2 * Generic hugetlb support.
3 * (C) William Irwin, 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(vma
->vm_file
->f_dentry
->d_inode
);
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_instantion_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 << (hstate
->order
+ PAGE_SHIFT
);
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 /* Decrement the reserved pages in the hugepage pool by one */
438 static void decrement_hugepage_resv_vma(struct hstate
*h
,
439 struct vm_area_struct
*vma
)
441 if (vma
->vm_flags
& VM_NORESERVE
)
444 if (vma
->vm_flags
& VM_MAYSHARE
) {
445 /* Shared mappings always use reserves */
446 h
->resv_huge_pages
--;
447 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
449 * Only the process that called mmap() has reserves for
452 h
->resv_huge_pages
--;
456 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
457 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
459 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
460 if (!(vma
->vm_flags
& VM_MAYSHARE
))
461 vma
->vm_private_data
= (void *)0;
464 /* Returns true if the VMA has associated reserve pages */
465 static int vma_has_reserves(struct vm_area_struct
*vma
)
467 if (vma
->vm_flags
& VM_MAYSHARE
)
469 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
474 static void copy_gigantic_page(struct page
*dst
, struct page
*src
)
477 struct hstate
*h
= page_hstate(src
);
478 struct page
*dst_base
= dst
;
479 struct page
*src_base
= src
;
481 for (i
= 0; i
< pages_per_huge_page(h
); ) {
483 copy_highpage(dst
, src
);
486 dst
= mem_map_next(dst
, dst_base
, i
);
487 src
= mem_map_next(src
, src_base
, i
);
491 void copy_huge_page(struct page
*dst
, struct page
*src
)
494 struct hstate
*h
= page_hstate(src
);
496 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
497 copy_gigantic_page(dst
, src
);
502 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
504 copy_highpage(dst
+ i
, src
+ i
);
508 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
510 int nid
= page_to_nid(page
);
511 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
512 h
->free_huge_pages
++;
513 h
->free_huge_pages_node
[nid
]++;
516 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
520 if (list_empty(&h
->hugepage_freelists
[nid
]))
522 page
= list_entry(h
->hugepage_freelists
[nid
].next
, struct page
, lru
);
523 list_move(&page
->lru
, &h
->hugepage_activelist
);
524 set_page_refcounted(page
);
525 h
->free_huge_pages
--;
526 h
->free_huge_pages_node
[nid
]--;
530 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
531 struct vm_area_struct
*vma
,
532 unsigned long address
, int avoid_reserve
)
534 struct page
*page
= NULL
;
535 struct mempolicy
*mpol
;
536 nodemask_t
*nodemask
;
537 struct zonelist
*zonelist
;
540 unsigned int cpuset_mems_cookie
;
543 cpuset_mems_cookie
= get_mems_allowed();
544 zonelist
= huge_zonelist(vma
, address
,
545 htlb_alloc_mask
, &mpol
, &nodemask
);
547 * A child process with MAP_PRIVATE mappings created by their parent
548 * have no page reserves. This check ensures that reservations are
549 * not "stolen". The child may still get SIGKILLed
551 if (!vma_has_reserves(vma
) &&
552 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
555 /* If reserves cannot be used, ensure enough pages are in the pool */
556 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
559 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
560 MAX_NR_ZONES
- 1, nodemask
) {
561 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
562 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
565 decrement_hugepage_resv_vma(h
, vma
);
572 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
581 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
585 VM_BUG_ON(h
->order
>= MAX_ORDER
);
588 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
589 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
590 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
591 1 << PG_referenced
| 1 << PG_dirty
|
592 1 << PG_active
| 1 << PG_reserved
|
593 1 << PG_private
| 1 << PG_writeback
);
595 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
596 set_compound_page_dtor(page
, NULL
);
597 set_page_refcounted(page
);
598 arch_release_hugepage(page
);
599 __free_pages(page
, huge_page_order(h
));
602 struct hstate
*size_to_hstate(unsigned long size
)
607 if (huge_page_size(h
) == size
)
613 static void free_huge_page(struct page
*page
)
616 * Can't pass hstate in here because it is called from the
617 * compound page destructor.
619 struct hstate
*h
= page_hstate(page
);
620 int nid
= page_to_nid(page
);
621 struct hugepage_subpool
*spool
=
622 (struct hugepage_subpool
*)page_private(page
);
624 set_page_private(page
, 0);
625 page
->mapping
= NULL
;
626 BUG_ON(page_count(page
));
627 BUG_ON(page_mapcount(page
));
629 spin_lock(&hugetlb_lock
);
630 hugetlb_cgroup_uncharge_page(hstate_index(h
),
631 pages_per_huge_page(h
), page
);
632 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
633 /* remove the page from active list */
634 list_del(&page
->lru
);
635 update_and_free_page(h
, page
);
636 h
->surplus_huge_pages
--;
637 h
->surplus_huge_pages_node
[nid
]--;
639 arch_clear_hugepage_flags(page
);
640 enqueue_huge_page(h
, page
);
642 spin_unlock(&hugetlb_lock
);
643 hugepage_subpool_put_pages(spool
, 1);
646 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
648 INIT_LIST_HEAD(&page
->lru
);
649 set_compound_page_dtor(page
, free_huge_page
);
650 spin_lock(&hugetlb_lock
);
651 set_hugetlb_cgroup(page
, NULL
);
653 h
->nr_huge_pages_node
[nid
]++;
654 spin_unlock(&hugetlb_lock
);
655 put_page(page
); /* free it into the hugepage allocator */
658 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
661 int nr_pages
= 1 << order
;
662 struct page
*p
= page
+ 1;
664 /* we rely on prep_new_huge_page to set the destructor */
665 set_compound_order(page
, order
);
667 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
669 set_page_count(p
, 0);
670 p
->first_page
= page
;
675 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
676 * transparent huge pages. See the PageTransHuge() documentation for more
679 int PageHuge(struct page
*page
)
681 compound_page_dtor
*dtor
;
683 if (!PageCompound(page
))
686 page
= compound_head(page
);
687 dtor
= get_compound_page_dtor(page
);
689 return dtor
== free_huge_page
;
691 EXPORT_SYMBOL_GPL(PageHuge
);
693 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
697 if (h
->order
>= MAX_ORDER
)
700 page
= alloc_pages_exact_node(nid
,
701 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
702 __GFP_REPEAT
|__GFP_NOWARN
,
705 if (arch_prepare_hugepage(page
)) {
706 __free_pages(page
, huge_page_order(h
));
709 prep_new_huge_page(h
, page
, nid
);
716 * common helper functions for hstate_next_node_to_{alloc|free}.
717 * We may have allocated or freed a huge page based on a different
718 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
719 * be outside of *nodes_allowed. Ensure that we use an allowed
720 * node for alloc or free.
722 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
724 nid
= next_node(nid
, *nodes_allowed
);
725 if (nid
== MAX_NUMNODES
)
726 nid
= first_node(*nodes_allowed
);
727 VM_BUG_ON(nid
>= MAX_NUMNODES
);
732 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
734 if (!node_isset(nid
, *nodes_allowed
))
735 nid
= next_node_allowed(nid
, nodes_allowed
);
740 * returns the previously saved node ["this node"] from which to
741 * allocate a persistent huge page for the pool and advance the
742 * next node from which to allocate, handling wrap at end of node
745 static int hstate_next_node_to_alloc(struct hstate
*h
,
746 nodemask_t
*nodes_allowed
)
750 VM_BUG_ON(!nodes_allowed
);
752 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
753 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
758 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
765 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
766 next_nid
= start_nid
;
769 page
= alloc_fresh_huge_page_node(h
, next_nid
);
774 next_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
775 } while (next_nid
!= start_nid
);
778 count_vm_event(HTLB_BUDDY_PGALLOC
);
780 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
786 * helper for free_pool_huge_page() - return the previously saved
787 * node ["this node"] from which to free a huge page. Advance the
788 * next node id whether or not we find a free huge page to free so
789 * that the next attempt to free addresses the next node.
