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>
26 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/hugetlb_cgroup.h>
32 #include <linux/node.h>
35 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
36 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
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, nr_huge_pages, and free_huge_pages
53 DEFINE_SPINLOCK(hugetlb_lock
);
55 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
57 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
59 spin_unlock(&spool
->lock
);
61 /* If no pages are used, and no other handles to the subpool
62 * remain, free the subpool the subpool remain */
67 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
69 struct hugepage_subpool
*spool
;
71 spool
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
75 spin_lock_init(&spool
->lock
);
77 spool
->max_hpages
= nr_blocks
;
78 spool
->used_hpages
= 0;
83 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
85 spin_lock(&spool
->lock
);
86 BUG_ON(!spool
->count
);
88 unlock_or_release_subpool(spool
);
91 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
99 spin_lock(&spool
->lock
);
100 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
101 spool
->used_hpages
+= delta
;
105 spin_unlock(&spool
->lock
);
110 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
116 spin_lock(&spool
->lock
);
117 spool
->used_hpages
-= delta
;
118 /* If hugetlbfs_put_super couldn't free spool due to
119 * an outstanding quota reference, free it now. */
120 unlock_or_release_subpool(spool
);
123 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
125 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
128 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
130 return subpool_inode(file_inode(vma
->vm_file
));
134 * Region tracking -- allows tracking of reservations and instantiated pages
135 * across the pages in a mapping.
137 * The region data structures are protected by a combination of the mmap_sem
138 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
139 * must either hold the mmap_sem for write, or the mmap_sem for read and
140 * the hugetlb_instantiation_mutex:
142 * down_write(&mm->mmap_sem);
144 * down_read(&mm->mmap_sem);
145 * mutex_lock(&hugetlb_instantiation_mutex);
148 struct list_head link
;
153 static long region_add(struct list_head
*head
, long f
, long t
)
155 struct file_region
*rg
, *nrg
, *trg
;
157 /* Locate the region we are either in or before. */
158 list_for_each_entry(rg
, head
, link
)
162 /* Round our left edge to the current segment if it encloses us. */
166 /* Check for and consume any regions we now overlap with. */
168 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
169 if (&rg
->link
== head
)
174 /* If this area reaches higher then extend our area to
175 * include it completely. If this is not the first area
176 * which we intend to reuse, free it. */
189 static long region_chg(struct list_head
*head
, long f
, long t
)
191 struct file_region
*rg
, *nrg
;
194 /* Locate the region we are before or in. */
195 list_for_each_entry(rg
, head
, link
)
199 /* If we are below the current region then a new region is required.
200 * Subtle, allocate a new region at the position but make it zero
201 * size such that we can guarantee to record the reservation. */
202 if (&rg
->link
== head
|| t
< rg
->from
) {
203 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
208 INIT_LIST_HEAD(&nrg
->link
);
209 list_add(&nrg
->link
, rg
->link
.prev
);
214 /* Round our left edge to the current segment if it encloses us. */
219 /* Check for and consume any regions we now overlap with. */
220 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
221 if (&rg
->link
== head
)
226 /* We overlap with this area, if it extends further than
227 * us then we must extend ourselves. Account for its
228 * existing reservation. */
233 chg
-= rg
->to
- rg
->from
;
238 static long region_truncate(struct list_head
*head
, long end
)
240 struct file_region
*rg
, *trg
;
243 /* Locate the region we are either in or before. */
244 list_for_each_entry(rg
, head
, link
)
247 if (&rg
->link
== head
)
250 /* If we are in the middle of a region then adjust it. */
251 if (end
> rg
->from
) {
254 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
257 /* Drop any remaining regions. */
258 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
259 if (&rg
->link
== head
)
261 chg
+= rg
->to
- rg
->from
;
268 static long region_count(struct list_head
*head
, long f
, long t
)
270 struct file_region
*rg
;
273 /* Locate each segment we overlap with, and count that overlap. */
274 list_for_each_entry(rg
, head
, link
) {
283 seg_from
= max(rg
->from
, f
);
284 seg_to
= min(rg
->to
, t
);
286 chg
+= seg_to
- seg_from
;
293 * Convert the address within this vma to the page offset within
294 * the mapping, in pagecache page units; huge pages here.
296 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
297 struct vm_area_struct
*vma
, unsigned long address
)
299 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
300 (vma
->vm_pgoff
>> huge_page_order(h
));
303 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
304 unsigned long address
)
306 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
310 * Return the size of the pages allocated when backing a VMA. In the majority
311 * cases this will be same size as used by the page table entries.
313 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
315 struct hstate
*hstate
;
317 if (!is_vm_hugetlb_page(vma
))
320 hstate
= hstate_vma(vma
);
322 return 1UL << huge_page_shift(hstate
);
324 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
327 * Return the page size being used by the MMU to back a VMA. In the majority
328 * of cases, the page size used by the kernel matches the MMU size. On
329 * architectures where it differs, an architecture-specific version of this
330 * function is required.
332 #ifndef vma_mmu_pagesize
333 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
335 return vma_kernel_pagesize(vma
);
340 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
341 * bits of the reservation map pointer, which are always clear due to
344 #define HPAGE_RESV_OWNER (1UL << 0)
345 #define HPAGE_RESV_UNMAPPED (1UL << 1)
346 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
349 * These helpers are used to track how many pages are reserved for
350 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
351 * is guaranteed to have their future faults succeed.
353 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
354 * the reserve counters are updated with the hugetlb_lock held. It is safe
355 * to reset the VMA at fork() time as it is not in use yet and there is no
356 * chance of the global counters getting corrupted as a result of the values.
358 * The private mapping reservation is represented in a subtly different
359 * manner to a shared mapping. A shared mapping has a region map associated
360 * with the underlying file, this region map represents the backing file
361 * pages which have ever had a reservation assigned which this persists even
362 * after the page is instantiated. A private mapping has a region map
363 * associated with the original mmap which is attached to all VMAs which
364 * reference it, this region map represents those offsets which have consumed
365 * reservation ie. where pages have been instantiated.
367 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
369 return (unsigned long)vma
->vm_private_data
;
372 static void set_vma_private_data(struct vm_area_struct
*vma
,
375 vma
->vm_private_data
= (void *)value
;
380 struct list_head regions
;
383 static struct resv_map
*resv_map_alloc(void)
385 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
389 kref_init(&resv_map
->refs
);
390 INIT_LIST_HEAD(&resv_map
->regions
);
395 static void resv_map_release(struct kref
*ref
)
397 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
399 /* Clear out any active regions before we release the map. */
400 region_truncate(&resv_map
->regions
, 0);
404 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
406 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
407 if (!(vma
->vm_flags
& VM_MAYSHARE
))
408 return (struct resv_map
*)(get_vma_private_data(vma
) &
413 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
415 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
416 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
418 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
419 HPAGE_RESV_MASK
) | (unsigned long)map
);
422 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
424 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
425 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
427 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
430 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
432 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
434 return (get_vma_private_data(vma
) & flag
) != 0;
437 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
438 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
440 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
441 if (!(vma
->vm_flags
& VM_MAYSHARE
))
442 vma
->vm_private_data
= (void *)0;
445 /* Returns true if the VMA has associated reserve pages */
446 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
448 if (vma
->vm_flags
& VM_NORESERVE
) {
450 * This address is already reserved by other process(chg == 0),
451 * so, we should decrement reserved count. Without decrementing,
452 * reserve count remains after releasing inode, because this
453 * allocated page will go into page cache and is regarded as
454 * coming from reserved pool in releasing step. Currently, we
455 * don't have any other solution to deal with this situation
456 * properly, so add work-around here.
458 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
464 /* Shared mappings always use reserves */
465 if (vma
->vm_flags
& VM_MAYSHARE
)
469 * Only the process that called mmap() has reserves for
472 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
478 static void copy_gigantic_page(struct page
*dst
, struct page
*src
)
481 struct hstate
*h
= page_hstate(src
);
482 struct page
*dst_base
= dst
;
483 struct page
*src_base
= src
;
485 for (i
= 0; i
< pages_per_huge_page(h
); ) {
487 copy_highpage(dst
, src
);
490 dst
= mem_map_next(dst
, dst_base
, i
);
491 src
= mem_map_next(src
, src_base
, i
);
495 void copy_huge_page(struct page
*dst
, struct page
*src
)
498 struct hstate
*h
= page_hstate(src
);
500 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
501 copy_gigantic_page(dst
, src
);
506 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
508 copy_highpage(dst
+ i
, src
+ i
);
512 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
514 int nid
= page_to_nid(page
);
515 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
516 h
->free_huge_pages
++;
517 h
->free_huge_pages_node
[nid
]++;
520 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
524 if (list_empty(&h
->hugepage_freelists
[nid
]))
526 page
= list_entry(h
->hugepage_freelists
[nid
].next
, struct page
, lru
);
527 list_move(&page
->lru
, &h
->hugepage_activelist
);
528 set_page_refcounted(page
);
529 h
->free_huge_pages
--;
530 h
->free_huge_pages_node
[nid
]--;
534 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
535 struct vm_area_struct
*vma
,
536 unsigned long address
, int avoid_reserve
,
539 struct page
*page
= NULL
;
540 struct mempolicy
*mpol
;
541 nodemask_t
*nodemask
;
542 struct zonelist
*zonelist
;
545 unsigned int cpuset_mems_cookie
;
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
552 if (!vma_has_reserves(vma
, chg
) &&
553 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
556 /* If reserves cannot be used, ensure enough pages are in the pool */
557 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
561 cpuset_mems_cookie
= get_mems_allowed();
562 zonelist
= huge_zonelist(vma
, address
,
563 htlb_alloc_mask
, &mpol
, &nodemask
);
565 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
566 MAX_NR_ZONES
- 1, nodemask
) {
567 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
568 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
572 if (!vma_has_reserves(vma
, chg
))
575 h
->resv_huge_pages
--;
582 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
590 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
594 VM_BUG_ON(h
->order
>= MAX_ORDER
);
597 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
598 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
599 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
600 1 << PG_referenced
| 1 << PG_dirty
|
601 1 << PG_active
| 1 << PG_reserved
|
602 1 << PG_private
| 1 << PG_writeback
);
604 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
605 set_compound_page_dtor(page
, NULL
);
606 set_page_refcounted(page
);
607 arch_release_hugepage(page
);
608 __free_pages(page
, huge_page_order(h
));
611 struct hstate
*size_to_hstate(unsigned long size
)
616 if (huge_page_size(h
) == size
)
622 static void free_huge_page(struct page
*page
)
625 * Can't pass hstate in here because it is called from the
626 * compound page destructor.
