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
577 cpuset_mems_cookie
= get_mems_allowed();
578 zonelist
= huge_zonelist(vma
, address
,
579 htlb_alloc_mask(h
), &mpol
, &nodemask
);
581 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
582 MAX_NR_ZONES
- 1, nodemask
) {
583 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask(h
))) {
584 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
588 if (!vma_has_reserves(vma
, chg
))
591 SetPagePrivate(page
);
592 h
->resv_huge_pages
--;
599 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
607 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
611 VM_BUG_ON(h
->order
>= MAX_ORDER
);
614 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
615 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
616 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
617 1 << PG_referenced
| 1 << PG_dirty
|
618 1 << PG_active
| 1 << PG_reserved
|
619 1 << PG_private
| 1 << PG_writeback
);
621 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
622 set_compound_page_dtor(page
, NULL
);
623 set_page_refcounted(page
);
624 arch_release_hugepage(page
);
625 __free_pages(page
, huge_page_order(h
));
628 struct hstate
*size_to_hstate(unsigned long size
)
633 if (huge_page_size(h
) == size
)
639 static void free_huge_page(struct page
*page
)
642 * Can't pass hstate in here because it is called from the
643 * compound page destructor.
645 struct hstate
*h
= page_hstate(page
);
646 int nid
= page_to_nid(page
);
647 struct hugepage_subpool
*spool
=
648 (struct hugepage_subpool
*)page_private(page
);
649 bool restore_reserve
;
651 set_page_private(page
, 0);
652 page
->mapping
= NULL
;
653 BUG_ON(page_count(page
));
654 BUG_ON(page_mapcount(page
));
655 restore_reserve
= PagePrivate(page
);
656 ClearPagePrivate(page
);
658 spin_lock(&hugetlb_lock
);
659 hugetlb_cgroup_uncharge_page(hstate_index(h
),
660 pages_per_huge_page(h
), page
);
662 h
->resv_huge_pages
++;
664 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
665 /* remove the page from active list */
666 list_del(&page
->lru
);
667 update_and_free_page(h
, page
);
668 h
->surplus_huge_pages
--;
669 h
->surplus_huge_pages_node
[nid
]--;
671 arch_clear_hugepage_flags(page
);
672 enqueue_huge_page(h
, page
);
674 spin_unlock(&hugetlb_lock
);
675 hugepage_subpool_put_pages(spool
, 1);
678 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
680 INIT_LIST_HEAD(&page
->lru
);
681 set_compound_page_dtor(page
, free_huge_page
);
682 spin_lock(&hugetlb_lock
);
683 set_hugetlb_cgroup(page
, NULL
);
685 h
->nr_huge_pages_node
[nid
]++;
686 spin_unlock(&hugetlb_lock
);
687 put_page(page
); /* free it into the hugepage allocator */
690 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
693 int nr_pages
= 1 << order
;
694 struct page
*p
= page
+ 1;
696 /* we rely on prep_new_huge_page to set the destructor */
697 set_compound_order(page
, order
);
699 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
701 set_page_count(p
, 0);
702 p
->first_page
= page
;
707 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
708 * transparent huge pages. See the PageTransHuge() documentation for more
711 int PageHuge(struct page
*page
)
713 compound_page_dtor
*dtor
;
715 if (!PageCompound(page
))
718 page
= compound_head(page
);
719 dtor
= get_compound_page_dtor(page
);
721 return dtor
== free_huge_page
;
723 EXPORT_SYMBOL_GPL(PageHuge
);
725 pgoff_t
__basepage_index(struct page
*page
)
727 struct page
*page_head
= compound_head(page
);
728 pgoff_t index
= page_index(page_head
);
729 unsigned long compound_idx
;
731 if (!PageHuge(page_head
))
732 return page_index(page
);
734 if (compound_order(page_head
) >= MAX_ORDER
)
735 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
737 compound_idx
= page
- page_head
;
739 return (index
<< compound_order(page_head
)) + compound_idx
;
742 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
746 if (h
->order
>= MAX_ORDER
)
749 page
= alloc_pages_exact_node(nid
,
750 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
751 __GFP_REPEAT
|__GFP_NOWARN
,
754 if (arch_prepare_hugepage(page
)) {
755 __free_pages(page
, huge_page_order(h
));
758 prep_new_huge_page(h
, page
, nid
);
765 * common helper functions for hstate_next_node_to_{alloc|free}.
766 * We may have allocated or freed a huge page based on a different
767 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
768 * be outside of *nodes_allowed. Ensure that we use an allowed
769 * node for alloc or free.
771 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
773 nid
= next_node(nid
, *nodes_allowed
);
774 if (nid
== MAX_NUMNODES
)
775 nid
= first_node(*nodes_allowed
);
776 VM_BUG_ON(nid
>= MAX_NUMNODES
);
781 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
783 if (!node_isset(nid
, *nodes_allowed
))
784 nid
= next_node_allowed(nid
, nodes_allowed
);
789 * returns the previously saved node ["this node"] from which to
790 * allocate a persistent huge page for the pool and advance the
791 * next node from which to allocate, handling wrap at end of node
794 static int hstate_next_node_to_alloc(struct hstate
*h
,
795 nodemask_t
*nodes_allowed
)
799 VM_BUG_ON(!nodes_allowed
);
801 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
802 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
808 * helper for free_pool_huge_page() - return the previously saved
809 * node ["this node"] from which to free a huge page. Advance the
810 * next node id whether or not we find a free huge page to free so
811 * that the next attempt to free addresses the next node.
813 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
817 VM_BUG_ON(!nodes_allowed
);
819 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
820 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
825 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
826 for (nr_nodes = nodes_weight(*mask); \
828 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
831 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
832 for (nr_nodes = nodes_weight(*mask); \
834 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
837 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
843 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
844 page
= alloc_fresh_huge_page_node(h
, node
);
852 count_vm_event(HTLB_BUDDY_PGALLOC
);
854 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
860 * Free huge page from pool from next node to free.
861 * Attempt to keep persistent huge pages more or less
862 * balanced over allowed nodes.
863 * Called with hugetlb_lock locked.
865 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
871 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
873 * If we're returning unused surplus pages, only examine
874 * nodes with surplus pages.
876 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
877 !list_empty(&h
->hugepage_freelists
[node
])) {
879 list_entry(h
->hugepage_freelists
[node
].next
,
881 list_del(&page
->lru
);
882 h
->free_huge_pages
--;
883 h
->free_huge_pages_node
[node
]--;
885 h
->surplus_huge_pages
--;
886 h
->surplus_huge_pages_node
[node
]--;
888 update_and_free_page(h
, page
);
898 * Dissolve a given free hugepage into free buddy pages. This function does
899 * nothing for in-use (including surplus) hugepages.
901 static void dissolve_free_huge_page(struct page
*page
)
903 spin_lock(&hugetlb_lock
);
904 if (PageHuge(page
) && !page_count(page
)) {
905 struct hstate
*h
= page_hstate(page
);
906 int nid
= page_to_nid(page
);
907 list_del(&page
->lru
);
908 h
->free_huge_pages
--;
909 h
->free_huge_pages_node
[nid
]--;
910 update_and_free_page(h
, page
);
912 spin_unlock(&hugetlb_lock
);
916 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
917 * make specified memory blocks removable from the system.
918 * Note that start_pfn should aligned with (minimum) hugepage size.
920 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
922 unsigned int order
= 8 * sizeof(void *);
926 /* Set scan step to minimum hugepage size */
928 if (order
> huge_page_order(h
))
929 order
= huge_page_order(h
);
930 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
931 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
932 dissolve_free_huge_page(pfn_to_page(pfn
));
935 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
940 if (h
->order
>= MAX_ORDER
)
944 * Assume we will successfully allocate the surplus page to
945 * prevent racing processes from causing the surplus to exceed
948 * This however introduces a different race, where a process B
949 * tries to grow the static hugepage pool while alloc_pages() is
950 * called by process A. B will only examine the per-node
951 * counters in determining if surplus huge pages can be
952 * converted to normal huge pages in adjust_pool_surplus(). A
953 * won't be able to increment the per-node counter, until the
954 * lock is dropped by B, but B doesn't drop hugetlb_lock until
955 * no more huge pages can be converted from surplus to normal
956 * state (and doesn't try to convert again). Thus, we have a
957 * case where a surplus huge page exists, the pool is grown, and
958 * the surplus huge page still exists after, even though it
959 * should just have been converted to a normal huge page. This
960 * does not leak memory, though, as the hugepage will be freed
961 * once it is out of use. It also does not allow the counters to
962 * go out of whack in adjust_pool_surplus() as we don't modify
963 * the node values until we've gotten the hugepage and only the
964 * per-node value is checked there.