791 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
795 VM_BUG_ON(!nodes_allowed
);
797 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
798 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
804 * Free huge page from pool from next node to free.
805 * Attempt to keep persistent huge pages more or less
806 * balanced over allowed nodes.
807 * Called with hugetlb_lock locked.
809 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
816 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
817 next_nid
= start_nid
;
821 * If we're returning unused surplus pages, only examine
822 * nodes with surplus pages.
824 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[next_nid
]) &&
825 !list_empty(&h
->hugepage_freelists
[next_nid
])) {
827 list_entry(h
->hugepage_freelists
[next_nid
].next
,
829 list_del(&page
->lru
);
830 h
->free_huge_pages
--;
831 h
->free_huge_pages_node
[next_nid
]--;
833 h
->surplus_huge_pages
--;
834 h
->surplus_huge_pages_node
[next_nid
]--;
836 update_and_free_page(h
, page
);
840 next_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
841 } while (next_nid
!= start_nid
);
846 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
851 if (h
->order
>= MAX_ORDER
)
855 * Assume we will successfully allocate the surplus page to
856 * prevent racing processes from causing the surplus to exceed
859 * This however introduces a different race, where a process B
860 * tries to grow the static hugepage pool while alloc_pages() is
861 * called by process A. B will only examine the per-node
862 * counters in determining if surplus huge pages can be
863 * converted to normal huge pages in adjust_pool_surplus(). A
864 * won't be able to increment the per-node counter, until the
865 * lock is dropped by B, but B doesn't drop hugetlb_lock until
866 * no more huge pages can be converted from surplus to normal
867 * state (and doesn't try to convert again). Thus, we have a
868 * case where a surplus huge page exists, the pool is grown, and
869 * the surplus huge page still exists after, even though it
870 * should just have been converted to a normal huge page. This
871 * does not leak memory, though, as the hugepage will be freed
872 * once it is out of use. It also does not allow the counters to
873 * go out of whack in adjust_pool_surplus() as we don't modify
874 * the node values until we've gotten the hugepage and only the
875 * per-node value is checked there.
877 spin_lock(&hugetlb_lock
);
878 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
879 spin_unlock(&hugetlb_lock
);
883 h
->surplus_huge_pages
++;
885 spin_unlock(&hugetlb_lock
);
887 if (nid
== NUMA_NO_NODE
)
888 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
889 __GFP_REPEAT
|__GFP_NOWARN
,
892 page
= alloc_pages_exact_node(nid
,
893 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
894 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
896 if (page
&& arch_prepare_hugepage(page
)) {
897 __free_pages(page
, huge_page_order(h
));
901 spin_lock(&hugetlb_lock
);
903 INIT_LIST_HEAD(&page
->lru
);
904 r_nid
= page_to_nid(page
);
905 set_compound_page_dtor(page
, free_huge_page
);
906 set_hugetlb_cgroup(page
, NULL
);
908 * We incremented the global counters already
910 h
->nr_huge_pages_node
[r_nid
]++;
911 h
->surplus_huge_pages_node
[r_nid
]++;
912 __count_vm_event(HTLB_BUDDY_PGALLOC
);
915 h
->surplus_huge_pages
--;
916 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
918 spin_unlock(&hugetlb_lock
);
924 * This allocation function is useful in the context where vma is irrelevant.
925 * E.g. soft-offlining uses this function because it only cares physical
926 * address of error page.
928 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
932 spin_lock(&hugetlb_lock
);
933 page
= dequeue_huge_page_node(h
, nid
);
934 spin_unlock(&hugetlb_lock
);
937 page
= alloc_buddy_huge_page(h
, nid
);
943 * Increase the hugetlb pool such that it can accommodate a reservation
946 static int gather_surplus_pages(struct hstate
*h
, int delta
)
948 struct list_head surplus_list
;
949 struct page
*page
, *tmp
;
951 int needed
, allocated
;
952 bool alloc_ok
= true;
954 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
956 h
->resv_huge_pages
+= delta
;
961 INIT_LIST_HEAD(&surplus_list
);
965 spin_unlock(&hugetlb_lock
);
966 for (i
= 0; i
< needed
; i
++) {
967 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
972 list_add(&page
->lru
, &surplus_list
);
977 * After retaking hugetlb_lock, we need to recalculate 'needed'
978 * because either resv_huge_pages or free_huge_pages may have changed.
980 spin_lock(&hugetlb_lock
);
981 needed
= (h
->resv_huge_pages
+ delta
) -
982 (h
->free_huge_pages
+ allocated
);
987 * We were not able to allocate enough pages to
988 * satisfy the entire reservation so we free what
989 * we've allocated so far.
994 * The surplus_list now contains _at_least_ the number of extra pages
995 * needed to accommodate the reservation. Add the appropriate number
996 * of pages to the hugetlb pool and free the extras back to the buddy
997 * allocator. Commit the entire reservation here to prevent another
998 * process from stealing the pages as they are added to the pool but
999 * before they are reserved.
1001 needed
+= allocated
;
1002 h
->resv_huge_pages
+= delta
;
1005 /* Free the needed pages to the hugetlb pool */
1006 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1010 * This page is now managed by the hugetlb allocator and has
1011 * no users -- drop the buddy allocator's reference.
1013 put_page_testzero(page
);
1014 VM_BUG_ON(page_count(page
));
1015 enqueue_huge_page(h
, page
);
1018 spin_unlock(&hugetlb_lock
);
1020 /* Free unnecessary surplus pages to the buddy allocator */
1021 if (!list_empty(&surplus_list
)) {
1022 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1026 spin_lock(&hugetlb_lock
);
1032 * When releasing a hugetlb pool reservation, any surplus pages that were
1033 * allocated to satisfy the reservation must be explicitly freed if they were
1035 * Called with hugetlb_lock held.
1037 static void return_unused_surplus_pages(struct hstate
*h
,
1038 unsigned long unused_resv_pages
)
1040 unsigned long nr_pages
;
1042 /* Uncommit the reservation */
1043 h
->resv_huge_pages
-= unused_resv_pages
;
1045 /* Cannot return gigantic pages currently */
1046 if (h
->order
>= MAX_ORDER
)
1049 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1052 * We want to release as many surplus pages as possible, spread
1053 * evenly across all nodes with memory. Iterate across these nodes
1054 * until we can no longer free unreserved surplus pages. This occurs
1055 * when the nodes with surplus pages have no free pages.
1056 * free_pool_huge_page() will balance the the freed pages across the
1057 * on-line nodes with memory and will handle the hstate accounting.