628 struct hstate
*h
= page_hstate(page
);
629 int nid
= page_to_nid(page
);
630 struct hugepage_subpool
*spool
=
631 (struct hugepage_subpool
*)page_private(page
);
633 set_page_private(page
, 0);
634 page
->mapping
= NULL
;
635 BUG_ON(page_count(page
));
636 BUG_ON(page_mapcount(page
));
638 spin_lock(&hugetlb_lock
);
639 hugetlb_cgroup_uncharge_page(hstate_index(h
),
640 pages_per_huge_page(h
), page
);
641 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
642 /* remove the page from active list */
643 list_del(&page
->lru
);
644 update_and_free_page(h
, page
);
645 h
->surplus_huge_pages
--;
646 h
->surplus_huge_pages_node
[nid
]--;
648 arch_clear_hugepage_flags(page
);
649 enqueue_huge_page(h
, page
);
651 spin_unlock(&hugetlb_lock
);
652 hugepage_subpool_put_pages(spool
, 1);
655 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
657 INIT_LIST_HEAD(&page
->lru
);
658 set_compound_page_dtor(page
, free_huge_page
);
659 spin_lock(&hugetlb_lock
);
660 set_hugetlb_cgroup(page
, NULL
);
662 h
->nr_huge_pages_node
[nid
]++;
663 spin_unlock(&hugetlb_lock
);
664 put_page(page
); /* free it into the hugepage allocator */
667 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
670 int nr_pages
= 1 << order
;
671 struct page
*p
= page
+ 1;
673 /* we rely on prep_new_huge_page to set the destructor */
674 set_compound_order(page
, order
);
676 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
678 set_page_count(p
, 0);
679 p
->first_page
= page
;
684 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
685 * transparent huge pages. See the PageTransHuge() documentation for more
688 int PageHuge(struct page
*page
)
690 compound_page_dtor
*dtor
;
692 if (!PageCompound(page
))
695 page
= compound_head(page
);
696 dtor
= get_compound_page_dtor(page
);
698 return dtor
== free_huge_page
;
700 EXPORT_SYMBOL_GPL(PageHuge
);
702 pgoff_t
__basepage_index(struct page
*page
)
704 struct page
*page_head
= compound_head(page
);
705 pgoff_t index
= page_index(page_head
);
706 unsigned long compound_idx
;
708 if (!PageHuge(page_head
))
709 return page_index(page
);
711 if (compound_order(page_head
) >= MAX_ORDER
)
712 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
714 compound_idx
= page
- page_head
;
716 return (index
<< compound_order(page_head
)) + compound_idx
;
719 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
723 if (h
->order
>= MAX_ORDER
)
726 page
= alloc_pages_exact_node(nid
,
727 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
728 __GFP_REPEAT
|__GFP_NOWARN
,
731 if (arch_prepare_hugepage(page
)) {
732 __free_pages(page
, huge_page_order(h
));
735 prep_new_huge_page(h
, page
, nid
);
742 * common helper functions for hstate_next_node_to_{alloc|free}.
743 * We may have allocated or freed a huge page based on a different
744 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
745 * be outside of *nodes_allowed. Ensure that we use an allowed
746 * node for alloc or free.
748 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
750 nid
= next_node(nid
, *nodes_allowed
);
751 if (nid
== MAX_NUMNODES
)
752 nid
= first_node(*nodes_allowed
);
753 VM_BUG_ON(nid
>= MAX_NUMNODES
);
758 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
760 if (!node_isset(nid
, *nodes_allowed
))
761 nid
= next_node_allowed(nid
, nodes_allowed
);
766 * returns the previously saved node ["this node"] from which to
767 * allocate a persistent huge page for the pool and advance the
768 * next node from which to allocate, handling wrap at end of node
771 static int hstate_next_node_to_alloc(struct hstate
*h
,
772 nodemask_t
*nodes_allowed
)
776 VM_BUG_ON(!nodes_allowed
);
778 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
779 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
785 * helper for free_pool_huge_page() - return the previously saved
786 * node ["this node"] from which to free a huge page. Advance the
787 * next node id whether or not we find a free huge page to free so
788 * that the next attempt to free addresses the next node.
790 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
794 VM_BUG_ON(!nodes_allowed
);
796 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
797 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
802 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
803 for (nr_nodes = nodes_weight(*mask); \
805 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
808 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
809 for (nr_nodes = nodes_weight(*mask); \
811 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
814 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
820 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
821 page
= alloc_fresh_huge_page_node(h
, node
);
829 count_vm_event(HTLB_BUDDY_PGALLOC
);
831 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
837 * Free huge page from pool from next node to free.
838 * Attempt to keep persistent huge pages more or less
839 * balanced over allowed nodes.
840 * Called with hugetlb_lock locked.
842 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
848 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
850 * If we're returning unused surplus pages, only examine
851 * nodes with surplus pages.
853 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
854 !list_empty(&h
->hugepage_freelists
[node
])) {
856 list_entry(h
->hugepage_freelists
[node
].next
,
858 list_del(&page
->lru
);
859 h
->free_huge_pages
--;
860 h
->free_huge_pages_node
[node
]--;
862 h
->surplus_huge_pages
--;
863 h
->surplus_huge_pages_node
[node
]--;
865 update_and_free_page(h
, page
);
874 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
879 if (h
->order
>= MAX_ORDER
)
883 * Assume we will successfully allocate the surplus page to
884 * prevent racing processes from causing the surplus to exceed
887 * This however introduces a different race, where a process B
888 * tries to grow the static hugepage pool while alloc_pages() is
889 * called by process A. B will only examine the per-node
890 * counters in determining if surplus huge pages can be
891 * converted to normal huge pages in adjust_pool_surplus(). A
892 * won't be able to increment the per-node counter, until the
893 * lock is dropped by B, but B doesn't drop hugetlb_lock until
894 * no more huge pages can be converted from surplus to normal
895 * state (and doesn't try to convert again). Thus, we have a
896 * case where a surplus huge page exists, the pool is grown, and
897 * the surplus huge page still exists after, even though it
898 * should just have been converted to a normal huge page. This
899 * does not leak memory, though, as the hugepage will be freed
900 * once it is out of use. It also does not allow the counters to
901 * go out of whack in adjust_pool_surplus() as we don't modify
902 * the node values until we've gotten the hugepage and only the
903 * per-node value is checked there.
905 spin_lock(&hugetlb_lock
);
906 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
907 spin_unlock(&hugetlb_lock
);
911 h
->surplus_huge_pages
++;
913 spin_unlock(&hugetlb_lock
);
915 if (nid
== NUMA_NO_NODE
)
916 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
917 __GFP_REPEAT
|__GFP_NOWARN
,
920 page
= alloc_pages_exact_node(nid
,
921 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
922 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
924 if (page
&& arch_prepare_hugepage(page
)) {
925 __free_pages(page
, huge_page_order(h
));
929 spin_lock(&hugetlb_lock
);
931 INIT_LIST_HEAD(&page
->lru
);
932 r_nid
= page_to_nid(page
);
933 set_compound_page_dtor(page
, free_huge_page
);
934 set_hugetlb_cgroup(page
, NULL
);
936 * We incremented the global counters already
938 h
->nr_huge_pages_node
[r_nid
]++;
939 h
->surplus_huge_pages_node
[r_nid
]++;
940 __count_vm_event(HTLB_BUDDY_PGALLOC
);
943 h
->surplus_huge_pages
--;
944 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
946 spin_unlock(&hugetlb_lock
);
952 * This allocation function is useful in the context where vma is irrelevant.
953 * E.g. soft-offlining uses this function because it only cares physical
954 * address of error page.
