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 /* 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 * A child process with MAP_PRIVATE mappings created by their parent
544 * have no page reserves. This check ensures that reservations are
545 * not "stolen". The child may still get SIGKILLed
547 if (!vma_has_reserves(vma
) &&
548 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
551 /* If reserves cannot be used, ensure enough pages are in the pool */
552 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
556 cpuset_mems_cookie
= get_mems_allowed();
557 zonelist
= huge_zonelist(vma
, address
,
558 htlb_alloc_mask
, &mpol
, &nodemask
);
560 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
561 MAX_NR_ZONES
- 1, nodemask
) {
562 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
563 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
566 decrement_hugepage_resv_vma(h
, vma
);
573 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 pgoff_t
__basepage_index(struct page
*page
)
695 struct page
*page_head
= compound_head(page
);
696 pgoff_t index
= page_index(page_head
);
697 unsigned long compound_idx
;
699 if (!PageHuge(page_head
))
700 return page_index(page
);
702 if (compound_order(page_head
) >= MAX_ORDER
)
703 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
705 compound_idx
= page
- page_head
;
707 return (index
<< compound_order(page_head
)) + compound_idx
;
710 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
714 if (h
->order
>= MAX_ORDER
)
717 page
= alloc_pages_exact_node(nid
,
718 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
719 __GFP_REPEAT
|__GFP_NOWARN
,
722 if (arch_prepare_hugepage(page
)) {
723 __free_pages(page
, huge_page_order(h
));
726 prep_new_huge_page(h
, page
, nid
);
733 * common helper functions for hstate_next_node_to_{alloc|free}.
734 * We may have allocated or freed a huge page based on a different
735 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
736 * be outside of *nodes_allowed. Ensure that we use an allowed
737 * node for alloc or free.
739 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
741 nid
= next_node(nid
, *nodes_allowed
);
742 if (nid
== MAX_NUMNODES
)
743 nid
= first_node(*nodes_allowed
);
744 VM_BUG_ON(nid
>= MAX_NUMNODES
);
749 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
751 if (!node_isset(nid
, *nodes_allowed
))
752 nid
= next_node_allowed(nid
, nodes_allowed
);
757 * returns the previously saved node ["this node"] from which to
758 * allocate a persistent huge page for the pool and advance the
759 * next node from which to allocate, handling wrap at end of node
762 static int hstate_next_node_to_alloc(struct hstate
*h
,
763 nodemask_t
*nodes_allowed
)
767 VM_BUG_ON(!nodes_allowed
);
769 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
770 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
776 * helper for free_pool_huge_page() - return the previously saved
777 * node ["this node"] from which to free a huge page. Advance the
778 * next node id whether or not we find a free huge page to free so
779 * that the next attempt to free addresses the next node.
781 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
785 VM_BUG_ON(!nodes_allowed
);
787 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
788 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
793 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
794 for (nr_nodes = nodes_weight(*mask); \
796 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
799 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
800 for (nr_nodes = nodes_weight(*mask); \
802 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
805 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
811 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
812 page
= alloc_fresh_huge_page_node(h
, node
);
820 count_vm_event(HTLB_BUDDY_PGALLOC
);
822 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
828 * Free huge page from pool from next node to free.
829 * Attempt to keep persistent huge pages more or less
830 * balanced over allowed nodes.
831 * Called with hugetlb_lock locked.
833 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
839 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
841 * If we're returning unused surplus pages, only examine
842 * nodes with surplus pages.
844 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
845 !list_empty(&h
->hugepage_freelists
[node
])) {
847 list_entry(h
->hugepage_freelists
[node
].next
,
849 list_del(&page
->lru
);
850 h
->free_huge_pages
--;
851 h
->free_huge_pages_node
[node
]--;
853 h
->surplus_huge_pages
--;
854 h
->surplus_huge_pages_node
[node
]--;
856 update_and_free_page(h
, page
);
865 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
870 if (h
->order
>= MAX_ORDER
)
874 * Assume we will successfully allocate the surplus page to
875 * prevent racing processes from causing the surplus to exceed
878 * This however introduces a different race, where a process B
879 * tries to grow the static hugepage pool while alloc_pages() is
880 * called by process A. B will only examine the per-node
881 * counters in determining if surplus huge pages can be
882 * converted to normal huge pages in adjust_pool_surplus(). A
883 * won't be able to increment the per-node counter, until the
884 * lock is dropped by B, but B doesn't drop hugetlb_lock until
885 * no more huge pages can be converted from surplus to normal
886 * state (and doesn't try to convert again). Thus, we have a
887 * case where a surplus huge page exists, the pool is grown, and
888 * the surplus huge page still exists after, even though it
889 * should just have been converted to a normal huge page. This
890 * does not leak memory, though, as the hugepage will be freed
891 * once it is out of use. It also does not allow the counters to
892 * go out of whack in adjust_pool_surplus() as we don't modify
893 * the node values until we've gotten the hugepage and only the
894 * per-node value is checked there.
896 spin_lock(&hugetlb_lock
);
897 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
898 spin_unlock(&hugetlb_lock
);
902 h
->surplus_huge_pages
++;
904 spin_unlock(&hugetlb_lock
);
906 if (nid
== NUMA_NO_NODE
)
907 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
908 __GFP_REPEAT
|__GFP_NOWARN
,
911 page
= alloc_pages_exact_node(nid
,
912 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
913 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
915 if (page
&& arch_prepare_hugepage(page
)) {
916 __free_pages(page
, huge_page_order(h
));
920 spin_lock(&hugetlb_lock
);
922 INIT_LIST_HEAD(&page
->lru
);
923 r_nid
= page_to_nid(page
);
924 set_compound_page_dtor(page
, free_huge_page
);
925 set_hugetlb_cgroup(page
, NULL
);
927 * We incremented the global counters already
929 h
->nr_huge_pages_node
[r_nid
]++;
930 h
->surplus_huge_pages_node
[r_nid
]++;
931 __count_vm_event(HTLB_BUDDY_PGALLOC
);
934 h
->surplus_huge_pages
--;
935 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
937 spin_unlock(&hugetlb_lock
);
943 * This allocation function is useful in the context where vma is irrelevant.
944 * E.g. soft-offlining uses this function because it only cares physical
945 * address of error page.
947 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
951 spin_lock(&hugetlb_lock
);
952 page
= dequeue_huge_page_node(h
, nid
);
953 spin_unlock(&hugetlb_lock
);
956 page
= alloc_buddy_huge_page(h
, nid
);
962 * Increase the hugetlb pool such that it can accommodate a reservation
965 static int gather_surplus_pages(struct hstate
*h
, int delta
)
967 struct list_head surplus_list
;
968 struct page
*page
, *tmp
;
970 int needed
, allocated
;
971 bool alloc_ok
= true;
973 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
975 h
->resv_huge_pages
+= delta
;
980 INIT_LIST_HEAD(&surplus_list
);
984 spin_unlock(&hugetlb_lock
);
985 for (i
= 0; i
< needed
; i
++) {
986 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
991 list_add(&page
->lru
, &surplus_list
);
996 * After retaking hugetlb_lock, we need to recalculate 'needed'
997 * because either resv_huge_pages or free_huge_pages may have changed.