1059 while (nr_pages
--) {
1060 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1066 * Determine if the huge page at addr within the vma has an associated
1067 * reservation. Where it does not we will need to logically increase
1068 * reservation and actually increase subpool usage before an allocation
1069 * can occur. Where any new reservation would be required the
1070 * reservation change is prepared, but not committed. Once the page
1071 * has been allocated from the subpool and instantiated the change should
1072 * be committed via vma_commit_reservation. No action is required on
1075 static long vma_needs_reservation(struct hstate
*h
,
1076 struct vm_area_struct
*vma
, unsigned long addr
)
1078 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1079 struct inode
*inode
= mapping
->host
;
1081 if (vma
->vm_flags
& VM_MAYSHARE
) {
1082 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1083 return region_chg(&inode
->i_mapping
->private_list
,
1086 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1091 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1092 struct resv_map
*reservations
= vma_resv_map(vma
);
1094 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
1100 static void vma_commit_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 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1110 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1111 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1112 struct resv_map
*reservations
= vma_resv_map(vma
);
1114 /* Mark this page used in the map. */
1115 region_add(&reservations
->regions
, idx
, idx
+ 1);
1119 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1120 unsigned long addr
, int avoid_reserve
)
1122 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1123 struct hstate
*h
= hstate_vma(vma
);
1127 struct hugetlb_cgroup
*h_cg
;
1129 idx
= hstate_index(h
);
1131 * Processes that did not create the mapping will have no
1132 * reserves and will not have accounted against subpool
1133 * limit. Check that the subpool limit can be made before
1134 * satisfying the allocation MAP_NORESERVE mappings may also
1135 * need pages and subpool limit allocated allocated if no reserve
1138 chg
= vma_needs_reservation(h
, vma
, addr
);
1140 return ERR_PTR(-ENOMEM
);
1142 if (hugepage_subpool_get_pages(spool
, chg
))
1143 return ERR_PTR(-ENOSPC
);
1145 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1147 hugepage_subpool_put_pages(spool
, chg
);
1148 return ERR_PTR(-ENOSPC
);
1150 spin_lock(&hugetlb_lock
);
1151 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
1153 /* update page cgroup details */
1154 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
),
1156 spin_unlock(&hugetlb_lock
);
1158 spin_unlock(&hugetlb_lock
);
1159 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1161 hugetlb_cgroup_uncharge_cgroup(idx
,
1162 pages_per_huge_page(h
),
1164 hugepage_subpool_put_pages(spool
, chg
);
1165 return ERR_PTR(-ENOSPC
);
1167 spin_lock(&hugetlb_lock
);
1168 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
),
1170 list_move(&page
->lru
, &h
->hugepage_activelist
);
1171 spin_unlock(&hugetlb_lock
);
1174 set_page_private(page
, (unsigned long)spool
);
1176 vma_commit_reservation(h
, vma
, addr
);
1180 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1182 struct huge_bootmem_page
*m
;
1183 int nr_nodes
= nodes_weight(node_states
[N_MEMORY
]);
1188 addr
= __alloc_bootmem_node_nopanic(
1189 NODE_DATA(hstate_next_node_to_alloc(h
,
1190 &node_states
[N_MEMORY
])),
1191 huge_page_size(h
), huge_page_size(h
), 0);
1195 * Use the beginning of the huge page to store the
1196 * huge_bootmem_page struct (until gather_bootmem
1197 * puts them into the mem_map).
1207 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1208 /* Put them into a private list first because mem_map is not up yet */
1209 list_add(&m
->list
, &huge_boot_pages
);
1214 static void prep_compound_huge_page(struct page
*page
, int order
)
1216 if (unlikely(order
> (MAX_ORDER
- 1)))
1217 prep_compound_gigantic_page(page
, order
);
1219 prep_compound_page(page
, order
);
1222 /* Put bootmem huge pages into the standard lists after mem_map is up */
1223 static void __init
gather_bootmem_prealloc(void)
1225 struct huge_bootmem_page
*m
;
1227 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1228 struct hstate
*h
= m
->hstate
;
1231 #ifdef CONFIG_HIGHMEM
1232 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1233 free_bootmem_late((unsigned long)m
,
1234 sizeof(struct huge_bootmem_page
));
1236 page
= virt_to_page(m
);
1238 __ClearPageReserved(page
);
1239 WARN_ON(page_count(page
) != 1);
1240 prep_compound_huge_page(page
, h
->order
);
1241 prep_new_huge_page(h
, page
, page_to_nid(page
));
1243 * If we had gigantic hugepages allocated at boot time, we need
1244 * to restore the 'stolen' pages to totalram_pages in order to
1245 * fix confusing memory reports from free(1) and another
1246 * side-effects, like CommitLimit going negative.
1248 if (h
->order
> (MAX_ORDER
- 1))
1249 totalram_pages
+= 1 << h
->order
;
1253 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1257 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1258 if (h
->order
>= MAX_ORDER
) {
1259 if (!alloc_bootmem_huge_page(h
))
1261 } else if (!alloc_fresh_huge_page(h
,
1262 &node_states
[N_MEMORY
]))
1265 h
->max_huge_pages
= i
;
1268 static void __init
hugetlb_init_hstates(void)
1272 for_each_hstate(h
) {
1273 /* oversize hugepages were init'ed in early boot */
1274 if (h
->order
< MAX_ORDER
)
1275 hugetlb_hstate_alloc_pages(h
);
1279 static char * __init
memfmt(char *buf
, unsigned long n
)
1281 if (n
>= (1UL << 30))
1282 sprintf(buf
, "%lu GB", n
>> 30);
1283 else if (n
>= (1UL << 20))
1284 sprintf(buf
, "%lu MB", n
>> 20);
1286 sprintf(buf
, "%lu KB", n
>> 10);
1290 static void __init
report_hugepages(void)
1294 for_each_hstate(h
) {
1296 printk(KERN_INFO
"HugeTLB registered %s page size, "
1297 "pre-allocated %ld pages\n",
1298 memfmt(buf
, huge_page_size(h
)),
1299 h
->free_huge_pages
);
1303 #ifdef CONFIG_HIGHMEM
1304 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1305 nodemask_t
*nodes_allowed
)
1309 if (h
->order
>= MAX_ORDER
)
1312 for_each_node_mask(i
, *nodes_allowed
) {
1313 struct page
*page
, *next
;
1314 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1315 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1316 if (count
>= h
->nr_huge_pages
)
1318 if (PageHighMem(page
))
1320 list_del(&page
->lru
);
1321 update_and_free_page(h
, page
);
1322 h
->free_huge_pages
--;
1323 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1328 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1329 nodemask_t
*nodes_allowed
)
1335 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1336 * balanced by operating on them in a round-robin fashion.
1337 * Returns 1 if an adjustment was made.
1339 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1342 int start_nid
, next_nid
;
1345 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1348 start_nid
= hstate_next_node_to_alloc(h
, nodes_allowed
);
1350 start_nid
= hstate_next_node_to_free(h
, nodes_allowed
);
1351 next_nid
= start_nid
;
1357 * To shrink on this node, there must be a surplus page
1359 if (!h
->surplus_huge_pages_node
[nid
]) {
1360 next_nid
= hstate_next_node_to_alloc(h
,
1367 * Surplus cannot exceed the total number of pages
1369 if (h
->surplus_huge_pages_node
[nid
] >=
1370 h
->nr_huge_pages_node
[nid
]) {
1371 next_nid
= hstate_next_node_to_free(h
,
1377 h
->surplus_huge_pages
+= delta
;
1378 h
->surplus_huge_pages_node
[nid
] += delta
;
1381 } while (next_nid
!= start_nid
);
1386 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1387 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1388 nodemask_t
*nodes_allowed
)
1390 unsigned long min_count
, ret
;
1392 if (h
->order
>= MAX_ORDER
)
1393 return h
->max_huge_pages
;
1396 * Increase the pool size
1397 * First take pages out of surplus state. Then make up the
1398 * remaining difference by allocating fresh huge pages.
1400 * We might race with alloc_buddy_huge_page() here and be unable
1401 * to convert a surplus huge page to a normal huge page. That is
1402 * not critical, though, it just means the overall size of the
1403 * pool might be one hugepage larger than it needs to be, but
1404 * within all the constraints specified by the sysctls.
1406 spin_lock(&hugetlb_lock
);
1407 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1408 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1412 while (count
> persistent_huge_pages(h
)) {
1414 * If this allocation races such that we no longer need the
1415 * page, free_huge_page will handle it by freeing the page
1416 * and reducing the surplus.
1418 spin_unlock(&hugetlb_lock
);
1419 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1420 spin_lock(&hugetlb_lock
);
1424 /* Bail for signals. Probably ctrl-c from user */
1425 if (signal_pending(current
))
1430 * Decrease the pool size
1431 * First return free pages to the buddy allocator (being careful
1432 * to keep enough around to satisfy reservations). Then place
1433 * pages into surplus state as needed so the pool will shrink
1434 * to the desired size as pages become free.
1436 * By placing pages into the surplus state independent of the
1437 * overcommit value, we are allowing the surplus pool size to
1438 * exceed overcommit. There are few sane options here. Since
1439 * alloc_buddy_huge_page() is checking the global counter,
1440 * though, we'll note that we're not allowed to exceed surplus
1441 * and won't grow the pool anywhere else. Not until one of the
1442 * sysctls are changed, or the surplus pages go out of use.