956 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
960 spin_lock(&hugetlb_lock
);
961 page
= dequeue_huge_page_node(h
, nid
);
962 spin_unlock(&hugetlb_lock
);
965 page
= alloc_buddy_huge_page(h
, nid
);
971 * Increase the hugetlb pool such that it can accommodate a reservation
974 static int gather_surplus_pages(struct hstate
*h
, int delta
)
976 struct list_head surplus_list
;
977 struct page
*page
, *tmp
;
979 int needed
, allocated
;
980 bool alloc_ok
= true;
982 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
984 h
->resv_huge_pages
+= delta
;
989 INIT_LIST_HEAD(&surplus_list
);
993 spin_unlock(&hugetlb_lock
);
994 for (i
= 0; i
< needed
; i
++) {
995 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1000 list_add(&page
->lru
, &surplus_list
);
1005 * After retaking hugetlb_lock, we need to recalculate 'needed'
1006 * because either resv_huge_pages or free_huge_pages may have changed.
1008 spin_lock(&hugetlb_lock
);
1009 needed
= (h
->resv_huge_pages
+ delta
) -
1010 (h
->free_huge_pages
+ allocated
);
1015 * We were not able to allocate enough pages to
1016 * satisfy the entire reservation so we free what
1017 * we've allocated so far.
1022 * The surplus_list now contains _at_least_ the number of extra pages
1023 * needed to accommodate the reservation. Add the appropriate number
1024 * of pages to the hugetlb pool and free the extras back to the buddy
1025 * allocator. Commit the entire reservation here to prevent another
1026 * process from stealing the pages as they are added to the pool but
1027 * before they are reserved.
1029 needed
+= allocated
;
1030 h
->resv_huge_pages
+= delta
;
1033 /* Free the needed pages to the hugetlb pool */
1034 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1038 * This page is now managed by the hugetlb allocator and has
1039 * no users -- drop the buddy allocator's reference.
1041 put_page_testzero(page
);
1042 VM_BUG_ON(page_count(page
));
1043 enqueue_huge_page(h
, page
);
1046 spin_unlock(&hugetlb_lock
);
1048 /* Free unnecessary surplus pages to the buddy allocator */
1049 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1051 spin_lock(&hugetlb_lock
);
1057 * When releasing a hugetlb pool reservation, any surplus pages that were
1058 * allocated to satisfy the reservation must be explicitly freed if they were
1060 * Called with hugetlb_lock held.
1062 static void return_unused_surplus_pages(struct hstate
*h
,
1063 unsigned long unused_resv_pages
)
1065 unsigned long nr_pages
;
1067 /* Uncommit the reservation */
1068 h
->resv_huge_pages
-= unused_resv_pages
;
1070 /* Cannot return gigantic pages currently */
1071 if (h
->order
>= MAX_ORDER
)
1074 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1077 * We want to release as many surplus pages as possible, spread
1078 * evenly across all nodes with memory. Iterate across these nodes
1079 * until we can no longer free unreserved surplus pages. This occurs
1080 * when the nodes with surplus pages have no free pages.
1081 * free_pool_huge_page() will balance the the freed pages across the
1082 * on-line nodes with memory and will handle the hstate accounting.
1084 while (nr_pages
--) {
1085 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1091 * Determine if the huge page at addr within the vma has an associated
1092 * reservation. Where it does not we will need to logically increase
1093 * reservation and actually increase subpool usage before an allocation
1094 * can occur. Where any new reservation would be required the
1095 * reservation change is prepared, but not committed. Once the page
1096 * has been allocated from the subpool and instantiated the change should
1097 * be committed via vma_commit_reservation. No action is required on
1100 static long vma_needs_reservation(struct hstate
*h
,
1101 struct vm_area_struct
*vma
, unsigned long addr
)
1103 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1104 struct inode
*inode
= mapping
->host
;
1106 if (vma
->vm_flags
& VM_MAYSHARE
) {
1107 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1108 return region_chg(&inode
->i_mapping
->private_list
,
1111 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1116 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1117 struct resv_map
*reservations
= vma_resv_map(vma
);
1119 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
1125 static void vma_commit_reservation(struct hstate
*h
,
1126 struct vm_area_struct
*vma
, unsigned long addr
)
1128 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1129 struct inode
*inode
= mapping
->host
;
1131 if (vma
->vm_flags
& VM_MAYSHARE
) {
1132 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1133 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1135 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1136 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1137 struct resv_map
*reservations
= vma_resv_map(vma
);
1139 /* Mark this page used in the map. */
1140 region_add(&reservations
->regions
, idx
, idx
+ 1);
1144 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1145 unsigned long addr
, int avoid_reserve
)
1147 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1148 struct hstate
*h
= hstate_vma(vma
);
1152 struct hugetlb_cgroup
*h_cg
;
1154 idx
= hstate_index(h
);
1156 * Processes that did not create the mapping will have no
1157 * reserves and will not have accounted against subpool
1158 * limit. Check that the subpool limit can be made before
1159 * satisfying the allocation MAP_NORESERVE mappings may also
1160 * need pages and subpool limit allocated allocated if no reserve
1163 chg
= vma_needs_reservation(h
, vma
, addr
);
1165 return ERR_PTR(-ENOMEM
);
1167 if (hugepage_subpool_get_pages(spool
, chg
))
1168 return ERR_PTR(-ENOSPC
);
1170 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1172 hugepage_subpool_put_pages(spool
, chg
);
1173 return ERR_PTR(-ENOSPC
);
1175 spin_lock(&hugetlb_lock
);
1176 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1178 spin_unlock(&hugetlb_lock
);
1179 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1181 hugetlb_cgroup_uncharge_cgroup(idx
,
1182 pages_per_huge_page(h
),
1184 hugepage_subpool_put_pages(spool
, chg
);
1185 return ERR_PTR(-ENOSPC
);
1187 spin_lock(&hugetlb_lock
);
1188 list_move(&page
->lru
, &h
->hugepage_activelist
);
1191 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1192 spin_unlock(&hugetlb_lock
);
1194 set_page_private(page
, (unsigned long)spool
);
1196 vma_commit_reservation(h
, vma
, addr
);
1200 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1202 struct huge_bootmem_page
*m
;
1205 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1208 addr
= __alloc_bootmem_node_nopanic(NODE_DATA(node
),
1209 huge_page_size(h
), huge_page_size(h
), 0);
1213 * Use the beginning of the huge page to store the
1214 * huge_bootmem_page struct (until gather_bootmem
1215 * puts them into the mem_map).
1224 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1225 /* Put them into a private list first because mem_map is not up yet */
1226 list_add(&m
->list
, &huge_boot_pages
);
1231 static void prep_compound_huge_page(struct page
*page
, int order
)
1233 if (unlikely(order
> (MAX_ORDER
- 1)))
1234 prep_compound_gigantic_page(page
, order
);
1236 prep_compound_page(page
, order
);
1239 /* Put bootmem huge pages into the standard lists after mem_map is up */
1240 static void __init
gather_bootmem_prealloc(void)
1242 struct huge_bootmem_page
*m
;
1244 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1245 struct hstate
*h
= m
->hstate
;
1248 #ifdef CONFIG_HIGHMEM
1249 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1250 free_bootmem_late((unsigned long)m
,
1251 sizeof(struct huge_bootmem_page
));
1253 page
= virt_to_page(m
);
1255 __ClearPageReserved(page
);
1256 WARN_ON(page_count(page
) != 1);
1257 prep_compound_huge_page(page
, h
->order
);
1258 prep_new_huge_page(h
, page
, page_to_nid(page
));
1260 * If we had gigantic hugepages allocated at boot time, we need
1261 * to restore the 'stolen' pages to totalram_pages in order to
1262 * fix confusing memory reports from free(1) and another
1263 * side-effects, like CommitLimit going negative.
1265 if (h
->order
> (MAX_ORDER
- 1))
1266 adjust_managed_page_count(page
, 1 << h
->order
);
1270 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1274 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1275 if (h
->order
>= MAX_ORDER
) {
1276 if (!alloc_bootmem_huge_page(h
))
1278 } else if (!alloc_fresh_huge_page(h
,
1279 &node_states
[N_MEMORY
]))
1282 h
->max_huge_pages
= i
;
1285 static void __init
hugetlb_init_hstates(void)
1289 for_each_hstate(h
) {
1290 /* oversize hugepages were init'ed in early boot */
1291 if (h
->order
< MAX_ORDER
)
1292 hugetlb_hstate_alloc_pages(h
);
1296 static char * __init
memfmt(char *buf
, unsigned long n
)
1298 if (n
>= (1UL << 30))
1299 sprintf(buf
, "%lu GB", n
>> 30);
1300 else if (n
>= (1UL << 20))
1301 sprintf(buf
, "%lu MB", n
>> 20);
1303 sprintf(buf
, "%lu KB", n
>> 10);
1307 static void __init
report_hugepages(void)
1311 for_each_hstate(h
) {
1313 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1314 memfmt(buf
, huge_page_size(h
)),
1315 h
->free_huge_pages
);
1319 #ifdef CONFIG_HIGHMEM
1320 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1321 nodemask_t
*nodes_allowed
)
1325 if (h
->order
>= MAX_ORDER
)
1328 for_each_node_mask(i
, *nodes_allowed
) {
1329 struct page
*page
, *next
;
1330 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1331 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1332 if (count
>= h
->nr_huge_pages
)
1334 if (PageHighMem(page
))
1336 list_del(&page
->lru
);
1337 update_and_free_page(h
, page
);
1338 h
->free_huge_pages
--;
1339 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1344 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1345 nodemask_t
*nodes_allowed
)
1351 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1352 * balanced by operating on them in a round-robin fashion.