966 spin_lock(&hugetlb_lock
);
967 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
968 spin_unlock(&hugetlb_lock
);
972 h
->surplus_huge_pages
++;
974 spin_unlock(&hugetlb_lock
);
976 if (nid
== NUMA_NO_NODE
)
977 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
978 __GFP_REPEAT
|__GFP_NOWARN
,
981 page
= alloc_pages_exact_node(nid
,
982 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
983 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
985 if (page
&& arch_prepare_hugepage(page
)) {
986 __free_pages(page
, huge_page_order(h
));
990 spin_lock(&hugetlb_lock
);
992 INIT_LIST_HEAD(&page
->lru
);
993 r_nid
= page_to_nid(page
);
994 set_compound_page_dtor(page
, free_huge_page
);
995 set_hugetlb_cgroup(page
, NULL
);
997 * We incremented the global counters already
999 h
->nr_huge_pages_node
[r_nid
]++;
1000 h
->surplus_huge_pages_node
[r_nid
]++;
1001 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1004 h
->surplus_huge_pages
--;
1005 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1007 spin_unlock(&hugetlb_lock
);
1013 * This allocation function is useful in the context where vma is irrelevant.
1014 * E.g. soft-offlining uses this function because it only cares physical
1015 * address of error page.
1017 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1019 struct page
*page
= NULL
;
1021 spin_lock(&hugetlb_lock
);
1022 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1023 page
= dequeue_huge_page_node(h
, nid
);
1024 spin_unlock(&hugetlb_lock
);
1027 page
= alloc_buddy_huge_page(h
, nid
);
1033 * Increase the hugetlb pool such that it can accommodate a reservation
1036 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1038 struct list_head surplus_list
;
1039 struct page
*page
, *tmp
;
1041 int needed
, allocated
;
1042 bool alloc_ok
= true;
1044 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1046 h
->resv_huge_pages
+= delta
;
1051 INIT_LIST_HEAD(&surplus_list
);
1055 spin_unlock(&hugetlb_lock
);
1056 for (i
= 0; i
< needed
; i
++) {
1057 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1062 list_add(&page
->lru
, &surplus_list
);
1067 * After retaking hugetlb_lock, we need to recalculate 'needed'
1068 * because either resv_huge_pages or free_huge_pages may have changed.
1070 spin_lock(&hugetlb_lock
);
1071 needed
= (h
->resv_huge_pages
+ delta
) -
1072 (h
->free_huge_pages
+ allocated
);
1077 * We were not able to allocate enough pages to
1078 * satisfy the entire reservation so we free what
1079 * we've allocated so far.
1084 * The surplus_list now contains _at_least_ the number of extra pages
1085 * needed to accommodate the reservation. Add the appropriate number
1086 * of pages to the hugetlb pool and free the extras back to the buddy
1087 * allocator. Commit the entire reservation here to prevent another
1088 * process from stealing the pages as they are added to the pool but
1089 * before they are reserved.
1091 needed
+= allocated
;
1092 h
->resv_huge_pages
+= delta
;
1095 /* Free the needed pages to the hugetlb pool */
1096 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1100 * This page is now managed by the hugetlb allocator and has
1101 * no users -- drop the buddy allocator's reference.
1103 put_page_testzero(page
);
1104 VM_BUG_ON(page_count(page
));
1105 enqueue_huge_page(h
, page
);
1108 spin_unlock(&hugetlb_lock
);
1110 /* Free unnecessary surplus pages to the buddy allocator */
1111 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1113 spin_lock(&hugetlb_lock
);
1119 * When releasing a hugetlb pool reservation, any surplus pages that were
1120 * allocated to satisfy the reservation must be explicitly freed if they were
1122 * Called with hugetlb_lock held.
1124 static void return_unused_surplus_pages(struct hstate
*h
,
1125 unsigned long unused_resv_pages
)
1127 unsigned long nr_pages
;
1129 /* Uncommit the reservation */
1130 h
->resv_huge_pages
-= unused_resv_pages
;
1132 /* Cannot return gigantic pages currently */
1133 if (h
->order
>= MAX_ORDER
)
1136 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1139 * We want to release as many surplus pages as possible, spread
1140 * evenly across all nodes with memory. Iterate across these nodes
1141 * until we can no longer free unreserved surplus pages. This occurs
1142 * when the nodes with surplus pages have no free pages.
1143 * free_pool_huge_page() will balance the the freed pages across the
1144 * on-line nodes with memory and will handle the hstate accounting.
1146 while (nr_pages
--) {
1147 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1153 * Determine if the huge page at addr within the vma has an associated
1154 * reservation. Where it does not we will need to logically increase
1155 * reservation and actually increase subpool usage before an allocation
1156 * can occur. Where any new reservation would be required the
1157 * reservation change is prepared, but not committed. Once the page
1158 * has been allocated from the subpool and instantiated the change should
1159 * be committed via vma_commit_reservation. No action is required on
1162 static long vma_needs_reservation(struct hstate
*h
,
1163 struct vm_area_struct
*vma
, unsigned long addr
)
1165 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1166 struct inode
*inode
= mapping
->host
;
1168 if (vma
->vm_flags
& VM_MAYSHARE
) {
1169 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1170 return region_chg(&inode
->i_mapping
->private_list
,
1173 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1178 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1179 struct resv_map
*resv
= vma_resv_map(vma
);
1181 err
= region_chg(&resv
->regions
, idx
, idx
+ 1);
1187 static void vma_commit_reservation(struct hstate
*h
,
1188 struct vm_area_struct
*vma
, unsigned long addr
)
1190 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1191 struct inode
*inode
= mapping
->host
;
1193 if (vma
->vm_flags
& VM_MAYSHARE
) {
1194 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1195 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1197 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1198 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1199 struct resv_map
*resv
= vma_resv_map(vma
);
1201 /* Mark this page used in the map. */
1202 region_add(&resv
->regions
, idx
, idx
+ 1);
1206 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1207 unsigned long addr
, int avoid_reserve
)
1209 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1210 struct hstate
*h
= hstate_vma(vma
);
1214 struct hugetlb_cgroup
*h_cg
;
1216 idx
= hstate_index(h
);
1218 * Processes that did not create the mapping will have no
1219 * reserves and will not have accounted against subpool
1220 * limit. Check that the subpool limit can be made before
1221 * satisfying the allocation MAP_NORESERVE mappings may also
1222 * need pages and subpool limit allocated allocated if no reserve
1225 chg
= vma_needs_reservation(h
, vma
, addr
);
1227 return ERR_PTR(-ENOMEM
);
1228 if (chg
|| avoid_reserve
)
1229 if (hugepage_subpool_get_pages(spool
, 1))
1230 return ERR_PTR(-ENOSPC
);
1232 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1234 if (chg
|| avoid_reserve
)
1235 hugepage_subpool_put_pages(spool
, 1);
1236 return ERR_PTR(-ENOSPC
);
1238 spin_lock(&hugetlb_lock
);
1239 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1241 spin_unlock(&hugetlb_lock
);
1242 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1244 hugetlb_cgroup_uncharge_cgroup(idx
,
1245 pages_per_huge_page(h
),
1247 if (chg
|| avoid_reserve
)
1248 hugepage_subpool_put_pages(spool
, 1);
1249 return ERR_PTR(-ENOSPC
);
1251 spin_lock(&hugetlb_lock
);
1252 list_move(&page
->lru
, &h
->hugepage_activelist
);
1255 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1256 spin_unlock(&hugetlb_lock
);
1258 set_page_private(page
, (unsigned long)spool
);
1260 vma_commit_reservation(h
, vma
, addr
);
1265 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1266 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1267 * where no ERR_VALUE is expected to be returned.
1269 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1270 unsigned long addr
, int avoid_reserve
)
1272 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1278 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1280 struct huge_bootmem_page
*m
;
1283 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1286 addr
= __alloc_bootmem_node_nopanic(NODE_DATA(node
),
1287 huge_page_size(h
), huge_page_size(h
), 0);
1291 * Use the beginning of the huge page to store the
1292 * huge_bootmem_page struct (until gather_bootmem
1293 * puts them into the mem_map).