1444 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1445 min_count
= max(count
, min_count
);
1446 try_to_free_low(h
, min_count
, nodes_allowed
);
1447 while (min_count
< persistent_huge_pages(h
)) {
1448 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1451 while (count
< persistent_huge_pages(h
)) {
1452 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1456 ret
= persistent_huge_pages(h
);
1457 spin_unlock(&hugetlb_lock
);
1461 #define HSTATE_ATTR_RO(_name) \
1462 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1464 #define HSTATE_ATTR(_name) \
1465 static struct kobj_attribute _name##_attr = \
1466 __ATTR(_name, 0644, _name##_show, _name##_store)
1468 static struct kobject
*hugepages_kobj
;
1469 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1471 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1473 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1477 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1478 if (hstate_kobjs
[i
] == kobj
) {
1480 *nidp
= NUMA_NO_NODE
;
1484 return kobj_to_node_hstate(kobj
, nidp
);
1487 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1488 struct kobj_attribute
*attr
, char *buf
)
1491 unsigned long nr_huge_pages
;
1494 h
= kobj_to_hstate(kobj
, &nid
);
1495 if (nid
== NUMA_NO_NODE
)
1496 nr_huge_pages
= h
->nr_huge_pages
;
1498 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1500 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1503 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1504 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1505 const char *buf
, size_t len
)
1509 unsigned long count
;
1511 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1513 err
= strict_strtoul(buf
, 10, &count
);
1517 h
= kobj_to_hstate(kobj
, &nid
);
1518 if (h
->order
>= MAX_ORDER
) {
1523 if (nid
== NUMA_NO_NODE
) {
1525 * global hstate attribute
1527 if (!(obey_mempolicy
&&
1528 init_nodemask_of_mempolicy(nodes_allowed
))) {
1529 NODEMASK_FREE(nodes_allowed
);
1530 nodes_allowed
= &node_states
[N_MEMORY
];
1532 } else if (nodes_allowed
) {
1534 * per node hstate attribute: adjust count to global,
1535 * but restrict alloc/free to the specified node.
1537 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1538 init_nodemask_of_node(nodes_allowed
, nid
);
1540 nodes_allowed
= &node_states
[N_MEMORY
];
1542 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1544 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1545 NODEMASK_FREE(nodes_allowed
);
1549 NODEMASK_FREE(nodes_allowed
);
1553 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1554 struct kobj_attribute
*attr
, char *buf
)
1556 return nr_hugepages_show_common(kobj
, attr
, buf
);
1559 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1560 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1562 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1564 HSTATE_ATTR(nr_hugepages
);
1569 * hstate attribute for optionally mempolicy-based constraint on persistent
1570 * huge page alloc/free.
1572 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1573 struct kobj_attribute
*attr
, char *buf
)
1575 return nr_hugepages_show_common(kobj
, attr
, buf
);
1578 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1579 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1581 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1583 HSTATE_ATTR(nr_hugepages_mempolicy
);
1587 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1588 struct kobj_attribute
*attr
, char *buf
)
1590 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1591 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1594 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1595 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1598 unsigned long input
;
1599 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1601 if (h
->order
>= MAX_ORDER
)
1604 err
= strict_strtoul(buf
, 10, &input
);
1608 spin_lock(&hugetlb_lock
);
1609 h
->nr_overcommit_huge_pages
= input
;
1610 spin_unlock(&hugetlb_lock
);
1614 HSTATE_ATTR(nr_overcommit_hugepages
);
1616 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1617 struct kobj_attribute
*attr
, char *buf
)
1620 unsigned long free_huge_pages
;
1623 h
= kobj_to_hstate(kobj
, &nid
);
1624 if (nid
== NUMA_NO_NODE
)
1625 free_huge_pages
= h
->free_huge_pages
;
1627 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1629 return sprintf(buf
, "%lu\n", free_huge_pages
);
1631 HSTATE_ATTR_RO(free_hugepages
);
1633 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1634 struct kobj_attribute
*attr
, char *buf
)
1636 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1637 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1639 HSTATE_ATTR_RO(resv_hugepages
);
1641 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1642 struct kobj_attribute
*attr
, char *buf
)
1645 unsigned long surplus_huge_pages
;
1648 h
= kobj_to_hstate(kobj
, &nid
);
1649 if (nid
== NUMA_NO_NODE
)
1650 surplus_huge_pages
= h
->surplus_huge_pages
;
1652 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1654 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1656 HSTATE_ATTR_RO(surplus_hugepages
);
1658 static struct attribute
*hstate_attrs
[] = {
1659 &nr_hugepages_attr
.attr
,
1660 &nr_overcommit_hugepages_attr
.attr
,
1661 &free_hugepages_attr
.attr
,
1662 &resv_hugepages_attr
.attr
,
1663 &surplus_hugepages_attr
.attr
,
1665 &nr_hugepages_mempolicy_attr
.attr
,
1670 static struct attribute_group hstate_attr_group
= {
1671 .attrs
= hstate_attrs
,
1674 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1675 struct kobject
**hstate_kobjs
,
1676 struct attribute_group
*hstate_attr_group
)
1679 int hi
= hstate_index(h
);
1681 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1682 if (!hstate_kobjs
[hi
])
1685 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1687 kobject_put(hstate_kobjs
[hi
]);
1692 static void __init
hugetlb_sysfs_init(void)
1697 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1698 if (!hugepages_kobj
)
1701 for_each_hstate(h
) {
1702 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1703 hstate_kobjs
, &hstate_attr_group
);
1705 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s",
1713 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1714 * with node devices in node_devices[] using a parallel array. The array
1715 * index of a node device or _hstate == node id.
1716 * This is here to avoid any static dependency of the node device driver, in
1717 * the base kernel, on the hugetlb module.
1719 struct node_hstate
{
1720 struct kobject
*hugepages_kobj
;
1721 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1723 struct node_hstate node_hstates
[MAX_NUMNODES
];
1726 * A subset of global hstate attributes for node devices
1728 static struct attribute
*per_node_hstate_attrs
[] = {
1729 &nr_hugepages_attr
.attr
,
1730 &free_hugepages_attr
.attr
,
1731 &surplus_hugepages_attr
.attr
,
1735 static struct attribute_group per_node_hstate_attr_group
= {
1736 .attrs
= per_node_hstate_attrs
,
1740 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1741 * Returns node id via non-NULL nidp.
1743 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1747 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1748 struct node_hstate
*nhs
= &node_hstates
[nid
];
1750 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1751 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1763 * Unregister hstate attributes from a single node device.
1764 * No-op if no hstate attributes attached.
1766 void hugetlb_unregister_node(struct node
*node
)
1769 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1771 if (!nhs
->hugepages_kobj
)
1772 return; /* no hstate attributes */
1774 for_each_hstate(h
) {
1775 int idx
= hstate_index(h
);
1776 if (nhs
->hstate_kobjs
[idx
]) {
1777 kobject_put(nhs
->hstate_kobjs
[idx
]);
1778 nhs
->hstate_kobjs
[idx
] = NULL
;
1782 kobject_put(nhs
->hugepages_kobj
);
1783 nhs
->hugepages_kobj
= NULL
;
1787 * hugetlb module exit: unregister hstate attributes from node devices
1790 static void hugetlb_unregister_all_nodes(void)
1795 * disable node device registrations.
1797 register_hugetlbfs_with_node(NULL
, NULL
);
1800 * remove hstate attributes from any nodes that have them.
1802 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1803 hugetlb_unregister_node(node_devices
[nid
]);
1807 * Register hstate attributes for a single node device.
1808 * No-op if attributes already registered.
1810 void hugetlb_register_node(struct node
*node
)
1813 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1816 if (nhs
->hugepages_kobj
)
1817 return; /* already allocated */
1819 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1821 if (!nhs
->hugepages_kobj
)
1824 for_each_hstate(h
) {
1825 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1827 &per_node_hstate_attr_group
);
1829 printk(KERN_ERR
"Hugetlb: Unable to add hstate %s"
1831 h
->name
, node
->dev
.id
);
1832 hugetlb_unregister_node(node
);
1839 * hugetlb init time: register hstate attributes for all registered node
1840 * devices of nodes that have memory. All on-line nodes should have
1841 * registered their associated device by this time.