999 spin_lock(&hugetlb_lock
);
1000 needed
= (h
->resv_huge_pages
+ delta
) -
1001 (h
->free_huge_pages
+ allocated
);
1006 * We were not able to allocate enough pages to
1007 * satisfy the entire reservation so we free what
1008 * we've allocated so far.
1013 * The surplus_list now contains _at_least_ the number of extra pages
1014 * needed to accommodate the reservation. Add the appropriate number
1015 * of pages to the hugetlb pool and free the extras back to the buddy
1016 * allocator. Commit the entire reservation here to prevent another
1017 * process from stealing the pages as they are added to the pool but
1018 * before they are reserved.
1020 needed
+= allocated
;
1021 h
->resv_huge_pages
+= delta
;
1024 /* Free the needed pages to the hugetlb pool */
1025 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1029 * This page is now managed by the hugetlb allocator and has
1030 * no users -- drop the buddy allocator's reference.
1032 put_page_testzero(page
);
1033 VM_BUG_ON(page_count(page
));
1034 enqueue_huge_page(h
, page
);
1037 spin_unlock(&hugetlb_lock
);
1039 /* Free unnecessary surplus pages to the buddy allocator */
1040 if (!list_empty(&surplus_list
)) {
1041 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1045 spin_lock(&hugetlb_lock
);
1051 * When releasing a hugetlb pool reservation, any surplus pages that were
1052 * allocated to satisfy the reservation must be explicitly freed if they were
1054 * Called with hugetlb_lock held.
1056 static void return_unused_surplus_pages(struct hstate
*h
,
1057 unsigned long unused_resv_pages
)
1059 unsigned long nr_pages
;
1061 /* Uncommit the reservation */
1062 h
->resv_huge_pages
-= unused_resv_pages
;
1064 /* Cannot return gigantic pages currently */
1065 if (h
->order
>= MAX_ORDER
)
1068 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1071 * We want to release as many surplus pages as possible, spread
1072 * evenly across all nodes with memory. Iterate across these nodes
1073 * until we can no longer free unreserved surplus pages. This occurs
1074 * when the nodes with surplus pages have no free pages.
1075 * free_pool_huge_page() will balance the the freed pages across the
1076 * on-line nodes with memory and will handle the hstate accounting.
1078 while (nr_pages
--) {
1079 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1085 * Determine if the huge page at addr within the vma has an associated
1086 * reservation. Where it does not we will need to logically increase
1087 * reservation and actually increase subpool usage before an allocation
1088 * can occur. Where any new reservation would be required the
1089 * reservation change is prepared, but not committed. Once the page
1090 * has been allocated from the subpool and instantiated the change should
1091 * be committed via vma_commit_reservation. No action is required on
1094 static long vma_needs_reservation(struct hstate
*h
,
1095 struct vm_area_struct
*vma
, unsigned long addr
)
1097 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1098 struct inode
*inode
= mapping
->host
;
1100 if (vma
->vm_flags
& VM_MAYSHARE
) {
1101 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1102 return region_chg(&inode
->i_mapping
->private_list
,
1105 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1110 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1111 struct resv_map
*reservations
= vma_resv_map(vma
);
1113 err
= region_chg(&reservations
->regions
, idx
, idx
+ 1);
1119 static void vma_commit_reservation(struct hstate
*h
,
1120 struct vm_area_struct
*vma
, unsigned long addr
)
1122 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1123 struct inode
*inode
= mapping
->host
;
1125 if (vma
->vm_flags
& VM_MAYSHARE
) {
1126 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1127 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1129 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1130 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1131 struct resv_map
*reservations
= vma_resv_map(vma
);
1133 /* Mark this page used in the map. */
1134 region_add(&reservations
->regions
, idx
, idx
+ 1);
1138 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1139 unsigned long addr
, int avoid_reserve
)
1141 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1142 struct hstate
*h
= hstate_vma(vma
);
1146 struct hugetlb_cgroup
*h_cg
;
1148 idx
= hstate_index(h
);
1150 * Processes that did not create the mapping will have no
1151 * reserves and will not have accounted against subpool
1152 * limit. Check that the subpool limit can be made before
1153 * satisfying the allocation MAP_NORESERVE mappings may also
1154 * need pages and subpool limit allocated allocated if no reserve
1157 chg
= vma_needs_reservation(h
, vma
, addr
);
1159 return ERR_PTR(-ENOMEM
);
1161 if (hugepage_subpool_get_pages(spool
, chg
))
1162 return ERR_PTR(-ENOSPC
);
1164 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1166 hugepage_subpool_put_pages(spool
, chg
);
1167 return ERR_PTR(-ENOSPC
);
1169 spin_lock(&hugetlb_lock
);
1170 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
);
1172 spin_unlock(&hugetlb_lock
);
1173 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1175 hugetlb_cgroup_uncharge_cgroup(idx
,
1176 pages_per_huge_page(h
),
1178 hugepage_subpool_put_pages(spool
, chg
);
1179 return ERR_PTR(-ENOSPC
);
1181 spin_lock(&hugetlb_lock
);
1182 list_move(&page
->lru
, &h
->hugepage_activelist
);
1185 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1186 spin_unlock(&hugetlb_lock
);
1188 set_page_private(page
, (unsigned long)spool
);
1190 vma_commit_reservation(h
, vma
, addr
);
1194 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1196 struct huge_bootmem_page
*m
;
1199 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1202 addr
= __alloc_bootmem_node_nopanic(NODE_DATA(node
),
1203 huge_page_size(h
), huge_page_size(h
), 0);
1207 * Use the beginning of the huge page to store the
1208 * huge_bootmem_page struct (until gather_bootmem
1209 * puts them into the mem_map).
1218 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1219 /* Put them into a private list first because mem_map is not up yet */
1220 list_add(&m
->list
, &huge_boot_pages
);
1225 static void prep_compound_huge_page(struct page
*page
, int order
)
1227 if (unlikely(order
> (MAX_ORDER
- 1)))
1228 prep_compound_gigantic_page(page
, order
);
1230 prep_compound_page(page
, order
);
1233 /* Put bootmem huge pages into the standard lists after mem_map is up */
1234 static void __init
gather_bootmem_prealloc(void)
1236 struct huge_bootmem_page
*m
;
1238 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1239 struct hstate
*h
= m
->hstate
;
1242 #ifdef CONFIG_HIGHMEM
1243 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1244 free_bootmem_late((unsigned long)m
,
1245 sizeof(struct huge_bootmem_page
));
1247 page
= virt_to_page(m
);
1249 __ClearPageReserved(page
);
1250 WARN_ON(page_count(page
) != 1);
1251 prep_compound_huge_page(page
, h
->order
);
1252 prep_new_huge_page(h
, page
, page_to_nid(page
));
1254 * If we had gigantic hugepages allocated at boot time, we need
1255 * to restore the 'stolen' pages to totalram_pages in order to
1256 * fix confusing memory reports from free(1) and another
1257 * side-effects, like CommitLimit going negative.
