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/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
38 int hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 __initdata
LIST_HEAD(huge_boot_pages
);
46 /* for command line parsing */
47 static struct hstate
* __initdata parsed_hstate
;
48 static unsigned long __initdata default_hstate_max_huge_pages
;
49 static unsigned long __initdata default_hstate_size
;
52 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
55 DEFINE_SPINLOCK(hugetlb_lock
);
58 * Serializes faults on the same logical page. This is used to
59 * prevent spurious OOMs when the hugepage pool is fully utilized.
61 static int num_fault_mutexes
;
62 static struct mutex
*htlb_fault_mutex_table ____cacheline_aligned_in_smp
;
64 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
66 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
68 spin_unlock(&spool
->lock
);
70 /* If no pages are used, and no other handles to the subpool
71 * remain, free the subpool the subpool remain */
76 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
78 struct hugepage_subpool
*spool
;
80 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
84 spin_lock_init(&spool
->lock
);
86 spool
->max_hpages
= nr_blocks
;
91 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
93 spin_lock(&spool
->lock
);
94 BUG_ON(!spool
->count
);
96 unlock_or_release_subpool(spool
);
99 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
107 spin_lock(&spool
->lock
);
108 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
109 spool
->used_hpages
+= delta
;
113 spin_unlock(&spool
->lock
);
118 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
124 spin_lock(&spool
->lock
);
125 spool
->used_hpages
-= delta
;
126 /* If hugetlbfs_put_super couldn't free spool due to
127 * an outstanding quota reference, free it now. */
128 unlock_or_release_subpool(spool
);
131 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
133 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
136 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
138 return subpool_inode(file_inode(vma
->vm_file
));
142 * Region tracking -- allows tracking of reservations and instantiated pages
143 * across the pages in a mapping.
145 * The region data structures are embedded into a resv_map and
146 * protected by a resv_map's lock
149 struct list_head link
;
154 static long region_add(struct resv_map
*resv
, long f
, long t
)
156 struct list_head
*head
= &resv
->regions
;
157 struct file_region
*rg
, *nrg
, *trg
;
159 spin_lock(&resv
->lock
);
160 /* Locate the region we are either in or before. */
161 list_for_each_entry(rg
, head
, link
)
165 /* Round our left edge to the current segment if it encloses us. */
169 /* Check for and consume any regions we now overlap with. */
171 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
172 if (&rg
->link
== head
)
177 /* If this area reaches higher then extend our area to
178 * include it completely. If this is not the first area
179 * which we intend to reuse, free it. */
189 spin_unlock(&resv
->lock
);
193 static long region_chg(struct resv_map
*resv
, long f
, long t
)
195 struct list_head
*head
= &resv
->regions
;
196 struct file_region
*rg
, *nrg
= NULL
;
200 spin_lock(&resv
->lock
);
201 /* Locate the region we are before or in. */
202 list_for_each_entry(rg
, head
, link
)
206 /* If we are below the current region then a new region is required.
207 * Subtle, allocate a new region at the position but make it zero
208 * size such that we can guarantee to record the reservation. */
209 if (&rg
->link
== head
|| t
< rg
->from
) {
211 spin_unlock(&resv
->lock
);
212 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
218 INIT_LIST_HEAD(&nrg
->link
);
222 list_add(&nrg
->link
, rg
->link
.prev
);
227 /* Round our left edge to the current segment if it encloses us. */
232 /* Check for and consume any regions we now overlap with. */
233 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
234 if (&rg
->link
== head
)
239 /* We overlap with this area, if it extends further than
240 * us then we must extend ourselves. Account for its
241 * existing reservation. */
246 chg
-= rg
->to
- rg
->from
;
250 spin_unlock(&resv
->lock
);
251 /* We already know we raced and no longer need the new region */
255 spin_unlock(&resv
->lock
);
259 static long region_truncate(struct resv_map
*resv
, long end
)
261 struct list_head
*head
= &resv
->regions
;
262 struct file_region
*rg
, *trg
;
265 spin_lock(&resv
->lock
);
266 /* Locate the region we are either in or before. */
267 list_for_each_entry(rg
, head
, link
)
270 if (&rg
->link
== head
)
273 /* If we are in the middle of a region then adjust it. */
274 if (end
> rg
->from
) {
277 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
280 /* Drop any remaining regions. */
281 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
282 if (&rg
->link
== head
)
284 chg
+= rg
->to
- rg
->from
;
290 spin_unlock(&resv
->lock
);
294 static long region_count(struct resv_map
*resv
, long f
, long t
)
296 struct list_head
*head
= &resv
->regions
;
297 struct file_region
*rg
;
300 spin_lock(&resv
->lock
);
301 /* Locate each segment we overlap with, and count that overlap. */
302 list_for_each_entry(rg
, head
, link
) {
311 seg_from
= max(rg
->from
, f
);
312 seg_to
= min(rg
->to
, t
);
314 chg
+= seg_to
- seg_from
;
316 spin_unlock(&resv
->lock
);
322 * Convert the address within this vma to the page offset within
323 * the mapping, in pagecache page units; huge pages here.
325 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
326 struct vm_area_struct
*vma
, unsigned long address
)
328 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
329 (vma
->vm_pgoff
>> huge_page_order(h
));
332 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
333 unsigned long address
)
335 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
339 * Return the size of the pages allocated when backing a VMA. In the majority
340 * cases this will be same size as used by the page table entries.
342 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
344 struct hstate
*hstate
;
346 if (!is_vm_hugetlb_page(vma
))
349 hstate
= hstate_vma(vma
);
351 return 1UL << huge_page_shift(hstate
);
353 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
356 * Return the page size being used by the MMU to back a VMA. In the majority
357 * of cases, the page size used by the kernel matches the MMU size. On
358 * architectures where it differs, an architecture-specific version of this
359 * function is required.
361 #ifndef vma_mmu_pagesize
362 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
364 return vma_kernel_pagesize(vma
);
369 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
370 * bits of the reservation map pointer, which are always clear due to
373 #define HPAGE_RESV_OWNER (1UL << 0)
374 #define HPAGE_RESV_UNMAPPED (1UL << 1)
375 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
378 * These helpers are used to track how many pages are reserved for
379 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
380 * is guaranteed to have their future faults succeed.
382 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
383 * the reserve counters are updated with the hugetlb_lock held. It is safe
384 * to reset the VMA at fork() time as it is not in use yet and there is no
385 * chance of the global counters getting corrupted as a result of the values.
387 * The private mapping reservation is represented in a subtly different
388 * manner to a shared mapping. A shared mapping has a region map associated
389 * with the underlying file, this region map represents the backing file
390 * pages which have ever had a reservation assigned which this persists even
391 * after the page is instantiated. A private mapping has a region map
392 * associated with the original mmap which is attached to all VMAs which
393 * reference it, this region map represents those offsets which have consumed
394 * reservation ie. where pages have been instantiated.
396 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
398 return (unsigned long)vma
->vm_private_data
;
401 static void set_vma_private_data(struct vm_area_struct
*vma
,
404 vma
->vm_private_data
= (void *)value
;
407 struct resv_map
*resv_map_alloc(void)
409 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
413 kref_init(&resv_map
->refs
);
414 spin_lock_init(&resv_map
->lock
);
415 INIT_LIST_HEAD(&resv_map
->regions
);
420 void resv_map_release(struct kref
*ref
)
422 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
424 /* Clear out any active regions before we release the map. */
425 region_truncate(resv_map
, 0);
429 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
431 return inode
->i_mapping
->private_data
;
434 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
436 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
437 if (vma
->vm_flags
& VM_MAYSHARE
) {
438 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
439 struct inode
*inode
= mapping
->host
;
441 return inode_resv_map(inode
);
444 return (struct resv_map
*)(get_vma_private_data(vma
) &
449 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
451 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
452 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
454 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
455 HPAGE_RESV_MASK
) | (unsigned long)map
);
458 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
460 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
461 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
463 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
466 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
468 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
470 return (get_vma_private_data(vma
) & flag
) != 0;
473 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
474 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
476 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
477 if (!(vma
->vm_flags
& VM_MAYSHARE
))
478 vma
->vm_private_data
= (void *)0;
481 /* Returns true if the VMA has associated reserve pages */
482 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
484 if (vma
->vm_flags
& VM_NORESERVE
) {
486 * This address is already reserved by other process(chg == 0),
487 * so, we should decrement reserved count. Without decrementing,
488 * reserve count remains after releasing inode, because this
489 * allocated page will go into page cache and is regarded as
490 * coming from reserved pool in releasing step. Currently, we
491 * don't have any other solution to deal with this situation
492 * properly, so add work-around here.