1843 static void hugetlb_register_all_nodes(void)
1847 for_each_node_state(nid
, N_MEMORY
) {
1848 struct node
*node
= node_devices
[nid
];
1849 if (node
->dev
.id
== nid
)
1850 hugetlb_register_node(node
);
1854 * Let the node device driver know we're here so it can
1855 * [un]register hstate attributes on node hotplug.
1857 register_hugetlbfs_with_node(hugetlb_register_node
,
1858 hugetlb_unregister_node
);
1860 #else /* !CONFIG_NUMA */
1862 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1870 static void hugetlb_unregister_all_nodes(void) { }
1872 static void hugetlb_register_all_nodes(void) { }
1876 static void __exit
hugetlb_exit(void)
1880 hugetlb_unregister_all_nodes();
1882 for_each_hstate(h
) {
1883 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1886 kobject_put(hugepages_kobj
);
1888 module_exit(hugetlb_exit
);
1890 static int __init
hugetlb_init(void)
1892 /* Some platform decide whether they support huge pages at boot
1893 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1894 * there is no such support
1896 if (HPAGE_SHIFT
== 0)
1899 if (!size_to_hstate(default_hstate_size
)) {
1900 default_hstate_size
= HPAGE_SIZE
;
1901 if (!size_to_hstate(default_hstate_size
))
1902 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1904 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1905 if (default_hstate_max_huge_pages
)
1906 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1908 hugetlb_init_hstates();
1910 gather_bootmem_prealloc();
1914 hugetlb_sysfs_init();
1916 hugetlb_register_all_nodes();
1920 module_init(hugetlb_init
);
1922 /* Should be called on processing a hugepagesz=... option */
1923 void __init
hugetlb_add_hstate(unsigned order
)
1928 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1929 printk(KERN_WARNING
"hugepagesz= specified twice, ignoring\n");
1932 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1934 h
= &hstates
[hugetlb_max_hstate
++];
1936 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1937 h
->nr_huge_pages
= 0;
1938 h
->free_huge_pages
= 0;
1939 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1940 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1941 INIT_LIST_HEAD(&h
->hugepage_activelist
);
1942 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
1943 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
1944 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1945 huge_page_size(h
)/1024);
1947 * Add cgroup control files only if the huge page consists
1948 * of more than two normal pages. This is because we use
1949 * page[2].lru.next for storing cgoup details.
1951 if (order
>= HUGETLB_CGROUP_MIN_ORDER
)
1952 hugetlb_cgroup_file_init(hugetlb_max_hstate
- 1);
1957 static int __init
hugetlb_nrpages_setup(char *s
)
1960 static unsigned long *last_mhp
;
1963 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1964 * so this hugepages= parameter goes to the "default hstate".
1966 if (!hugetlb_max_hstate
)
1967 mhp
= &default_hstate_max_huge_pages
;
1969 mhp
= &parsed_hstate
->max_huge_pages
;
1971 if (mhp
== last_mhp
) {
1972 printk(KERN_WARNING
"hugepages= specified twice without "
1973 "interleaving hugepagesz=, ignoring\n");
1977 if (sscanf(s
, "%lu", mhp
) <= 0)
1981 * Global state is always initialized later in hugetlb_init.
1982 * But we need to allocate >= MAX_ORDER hstates here early to still
1983 * use the bootmem allocator.
1985 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1986 hugetlb_hstate_alloc_pages(parsed_hstate
);
1992 __setup("hugepages=", hugetlb_nrpages_setup
);
1994 static int __init
hugetlb_default_setup(char *s
)
1996 default_hstate_size
= memparse(s
, &s
);
1999 __setup("default_hugepagesz=", hugetlb_default_setup
);
2001 static unsigned int cpuset_mems_nr(unsigned int *array
)
2004 unsigned int nr
= 0;
2006 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2012 #ifdef CONFIG_SYSCTL
2013 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2014 struct ctl_table
*table
, int write
,
2015 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2017 struct hstate
*h
= &default_hstate
;
2021 tmp
= h
->max_huge_pages
;
2023 if (write
&& h
->order
>= MAX_ORDER
)
2027 table
->maxlen
= sizeof(unsigned long);
2028 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2033 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2034 GFP_KERNEL
| __GFP_NORETRY
);
2035 if (!(obey_mempolicy
&&
2036 init_nodemask_of_mempolicy(nodes_allowed
))) {
2037 NODEMASK_FREE(nodes_allowed
);
2038 nodes_allowed
= &node_states
[N_MEMORY
];
2040 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2042 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2043 NODEMASK_FREE(nodes_allowed
);
2049 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2050 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2053 return hugetlb_sysctl_handler_common(false, table
, write
,
2054 buffer
, length
, ppos
);
2058 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2059 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2061 return hugetlb_sysctl_handler_common(true, table
, write
,
2062 buffer
, length
, ppos
);
2064 #endif /* CONFIG_NUMA */
2066 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2067 void __user
*buffer
,
2068 size_t *length
, loff_t
*ppos
)
2070 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2071 if (hugepages_treat_as_movable
)
2072 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2074 htlb_alloc_mask
= GFP_HIGHUSER
;
2078 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2079 void __user
*buffer
,
2080 size_t *length
, loff_t
*ppos
)
2082 struct hstate
*h
= &default_hstate
;
2086 tmp
= h
->nr_overcommit_huge_pages
;
2088 if (write
&& h
->order
>= MAX_ORDER
)
2092 table
->maxlen
= sizeof(unsigned long);
2093 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2098 spin_lock(&hugetlb_lock
);
2099 h
->nr_overcommit_huge_pages
= tmp
;
2100 spin_unlock(&hugetlb_lock
);
2106 #endif /* CONFIG_SYSCTL */
2108 void hugetlb_report_meminfo(struct seq_file
*m
)
2110 struct hstate
*h
= &default_hstate
;
2112 "HugePages_Total: %5lu\n"
2113 "HugePages_Free: %5lu\n"
2114 "HugePages_Rsvd: %5lu\n"
2115 "HugePages_Surp: %5lu\n"
2116 "Hugepagesize: %8lu kB\n",
2120 h
->surplus_huge_pages
,
2121 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2124 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2126 struct hstate
*h
= &default_hstate
;
2128 "Node %d HugePages_Total: %5u\n"
2129 "Node %d HugePages_Free: %5u\n"
2130 "Node %d HugePages_Surp: %5u\n",
2131 nid
, h
->nr_huge_pages_node
[nid
],
2132 nid
, h
->free_huge_pages_node
[nid
],
2133 nid
, h
->surplus_huge_pages_node
[nid
]);
2136 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2137 unsigned long hugetlb_total_pages(void)
2139 struct hstate
*h
= &default_hstate
;
2140 return h
->nr_huge_pages
* pages_per_huge_page(h
);
2143 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2147 spin_lock(&hugetlb_lock
);
2149 * When cpuset is configured, it breaks the strict hugetlb page
2150 * reservation as the accounting is done on a global variable. Such
2151 * reservation is completely rubbish in the presence of cpuset because
2152 * the reservation is not checked against page availability for the
2153 * current cpuset. Application can still potentially OOM'ed by kernel
2154 * with lack of free htlb page in cpuset that the task is in.
2155 * Attempt to enforce strict accounting with cpuset is almost
2156 * impossible (or too ugly) because cpuset is too fluid that
2157 * task or memory node can be dynamically moved between cpusets.
2159 * The change of semantics for shared hugetlb mapping with cpuset is
2160 * undesirable. However, in order to preserve some of the semantics,
2161 * we fall back to check against current free page availability as
2162 * a best attempt and hopefully to minimize the impact of changing
2163 * semantics that cpuset has.