1353 * Returns 1 if an adjustment was made.
1355 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1360 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1363 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1364 if (h
->surplus_huge_pages_node
[node
])
1368 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1369 if (h
->surplus_huge_pages_node
[node
] <
1370 h
->nr_huge_pages_node
[node
])
1377 h
->surplus_huge_pages
+= delta
;
1378 h
->surplus_huge_pages_node
[node
] += delta
;
1382 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1383 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1384 nodemask_t
*nodes_allowed
)
1386 unsigned long min_count
, ret
;
1388 if (h
->order
>= MAX_ORDER
)
1389 return h
->max_huge_pages
;
1392 * Increase the pool size
1393 * First take pages out of surplus state. Then make up the
1394 * remaining difference by allocating fresh huge pages.
1396 * We might race with alloc_buddy_huge_page() here and be unable
1397 * to convert a surplus huge page to a normal huge page. That is
1398 * not critical, though, it just means the overall size of the
1399 * pool might be one hugepage larger than it needs to be, but
1400 * within all the constraints specified by the sysctls.
1402 spin_lock(&hugetlb_lock
);
1403 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1404 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1408 while (count
> persistent_huge_pages(h
)) {
1410 * If this allocation races such that we no longer need the
1411 * page, free_huge_page will handle it by freeing the page
1412 * and reducing the surplus.
1414 spin_unlock(&hugetlb_lock
);
1415 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1416 spin_lock(&hugetlb_lock
);
1420 /* Bail for signals. Probably ctrl-c from user */
1421 if (signal_pending(current
))
1426 * Decrease the pool size
1427 * First return free pages to the buddy allocator (being careful
1428 * to keep enough around to satisfy reservations). Then place
1429 * pages into surplus state as needed so the pool will shrink
1430 * to the desired size as pages become free.
1432 * By placing pages into the surplus state independent of the
1433 * overcommit value, we are allowing the surplus pool size to
1434 * exceed overcommit. There are few sane options here. Since
1435 * alloc_buddy_huge_page() is checking the global counter,
1436 * though, we'll note that we're not allowed to exceed surplus
1437 * and won't grow the pool anywhere else. Not until one of the
1438 * sysctls are changed, or the surplus pages go out of use.
1440 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1441 min_count
= max(count
, min_count
);
1442 try_to_free_low(h
, min_count
, nodes_allowed
);
1443 while (min_count
< persistent_huge_pages(h
)) {
1444 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1447 while (count
< persistent_huge_pages(h
)) {
1448 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1452 ret
= persistent_huge_pages(h
);
1453 spin_unlock(&hugetlb_lock
);
1457 #define HSTATE_ATTR_RO(_name) \
1458 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1460 #define HSTATE_ATTR(_name) \
1461 static struct kobj_attribute _name##_attr = \
1462 __ATTR(_name, 0644, _name##_show, _name##_store)
1464 static struct kobject
*hugepages_kobj
;
1465 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1467 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1469 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1473 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1474 if (hstate_kobjs
[i
] == kobj
) {
1476 *nidp
= NUMA_NO_NODE
;
1480 return kobj_to_node_hstate(kobj
, nidp
);
1483 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1484 struct kobj_attribute
*attr
, char *buf
)
1487 unsigned long nr_huge_pages
;
1490 h
= kobj_to_hstate(kobj
, &nid
);
1491 if (nid
== NUMA_NO_NODE
)
1492 nr_huge_pages
= h
->nr_huge_pages
;
1494 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1496 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1499 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1500 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1501 const char *buf
, size_t len
)
1505 unsigned long count
;
1507 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1509 err
= kstrtoul(buf
, 10, &count
);
1513 h
= kobj_to_hstate(kobj
, &nid
);
1514 if (h
->order
>= MAX_ORDER
) {
1519 if (nid
== NUMA_NO_NODE
) {
1521 * global hstate attribute
1523 if (!(obey_mempolicy
&&
1524 init_nodemask_of_mempolicy(nodes_allowed
))) {
1525 NODEMASK_FREE(nodes_allowed
);
1526 nodes_allowed
= &node_states
[N_MEMORY
];
1528 } else if (nodes_allowed
) {
1530 * per node hstate attribute: adjust count to global,
1531 * but restrict alloc/free to the specified node.
1533 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1534 init_nodemask_of_node(nodes_allowed
, nid
);
1536 nodes_allowed
= &node_states
[N_MEMORY
];
1538 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1540 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1541 NODEMASK_FREE(nodes_allowed
);
1545 NODEMASK_FREE(nodes_allowed
);
1549 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1550 struct kobj_attribute
*attr
, char *buf
)
1552 return nr_hugepages_show_common(kobj
, attr
, buf
);
1555 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1556 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1558 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1560 HSTATE_ATTR(nr_hugepages
);
1565 * hstate attribute for optionally mempolicy-based constraint on persistent
1566 * huge page alloc/free.
1568 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1569 struct kobj_attribute
*attr
, char *buf
)
1571 return nr_hugepages_show_common(kobj
, attr
, buf
);
1574 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1575 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1577 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1579 HSTATE_ATTR(nr_hugepages_mempolicy
);
1583 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1584 struct kobj_attribute
*attr
, char *buf
)
1586 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1587 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1590 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1591 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1594 unsigned long input
;
1595 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1597 if (h
->order
>= MAX_ORDER
)
1600 err
= kstrtoul(buf
, 10, &input
);
1604 spin_lock(&hugetlb_lock
);
1605 h
->nr_overcommit_huge_pages
= input
;
1606 spin_unlock(&hugetlb_lock
);
1610 HSTATE_ATTR(nr_overcommit_hugepages
);
1612 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1613 struct kobj_attribute
*attr
, char *buf
)
1616 unsigned long free_huge_pages
;
1619 h
= kobj_to_hstate(kobj
, &nid
);
1620 if (nid
== NUMA_NO_NODE
)
1621 free_huge_pages
= h
->free_huge_pages
;
1623 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1625 return sprintf(buf
, "%lu\n", free_huge_pages
);
1627 HSTATE_ATTR_RO(free_hugepages
);
1629 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1630 struct kobj_attribute
*attr
, char *buf
)
1632 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1633 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1635 HSTATE_ATTR_RO(resv_hugepages
);
1637 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1638 struct kobj_attribute
*attr
, char *buf
)
1641 unsigned long surplus_huge_pages
;
1644 h
= kobj_to_hstate(kobj
, &nid
);
1645 if (nid
== NUMA_NO_NODE
)
1646 surplus_huge_pages
= h
->surplus_huge_pages
;
1648 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1650 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1652 HSTATE_ATTR_RO(surplus_hugepages
);
1654 static struct attribute
*hstate_attrs
[] = {
1655 &nr_hugepages_attr
.attr
,
1656 &nr_overcommit_hugepages_attr
.attr
,
1657 &free_hugepages_attr
.attr
,
1658 &resv_hugepages_attr
.attr
,
1659 &surplus_hugepages_attr
.attr
,
1661 &nr_hugepages_mempolicy_attr
.attr
,
1666 static struct attribute_group hstate_attr_group
= {
1667 .attrs
= hstate_attrs
,
1670 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1671 struct kobject
**hstate_kobjs
,
1672 struct attribute_group
*hstate_attr_group
)
1675 int hi
= hstate_index(h
);
1677 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1678 if (!hstate_kobjs
[hi
])
1681 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1683 kobject_put(hstate_kobjs
[hi
]);
1688 static void __init
hugetlb_sysfs_init(void)
1693 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1694 if (!hugepages_kobj
)
1697 for_each_hstate(h
) {
1698 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1699 hstate_kobjs
, &hstate_attr_group
);
1701 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1708 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1709 * with node devices in node_devices[] using a parallel array. The array
1710 * index of a node device or _hstate == node id.
1711 * This is here to avoid any static dependency of the node device driver, in
1712 * the base kernel, on the hugetlb module.
1714 struct node_hstate
{
1715 struct kobject
*hugepages_kobj
;
1716 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1718 struct node_hstate node_hstates
[MAX_NUMNODES
];
1721 * A subset of global hstate attributes for node devices
1723 static struct attribute
*per_node_hstate_attrs
[] = {
1724 &nr_hugepages_attr
.attr
,
1725 &free_hugepages_attr
.attr
,
1726 &surplus_hugepages_attr
.attr
,
1730 static struct attribute_group per_node_hstate_attr_group
= {
1731 .attrs
= per_node_hstate_attrs
,
1735 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1736 * Returns node id via non-NULL nidp.
1738 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1742 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1743 struct node_hstate
*nhs
= &node_hstates
[nid
];
1745 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1746 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1758 * Unregister hstate attributes from a single node device.
1759 * No-op if no hstate attributes attached.