494 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
500 /* Shared mappings always use reserves */
501 if (vma
->vm_flags
& VM_MAYSHARE
)
505 * Only the process that called mmap() has reserves for
508 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
514 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
516 int nid
= page_to_nid(page
);
517 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
518 h
->free_huge_pages
++;
519 h
->free_huge_pages_node
[nid
]++;
522 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
526 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
527 if (!is_migrate_isolate_page(page
))
530 * if 'non-isolated free hugepage' not found on the list,
531 * the allocation fails.
533 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
535 list_move(&page
->lru
, &h
->hugepage_activelist
);
536 set_page_refcounted(page
);
537 h
->free_huge_pages
--;
538 h
->free_huge_pages_node
[nid
]--;
542 /* Movability of hugepages depends on migration support. */
543 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
545 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
546 return GFP_HIGHUSER_MOVABLE
;
551 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
552 struct vm_area_struct
*vma
,
553 unsigned long address
, int avoid_reserve
,
556 struct page
*page
= NULL
;
557 struct mempolicy
*mpol
;
558 nodemask_t
*nodemask
;
559 struct zonelist
*zonelist
;
562 unsigned int cpuset_mems_cookie
;
565 * A child process with MAP_PRIVATE mappings created by their parent
566 * have no page reserves. This check ensures that reservations are
567 * not "stolen". The child may still get SIGKILLed
569 if (!vma_has_reserves(vma
, chg
) &&
570 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
573 /* If reserves cannot be used, ensure enough pages are in the pool */
574 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
578 cpuset_mems_cookie
= read_mems_allowed_begin();
579 zonelist
= huge_zonelist(vma
, address
,
580 htlb_alloc_mask(h
), &mpol
, &nodemask
);
582 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
583 MAX_NR_ZONES
- 1, nodemask
) {
584 if (cpuset_zone_allowed(zone
, htlb_alloc_mask(h
))) {
585 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
589 if (!vma_has_reserves(vma
, chg
))
592 SetPagePrivate(page
);
593 h
->resv_huge_pages
--;
600 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
609 * common helper functions for hstate_next_node_to_{alloc|free}.
610 * We may have allocated or freed a huge page based on a different
611 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
612 * be outside of *nodes_allowed. Ensure that we use an allowed
613 * node for alloc or free.
615 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
617 nid
= next_node(nid
, *nodes_allowed
);
618 if (nid
== MAX_NUMNODES
)
619 nid
= first_node(*nodes_allowed
);
620 VM_BUG_ON(nid
>= MAX_NUMNODES
);
625 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
627 if (!node_isset(nid
, *nodes_allowed
))
628 nid
= next_node_allowed(nid
, nodes_allowed
);
633 * returns the previously saved node ["this node"] from which to
634 * allocate a persistent huge page for the pool and advance the
635 * next node from which to allocate, handling wrap at end of node
638 static int hstate_next_node_to_alloc(struct hstate
*h
,
639 nodemask_t
*nodes_allowed
)
643 VM_BUG_ON(!nodes_allowed
);
645 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
646 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
652 * helper for free_pool_huge_page() - return the previously saved
653 * node ["this node"] from which to free a huge page. Advance the
654 * next node id whether or not we find a free huge page to free so
655 * that the next attempt to free addresses the next node.
657 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
661 VM_BUG_ON(!nodes_allowed
);
663 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
664 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
669 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
670 for (nr_nodes = nodes_weight(*mask); \
672 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
675 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
676 for (nr_nodes = nodes_weight(*mask); \
678 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
681 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
682 static void destroy_compound_gigantic_page(struct page
*page
,
686 int nr_pages
= 1 << order
;
687 struct page
*p
= page
+ 1;
689 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
691 set_page_refcounted(p
);
692 p
->first_page
= NULL
;
695 set_compound_order(page
, 0);
696 __ClearPageHead(page
);
699 static void free_gigantic_page(struct page
*page
, unsigned order
)
701 free_contig_range(page_to_pfn(page
), 1 << order
);
704 static int __alloc_gigantic_page(unsigned long start_pfn
,
705 unsigned long nr_pages
)
707 unsigned long end_pfn
= start_pfn
+ nr_pages
;
708 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
);
711 static bool pfn_range_valid_gigantic(unsigned long start_pfn
,
712 unsigned long nr_pages
)
714 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
717 for (i
= start_pfn
; i
< end_pfn
; i
++) {
721 page
= pfn_to_page(i
);
723 if (PageReserved(page
))
726 if (page_count(page
) > 0)
736 static bool zone_spans_last_pfn(const struct zone
*zone
,
737 unsigned long start_pfn
, unsigned long nr_pages
)
739 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
740 return zone_spans_pfn(zone
, last_pfn
);
743 static struct page
*alloc_gigantic_page(int nid
, unsigned order
)
745 unsigned long nr_pages
= 1 << order
;
746 unsigned long ret
, pfn
, flags
;
749 z
= NODE_DATA(nid
)->node_zones
;
750 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
751 spin_lock_irqsave(&z
->lock
, flags
);
753 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
754 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
755 if (pfn_range_valid_gigantic(pfn
, nr_pages
)) {
757 * We release the zone lock here because
758 * alloc_contig_range() will also lock the zone
759 * at some point. If there's an allocation
760 * spinning on this lock, it may win the race
761 * and cause alloc_contig_range() to fail...
763 spin_unlock_irqrestore(&z
->lock
, flags
);
764 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
766 return pfn_to_page(pfn
);
767 spin_lock_irqsave(&z
->lock
, flags
);
772 spin_unlock_irqrestore(&z
->lock
, flags
);
778 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
779 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
);
781 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
785 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
787 prep_compound_gigantic_page(page
, huge_page_order(h
));
788 prep_new_huge_page(h
, page
, nid
);
794 static int alloc_fresh_gigantic_page(struct hstate
*h
,
795 nodemask_t
*nodes_allowed
)
797 struct page
*page
= NULL
;
800 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
801 page
= alloc_fresh_gigantic_page_node(h
, node
);
809 static inline bool gigantic_page_supported(void) { return true; }
811 static inline bool gigantic_page_supported(void) { return false; }
812 static inline void free_gigantic_page(struct page
*page
, unsigned order
) { }
813 static inline void destroy_compound_gigantic_page(struct page
*page
,
814 unsigned long order
) { }
815 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
816 nodemask_t
*nodes_allowed
) { return 0; }
819 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
823 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
827 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
828 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
829 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
830 1 << PG_referenced
| 1 << PG_dirty
|
831 1 << PG_active
| 1 << PG_private
|
834 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
835 set_compound_page_dtor(page
, NULL
);
836 set_page_refcounted(page
);
837 if (hstate_is_gigantic(h
)) {
838 destroy_compound_gigantic_page(page
, huge_page_order(h
));
839 free_gigantic_page(page
, huge_page_order(h
));
841 arch_release_hugepage(page
);
842 __free_pages(page
, huge_page_order(h
));
846 struct hstate
*size_to_hstate(unsigned long size
)
851 if (huge_page_size(h
) == size
)
857 void free_huge_page(struct page
*page
)
860 * Can't pass hstate in here because it is called from the
861 * compound page destructor.