2166 if (gather_surplus_pages(h
, delta
) < 0)
2169 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2170 return_unused_surplus_pages(h
, delta
);
2177 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2180 spin_unlock(&hugetlb_lock
);
2184 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2186 struct resv_map
*reservations
= vma_resv_map(vma
);
2189 * This new VMA should share its siblings reservation map if present.
2190 * The VMA will only ever have a valid reservation map pointer where
2191 * it is being copied for another still existing VMA. As that VMA
2192 * has a reference to the reservation map it cannot disappear until
2193 * after this open call completes. It is therefore safe to take a
2194 * new reference here without additional locking.
2197 kref_get(&reservations
->refs
);
2200 static void resv_map_put(struct vm_area_struct
*vma
)
2202 struct resv_map
*reservations
= vma_resv_map(vma
);
2206 kref_put(&reservations
->refs
, resv_map_release
);
2209 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2211 struct hstate
*h
= hstate_vma(vma
);
2212 struct resv_map
*reservations
= vma_resv_map(vma
);
2213 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2214 unsigned long reserve
;
2215 unsigned long start
;
2219 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2220 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2222 reserve
= (end
- start
) -
2223 region_count(&reservations
->regions
, start
, end
);
2228 hugetlb_acct_memory(h
, -reserve
);
2229 hugepage_subpool_put_pages(spool
, reserve
);
2235 * We cannot handle pagefaults against hugetlb pages at all. They cause
2236 * handle_mm_fault() to try to instantiate regular-sized pages in the
2237 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2240 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2246 const struct vm_operations_struct hugetlb_vm_ops
= {
2247 .fault
= hugetlb_vm_op_fault
,
2248 .open
= hugetlb_vm_op_open
,
2249 .close
= hugetlb_vm_op_close
,
2252 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2259 pte_mkwrite(pte_mkdirty(mk_pte(page
, vma
->vm_page_prot
)));
2261 entry
= huge_pte_wrprotect(mk_pte(page
, vma
->vm_page_prot
));
2263 entry
= pte_mkyoung(entry
);
2264 entry
= pte_mkhuge(entry
);
2265 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2270 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2271 unsigned long address
, pte_t
*ptep
)
2275 entry
= pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep
)));
2276 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2277 update_mmu_cache(vma
, address
, ptep
);
2281 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2282 struct vm_area_struct
*vma
)
2284 pte_t
*src_pte
, *dst_pte
, entry
;
2285 struct page
*ptepage
;
2288 struct hstate
*h
= hstate_vma(vma
);
2289 unsigned long sz
= huge_page_size(h
);
2291 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2293 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2294 src_pte
= huge_pte_offset(src
, addr
);
2297 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2301 /* If the pagetables are shared don't copy or take references */
2302 if (dst_pte
== src_pte
)
2305 spin_lock(&dst
->page_table_lock
);
2306 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2307 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2309 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2310 entry
= huge_ptep_get(src_pte
);
2311 ptepage
= pte_page(entry
);
2313 page_dup_rmap(ptepage
);
2314 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2316 spin_unlock(&src
->page_table_lock
);
2317 spin_unlock(&dst
->page_table_lock
);
2325 static int is_hugetlb_entry_migration(pte_t pte
)
2329 if (huge_pte_none(pte
) || pte_present(pte
))
2331 swp
= pte_to_swp_entry(pte
);
2332 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2338 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2342 if (huge_pte_none(pte
) || pte_present(pte
))
2344 swp
= pte_to_swp_entry(pte
);
2345 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2351 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2352 unsigned long start
, unsigned long end
,
2353 struct page
*ref_page
)
2355 int force_flush
= 0;
2356 struct mm_struct
*mm
= vma
->vm_mm
;
2357 unsigned long address
;
2361 struct hstate
*h
= hstate_vma(vma
);
2362 unsigned long sz
= huge_page_size(h
);
2363 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2364 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2366 WARN_ON(!is_vm_hugetlb_page(vma
));
2367 BUG_ON(start
& ~huge_page_mask(h
));
2368 BUG_ON(end
& ~huge_page_mask(h
));
2370 tlb_start_vma(tlb
, vma
);
2371 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2373 spin_lock(&mm
->page_table_lock
);
2374 for (address
= start
; address
< end
; address
+= sz
) {
2375 ptep
= huge_pte_offset(mm
, address
);
2379 if (huge_pmd_unshare(mm
, &address
, ptep
))
2382 pte
= huge_ptep_get(ptep
);
2383 if (huge_pte_none(pte
))
2387 * HWPoisoned hugepage is already unmapped and dropped reference
2389 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
)))
2392 page
= pte_page(pte
);
2394 * If a reference page is supplied, it is because a specific
2395 * page is being unmapped, not a range. Ensure the page we
2396 * are about to unmap is the actual page of interest.
2399 if (page
!= ref_page
)
2403 * Mark the VMA as having unmapped its page so that
2404 * future faults in this VMA will fail rather than
2405 * looking like data was lost
2407 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2410 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2411 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2413 set_page_dirty(page
);
2415 page_remove_rmap(page
);
2416 force_flush
= !__tlb_remove_page(tlb
, page
);
2419 /* Bail out after unmapping reference page if supplied */
2423 spin_unlock(&mm
->page_table_lock
);
2425 * mmu_gather ran out of room to batch pages, we break out of
2426 * the PTE lock to avoid doing the potential expensive TLB invalidate
2427 * and page-free while holding it.
2432 if (address
< end
&& !ref_page
)
2435 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2436 tlb_end_vma(tlb
, vma
);
2439 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2440 struct vm_area_struct
*vma
, unsigned long start
,
2441 unsigned long end
, struct page
*ref_page
)
2443 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2446 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2447 * test will fail on a vma being torn down, and not grab a page table
2448 * on its way out. We're lucky that the flag has such an appropriate
2449 * name, and can in fact be safely cleared here. We could clear it
2450 * before the __unmap_hugepage_range above, but all that's necessary
2451 * is to clear it before releasing the i_mmap_mutex. This works
2452 * because in the context this is called, the VMA is about to be
2453 * destroyed and the i_mmap_mutex is held.
2455 vma
->vm_flags
&= ~VM_MAYSHARE
;
2458 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2459 unsigned long end
, struct page
*ref_page
)
2461 struct mm_struct
*mm
;
2462 struct mmu_gather tlb
;
2466 tlb_gather_mmu(&tlb
, mm
, 0);
2467 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2468 tlb_finish_mmu(&tlb
, start
, end
);
2472 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2473 * mappping it owns the reserve page for. The intention is to unmap the page
2474 * from other VMAs and let the children be SIGKILLed if they are faulting the
2477 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2478 struct page
*page
, unsigned long address
)
2480 struct hstate
*h
= hstate_vma(vma
);
2481 struct vm_area_struct
*iter_vma
;
2482 struct address_space
*mapping
;
2486 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2487 * from page cache lookup which is in HPAGE_SIZE units.
2489 address
= address
& huge_page_mask(h
);
2490 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2492 mapping
= vma
->vm_file
->f_dentry
->d_inode
->i_mapping
;
2495 * Take the mapping lock for the duration of the table walk. As
2496 * this mapping should be shared between all the VMAs,
2497 * __unmap_hugepage_range() is called as the lock is already held
2499 mutex_lock(&mapping
->i_mmap_mutex
);
2500 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2501 /* Do not unmap the current VMA */
2502 if (iter_vma
== vma
)
2506 * Unmap the page from other VMAs without their own reserves.
2507 * They get marked to be SIGKILLed if they fault in these
2508 * areas. This is because a future no-page fault on this VMA
2509 * could insert a zeroed page instead of the data existing
2510 * from the time of fork. This would look like data corruption
2512 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2513 unmap_hugepage_range(iter_vma
, address
,
2514 address
+ huge_page_size(h
), page
);
2516 mutex_unlock(&mapping
->i_mmap_mutex
);
2522 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2523 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2524 * cannot race with other handlers or page migration.
2525 * Keep the pte_same checks anyway to make transition from the mutex easier.