1259 if (h
->order
> (MAX_ORDER
- 1))
1260 adjust_managed_page_count(page
, 1 << h
->order
);
1264 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1268 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1269 if (h
->order
>= MAX_ORDER
) {
1270 if (!alloc_bootmem_huge_page(h
))
1272 } else if (!alloc_fresh_huge_page(h
,
1273 &node_states
[N_MEMORY
]))
1276 h
->max_huge_pages
= i
;
1279 static void __init
hugetlb_init_hstates(void)
1283 for_each_hstate(h
) {
1284 /* oversize hugepages were init'ed in early boot */
1285 if (h
->order
< MAX_ORDER
)
1286 hugetlb_hstate_alloc_pages(h
);
1290 static char * __init
memfmt(char *buf
, unsigned long n
)
1292 if (n
>= (1UL << 30))
1293 sprintf(buf
, "%lu GB", n
>> 30);
1294 else if (n
>= (1UL << 20))
1295 sprintf(buf
, "%lu MB", n
>> 20);
1297 sprintf(buf
, "%lu KB", n
>> 10);
1301 static void __init
report_hugepages(void)
1305 for_each_hstate(h
) {
1307 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1308 memfmt(buf
, huge_page_size(h
)),
1309 h
->free_huge_pages
);
1313 #ifdef CONFIG_HIGHMEM
1314 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1315 nodemask_t
*nodes_allowed
)
1319 if (h
->order
>= MAX_ORDER
)
1322 for_each_node_mask(i
, *nodes_allowed
) {
1323 struct page
*page
, *next
;
1324 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1325 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1326 if (count
>= h
->nr_huge_pages
)
1328 if (PageHighMem(page
))
1330 list_del(&page
->lru
);
1331 update_and_free_page(h
, page
);
1332 h
->free_huge_pages
--;
1333 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1338 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1339 nodemask_t
*nodes_allowed
)
1345 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1346 * balanced by operating on them in a round-robin fashion.
1347 * Returns 1 if an adjustment was made.
1349 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1354 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1357 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1358 if (h
->surplus_huge_pages_node
[node
])
1362 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1363 if (h
->surplus_huge_pages_node
[node
] <
1364 h
->nr_huge_pages_node
[node
])
1371 h
->surplus_huge_pages
+= delta
;
1372 h
->surplus_huge_pages_node
[node
] += delta
;
1376 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1377 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1378 nodemask_t
*nodes_allowed
)
1380 unsigned long min_count
, ret
;
1382 if (h
->order
>= MAX_ORDER
)
1383 return h
->max_huge_pages
;
1386 * Increase the pool size
1387 * First take pages out of surplus state. Then make up the
1388 * remaining difference by allocating fresh huge pages.
1390 * We might race with alloc_buddy_huge_page() here and be unable
1391 * to convert a surplus huge page to a normal huge page. That is
1392 * not critical, though, it just means the overall size of the
1393 * pool might be one hugepage larger than it needs to be, but
1394 * within all the constraints specified by the sysctls.
1396 spin_lock(&hugetlb_lock
);
1397 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1398 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1402 while (count
> persistent_huge_pages(h
)) {
1404 * If this allocation races such that we no longer need the
1405 * page, free_huge_page will handle it by freeing the page
1406 * and reducing the surplus.
1408 spin_unlock(&hugetlb_lock
);
1409 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1410 spin_lock(&hugetlb_lock
);
1414 /* Bail for signals. Probably ctrl-c from user */
1415 if (signal_pending(current
))
1420 * Decrease the pool size
1421 * First return free pages to the buddy allocator (being careful
1422 * to keep enough around to satisfy reservations). Then place
1423 * pages into surplus state as needed so the pool will shrink
1424 * to the desired size as pages become free.
1426 * By placing pages into the surplus state independent of the
1427 * overcommit value, we are allowing the surplus pool size to
1428 * exceed overcommit. There are few sane options here. Since
1429 * alloc_buddy_huge_page() is checking the global counter,
1430 * though, we'll note that we're not allowed to exceed surplus
1431 * and won't grow the pool anywhere else. Not until one of the
1432 * sysctls are changed, or the surplus pages go out of use.
1434 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1435 min_count
= max(count
, min_count
);
1436 try_to_free_low(h
, min_count
, nodes_allowed
);
1437 while (min_count
< persistent_huge_pages(h
)) {
1438 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1441 while (count
< persistent_huge_pages(h
)) {
1442 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1446 ret
= persistent_huge_pages(h
);
1447 spin_unlock(&hugetlb_lock
);
1451 #define HSTATE_ATTR_RO(_name) \
1452 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1454 #define HSTATE_ATTR(_name) \
1455 static struct kobj_attribute _name##_attr = \
1456 __ATTR(_name, 0644, _name##_show, _name##_store)
1458 static struct kobject
*hugepages_kobj
;
1459 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1461 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1463 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1467 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1468 if (hstate_kobjs
[i
] == kobj
) {
1470 *nidp
= NUMA_NO_NODE
;
1474 return kobj_to_node_hstate(kobj
, nidp
);
1477 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1478 struct kobj_attribute
*attr
, char *buf
)
1481 unsigned long nr_huge_pages
;
1484 h
= kobj_to_hstate(kobj
, &nid
);
1485 if (nid
== NUMA_NO_NODE
)
1486 nr_huge_pages
= h
->nr_huge_pages
;
1488 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1490 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1493 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1494 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1495 const char *buf
, size_t len
)
1499 unsigned long count
;
1501 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1503 err
= kstrtoul(buf
, 10, &count
);
1507 h
= kobj_to_hstate(kobj
, &nid
);
1508 if (h
->order
>= MAX_ORDER
) {
1513 if (nid
== NUMA_NO_NODE
) {
1515 * global hstate attribute
1517 if (!(obey_mempolicy
&&
1518 init_nodemask_of_mempolicy(nodes_allowed
))) {
1519 NODEMASK_FREE(nodes_allowed
);
1520 nodes_allowed
= &node_states
[N_MEMORY
];
1522 } else if (nodes_allowed
) {
1524 * per node hstate attribute: adjust count to global,
1525 * but restrict alloc/free to the specified node.
1527 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1528 init_nodemask_of_node(nodes_allowed
, nid
);
1530 nodes_allowed
= &node_states
[N_MEMORY
];
1532 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1534 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1535 NODEMASK_FREE(nodes_allowed
);
1539 NODEMASK_FREE(nodes_allowed
);
1543 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1544 struct kobj_attribute
*attr
, char *buf
)
1546 return nr_hugepages_show_common(kobj
, attr
, buf
);
1549 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1550 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1552 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1554 HSTATE_ATTR(nr_hugepages
);
1559 * hstate attribute for optionally mempolicy-based constraint on persistent
1560 * huge page alloc/free.