1302 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1303 /* Put them into a private list first because mem_map is not up yet */
1304 list_add(&m
->list
, &huge_boot_pages
);
1309 static void prep_compound_huge_page(struct page
*page
, int order
)
1311 if (unlikely(order
> (MAX_ORDER
- 1)))
1312 prep_compound_gigantic_page(page
, order
);
1314 prep_compound_page(page
, order
);
1317 /* Put bootmem huge pages into the standard lists after mem_map is up */
1318 static void __init
gather_bootmem_prealloc(void)
1320 struct huge_bootmem_page
*m
;
1322 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1323 struct hstate
*h
= m
->hstate
;
1326 #ifdef CONFIG_HIGHMEM
1327 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1328 free_bootmem_late((unsigned long)m
,
1329 sizeof(struct huge_bootmem_page
));
1331 page
= virt_to_page(m
);
1333 __ClearPageReserved(page
);
1334 WARN_ON(page_count(page
) != 1);
1335 prep_compound_huge_page(page
, h
->order
);
1336 prep_new_huge_page(h
, page
, page_to_nid(page
));
1338 * If we had gigantic hugepages allocated at boot time, we need
1339 * to restore the 'stolen' pages to totalram_pages in order to
1340 * fix confusing memory reports from free(1) and another
1341 * side-effects, like CommitLimit going negative.
1343 if (h
->order
> (MAX_ORDER
- 1))
1344 adjust_managed_page_count(page
, 1 << h
->order
);
1348 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1352 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1353 if (h
->order
>= MAX_ORDER
) {
1354 if (!alloc_bootmem_huge_page(h
))
1356 } else if (!alloc_fresh_huge_page(h
,
1357 &node_states
[N_MEMORY
]))
1360 h
->max_huge_pages
= i
;
1363 static void __init
hugetlb_init_hstates(void)
1367 for_each_hstate(h
) {
1368 /* oversize hugepages were init'ed in early boot */
1369 if (h
->order
< MAX_ORDER
)
1370 hugetlb_hstate_alloc_pages(h
);
1374 static char * __init
memfmt(char *buf
, unsigned long n
)
1376 if (n
>= (1UL << 30))
1377 sprintf(buf
, "%lu GB", n
>> 30);
1378 else if (n
>= (1UL << 20))
1379 sprintf(buf
, "%lu MB", n
>> 20);
1381 sprintf(buf
, "%lu KB", n
>> 10);
1385 static void __init
report_hugepages(void)
1389 for_each_hstate(h
) {
1391 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1392 memfmt(buf
, huge_page_size(h
)),
1393 h
->free_huge_pages
);
1397 #ifdef CONFIG_HIGHMEM
1398 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1399 nodemask_t
*nodes_allowed
)
1403 if (h
->order
>= MAX_ORDER
)
1406 for_each_node_mask(i
, *nodes_allowed
) {
1407 struct page
*page
, *next
;
1408 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1409 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1410 if (count
>= h
->nr_huge_pages
)
1412 if (PageHighMem(page
))
1414 list_del(&page
->lru
);
1415 update_and_free_page(h
, page
);
1416 h
->free_huge_pages
--;
1417 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1422 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1423 nodemask_t
*nodes_allowed
)
1429 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1430 * balanced by operating on them in a round-robin fashion.
1431 * Returns 1 if an adjustment was made.
1433 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1438 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1441 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1442 if (h
->surplus_huge_pages_node
[node
])
1446 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1447 if (h
->surplus_huge_pages_node
[node
] <
1448 h
->nr_huge_pages_node
[node
])
1455 h
->surplus_huge_pages
+= delta
;
1456 h
->surplus_huge_pages_node
[node
] += delta
;
1460 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1461 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1462 nodemask_t
*nodes_allowed
)
1464 unsigned long min_count
, ret
;
1466 if (h
->order
>= MAX_ORDER
)
1467 return h
->max_huge_pages
;
1470 * Increase the pool size
1471 * First take pages out of surplus state. Then make up the
1472 * remaining difference by allocating fresh huge pages.
1474 * We might race with alloc_buddy_huge_page() here and be unable
1475 * to convert a surplus huge page to a normal huge page. That is
1476 * not critical, though, it just means the overall size of the
1477 * pool might be one hugepage larger than it needs to be, but
1478 * within all the constraints specified by the sysctls.
1480 spin_lock(&hugetlb_lock
);
1481 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1482 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1486 while (count
> persistent_huge_pages(h
)) {
1488 * If this allocation races such that we no longer need the
1489 * page, free_huge_page will handle it by freeing the page
1490 * and reducing the surplus.
1492 spin_unlock(&hugetlb_lock
);
1493 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1494 spin_lock(&hugetlb_lock
);
1498 /* Bail for signals. Probably ctrl-c from user */
1499 if (signal_pending(current
))
1504 * Decrease the pool size
1505 * First return free pages to the buddy allocator (being careful
1506 * to keep enough around to satisfy reservations). Then place
1507 * pages into surplus state as needed so the pool will shrink
1508 * to the desired size as pages become free.
1510 * By placing pages into the surplus state independent of the
1511 * overcommit value, we are allowing the surplus pool size to
1512 * exceed overcommit. There are few sane options here. Since
1513 * alloc_buddy_huge_page() is checking the global counter,
1514 * though, we'll note that we're not allowed to exceed surplus
1515 * and won't grow the pool anywhere else. Not until one of the
1516 * sysctls are changed, or the surplus pages go out of use.
1518 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1519 min_count
= max(count
, min_count
);
1520 try_to_free_low(h
, min_count
, nodes_allowed
);
1521 while (min_count
< persistent_huge_pages(h
)) {
1522 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1525 while (count
< persistent_huge_pages(h
)) {
1526 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1530 ret
= persistent_huge_pages(h
);
1531 spin_unlock(&hugetlb_lock
);
1535 #define HSTATE_ATTR_RO(_name) \
1536 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1538 #define HSTATE_ATTR(_name) \
1539 static struct kobj_attribute _name##_attr = \
1540 __ATTR(_name, 0644, _name##_show, _name##_store)
1542 static struct kobject
*hugepages_kobj
;
1543 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1545 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1547 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1551 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1552 if (hstate_kobjs
[i
] == kobj
) {
1554 *nidp
= NUMA_NO_NODE
;
1558 return kobj_to_node_hstate(kobj
, nidp
);
1561 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1562 struct kobj_attribute
*attr
, char *buf
)
1565 unsigned long nr_huge_pages
;
1568 h
= kobj_to_hstate(kobj
, &nid
);
1569 if (nid
== NUMA_NO_NODE
)
1570 nr_huge_pages
= h
->nr_huge_pages
;
1572 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1574 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1577 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1578 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1579 const char *buf
, size_t len
)
1583 unsigned long count
;
1585 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1587 err
= kstrtoul(buf
, 10, &count
);
1591 h
= kobj_to_hstate(kobj
, &nid
);
1592 if (h
->order
>= MAX_ORDER
) {
1597 if (nid
== NUMA_NO_NODE
) {
1599 * global hstate attribute
1601 if (!(obey_mempolicy
&&
1602 init_nodemask_of_mempolicy(nodes_allowed
))) {
1603 NODEMASK_FREE(nodes_allowed
);
1604 nodes_allowed
= &node_states
[N_MEMORY
];
1606 } else if (nodes_allowed
) {
1608 * per node hstate attribute: adjust count to global,
1609 * but restrict alloc/free to the specified node.
1611 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1612 init_nodemask_of_node(nodes_allowed
, nid
);
1614 nodes_allowed
= &node_states
[N_MEMORY
];
1616 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1618 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1619 NODEMASK_FREE(nodes_allowed
);
1623 NODEMASK_FREE(nodes_allowed
);
1627 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1628 struct kobj_attribute
*attr
, char *buf
)
1630 return nr_hugepages_show_common(kobj
, attr
, buf
);
1633 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1634 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1636 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1638 HSTATE_ATTR(nr_hugepages
);
1643 * hstate attribute for optionally mempolicy-based constraint on persistent
1644 * huge page alloc/free.