2527 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2528 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2529 struct page
*pagecache_page
)
2531 struct hstate
*h
= hstate_vma(vma
);
2532 struct page
*old_page
, *new_page
;
2534 int outside_reserve
= 0;
2535 unsigned long mmun_start
; /* For mmu_notifiers */
2536 unsigned long mmun_end
; /* For mmu_notifiers */
2538 old_page
= pte_page(pte
);
2541 /* If no-one else is actually using this page, avoid the copy
2542 * and just make the page writable */
2543 avoidcopy
= (page_mapcount(old_page
) == 1);
2545 if (PageAnon(old_page
))
2546 page_move_anon_rmap(old_page
, vma
, address
);
2547 set_huge_ptep_writable(vma
, address
, ptep
);
2552 * If the process that created a MAP_PRIVATE mapping is about to
2553 * perform a COW due to a shared page count, attempt to satisfy
2554 * the allocation without using the existing reserves. The pagecache
2555 * page is used to determine if the reserve at this address was
2556 * consumed or not. If reserves were used, a partial faulted mapping
2557 * at the time of fork() could consume its reserves on COW instead
2558 * of the full address range.
2560 if (!(vma
->vm_flags
& VM_MAYSHARE
) &&
2561 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2562 old_page
!= pagecache_page
)
2563 outside_reserve
= 1;
2565 page_cache_get(old_page
);
2567 /* Drop page_table_lock as buddy allocator may be called */
2568 spin_unlock(&mm
->page_table_lock
);
2569 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2571 if (IS_ERR(new_page
)) {
2572 long err
= PTR_ERR(new_page
);
2573 page_cache_release(old_page
);
2576 * If a process owning a MAP_PRIVATE mapping fails to COW,
2577 * it is due to references held by a child and an insufficient
2578 * huge page pool. To guarantee the original mappers
2579 * reliability, unmap the page from child processes. The child
2580 * may get SIGKILLed if it later faults.
2582 if (outside_reserve
) {
2583 BUG_ON(huge_pte_none(pte
));
2584 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2585 BUG_ON(huge_pte_none(pte
));
2586 spin_lock(&mm
->page_table_lock
);
2587 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2588 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2589 goto retry_avoidcopy
;
2591 * race occurs while re-acquiring page_table_lock, and
2599 /* Caller expects lock to be held */
2600 spin_lock(&mm
->page_table_lock
);
2602 return VM_FAULT_OOM
;
2604 return VM_FAULT_SIGBUS
;
2608 * When the original hugepage is shared one, it does not have
2609 * anon_vma prepared.
2611 if (unlikely(anon_vma_prepare(vma
))) {
2612 page_cache_release(new_page
);
2613 page_cache_release(old_page
);
2614 /* Caller expects lock to be held */
2615 spin_lock(&mm
->page_table_lock
);
2616 return VM_FAULT_OOM
;
2619 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2620 pages_per_huge_page(h
));
2621 __SetPageUptodate(new_page
);
2623 mmun_start
= address
& huge_page_mask(h
);
2624 mmun_end
= mmun_start
+ huge_page_size(h
);
2625 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2627 * Retake the page_table_lock to check for racing updates
2628 * before the page tables are altered
2630 spin_lock(&mm
->page_table_lock
);
2631 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2632 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2634 huge_ptep_clear_flush(vma
, address
, ptep
);
2635 set_huge_pte_at(mm
, address
, ptep
,
2636 make_huge_pte(vma
, new_page
, 1));
2637 page_remove_rmap(old_page
);
2638 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2639 /* Make the old page be freed below */
2640 new_page
= old_page
;
2642 spin_unlock(&mm
->page_table_lock
);
2643 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2644 /* Caller expects lock to be held */
2645 spin_lock(&mm
->page_table_lock
);
2646 page_cache_release(new_page
);
2647 page_cache_release(old_page
);
2651 /* Return the pagecache page at a given address within a VMA */
2652 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2653 struct vm_area_struct
*vma
, unsigned long address
)
2655 struct address_space
*mapping
;
2658 mapping
= vma
->vm_file
->f_mapping
;
2659 idx
= vma_hugecache_offset(h
, vma
, address
);
2661 return find_lock_page(mapping
, idx
);
2665 * Return whether there is a pagecache page to back given address within VMA.
2666 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2668 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2669 struct vm_area_struct
*vma
, unsigned long address
)
2671 struct address_space
*mapping
;
2675 mapping
= vma
->vm_file
->f_mapping
;
2676 idx
= vma_hugecache_offset(h
, vma
, address
);
2678 page
= find_get_page(mapping
, idx
);
2681 return page
!= NULL
;
2684 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2685 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2687 struct hstate
*h
= hstate_vma(vma
);
2688 int ret
= VM_FAULT_SIGBUS
;
2693 struct address_space
*mapping
;
2697 * Currently, we are forced to kill the process in the event the
2698 * original mapper has unmapped pages from the child due to a failed
2699 * COW. Warn that such a situation has occurred as it may not be obvious
2701 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2703 "PID %d killed due to inadequate hugepage pool\n",
2708 mapping
= vma
->vm_file
->f_mapping
;
2709 idx
= vma_hugecache_offset(h
, vma
, address
);
2712 * Use page lock to guard against racing truncation
2713 * before we get page_table_lock.
2716 page
= find_lock_page(mapping
, idx
);
2718 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2721 page
= alloc_huge_page(vma
, address
, 0);
2723 ret
= PTR_ERR(page
);
2727 ret
= VM_FAULT_SIGBUS
;
2730 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2731 __SetPageUptodate(page
);
2733 if (vma
->vm_flags
& VM_MAYSHARE
) {
2735 struct inode
*inode
= mapping
->host
;
2737 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2745 spin_lock(&inode
->i_lock
);
2746 inode
->i_blocks
+= blocks_per_huge_page(h
);
2747 spin_unlock(&inode
->i_lock
);
2750 if (unlikely(anon_vma_prepare(vma
))) {
2752 goto backout_unlocked
;
2758 * If memory error occurs between mmap() and fault, some process
2759 * don't have hwpoisoned swap entry for errored virtual address.
2760 * So we need to block hugepage fault by PG_hwpoison bit check.
2762 if (unlikely(PageHWPoison(page
))) {
2763 ret
= VM_FAULT_HWPOISON
|
2764 VM_FAULT_SET_HINDEX(hstate_index(h
));
2765 goto backout_unlocked
;
2770 * If we are going to COW a private mapping later, we examine the
2771 * pending reservations for this page now. This will ensure that
2772 * any allocations necessary to record that reservation occur outside
2775 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2776 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2778 goto backout_unlocked
;
2781 spin_lock(&mm
->page_table_lock
);
2782 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2787 if (!huge_pte_none(huge_ptep_get(ptep
)))
2791 hugepage_add_new_anon_rmap(page
, vma
, address
);
2793 page_dup_rmap(page
);
2794 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2795 && (vma
->vm_flags
& VM_SHARED
)));
2796 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2798 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2799 /* Optimization, do the COW without a second fault */
2800 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2803 spin_unlock(&mm
->page_table_lock
);
2809 spin_unlock(&mm
->page_table_lock
);
2816 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2817 unsigned long address
, unsigned int flags
)
2822 struct page
*page
= NULL
;
2823 struct page
*pagecache_page
= NULL
;
2824 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2825 struct hstate
*h
= hstate_vma(vma
);
2827 address
&= huge_page_mask(h
);
2829 ptep
= huge_pte_offset(mm
, address
);
2831 entry
= huge_ptep_get(ptep
);
2832 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2833 migration_entry_wait(mm
, (pmd_t
*)ptep
, address
);
2835 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2836 return VM_FAULT_HWPOISON_LARGE
|
2837 VM_FAULT_SET_HINDEX(hstate_index(h
));
2840 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2842 return VM_FAULT_OOM
;
2845 * Serialize hugepage allocation and instantiation, so that we don't
2846 * get spurious allocation failures if two CPUs race to instantiate
2847 * the same page in the page cache.