1562 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1563 struct kobj_attribute
*attr
, char *buf
)
1565 return nr_hugepages_show_common(kobj
, attr
, buf
);
1568 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1569 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1571 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1573 HSTATE_ATTR(nr_hugepages_mempolicy
);
1577 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1578 struct kobj_attribute
*attr
, char *buf
)
1580 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1581 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1584 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1585 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1588 unsigned long input
;
1589 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1591 if (h
->order
>= MAX_ORDER
)
1594 err
= kstrtoul(buf
, 10, &input
);
1598 spin_lock(&hugetlb_lock
);
1599 h
->nr_overcommit_huge_pages
= input
;
1600 spin_unlock(&hugetlb_lock
);
1604 HSTATE_ATTR(nr_overcommit_hugepages
);
1606 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1607 struct kobj_attribute
*attr
, char *buf
)
1610 unsigned long free_huge_pages
;
1613 h
= kobj_to_hstate(kobj
, &nid
);
1614 if (nid
== NUMA_NO_NODE
)
1615 free_huge_pages
= h
->free_huge_pages
;
1617 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1619 return sprintf(buf
, "%lu\n", free_huge_pages
);
1621 HSTATE_ATTR_RO(free_hugepages
);
1623 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1624 struct kobj_attribute
*attr
, char *buf
)
1626 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1627 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1629 HSTATE_ATTR_RO(resv_hugepages
);
1631 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1632 struct kobj_attribute
*attr
, char *buf
)
1635 unsigned long surplus_huge_pages
;
1638 h
= kobj_to_hstate(kobj
, &nid
);
1639 if (nid
== NUMA_NO_NODE
)
1640 surplus_huge_pages
= h
->surplus_huge_pages
;
1642 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1644 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1646 HSTATE_ATTR_RO(surplus_hugepages
);
1648 static struct attribute
*hstate_attrs
[] = {
1649 &nr_hugepages_attr
.attr
,
1650 &nr_overcommit_hugepages_attr
.attr
,
1651 &free_hugepages_attr
.attr
,
1652 &resv_hugepages_attr
.attr
,
1653 &surplus_hugepages_attr
.attr
,
1655 &nr_hugepages_mempolicy_attr
.attr
,
1660 static struct attribute_group hstate_attr_group
= {
1661 .attrs
= hstate_attrs
,
1664 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1665 struct kobject
**hstate_kobjs
,
1666 struct attribute_group
*hstate_attr_group
)
1669 int hi
= hstate_index(h
);
1671 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1672 if (!hstate_kobjs
[hi
])
1675 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1677 kobject_put(hstate_kobjs
[hi
]);
1682 static void __init
hugetlb_sysfs_init(void)
1687 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1688 if (!hugepages_kobj
)
1691 for_each_hstate(h
) {
1692 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1693 hstate_kobjs
, &hstate_attr_group
);
1695 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1702 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1703 * with node devices in node_devices[] using a parallel array. The array
1704 * index of a node device or _hstate == node id.
1705 * This is here to avoid any static dependency of the node device driver, in
1706 * the base kernel, on the hugetlb module.
1708 struct node_hstate
{
1709 struct kobject
*hugepages_kobj
;
1710 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1712 struct node_hstate node_hstates
[MAX_NUMNODES
];
1715 * A subset of global hstate attributes for node devices
1717 static struct attribute
*per_node_hstate_attrs
[] = {
1718 &nr_hugepages_attr
.attr
,
1719 &free_hugepages_attr
.attr
,
1720 &surplus_hugepages_attr
.attr
,
1724 static struct attribute_group per_node_hstate_attr_group
= {
1725 .attrs
= per_node_hstate_attrs
,
1729 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1730 * Returns node id via non-NULL nidp.
1732 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1736 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1737 struct node_hstate
*nhs
= &node_hstates
[nid
];
1739 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1740 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1752 * Unregister hstate attributes from a single node device.
1753 * No-op if no hstate attributes attached.
1755 static void hugetlb_unregister_node(struct node
*node
)
1758 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1760 if (!nhs
->hugepages_kobj
)
1761 return; /* no hstate attributes */
1763 for_each_hstate(h
) {
1764 int idx
= hstate_index(h
);
1765 if (nhs
->hstate_kobjs
[idx
]) {
1766 kobject_put(nhs
->hstate_kobjs
[idx
]);
1767 nhs
->hstate_kobjs
[idx
] = NULL
;
1771 kobject_put(nhs
->hugepages_kobj
);
1772 nhs
->hugepages_kobj
= NULL
;
1776 * hugetlb module exit: unregister hstate attributes from node devices
1779 static void hugetlb_unregister_all_nodes(void)
1784 * disable node device registrations.
1786 register_hugetlbfs_with_node(NULL
, NULL
);
1789 * remove hstate attributes from any nodes that have them.
1791 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1792 hugetlb_unregister_node(node_devices
[nid
]);
1796 * Register hstate attributes for a single node device.
1797 * No-op if attributes already registered.
1799 static void hugetlb_register_node(struct node
*node
)
1802 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1805 if (nhs
->hugepages_kobj
)
1806 return; /* already allocated */
1808 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1810 if (!nhs
->hugepages_kobj
)
1813 for_each_hstate(h
) {
1814 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1816 &per_node_hstate_attr_group
);
1818 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1819 h
->name
, node
->dev
.id
);
1820 hugetlb_unregister_node(node
);
1827 * hugetlb init time: register hstate attributes for all registered node
1828 * devices of nodes that have memory. All on-line nodes should have
1829 * registered their associated device by this time.
1831 static void hugetlb_register_all_nodes(void)
1835 for_each_node_state(nid
, N_MEMORY
) {
1836 struct node
*node
= node_devices
[nid
];
1837 if (node
->dev
.id
== nid
)
1838 hugetlb_register_node(node
);
1842 * Let the node device driver know we're here so it can
1843 * [un]register hstate attributes on node hotplug.
1845 register_hugetlbfs_with_node(hugetlb_register_node
,
1846 hugetlb_unregister_node
);
1848 #else /* !CONFIG_NUMA */
1850 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1858 static void hugetlb_unregister_all_nodes(void) { }
1860 static void hugetlb_register_all_nodes(void) { }
1864 static void __exit
hugetlb_exit(void)
1868 hugetlb_unregister_all_nodes();
1870 for_each_hstate(h
) {
1871 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1874 kobject_put(hugepages_kobj
);
1876 module_exit(hugetlb_exit
);
1878 static int __init
hugetlb_init(void)
1880 /* Some platform decide whether they support huge pages at boot
1881 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1882 * there is no such support
1884 if (HPAGE_SHIFT
== 0)
1887 if (!size_to_hstate(default_hstate_size
)) {
1888 default_hstate_size
= HPAGE_SIZE
;
1889 if (!size_to_hstate(default_hstate_size
))
1890 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1892 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1893 if (default_hstate_max_huge_pages
)
1894 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1896 hugetlb_init_hstates();
1897 gather_bootmem_prealloc();
1900 hugetlb_sysfs_init();
1901 hugetlb_register_all_nodes();
1902 hugetlb_cgroup_file_init();
1906 module_init(hugetlb_init
);
1908 /* Should be called on processing a hugepagesz=... option */
1909 void __init
hugetlb_add_hstate(unsigned order
)
1914 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1915 pr_warning("hugepagesz= specified twice, ignoring\n");
1918 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1920 h
= &hstates
[hugetlb_max_hstate
++];
1922 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1923 h
->nr_huge_pages
= 0;
1924 h
->free_huge_pages
= 0;
1925 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1926 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1927 INIT_LIST_HEAD(&h
->hugepage_activelist
);
1928 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
1929 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
1930 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1931 huge_page_size(h
)/1024);
1936 static int __init
hugetlb_nrpages_setup(char *s
)
1939 static unsigned long *last_mhp
;
1942 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1943 * so this hugepages= parameter goes to the "default hstate".