863 struct hstate
*h
= page_hstate(page
);
864 int nid
= page_to_nid(page
);
865 struct hugepage_subpool
*spool
=
866 (struct hugepage_subpool
*)page_private(page
);
867 bool restore_reserve
;
869 set_page_private(page
, 0);
870 page
->mapping
= NULL
;
871 BUG_ON(page_count(page
));
872 BUG_ON(page_mapcount(page
));
873 restore_reserve
= PagePrivate(page
);
874 ClearPagePrivate(page
);
876 spin_lock(&hugetlb_lock
);
877 hugetlb_cgroup_uncharge_page(hstate_index(h
),
878 pages_per_huge_page(h
), page
);
880 h
->resv_huge_pages
++;
882 if (h
->surplus_huge_pages_node
[nid
]) {
883 /* remove the page from active list */
884 list_del(&page
->lru
);
885 update_and_free_page(h
, page
);
886 h
->surplus_huge_pages
--;
887 h
->surplus_huge_pages_node
[nid
]--;
889 arch_clear_hugepage_flags(page
);
890 enqueue_huge_page(h
, page
);
892 spin_unlock(&hugetlb_lock
);
893 hugepage_subpool_put_pages(spool
, 1);
896 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
898 INIT_LIST_HEAD(&page
->lru
);
899 set_compound_page_dtor(page
, free_huge_page
);
900 spin_lock(&hugetlb_lock
);
901 set_hugetlb_cgroup(page
, NULL
);
903 h
->nr_huge_pages_node
[nid
]++;
904 spin_unlock(&hugetlb_lock
);
905 put_page(page
); /* free it into the hugepage allocator */
908 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
911 int nr_pages
= 1 << order
;
912 struct page
*p
= page
+ 1;
914 /* we rely on prep_new_huge_page to set the destructor */
915 set_compound_order(page
, order
);
917 __ClearPageReserved(page
);
918 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
920 * For gigantic hugepages allocated through bootmem at
921 * boot, it's safer to be consistent with the not-gigantic
922 * hugepages and clear the PG_reserved bit from all tail pages
923 * too. Otherwse drivers using get_user_pages() to access tail
924 * pages may get the reference counting wrong if they see
925 * PG_reserved set on a tail page (despite the head page not
926 * having PG_reserved set). Enforcing this consistency between
927 * head and tail pages allows drivers to optimize away a check
928 * on the head page when they need know if put_page() is needed
929 * after get_user_pages().
931 __ClearPageReserved(p
);
932 set_page_count(p
, 0);
933 p
->first_page
= page
;
934 /* Make sure p->first_page is always valid for PageTail() */
941 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
942 * transparent huge pages. See the PageTransHuge() documentation for more
945 int PageHuge(struct page
*page
)
947 if (!PageCompound(page
))
950 page
= compound_head(page
);
951 return get_compound_page_dtor(page
) == free_huge_page
;
953 EXPORT_SYMBOL_GPL(PageHuge
);
956 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
957 * normal or transparent huge pages.
959 int PageHeadHuge(struct page
*page_head
)
961 if (!PageHead(page_head
))
964 return get_compound_page_dtor(page_head
) == free_huge_page
;
967 pgoff_t
__basepage_index(struct page
*page
)
969 struct page
*page_head
= compound_head(page
);
970 pgoff_t index
= page_index(page_head
);
971 unsigned long compound_idx
;
973 if (!PageHuge(page_head
))
974 return page_index(page
);
976 if (compound_order(page_head
) >= MAX_ORDER
)
977 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
979 compound_idx
= page
- page_head
;
981 return (index
<< compound_order(page_head
)) + compound_idx
;
984 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
988 page
= alloc_pages_exact_node(nid
,
989 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
990 __GFP_REPEAT
|__GFP_NOWARN
,
993 if (arch_prepare_hugepage(page
)) {
994 __free_pages(page
, huge_page_order(h
));
997 prep_new_huge_page(h
, page
, nid
);
1003 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1009 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1010 page
= alloc_fresh_huge_page_node(h
, node
);
1018 count_vm_event(HTLB_BUDDY_PGALLOC
);
1020 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1026 * Free huge page from pool from next node to free.
1027 * Attempt to keep persistent huge pages more or less
1028 * balanced over allowed nodes.
1029 * Called with hugetlb_lock locked.
1031 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1037 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1039 * If we're returning unused surplus pages, only examine
1040 * nodes with surplus pages.
1042 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1043 !list_empty(&h
->hugepage_freelists
[node
])) {
1045 list_entry(h
->hugepage_freelists
[node
].next
,
1047 list_del(&page
->lru
);
1048 h
->free_huge_pages
--;
1049 h
->free_huge_pages_node
[node
]--;
1051 h
->surplus_huge_pages
--;
1052 h
->surplus_huge_pages_node
[node
]--;
1054 update_and_free_page(h
, page
);
1064 * Dissolve a given free hugepage into free buddy pages. This function does
1065 * nothing for in-use (including surplus) hugepages.
1067 static void dissolve_free_huge_page(struct page
*page
)
1069 spin_lock(&hugetlb_lock
);
1070 if (PageHuge(page
) && !page_count(page
)) {
1071 struct hstate
*h
= page_hstate(page
);
1072 int nid
= page_to_nid(page
);
1073 list_del(&page
->lru
);
1074 h
->free_huge_pages
--;
1075 h
->free_huge_pages_node
[nid
]--;
1076 update_and_free_page(h
, page
);
1078 spin_unlock(&hugetlb_lock
);
1082 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1083 * make specified memory blocks removable from the system.
1084 * Note that start_pfn should aligned with (minimum) hugepage size.
1086 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1088 unsigned int order
= 8 * sizeof(void *);
1092 if (!hugepages_supported())
1095 /* Set scan step to minimum hugepage size */
1097 if (order
> huge_page_order(h
))
1098 order
= huge_page_order(h
);
1099 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
1100 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
1101 dissolve_free_huge_page(pfn_to_page(pfn
));
1104 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
1109 if (hstate_is_gigantic(h
))
1113 * Assume we will successfully allocate the surplus page to
1114 * prevent racing processes from causing the surplus to exceed
1117 * This however introduces a different race, where a process B
1118 * tries to grow the static hugepage pool while alloc_pages() is
1119 * called by process A. B will only examine the per-node
1120 * counters in determining if surplus huge pages can be
1121 * converted to normal huge pages in adjust_pool_surplus(). A
1122 * won't be able to increment the per-node counter, until the
1123 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1124 * no more huge pages can be converted from surplus to normal
1125 * state (and doesn't try to convert again). Thus, we have a
1126 * case where a surplus huge page exists, the pool is grown, and
1127 * the surplus huge page still exists after, even though it
1128 * should just have been converted to a normal huge page. This
1129 * does not leak memory, though, as the hugepage will be freed
1130 * once it is out of use. It also does not allow the counters to
1131 * go out of whack in adjust_pool_surplus() as we don't modify
1132 * the node values until we've gotten the hugepage and only the
1133 * per-node value is checked there.
1135 spin_lock(&hugetlb_lock
);
1136 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1137 spin_unlock(&hugetlb_lock
);
1141 h
->surplus_huge_pages
++;
1143 spin_unlock(&hugetlb_lock
);
1145 if (nid
== NUMA_NO_NODE
)
1146 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
1147 __GFP_REPEAT
|__GFP_NOWARN
,
1148 huge_page_order(h
));
1150 page
= alloc_pages_exact_node(nid
,
1151 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1152 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
1154 if (page
&& arch_prepare_hugepage(page
)) {
1155 __free_pages(page
, huge_page_order(h
));
1159 spin_lock(&hugetlb_lock
);
1161 INIT_LIST_HEAD(&page
->lru
);
1162 r_nid
= page_to_nid(page
);
1163 set_compound_page_dtor(page
, free_huge_page
);
1164 set_hugetlb_cgroup(page
, NULL
);
1166 * We incremented the global counters already
1168 h
->nr_huge_pages_node
[r_nid
]++;
1169 h
->surplus_huge_pages_node
[r_nid
]++;
1170 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1173 h
->surplus_huge_pages
--;
1174 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1176 spin_unlock(&hugetlb_lock
);
1182 * This allocation function is useful in the context where vma is irrelevant.