1646 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1647 struct kobj_attribute
*attr
, char *buf
)
1649 return nr_hugepages_show_common(kobj
, attr
, buf
);
1652 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1653 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1655 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1657 HSTATE_ATTR(nr_hugepages_mempolicy
);
1661 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1662 struct kobj_attribute
*attr
, char *buf
)
1664 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1665 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1668 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1669 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1672 unsigned long input
;
1673 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1675 if (h
->order
>= MAX_ORDER
)
1678 err
= kstrtoul(buf
, 10, &input
);
1682 spin_lock(&hugetlb_lock
);
1683 h
->nr_overcommit_huge_pages
= input
;
1684 spin_unlock(&hugetlb_lock
);
1688 HSTATE_ATTR(nr_overcommit_hugepages
);
1690 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1691 struct kobj_attribute
*attr
, char *buf
)
1694 unsigned long free_huge_pages
;
1697 h
= kobj_to_hstate(kobj
, &nid
);
1698 if (nid
== NUMA_NO_NODE
)
1699 free_huge_pages
= h
->free_huge_pages
;
1701 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1703 return sprintf(buf
, "%lu\n", free_huge_pages
);
1705 HSTATE_ATTR_RO(free_hugepages
);
1707 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1708 struct kobj_attribute
*attr
, char *buf
)
1710 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1711 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1713 HSTATE_ATTR_RO(resv_hugepages
);
1715 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1716 struct kobj_attribute
*attr
, char *buf
)
1719 unsigned long surplus_huge_pages
;
1722 h
= kobj_to_hstate(kobj
, &nid
);
1723 if (nid
== NUMA_NO_NODE
)
1724 surplus_huge_pages
= h
->surplus_huge_pages
;
1726 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1728 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1730 HSTATE_ATTR_RO(surplus_hugepages
);
1732 static struct attribute
*hstate_attrs
[] = {
1733 &nr_hugepages_attr
.attr
,
1734 &nr_overcommit_hugepages_attr
.attr
,
1735 &free_hugepages_attr
.attr
,
1736 &resv_hugepages_attr
.attr
,
1737 &surplus_hugepages_attr
.attr
,
1739 &nr_hugepages_mempolicy_attr
.attr
,
1744 static struct attribute_group hstate_attr_group
= {
1745 .attrs
= hstate_attrs
,
1748 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1749 struct kobject
**hstate_kobjs
,
1750 struct attribute_group
*hstate_attr_group
)
1753 int hi
= hstate_index(h
);
1755 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1756 if (!hstate_kobjs
[hi
])
1759 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1761 kobject_put(hstate_kobjs
[hi
]);
1766 static void __init
hugetlb_sysfs_init(void)
1771 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1772 if (!hugepages_kobj
)
1775 for_each_hstate(h
) {
1776 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1777 hstate_kobjs
, &hstate_attr_group
);
1779 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1786 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1787 * with node devices in node_devices[] using a parallel array. The array
1788 * index of a node device or _hstate == node id.
1789 * This is here to avoid any static dependency of the node device driver, in
1790 * the base kernel, on the hugetlb module.
1792 struct node_hstate
{
1793 struct kobject
*hugepages_kobj
;
1794 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1796 struct node_hstate node_hstates
[MAX_NUMNODES
];
1799 * A subset of global hstate attributes for node devices
1801 static struct attribute
*per_node_hstate_attrs
[] = {
1802 &nr_hugepages_attr
.attr
,
1803 &free_hugepages_attr
.attr
,
1804 &surplus_hugepages_attr
.attr
,
1808 static struct attribute_group per_node_hstate_attr_group
= {
1809 .attrs
= per_node_hstate_attrs
,
1813 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1814 * Returns node id via non-NULL nidp.
1816 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1820 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1821 struct node_hstate
*nhs
= &node_hstates
[nid
];
1823 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1824 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1836 * Unregister hstate attributes from a single node device.
1837 * No-op if no hstate attributes attached.
1839 static void hugetlb_unregister_node(struct node
*node
)
1842 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1844 if (!nhs
->hugepages_kobj
)
1845 return; /* no hstate attributes */
1847 for_each_hstate(h
) {
1848 int idx
= hstate_index(h
);
1849 if (nhs
->hstate_kobjs
[idx
]) {
1850 kobject_put(nhs
->hstate_kobjs
[idx
]);
1851 nhs
->hstate_kobjs
[idx
] = NULL
;
1855 kobject_put(nhs
->hugepages_kobj
);
1856 nhs
->hugepages_kobj
= NULL
;
1860 * hugetlb module exit: unregister hstate attributes from node devices
1863 static void hugetlb_unregister_all_nodes(void)
1868 * disable node device registrations.
1870 register_hugetlbfs_with_node(NULL
, NULL
);
1873 * remove hstate attributes from any nodes that have them.
1875 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1876 hugetlb_unregister_node(node_devices
[nid
]);
1880 * Register hstate attributes for a single node device.
1881 * No-op if attributes already registered.
1883 static void hugetlb_register_node(struct node
*node
)
1886 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1889 if (nhs
->hugepages_kobj
)
1890 return; /* already allocated */
1892 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1894 if (!nhs
->hugepages_kobj
)
1897 for_each_hstate(h
) {
1898 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1900 &per_node_hstate_attr_group
);
1902 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1903 h
->name
, node
->dev
.id
);
1904 hugetlb_unregister_node(node
);
1911 * hugetlb init time: register hstate attributes for all registered node
1912 * devices of nodes that have memory. All on-line nodes should have
1913 * registered their associated device by this time.
1915 static void hugetlb_register_all_nodes(void)
1919 for_each_node_state(nid
, N_MEMORY
) {
1920 struct node
*node
= node_devices
[nid
];
1921 if (node
->dev
.id
== nid
)
1922 hugetlb_register_node(node
);
1926 * Let the node device driver know we're here so it can
1927 * [un]register hstate attributes on node hotplug.
1929 register_hugetlbfs_with_node(hugetlb_register_node
,
1930 hugetlb_unregister_node
);
1932 #else /* !CONFIG_NUMA */
1934 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1942 static void hugetlb_unregister_all_nodes(void) { }
1944 static void hugetlb_register_all_nodes(void) { }
1948 static void __exit
hugetlb_exit(void)
1952 hugetlb_unregister_all_nodes();
1954 for_each_hstate(h
) {
1955 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1958 kobject_put(hugepages_kobj
);
1960 module_exit(hugetlb_exit
);
1962 static int __init
hugetlb_init(void)
1964 /* Some platform decide whether they support huge pages at boot
1965 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1966 * there is no such support
1968 if (HPAGE_SHIFT
== 0)
1971 if (!size_to_hstate(default_hstate_size
)) {
1972 default_hstate_size
= HPAGE_SIZE
;
1973 if (!size_to_hstate(default_hstate_size
))
1974 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1976 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1977 if (default_hstate_max_huge_pages
)
1978 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1980 hugetlb_init_hstates();
1981 gather_bootmem_prealloc();
1984 hugetlb_sysfs_init();
1985 hugetlb_register_all_nodes();
1986 hugetlb_cgroup_file_init();
1990 module_init(hugetlb_init
);
1992 /* Should be called on processing a hugepagesz=... option */
1993 void __init
hugetlb_add_hstate(unsigned order
)
1998 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1999 pr_warning("hugepagesz= specified twice, ignoring\n");
2002 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2004 h
= &hstates
[hugetlb_max_hstate
++];
2006 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2007 h
->nr_huge_pages
= 0;
2008 h
->free_huge_pages
= 0;
2009 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2010 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2011 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2012 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2013 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2014 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2015 huge_page_size(h
)/1024);
2020 static int __init
hugetlb_nrpages_setup(char *s
)
2023 static unsigned long *last_mhp
;
2026 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2027 * so this hugepages= parameter goes to the "default hstate".
2029 if (!hugetlb_max_hstate
)
2030 mhp
= &default_hstate_max_huge_pages
;
2032 mhp
= &parsed_hstate
->max_huge_pages
;
2034 if (mhp
== last_mhp
) {
2035 pr_warning("hugepages= specified twice without "
2036 "interleaving hugepagesz=, ignoring\n");
2040 if (sscanf(s
, "%lu", mhp
) <= 0)
2044 * Global state is always initialized later in hugetlb_init.