1945 if (!hugetlb_max_hstate
)
1946 mhp
= &default_hstate_max_huge_pages
;
1948 mhp
= &parsed_hstate
->max_huge_pages
;
1950 if (mhp
== last_mhp
) {
1951 pr_warning("hugepages= specified twice without "
1952 "interleaving hugepagesz=, ignoring\n");
1956 if (sscanf(s
, "%lu", mhp
) <= 0)
1960 * Global state is always initialized later in hugetlb_init.
1961 * But we need to allocate >= MAX_ORDER hstates here early to still
1962 * use the bootmem allocator.
1964 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1965 hugetlb_hstate_alloc_pages(parsed_hstate
);
1971 __setup("hugepages=", hugetlb_nrpages_setup
);
1973 static int __init
hugetlb_default_setup(char *s
)
1975 default_hstate_size
= memparse(s
, &s
);
1978 __setup("default_hugepagesz=", hugetlb_default_setup
);
1980 static unsigned int cpuset_mems_nr(unsigned int *array
)
1983 unsigned int nr
= 0;
1985 for_each_node_mask(node
, cpuset_current_mems_allowed
)
1991 #ifdef CONFIG_SYSCTL
1992 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
1993 struct ctl_table
*table
, int write
,
1994 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1996 struct hstate
*h
= &default_hstate
;
2000 tmp
= h
->max_huge_pages
;
2002 if (write
&& h
->order
>= MAX_ORDER
)
2006 table
->maxlen
= sizeof(unsigned long);
2007 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2012 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2013 GFP_KERNEL
| __GFP_NORETRY
);
2014 if (!(obey_mempolicy
&&
2015 init_nodemask_of_mempolicy(nodes_allowed
))) {
2016 NODEMASK_FREE(nodes_allowed
);
2017 nodes_allowed
= &node_states
[N_MEMORY
];
2019 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2021 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2022 NODEMASK_FREE(nodes_allowed
);
2028 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2029 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2032 return hugetlb_sysctl_handler_common(false, table
, write
,
2033 buffer
, length
, ppos
);
2037 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2038 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2040 return hugetlb_sysctl_handler_common(true, table
, write
,
2041 buffer
, length
, ppos
);
2043 #endif /* CONFIG_NUMA */
2045 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2046 void __user
*buffer
,
2047 size_t *length
, loff_t
*ppos
)
2049 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2050 if (hugepages_treat_as_movable
)
2051 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2053 htlb_alloc_mask
= GFP_HIGHUSER
;
2057 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2058 void __user
*buffer
,
2059 size_t *length
, loff_t
*ppos
)
2061 struct hstate
*h
= &default_hstate
;
2065 tmp
= h
->nr_overcommit_huge_pages
;
2067 if (write
&& h
->order
>= MAX_ORDER
)
2071 table
->maxlen
= sizeof(unsigned long);
2072 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2077 spin_lock(&hugetlb_lock
);
2078 h
->nr_overcommit_huge_pages
= tmp
;
2079 spin_unlock(&hugetlb_lock
);
2085 #endif /* CONFIG_SYSCTL */
2087 void hugetlb_report_meminfo(struct seq_file
*m
)
2089 struct hstate
*h
= &default_hstate
;
2091 "HugePages_Total: %5lu\n"
2092 "HugePages_Free: %5lu\n"
2093 "HugePages_Rsvd: %5lu\n"
2094 "HugePages_Surp: %5lu\n"
2095 "Hugepagesize: %8lu kB\n",
2099 h
->surplus_huge_pages
,
2100 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2103 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2105 struct hstate
*h
= &default_hstate
;
2107 "Node %d HugePages_Total: %5u\n"
2108 "Node %d HugePages_Free: %5u\n"
2109 "Node %d HugePages_Surp: %5u\n",
2110 nid
, h
->nr_huge_pages_node
[nid
],
2111 nid
, h
->free_huge_pages_node
[nid
],
2112 nid
, h
->surplus_huge_pages_node
[nid
]);
2115 void hugetlb_show_meminfo(void)
2120 for_each_node_state(nid
, N_MEMORY
)
2122 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2124 h
->nr_huge_pages_node
[nid
],
2125 h
->free_huge_pages_node
[nid
],
2126 h
->surplus_huge_pages_node
[nid
],
2127 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2130 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2131 unsigned long hugetlb_total_pages(void)
2134 unsigned long nr_total_pages
= 0;
2137 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2138 return nr_total_pages
;
2141 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2145 spin_lock(&hugetlb_lock
);
2147 * When cpuset is configured, it breaks the strict hugetlb page
2148 * reservation as the accounting is done on a global variable. Such
2149 * reservation is completely rubbish in the presence of cpuset because
2150 * the reservation is not checked against page availability for the
2151 * current cpuset. Application can still potentially OOM'ed by kernel
2152 * with lack of free htlb page in cpuset that the task is in.
2153 * Attempt to enforce strict accounting with cpuset is almost
2154 * impossible (or too ugly) because cpuset is too fluid that
2155 * task or memory node can be dynamically moved between cpusets.
2157 * The change of semantics for shared hugetlb mapping with cpuset is
2158 * undesirable. However, in order to preserve some of the semantics,
2159 * we fall back to check against current free page availability as
2160 * a best attempt and hopefully to minimize the impact of changing
2161 * semantics that cpuset has.
2164 if (gather_surplus_pages(h
, delta
) < 0)
2167 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2168 return_unused_surplus_pages(h
, delta
);
2175 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2178 spin_unlock(&hugetlb_lock
);
2182 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2184 struct resv_map
*reservations
= vma_resv_map(vma
);
2187 * This new VMA should share its siblings reservation map if present.
2188 * The VMA will only ever have a valid reservation map pointer where
2189 * it is being copied for another still existing VMA. As that VMA
2190 * has a reference to the reservation map it cannot disappear until
2191 * after this open call completes. It is therefore safe to take a
2192 * new reference here without additional locking.