1183 * E.g. soft-offlining uses this function because it only cares physical
1184 * address of error page.
1186 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1188 struct page
*page
= NULL
;
1190 spin_lock(&hugetlb_lock
);
1191 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1192 page
= dequeue_huge_page_node(h
, nid
);
1193 spin_unlock(&hugetlb_lock
);
1196 page
= alloc_buddy_huge_page(h
, nid
);
1202 * Increase the hugetlb pool such that it can accommodate a reservation
1205 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1207 struct list_head surplus_list
;
1208 struct page
*page
, *tmp
;
1210 int needed
, allocated
;
1211 bool alloc_ok
= true;
1213 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1215 h
->resv_huge_pages
+= delta
;
1220 INIT_LIST_HEAD(&surplus_list
);
1224 spin_unlock(&hugetlb_lock
);
1225 for (i
= 0; i
< needed
; i
++) {
1226 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1231 list_add(&page
->lru
, &surplus_list
);
1236 * After retaking hugetlb_lock, we need to recalculate 'needed'
1237 * because either resv_huge_pages or free_huge_pages may have changed.
1239 spin_lock(&hugetlb_lock
);
1240 needed
= (h
->resv_huge_pages
+ delta
) -
1241 (h
->free_huge_pages
+ allocated
);
1246 * We were not able to allocate enough pages to
1247 * satisfy the entire reservation so we free what
1248 * we've allocated so far.
1253 * The surplus_list now contains _at_least_ the number of extra pages
1254 * needed to accommodate the reservation. Add the appropriate number
1255 * of pages to the hugetlb pool and free the extras back to the buddy
1256 * allocator. Commit the entire reservation here to prevent another
1257 * process from stealing the pages as they are added to the pool but
1258 * before they are reserved.
1260 needed
+= allocated
;
1261 h
->resv_huge_pages
+= delta
;
1264 /* Free the needed pages to the hugetlb pool */
1265 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1269 * This page is now managed by the hugetlb allocator and has
1270 * no users -- drop the buddy allocator's reference.
1272 put_page_testzero(page
);
1273 VM_BUG_ON_PAGE(page_count(page
), page
);
1274 enqueue_huge_page(h
, page
);
1277 spin_unlock(&hugetlb_lock
);
1279 /* Free unnecessary surplus pages to the buddy allocator */
1280 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1282 spin_lock(&hugetlb_lock
);
1288 * When releasing a hugetlb pool reservation, any surplus pages that were
1289 * allocated to satisfy the reservation must be explicitly freed if they were
1291 * Called with hugetlb_lock held.
1293 static void return_unused_surplus_pages(struct hstate
*h
,
1294 unsigned long unused_resv_pages
)
1296 unsigned long nr_pages
;
1298 /* Uncommit the reservation */
1299 h
->resv_huge_pages
-= unused_resv_pages
;
1301 /* Cannot return gigantic pages currently */
1302 if (hstate_is_gigantic(h
))
1305 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1308 * We want to release as many surplus pages as possible, spread
1309 * evenly across all nodes with memory. Iterate across these nodes
1310 * until we can no longer free unreserved surplus pages. This occurs
1311 * when the nodes with surplus pages have no free pages.
1312 * free_pool_huge_page() will balance the the freed pages across the
1313 * on-line nodes with memory and will handle the hstate accounting.
1315 while (nr_pages
--) {
1316 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1318 cond_resched_lock(&hugetlb_lock
);
1323 * Determine if the huge page at addr within the vma has an associated
1324 * reservation. Where it does not we will need to logically increase
1325 * reservation and actually increase subpool usage before an allocation
1326 * can occur. Where any new reservation would be required the
1327 * reservation change is prepared, but not committed. Once the page
1328 * has been allocated from the subpool and instantiated the change should
1329 * be committed via vma_commit_reservation. No action is required on
1332 static long vma_needs_reservation(struct hstate
*h
,
1333 struct vm_area_struct
*vma
, unsigned long addr
)
1335 struct resv_map
*resv
;
1339 resv
= vma_resv_map(vma
);
1343 idx
= vma_hugecache_offset(h
, vma
, addr
);
1344 chg
= region_chg(resv
, idx
, idx
+ 1);
1346 if (vma
->vm_flags
& VM_MAYSHARE
)
1349 return chg
< 0 ? chg
: 0;
1351 static void vma_commit_reservation(struct hstate
*h
,
1352 struct vm_area_struct
*vma
, unsigned long addr
)
1354 struct resv_map
*resv
;
1357 resv
= vma_resv_map(vma
);
1361 idx
= vma_hugecache_offset(h
, vma
, addr
);
1362 region_add(resv
, idx
, idx
+ 1);
1365 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1366 unsigned long addr
, int avoid_reserve
)
1368 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1369 struct hstate
*h
= hstate_vma(vma
);
1373 struct hugetlb_cgroup
*h_cg
;
1375 idx
= hstate_index(h
);
1377 * Processes that did not create the mapping will have no
1378 * reserves and will not have accounted against subpool
1379 * limit. Check that the subpool limit can be made before
1380 * satisfying the allocation MAP_NORESERVE mappings may also
1381 * need pages and subpool limit allocated allocated if no reserve
1384 chg
= vma_needs_reservation(h
, vma
, addr
);
1386 return ERR_PTR(-ENOMEM
);
1387 if (chg
|| avoid_reserve
)
1388 if (hugepage_subpool_get_pages(spool
, 1))
1389 return ERR_PTR(-ENOSPC
);
1391 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1393 goto out_subpool_put
;
1395 spin_lock(&hugetlb_lock
);
1396 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1398 spin_unlock(&hugetlb_lock
);
1399 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1401 goto out_uncharge_cgroup
;
1403 spin_lock(&hugetlb_lock
);
1404 list_move(&page
->lru
, &h
->hugepage_activelist
);
1407 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1408 spin_unlock(&hugetlb_lock
);
1410 set_page_private(page
, (unsigned long)spool
);
1412 vma_commit_reservation(h
, vma
, addr
);
1415 out_uncharge_cgroup
:
1416 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
1418 if (chg
|| avoid_reserve
)
1419 hugepage_subpool_put_pages(spool
, 1);
1420 return ERR_PTR(-ENOSPC
);
1424 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1425 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1426 * where no ERR_VALUE is expected to be returned.
1428 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1429 unsigned long addr
, int avoid_reserve
)
1431 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1437 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1439 struct huge_bootmem_page
*m
;
1442 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1445 addr
= memblock_virt_alloc_try_nid_nopanic(
1446 huge_page_size(h
), huge_page_size(h
),
1447 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1450 * Use the beginning of the huge page to store the
1451 * huge_bootmem_page struct (until gather_bootmem
1452 * puts them into the mem_map).
1461 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
1462 /* Put them into a private list first because mem_map is not up yet */
1463 list_add(&m
->list
, &huge_boot_pages
);
1468 static void __init
prep_compound_huge_page(struct page
*page
, int order
)
1470 if (unlikely(order
> (MAX_ORDER
- 1)))
1471 prep_compound_gigantic_page(page
, order
);
1473 prep_compound_page(page
, order
);
1476 /* Put bootmem huge pages into the standard lists after mem_map is up */
1477 static void __init
gather_bootmem_prealloc(void)
1479 struct huge_bootmem_page
*m
;
1481 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1482 struct hstate
*h
= m
->hstate
;
1485 #ifdef CONFIG_HIGHMEM
1486 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1487 memblock_free_late(__pa(m
),
1488 sizeof(struct huge_bootmem_page
));
1490 page
= virt_to_page(m
);
1492 WARN_ON(page_count(page
) != 1);
1493 prep_compound_huge_page(page
, h
->order
);
1494 WARN_ON(PageReserved(page
));
1495 prep_new_huge_page(h
, page
, page_to_nid(page
));
1497 * If we had gigantic hugepages allocated at boot time, we need
1498 * to restore the 'stolen' pages to totalram_pages in order to
1499 * fix confusing memory reports from free(1) and another
1500 * side-effects, like CommitLimit going negative.