1761 static void hugetlb_unregister_node(struct node
*node
)
1764 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1766 if (!nhs
->hugepages_kobj
)
1767 return; /* no hstate attributes */
1769 for_each_hstate(h
) {
1770 int idx
= hstate_index(h
);
1771 if (nhs
->hstate_kobjs
[idx
]) {
1772 kobject_put(nhs
->hstate_kobjs
[idx
]);
1773 nhs
->hstate_kobjs
[idx
] = NULL
;
1777 kobject_put(nhs
->hugepages_kobj
);
1778 nhs
->hugepages_kobj
= NULL
;
1782 * hugetlb module exit: unregister hstate attributes from node devices
1785 static void hugetlb_unregister_all_nodes(void)
1790 * disable node device registrations.
1792 register_hugetlbfs_with_node(NULL
, NULL
);
1795 * remove hstate attributes from any nodes that have them.
1797 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1798 hugetlb_unregister_node(node_devices
[nid
]);
1802 * Register hstate attributes for a single node device.
1803 * No-op if attributes already registered.
1805 static void hugetlb_register_node(struct node
*node
)
1808 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1811 if (nhs
->hugepages_kobj
)
1812 return; /* already allocated */
1814 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1816 if (!nhs
->hugepages_kobj
)
1819 for_each_hstate(h
) {
1820 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1822 &per_node_hstate_attr_group
);
1824 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1825 h
->name
, node
->dev
.id
);
1826 hugetlb_unregister_node(node
);
1833 * hugetlb init time: register hstate attributes for all registered node
1834 * devices of nodes that have memory. All on-line nodes should have
1835 * registered their associated device by this time.
1837 static void hugetlb_register_all_nodes(void)
1841 for_each_node_state(nid
, N_MEMORY
) {
1842 struct node
*node
= node_devices
[nid
];
1843 if (node
->dev
.id
== nid
)
1844 hugetlb_register_node(node
);
1848 * Let the node device driver know we're here so it can
1849 * [un]register hstate attributes on node hotplug.
1851 register_hugetlbfs_with_node(hugetlb_register_node
,
1852 hugetlb_unregister_node
);
1854 #else /* !CONFIG_NUMA */
1856 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1864 static void hugetlb_unregister_all_nodes(void) { }
1866 static void hugetlb_register_all_nodes(void) { }
1870 static void __exit
hugetlb_exit(void)
1874 hugetlb_unregister_all_nodes();
1876 for_each_hstate(h
) {
1877 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1880 kobject_put(hugepages_kobj
);
1882 module_exit(hugetlb_exit
);
1884 static int __init
hugetlb_init(void)
1886 /* Some platform decide whether they support huge pages at boot
1887 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1888 * there is no such support
1890 if (HPAGE_SHIFT
== 0)
1893 if (!size_to_hstate(default_hstate_size
)) {
1894 default_hstate_size
= HPAGE_SIZE
;
1895 if (!size_to_hstate(default_hstate_size
))
1896 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1898 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1899 if (default_hstate_max_huge_pages
)
1900 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1902 hugetlb_init_hstates();
1903 gather_bootmem_prealloc();
1906 hugetlb_sysfs_init();
1907 hugetlb_register_all_nodes();
1908 hugetlb_cgroup_file_init();
1912 module_init(hugetlb_init
);
1914 /* Should be called on processing a hugepagesz=... option */
1915 void __init
hugetlb_add_hstate(unsigned order
)
1920 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1921 pr_warning("hugepagesz= specified twice, ignoring\n");
1924 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1926 h
= &hstates
[hugetlb_max_hstate
++];
1928 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1929 h
->nr_huge_pages
= 0;
1930 h
->free_huge_pages
= 0;
1931 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1932 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1933 INIT_LIST_HEAD(&h
->hugepage_activelist
);
1934 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
1935 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
1936 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1937 huge_page_size(h
)/1024);
1942 static int __init
hugetlb_nrpages_setup(char *s
)
1945 static unsigned long *last_mhp
;
1948 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1949 * so this hugepages= parameter goes to the "default hstate".
1951 if (!hugetlb_max_hstate
)
1952 mhp
= &default_hstate_max_huge_pages
;
1954 mhp
= &parsed_hstate
->max_huge_pages
;
1956 if (mhp
== last_mhp
) {
1957 pr_warning("hugepages= specified twice without "
1958 "interleaving hugepagesz=, ignoring\n");
1962 if (sscanf(s
, "%lu", mhp
) <= 0)
1966 * Global state is always initialized later in hugetlb_init.
1967 * But we need to allocate >= MAX_ORDER hstates here early to still
1968 * use the bootmem allocator.
1970 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1971 hugetlb_hstate_alloc_pages(parsed_hstate
);
1977 __setup("hugepages=", hugetlb_nrpages_setup
);
1979 static int __init
hugetlb_default_setup(char *s
)
1981 default_hstate_size
= memparse(s
, &s
);
1984 __setup("default_hugepagesz=", hugetlb_default_setup
);
1986 static unsigned int cpuset_mems_nr(unsigned int *array
)
1989 unsigned int nr
= 0;
1991 for_each_node_mask(node
, cpuset_current_mems_allowed
)
1997 #ifdef CONFIG_SYSCTL
1998 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
1999 struct ctl_table
*table
, int write
,
2000 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2002 struct hstate
*h
= &default_hstate
;
2006 tmp
= h
->max_huge_pages
;
2008 if (write
&& h
->order
>= MAX_ORDER
)
2012 table
->maxlen
= sizeof(unsigned long);
2013 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2018 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2019 GFP_KERNEL
| __GFP_NORETRY
);
2020 if (!(obey_mempolicy
&&
2021 init_nodemask_of_mempolicy(nodes_allowed
))) {
2022 NODEMASK_FREE(nodes_allowed
);
2023 nodes_allowed
= &node_states
[N_MEMORY
];
2025 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2027 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2028 NODEMASK_FREE(nodes_allowed
);
2034 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2035 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2038 return hugetlb_sysctl_handler_common(false, table
, write
,
2039 buffer
, length
, ppos
);
2043 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2044 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2046 return hugetlb_sysctl_handler_common(true, table
, write
,
2047 buffer
, length
, ppos
);
2049 #endif /* CONFIG_NUMA */
2051 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2052 void __user
*buffer
,
2053 size_t *length
, loff_t
*ppos
)
2055 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2056 if (hugepages_treat_as_movable
)
2057 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2059 htlb_alloc_mask
= GFP_HIGHUSER
;
2063 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2064 void __user
*buffer
,
2065 size_t *length
, loff_t
*ppos
)
2067 struct hstate
*h
= &default_hstate
;
2071 tmp
= h
->nr_overcommit_huge_pages
;
2073 if (write
&& h
->order
>= MAX_ORDER
)
2077 table
->maxlen
= sizeof(unsigned long);
2078 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2083 spin_lock(&hugetlb_lock
);
2084 h
->nr_overcommit_huge_pages
= tmp
;
2085 spin_unlock(&hugetlb_lock
);
2091 #endif /* CONFIG_SYSCTL */
2093 void hugetlb_report_meminfo(struct seq_file
*m
)
2095 struct hstate
*h
= &default_hstate
;
2097 "HugePages_Total: %5lu\n"
2098 "HugePages_Free: %5lu\n"
2099 "HugePages_Rsvd: %5lu\n"
2100 "HugePages_Surp: %5lu\n"
2101 "Hugepagesize: %8lu kB\n",
2105 h
->surplus_huge_pages
,
2106 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2109 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2111 struct hstate
*h
= &default_hstate
;
2113 "Node %d HugePages_Total: %5u\n"
2114 "Node %d HugePages_Free: %5u\n"
2115 "Node %d HugePages_Surp: %5u\n",
2116 nid
, h
->nr_huge_pages_node
[nid
],
2117 nid
, h
->free_huge_pages_node
[nid
],
2118 nid
, h
->surplus_huge_pages_node
[nid
]);
2121 void hugetlb_show_meminfo(void)
2126 for_each_node_state(nid
, N_MEMORY
)
2128 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2130 h
->nr_huge_pages_node
[nid
],
2131 h
->free_huge_pages_node
[nid
],
2132 h
->surplus_huge_pages_node
[nid
],
2133 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2136 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2137 unsigned long hugetlb_total_pages(void)
2140 unsigned long nr_total_pages
= 0;
2143 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2144 return nr_total_pages
;
2147 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2151 spin_lock(&hugetlb_lock
);
2153 * When cpuset is configured, it breaks the strict hugetlb page
2154 * reservation as the accounting is done on a global variable. Such
2155 * reservation is completely rubbish in the presence of cpuset because
2156 * the reservation is not checked against page availability for the
2157 * current cpuset. Application can still potentially OOM'ed by kernel
2158 * with lack of free htlb page in cpuset that the task is in.
2159 * Attempt to enforce strict accounting with cpuset is almost
2160 * impossible (or too ugly) because cpuset is too fluid that
2161 * task or memory node can be dynamically moved between cpusets.
2163 * The change of semantics for shared hugetlb mapping with cpuset is
2164 * undesirable. However, in order to preserve some of the semantics,
2165 * we fall back to check against current free page availability as
2166 * a best attempt and hopefully to minimize the impact of changing
2167 * semantics that cpuset has.