2045 * But we need to allocate >= MAX_ORDER hstates here early to still
2046 * use the bootmem allocator.
2048 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2049 hugetlb_hstate_alloc_pages(parsed_hstate
);
2055 __setup("hugepages=", hugetlb_nrpages_setup
);
2057 static int __init
hugetlb_default_setup(char *s
)
2059 default_hstate_size
= memparse(s
, &s
);
2062 __setup("default_hugepagesz=", hugetlb_default_setup
);
2064 static unsigned int cpuset_mems_nr(unsigned int *array
)
2067 unsigned int nr
= 0;
2069 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2075 #ifdef CONFIG_SYSCTL
2076 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2077 struct ctl_table
*table
, int write
,
2078 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2080 struct hstate
*h
= &default_hstate
;
2084 tmp
= h
->max_huge_pages
;
2086 if (write
&& h
->order
>= MAX_ORDER
)
2090 table
->maxlen
= sizeof(unsigned long);
2091 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2096 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2097 GFP_KERNEL
| __GFP_NORETRY
);
2098 if (!(obey_mempolicy
&&
2099 init_nodemask_of_mempolicy(nodes_allowed
))) {
2100 NODEMASK_FREE(nodes_allowed
);
2101 nodes_allowed
= &node_states
[N_MEMORY
];
2103 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2105 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2106 NODEMASK_FREE(nodes_allowed
);
2112 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2113 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2116 return hugetlb_sysctl_handler_common(false, table
, write
,
2117 buffer
, length
, ppos
);
2121 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2122 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2124 return hugetlb_sysctl_handler_common(true, table
, write
,
2125 buffer
, length
, ppos
);
2127 #endif /* CONFIG_NUMA */
2129 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2130 void __user
*buffer
,
2131 size_t *length
, loff_t
*ppos
)
2133 struct hstate
*h
= &default_hstate
;
2137 tmp
= h
->nr_overcommit_huge_pages
;
2139 if (write
&& h
->order
>= MAX_ORDER
)
2143 table
->maxlen
= sizeof(unsigned long);
2144 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2149 spin_lock(&hugetlb_lock
);
2150 h
->nr_overcommit_huge_pages
= tmp
;
2151 spin_unlock(&hugetlb_lock
);
2157 #endif /* CONFIG_SYSCTL */
2159 void hugetlb_report_meminfo(struct seq_file
*m
)
2161 struct hstate
*h
= &default_hstate
;
2163 "HugePages_Total: %5lu\n"
2164 "HugePages_Free: %5lu\n"
2165 "HugePages_Rsvd: %5lu\n"
2166 "HugePages_Surp: %5lu\n"
2167 "Hugepagesize: %8lu kB\n",
2171 h
->surplus_huge_pages
,
2172 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2175 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2177 struct hstate
*h
= &default_hstate
;
2179 "Node %d HugePages_Total: %5u\n"
2180 "Node %d HugePages_Free: %5u\n"
2181 "Node %d HugePages_Surp: %5u\n",
2182 nid
, h
->nr_huge_pages_node
[nid
],
2183 nid
, h
->free_huge_pages_node
[nid
],
2184 nid
, h
->surplus_huge_pages_node
[nid
]);
2187 void hugetlb_show_meminfo(void)
2192 for_each_node_state(nid
, N_MEMORY
)
2194 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2196 h
->nr_huge_pages_node
[nid
],
2197 h
->free_huge_pages_node
[nid
],
2198 h
->surplus_huge_pages_node
[nid
],
2199 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2202 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2203 unsigned long hugetlb_total_pages(void)
2206 unsigned long nr_total_pages
= 0;
2209 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2210 return nr_total_pages
;
2213 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2217 spin_lock(&hugetlb_lock
);
2219 * When cpuset is configured, it breaks the strict hugetlb page
2220 * reservation as the accounting is done on a global variable. Such
2221 * reservation is completely rubbish in the presence of cpuset because
2222 * the reservation is not checked against page availability for the
2223 * current cpuset. Application can still potentially OOM'ed by kernel
2224 * with lack of free htlb page in cpuset that the task is in.
2225 * Attempt to enforce strict accounting with cpuset is almost
2226 * impossible (or too ugly) because cpuset is too fluid that
2227 * task or memory node can be dynamically moved between cpusets.
2229 * The change of semantics for shared hugetlb mapping with cpuset is
2230 * undesirable. However, in order to preserve some of the semantics,
2231 * we fall back to check against current free page availability as
2232 * a best attempt and hopefully to minimize the impact of changing
2233 * semantics that cpuset has.
2236 if (gather_surplus_pages(h
, delta
) < 0)
2239 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2240 return_unused_surplus_pages(h
, delta
);
2247 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2250 spin_unlock(&hugetlb_lock
);
2254 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2256 struct resv_map
*resv
= vma_resv_map(vma
);
2259 * This new VMA should share its siblings reservation map if present.
2260 * The VMA will only ever have a valid reservation map pointer where
2261 * it is being copied for another still existing VMA. As that VMA
2262 * has a reference to the reservation map it cannot disappear until
2263 * after this open call completes. It is therefore safe to take a
2264 * new reference here without additional locking.
2267 kref_get(&resv
->refs
);
2270 static void resv_map_put(struct vm_area_struct
*vma
)
2272 struct resv_map
*resv
= vma_resv_map(vma
);
2276 kref_put(&resv
->refs
, resv_map_release
);
2279 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2281 struct hstate
*h
= hstate_vma(vma
);
2282 struct resv_map
*resv
= vma_resv_map(vma
);
2283 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2284 unsigned long reserve
;
2285 unsigned long start
;
2289 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2290 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2292 reserve
= (end
- start
) -
2293 region_count(&resv
->regions
, start
, end
);
2298 hugetlb_acct_memory(h
, -reserve
);
2299 hugepage_subpool_put_pages(spool
, reserve
);
2305 * We cannot handle pagefaults against hugetlb pages at all. They cause
2306 * handle_mm_fault() to try to instantiate regular-sized pages in the
2307 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2310 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2316 const struct vm_operations_struct hugetlb_vm_ops
= {
2317 .fault
= hugetlb_vm_op_fault
,
2318 .open
= hugetlb_vm_op_open
,
2319 .close
= hugetlb_vm_op_close
,
2322 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2328 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2329 vma
->vm_page_prot
)));
2331 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2332 vma
->vm_page_prot
));
2334 entry
= pte_mkyoung(entry
);
2335 entry
= pte_mkhuge(entry
);
2336 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2341 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2342 unsigned long address
, pte_t
*ptep
)
2346 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2347 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2348 update_mmu_cache(vma
, address
, ptep
);
2352 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2353 struct vm_area_struct
*vma
)
2355 pte_t
*src_pte
, *dst_pte
, entry
;
2356 struct page
*ptepage
;
2359 struct hstate
*h
= hstate_vma(vma
);
2360 unsigned long sz
= huge_page_size(h
);
2362 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2364 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2365 src_pte
= huge_pte_offset(src
, addr
);
2368 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2372 /* If the pagetables are shared don't copy or take references */
2373 if (dst_pte
== src_pte
)
2376 spin_lock(&dst
->page_table_lock
);
2377 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2378 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2380 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2381 entry
= huge_ptep_get(src_pte
);
2382 ptepage
= pte_page(entry
);
2384 page_dup_rmap(ptepage
);
2385 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2387 spin_unlock(&src
->page_table_lock
);
2388 spin_unlock(&dst
->page_table_lock
);
2396 static int is_hugetlb_entry_migration(pte_t pte
)
2400 if (huge_pte_none(pte
) || pte_present(pte
))
2402 swp
= pte_to_swp_entry(pte
);
2403 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2409 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2413 if (huge_pte_none(pte
) || pte_present(pte
))
2415 swp
= pte_to_swp_entry(pte
);
2416 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2422 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2423 unsigned long start
, unsigned long end
,
2424 struct page
*ref_page
)
2426 int force_flush
= 0;
2427 struct mm_struct
*mm
= vma
->vm_mm
;
2428 unsigned long address
;
2432 struct hstate
*h
= hstate_vma(vma
);
2433 unsigned long sz
= huge_page_size(h
);
2434 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2435 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2437 WARN_ON(!is_vm_hugetlb_page(vma
));
2438 BUG_ON(start
& ~huge_page_mask(h
));
2439 BUG_ON(end
& ~huge_page_mask(h
));
2441 tlb_start_vma(tlb
, vma
);
2442 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2444 spin_lock(&mm
->page_table_lock
);
2445 for (address
= start
; address
< end
; address
+= sz
) {
2446 ptep
= huge_pte_offset(mm
, address
);
2450 if (huge_pmd_unshare(mm
, &address
, ptep
))
2453 pte
= huge_ptep_get(ptep
);
2454 if (huge_pte_none(pte
))
2458 * HWPoisoned hugepage is already unmapped and dropped reference
2460 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2461 huge_pte_clear(mm
, address
, ptep
);
2465 page
= pte_page(pte
);
2467 * If a reference page is supplied, it is because a specific
2468 * page is being unmapped, not a range. Ensure the page we
2469 * are about to unmap is the actual page of interest.