1502 if (hstate_is_gigantic(h
))
1503 adjust_managed_page_count(page
, 1 << h
->order
);
1507 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1511 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1512 if (hstate_is_gigantic(h
)) {
1513 if (!alloc_bootmem_huge_page(h
))
1515 } else if (!alloc_fresh_huge_page(h
,
1516 &node_states
[N_MEMORY
]))
1519 h
->max_huge_pages
= i
;
1522 static void __init
hugetlb_init_hstates(void)
1526 for_each_hstate(h
) {
1527 /* oversize hugepages were init'ed in early boot */
1528 if (!hstate_is_gigantic(h
))
1529 hugetlb_hstate_alloc_pages(h
);
1533 static char * __init
memfmt(char *buf
, unsigned long n
)
1535 if (n
>= (1UL << 30))
1536 sprintf(buf
, "%lu GB", n
>> 30);
1537 else if (n
>= (1UL << 20))
1538 sprintf(buf
, "%lu MB", n
>> 20);
1540 sprintf(buf
, "%lu KB", n
>> 10);
1544 static void __init
report_hugepages(void)
1548 for_each_hstate(h
) {
1550 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1551 memfmt(buf
, huge_page_size(h
)),
1552 h
->free_huge_pages
);
1556 #ifdef CONFIG_HIGHMEM
1557 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1558 nodemask_t
*nodes_allowed
)
1562 if (hstate_is_gigantic(h
))
1565 for_each_node_mask(i
, *nodes_allowed
) {
1566 struct page
*page
, *next
;
1567 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1568 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1569 if (count
>= h
->nr_huge_pages
)
1571 if (PageHighMem(page
))
1573 list_del(&page
->lru
);
1574 update_and_free_page(h
, page
);
1575 h
->free_huge_pages
--;
1576 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1581 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1582 nodemask_t
*nodes_allowed
)
1588 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1589 * balanced by operating on them in a round-robin fashion.
1590 * Returns 1 if an adjustment was made.
1592 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1597 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1600 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1601 if (h
->surplus_huge_pages_node
[node
])
1605 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1606 if (h
->surplus_huge_pages_node
[node
] <
1607 h
->nr_huge_pages_node
[node
])
1614 h
->surplus_huge_pages
+= delta
;
1615 h
->surplus_huge_pages_node
[node
] += delta
;
1619 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1620 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1621 nodemask_t
*nodes_allowed
)
1623 unsigned long min_count
, ret
;
1625 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1626 return h
->max_huge_pages
;
1629 * Increase the pool size
1630 * First take pages out of surplus state. Then make up the
1631 * remaining difference by allocating fresh huge pages.
1633 * We might race with alloc_buddy_huge_page() here and be unable
1634 * to convert a surplus huge page to a normal huge page. That is
1635 * not critical, though, it just means the overall size of the
1636 * pool might be one hugepage larger than it needs to be, but
1637 * within all the constraints specified by the sysctls.
1639 spin_lock(&hugetlb_lock
);
1640 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1641 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1645 while (count
> persistent_huge_pages(h
)) {
1647 * If this allocation races such that we no longer need the
1648 * page, free_huge_page will handle it by freeing the page
1649 * and reducing the surplus.
1651 spin_unlock(&hugetlb_lock
);
1652 if (hstate_is_gigantic(h
))
1653 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
1655 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1656 spin_lock(&hugetlb_lock
);
1660 /* Bail for signals. Probably ctrl-c from user */
1661 if (signal_pending(current
))
1666 * Decrease the pool size
1667 * First return free pages to the buddy allocator (being careful
1668 * to keep enough around to satisfy reservations). Then place
1669 * pages into surplus state as needed so the pool will shrink
1670 * to the desired size as pages become free.
1672 * By placing pages into the surplus state independent of the
1673 * overcommit value, we are allowing the surplus pool size to
1674 * exceed overcommit. There are few sane options here. Since
1675 * alloc_buddy_huge_page() is checking the global counter,
1676 * though, we'll note that we're not allowed to exceed surplus
1677 * and won't grow the pool anywhere else. Not until one of the
1678 * sysctls are changed, or the surplus pages go out of use.
1680 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1681 min_count
= max(count
, min_count
);
1682 try_to_free_low(h
, min_count
, nodes_allowed
);
1683 while (min_count
< persistent_huge_pages(h
)) {
1684 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1686 cond_resched_lock(&hugetlb_lock
);
1688 while (count
< persistent_huge_pages(h
)) {
1689 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1693 ret
= persistent_huge_pages(h
);
1694 spin_unlock(&hugetlb_lock
);
1698 #define HSTATE_ATTR_RO(_name) \
1699 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1701 #define HSTATE_ATTR(_name) \
1702 static struct kobj_attribute _name##_attr = \
1703 __ATTR(_name, 0644, _name##_show, _name##_store)
1705 static struct kobject
*hugepages_kobj
;
1706 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1708 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1710 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1714 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1715 if (hstate_kobjs
[i
] == kobj
) {
1717 *nidp
= NUMA_NO_NODE
;
1721 return kobj_to_node_hstate(kobj
, nidp
);
1724 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1725 struct kobj_attribute
*attr
, char *buf
)
1728 unsigned long nr_huge_pages
;
1731 h
= kobj_to_hstate(kobj
, &nid
);
1732 if (nid
== NUMA_NO_NODE
)
1733 nr_huge_pages
= h
->nr_huge_pages
;
1735 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1737 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1740 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
1741 struct hstate
*h
, int nid
,
1742 unsigned long count
, size_t len
)
1745 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1747 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
1752 if (nid
== NUMA_NO_NODE
) {
1754 * global hstate attribute
1756 if (!(obey_mempolicy
&&
1757 init_nodemask_of_mempolicy(nodes_allowed
))) {
1758 NODEMASK_FREE(nodes_allowed
);
1759 nodes_allowed
= &node_states
[N_MEMORY
];
1761 } else if (nodes_allowed
) {
1763 * per node hstate attribute: adjust count to global,
1764 * but restrict alloc/free to the specified node.
1766 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1767 init_nodemask_of_node(nodes_allowed
, nid
);
1769 nodes_allowed
= &node_states
[N_MEMORY
];
1771 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1773 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1774 NODEMASK_FREE(nodes_allowed
);
1778 NODEMASK_FREE(nodes_allowed
);
1782 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1783 struct kobject
*kobj
, const char *buf
,
1787 unsigned long count
;
1791 err
= kstrtoul(buf
, 10, &count
);
1795 h
= kobj_to_hstate(kobj
, &nid
);
1796 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
1799 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1800 struct kobj_attribute
*attr
, char *buf
)
1802 return nr_hugepages_show_common(kobj
, attr
, buf
);
1805 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1806 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1808 return nr_hugepages_store_common(false, kobj
, buf
, len
);
1810 HSTATE_ATTR(nr_hugepages
);
1815 * hstate attribute for optionally mempolicy-based constraint on persistent
1816 * huge page alloc/free.
