2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
27 #include <asm/pgtable.h>
31 #include <linux/hugetlb.h>
32 #include <linux/hugetlb_cgroup.h>
33 #include <linux/node.h>
36 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
37 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
38 unsigned long hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 __initdata
LIST_HEAD(huge_boot_pages
);
46 /* for command line parsing */
47 static struct hstate
* __initdata parsed_hstate
;
48 static unsigned long __initdata default_hstate_max_huge_pages
;
49 static unsigned long __initdata default_hstate_size
;
52 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
55 DEFINE_SPINLOCK(hugetlb_lock
);
57 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
59 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
61 spin_unlock(&spool
->lock
);
63 /* If no pages are used, and no other handles to the subpool
64 * remain, free the subpool the subpool remain */
69 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
71 struct hugepage_subpool
*spool
;
73 spool
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
77 spin_lock_init(&spool
->lock
);
79 spool
->max_hpages
= nr_blocks
;
80 spool
->used_hpages
= 0;
85 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
87 spin_lock(&spool
->lock
);
88 BUG_ON(!spool
->count
);
90 unlock_or_release_subpool(spool
);
93 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
101 spin_lock(&spool
->lock
);
102 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
103 spool
->used_hpages
+= delta
;
107 spin_unlock(&spool
->lock
);
112 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
118 spin_lock(&spool
->lock
);
119 spool
->used_hpages
-= delta
;
120 /* If hugetlbfs_put_super couldn't free spool due to
121 * an outstanding quota reference, free it now. */
122 unlock_or_release_subpool(spool
);
125 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
127 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
130 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
132 return subpool_inode(file_inode(vma
->vm_file
));
136 * Region tracking -- allows tracking of reservations and instantiated pages
137 * across the pages in a mapping.
139 * The region data structures are protected by a combination of the mmap_sem
140 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
141 * must either hold the mmap_sem for write, or the mmap_sem for read and
142 * the hugetlb_instantiation_mutex:
144 * down_write(&mm->mmap_sem);
146 * down_read(&mm->mmap_sem);
147 * mutex_lock(&hugetlb_instantiation_mutex);
150 struct list_head link
;
155 static long region_add(struct list_head
*head
, long f
, long t
)
157 struct file_region
*rg
, *nrg
, *trg
;
159 /* Locate the region we are either in or before. */
160 list_for_each_entry(rg
, head
, link
)
164 /* Round our left edge to the current segment if it encloses us. */
168 /* Check for and consume any regions we now overlap with. */
170 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
171 if (&rg
->link
== head
)
176 /* If this area reaches higher then extend our area to
177 * include it completely. If this is not the first area
178 * which we intend to reuse, free it. */
191 static long region_chg(struct list_head
*head
, long f
, long t
)
193 struct file_region
*rg
, *nrg
;
196 /* Locate the region we are before or in. */
197 list_for_each_entry(rg
, head
, link
)
201 /* If we are below the current region then a new region is required.
202 * Subtle, allocate a new region at the position but make it zero
203 * size such that we can guarantee to record the reservation. */
204 if (&rg
->link
== head
|| t
< rg
->from
) {
205 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
210 INIT_LIST_HEAD(&nrg
->link
);
211 list_add(&nrg
->link
, rg
->link
.prev
);
216 /* Round our left edge to the current segment if it encloses us. */
221 /* Check for and consume any regions we now overlap with. */
222 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
223 if (&rg
->link
== head
)
228 /* We overlap with this area, if it extends further than
229 * us then we must extend ourselves. Account for its
230 * existing reservation. */
235 chg
-= rg
->to
- rg
->from
;
240 static long region_truncate(struct list_head
*head
, long end
)
242 struct file_region
*rg
, *trg
;
245 /* Locate the region we are either in or before. */
246 list_for_each_entry(rg
, head
, link
)
249 if (&rg
->link
== head
)
252 /* If we are in the middle of a region then adjust it. */
253 if (end
> rg
->from
) {
256 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
259 /* Drop any remaining regions. */
260 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
261 if (&rg
->link
== head
)
263 chg
+= rg
->to
- rg
->from
;
270 static long region_count(struct list_head
*head
, long f
, long t
)
272 struct file_region
*rg
;
275 /* Locate each segment we overlap with, and count that overlap. */
276 list_for_each_entry(rg
, head
, link
) {
285 seg_from
= max(rg
->from
, f
);
286 seg_to
= min(rg
->to
, t
);
288 chg
+= seg_to
- seg_from
;
295 * Convert the address within this vma to the page offset within
296 * the mapping, in pagecache page units; huge pages here.
298 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
299 struct vm_area_struct
*vma
, unsigned long address
)
301 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
302 (vma
->vm_pgoff
>> huge_page_order(h
));
305 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
306 unsigned long address
)
308 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
312 * Return the size of the pages allocated when backing a VMA. In the majority
313 * cases this will be same size as used by the page table entries.
315 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
317 struct hstate
*hstate
;
319 if (!is_vm_hugetlb_page(vma
))
322 hstate
= hstate_vma(vma
);
324 return 1UL << huge_page_shift(hstate
);
326 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
329 * Return the page size being used by the MMU to back a VMA. In the majority
330 * of cases, the page size used by the kernel matches the MMU size. On
331 * architectures where it differs, an architecture-specific version of this
332 * function is required.
334 #ifndef vma_mmu_pagesize
335 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
337 return vma_kernel_pagesize(vma
);
342 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
343 * bits of the reservation map pointer, which are always clear due to
346 #define HPAGE_RESV_OWNER (1UL << 0)
347 #define HPAGE_RESV_UNMAPPED (1UL << 1)
348 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
351 * These helpers are used to track how many pages are reserved for
352 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
353 * is guaranteed to have their future faults succeed.
355 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
356 * the reserve counters are updated with the hugetlb_lock held. It is safe
357 * to reset the VMA at fork() time as it is not in use yet and there is no
358 * chance of the global counters getting corrupted as a result of the values.
360 * The private mapping reservation is represented in a subtly different
361 * manner to a shared mapping. A shared mapping has a region map associated
362 * with the underlying file, this region map represents the backing file
363 * pages which have ever had a reservation assigned which this persists even
364 * after the page is instantiated. A private mapping has a region map
365 * associated with the original mmap which is attached to all VMAs which
366 * reference it, this region map represents those offsets which have consumed
367 * reservation ie. where pages have been instantiated.
369 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
371 return (unsigned long)vma
->vm_private_data
;
374 static void set_vma_private_data(struct vm_area_struct
*vma
,
377 vma
->vm_private_data
= (void *)value
;
382 struct list_head regions
;
385 static struct resv_map
*resv_map_alloc(void)
387 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
391 kref_init(&resv_map
->refs
);
392 INIT_LIST_HEAD(&resv_map
->regions
);
397 static void resv_map_release(struct kref
*ref
)
399 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
401 /* Clear out any active regions before we release the map. */
402 region_truncate(&resv_map
->regions
, 0);
406 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
408 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
409 if (!(vma
->vm_flags
& VM_MAYSHARE
))
410 return (struct resv_map
*)(get_vma_private_data(vma
) &
415 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
417 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
418 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
420 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
421 HPAGE_RESV_MASK
) | (unsigned long)map
);
424 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
426 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
427 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
429 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
432 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
434 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
436 return (get_vma_private_data(vma
) & flag
) != 0;
439 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
440 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
442 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
443 if (!(vma
->vm_flags
& VM_MAYSHARE
))
444 vma
->vm_private_data
= (void *)0;
447 /* Returns true if the VMA has associated reserve pages */
448 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
450 if (vma
->vm_flags
& VM_NORESERVE
) {
452 * This address is already reserved by other process(chg == 0),
453 * so, we should decrement reserved count. Without decrementing,
454 * reserve count remains after releasing inode, because this
455 * allocated page will go into page cache and is regarded as
456 * coming from reserved pool in releasing step. Currently, we
457 * don't have any other solution to deal with this situation
458 * properly, so add work-around here.
460 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
466 /* Shared mappings always use reserves */
467 if (vma
->vm_flags
& VM_MAYSHARE
)
471 * Only the process that called mmap() has reserves for
474 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
480 static void copy_gigantic_page(struct page
*dst
, struct page
*src
)
483 struct hstate
*h
= page_hstate(src
);
484 struct page
*dst_base
= dst
;
485 struct page
*src_base
= src
;
487 for (i
= 0; i
< pages_per_huge_page(h
); ) {
489 copy_highpage(dst
, src
);
492 dst
= mem_map_next(dst
, dst_base
, i
);
493 src
= mem_map_next(src
, src_base
, i
);
497 void copy_huge_page(struct page
*dst
, struct page
*src
)
500 struct hstate
*h
= page_hstate(src
);
502 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
503 copy_gigantic_page(dst
, src
);
508 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
510 copy_highpage(dst
+ i
, src
+ i
);
514 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
516 int nid
= page_to_nid(page
);
517 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
518 h
->free_huge_pages
++;
519 h
->free_huge_pages_node
[nid
]++;
522 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
526 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
527 if (!is_migrate_isolate_page(page
))
530 * if 'non-isolated free hugepage' not found on the list,
531 * the allocation fails.