2170 if (gather_surplus_pages(h
, delta
) < 0)
2173 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2174 return_unused_surplus_pages(h
, delta
);
2181 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2184 spin_unlock(&hugetlb_lock
);
2188 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2190 struct resv_map
*reservations
= vma_resv_map(vma
);
2193 * This new VMA should share its siblings reservation map if present.
2194 * The VMA will only ever have a valid reservation map pointer where
2195 * it is being copied for another still existing VMA. As that VMA
2196 * has a reference to the reservation map it cannot disappear until
2197 * after this open call completes. It is therefore safe to take a
2198 * new reference here without additional locking.
2201 kref_get(&reservations
->refs
);
2204 static void resv_map_put(struct vm_area_struct
*vma
)
2206 struct resv_map
*reservations
= vma_resv_map(vma
);
2210 kref_put(&reservations
->refs
, resv_map_release
);
2213 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2215 struct hstate
*h
= hstate_vma(vma
);
2216 struct resv_map
*reservations
= vma_resv_map(vma
);
2217 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2218 unsigned long reserve
;
2219 unsigned long start
;
2223 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2224 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2226 reserve
= (end
- start
) -
2227 region_count(&reservations
->regions
, start
, end
);
2232 hugetlb_acct_memory(h
, -reserve
);
2233 hugepage_subpool_put_pages(spool
, reserve
);
2239 * We cannot handle pagefaults against hugetlb pages at all. They cause
2240 * handle_mm_fault() to try to instantiate regular-sized pages in the
2241 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2244 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2250 const struct vm_operations_struct hugetlb_vm_ops
= {
2251 .fault
= hugetlb_vm_op_fault
,
2252 .open
= hugetlb_vm_op_open
,
2253 .close
= hugetlb_vm_op_close
,
2256 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2262 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2263 vma
->vm_page_prot
)));
2265 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2266 vma
->vm_page_prot
));
2268 entry
= pte_mkyoung(entry
);
2269 entry
= pte_mkhuge(entry
);
2270 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2275 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2276 unsigned long address
, pte_t
*ptep
)
2280 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2281 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2282 update_mmu_cache(vma
, address
, ptep
);
2286 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2287 struct vm_area_struct
*vma
)
2289 pte_t
*src_pte
, *dst_pte
, entry
;
2290 struct page
*ptepage
;
2293 struct hstate
*h
= hstate_vma(vma
);
2294 unsigned long sz
= huge_page_size(h
);
2296 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2298 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2299 src_pte
= huge_pte_offset(src
, addr
);
2302 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2306 /* If the pagetables are shared don't copy or take references */
2307 if (dst_pte
== src_pte
)
2310 spin_lock(&dst
->page_table_lock
);
2311 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2312 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2314 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2315 entry
= huge_ptep_get(src_pte
);
2316 ptepage
= pte_page(entry
);
2318 page_dup_rmap(ptepage
);
2319 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2321 spin_unlock(&src
->page_table_lock
);
2322 spin_unlock(&dst
->page_table_lock
);
2330 static int is_hugetlb_entry_migration(pte_t pte
)
2334 if (huge_pte_none(pte
) || pte_present(pte
))
2336 swp
= pte_to_swp_entry(pte
);
2337 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2343 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2347 if (huge_pte_none(pte
) || pte_present(pte
))
2349 swp
= pte_to_swp_entry(pte
);
2350 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2356 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2357 unsigned long start
, unsigned long end
,
2358 struct page
*ref_page
)
2360 int force_flush
= 0;
2361 struct mm_struct
*mm
= vma
->vm_mm
;
2362 unsigned long address
;
2366 struct hstate
*h
= hstate_vma(vma
);
2367 unsigned long sz
= huge_page_size(h
);
2368 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2369 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2371 WARN_ON(!is_vm_hugetlb_page(vma
));
2372 BUG_ON(start
& ~huge_page_mask(h
));
2373 BUG_ON(end
& ~huge_page_mask(h
));
2375 tlb_start_vma(tlb
, vma
);
2376 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2378 spin_lock(&mm
->page_table_lock
);
2379 for (address
= start
; address
< end
; address
+= sz
) {
2380 ptep
= huge_pte_offset(mm
, address
);
2384 if (huge_pmd_unshare(mm
, &address
, ptep
))
2387 pte
= huge_ptep_get(ptep
);
2388 if (huge_pte_none(pte
))
2392 * HWPoisoned hugepage is already unmapped and dropped reference
2394 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2395 huge_pte_clear(mm
, address
, ptep
);
2399 page
= pte_page(pte
);
2401 * If a reference page is supplied, it is because a specific
2402 * page is being unmapped, not a range. Ensure the page we
2403 * are about to unmap is the actual page of interest.
2406 if (page
!= ref_page
)
2410 * Mark the VMA as having unmapped its page so that
2411 * future faults in this VMA will fail rather than
2412 * looking like data was lost
2414 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2417 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2418 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2419 if (huge_pte_dirty(pte
))
2420 set_page_dirty(page
);
2422 page_remove_rmap(page
);
2423 force_flush
= !__tlb_remove_page(tlb
, page
);
2426 /* Bail out after unmapping reference page if supplied */
2430 spin_unlock(&mm
->page_table_lock
);
2432 * mmu_gather ran out of room to batch pages, we break out of
2433 * the PTE lock to avoid doing the potential expensive TLB invalidate
2434 * and page-free while holding it.
2439 if (address
< end
&& !ref_page
)
2442 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2443 tlb_end_vma(tlb
, vma
);
2446 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2447 struct vm_area_struct
*vma
, unsigned long start
,
2448 unsigned long end
, struct page
*ref_page
)
2450 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2453 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2454 * test will fail on a vma being torn down, and not grab a page table
2455 * on its way out. We're lucky that the flag has such an appropriate
2456 * name, and can in fact be safely cleared here. We could clear it
2457 * before the __unmap_hugepage_range above, but all that's necessary
2458 * is to clear it before releasing the i_mmap_mutex. This works
2459 * because in the context this is called, the VMA is about to be
2460 * destroyed and the i_mmap_mutex is held.
2462 vma
->vm_flags
&= ~VM_MAYSHARE
;
2465 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2466 unsigned long end
, struct page
*ref_page
)
2468 struct mm_struct
*mm
;
2469 struct mmu_gather tlb
;
2473 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2474 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2475 tlb_finish_mmu(&tlb
, start
, end
);
2479 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2480 * mappping it owns the reserve page for. The intention is to unmap the page
2481 * from other VMAs and let the children be SIGKILLed if they are faulting the
2484 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2485 struct page
*page
, unsigned long address
)
2487 struct hstate
*h
= hstate_vma(vma
);
2488 struct vm_area_struct
*iter_vma
;
2489 struct address_space
*mapping
;
2493 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2494 * from page cache lookup which is in HPAGE_SIZE units.
2496 address
= address
& huge_page_mask(h
);
2497 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2499 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2502 * Take the mapping lock for the duration of the table walk. As
2503 * this mapping should be shared between all the VMAs,
2504 * __unmap_hugepage_range() is called as the lock is already held
2506 mutex_lock(&mapping
->i_mmap_mutex
);
2507 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2508 /* Do not unmap the current VMA */
2509 if (iter_vma
== vma
)
2513 * Unmap the page from other VMAs without their own reserves.
2514 * They get marked to be SIGKILLed if they fault in these
2515 * areas. This is because a future no-page fault on this VMA
2516 * could insert a zeroed page instead of the data existing
2517 * from the time of fork. This would look like data corruption
2519 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2520 unmap_hugepage_range(iter_vma
, address
,
2521 address
+ huge_page_size(h
), page
);
2523 mutex_unlock(&mapping
->i_mmap_mutex
);
2529 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2530 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2531 * cannot race with other handlers or page migration.
2532 * Keep the pte_same checks anyway to make transition from the mutex easier.
2534 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2535 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2536 struct page
*pagecache_page
)
2538 struct hstate
*h
= hstate_vma(vma
);
2539 struct page
*old_page
, *new_page
;
2540 int outside_reserve
= 0;
2541 unsigned long mmun_start
; /* For mmu_notifiers */
2542 unsigned long mmun_end
; /* For mmu_notifiers */
2544 old_page
= pte_page(pte
);
2547 /* If no-one else is actually using this page, avoid the copy
2548 * and just make the page writable */
2549 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2550 page_move_anon_rmap(old_page
, vma
, address
);
2551 set_huge_ptep_writable(vma
, address
, ptep
);
2556 * If the process that created a MAP_PRIVATE mapping is about to
2557 * perform a COW due to a shared page count, attempt to satisfy
2558 * the allocation without using the existing reserves. The pagecache
2559 * page is used to determine if the reserve at this address was
2560 * consumed or not. If reserves were used, a partial faulted mapping
2561 * at the time of fork() could consume its reserves on COW instead
2562 * of the full address range.
2564 if (!(vma
->vm_flags
& VM_MAYSHARE
) &&
2565 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2566 old_page
!= pagecache_page
)
2567 outside_reserve
= 1;
2569 page_cache_get(old_page
);
2571 /* Drop page_table_lock as buddy allocator may be called */
2572 spin_unlock(&mm
->page_table_lock
);
2573 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2575 if (IS_ERR(new_page
)) {
2576 long err
= PTR_ERR(new_page
);
2577 page_cache_release(old_page
);
2580 * If a process owning a MAP_PRIVATE mapping fails to COW,
2581 * it is due to references held by a child and an insufficient
2582 * huge page pool. To guarantee the original mappers
2583 * reliability, unmap the page from child processes. The child
2584 * may get SIGKILLed if it later faults.