1818 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1819 struct kobj_attribute
*attr
, char *buf
)
1821 return nr_hugepages_show_common(kobj
, attr
, buf
);
1824 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1825 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1827 return nr_hugepages_store_common(true, kobj
, buf
, len
);
1829 HSTATE_ATTR(nr_hugepages_mempolicy
);
1833 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1834 struct kobj_attribute
*attr
, char *buf
)
1836 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1837 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1840 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1841 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1844 unsigned long input
;
1845 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1847 if (hstate_is_gigantic(h
))
1850 err
= kstrtoul(buf
, 10, &input
);
1854 spin_lock(&hugetlb_lock
);
1855 h
->nr_overcommit_huge_pages
= input
;
1856 spin_unlock(&hugetlb_lock
);
1860 HSTATE_ATTR(nr_overcommit_hugepages
);
1862 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1863 struct kobj_attribute
*attr
, char *buf
)
1866 unsigned long free_huge_pages
;
1869 h
= kobj_to_hstate(kobj
, &nid
);
1870 if (nid
== NUMA_NO_NODE
)
1871 free_huge_pages
= h
->free_huge_pages
;
1873 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1875 return sprintf(buf
, "%lu\n", free_huge_pages
);
1877 HSTATE_ATTR_RO(free_hugepages
);
1879 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1880 struct kobj_attribute
*attr
, char *buf
)
1882 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1883 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1885 HSTATE_ATTR_RO(resv_hugepages
);
1887 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1888 struct kobj_attribute
*attr
, char *buf
)
1891 unsigned long surplus_huge_pages
;
1894 h
= kobj_to_hstate(kobj
, &nid
);
1895 if (nid
== NUMA_NO_NODE
)
1896 surplus_huge_pages
= h
->surplus_huge_pages
;
1898 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1900 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1902 HSTATE_ATTR_RO(surplus_hugepages
);
1904 static struct attribute
*hstate_attrs
[] = {
1905 &nr_hugepages_attr
.attr
,
1906 &nr_overcommit_hugepages_attr
.attr
,
1907 &free_hugepages_attr
.attr
,
1908 &resv_hugepages_attr
.attr
,
1909 &surplus_hugepages_attr
.attr
,
1911 &nr_hugepages_mempolicy_attr
.attr
,
1916 static struct attribute_group hstate_attr_group
= {
1917 .attrs
= hstate_attrs
,
1920 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1921 struct kobject
**hstate_kobjs
,
1922 struct attribute_group
*hstate_attr_group
)
1925 int hi
= hstate_index(h
);
1927 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1928 if (!hstate_kobjs
[hi
])
1931 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1933 kobject_put(hstate_kobjs
[hi
]);
1938 static void __init
hugetlb_sysfs_init(void)
1943 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1944 if (!hugepages_kobj
)
1947 for_each_hstate(h
) {
1948 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1949 hstate_kobjs
, &hstate_attr_group
);
1951 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1958 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1959 * with node devices in node_devices[] using a parallel array. The array
1960 * index of a node device or _hstate == node id.
1961 * This is here to avoid any static dependency of the node device driver, in
1962 * the base kernel, on the hugetlb module.
1964 struct node_hstate
{
1965 struct kobject
*hugepages_kobj
;
1966 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1968 struct node_hstate node_hstates
[MAX_NUMNODES
];
1971 * A subset of global hstate attributes for node devices
1973 static struct attribute
*per_node_hstate_attrs
[] = {
1974 &nr_hugepages_attr
.attr
,
1975 &free_hugepages_attr
.attr
,
1976 &surplus_hugepages_attr
.attr
,
1980 static struct attribute_group per_node_hstate_attr_group
= {
1981 .attrs
= per_node_hstate_attrs
,
1985 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1986 * Returns node id via non-NULL nidp.
1988 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1992 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1993 struct node_hstate
*nhs
= &node_hstates
[nid
];
1995 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1996 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2008 * Unregister hstate attributes from a single node device.
2009 * No-op if no hstate attributes attached.
2011 static void hugetlb_unregister_node(struct node
*node
)
2014 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2016 if (!nhs
->hugepages_kobj
)
2017 return; /* no hstate attributes */
2019 for_each_hstate(h
) {
2020 int idx
= hstate_index(h
);
2021 if (nhs
->hstate_kobjs
[idx
]) {
2022 kobject_put(nhs
->hstate_kobjs
[idx
]);
2023 nhs
->hstate_kobjs
[idx
] = NULL
;
2027 kobject_put(nhs
->hugepages_kobj
);
2028 nhs
->hugepages_kobj
= NULL
;
2032 * hugetlb module exit: unregister hstate attributes from node devices
2035 static void hugetlb_unregister_all_nodes(void)
2040 * disable node device registrations.
2042 register_hugetlbfs_with_node(NULL
, NULL
);
2045 * remove hstate attributes from any nodes that have them.
2047 for (nid
= 0; nid
< nr_node_ids
; nid
++)
2048 hugetlb_unregister_node(node_devices
[nid
]);
2052 * Register hstate attributes for a single node device.
2053 * No-op if attributes already registered.
2055 static void hugetlb_register_node(struct node
*node
)
2058 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2061 if (nhs
->hugepages_kobj
)
2062 return; /* already allocated */
2064 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2066 if (!nhs
->hugepages_kobj
)
2069 for_each_hstate(h
) {
2070 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2072 &per_node_hstate_attr_group
);
2074 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2075 h
->name
, node
->dev
.id
);
2076 hugetlb_unregister_node(node
);
2083 * hugetlb init time: register hstate attributes for all registered node
2084 * devices of nodes that have memory. All on-line nodes should have
2085 * registered their associated device by this time.
2087 static void __init
hugetlb_register_all_nodes(void)
2091 for_each_node_state(nid
, N_MEMORY
) {
2092 struct node
*node
= node_devices
[nid
];
2093 if (node
->dev
.id
== nid
)
2094 hugetlb_register_node(node
);
2098 * Let the node device driver know we're here so it can
2099 * [un]register hstate attributes on node hotplug.
2101 register_hugetlbfs_with_node(hugetlb_register_node
,
2102 hugetlb_unregister_node
);
2104 #else /* !CONFIG_NUMA */
2106 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2114 static void hugetlb_unregister_all_nodes(void) { }
2116 static void hugetlb_register_all_nodes(void) { }
2120 static void __exit
hugetlb_exit(void)
2124 hugetlb_unregister_all_nodes();
2126 for_each_hstate(h
) {
2127 kobject_put(hstate_kobjs
[hstate_index(h
)]);
2130 kobject_put(hugepages_kobj
);
2131 kfree(htlb_fault_mutex_table
);
2133 module_exit(hugetlb_exit
);
2135 static int __init
hugetlb_init(void)
2139 if (!hugepages_supported())
2142 if (!size_to_hstate(default_hstate_size
)) {
2143 default_hstate_size
= HPAGE_SIZE
;
2144 if (!size_to_hstate(default_hstate_size
))
2145 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2147 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2148 if (default_hstate_max_huge_pages
)
2149 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2151 hugetlb_init_hstates();
2152 gather_bootmem_prealloc();
2155 hugetlb_sysfs_init();
2156 hugetlb_register_all_nodes();
2157 hugetlb_cgroup_file_init();
2160 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2162 num_fault_mutexes
= 1;
2164 htlb_fault_mutex_table
=
2165 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2166 BUG_ON(!htlb_fault_mutex_table
);
2168 for (i
= 0; i
< num_fault_mutexes
; i
++)
2169 mutex_init(&htlb_fault_mutex_table
[i
]);
2172 module_init(hugetlb_init
);
2174 /* Should be called on processing a hugepagesz=... option */
2175 void __init
hugetlb_add_hstate(unsigned order
)
2180 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2181 pr_warning("hugepagesz= specified twice, ignoring\n");
2184 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2186 h
= &hstates
[hugetlb_max_hstate
++];
2188 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2189 h
->nr_huge_pages
= 0;
2190 h
->free_huge_pages
= 0;
2191 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2192 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2193 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2194 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2195 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2196 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2197 huge_page_size(h
)/1024);
2202 static int __init
hugetlb_nrpages_setup(char *s
)
2205 static unsigned long *last_mhp
;
2208 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2209 * so this hugepages= parameter goes to the "default hstate".
2211 if (!hugetlb_max_hstate
)
2212 mhp
= &default_hstate_max_huge_pages
;
2214 mhp
= &parsed_hstate
->max_huge_pages
;
2216 if (mhp
== last_mhp
) {
2217 pr_warning("hugepages= specified twice without "
2218 "interleaving hugepagesz=, ignoring\n");
2222 if (sscanf(s
, "%lu", mhp
) <= 0)
2226 * Global state is always initialized later in hugetlb_init.
2227 * But we need to allocate >= MAX_ORDER hstates here early to still
2228 * use the bootmem allocator.