533 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
535 list_move(&page
->lru
, &h
->hugepage_activelist
);
536 set_page_refcounted(page
);
537 h
->free_huge_pages
--;
538 h
->free_huge_pages_node
[nid
]--;
542 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
543 struct vm_area_struct
*vma
,
544 unsigned long address
, int avoid_reserve
,
547 struct page
*page
= NULL
;
548 struct mempolicy
*mpol
;
549 nodemask_t
*nodemask
;
550 struct zonelist
*zonelist
;
553 unsigned int cpuset_mems_cookie
;
556 * A child process with MAP_PRIVATE mappings created by their parent
557 * have no page reserves. This check ensures that reservations are
558 * not "stolen". The child may still get SIGKILLed
560 if (!vma_has_reserves(vma
, chg
) &&
561 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
564 /* If reserves cannot be used, ensure enough pages are in the pool */
565 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
569 cpuset_mems_cookie
= get_mems_allowed();
570 zonelist
= huge_zonelist(vma
, address
,
571 htlb_alloc_mask
, &mpol
, &nodemask
);
573 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
574 MAX_NR_ZONES
- 1, nodemask
) {
575 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
576 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
580 if (!vma_has_reserves(vma
, chg
))
583 SetPagePrivate(page
);
584 h
->resv_huge_pages
--;
591 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
599 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
603 VM_BUG_ON(h
->order
>= MAX_ORDER
);
606 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
607 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
608 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
609 1 << PG_referenced
| 1 << PG_dirty
|
610 1 << PG_active
| 1 << PG_reserved
|
611 1 << PG_private
| 1 << PG_writeback
);
613 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
614 set_compound_page_dtor(page
, NULL
);
615 set_page_refcounted(page
);
616 arch_release_hugepage(page
);
617 __free_pages(page
, huge_page_order(h
));
620 struct hstate
*size_to_hstate(unsigned long size
)
625 if (huge_page_size(h
) == size
)
631 static void free_huge_page(struct page
*page
)
634 * Can't pass hstate in here because it is called from the
635 * compound page destructor.
637 struct hstate
*h
= page_hstate(page
);
638 int nid
= page_to_nid(page
);
639 struct hugepage_subpool
*spool
=
640 (struct hugepage_subpool
*)page_private(page
);
641 bool restore_reserve
;
643 set_page_private(page
, 0);
644 page
->mapping
= NULL
;
645 BUG_ON(page_count(page
));
646 BUG_ON(page_mapcount(page
));
647 restore_reserve
= PagePrivate(page
);
649 spin_lock(&hugetlb_lock
);
650 hugetlb_cgroup_uncharge_page(hstate_index(h
),
651 pages_per_huge_page(h
), page
);
653 h
->resv_huge_pages
++;
655 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
656 /* remove the page from active list */
657 list_del(&page
->lru
);
658 update_and_free_page(h
, page
);
659 h
->surplus_huge_pages
--;
660 h
->surplus_huge_pages_node
[nid
]--;
662 arch_clear_hugepage_flags(page
);
663 enqueue_huge_page(h
, page
);
665 spin_unlock(&hugetlb_lock
);
666 hugepage_subpool_put_pages(spool
, 1);
669 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
671 INIT_LIST_HEAD(&page
->lru
);
672 set_compound_page_dtor(page
, free_huge_page
);
673 spin_lock(&hugetlb_lock
);
674 set_hugetlb_cgroup(page
, NULL
);
676 h
->nr_huge_pages_node
[nid
]++;
677 spin_unlock(&hugetlb_lock
);
678 put_page(page
); /* free it into the hugepage allocator */
681 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
684 int nr_pages
= 1 << order
;
685 struct page
*p
= page
+ 1;
687 /* we rely on prep_new_huge_page to set the destructor */
688 set_compound_order(page
, order
);
690 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
692 set_page_count(p
, 0);
693 p
->first_page
= page
;
698 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
699 * transparent huge pages. See the PageTransHuge() documentation for more
702 int PageHuge(struct page
*page
)
704 compound_page_dtor
*dtor
;
706 if (!PageCompound(page
))
709 page
= compound_head(page
);
710 dtor
= get_compound_page_dtor(page
);
712 return dtor
== free_huge_page
;
714 EXPORT_SYMBOL_GPL(PageHuge
);
716 pgoff_t
__basepage_index(struct page
*page
)
718 struct page
*page_head
= compound_head(page
);
719 pgoff_t index
= page_index(page_head
);
720 unsigned long compound_idx
;
722 if (!PageHuge(page_head
))
723 return page_index(page
);
725 if (compound_order(page_head
) >= MAX_ORDER
)
726 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
728 compound_idx
= page
- page_head
;
730 return (index
<< compound_order(page_head
)) + compound_idx
;
733 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
737 if (h
->order
>= MAX_ORDER
)
740 page
= alloc_pages_exact_node(nid
,
741 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
742 __GFP_REPEAT
|__GFP_NOWARN
,
745 if (arch_prepare_hugepage(page
)) {
746 __free_pages(page
, huge_page_order(h
));
749 prep_new_huge_page(h
, page
, nid
);
756 * common helper functions for hstate_next_node_to_{alloc|free}.
757 * We may have allocated or freed a huge page based on a different
758 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
759 * be outside of *nodes_allowed. Ensure that we use an allowed
760 * node for alloc or free.
762 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
764 nid
= next_node(nid
, *nodes_allowed
);
765 if (nid
== MAX_NUMNODES
)
766 nid
= first_node(*nodes_allowed
);
767 VM_BUG_ON(nid
>= MAX_NUMNODES
);
772 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
774 if (!node_isset(nid
, *nodes_allowed
))
775 nid
= next_node_allowed(nid
, nodes_allowed
);
780 * returns the previously saved node ["this node"] from which to
781 * allocate a persistent huge page for the pool and advance the
782 * next node from which to allocate, handling wrap at end of node
785 static int hstate_next_node_to_alloc(struct hstate
*h
,
786 nodemask_t
*nodes_allowed
)
790 VM_BUG_ON(!nodes_allowed
);
792 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
793 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
799 * helper for free_pool_huge_page() - return the previously saved
800 * node ["this node"] from which to free a huge page. Advance the
801 * next node id whether or not we find a free huge page to free so
802 * that the next attempt to free addresses the next node.
804 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
808 VM_BUG_ON(!nodes_allowed
);
810 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
811 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
816 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
817 for (nr_nodes = nodes_weight(*mask); \
819 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
822 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
823 for (nr_nodes = nodes_weight(*mask); \
825 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
828 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
834 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
835 page
= alloc_fresh_huge_page_node(h
, node
);
843 count_vm_event(HTLB_BUDDY_PGALLOC
);
845 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
851 * Free huge page from pool from next node to free.
852 * Attempt to keep persistent huge pages more or less
853 * balanced over allowed nodes.
854 * Called with hugetlb_lock locked.
856 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
862 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
864 * If we're returning unused surplus pages, only examine
865 * nodes with surplus pages.
867 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
868 !list_empty(&h
->hugepage_freelists
[node
])) {
870 list_entry(h
->hugepage_freelists
[node
].next
,
872 list_del(&page
->lru
);
873 h
->free_huge_pages
--;
874 h
->free_huge_pages_node
[node
]--;
876 h
->surplus_huge_pages
--;
877 h
->surplus_huge_pages_node
[node
]--;
879 update_and_free_page(h
, page
);
889 * Dissolve a given free hugepage into free buddy pages. This function does
890 * nothing for in-use (including surplus) hugepages.
892 static void dissolve_free_huge_page(struct page
*page
)
894 spin_lock(&hugetlb_lock
);
895 if (PageHuge(page
) && !page_count(page
)) {
896 struct hstate
*h
= page_hstate(page
);
897 int nid
= page_to_nid(page
);
898 list_del(&page
->lru
);
899 h
->free_huge_pages
--;
900 h
->free_huge_pages_node
[nid
]--;
901 update_and_free_page(h
, page
);
903 spin_unlock(&hugetlb_lock
);
907 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
908 * make specified memory blocks removable from the system.
909 * Note that start_pfn should aligned with (minimum) hugepage size.
911 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
913 unsigned int order
= 8 * sizeof(void *);
917 /* Set scan step to minimum hugepage size */
919 if (order
> huge_page_order(h
))
920 order
= huge_page_order(h
);
921 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
922 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
923 dissolve_free_huge_page(pfn_to_page(pfn
));
926 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
931 if (h
->order
>= MAX_ORDER
)
935 * Assume we will successfully allocate the surplus page to
936 * prevent racing processes from causing the surplus to exceed
939 * This however introduces a different race, where a process B
940 * tries to grow the static hugepage pool while alloc_pages() is
941 * called by process A. B will only examine the per-node
942 * counters in determining if surplus huge pages can be
943 * converted to normal huge pages in adjust_pool_surplus(). A
944 * won't be able to increment the per-node counter, until the
945 * lock is dropped by B, but B doesn't drop hugetlb_lock until
946 * no more huge pages can be converted from surplus to normal
947 * state (and doesn't try to convert again). Thus, we have a
948 * case where a surplus huge page exists, the pool is grown, and
949 * the surplus huge page still exists after, even though it
950 * should just have been converted to a normal huge page. This
951 * does not leak memory, though, as the hugepage will be freed
952 * once it is out of use. It also does not allow the counters to
953 * go out of whack in adjust_pool_surplus() as we don't modify
954 * the node values until we've gotten the hugepage and only the
955 * per-node value is checked there.