2849 mutex_lock(&hugetlb_instantiation_mutex
);
2850 entry
= huge_ptep_get(ptep
);
2851 if (huge_pte_none(entry
)) {
2852 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2859 * If we are going to COW the mapping later, we examine the pending
2860 * reservations for this page now. This will ensure that any
2861 * allocations necessary to record that reservation occur outside the
2862 * spinlock. For private mappings, we also lookup the pagecache
2863 * page now as it is used to determine if a reservation has been
2866 if ((flags
& FAULT_FLAG_WRITE
) && !pte_write(entry
)) {
2867 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2872 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2873 pagecache_page
= hugetlbfs_pagecache_page(h
,
2878 * hugetlb_cow() requires page locks of pte_page(entry) and
2879 * pagecache_page, so here we need take the former one
2880 * when page != pagecache_page or !pagecache_page.
2881 * Note that locking order is always pagecache_page -> page,
2882 * so no worry about deadlock.
2884 page
= pte_page(entry
);
2886 if (page
!= pagecache_page
)
2889 spin_lock(&mm
->page_table_lock
);
2890 /* Check for a racing update before calling hugetlb_cow */
2891 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2892 goto out_page_table_lock
;
2895 if (flags
& FAULT_FLAG_WRITE
) {
2896 if (!pte_write(entry
)) {
2897 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2899 goto out_page_table_lock
;
2901 entry
= pte_mkdirty(entry
);
2903 entry
= pte_mkyoung(entry
);
2904 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2905 flags
& FAULT_FLAG_WRITE
))
2906 update_mmu_cache(vma
, address
, ptep
);
2908 out_page_table_lock
:
2909 spin_unlock(&mm
->page_table_lock
);
2911 if (pagecache_page
) {
2912 unlock_page(pagecache_page
);
2913 put_page(pagecache_page
);
2915 if (page
!= pagecache_page
)
2920 mutex_unlock(&hugetlb_instantiation_mutex
);
2925 /* Can be overriden by architectures */
2926 __attribute__((weak
)) struct page
*
2927 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
2928 pud_t
*pud
, int write
)
2934 int follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2935 struct page
**pages
, struct vm_area_struct
**vmas
,
2936 unsigned long *position
, int *length
, int i
,
2939 unsigned long pfn_offset
;
2940 unsigned long vaddr
= *position
;
2941 int remainder
= *length
;
2942 struct hstate
*h
= hstate_vma(vma
);
2944 spin_lock(&mm
->page_table_lock
);
2945 while (vaddr
< vma
->vm_end
&& remainder
) {
2951 * Some archs (sparc64, sh*) have multiple pte_ts to
2952 * each hugepage. We have to make sure we get the
2953 * first, for the page indexing below to work.
2955 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2956 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
2959 * When coredumping, it suits get_dump_page if we just return
2960 * an error where there's an empty slot with no huge pagecache
2961 * to back it. This way, we avoid allocating a hugepage, and
2962 * the sparse dumpfile avoids allocating disk blocks, but its
2963 * huge holes still show up with zeroes where they need to be.
2965 if (absent
&& (flags
& FOLL_DUMP
) &&
2966 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
2972 ((flags
& FOLL_WRITE
) && !pte_write(huge_ptep_get(pte
)))) {
2975 spin_unlock(&mm
->page_table_lock
);
2976 ret
= hugetlb_fault(mm
, vma
, vaddr
,
2977 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
2978 spin_lock(&mm
->page_table_lock
);
2979 if (!(ret
& VM_FAULT_ERROR
))
2986 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2987 page
= pte_page(huge_ptep_get(pte
));
2990 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3001 if (vaddr
< vma
->vm_end
&& remainder
&&
3002 pfn_offset
< pages_per_huge_page(h
)) {
3004 * We use pfn_offset to avoid touching the pageframes
3005 * of this compound page.
3010 spin_unlock(&mm
->page_table_lock
);
3011 *length
= remainder
;
3014 return i
? i
: -EFAULT
;
3017 void hugetlb_change_protection(struct vm_area_struct
*vma
,
3018 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3020 struct mm_struct
*mm
= vma
->vm_mm
;
3021 unsigned long start
= address
;
3024 struct hstate
*h
= hstate_vma(vma
);
3026 BUG_ON(address
>= end
);
3027 flush_cache_range(vma
, address
, end
);
3029 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3030 spin_lock(&mm
->page_table_lock
);
3031 for (; address
< end
; address
+= huge_page_size(h
)) {
3032 ptep
= huge_pte_offset(mm
, address
);
3035 if (huge_pmd_unshare(mm
, &address
, ptep
))
3037 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3038 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3039 pte
= pte_mkhuge(pte_modify(pte
, newprot
));
3040 set_huge_pte_at(mm
, address
, ptep
, pte
);
3043 spin_unlock(&mm
->page_table_lock
);
3045 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3046 * may have cleared our pud entry and done put_page on the page table:
3047 * once we release i_mmap_mutex, another task can do the final put_page
3048 * and that page table be reused and filled with junk.
3050 flush_tlb_range(vma
, start
, end
);
3051 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3054 int hugetlb_reserve_pages(struct inode
*inode
,
3056 struct vm_area_struct
*vma
,
3057 vm_flags_t vm_flags
)
3060 struct hstate
*h
= hstate_inode(inode
);
3061 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3064 * Only apply hugepage reservation if asked. At fault time, an
3065 * attempt will be made for VM_NORESERVE to allocate a page
3066 * without using reserves
3068 if (vm_flags
& VM_NORESERVE
)
3072 * Shared mappings base their reservation on the number of pages that
3073 * are already allocated on behalf of the file. Private mappings need
3074 * to reserve the full area even if read-only as mprotect() may be
3075 * called to make the mapping read-write. Assume !vma is a shm mapping
3077 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3078 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3080 struct resv_map
*resv_map
= resv_map_alloc();
3086 set_vma_resv_map(vma
, resv_map
);
3087 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3095 /* There must be enough pages in the subpool for the mapping */
3096 if (hugepage_subpool_get_pages(spool
, chg
)) {
3102 * Check enough hugepages are available for the reservation.
3103 * Hand the pages back to the subpool if there are not
3105 ret
= hugetlb_acct_memory(h
, chg
);
3107 hugepage_subpool_put_pages(spool
, chg
);
3112 * Account for the reservations made. Shared mappings record regions
3113 * that have reservations as they are shared by multiple VMAs.
3114 * When the last VMA disappears, the region map says how much
3115 * the reservation was and the page cache tells how much of
3116 * the reservation was consumed. Private mappings are per-VMA and
3117 * only the consumed reservations are tracked. When the VMA
3118 * disappears, the original reservation is the VMA size and the
3119 * consumed reservations are stored in the map. Hence, nothing
3120 * else has to be done for private mappings here
3122 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3123 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3131 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3133 struct hstate
*h
= hstate_inode(inode
);
3134 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3135 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3137 spin_lock(&inode
->i_lock
);
3138 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3139 spin_unlock(&inode
->i_lock
);
3141 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3142 hugetlb_acct_memory(h
, -(chg
- freed
));
3145 #ifdef CONFIG_MEMORY_FAILURE
3147 /* Should be called in hugetlb_lock */
3148 static int is_hugepage_on_freelist(struct page
*hpage
)
3152 struct hstate
*h
= page_hstate(hpage
);
3153 int nid
= page_to_nid(hpage
);
3155 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3162 * This function is called from memory failure code.
3163 * Assume the caller holds page lock of the head page.
3165 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3167 struct hstate
*h
= page_hstate(hpage
);
3168 int nid
= page_to_nid(hpage
);
3171 spin_lock(&hugetlb_lock
);
3172 if (is_hugepage_on_freelist(hpage
)) {
3173 list_del(&hpage
->lru
);
3174 set_page_refcounted(hpage
);
3175 h
->free_huge_pages
--;
3176 h
->free_huge_pages_node
[nid
]--;
3179 spin_unlock(&hugetlb_lock
);