2472 if (page
!= ref_page
)
2476 * Mark the VMA as having unmapped its page so that
2477 * future faults in this VMA will fail rather than
2478 * looking like data was lost
2480 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2483 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2484 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2485 if (huge_pte_dirty(pte
))
2486 set_page_dirty(page
);
2488 page_remove_rmap(page
);
2489 force_flush
= !__tlb_remove_page(tlb
, page
);
2492 /* Bail out after unmapping reference page if supplied */
2496 spin_unlock(&mm
->page_table_lock
);
2498 * mmu_gather ran out of room to batch pages, we break out of
2499 * the PTE lock to avoid doing the potential expensive TLB invalidate
2500 * and page-free while holding it.
2505 if (address
< end
&& !ref_page
)
2508 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2509 tlb_end_vma(tlb
, vma
);
2512 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2513 struct vm_area_struct
*vma
, unsigned long start
,
2514 unsigned long end
, struct page
*ref_page
)
2516 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2519 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2520 * test will fail on a vma being torn down, and not grab a page table
2521 * on its way out. We're lucky that the flag has such an appropriate
2522 * name, and can in fact be safely cleared here. We could clear it
2523 * before the __unmap_hugepage_range above, but all that's necessary
2524 * is to clear it before releasing the i_mmap_mutex. This works
2525 * because in the context this is called, the VMA is about to be
2526 * destroyed and the i_mmap_mutex is held.
2528 vma
->vm_flags
&= ~VM_MAYSHARE
;
2531 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2532 unsigned long end
, struct page
*ref_page
)
2534 struct mm_struct
*mm
;
2535 struct mmu_gather tlb
;
2539 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2540 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2541 tlb_finish_mmu(&tlb
, start
, end
);
2545 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2546 * mappping it owns the reserve page for. The intention is to unmap the page
2547 * from other VMAs and let the children be SIGKILLed if they are faulting the
2550 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2551 struct page
*page
, unsigned long address
)
2553 struct hstate
*h
= hstate_vma(vma
);
2554 struct vm_area_struct
*iter_vma
;
2555 struct address_space
*mapping
;
2559 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2560 * from page cache lookup which is in HPAGE_SIZE units.
2562 address
= address
& huge_page_mask(h
);
2563 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2565 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2568 * Take the mapping lock for the duration of the table walk. As
2569 * this mapping should be shared between all the VMAs,
2570 * __unmap_hugepage_range() is called as the lock is already held
2572 mutex_lock(&mapping
->i_mmap_mutex
);
2573 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2574 /* Do not unmap the current VMA */
2575 if (iter_vma
== vma
)
2579 * Unmap the page from other VMAs without their own reserves.
2580 * They get marked to be SIGKILLed if they fault in these
2581 * areas. This is because a future no-page fault on this VMA
2582 * could insert a zeroed page instead of the data existing
2583 * from the time of fork. This would look like data corruption
2585 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2586 unmap_hugepage_range(iter_vma
, address
,
2587 address
+ huge_page_size(h
), page
);
2589 mutex_unlock(&mapping
->i_mmap_mutex
);
2595 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2596 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2597 * cannot race with other handlers or page migration.
2598 * Keep the pte_same checks anyway to make transition from the mutex easier.
2600 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2601 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2602 struct page
*pagecache_page
)
2604 struct hstate
*h
= hstate_vma(vma
);
2605 struct page
*old_page
, *new_page
;
2606 int outside_reserve
= 0;
2607 unsigned long mmun_start
; /* For mmu_notifiers */
2608 unsigned long mmun_end
; /* For mmu_notifiers */
2610 old_page
= pte_page(pte
);
2613 /* If no-one else is actually using this page, avoid the copy
2614 * and just make the page writable */
2615 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2616 page_move_anon_rmap(old_page
, vma
, address
);
2617 set_huge_ptep_writable(vma
, address
, ptep
);
2622 * If the process that created a MAP_PRIVATE mapping is about to
2623 * perform a COW due to a shared page count, attempt to satisfy
2624 * the allocation without using the existing reserves. The pagecache
2625 * page is used to determine if the reserve at this address was
2626 * consumed or not. If reserves were used, a partial faulted mapping
2627 * at the time of fork() could consume its reserves on COW instead
2628 * of the full address range.
2630 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2631 old_page
!= pagecache_page
)
2632 outside_reserve
= 1;
2634 page_cache_get(old_page
);
2636 /* Drop page_table_lock as buddy allocator may be called */
2637 spin_unlock(&mm
->page_table_lock
);
2638 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2640 if (IS_ERR(new_page
)) {
2641 long err
= PTR_ERR(new_page
);
2642 page_cache_release(old_page
);
2645 * If a process owning a MAP_PRIVATE mapping fails to COW,
2646 * it is due to references held by a child and an insufficient
2647 * huge page pool. To guarantee the original mappers
2648 * reliability, unmap the page from child processes. The child
2649 * may get SIGKILLed if it later faults.
2651 if (outside_reserve
) {
2652 BUG_ON(huge_pte_none(pte
));
2653 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2654 BUG_ON(huge_pte_none(pte
));
2655 spin_lock(&mm
->page_table_lock
);
2656 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2657 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2658 goto retry_avoidcopy
;
2660 * race occurs while re-acquiring page_table_lock, and
2668 /* Caller expects lock to be held */
2669 spin_lock(&mm
->page_table_lock
);
2671 return VM_FAULT_OOM
;
2673 return VM_FAULT_SIGBUS
;
2677 * When the original hugepage is shared one, it does not have
2678 * anon_vma prepared.
2680 if (unlikely(anon_vma_prepare(vma
))) {
2681 page_cache_release(new_page
);
2682 page_cache_release(old_page
);
2683 /* Caller expects lock to be held */
2684 spin_lock(&mm
->page_table_lock
);
2685 return VM_FAULT_OOM
;
2688 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2689 pages_per_huge_page(h
));
2690 __SetPageUptodate(new_page
);
2692 mmun_start
= address
& huge_page_mask(h
);
2693 mmun_end
= mmun_start
+ huge_page_size(h
);
2694 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2696 * Retake the page_table_lock to check for racing updates
2697 * before the page tables are altered
2699 spin_lock(&mm
->page_table_lock
);
2700 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2701 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2702 ClearPagePrivate(new_page
);
2705 huge_ptep_clear_flush(vma
, address
, ptep
);
2706 set_huge_pte_at(mm
, address
, ptep
,
2707 make_huge_pte(vma
, new_page
, 1));
2708 page_remove_rmap(old_page
);
2709 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2710 /* Make the old page be freed below */
2711 new_page
= old_page
;
2713 spin_unlock(&mm
->page_table_lock
);
2714 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2715 page_cache_release(new_page
);
2716 page_cache_release(old_page
);
2718 /* Caller expects lock to be held */
2719 spin_lock(&mm
->page_table_lock
);
2723 /* Return the pagecache page at a given address within a VMA */
2724 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2725 struct vm_area_struct
*vma
, unsigned long address
)
2727 struct address_space
*mapping
;
2730 mapping
= vma
->vm_file
->f_mapping
;
2731 idx
= vma_hugecache_offset(h
, vma
, address
);
2733 return find_lock_page(mapping
, idx
);
2737 * Return whether there is a pagecache page to back given address within VMA.