957 spin_lock(&hugetlb_lock
);
958 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
959 spin_unlock(&hugetlb_lock
);
963 h
->surplus_huge_pages
++;
965 spin_unlock(&hugetlb_lock
);
967 if (nid
== NUMA_NO_NODE
)
968 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
969 __GFP_REPEAT
|__GFP_NOWARN
,
972 page
= alloc_pages_exact_node(nid
,
973 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
974 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
976 if (page
&& arch_prepare_hugepage(page
)) {
977 __free_pages(page
, huge_page_order(h
));
981 spin_lock(&hugetlb_lock
);
983 INIT_LIST_HEAD(&page
->lru
);
984 r_nid
= page_to_nid(page
);
985 set_compound_page_dtor(page
, free_huge_page
);
986 set_hugetlb_cgroup(page
, NULL
);
988 * We incremented the global counters already
990 h
->nr_huge_pages_node
[r_nid
]++;
991 h
->surplus_huge_pages_node
[r_nid
]++;
992 __count_vm_event(HTLB_BUDDY_PGALLOC
);
995 h
->surplus_huge_pages
--;
996 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
998 spin_unlock(&hugetlb_lock
);
1004 * This allocation function is useful in the context where vma is irrelevant.
1005 * E.g. soft-offlining uses this function because it only cares physical
1006 * address of error page.
1008 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1010 struct page
*page
= NULL
;
1012 spin_lock(&hugetlb_lock
);
1013 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1014 page
= dequeue_huge_page_node(h
, nid
);
1015 spin_unlock(&hugetlb_lock
);
1018 page
= alloc_buddy_huge_page(h
, nid
);
1024 * Increase the hugetlb pool such that it can accommodate a reservation
1027 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1029 struct list_head surplus_list
;
1030 struct page
*page
, *tmp
;
1032 int needed
, allocated
;
1033 bool alloc_ok
= true;
1035 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1037 h
->resv_huge_pages
+= delta
;
1042 INIT_LIST_HEAD(&surplus_list
);
1046 spin_unlock(&hugetlb_lock
);
1047 for (i
= 0; i
< needed
; i
++) {
1048 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1053 list_add(&page
->lru
, &surplus_list
);
1058 * After retaking hugetlb_lock, we need to recalculate 'needed'
1059 * because either resv_huge_pages or free_huge_pages may have changed.
1061 spin_lock(&hugetlb_lock
);
1062 needed
= (h
->resv_huge_pages
+ delta
) -
1063 (h
->free_huge_pages
+ allocated
);
1068 * We were not able to allocate enough pages to
1069 * satisfy the entire reservation so we free what
1070 * we've allocated so far.
1075 * The surplus_list now contains _at_least_ the number of extra pages
1076 * needed to accommodate the reservation. Add the appropriate number
1077 * of pages to the hugetlb pool and free the extras back to the buddy
1078 * allocator. Commit the entire reservation here to prevent another
1079 * process from stealing the pages as they are added to the pool but
1080 * before they are reserved.
1082 needed
+= allocated
;
1083 h
->resv_huge_pages
+= delta
;
1086 /* Free the needed pages to the hugetlb pool */
1087 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1091 * This page is now managed by the hugetlb allocator and has
1092 * no users -- drop the buddy allocator's reference.
1094 put_page_testzero(page
);
1095 VM_BUG_ON(page_count(page
));
1096 enqueue_huge_page(h
, page
);
1099 spin_unlock(&hugetlb_lock
);
1101 /* Free unnecessary surplus pages to the buddy allocator */
1102 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1104 spin_lock(&hugetlb_lock
);
1110 * When releasing a hugetlb pool reservation, any surplus pages that were
1111 * allocated to satisfy the reservation must be explicitly freed if they were
1113 * Called with hugetlb_lock held.
1115 static void return_unused_surplus_pages(struct hstate
*h
,
1116 unsigned long unused_resv_pages
)
1118 unsigned long nr_pages
;
1120 /* Uncommit the reservation */
1121 h
->resv_huge_pages
-= unused_resv_pages
;
1123 /* Cannot return gigantic pages currently */
1124 if (h
->order
>= MAX_ORDER
)
1127 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1130 * We want to release as many surplus pages as possible, spread
1131 * evenly across all nodes with memory. Iterate across these nodes
1132 * until we can no longer free unreserved surplus pages. This occurs
1133 * when the nodes with surplus pages have no free pages.
1134 * free_pool_huge_page() will balance the the freed pages across the
1135 * on-line nodes with memory and will handle the hstate accounting.
1137 while (nr_pages
--) {
1138 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1144 * Determine if the huge page at addr within the vma has an associated
1145 * reservation. Where it does not we will need to logically increase
1146 * reservation and actually increase subpool usage before an allocation
1147 * can occur. Where any new reservation would be required the
1148 * reservation change is prepared, but not committed. Once the page
1149 * has been allocated from the subpool and instantiated the change should
1150 * be committed via vma_commit_reservation. No action is required on
1153 static long vma_needs_reservation(struct hstate
*h
,
1154 struct vm_area_struct
*vma
, unsigned long addr
)
1156 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1157 struct inode
*inode
= mapping
->host
;
1159 if (vma
->vm_flags
& VM_MAYSHARE
) {
1160 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1161 return region_chg(&inode
->i_mapping
->private_list
,
1164 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1169 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1170 struct resv_map
*resv
= vma_resv_map(vma
);
1172 err
= region_chg(&resv
->regions
, idx
, idx
+ 1);
1178 static void vma_commit_reservation(struct hstate
*h
,
1179 struct vm_area_struct
*vma
, unsigned long addr
)
1181 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1182 struct inode
*inode
= mapping
->host
;
1184 if (vma
->vm_flags
& VM_MAYSHARE
) {
1185 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1186 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1188 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1189 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1190 struct resv_map
*resv
= vma_resv_map(vma
);
1192 /* Mark this page used in the map. */
1193 region_add(&resv
->regions
, idx
, idx
+ 1);
1197 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1198 unsigned long addr
, int avoid_reserve
)
1200 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1201 struct hstate
*h
= hstate_vma(vma
);
1205 struct hugetlb_cgroup
*h_cg
;
1207 idx
= hstate_index(h
);
1209 * Processes that did not create the mapping will have no
1210 * reserves and will not have accounted against subpool
1211 * limit. Check that the subpool limit can be made before
1212 * satisfying the allocation MAP_NORESERVE mappings may also
1213 * need pages and subpool limit allocated allocated if no reserve
1216 chg
= vma_needs_reservation(h
, vma
, addr
);
1218 return ERR_PTR(-ENOMEM
);
1219 if (chg
|| avoid_reserve
)
1220 if (hugepage_subpool_get_pages(spool
, 1))
1221 return ERR_PTR(-ENOSPC
);
1223 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1225 if (chg
|| avoid_reserve
)
1226 hugepage_subpool_put_pages(spool
, 1);
1227 return ERR_PTR(-ENOSPC
);
1229 spin_lock(&hugetlb_lock
);
1230 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1232 spin_unlock(&hugetlb_lock
);
1233 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1235 hugetlb_cgroup_uncharge_cgroup(idx
,
1236 pages_per_huge_page(h
),
1238 if (chg
|| avoid_reserve
)
1239 hugepage_subpool_put_pages(spool
, 1);
1240 return ERR_PTR(-ENOSPC
);
1242 spin_lock(&hugetlb_lock
);
1243 list_move(&page
->lru
, &h
->hugepage_activelist
);
1246 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1247 spin_unlock(&hugetlb_lock
);
1249 set_page_private(page
, (unsigned long)spool
);
1251 vma_commit_reservation(h
, vma
, addr
);
1256 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1257 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1258 * where no ERR_VALUE is expected to be returned.
1260 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1261 unsigned long addr
, int avoid_reserve
)
1263 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1269 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1271 struct huge_bootmem_page
*m
;
1274 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1277 addr
= __alloc_bootmem_node_nopanic(NODE_DATA(node
),
1278 huge_page_size(h
), huge_page_size(h
), 0);
1282 * Use the beginning of the huge page to store the
1283 * huge_bootmem_page struct (until gather_bootmem
1284 * puts them into the mem_map).
1293 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1294 /* Put them into a private list first because mem_map is not up yet */
1295 list_add(&m
->list
, &huge_boot_pages
);
1300 static void prep_compound_huge_page(struct page
*page
, int order
)
1302 if (unlikely(order
> (MAX_ORDER
- 1)))
1303 prep_compound_gigantic_page(page
, order
);
1305 prep_compound_page(page
, order
);
1308 /* Put bootmem huge pages into the standard lists after mem_map is up */
1309 static void __init
gather_bootmem_prealloc(void)
1311 struct huge_bootmem_page
*m
;
1313 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1314 struct hstate
*h
= m
->hstate
;
1317 #ifdef CONFIG_HIGHMEM
1318 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1319 free_bootmem_late((unsigned long)m
,
1320 sizeof(struct huge_bootmem_page
));
1322 page
= virt_to_page(m
);
1324 __ClearPageReserved(page
);
1325 WARN_ON(page_count(page
) != 1);
1326 prep_compound_huge_page(page
, h
->order
);
1327 prep_new_huge_page(h
, page
, page_to_nid(page
));
1329 * If we had gigantic hugepages allocated at boot time, we need
1330 * to restore the 'stolen' pages to totalram_pages in order to
1331 * fix confusing memory reports from free(1) and another
1332 * side-effects, like CommitLimit going negative.