2195 kref_get(&reservations
->refs
);
2198 static void resv_map_put(struct vm_area_struct
*vma
)
2200 struct resv_map
*reservations
= vma_resv_map(vma
);
2204 kref_put(&reservations
->refs
, resv_map_release
);
2207 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2209 struct hstate
*h
= hstate_vma(vma
);
2210 struct resv_map
*reservations
= vma_resv_map(vma
);
2211 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2212 unsigned long reserve
;
2213 unsigned long start
;
2217 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2218 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2220 reserve
= (end
- start
) -
2221 region_count(&reservations
->regions
, start
, end
);
2226 hugetlb_acct_memory(h
, -reserve
);
2227 hugepage_subpool_put_pages(spool
, reserve
);
2233 * We cannot handle pagefaults against hugetlb pages at all. They cause
2234 * handle_mm_fault() to try to instantiate regular-sized pages in the
2235 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2238 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2244 const struct vm_operations_struct hugetlb_vm_ops
= {
2245 .fault
= hugetlb_vm_op_fault
,
2246 .open
= hugetlb_vm_op_open
,
2247 .close
= hugetlb_vm_op_close
,
2250 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2256 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2257 vma
->vm_page_prot
)));
2259 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2260 vma
->vm_page_prot
));
2262 entry
= pte_mkyoung(entry
);
2263 entry
= pte_mkhuge(entry
);
2264 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2269 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2270 unsigned long address
, pte_t
*ptep
)
2274 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2275 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2276 update_mmu_cache(vma
, address
, ptep
);
2280 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2281 struct vm_area_struct
*vma
)
2283 pte_t
*src_pte
, *dst_pte
, entry
;
2284 struct page
*ptepage
;
2287 struct hstate
*h
= hstate_vma(vma
);
2288 unsigned long sz
= huge_page_size(h
);
2290 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2292 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2293 src_pte
= huge_pte_offset(src
, addr
);
2296 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2300 /* If the pagetables are shared don't copy or take references */
2301 if (dst_pte
== src_pte
)
2304 spin_lock(&dst
->page_table_lock
);
2305 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2306 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2308 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2309 entry
= huge_ptep_get(src_pte
);
2310 ptepage
= pte_page(entry
);
2312 page_dup_rmap(ptepage
);
2313 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2315 spin_unlock(&src
->page_table_lock
);
2316 spin_unlock(&dst
->page_table_lock
);
2324 static int is_hugetlb_entry_migration(pte_t pte
)
2328 if (huge_pte_none(pte
) || pte_present(pte
))
2330 swp
= pte_to_swp_entry(pte
);
2331 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2337 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2341 if (huge_pte_none(pte
) || pte_present(pte
))
2343 swp
= pte_to_swp_entry(pte
);
2344 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2350 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2351 unsigned long start
, unsigned long end
,
2352 struct page
*ref_page
)
2354 int force_flush
= 0;
2355 struct mm_struct
*mm
= vma
->vm_mm
;
2356 unsigned long address
;
2360 struct hstate
*h
= hstate_vma(vma
);
2361 unsigned long sz
= huge_page_size(h
);
2362 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2363 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2365 WARN_ON(!is_vm_hugetlb_page(vma
));
2366 BUG_ON(start
& ~huge_page_mask(h
));
2367 BUG_ON(end
& ~huge_page_mask(h
));
2369 tlb_start_vma(tlb
, vma
);
2370 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2372 spin_lock(&mm
->page_table_lock
);
2373 for (address
= start
; address
< end
; address
+= sz
) {
2374 ptep
= huge_pte_offset(mm
, address
);
2378 if (huge_pmd_unshare(mm
, &address
, ptep
))
2381 pte
= huge_ptep_get(ptep
);
2382 if (huge_pte_none(pte
))
2386 * HWPoisoned hugepage is already unmapped and dropped reference
2388 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2389 huge_pte_clear(mm
, address
, ptep
);
2393 page
= pte_page(pte
);
2395 * If a reference page is supplied, it is because a specific
2396 * page is being unmapped, not a range. Ensure the page we
2397 * are about to unmap is the actual page of interest.
2400 if (page
!= ref_page
)
2404 * Mark the VMA as having unmapped its page so that
2405 * future faults in this VMA will fail rather than
2406 * looking like data was lost
2408 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2411 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2412 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2413 if (huge_pte_dirty(pte
))
2414 set_page_dirty(page
);
2416 page_remove_rmap(page
);
2417 force_flush
= !__tlb_remove_page(tlb
, page
);
2420 /* Bail out after unmapping reference page if supplied */
2424 spin_unlock(&mm
->page_table_lock
);
2426 * mmu_gather ran out of room to batch pages, we break out of
2427 * the PTE lock to avoid doing the potential expensive TLB invalidate
2428 * and page-free while holding it.
2433 if (address
< end
&& !ref_page
)
2436 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2437 tlb_end_vma(tlb
, vma
);
2440 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2441 struct vm_area_struct
*vma
, unsigned long start
,
2442 unsigned long end
, struct page
*ref_page
)
2444 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2447 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2448 * test will fail on a vma being torn down, and not grab a page table
2449 * on its way out. We're lucky that the flag has such an appropriate
2450 * name, and can in fact be safely cleared here. We could clear it
2451 * before the __unmap_hugepage_range above, but all that's necessary
2452 * is to clear it before releasing the i_mmap_mutex. This works
2453 * because in the context this is called, the VMA is about to be
2454 * destroyed and the i_mmap_mutex is held.
2456 vma
->vm_flags
&= ~VM_MAYSHARE
;
2459 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2460 unsigned long end
, struct page
*ref_page
)
2462 struct mm_struct
*mm
;
2463 struct mmu_gather tlb
;
2467 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2468 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2469 tlb_finish_mmu(&tlb
, start
, end
);
2473 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2474 * mappping it owns the reserve page for. The intention is to unmap the page
2475 * from other VMAs and let the children be SIGKILLed if they are faulting the
2478 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2479 struct page
*page
, unsigned long address
)
2481 struct hstate
*h
= hstate_vma(vma
);
2482 struct vm_area_struct
*iter_vma
;
2483 struct address_space
*mapping
;
2487 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2488 * from page cache lookup which is in HPAGE_SIZE units.
2490 address
= address
& huge_page_mask(h
);
2491 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2493 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2496 * Take the mapping lock for the duration of the table walk. As
2497 * this mapping should be shared between all the VMAs,
2498 * __unmap_hugepage_range() is called as the lock is already held
2500 mutex_lock(&mapping
->i_mmap_mutex
);
2501 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2502 /* Do not unmap the current VMA */
2503 if (iter_vma
== vma
)
2507 * Unmap the page from other VMAs without their own reserves.
2508 * They get marked to be SIGKILLed if they fault in these
2509 * areas. This is because a future no-page fault on this VMA
2510 * could insert a zeroed page instead of the data existing
2511 * from the time of fork. This would look like data corruption
2513 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2514 unmap_hugepage_range(iter_vma
, address
,
2515 address
+ huge_page_size(h
), page
);
2517 mutex_unlock(&mapping
->i_mmap_mutex
);
2523 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2524 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2525 * cannot race with other handlers or page migration.
2526 * Keep the pte_same checks anyway to make transition from the mutex easier.
2528 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2529 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2530 struct page
*pagecache_page
)
2532 struct hstate
*h
= hstate_vma(vma
);
2533 struct page
*old_page
, *new_page
;
2535 int outside_reserve
= 0;
2536 unsigned long mmun_start
; /* For mmu_notifiers */
2537 unsigned long mmun_end
; /* For mmu_notifiers */
2539 old_page
= pte_page(pte
);
2542 /* If no-one else is actually using this page, avoid the copy
2543 * and just make the page writable */
2544 avoidcopy
= (page_mapcount(old_page
) == 1);
2546 if (PageAnon(old_page
))
2547 page_move_anon_rmap(old_page
, vma
, address
);
2548 set_huge_ptep_writable(vma
, address
, ptep
);
2553 * If the process that created a MAP_PRIVATE mapping is about to
2554 * perform a COW due to a shared page count, attempt to satisfy
2555 * the allocation without using the existing reserves. The pagecache
2556 * page is used to determine if the reserve at this address was
2557 * consumed or not. If reserves were used, a partial faulted mapping
2558 * at the time of fork() could consume its reserves on COW instead
2559 * of the full address range.