2586 if (outside_reserve
) {
2587 BUG_ON(huge_pte_none(pte
));
2588 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2589 BUG_ON(huge_pte_none(pte
));
2590 spin_lock(&mm
->page_table_lock
);
2591 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2592 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2593 goto retry_avoidcopy
;
2595 * race occurs while re-acquiring page_table_lock, and
2603 /* Caller expects lock to be held */
2604 spin_lock(&mm
->page_table_lock
);
2606 return VM_FAULT_OOM
;
2608 return VM_FAULT_SIGBUS
;
2612 * When the original hugepage is shared one, it does not have
2613 * anon_vma prepared.
2615 if (unlikely(anon_vma_prepare(vma
))) {
2616 page_cache_release(new_page
);
2617 page_cache_release(old_page
);
2618 /* Caller expects lock to be held */
2619 spin_lock(&mm
->page_table_lock
);
2620 return VM_FAULT_OOM
;
2623 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2624 pages_per_huge_page(h
));
2625 __SetPageUptodate(new_page
);
2627 mmun_start
= address
& huge_page_mask(h
);
2628 mmun_end
= mmun_start
+ huge_page_size(h
);
2629 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2631 * Retake the page_table_lock to check for racing updates
2632 * before the page tables are altered
2634 spin_lock(&mm
->page_table_lock
);
2635 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2636 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2638 huge_ptep_clear_flush(vma
, address
, ptep
);
2639 set_huge_pte_at(mm
, address
, ptep
,
2640 make_huge_pte(vma
, new_page
, 1));
2641 page_remove_rmap(old_page
);
2642 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2643 /* Make the old page be freed below */
2644 new_page
= old_page
;
2646 spin_unlock(&mm
->page_table_lock
);
2647 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2648 /* Caller expects lock to be held */
2649 spin_lock(&mm
->page_table_lock
);
2650 page_cache_release(new_page
);
2651 page_cache_release(old_page
);
2655 /* Return the pagecache page at a given address within a VMA */
2656 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2657 struct vm_area_struct
*vma
, unsigned long address
)
2659 struct address_space
*mapping
;
2662 mapping
= vma
->vm_file
->f_mapping
;
2663 idx
= vma_hugecache_offset(h
, vma
, address
);
2665 return find_lock_page(mapping
, idx
);
2669 * Return whether there is a pagecache page to back given address within VMA.
2670 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2672 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2673 struct vm_area_struct
*vma
, unsigned long address
)
2675 struct address_space
*mapping
;
2679 mapping
= vma
->vm_file
->f_mapping
;
2680 idx
= vma_hugecache_offset(h
, vma
, address
);
2682 page
= find_get_page(mapping
, idx
);
2685 return page
!= NULL
;
2688 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2689 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2691 struct hstate
*h
= hstate_vma(vma
);
2692 int ret
= VM_FAULT_SIGBUS
;
2697 struct address_space
*mapping
;
2701 * Currently, we are forced to kill the process in the event the
2702 * original mapper has unmapped pages from the child due to a failed
2703 * COW. Warn that such a situation has occurred as it may not be obvious
2705 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2706 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2711 mapping
= vma
->vm_file
->f_mapping
;
2712 idx
= vma_hugecache_offset(h
, vma
, address
);
2715 * Use page lock to guard against racing truncation
2716 * before we get page_table_lock.
2719 page
= find_lock_page(mapping
, idx
);
2721 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2724 page
= alloc_huge_page(vma
, address
, 0);
2726 ret
= PTR_ERR(page
);
2730 ret
= VM_FAULT_SIGBUS
;
2733 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2734 __SetPageUptodate(page
);
2736 if (vma
->vm_flags
& VM_MAYSHARE
) {
2738 struct inode
*inode
= mapping
->host
;
2740 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2748 spin_lock(&inode
->i_lock
);
2749 inode
->i_blocks
+= blocks_per_huge_page(h
);
2750 spin_unlock(&inode
->i_lock
);
2753 if (unlikely(anon_vma_prepare(vma
))) {
2755 goto backout_unlocked
;
2761 * If memory error occurs between mmap() and fault, some process
2762 * don't have hwpoisoned swap entry for errored virtual address.
2763 * So we need to block hugepage fault by PG_hwpoison bit check.
2765 if (unlikely(PageHWPoison(page
))) {
2766 ret
= VM_FAULT_HWPOISON
|
2767 VM_FAULT_SET_HINDEX(hstate_index(h
));
2768 goto backout_unlocked
;
2773 * If we are going to COW a private mapping later, we examine the
2774 * pending reservations for this page now. This will ensure that
2775 * any allocations necessary to record that reservation occur outside
2778 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2779 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2781 goto backout_unlocked
;
2784 spin_lock(&mm
->page_table_lock
);
2785 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2790 if (!huge_pte_none(huge_ptep_get(ptep
)))
2794 hugepage_add_new_anon_rmap(page
, vma
, address
);
2796 page_dup_rmap(page
);
2797 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2798 && (vma
->vm_flags
& VM_SHARED
)));
2799 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2801 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2802 /* Optimization, do the COW without a second fault */
2803 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2806 spin_unlock(&mm
->page_table_lock
);
2812 spin_unlock(&mm
->page_table_lock
);
2819 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2820 unsigned long address
, unsigned int flags
)
2825 struct page
*page
= NULL
;
2826 struct page
*pagecache_page
= NULL
;
2827 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2828 struct hstate
*h
= hstate_vma(vma
);
2830 address
&= huge_page_mask(h
);
2832 ptep
= huge_pte_offset(mm
, address
);
2834 entry
= huge_ptep_get(ptep
);
2835 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2836 migration_entry_wait_huge(mm
, ptep
);
2838 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2839 return VM_FAULT_HWPOISON_LARGE
|
2840 VM_FAULT_SET_HINDEX(hstate_index(h
));
2843 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2845 return VM_FAULT_OOM
;
2848 * Serialize hugepage allocation and instantiation, so that we don't
2849 * get spurious allocation failures if two CPUs race to instantiate
2850 * the same page in the page cache.
2852 mutex_lock(&hugetlb_instantiation_mutex
);
2853 entry
= huge_ptep_get(ptep
);
2854 if (huge_pte_none(entry
)) {
2855 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2862 * If we are going to COW the mapping later, we examine the pending
2863 * reservations for this page now. This will ensure that any
2864 * allocations necessary to record that reservation occur outside the
2865 * spinlock. For private mappings, we also lookup the pagecache
2866 * page now as it is used to determine if a reservation has been
2869 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2870 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2875 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2876 pagecache_page
= hugetlbfs_pagecache_page(h
,
2881 * hugetlb_cow() requires page locks of pte_page(entry) and
2882 * pagecache_page, so here we need take the former one
2883 * when page != pagecache_page or !pagecache_page.
2884 * Note that locking order is always pagecache_page -> page,
2885 * so no worry about deadlock.
2887 page
= pte_page(entry
);
2889 if (page
!= pagecache_page
)
2892 spin_lock(&mm
->page_table_lock
);
2893 /* Check for a racing update before calling hugetlb_cow */
2894 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2895 goto out_page_table_lock
;
2898 if (flags
& FAULT_FLAG_WRITE
) {
2899 if (!huge_pte_write(entry
)) {
2900 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2902 goto out_page_table_lock
;
2904 entry
= huge_pte_mkdirty(entry
);
2906 entry
= pte_mkyoung(entry
);
2907 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2908 flags
& FAULT_FLAG_WRITE
))
2909 update_mmu_cache(vma
, address
, ptep
);
2911 out_page_table_lock
:
2912 spin_unlock(&mm
->page_table_lock
);
2914 if (pagecache_page
) {
2915 unlock_page(pagecache_page
);
2916 put_page(pagecache_page
);
2918 if (page
!= pagecache_page
)
2923 mutex_unlock(&hugetlb_instantiation_mutex
);
2928 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2929 struct page
**pages
, struct vm_area_struct
**vmas
,
2930 unsigned long *position
, unsigned long *nr_pages
,
2931 long i
, unsigned int flags
)
2933 unsigned long pfn_offset
;
2934 unsigned long vaddr
= *position
;
2935 unsigned long remainder
= *nr_pages
;
2936 struct hstate
*h
= hstate_vma(vma
);
2938 spin_lock(&mm
->page_table_lock
);
2939 while (vaddr
< vma
->vm_end
&& remainder
) {
2945 * Some archs (sparc64, sh*) have multiple pte_ts to
2946 * each hugepage. We have to make sure we get the
2947 * first, for the page indexing below to work.
2949 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2950 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
2953 * When coredumping, it suits get_dump_page if we just return
2954 * an error where there's an empty slot with no huge pagecache
2955 * to back it. This way, we avoid allocating a hugepage, and
2956 * the sparse dumpfile avoids allocating disk blocks, but its
2957 * huge holes still show up with zeroes where they need to be.