1334 if (h
->order
> (MAX_ORDER
- 1))
1335 adjust_managed_page_count(page
, 1 << h
->order
);
1339 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1343 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1344 if (h
->order
>= MAX_ORDER
) {
1345 if (!alloc_bootmem_huge_page(h
))
1347 } else if (!alloc_fresh_huge_page(h
,
1348 &node_states
[N_MEMORY
]))
1351 h
->max_huge_pages
= i
;
1354 static void __init
hugetlb_init_hstates(void)
1358 for_each_hstate(h
) {
1359 /* oversize hugepages were init'ed in early boot */
1360 if (h
->order
< MAX_ORDER
)
1361 hugetlb_hstate_alloc_pages(h
);
1365 static char * __init
memfmt(char *buf
, unsigned long n
)
1367 if (n
>= (1UL << 30))
1368 sprintf(buf
, "%lu GB", n
>> 30);
1369 else if (n
>= (1UL << 20))
1370 sprintf(buf
, "%lu MB", n
>> 20);
1372 sprintf(buf
, "%lu KB", n
>> 10);
1376 static void __init
report_hugepages(void)
1380 for_each_hstate(h
) {
1382 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1383 memfmt(buf
, huge_page_size(h
)),
1384 h
->free_huge_pages
);
1388 #ifdef CONFIG_HIGHMEM
1389 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1390 nodemask_t
*nodes_allowed
)
1394 if (h
->order
>= MAX_ORDER
)
1397 for_each_node_mask(i
, *nodes_allowed
) {
1398 struct page
*page
, *next
;
1399 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1400 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1401 if (count
>= h
->nr_huge_pages
)
1403 if (PageHighMem(page
))
1405 list_del(&page
->lru
);
1406 update_and_free_page(h
, page
);
1407 h
->free_huge_pages
--;
1408 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1413 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1414 nodemask_t
*nodes_allowed
)
1420 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1421 * balanced by operating on them in a round-robin fashion.
1422 * Returns 1 if an adjustment was made.
1424 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1429 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1432 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1433 if (h
->surplus_huge_pages_node
[node
])
1437 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1438 if (h
->surplus_huge_pages_node
[node
] <
1439 h
->nr_huge_pages_node
[node
])
1446 h
->surplus_huge_pages
+= delta
;
1447 h
->surplus_huge_pages_node
[node
] += delta
;
1451 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1452 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1453 nodemask_t
*nodes_allowed
)
1455 unsigned long min_count
, ret
;
1457 if (h
->order
>= MAX_ORDER
)
1458 return h
->max_huge_pages
;
1461 * Increase the pool size
1462 * First take pages out of surplus state. Then make up the
1463 * remaining difference by allocating fresh huge pages.
1465 * We might race with alloc_buddy_huge_page() here and be unable
1466 * to convert a surplus huge page to a normal huge page. That is
1467 * not critical, though, it just means the overall size of the
1468 * pool might be one hugepage larger than it needs to be, but
1469 * within all the constraints specified by the sysctls.
1471 spin_lock(&hugetlb_lock
);
1472 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1473 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1477 while (count
> persistent_huge_pages(h
)) {
1479 * If this allocation races such that we no longer need the
1480 * page, free_huge_page will handle it by freeing the page
1481 * and reducing the surplus.
1483 spin_unlock(&hugetlb_lock
);
1484 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1485 spin_lock(&hugetlb_lock
);
1489 /* Bail for signals. Probably ctrl-c from user */
1490 if (signal_pending(current
))
1495 * Decrease the pool size
1496 * First return free pages to the buddy allocator (being careful
1497 * to keep enough around to satisfy reservations). Then place
1498 * pages into surplus state as needed so the pool will shrink
1499 * to the desired size as pages become free.
1501 * By placing pages into the surplus state independent of the
1502 * overcommit value, we are allowing the surplus pool size to
1503 * exceed overcommit. There are few sane options here. Since
1504 * alloc_buddy_huge_page() is checking the global counter,
1505 * though, we'll note that we're not allowed to exceed surplus
1506 * and won't grow the pool anywhere else. Not until one of the
1507 * sysctls are changed, or the surplus pages go out of use.
1509 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1510 min_count
= max(count
, min_count
);
1511 try_to_free_low(h
, min_count
, nodes_allowed
);
1512 while (min_count
< persistent_huge_pages(h
)) {
1513 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1516 while (count
< persistent_huge_pages(h
)) {
1517 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1521 ret
= persistent_huge_pages(h
);
1522 spin_unlock(&hugetlb_lock
);
1526 #define HSTATE_ATTR_RO(_name) \
1527 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1529 #define HSTATE_ATTR(_name) \
1530 static struct kobj_attribute _name##_attr = \
1531 __ATTR(_name, 0644, _name##_show, _name##_store)
1533 static struct kobject
*hugepages_kobj
;
1534 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1536 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1538 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1542 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1543 if (hstate_kobjs
[i
] == kobj
) {
1545 *nidp
= NUMA_NO_NODE
;
1549 return kobj_to_node_hstate(kobj
, nidp
);
1552 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1553 struct kobj_attribute
*attr
, char *buf
)
1556 unsigned long nr_huge_pages
;
1559 h
= kobj_to_hstate(kobj
, &nid
);
1560 if (nid
== NUMA_NO_NODE
)
1561 nr_huge_pages
= h
->nr_huge_pages
;
1563 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1565 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1568 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1569 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1570 const char *buf
, size_t len
)
1574 unsigned long count
;
1576 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1578 err
= kstrtoul(buf
, 10, &count
);
1582 h
= kobj_to_hstate(kobj
, &nid
);
1583 if (h
->order
>= MAX_ORDER
) {
1588 if (nid
== NUMA_NO_NODE
) {
1590 * global hstate attribute
1592 if (!(obey_mempolicy
&&
1593 init_nodemask_of_mempolicy(nodes_allowed
))) {
1594 NODEMASK_FREE(nodes_allowed
);
1595 nodes_allowed
= &node_states
[N_MEMORY
];
1597 } else if (nodes_allowed
) {
1599 * per node hstate attribute: adjust count to global,
1600 * but restrict alloc/free to the specified node.
1602 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1603 init_nodemask_of_node(nodes_allowed
, nid
);
1605 nodes_allowed
= &node_states
[N_MEMORY
];
1607 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1609 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1610 NODEMASK_FREE(nodes_allowed
);
1614 NODEMASK_FREE(nodes_allowed
);
1618 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1619 struct kobj_attribute
*attr
, char *buf
)
1621 return nr_hugepages_show_common(kobj
, attr
, buf
);
1624 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1625 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1627 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1629 HSTATE_ATTR(nr_hugepages
);
1634 * hstate attribute for optionally mempolicy-based constraint on persistent
1635 * huge page alloc/free.
1637 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1638 struct kobj_attribute
*attr
, char *buf
)
1640 return nr_hugepages_show_common(kobj
, attr
, buf
);
1643 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1644 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1646 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1648 HSTATE_ATTR(nr_hugepages_mempolicy
);
1652 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1653 struct kobj_attribute
*attr
, char *buf
)
1655 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1656 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1659 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1660 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1663 unsigned long input
;
1664 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1666 if (h
->order
>= MAX_ORDER
)
1669 err
= kstrtoul(buf
, 10, &input
);
1673 spin_lock(&hugetlb_lock
);
1674 h
->nr_overcommit_huge_pages
= input
;
1675 spin_unlock(&hugetlb_lock
);
1679 HSTATE_ATTR(nr_overcommit_hugepages
);
1681 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1682 struct kobj_attribute
*attr
, char *buf
)
1685 unsigned long free_huge_pages
;
1688 h
= kobj_to_hstate(kobj
, &nid
);
1689 if (nid
== NUMA_NO_NODE
)
1690 free_huge_pages
= h
->free_huge_pages
;
1692 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1694 return sprintf(buf
, "%lu\n", free_huge_pages
);
1696 HSTATE_ATTR_RO(free_hugepages
);
1698 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1699 struct kobj_attribute
*attr
, char *buf
)
1701 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1702 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1704 HSTATE_ATTR_RO(resv_hugepages
);
1706 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1707 struct kobj_attribute
*attr
, char *buf
)
1710 unsigned long surplus_huge_pages
;
1713 h
= kobj_to_hstate(kobj
, &nid
);
1714 if (nid
== NUMA_NO_NODE
)
1715 surplus_huge_pages
= h
->surplus_huge_pages
;
1717 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1719 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1721 HSTATE_ATTR_RO(surplus_hugepages
);
1723 static struct attribute
*hstate_attrs
[] = {
1724 &nr_hugepages_attr
.attr
,
1725 &nr_overcommit_hugepages_attr
.attr
,
1726 &free_hugepages_attr
.attr
,
1727 &resv_hugepages_attr
.attr
,
1728 &surplus_hugepages_attr
.attr
,
1730 &nr_hugepages_mempolicy_attr
.attr
,
1735 static struct attribute_group hstate_attr_group
= {
1736 .attrs
= hstate_attrs
,
1739 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1740 struct kobject
**hstate_kobjs
,
1741 struct attribute_group
*hstate_attr_group
)
1744 int hi
= hstate_index(h
);
1746 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1747 if (!hstate_kobjs
[hi
])
1750 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1752 kobject_put(hstate_kobjs
[hi
]);
1757 static void __init
hugetlb_sysfs_init(void)
1762 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1763 if (!hugepages_kobj
)
1766 for_each_hstate(h
) {
1767 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1768 hstate_kobjs
, &hstate_attr_group
);
1770 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1777 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1778 * with node devices in node_devices[] using a parallel array. The array
1779 * index of a node device or _hstate == node id.
1780 * This is here to avoid any static dependency of the node device driver, in
1781 * the base kernel, on the hugetlb module.
1783 struct node_hstate
{
1784 struct kobject
*hugepages_kobj
;
1785 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1787 struct node_hstate node_hstates
[MAX_NUMNODES
];
1790 * A subset of global hstate attributes for node devices
1792 static struct attribute
*per_node_hstate_attrs
[] = {
1793 &nr_hugepages_attr
.attr
,
1794 &free_hugepages_attr
.attr
,
1795 &surplus_hugepages_attr
.attr
,
1799 static struct attribute_group per_node_hstate_attr_group
= {
1800 .attrs
= per_node_hstate_attrs
,
1804 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1805 * Returns node id via non-NULL nidp.