2230 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2231 hugetlb_hstate_alloc_pages(parsed_hstate
);
2237 __setup("hugepages=", hugetlb_nrpages_setup
);
2239 static int __init
hugetlb_default_setup(char *s
)
2241 default_hstate_size
= memparse(s
, &s
);
2244 __setup("default_hugepagesz=", hugetlb_default_setup
);
2246 static unsigned int cpuset_mems_nr(unsigned int *array
)
2249 unsigned int nr
= 0;
2251 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2257 #ifdef CONFIG_SYSCTL
2258 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2259 struct ctl_table
*table
, int write
,
2260 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2262 struct hstate
*h
= &default_hstate
;
2263 unsigned long tmp
= h
->max_huge_pages
;
2266 if (!hugepages_supported())
2270 table
->maxlen
= sizeof(unsigned long);
2271 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2276 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2277 NUMA_NO_NODE
, tmp
, *length
);
2282 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2283 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2286 return hugetlb_sysctl_handler_common(false, table
, write
,
2287 buffer
, length
, ppos
);
2291 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2292 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2294 return hugetlb_sysctl_handler_common(true, table
, write
,
2295 buffer
, length
, ppos
);
2297 #endif /* CONFIG_NUMA */
2299 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2300 void __user
*buffer
,
2301 size_t *length
, loff_t
*ppos
)
2303 struct hstate
*h
= &default_hstate
;
2307 if (!hugepages_supported())
2310 tmp
= h
->nr_overcommit_huge_pages
;
2312 if (write
&& hstate_is_gigantic(h
))
2316 table
->maxlen
= sizeof(unsigned long);
2317 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2322 spin_lock(&hugetlb_lock
);
2323 h
->nr_overcommit_huge_pages
= tmp
;
2324 spin_unlock(&hugetlb_lock
);
2330 #endif /* CONFIG_SYSCTL */
2332 void hugetlb_report_meminfo(struct seq_file
*m
)
2334 struct hstate
*h
= &default_hstate
;
2335 if (!hugepages_supported())
2338 "HugePages_Total: %5lu\n"
2339 "HugePages_Free: %5lu\n"
2340 "HugePages_Rsvd: %5lu\n"
2341 "HugePages_Surp: %5lu\n"
2342 "Hugepagesize: %8lu kB\n",
2346 h
->surplus_huge_pages
,
2347 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2350 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2352 struct hstate
*h
= &default_hstate
;
2353 if (!hugepages_supported())
2356 "Node %d HugePages_Total: %5u\n"
2357 "Node %d HugePages_Free: %5u\n"
2358 "Node %d HugePages_Surp: %5u\n",
2359 nid
, h
->nr_huge_pages_node
[nid
],
2360 nid
, h
->free_huge_pages_node
[nid
],
2361 nid
, h
->surplus_huge_pages_node
[nid
]);
2364 void hugetlb_show_meminfo(void)
2369 if (!hugepages_supported())
2372 for_each_node_state(nid
, N_MEMORY
)
2374 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2376 h
->nr_huge_pages_node
[nid
],
2377 h
->free_huge_pages_node
[nid
],
2378 h
->surplus_huge_pages_node
[nid
],
2379 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2382 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2383 unsigned long hugetlb_total_pages(void)
2386 unsigned long nr_total_pages
= 0;
2389 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2390 return nr_total_pages
;
2393 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2397 spin_lock(&hugetlb_lock
);
2399 * When cpuset is configured, it breaks the strict hugetlb page
2400 * reservation as the accounting is done on a global variable. Such
2401 * reservation is completely rubbish in the presence of cpuset because
2402 * the reservation is not checked against page availability for the
2403 * current cpuset. Application can still potentially OOM'ed by kernel
2404 * with lack of free htlb page in cpuset that the task is in.
2405 * Attempt to enforce strict accounting with cpuset is almost
2406 * impossible (or too ugly) because cpuset is too fluid that
2407 * task or memory node can be dynamically moved between cpusets.
2409 * The change of semantics for shared hugetlb mapping with cpuset is
2410 * undesirable. However, in order to preserve some of the semantics,
2411 * we fall back to check against current free page availability as
2412 * a best attempt and hopefully to minimize the impact of changing
2413 * semantics that cpuset has.
2416 if (gather_surplus_pages(h
, delta
) < 0)
2419 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2420 return_unused_surplus_pages(h
, delta
);
2427 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2430 spin_unlock(&hugetlb_lock
);
2434 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2436 struct resv_map
*resv
= vma_resv_map(vma
);
2439 * This new VMA should share its siblings reservation map if present.
2440 * The VMA will only ever have a valid reservation map pointer where
2441 * it is being copied for another still existing VMA. As that VMA
2442 * has a reference to the reservation map it cannot disappear until
2443 * after this open call completes. It is therefore safe to take a
2444 * new reference here without additional locking.
2446 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2447 kref_get(&resv
->refs
);
2450 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2452 struct hstate
*h
= hstate_vma(vma
);
2453 struct resv_map
*resv
= vma_resv_map(vma
);
2454 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2455 unsigned long reserve
, start
, end
;
2457 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2460 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2461 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2463 reserve
= (end
- start
) - region_count(resv
, start
, end
);
2465 kref_put(&resv
->refs
, resv_map_release
);
2468 hugetlb_acct_memory(h
, -reserve
);
2469 hugepage_subpool_put_pages(spool
, reserve
);
2474 * We cannot handle pagefaults against hugetlb pages at all. They cause
2475 * handle_mm_fault() to try to instantiate regular-sized pages in the
2476 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2479 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2485 const struct vm_operations_struct hugetlb_vm_ops
= {
2486 .fault
= hugetlb_vm_op_fault
,
2487 .open
= hugetlb_vm_op_open
,
2488 .close
= hugetlb_vm_op_close
,
2491 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2497 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2498 vma
->vm_page_prot
)));
2500 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2501 vma
->vm_page_prot
));
2503 entry
= pte_mkyoung(entry
);
2504 entry
= pte_mkhuge(entry
);
2505 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2510 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2511 unsigned long address
, pte_t
*ptep
)
2515 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2516 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2517 update_mmu_cache(vma
, address
, ptep
);
2520 static int is_hugetlb_entry_migration(pte_t pte
)
2524 if (huge_pte_none(pte
) || pte_present(pte
))
2526 swp
= pte_to_swp_entry(pte
);
2527 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2533 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2537 if (huge_pte_none(pte
) || pte_present(pte
))
2539 swp
= pte_to_swp_entry(pte
);
2540 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2546 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2547 struct vm_area_struct
*vma
)
2549 pte_t
*src_pte
, *dst_pte
, entry
;
2550 struct page
*ptepage
;
2553 struct hstate
*h
= hstate_vma(vma
);
2554 unsigned long sz
= huge_page_size(h
);
2555 unsigned long mmun_start
; /* For mmu_notifiers */
2556 unsigned long mmun_end
; /* For mmu_notifiers */
2559 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2561 mmun_start
= vma
->vm_start
;
2562 mmun_end
= vma
->vm_end
;
2564 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
2566 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2567 spinlock_t
*src_ptl
, *dst_ptl
;
2568 src_pte
= huge_pte_offset(src
, addr
);
2571 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2577 /* If the pagetables are shared don't copy or take references */
2578 if (dst_pte
== src_pte
)
2581 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
2582 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
2583 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
2584 entry
= huge_ptep_get(src_pte
);
2585 if (huge_pte_none(entry
)) { /* skip none entry */
2587 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
2588 is_hugetlb_entry_hwpoisoned(entry
))) {
2589 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
2591 if (is_write_migration_entry(swp_entry
) && cow
) {
2593 * COW mappings require pages in both
2594 * parent and child to be set to read.