2561 if (!(vma
->vm_flags
& VM_MAYSHARE
) &&
2562 is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2563 old_page
!= pagecache_page
)
2564 outside_reserve
= 1;
2566 page_cache_get(old_page
);
2568 /* Drop page_table_lock as buddy allocator may be called */
2569 spin_unlock(&mm
->page_table_lock
);
2570 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2572 if (IS_ERR(new_page
)) {
2573 long err
= PTR_ERR(new_page
);
2574 page_cache_release(old_page
);
2577 * If a process owning a MAP_PRIVATE mapping fails to COW,
2578 * it is due to references held by a child and an insufficient
2579 * huge page pool. To guarantee the original mappers
2580 * reliability, unmap the page from child processes. The child
2581 * may get SIGKILLed if it later faults.
2583 if (outside_reserve
) {
2584 BUG_ON(huge_pte_none(pte
));
2585 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2586 BUG_ON(huge_pte_none(pte
));
2587 spin_lock(&mm
->page_table_lock
);
2588 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2589 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2590 goto retry_avoidcopy
;
2592 * race occurs while re-acquiring page_table_lock, and
2600 /* Caller expects lock to be held */
2601 spin_lock(&mm
->page_table_lock
);
2603 return VM_FAULT_OOM
;
2605 return VM_FAULT_SIGBUS
;
2609 * When the original hugepage is shared one, it does not have
2610 * anon_vma prepared.
2612 if (unlikely(anon_vma_prepare(vma
))) {
2613 page_cache_release(new_page
);
2614 page_cache_release(old_page
);
2615 /* Caller expects lock to be held */
2616 spin_lock(&mm
->page_table_lock
);
2617 return VM_FAULT_OOM
;
2620 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2621 pages_per_huge_page(h
));
2622 __SetPageUptodate(new_page
);
2624 mmun_start
= address
& huge_page_mask(h
);
2625 mmun_end
= mmun_start
+ huge_page_size(h
);
2626 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2628 * Retake the page_table_lock to check for racing updates
2629 * before the page tables are altered
2631 spin_lock(&mm
->page_table_lock
);
2632 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2633 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2635 huge_ptep_clear_flush(vma
, address
, ptep
);
2636 set_huge_pte_at(mm
, address
, ptep
,
2637 make_huge_pte(vma
, new_page
, 1));
2638 page_remove_rmap(old_page
);
2639 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2640 /* Make the old page be freed below */
2641 new_page
= old_page
;
2643 spin_unlock(&mm
->page_table_lock
);
2644 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2645 /* Caller expects lock to be held */
2646 spin_lock(&mm
->page_table_lock
);
2647 page_cache_release(new_page
);
2648 page_cache_release(old_page
);
2652 /* Return the pagecache page at a given address within a VMA */
2653 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2654 struct vm_area_struct
*vma
, unsigned long address
)
2656 struct address_space
*mapping
;
2659 mapping
= vma
->vm_file
->f_mapping
;
2660 idx
= vma_hugecache_offset(h
, vma
, address
);
2662 return find_lock_page(mapping
, idx
);
2666 * Return whether there is a pagecache page to back given address within VMA.
2667 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2669 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2670 struct vm_area_struct
*vma
, unsigned long address
)
2672 struct address_space
*mapping
;
2676 mapping
= vma
->vm_file
->f_mapping
;
2677 idx
= vma_hugecache_offset(h
, vma
, address
);
2679 page
= find_get_page(mapping
, idx
);
2682 return page
!= NULL
;
2685 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2686 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2688 struct hstate
*h
= hstate_vma(vma
);
2689 int ret
= VM_FAULT_SIGBUS
;
2694 struct address_space
*mapping
;
2698 * Currently, we are forced to kill the process in the event the
2699 * original mapper has unmapped pages from the child due to a failed
2700 * COW. Warn that such a situation has occurred as it may not be obvious
2702 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2703 pr_warning("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_huge(mm
, ptep
);
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
) && !huge_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 (!huge_pte_write(entry
)) {
2897 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2899 goto out_page_table_lock
;
2901 entry
= huge_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 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2926 struct page
**pages
, struct vm_area_struct
**vmas
,
2927 unsigned long *position
, unsigned long *nr_pages
,
2928 long i
, unsigned int flags
)
2930 unsigned long pfn_offset
;
2931 unsigned long vaddr
= *position
;
2932 unsigned long remainder
= *nr_pages
;
2933 struct hstate
*h
= hstate_vma(vma
);
2935 spin_lock(&mm
->page_table_lock
);
2936 while (vaddr
< vma
->vm_end
&& remainder
) {
2942 * Some archs (sparc64, sh*) have multiple pte_ts to
2943 * each hugepage. We have to make sure we get the
2944 * first, for the page indexing below to work.
2946 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2947 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
2950 * When coredumping, it suits get_dump_page if we just return
2951 * an error where there's an empty slot with no huge pagecache
2952 * to back it. This way, we avoid allocating a hugepage, and
2953 * the sparse dumpfile avoids allocating disk blocks, but its
2954 * huge holes still show up with zeroes where they need to be.
2956 if (absent
&& (flags
& FOLL_DUMP
) &&
2957 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
2963 * We need call hugetlb_fault for both hugepages under migration
2964 * (in which case hugetlb_fault waits for the migration,) and
2965 * hwpoisoned hugepages (in which case we need to prevent the
2966 * caller from accessing to them.) In order to do this, we use
2967 * here is_swap_pte instead of is_hugetlb_entry_migration and
2968 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2969 * both cases, and because we can't follow correct pages
2970 * directly from any kind of swap entries.
2972 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
2973 ((flags
& FOLL_WRITE
) &&
2974 !huge_pte_write(huge_ptep_get(pte
)))) {
2977 spin_unlock(&mm
->page_table_lock
);
2978 ret
= hugetlb_fault(mm
, vma
, vaddr
,
2979 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
2980 spin_lock(&mm
->page_table_lock
);
2981 if (!(ret
& VM_FAULT_ERROR
))
2988 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
2989 page
= pte_page(huge_ptep_get(pte
));
2992 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3003 if (vaddr
< vma
->vm_end
&& remainder
&&
3004 pfn_offset
< pages_per_huge_page(h
)) {
3006 * We use pfn_offset to avoid touching the pageframes
3007 * of this compound page.
3012 spin_unlock(&mm
->page_table_lock
);
3013 *nr_pages
= remainder
;
3016 return i
? i
: -EFAULT
;
3019 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3020 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3022 struct mm_struct
*mm
= vma
->vm_mm
;
3023 unsigned long start
= address
;
3026 struct hstate
*h
= hstate_vma(vma
);
3027 unsigned long pages
= 0;
3029 BUG_ON(address
>= end
);
3030 flush_cache_range(vma
, address
, end
);
3032 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3033 spin_lock(&mm
->page_table_lock
);
3034 for (; address
< end
; address
+= huge_page_size(h
)) {
3035 ptep
= huge_pte_offset(mm
, address
);
3038 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3042 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3043 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3044 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3045 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3046 set_huge_pte_at(mm
, address
, ptep
, pte
);
3050 spin_unlock(&mm
->page_table_lock
);
3052 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3053 * may have cleared our pud entry and done put_page on the page table:
3054 * once we release i_mmap_mutex, another task can do the final put_page
3055 * and that page table be reused and filled with junk.