1807 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1811 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1812 struct node_hstate
*nhs
= &node_hstates
[nid
];
1814 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1815 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1827 * Unregister hstate attributes from a single node device.
1828 * No-op if no hstate attributes attached.
1830 static void hugetlb_unregister_node(struct node
*node
)
1833 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1835 if (!nhs
->hugepages_kobj
)
1836 return; /* no hstate attributes */
1838 for_each_hstate(h
) {
1839 int idx
= hstate_index(h
);
1840 if (nhs
->hstate_kobjs
[idx
]) {
1841 kobject_put(nhs
->hstate_kobjs
[idx
]);
1842 nhs
->hstate_kobjs
[idx
] = NULL
;
1846 kobject_put(nhs
->hugepages_kobj
);
1847 nhs
->hugepages_kobj
= NULL
;
1851 * hugetlb module exit: unregister hstate attributes from node devices
1854 static void hugetlb_unregister_all_nodes(void)
1859 * disable node device registrations.
1861 register_hugetlbfs_with_node(NULL
, NULL
);
1864 * remove hstate attributes from any nodes that have them.
1866 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1867 hugetlb_unregister_node(node_devices
[nid
]);
1871 * Register hstate attributes for a single node device.
1872 * No-op if attributes already registered.
1874 static void hugetlb_register_node(struct node
*node
)
1877 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1880 if (nhs
->hugepages_kobj
)
1881 return; /* already allocated */
1883 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1885 if (!nhs
->hugepages_kobj
)
1888 for_each_hstate(h
) {
1889 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1891 &per_node_hstate_attr_group
);
1893 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1894 h
->name
, node
->dev
.id
);
1895 hugetlb_unregister_node(node
);
1902 * hugetlb init time: register hstate attributes for all registered node
1903 * devices of nodes that have memory. All on-line nodes should have
1904 * registered their associated device by this time.
1906 static void hugetlb_register_all_nodes(void)
1910 for_each_node_state(nid
, N_MEMORY
) {
1911 struct node
*node
= node_devices
[nid
];
1912 if (node
->dev
.id
== nid
)
1913 hugetlb_register_node(node
);
1917 * Let the node device driver know we're here so it can
1918 * [un]register hstate attributes on node hotplug.
1920 register_hugetlbfs_with_node(hugetlb_register_node
,
1921 hugetlb_unregister_node
);
1923 #else /* !CONFIG_NUMA */
1925 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1933 static void hugetlb_unregister_all_nodes(void) { }
1935 static void hugetlb_register_all_nodes(void) { }
1939 static void __exit
hugetlb_exit(void)
1943 hugetlb_unregister_all_nodes();
1945 for_each_hstate(h
) {
1946 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1949 kobject_put(hugepages_kobj
);
1951 module_exit(hugetlb_exit
);
1953 static int __init
hugetlb_init(void)
1955 /* Some platform decide whether they support huge pages at boot
1956 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1957 * there is no such support
1959 if (HPAGE_SHIFT
== 0)
1962 if (!size_to_hstate(default_hstate_size
)) {
1963 default_hstate_size
= HPAGE_SIZE
;
1964 if (!size_to_hstate(default_hstate_size
))
1965 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1967 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1968 if (default_hstate_max_huge_pages
)
1969 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1971 hugetlb_init_hstates();
1972 gather_bootmem_prealloc();
1975 hugetlb_sysfs_init();
1976 hugetlb_register_all_nodes();
1977 hugetlb_cgroup_file_init();
1981 module_init(hugetlb_init
);
1983 /* Should be called on processing a hugepagesz=... option */
1984 void __init
hugetlb_add_hstate(unsigned order
)
1989 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1990 pr_warning("hugepagesz= specified twice, ignoring\n");
1993 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1995 h
= &hstates
[hugetlb_max_hstate
++];
1997 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1998 h
->nr_huge_pages
= 0;
1999 h
->free_huge_pages
= 0;
2000 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2001 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2002 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2003 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2004 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2005 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2006 huge_page_size(h
)/1024);
2011 static int __init
hugetlb_nrpages_setup(char *s
)
2014 static unsigned long *last_mhp
;
2017 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2018 * so this hugepages= parameter goes to the "default hstate".
2020 if (!hugetlb_max_hstate
)
2021 mhp
= &default_hstate_max_huge_pages
;
2023 mhp
= &parsed_hstate
->max_huge_pages
;
2025 if (mhp
== last_mhp
) {
2026 pr_warning("hugepages= specified twice without "
2027 "interleaving hugepagesz=, ignoring\n");
2031 if (sscanf(s
, "%lu", mhp
) <= 0)
2035 * Global state is always initialized later in hugetlb_init.
2036 * But we need to allocate >= MAX_ORDER hstates here early to still
2037 * use the bootmem allocator.
2039 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2040 hugetlb_hstate_alloc_pages(parsed_hstate
);
2046 __setup("hugepages=", hugetlb_nrpages_setup
);
2048 static int __init
hugetlb_default_setup(char *s
)
2050 default_hstate_size
= memparse(s
, &s
);
2053 __setup("default_hugepagesz=", hugetlb_default_setup
);
2055 static unsigned int cpuset_mems_nr(unsigned int *array
)
2058 unsigned int nr
= 0;
2060 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2066 #ifdef CONFIG_SYSCTL
2067 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2068 struct ctl_table
*table
, int write
,
2069 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2071 struct hstate
*h
= &default_hstate
;
2075 tmp
= h
->max_huge_pages
;
2077 if (write
&& h
->order
>= MAX_ORDER
)
2081 table
->maxlen
= sizeof(unsigned long);
2082 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2087 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2088 GFP_KERNEL
| __GFP_NORETRY
);
2089 if (!(obey_mempolicy
&&
2090 init_nodemask_of_mempolicy(nodes_allowed
))) {
2091 NODEMASK_FREE(nodes_allowed
);
2092 nodes_allowed
= &node_states
[N_MEMORY
];
2094 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2096 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2097 NODEMASK_FREE(nodes_allowed
);
2103 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2104 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2107 return hugetlb_sysctl_handler_common(false, table
, write
,
2108 buffer
, length
, ppos
);
2112 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2113 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2115 return hugetlb_sysctl_handler_common(true, table
, write
,
2116 buffer
, length
, ppos
);
2118 #endif /* CONFIG_NUMA */
2120 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2121 void __user
*buffer
,
2122 size_t *length
, loff_t
*ppos
)
2124 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2125 if (hugepages_treat_as_movable
)
2126 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2128 htlb_alloc_mask
= GFP_HIGHUSER
;
2132 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2133 void __user
*buffer
,
2134 size_t *length
, loff_t
*ppos
)
2136 struct hstate
*h
= &default_hstate
;
2140 tmp
= h
->nr_overcommit_huge_pages
;
2142 if (write
&& h
->order
>= MAX_ORDER
)
2146 table
->maxlen
= sizeof(unsigned long);
2147 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2152 spin_lock(&hugetlb_lock
);
2153 h
->nr_overcommit_huge_pages
= tmp
;
2154 spin_unlock(&hugetlb_lock
);
2160 #endif /* CONFIG_SYSCTL */
2162 void hugetlb_report_meminfo(struct seq_file
*m
)
2164 struct hstate
*h
= &default_hstate
;
2166 "HugePages_Total: %5lu\n"
2167 "HugePages_Free: %5lu\n"
2168 "HugePages_Rsvd: %5lu\n"
2169 "HugePages_Surp: %5lu\n"
2170 "Hugepagesize: %8lu kB\n",
2174 h
->surplus_huge_pages
,
2175 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2178 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2180 struct hstate
*h
= &default_hstate
;
2182 "Node %d HugePages_Total: %5u\n"
2183 "Node %d HugePages_Free: %5u\n"
2184 "Node %d HugePages_Surp: %5u\n",
2185 nid
, h
->nr_huge_pages_node
[nid
],
2186 nid
, h
->free_huge_pages_node
[nid
],
2187 nid
, h
->surplus_huge_pages_node
[nid
]);
2190 void hugetlb_show_meminfo(void)
2195 for_each_node_state(nid
, N_MEMORY
)
2197 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2199 h
->nr_huge_pages_node
[nid
],
2200 h
->free_huge_pages_node
[nid
],
2201 h
->surplus_huge_pages_node
[nid
],
2202 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2205 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2206 unsigned long hugetlb_total_pages(void)
2209 unsigned long nr_total_pages
= 0;
2212 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2213 return nr_total_pages
;
2216 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2220 spin_lock(&hugetlb_lock
);
2222 * When cpuset is configured, it breaks the strict hugetlb page
2223 * reservation as the accounting is done on a global variable. Such
2224 * reservation is completely rubbish in the presence of cpuset because
2225 * the reservation is not checked against page availability for the
2226 * current cpuset. Application can still potentially OOM'ed by kernel
2227 * with lack of free htlb page in cpuset that the task is in.
2228 * Attempt to enforce strict accounting with cpuset is almost
2229 * impossible (or too ugly) because cpuset is too fluid that
2230 * task or memory node can be dynamically moved between cpusets.
2232 * The change of semantics for shared hugetlb mapping with cpuset is
2233 * undesirable. However, in order to preserve some of the semantics,
2234 * we fall back to check against current free page availability as
2235 * a best attempt and hopefully to minimize the impact of changing
2236 * semantics that cpuset has.