2738 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2740 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2741 struct vm_area_struct
*vma
, unsigned long address
)
2743 struct address_space
*mapping
;
2747 mapping
= vma
->vm_file
->f_mapping
;
2748 idx
= vma_hugecache_offset(h
, vma
, address
);
2750 page
= find_get_page(mapping
, idx
);
2753 return page
!= NULL
;
2756 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2757 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2759 struct hstate
*h
= hstate_vma(vma
);
2760 int ret
= VM_FAULT_SIGBUS
;
2765 struct address_space
*mapping
;
2769 * Currently, we are forced to kill the process in the event the
2770 * original mapper has unmapped pages from the child due to a failed
2771 * COW. Warn that such a situation has occurred as it may not be obvious
2773 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2774 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2779 mapping
= vma
->vm_file
->f_mapping
;
2780 idx
= vma_hugecache_offset(h
, vma
, address
);
2783 * Use page lock to guard against racing truncation
2784 * before we get page_table_lock.
2787 page
= find_lock_page(mapping
, idx
);
2789 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2792 page
= alloc_huge_page(vma
, address
, 0);
2794 ret
= PTR_ERR(page
);
2798 ret
= VM_FAULT_SIGBUS
;
2801 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2802 __SetPageUptodate(page
);
2804 if (vma
->vm_flags
& VM_MAYSHARE
) {
2806 struct inode
*inode
= mapping
->host
;
2808 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2815 ClearPagePrivate(page
);
2817 spin_lock(&inode
->i_lock
);
2818 inode
->i_blocks
+= blocks_per_huge_page(h
);
2819 spin_unlock(&inode
->i_lock
);
2822 if (unlikely(anon_vma_prepare(vma
))) {
2824 goto backout_unlocked
;
2830 * If memory error occurs between mmap() and fault, some process
2831 * don't have hwpoisoned swap entry for errored virtual address.
2832 * So we need to block hugepage fault by PG_hwpoison bit check.
2834 if (unlikely(PageHWPoison(page
))) {
2835 ret
= VM_FAULT_HWPOISON
|
2836 VM_FAULT_SET_HINDEX(hstate_index(h
));
2837 goto backout_unlocked
;
2842 * If we are going to COW a private mapping later, we examine the
2843 * pending reservations for this page now. This will ensure that
2844 * any allocations necessary to record that reservation occur outside
2847 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2848 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2850 goto backout_unlocked
;
2853 spin_lock(&mm
->page_table_lock
);
2854 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2859 if (!huge_pte_none(huge_ptep_get(ptep
)))
2863 ClearPagePrivate(page
);
2864 hugepage_add_new_anon_rmap(page
, vma
, address
);
2867 page_dup_rmap(page
);
2868 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2869 && (vma
->vm_flags
& VM_SHARED
)));
2870 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2872 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2873 /* Optimization, do the COW without a second fault */
2874 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2877 spin_unlock(&mm
->page_table_lock
);
2883 spin_unlock(&mm
->page_table_lock
);
2890 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2891 unsigned long address
, unsigned int flags
)
2896 struct page
*page
= NULL
;
2897 struct page
*pagecache_page
= NULL
;
2898 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2899 struct hstate
*h
= hstate_vma(vma
);
2901 address
&= huge_page_mask(h
);
2903 ptep
= huge_pte_offset(mm
, address
);
2905 entry
= huge_ptep_get(ptep
);
2906 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2907 migration_entry_wait_huge(mm
, ptep
);
2909 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2910 return VM_FAULT_HWPOISON_LARGE
|
2911 VM_FAULT_SET_HINDEX(hstate_index(h
));
2914 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2916 return VM_FAULT_OOM
;
2919 * Serialize hugepage allocation and instantiation, so that we don't
2920 * get spurious allocation failures if two CPUs race to instantiate
2921 * the same page in the page cache.
2923 mutex_lock(&hugetlb_instantiation_mutex
);
2924 entry
= huge_ptep_get(ptep
);
2925 if (huge_pte_none(entry
)) {
2926 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2933 * If we are going to COW the mapping later, we examine the pending
2934 * reservations for this page now. This will ensure that any
2935 * allocations necessary to record that reservation occur outside the
2936 * spinlock. For private mappings, we also lookup the pagecache
2937 * page now as it is used to determine if a reservation has been
2940 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2941 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2946 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2947 pagecache_page
= hugetlbfs_pagecache_page(h
,
2952 * hugetlb_cow() requires page locks of pte_page(entry) and
2953 * pagecache_page, so here we need take the former one
2954 * when page != pagecache_page or !pagecache_page.
2955 * Note that locking order is always pagecache_page -> page,
2956 * so no worry about deadlock.
2958 page
= pte_page(entry
);
2960 if (page
!= pagecache_page
)
2963 spin_lock(&mm
->page_table_lock
);
2964 /* Check for a racing update before calling hugetlb_cow */
2965 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2966 goto out_page_table_lock
;
2969 if (flags
& FAULT_FLAG_WRITE
) {
2970 if (!huge_pte_write(entry
)) {
2971 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2973 goto out_page_table_lock
;
2975 entry
= huge_pte_mkdirty(entry
);
2977 entry
= pte_mkyoung(entry
);
2978 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2979 flags
& FAULT_FLAG_WRITE
))
2980 update_mmu_cache(vma
, address
, ptep
);
2982 out_page_table_lock
:
2983 spin_unlock(&mm
->page_table_lock
);
2985 if (pagecache_page
) {
2986 unlock_page(pagecache_page
);
2987 put_page(pagecache_page
);
2989 if (page
!= pagecache_page
)
2994 mutex_unlock(&hugetlb_instantiation_mutex
);
2999 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3000 struct page
**pages
, struct vm_area_struct
**vmas
,
3001 unsigned long *position
, unsigned long *nr_pages
,
3002 long i
, unsigned int flags
)
3004 unsigned long pfn_offset
;
3005 unsigned long vaddr
= *position
;
3006 unsigned long remainder
= *nr_pages
;
3007 struct hstate
*h
= hstate_vma(vma
);
3009 spin_lock(&mm
->page_table_lock
);
3010 while (vaddr
< vma
->vm_end
&& remainder
) {
3016 * Some archs (sparc64, sh*) have multiple pte_ts to
3017 * each hugepage. We have to make sure we get the
3018 * first, for the page indexing below to work.
3020 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3021 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3024 * When coredumping, it suits get_dump_page if we just return
3025 * an error where there's an empty slot with no huge pagecache
3026 * to back it. This way, we avoid allocating a hugepage, and
3027 * the sparse dumpfile avoids allocating disk blocks, but its
3028 * huge holes still show up with zeroes where they need to be.
3030 if (absent
&& (flags
& FOLL_DUMP
) &&
3031 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3037 * We need call hugetlb_fault for both hugepages under migration
3038 * (in which case hugetlb_fault waits for the migration,) and
3039 * hwpoisoned hugepages (in which case we need to prevent the
3040 * caller from accessing to them.) In order to do this, we use
3041 * here is_swap_pte instead of is_hugetlb_entry_migration and
3042 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3043 * both cases, and because we can't follow correct pages
3044 * directly from any kind of swap entries.
3046 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3047 ((flags
& FOLL_WRITE
) &&
3048 !huge_pte_write(huge_ptep_get(pte
)))) {
3051 spin_unlock(&mm
->page_table_lock
);
3052 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3053 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3054 spin_lock(&mm
->page_table_lock
);
3055 if (!(ret
& VM_FAULT_ERROR
))
3062 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3063 page
= pte_page(huge_ptep_get(pte
));
3066 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3077 if (vaddr
< vma
->vm_end
&& remainder
&&
3078 pfn_offset
< pages_per_huge_page(h
)) {
3080 * We use pfn_offset to avoid touching the pageframes
3081 * of this compound page.
3086 spin_unlock(&mm
->page_table_lock
);
3087 *nr_pages
= remainder
;
3090 return i
? i
: -EFAULT
;
3093 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3094 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3096 struct mm_struct
*mm
= vma
->vm_mm
;
3097 unsigned long start
= address
;
3100 struct hstate
*h
= hstate_vma(vma
);
3101 unsigned long pages
= 0;
3103 BUG_ON(address
>= end
);
3104 flush_cache_range(vma
, address
, end
);
3106 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3107 spin_lock(&mm
->page_table_lock
);
3108 for (; address
< end
; address
+= huge_page_size(h
)) {
3109 ptep
= huge_pte_offset(mm
, address
);
3112 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3116 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3117 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3118 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3119 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3120 set_huge_pte_at(mm
, address
, ptep
, pte
);
3124 spin_unlock(&mm
->page_table_lock
);
3126 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3127 * may have cleared our pud entry and done put_page on the page table:
3128 * once we release i_mmap_mutex, another task can do the final put_page
3129 * and that page table be reused and filled with junk.