3057 flush_tlb_range(vma
, start
, end
);
3058 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3060 return pages
<< h
->order
;
3063 int hugetlb_reserve_pages(struct inode
*inode
,
3065 struct vm_area_struct
*vma
,
3066 vm_flags_t vm_flags
)
3069 struct hstate
*h
= hstate_inode(inode
);
3070 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3073 * Only apply hugepage reservation if asked. At fault time, an
3074 * attempt will be made for VM_NORESERVE to allocate a page
3075 * without using reserves
3077 if (vm_flags
& VM_NORESERVE
)
3081 * Shared mappings base their reservation on the number of pages that
3082 * are already allocated on behalf of the file. Private mappings need
3083 * to reserve the full area even if read-only as mprotect() may be
3084 * called to make the mapping read-write. Assume !vma is a shm mapping
3086 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3087 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3089 struct resv_map
*resv_map
= resv_map_alloc();
3095 set_vma_resv_map(vma
, resv_map
);
3096 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3104 /* There must be enough pages in the subpool for the mapping */
3105 if (hugepage_subpool_get_pages(spool
, chg
)) {
3111 * Check enough hugepages are available for the reservation.
3112 * Hand the pages back to the subpool if there are not
3114 ret
= hugetlb_acct_memory(h
, chg
);
3116 hugepage_subpool_put_pages(spool
, chg
);
3121 * Account for the reservations made. Shared mappings record regions
3122 * that have reservations as they are shared by multiple VMAs.
3123 * When the last VMA disappears, the region map says how much
3124 * the reservation was and the page cache tells how much of
3125 * the reservation was consumed. Private mappings are per-VMA and
3126 * only the consumed reservations are tracked. When the VMA
3127 * disappears, the original reservation is the VMA size and the
3128 * consumed reservations are stored in the map. Hence, nothing
3129 * else has to be done for private mappings here
3131 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3132 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3140 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3142 struct hstate
*h
= hstate_inode(inode
);
3143 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3144 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3146 spin_lock(&inode
->i_lock
);
3147 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3148 spin_unlock(&inode
->i_lock
);
3150 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3151 hugetlb_acct_memory(h
, -(chg
- freed
));
3154 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3155 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3156 struct vm_area_struct
*vma
,
3157 unsigned long addr
, pgoff_t idx
)
3159 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3161 unsigned long sbase
= saddr
& PUD_MASK
;
3162 unsigned long s_end
= sbase
+ PUD_SIZE
;
3164 /* Allow segments to share if only one is marked locked */
3165 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3166 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3169 * match the virtual addresses, permission and the alignment of the
3172 if (pmd_index(addr
) != pmd_index(saddr
) ||
3173 vm_flags
!= svm_flags
||
3174 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3180 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3182 unsigned long base
= addr
& PUD_MASK
;
3183 unsigned long end
= base
+ PUD_SIZE
;
3186 * check on proper vm_flags and page table alignment
3188 if (vma
->vm_flags
& VM_MAYSHARE
&&
3189 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3195 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3196 * and returns the corresponding pte. While this is not necessary for the
3197 * !shared pmd case because we can allocate the pmd later as well, it makes the
3198 * code much cleaner. pmd allocation is essential for the shared case because
3199 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3200 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3201 * bad pmd for sharing.
3203 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3205 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3206 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3207 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3209 struct vm_area_struct
*svma
;
3210 unsigned long saddr
;
3214 if (!vma_shareable(vma
, addr
))
3215 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3217 mutex_lock(&mapping
->i_mmap_mutex
);
3218 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3222 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3224 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3226 get_page(virt_to_page(spte
));
3235 spin_lock(&mm
->page_table_lock
);
3237 pud_populate(mm
, pud
,
3238 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3240 put_page(virt_to_page(spte
));
3241 spin_unlock(&mm
->page_table_lock
);
3243 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3244 mutex_unlock(&mapping
->i_mmap_mutex
);
3249 * unmap huge page backed by shared pte.
3251 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3252 * indicated by page_count > 1, unmap is achieved by clearing pud and
3253 * decrementing the ref count. If count == 1, the pte page is not shared.
3255 * called with vma->vm_mm->page_table_lock held.
3257 * returns: 1 successfully unmapped a shared pte page
3258 * 0 the underlying pte page is not shared, or it is the last user
3260 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3262 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3263 pud_t
*pud
= pud_offset(pgd
, *addr
);
3265 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3266 if (page_count(virt_to_page(ptep
)) == 1)
3270 put_page(virt_to_page(ptep
));
3271 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3274 #define want_pmd_share() (1)
3275 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3276 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3280 #define want_pmd_share() (0)
3281 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3283 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3284 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3285 unsigned long addr
, unsigned long sz
)
3291 pgd
= pgd_offset(mm
, addr
);
3292 pud
= pud_alloc(mm
, pgd
, addr
);
3294 if (sz
== PUD_SIZE
) {
3297 BUG_ON(sz
!= PMD_SIZE
);
3298 if (want_pmd_share() && pud_none(*pud
))
3299 pte
= huge_pmd_share(mm
, addr
, pud
);
3301 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3304 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3309 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3315 pgd
= pgd_offset(mm
, addr
);
3316 if (pgd_present(*pgd
)) {
3317 pud
= pud_offset(pgd
, addr
);
3318 if (pud_present(*pud
)) {
3320 return (pte_t
*)pud
;
3321 pmd
= pmd_offset(pud
, addr
);
3324 return (pte_t
*) pmd
;
3328 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3329 pmd_t
*pmd
, int write
)
3333 page
= pte_page(*(pte_t
*)pmd
);
3335 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3340 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3341 pud_t
*pud
, int write
)
3345 page
= pte_page(*(pte_t
*)pud
);
3347 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3351 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3353 /* Can be overriden by architectures */
3354 __attribute__((weak
)) struct page
*
3355 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3356 pud_t
*pud
, int write
)
3362 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3364 #ifdef CONFIG_MEMORY_FAILURE
3366 /* Should be called in hugetlb_lock */
3367 static int is_hugepage_on_freelist(struct page
*hpage
)
3371 struct hstate
*h
= page_hstate(hpage
);
3372 int nid
= page_to_nid(hpage
);
3374 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3381 * This function is called from memory failure code.
3382 * Assume the caller holds page lock of the head page.
3384 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3386 struct hstate
*h
= page_hstate(hpage
);
3387 int nid
= page_to_nid(hpage
);
3390 spin_lock(&hugetlb_lock
);
3391 if (is_hugepage_on_freelist(hpage
)) {
3393 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3394 * but dangling hpage->lru can trigger list-debug warnings
3395 * (this happens when we call unpoison_memory() on it),
3396 * so let it point to itself with list_del_init().
3398 list_del_init(&hpage
->lru
);
3399 set_page_refcounted(hpage
);
3400 h
->free_huge_pages
--;
3401 h
->free_huge_pages_node
[nid
]--;
3404 spin_unlock(&hugetlb_lock
);