2239 if (gather_surplus_pages(h
, delta
) < 0)
2242 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2243 return_unused_surplus_pages(h
, delta
);
2250 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2253 spin_unlock(&hugetlb_lock
);
2257 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2259 struct resv_map
*resv
= vma_resv_map(vma
);
2262 * This new VMA should share its siblings reservation map if present.
2263 * The VMA will only ever have a valid reservation map pointer where
2264 * it is being copied for another still existing VMA. As that VMA
2265 * has a reference to the reservation map it cannot disappear until
2266 * after this open call completes. It is therefore safe to take a
2267 * new reference here without additional locking.
2270 kref_get(&resv
->refs
);
2273 static void resv_map_put(struct vm_area_struct
*vma
)
2275 struct resv_map
*resv
= vma_resv_map(vma
);
2279 kref_put(&resv
->refs
, resv_map_release
);
2282 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2284 struct hstate
*h
= hstate_vma(vma
);
2285 struct resv_map
*resv
= vma_resv_map(vma
);
2286 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2287 unsigned long reserve
;
2288 unsigned long start
;
2292 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2293 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2295 reserve
= (end
- start
) -
2296 region_count(&resv
->regions
, start
, end
);
2301 hugetlb_acct_memory(h
, -reserve
);
2302 hugepage_subpool_put_pages(spool
, reserve
);
2308 * We cannot handle pagefaults against hugetlb pages at all. They cause
2309 * handle_mm_fault() to try to instantiate regular-sized pages in the
2310 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2313 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2319 const struct vm_operations_struct hugetlb_vm_ops
= {
2320 .fault
= hugetlb_vm_op_fault
,
2321 .open
= hugetlb_vm_op_open
,
2322 .close
= hugetlb_vm_op_close
,
2325 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2331 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2332 vma
->vm_page_prot
)));
2334 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2335 vma
->vm_page_prot
));
2337 entry
= pte_mkyoung(entry
);
2338 entry
= pte_mkhuge(entry
);
2339 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2344 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2345 unsigned long address
, pte_t
*ptep
)
2349 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2350 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2351 update_mmu_cache(vma
, address
, ptep
);
2355 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2356 struct vm_area_struct
*vma
)
2358 pte_t
*src_pte
, *dst_pte
, entry
;
2359 struct page
*ptepage
;
2362 struct hstate
*h
= hstate_vma(vma
);
2363 unsigned long sz
= huge_page_size(h
);
2365 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2367 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2368 src_pte
= huge_pte_offset(src
, addr
);
2371 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2375 /* If the pagetables are shared don't copy or take references */
2376 if (dst_pte
== src_pte
)
2379 spin_lock(&dst
->page_table_lock
);
2380 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2381 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2383 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2384 entry
= huge_ptep_get(src_pte
);
2385 ptepage
= pte_page(entry
);
2387 page_dup_rmap(ptepage
);
2388 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2390 spin_unlock(&src
->page_table_lock
);
2391 spin_unlock(&dst
->page_table_lock
);
2399 static int is_hugetlb_entry_migration(pte_t pte
)
2403 if (huge_pte_none(pte
) || pte_present(pte
))
2405 swp
= pte_to_swp_entry(pte
);
2406 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2412 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2416 if (huge_pte_none(pte
) || pte_present(pte
))
2418 swp
= pte_to_swp_entry(pte
);
2419 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2425 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2426 unsigned long start
, unsigned long end
,
2427 struct page
*ref_page
)
2429 int force_flush
= 0;
2430 struct mm_struct
*mm
= vma
->vm_mm
;
2431 unsigned long address
;
2435 struct hstate
*h
= hstate_vma(vma
);
2436 unsigned long sz
= huge_page_size(h
);
2437 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2438 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2440 WARN_ON(!is_vm_hugetlb_page(vma
));
2441 BUG_ON(start
& ~huge_page_mask(h
));
2442 BUG_ON(end
& ~huge_page_mask(h
));
2444 tlb_start_vma(tlb
, vma
);
2445 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2447 spin_lock(&mm
->page_table_lock
);
2448 for (address
= start
; address
< end
; address
+= sz
) {
2449 ptep
= huge_pte_offset(mm
, address
);
2453 if (huge_pmd_unshare(mm
, &address
, ptep
))
2456 pte
= huge_ptep_get(ptep
);
2457 if (huge_pte_none(pte
))
2461 * HWPoisoned hugepage is already unmapped and dropped reference
2463 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2464 huge_pte_clear(mm
, address
, ptep
);
2468 page
= pte_page(pte
);
2470 * If a reference page is supplied, it is because a specific
2471 * page is being unmapped, not a range. Ensure the page we
2472 * are about to unmap is the actual page of interest.
2475 if (page
!= ref_page
)
2479 * Mark the VMA as having unmapped its page so that
2480 * future faults in this VMA will fail rather than
2481 * looking like data was lost
2483 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2486 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2487 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2488 if (huge_pte_dirty(pte
))
2489 set_page_dirty(page
);
2491 page_remove_rmap(page
);
2492 force_flush
= !__tlb_remove_page(tlb
, page
);
2495 /* Bail out after unmapping reference page if supplied */
2499 spin_unlock(&mm
->page_table_lock
);
2501 * mmu_gather ran out of room to batch pages, we break out of
2502 * the PTE lock to avoid doing the potential expensive TLB invalidate
2503 * and page-free while holding it.
2508 if (address
< end
&& !ref_page
)
2511 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2512 tlb_end_vma(tlb
, vma
);
2515 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2516 struct vm_area_struct
*vma
, unsigned long start
,
2517 unsigned long end
, struct page
*ref_page
)
2519 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2522 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2523 * test will fail on a vma being torn down, and not grab a page table
2524 * on its way out. We're lucky that the flag has such an appropriate
2525 * name, and can in fact be safely cleared here. We could clear it
2526 * before the __unmap_hugepage_range above, but all that's necessary
2527 * is to clear it before releasing the i_mmap_mutex. This works
2528 * because in the context this is called, the VMA is about to be
2529 * destroyed and the i_mmap_mutex is held.
2531 vma
->vm_flags
&= ~VM_MAYSHARE
;
2534 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2535 unsigned long end
, struct page
*ref_page
)
2537 struct mm_struct
*mm
;
2538 struct mmu_gather tlb
;
2542 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2543 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2544 tlb_finish_mmu(&tlb
, start
, end
);
2548 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2549 * mappping it owns the reserve page for. The intention is to unmap the page
2550 * from other VMAs and let the children be SIGKILLed if they are faulting the
2553 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2554 struct page
*page
, unsigned long address
)
2556 struct hstate
*h
= hstate_vma(vma
);
2557 struct vm_area_struct
*iter_vma
;
2558 struct address_space
*mapping
;
2562 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2563 * from page cache lookup which is in HPAGE_SIZE units.
2565 address
= address
& huge_page_mask(h
);
2566 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2568 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2571 * Take the mapping lock for the duration of the table walk. As
2572 * this mapping should be shared between all the VMAs,
2573 * __unmap_hugepage_range() is called as the lock is already held
2575 mutex_lock(&mapping
->i_mmap_mutex
);
2576 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2577 /* Do not unmap the current VMA */
2578 if (iter_vma
== vma
)
2582 * Unmap the page from other VMAs without their own reserves.
2583 * They get marked to be SIGKILLed if they fault in these
2584 * areas. This is because a future no-page fault on this VMA
2585 * could insert a zeroed page instead of the data existing
2586 * from the time of fork. This would look like data corruption
2588 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2589 unmap_hugepage_range(iter_vma
, address
,
2590 address
+ huge_page_size(h
), page
);
2592 mutex_unlock(&mapping
->i_mmap_mutex
);
2598 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2599 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2600 * cannot race with other handlers or page migration.
2601 * Keep the pte_same checks anyway to make transition from the mutex easier.
2603 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2604 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2605 struct page
*pagecache_page
)
2607 struct hstate
*h
= hstate_vma(vma
);
2608 struct page
*old_page
, *new_page
;
2609 int outside_reserve
= 0;
2610 unsigned long mmun_start
; /* For mmu_notifiers */
2611 unsigned long mmun_end
; /* For mmu_notifiers */
2613 old_page
= pte_page(pte
);
2616 /* If no-one else is actually using this page, avoid the copy
2617 * and just make the page writable */
2618 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2619 page_move_anon_rmap(old_page
, vma
, address
);
2620 set_huge_ptep_writable(vma
, address
, ptep
);
2625 * If the process that created a MAP_PRIVATE mapping is about to
2626 * perform a COW due to a shared page count, attempt to satisfy
2627 * the allocation without using the existing reserves. The pagecache
2628 * page is used to determine if the reserve at this address was
2629 * consumed or not. If reserves were used, a partial faulted mapping
2630 * at the time of fork() could consume its reserves on COW instead
2631 * of the full address range.
2633 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2634 old_page
!= pagecache_page
)
2635 outside_reserve
= 1;
2637 page_cache_get(old_page
);
2639 /* Drop page_table_lock as buddy allocator may be called */
2640 spin_unlock(&mm
->page_table_lock
);
2641 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2643 if (IS_ERR(new_page
)) {
2644 long err
= PTR_ERR(new_page
);
2645 page_cache_release(old_page
);
2648 * If a process owning a MAP_PRIVATE mapping fails to COW,
2649 * it is due to references held by a child and an insufficient
2650 * huge page pool. To guarantee the original mappers
2651 * reliability, unmap the page from child processes. The child
2652 * may get SIGKILLed if it later faults.