2959 if (absent
&& (flags
& FOLL_DUMP
) &&
2960 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
2966 * We need call hugetlb_fault for both hugepages under migration
2967 * (in which case hugetlb_fault waits for the migration,) and
2968 * hwpoisoned hugepages (in which case we need to prevent the
2969 * caller from accessing to them.) In order to do this, we use
2970 * here is_swap_pte instead of is_hugetlb_entry_migration and
2971 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2972 * both cases, and because we can't follow correct pages
2973 * directly from any kind of swap entries.
2975 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
2976 ((flags
& FOLL_WRITE
) &&
2977 !huge_pte_write(huge_ptep_get(pte
)))) {
2980 spin_unlock(&mm
->page_table_lock
);
2981 ret
= hugetlb_fault(mm
, vma
, vaddr
,
2982 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
2983 spin_lock(&mm
->page_table_lock
);
2984 if (!(ret
& VM_FAULT_ERROR
))
2991 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2992 page
= pte_page(huge_ptep_get(pte
));
2995 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3006 if (vaddr
< vma
->vm_end
&& remainder
&&
3007 pfn_offset
< pages_per_huge_page(h
)) {
3009 * We use pfn_offset to avoid touching the pageframes
3010 * of this compound page.
3015 spin_unlock(&mm
->page_table_lock
);
3016 *nr_pages
= remainder
;
3019 return i
? i
: -EFAULT
;
3022 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3023 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3025 struct mm_struct
*mm
= vma
->vm_mm
;
3026 unsigned long start
= address
;
3029 struct hstate
*h
= hstate_vma(vma
);
3030 unsigned long pages
= 0;
3032 BUG_ON(address
>= end
);
3033 flush_cache_range(vma
, address
, end
);
3035 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3036 spin_lock(&mm
->page_table_lock
);
3037 for (; address
< end
; address
+= huge_page_size(h
)) {
3038 ptep
= huge_pte_offset(mm
, address
);
3041 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3045 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3046 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3047 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3048 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3049 set_huge_pte_at(mm
, address
, ptep
, pte
);
3053 spin_unlock(&mm
->page_table_lock
);
3055 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3056 * may have cleared our pud entry and done put_page on the page table:
3057 * once we release i_mmap_mutex, another task can do the final put_page
3058 * and that page table be reused and filled with junk.
3060 flush_tlb_range(vma
, start
, end
);
3061 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3063 return pages
<< h
->order
;
3066 int hugetlb_reserve_pages(struct inode
*inode
,
3068 struct vm_area_struct
*vma
,
3069 vm_flags_t vm_flags
)
3072 struct hstate
*h
= hstate_inode(inode
);
3073 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3076 * Only apply hugepage reservation if asked. At fault time, an
3077 * attempt will be made for VM_NORESERVE to allocate a page
3078 * without using reserves
3080 if (vm_flags
& VM_NORESERVE
)
3084 * Shared mappings base their reservation on the number of pages that
3085 * are already allocated on behalf of the file. Private mappings need
3086 * to reserve the full area even if read-only as mprotect() may be
3087 * called to make the mapping read-write. Assume !vma is a shm mapping
3089 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3090 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3092 struct resv_map
*resv_map
= resv_map_alloc();
3098 set_vma_resv_map(vma
, resv_map
);
3099 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3107 /* There must be enough pages in the subpool for the mapping */
3108 if (hugepage_subpool_get_pages(spool
, chg
)) {
3114 * Check enough hugepages are available for the reservation.
3115 * Hand the pages back to the subpool if there are not
3117 ret
= hugetlb_acct_memory(h
, chg
);
3119 hugepage_subpool_put_pages(spool
, chg
);
3124 * Account for the reservations made. Shared mappings record regions
3125 * that have reservations as they are shared by multiple VMAs.
3126 * When the last VMA disappears, the region map says how much
3127 * the reservation was and the page cache tells how much of
3128 * the reservation was consumed. Private mappings are per-VMA and
3129 * only the consumed reservations are tracked. When the VMA
3130 * disappears, the original reservation is the VMA size and the
3131 * consumed reservations are stored in the map. Hence, nothing
3132 * else has to be done for private mappings here
3134 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3135 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3143 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3145 struct hstate
*h
= hstate_inode(inode
);
3146 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3147 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3149 spin_lock(&inode
->i_lock
);
3150 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3151 spin_unlock(&inode
->i_lock
);
3153 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3154 hugetlb_acct_memory(h
, -(chg
- freed
));
3157 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3158 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3159 struct vm_area_struct
*vma
,
3160 unsigned long addr
, pgoff_t idx
)
3162 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3164 unsigned long sbase
= saddr
& PUD_MASK
;
3165 unsigned long s_end
= sbase
+ PUD_SIZE
;
3167 /* Allow segments to share if only one is marked locked */
3168 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3169 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3172 * match the virtual addresses, permission and the alignment of the
3175 if (pmd_index(addr
) != pmd_index(saddr
) ||
3176 vm_flags
!= svm_flags
||
3177 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3183 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3185 unsigned long base
= addr
& PUD_MASK
;
3186 unsigned long end
= base
+ PUD_SIZE
;
3189 * check on proper vm_flags and page table alignment
3191 if (vma
->vm_flags
& VM_MAYSHARE
&&
3192 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3198 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3199 * and returns the corresponding pte. While this is not necessary for the
3200 * !shared pmd case because we can allocate the pmd later as well, it makes the
3201 * code much cleaner. pmd allocation is essential for the shared case because
3202 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3203 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3204 * bad pmd for sharing.
3206 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3208 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3209 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3210 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3212 struct vm_area_struct
*svma
;
3213 unsigned long saddr
;
3217 if (!vma_shareable(vma
, addr
))
3218 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3220 mutex_lock(&mapping
->i_mmap_mutex
);
3221 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3225 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3227 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3229 get_page(virt_to_page(spte
));
3238 spin_lock(&mm
->page_table_lock
);
3240 pud_populate(mm
, pud
,
3241 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3243 put_page(virt_to_page(spte
));
3244 spin_unlock(&mm
->page_table_lock
);
3246 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3247 mutex_unlock(&mapping
->i_mmap_mutex
);
3252 * unmap huge page backed by shared pte.
3254 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3255 * indicated by page_count > 1, unmap is achieved by clearing pud and
3256 * decrementing the ref count. If count == 1, the pte page is not shared.
3258 * called with vma->vm_mm->page_table_lock held.
3260 * returns: 1 successfully unmapped a shared pte page
3261 * 0 the underlying pte page is not shared, or it is the last user
3263 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3265 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3266 pud_t
*pud
= pud_offset(pgd
, *addr
);
3268 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3269 if (page_count(virt_to_page(ptep
)) == 1)
3273 put_page(virt_to_page(ptep
));
3274 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3277 #define want_pmd_share() (1)
3278 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3279 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3283 #define want_pmd_share() (0)
3284 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3286 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3287 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3288 unsigned long addr
, unsigned long sz
)
3294 pgd
= pgd_offset(mm
, addr
);
3295 pud
= pud_alloc(mm
, pgd
, addr
);
3297 if (sz
== PUD_SIZE
) {
3300 BUG_ON(sz
!= PMD_SIZE
);
3301 if (want_pmd_share() && pud_none(*pud
))
3302 pte
= huge_pmd_share(mm
, addr
, pud
);
3304 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3307 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3312 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3318 pgd
= pgd_offset(mm
, addr
);
3319 if (pgd_present(*pgd
)) {
3320 pud
= pud_offset(pgd
, addr
);
3321 if (pud_present(*pud
)) {
3323 return (pte_t
*)pud
;
3324 pmd
= pmd_offset(pud
, addr
);
3327 return (pte_t
*) pmd
;
3331 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3332 pmd_t
*pmd
, int write
)
3336 page
= pte_page(*(pte_t
*)pmd
);
3338 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3343 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3344 pud_t
*pud
, int write
)
3348 page
= pte_page(*(pte_t
*)pud
);
3350 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3354 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3356 /* Can be overriden by architectures */
3357 __attribute__((weak
)) struct page
*
3358 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3359 pud_t
*pud
, int write
)
3365 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3367 #ifdef CONFIG_MEMORY_FAILURE
3369 /* Should be called in hugetlb_lock */
3370 static int is_hugepage_on_freelist(struct page
*hpage
)
3374 struct hstate
*h
= page_hstate(hpage
);
3375 int nid
= page_to_nid(hpage
);
3377 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3384 * This function is called from memory failure code.
3385 * Assume the caller holds page lock of the head page.
3387 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3389 struct hstate
*h
= page_hstate(hpage
);
3390 int nid
= page_to_nid(hpage
);
3393 spin_lock(&hugetlb_lock
);
3394 if (is_hugepage_on_freelist(hpage
)) {
3396 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3397 * but dangling hpage->lru can trigger list-debug warnings
3398 * (this happens when we call unpoison_memory() on it),
3399 * so let it point to itself with list_del_init().
3401 list_del_init(&hpage
->lru
);
3402 set_page_refcounted(hpage
);
3403 h
->free_huge_pages
--;
3404 h
->free_huge_pages_node
[nid
]--;
3407 spin_unlock(&hugetlb_lock
);