3131 flush_tlb_range(vma
, start
, end
);
3132 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3134 return pages
<< h
->order
;
3137 int hugetlb_reserve_pages(struct inode
*inode
,
3139 struct vm_area_struct
*vma
,
3140 vm_flags_t vm_flags
)
3143 struct hstate
*h
= hstate_inode(inode
);
3144 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3147 * Only apply hugepage reservation if asked. At fault time, an
3148 * attempt will be made for VM_NORESERVE to allocate a page
3149 * without using reserves
3151 if (vm_flags
& VM_NORESERVE
)
3155 * Shared mappings base their reservation on the number of pages that
3156 * are already allocated on behalf of the file. Private mappings need
3157 * to reserve the full area even if read-only as mprotect() may be
3158 * called to make the mapping read-write. Assume !vma is a shm mapping
3160 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3161 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3163 struct resv_map
*resv_map
= resv_map_alloc();
3169 set_vma_resv_map(vma
, resv_map
);
3170 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3178 /* There must be enough pages in the subpool for the mapping */
3179 if (hugepage_subpool_get_pages(spool
, chg
)) {
3185 * Check enough hugepages are available for the reservation.
3186 * Hand the pages back to the subpool if there are not
3188 ret
= hugetlb_acct_memory(h
, chg
);
3190 hugepage_subpool_put_pages(spool
, chg
);
3195 * Account for the reservations made. Shared mappings record regions
3196 * that have reservations as they are shared by multiple VMAs.
3197 * When the last VMA disappears, the region map says how much
3198 * the reservation was and the page cache tells how much of
3199 * the reservation was consumed. Private mappings are per-VMA and
3200 * only the consumed reservations are tracked. When the VMA
3201 * disappears, the original reservation is the VMA size and the
3202 * consumed reservations are stored in the map. Hence, nothing
3203 * else has to be done for private mappings here
3205 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3206 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3214 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3216 struct hstate
*h
= hstate_inode(inode
);
3217 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3218 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3220 spin_lock(&inode
->i_lock
);
3221 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3222 spin_unlock(&inode
->i_lock
);
3224 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3225 hugetlb_acct_memory(h
, -(chg
- freed
));
3228 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3229 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3230 struct vm_area_struct
*vma
,
3231 unsigned long addr
, pgoff_t idx
)
3233 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3235 unsigned long sbase
= saddr
& PUD_MASK
;
3236 unsigned long s_end
= sbase
+ PUD_SIZE
;
3238 /* Allow segments to share if only one is marked locked */
3239 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3240 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3243 * match the virtual addresses, permission and the alignment of the
3246 if (pmd_index(addr
) != pmd_index(saddr
) ||
3247 vm_flags
!= svm_flags
||
3248 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3254 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3256 unsigned long base
= addr
& PUD_MASK
;
3257 unsigned long end
= base
+ PUD_SIZE
;
3260 * check on proper vm_flags and page table alignment
3262 if (vma
->vm_flags
& VM_MAYSHARE
&&
3263 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3269 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3270 * and returns the corresponding pte. While this is not necessary for the
3271 * !shared pmd case because we can allocate the pmd later as well, it makes the
3272 * code much cleaner. pmd allocation is essential for the shared case because
3273 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3274 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3275 * bad pmd for sharing.
3277 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3279 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3280 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3281 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3283 struct vm_area_struct
*svma
;
3284 unsigned long saddr
;
3288 if (!vma_shareable(vma
, addr
))
3289 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3291 mutex_lock(&mapping
->i_mmap_mutex
);
3292 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3296 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3298 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3300 get_page(virt_to_page(spte
));
3309 spin_lock(&mm
->page_table_lock
);
3311 pud_populate(mm
, pud
,
3312 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3314 put_page(virt_to_page(spte
));
3315 spin_unlock(&mm
->page_table_lock
);
3317 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3318 mutex_unlock(&mapping
->i_mmap_mutex
);
3323 * unmap huge page backed by shared pte.
3325 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3326 * indicated by page_count > 1, unmap is achieved by clearing pud and
3327 * decrementing the ref count. If count == 1, the pte page is not shared.
3329 * called with vma->vm_mm->page_table_lock held.
3331 * returns: 1 successfully unmapped a shared pte page
3332 * 0 the underlying pte page is not shared, or it is the last user
3334 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3336 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3337 pud_t
*pud
= pud_offset(pgd
, *addr
);
3339 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3340 if (page_count(virt_to_page(ptep
)) == 1)
3344 put_page(virt_to_page(ptep
));
3345 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3348 #define want_pmd_share() (1)
3349 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3350 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3354 #define want_pmd_share() (0)
3355 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3357 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3358 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3359 unsigned long addr
, unsigned long sz
)
3365 pgd
= pgd_offset(mm
, addr
);
3366 pud
= pud_alloc(mm
, pgd
, addr
);
3368 if (sz
== PUD_SIZE
) {
3371 BUG_ON(sz
!= PMD_SIZE
);
3372 if (want_pmd_share() && pud_none(*pud
))
3373 pte
= huge_pmd_share(mm
, addr
, pud
);
3375 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3378 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3383 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3389 pgd
= pgd_offset(mm
, addr
);
3390 if (pgd_present(*pgd
)) {
3391 pud
= pud_offset(pgd
, addr
);
3392 if (pud_present(*pud
)) {
3394 return (pte_t
*)pud
;
3395 pmd
= pmd_offset(pud
, addr
);
3398 return (pte_t
*) pmd
;
3402 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3403 pmd_t
*pmd
, int write
)
3407 page
= pte_page(*(pte_t
*)pmd
);
3409 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3414 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3415 pud_t
*pud
, int write
)
3419 page
= pte_page(*(pte_t
*)pud
);
3421 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3425 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3427 /* Can be overriden by architectures */
3428 __attribute__((weak
)) struct page
*
3429 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3430 pud_t
*pud
, int write
)
3436 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3438 #ifdef CONFIG_MEMORY_FAILURE
3440 /* Should be called in hugetlb_lock */
3441 static int is_hugepage_on_freelist(struct page
*hpage
)
3445 struct hstate
*h
= page_hstate(hpage
);
3446 int nid
= page_to_nid(hpage
);
3448 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3455 * This function is called from memory failure code.
3456 * Assume the caller holds page lock of the head page.
3458 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3460 struct hstate
*h
= page_hstate(hpage
);
3461 int nid
= page_to_nid(hpage
);
3464 spin_lock(&hugetlb_lock
);
3465 if (is_hugepage_on_freelist(hpage
)) {
3467 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3468 * but dangling hpage->lru can trigger list-debug warnings
3469 * (this happens when we call unpoison_memory() on it),
3470 * so let it point to itself with list_del_init().
3472 list_del_init(&hpage
->lru
);
3473 set_page_refcounted(hpage
);
3474 h
->free_huge_pages
--;
3475 h
->free_huge_pages_node
[nid
]--;
3478 spin_unlock(&hugetlb_lock
);
3483 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3485 VM_BUG_ON(!PageHead(page
));
3486 if (!get_page_unless_zero(page
))
3488 spin_lock(&hugetlb_lock
);
3489 list_move_tail(&page
->lru
, list
);
3490 spin_unlock(&hugetlb_lock
);
3494 void putback_active_hugepage(struct page
*page
)
3496 VM_BUG_ON(!PageHead(page
));
3497 spin_lock(&hugetlb_lock
);
3498 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3499 spin_unlock(&hugetlb_lock
);
3503 bool is_hugepage_active(struct page
*page
)
3505 VM_BUG_ON(!PageHuge(page
));
3507 * This function can be called for a tail page because the caller,
3508 * scan_movable_pages, scans through a given pfn-range which typically
3509 * covers one memory block. In systems using gigantic hugepage (1GB
3510 * for x86_64,) a hugepage is larger than a memory block, and we don't
3511 * support migrating such large hugepages for now, so return false
3512 * when called for tail pages.
3517 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3518 * so we should return false for them.
3520 if (unlikely(PageHWPoison(page
)))
3522 return page_count(page
) > 0;