2654 if (outside_reserve
) {
2655 BUG_ON(huge_pte_none(pte
));
2656 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2657 BUG_ON(huge_pte_none(pte
));
2658 spin_lock(&mm
->page_table_lock
);
2659 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2660 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2661 goto retry_avoidcopy
;
2663 * race occurs while re-acquiring page_table_lock, and
2671 /* Caller expects lock to be held */
2672 spin_lock(&mm
->page_table_lock
);
2674 return VM_FAULT_OOM
;
2676 return VM_FAULT_SIGBUS
;
2680 * When the original hugepage is shared one, it does not have
2681 * anon_vma prepared.
2683 if (unlikely(anon_vma_prepare(vma
))) {
2684 page_cache_release(new_page
);
2685 page_cache_release(old_page
);
2686 /* Caller expects lock to be held */
2687 spin_lock(&mm
->page_table_lock
);
2688 return VM_FAULT_OOM
;
2691 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2692 pages_per_huge_page(h
));
2693 __SetPageUptodate(new_page
);
2695 mmun_start
= address
& huge_page_mask(h
);
2696 mmun_end
= mmun_start
+ huge_page_size(h
);
2697 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2699 * Retake the page_table_lock to check for racing updates
2700 * before the page tables are altered
2702 spin_lock(&mm
->page_table_lock
);
2703 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2704 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2705 ClearPagePrivate(new_page
);
2708 huge_ptep_clear_flush(vma
, address
, ptep
);
2709 set_huge_pte_at(mm
, address
, ptep
,
2710 make_huge_pte(vma
, new_page
, 1));
2711 page_remove_rmap(old_page
);
2712 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2713 /* Make the old page be freed below */
2714 new_page
= old_page
;
2716 spin_unlock(&mm
->page_table_lock
);
2717 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2718 page_cache_release(new_page
);
2719 page_cache_release(old_page
);
2721 /* Caller expects lock to be held */
2722 spin_lock(&mm
->page_table_lock
);
2726 /* Return the pagecache page at a given address within a VMA */
2727 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2728 struct vm_area_struct
*vma
, unsigned long address
)
2730 struct address_space
*mapping
;
2733 mapping
= vma
->vm_file
->f_mapping
;
2734 idx
= vma_hugecache_offset(h
, vma
, address
);
2736 return find_lock_page(mapping
, idx
);
2740 * Return whether there is a pagecache page to back given address within VMA.
2741 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2743 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2744 struct vm_area_struct
*vma
, unsigned long address
)
2746 struct address_space
*mapping
;
2750 mapping
= vma
->vm_file
->f_mapping
;
2751 idx
= vma_hugecache_offset(h
, vma
, address
);
2753 page
= find_get_page(mapping
, idx
);
2756 return page
!= NULL
;
2759 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2760 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2762 struct hstate
*h
= hstate_vma(vma
);
2763 int ret
= VM_FAULT_SIGBUS
;
2768 struct address_space
*mapping
;
2772 * Currently, we are forced to kill the process in the event the
2773 * original mapper has unmapped pages from the child due to a failed
2774 * COW. Warn that such a situation has occurred as it may not be obvious
2776 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2777 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2782 mapping
= vma
->vm_file
->f_mapping
;
2783 idx
= vma_hugecache_offset(h
, vma
, address
);
2786 * Use page lock to guard against racing truncation
2787 * before we get page_table_lock.
2790 page
= find_lock_page(mapping
, idx
);
2792 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2795 page
= alloc_huge_page(vma
, address
, 0);
2797 ret
= PTR_ERR(page
);
2801 ret
= VM_FAULT_SIGBUS
;
2804 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2805 __SetPageUptodate(page
);
2807 if (vma
->vm_flags
& VM_MAYSHARE
) {
2809 struct inode
*inode
= mapping
->host
;
2811 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2818 ClearPagePrivate(page
);
2820 spin_lock(&inode
->i_lock
);
2821 inode
->i_blocks
+= blocks_per_huge_page(h
);
2822 spin_unlock(&inode
->i_lock
);
2825 if (unlikely(anon_vma_prepare(vma
))) {
2827 goto backout_unlocked
;
2833 * If memory error occurs between mmap() and fault, some process
2834 * don't have hwpoisoned swap entry for errored virtual address.
2835 * So we need to block hugepage fault by PG_hwpoison bit check.
2837 if (unlikely(PageHWPoison(page
))) {
2838 ret
= VM_FAULT_HWPOISON
|
2839 VM_FAULT_SET_HINDEX(hstate_index(h
));
2840 goto backout_unlocked
;
2845 * If we are going to COW a private mapping later, we examine the
2846 * pending reservations for this page now. This will ensure that
2847 * any allocations necessary to record that reservation occur outside
2850 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2851 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2853 goto backout_unlocked
;
2856 spin_lock(&mm
->page_table_lock
);
2857 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2862 if (!huge_pte_none(huge_ptep_get(ptep
)))
2866 ClearPagePrivate(page
);
2867 hugepage_add_new_anon_rmap(page
, vma
, address
);
2870 page_dup_rmap(page
);
2871 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2872 && (vma
->vm_flags
& VM_SHARED
)));
2873 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2875 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2876 /* Optimization, do the COW without a second fault */
2877 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2880 spin_unlock(&mm
->page_table_lock
);
2886 spin_unlock(&mm
->page_table_lock
);
2893 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2894 unsigned long address
, unsigned int flags
)
2899 struct page
*page
= NULL
;
2900 struct page
*pagecache_page
= NULL
;
2901 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2902 struct hstate
*h
= hstate_vma(vma
);
2904 address
&= huge_page_mask(h
);
2906 ptep
= huge_pte_offset(mm
, address
);
2908 entry
= huge_ptep_get(ptep
);
2909 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2910 migration_entry_wait_huge(mm
, ptep
);
2912 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2913 return VM_FAULT_HWPOISON_LARGE
|
2914 VM_FAULT_SET_HINDEX(hstate_index(h
));
2917 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2919 return VM_FAULT_OOM
;
2922 * Serialize hugepage allocation and instantiation, so that we don't
2923 * get spurious allocation failures if two CPUs race to instantiate
2924 * the same page in the page cache.
2926 mutex_lock(&hugetlb_instantiation_mutex
);
2927 entry
= huge_ptep_get(ptep
);
2928 if (huge_pte_none(entry
)) {
2929 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2936 * If we are going to COW the mapping later, we examine the pending
2937 * reservations for this page now. This will ensure that any
2938 * allocations necessary to record that reservation occur outside the
2939 * spinlock. For private mappings, we also lookup the pagecache
2940 * page now as it is used to determine if a reservation has been
2943 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2944 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2949 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2950 pagecache_page
= hugetlbfs_pagecache_page(h
,
2955 * hugetlb_cow() requires page locks of pte_page(entry) and
2956 * pagecache_page, so here we need take the former one
2957 * when page != pagecache_page or !pagecache_page.
2958 * Note that locking order is always pagecache_page -> page,
2959 * so no worry about deadlock.
2961 page
= pte_page(entry
);
2963 if (page
!= pagecache_page
)
2966 spin_lock(&mm
->page_table_lock
);
2967 /* Check for a racing update before calling hugetlb_cow */
2968 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2969 goto out_page_table_lock
;
2972 if (flags
& FAULT_FLAG_WRITE
) {
2973 if (!huge_pte_write(entry
)) {
2974 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2976 goto out_page_table_lock
;
2978 entry
= huge_pte_mkdirty(entry
);
2980 entry
= pte_mkyoung(entry
);
2981 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2982 flags
& FAULT_FLAG_WRITE
))
2983 update_mmu_cache(vma
, address
, ptep
);
2985 out_page_table_lock
:
2986 spin_unlock(&mm
->page_table_lock
);
2988 if (pagecache_page
) {
2989 unlock_page(pagecache_page
);
2990 put_page(pagecache_page
);
2992 if (page
!= pagecache_page
)
2997 mutex_unlock(&hugetlb_instantiation_mutex
);
3002 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3003 struct page
**pages
, struct vm_area_struct
**vmas
,
3004 unsigned long *position
, unsigned long *nr_pages
,
3005 long i
, unsigned int flags
)
3007 unsigned long pfn_offset
;
3008 unsigned long vaddr
= *position
;
3009 unsigned long remainder
= *nr_pages
;
3010 struct hstate
*h
= hstate_vma(vma
);
3012 spin_lock(&mm
->page_table_lock
);
3013 while (vaddr
< vma
->vm_end
&& remainder
) {
3019 * Some archs (sparc64, sh*) have multiple pte_ts to
3020 * each hugepage. We have to make sure we get the
3021 * first, for the page indexing below to work.
3023 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3024 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3027 * When coredumping, it suits get_dump_page if we just return
3028 * an error where there's an empty slot with no huge pagecache
3029 * to back it. This way, we avoid allocating a hugepage, and
3030 * the sparse dumpfile avoids allocating disk blocks, but its
3031 * huge holes still show up with zeroes where they need to be.
3033 if (absent
&& (flags
& FOLL_DUMP
) &&
3034 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3040 * We need call hugetlb_fault for both hugepages under migration
3041 * (in which case hugetlb_fault waits for the migration,) and
3042 * hwpoisoned hugepages (in which case we need to prevent the
3043 * caller from accessing to them.) In order to do this, we use
3044 * here is_swap_pte instead of is_hugetlb_entry_migration and
3045 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3046 * both cases, and because we can't follow correct pages
3047 * directly from any kind of swap entries.
3049 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3050 ((flags
& FOLL_WRITE
) &&
3051 !huge_pte_write(huge_ptep_get(pte
)))) {
3054 spin_unlock(&mm
->page_table_lock
);
3055 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3056 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3057 spin_lock(&mm
->page_table_lock
);
3058 if (!(ret
& VM_FAULT_ERROR
))
3065 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3066 page
= pte_page(huge_ptep_get(pte
));
3069 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3080 if (vaddr
< vma
->vm_end
&& remainder
&&
3081 pfn_offset
< pages_per_huge_page(h
)) {
3083 * We use pfn_offset to avoid touching the pageframes
3084 * of this compound page.
3089 spin_unlock(&mm
->page_table_lock
);
3090 *nr_pages
= remainder
;
3093 return i
? i
: -EFAULT
;
3096 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3097 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3099 struct mm_struct
*mm
= vma
->vm_mm
;
3100 unsigned long start
= address
;
3103 struct hstate
*h
= hstate_vma(vma
);
3104 unsigned long pages
= 0;
3106 BUG_ON(address
>= end
);
3107 flush_cache_range(vma
, address
, end
);
3109 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3110 spin_lock(&mm
->page_table_lock
);
3111 for (; address
< end
; address
+= huge_page_size(h
)) {
3112 ptep
= huge_pte_offset(mm
, address
);
3115 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3119 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3120 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3121 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3122 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3123 set_huge_pte_at(mm
, address
, ptep
, pte
);
3127 spin_unlock(&mm
->page_table_lock
);
3129 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3130 * may have cleared our pud entry and done put_page on the page table:
3131 * once we release i_mmap_mutex, another task can do the final put_page
3132 * and that page table be reused and filled with junk.
3134 flush_tlb_range(vma
, start
, end
);
3135 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3137 return pages
<< h
->order
;
3140 int hugetlb_reserve_pages(struct inode
*inode
,
3142 struct vm_area_struct
*vma
,
3143 vm_flags_t vm_flags
)
3146 struct hstate
*h
= hstate_inode(inode
);
3147 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3150 * Only apply hugepage reservation if asked. At fault time, an
3151 * attempt will be made for VM_NORESERVE to allocate a page
3152 * without using reserves
3154 if (vm_flags
& VM_NORESERVE
)
3158 * Shared mappings base their reservation on the number of pages that
3159 * are already allocated on behalf of the file. Private mappings need
3160 * to reserve the full area even if read-only as mprotect() may be
3161 * called to make the mapping read-write. Assume !vma is a shm mapping
3163 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3164 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3166 struct resv_map
*resv_map
= resv_map_alloc();
3172 set_vma_resv_map(vma
, resv_map
);
3173 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3181 /* There must be enough pages in the subpool for the mapping */
3182 if (hugepage_subpool_get_pages(spool
, chg
)) {
3188 * Check enough hugepages are available for the reservation.
3189 * Hand the pages back to the subpool if there are not
3191 ret
= hugetlb_acct_memory(h
, chg
);
3193 hugepage_subpool_put_pages(spool
, chg
);
3198 * Account for the reservations made. Shared mappings record regions
3199 * that have reservations as they are shared by multiple VMAs.
3200 * When the last VMA disappears, the region map says how much
3201 * the reservation was and the page cache tells how much of
3202 * the reservation was consumed. Private mappings are per-VMA and
3203 * only the consumed reservations are tracked. When the VMA
3204 * disappears, the original reservation is the VMA size and the
3205 * consumed reservations are stored in the map. Hence, nothing
3206 * else has to be done for private mappings here
3208 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3209 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3217 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3219 struct hstate
*h
= hstate_inode(inode
);
3220 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3221 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3223 spin_lock(&inode
->i_lock
);
3224 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3225 spin_unlock(&inode
->i_lock
);
3227 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3228 hugetlb_acct_memory(h
, -(chg
- freed
));
3231 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3232 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3233 struct vm_area_struct
*vma
,
3234 unsigned long addr
, pgoff_t idx
)
3236 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3238 unsigned long sbase
= saddr
& PUD_MASK
;
3239 unsigned long s_end
= sbase
+ PUD_SIZE
;
3241 /* Allow segments to share if only one is marked locked */
3242 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3243 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3246 * match the virtual addresses, permission and the alignment of the
3249 if (pmd_index(addr
) != pmd_index(saddr
) ||
3250 vm_flags
!= svm_flags
||
3251 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3257 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3259 unsigned long base
= addr
& PUD_MASK
;
3260 unsigned long end
= base
+ PUD_SIZE
;
3263 * check on proper vm_flags and page table alignment
3265 if (vma
->vm_flags
& VM_MAYSHARE
&&
3266 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3272 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3273 * and returns the corresponding pte. While this is not necessary for the
3274 * !shared pmd case because we can allocate the pmd later as well, it makes the
3275 * code much cleaner. pmd allocation is essential for the shared case because
3276 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3277 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3278 * bad pmd for sharing.
3280 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3282 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3283 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3284 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3286 struct vm_area_struct
*svma
;
3287 unsigned long saddr
;
3291 if (!vma_shareable(vma
, addr
))
3292 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3294 mutex_lock(&mapping
->i_mmap_mutex
);
3295 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3299 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3301 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3303 get_page(virt_to_page(spte
));
3312 spin_lock(&mm
->page_table_lock
);
3314 pud_populate(mm
, pud
,
3315 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3317 put_page(virt_to_page(spte
));
3318 spin_unlock(&mm
->page_table_lock
);
3320 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3321 mutex_unlock(&mapping
->i_mmap_mutex
);
3326 * unmap huge page backed by shared pte.
3328 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3329 * indicated by page_count > 1, unmap is achieved by clearing pud and
3330 * decrementing the ref count. If count == 1, the pte page is not shared.
3332 * called with vma->vm_mm->page_table_lock held.
3334 * returns: 1 successfully unmapped a shared pte page
3335 * 0 the underlying pte page is not shared, or it is the last user
3337 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3339 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3340 pud_t
*pud
= pud_offset(pgd
, *addr
);
3342 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3343 if (page_count(virt_to_page(ptep
)) == 1)
3347 put_page(virt_to_page(ptep
));
3348 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3351 #define want_pmd_share() (1)
3352 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3353 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3357 #define want_pmd_share() (0)
3358 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3360 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3361 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3362 unsigned long addr
, unsigned long sz
)
3368 pgd
= pgd_offset(mm
, addr
);
3369 pud
= pud_alloc(mm
, pgd
, addr
);
3371 if (sz
== PUD_SIZE
) {
3374 BUG_ON(sz
!= PMD_SIZE
);
3375 if (want_pmd_share() && pud_none(*pud
))
3376 pte
= huge_pmd_share(mm
, addr
, pud
);
3378 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3381 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3386 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3392 pgd
= pgd_offset(mm
, addr
);
3393 if (pgd_present(*pgd
)) {
3394 pud
= pud_offset(pgd
, addr
);
3395 if (pud_present(*pud
)) {
3397 return (pte_t
*)pud
;
3398 pmd
= pmd_offset(pud
, addr
);
3401 return (pte_t
*) pmd
;
3405 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3406 pmd_t
*pmd
, int write
)
3410 page
= pte_page(*(pte_t
*)pmd
);
3412 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3417 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3418 pud_t
*pud
, int write
)
3422 page
= pte_page(*(pte_t
*)pud
);
3424 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3428 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3430 /* Can be overriden by architectures */
3431 __attribute__((weak
)) struct page
*
3432 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3433 pud_t
*pud
, int write
)
3439 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3441 #ifdef CONFIG_MEMORY_FAILURE
3443 /* Should be called in hugetlb_lock */
3444 static int is_hugepage_on_freelist(struct page
*hpage
)
3448 struct hstate
*h
= page_hstate(hpage
);
3449 int nid
= page_to_nid(hpage
);
3451 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3458 * This function is called from memory failure code.
3459 * Assume the caller holds page lock of the head page.
3461 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3463 struct hstate
*h
= page_hstate(hpage
);
3464 int nid
= page_to_nid(hpage
);
3467 spin_lock(&hugetlb_lock
);
3468 if (is_hugepage_on_freelist(hpage
)) {
3470 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3471 * but dangling hpage->lru can trigger list-debug warnings
3472 * (this happens when we call unpoison_memory() on it),
3473 * so let it point to itself with list_del_init().
3475 list_del_init(&hpage
->lru
);
3476 set_page_refcounted(hpage
);
3477 h
->free_huge_pages
--;
3478 h
->free_huge_pages_node
[nid
]--;
3481 spin_unlock(&hugetlb_lock
);
3486 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3488 VM_BUG_ON(!PageHead(page
));
3489 if (!get_page_unless_zero(page
))
3491 spin_lock(&hugetlb_lock
);
3492 list_move_tail(&page
->lru
, list
);
3493 spin_unlock(&hugetlb_lock
);
3497 void putback_active_hugepage(struct page
*page
)
3499 VM_BUG_ON(!PageHead(page
));
3500 spin_lock(&hugetlb_lock
);
3501 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3502 spin_unlock(&hugetlb_lock
);
3506 bool is_hugepage_active(struct page
*page
)
3508 VM_BUG_ON(!PageHuge(page
));
3510 * This function can be called for a tail page because the caller,
3511 * scan_movable_pages, scans through a given pfn-range which typically
3512 * covers one memory block. In systems using gigantic hugepage (1GB
3513 * for x86_64,) a hugepage is larger than a memory block, and we don't
3514 * support migrating such large hugepages for now, so return false
3515 * when called for tail pages.
3520 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3521 * so we should return false for them.
3523 if (unlikely(PageHWPoison(page
)))
3525 return page_count(page
) > 0;