2596 make_migration_entry_read(&swp_entry
);
2597 entry
= swp_entry_to_pte(swp_entry
);
2598 set_huge_pte_at(src
, addr
, src_pte
, entry
);
2600 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2603 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2604 mmu_notifier_invalidate_range(src
, mmun_start
,
2607 entry
= huge_ptep_get(src_pte
);
2608 ptepage
= pte_page(entry
);
2610 page_dup_rmap(ptepage
);
2611 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2613 spin_unlock(src_ptl
);
2614 spin_unlock(dst_ptl
);
2618 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
2623 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2624 unsigned long start
, unsigned long end
,
2625 struct page
*ref_page
)
2627 int force_flush
= 0;
2628 struct mm_struct
*mm
= vma
->vm_mm
;
2629 unsigned long address
;
2634 struct hstate
*h
= hstate_vma(vma
);
2635 unsigned long sz
= huge_page_size(h
);
2636 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2637 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2639 WARN_ON(!is_vm_hugetlb_page(vma
));
2640 BUG_ON(start
& ~huge_page_mask(h
));
2641 BUG_ON(end
& ~huge_page_mask(h
));
2643 tlb_start_vma(tlb
, vma
);
2644 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2647 for (; address
< end
; address
+= sz
) {
2648 ptep
= huge_pte_offset(mm
, address
);
2652 ptl
= huge_pte_lock(h
, mm
, ptep
);
2653 if (huge_pmd_unshare(mm
, &address
, ptep
))
2656 pte
= huge_ptep_get(ptep
);
2657 if (huge_pte_none(pte
))
2661 * Migrating hugepage or HWPoisoned hugepage is already
2662 * unmapped and its refcount is dropped, so just clear pte here.
2664 if (unlikely(!pte_present(pte
))) {
2665 huge_pte_clear(mm
, address
, ptep
);
2669 page
= pte_page(pte
);
2671 * If a reference page is supplied, it is because a specific
2672 * page is being unmapped, not a range. Ensure the page we
2673 * are about to unmap is the actual page of interest.
2676 if (page
!= ref_page
)
2680 * Mark the VMA as having unmapped its page so that
2681 * future faults in this VMA will fail rather than
2682 * looking like data was lost
2684 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2687 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2688 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2689 if (huge_pte_dirty(pte
))
2690 set_page_dirty(page
);
2692 page_remove_rmap(page
);
2693 force_flush
= !__tlb_remove_page(tlb
, page
);
2699 /* Bail out after unmapping reference page if supplied */
2708 * mmu_gather ran out of room to batch pages, we break out of
2709 * the PTE lock to avoid doing the potential expensive TLB invalidate
2710 * and page-free while holding it.
2715 if (address
< end
&& !ref_page
)
2718 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2719 tlb_end_vma(tlb
, vma
);
2722 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2723 struct vm_area_struct
*vma
, unsigned long start
,
2724 unsigned long end
, struct page
*ref_page
)
2726 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2729 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2730 * test will fail on a vma being torn down, and not grab a page table
2731 * on its way out. We're lucky that the flag has such an appropriate
2732 * name, and can in fact be safely cleared here. We could clear it
2733 * before the __unmap_hugepage_range above, but all that's necessary
2734 * is to clear it before releasing the i_mmap_rwsem. This works
2735 * because in the context this is called, the VMA is about to be
2736 * destroyed and the i_mmap_rwsem is held.
2738 vma
->vm_flags
&= ~VM_MAYSHARE
;
2741 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2742 unsigned long end
, struct page
*ref_page
)
2744 struct mm_struct
*mm
;
2745 struct mmu_gather tlb
;
2749 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2750 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2751 tlb_finish_mmu(&tlb
, start
, end
);
2755 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2756 * mappping it owns the reserve page for. The intention is to unmap the page
2757 * from other VMAs and let the children be SIGKILLed if they are faulting the
2760 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2761 struct page
*page
, unsigned long address
)
2763 struct hstate
*h
= hstate_vma(vma
);
2764 struct vm_area_struct
*iter_vma
;
2765 struct address_space
*mapping
;
2769 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2770 * from page cache lookup which is in HPAGE_SIZE units.
2772 address
= address
& huge_page_mask(h
);
2773 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2775 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2778 * Take the mapping lock for the duration of the table walk. As
2779 * this mapping should be shared between all the VMAs,
2780 * __unmap_hugepage_range() is called as the lock is already held
2782 i_mmap_lock_write(mapping
);
2783 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2784 /* Do not unmap the current VMA */
2785 if (iter_vma
== vma
)
2789 * Unmap the page from other VMAs without their own reserves.
2790 * They get marked to be SIGKILLed if they fault in these
2791 * areas. This is because a future no-page fault on this VMA
2792 * could insert a zeroed page instead of the data existing
2793 * from the time of fork. This would look like data corruption
2795 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2796 unmap_hugepage_range(iter_vma
, address
,
2797 address
+ huge_page_size(h
), page
);
2799 i_mmap_unlock_write(mapping
);
2803 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2804 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2805 * cannot race with other handlers or page migration.
2806 * Keep the pte_same checks anyway to make transition from the mutex easier.
2808 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2809 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2810 struct page
*pagecache_page
, spinlock_t
*ptl
)
2812 struct hstate
*h
= hstate_vma(vma
);
2813 struct page
*old_page
, *new_page
;
2814 int ret
= 0, outside_reserve
= 0;
2815 unsigned long mmun_start
; /* For mmu_notifiers */
2816 unsigned long mmun_end
; /* For mmu_notifiers */
2818 old_page
= pte_page(pte
);
2821 /* If no-one else is actually using this page, avoid the copy
2822 * and just make the page writable */
2823 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2824 page_move_anon_rmap(old_page
, vma
, address
);
2825 set_huge_ptep_writable(vma
, address
, ptep
);
2830 * If the process that created a MAP_PRIVATE mapping is about to
2831 * perform a COW due to a shared page count, attempt to satisfy
2832 * the allocation without using the existing reserves. The pagecache
2833 * page is used to determine if the reserve at this address was
2834 * consumed or not. If reserves were used, a partial faulted mapping
2835 * at the time of fork() could consume its reserves on COW instead
2836 * of the full address range.
2838 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2839 old_page
!= pagecache_page
)
2840 outside_reserve
= 1;
2842 page_cache_get(old_page
);
2845 * Drop page table lock as buddy allocator may be called. It will
2846 * be acquired again before returning to the caller, as expected.
2849 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2851 if (IS_ERR(new_page
)) {
2853 * If a process owning a MAP_PRIVATE mapping fails to COW,
2854 * it is due to references held by a child and an insufficient
2855 * huge page pool. To guarantee the original mappers
2856 * reliability, unmap the page from child processes. The child
2857 * may get SIGKILLed if it later faults.
2859 if (outside_reserve
) {
2860 page_cache_release(old_page
);
2861 BUG_ON(huge_pte_none(pte
));
2862 unmap_ref_private(mm
, vma
, old_page
, address
);
2863 BUG_ON(huge_pte_none(pte
));
2865 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2867 pte_same(huge_ptep_get(ptep
), pte
)))
2868 goto retry_avoidcopy
;
2870 * race occurs while re-acquiring page table
2871 * lock, and our job is done.
2876 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
2877 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
2878 goto out_release_old
;
2882 * When the original hugepage is shared one, it does not have
2883 * anon_vma prepared.
2885 if (unlikely(anon_vma_prepare(vma
))) {
2887 goto out_release_all
;
2890 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2891 pages_per_huge_page(h
));
2892 __SetPageUptodate(new_page
);
2894 mmun_start
= address
& huge_page_mask(h
);
2895 mmun_end
= mmun_start
+ huge_page_size(h
);
2896 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2899 * Retake the page table lock to check for racing updates
2900 * before the page tables are altered
2903 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2904 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
2905 ClearPagePrivate(new_page
);
2908 huge_ptep_clear_flush(vma
, address
, ptep
);
2909 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
2910 set_huge_pte_at(mm
, address
, ptep
,
2911 make_huge_pte(vma
, new_page
, 1));
2912 page_remove_rmap(old_page
);
2913 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2914 /* Make the old page be freed below */
2915 new_page
= old_page
;
2918 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2920 page_cache_release(new_page
);
2922 page_cache_release(old_page
);
2924 spin_lock(ptl
); /* Caller expects lock to be held */
2928 /* Return the pagecache page at a given address within a VMA */
2929 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2930 struct vm_area_struct
*vma
, unsigned long address
)
2932 struct address_space
*mapping
;
2935 mapping
= vma
->vm_file
->f_mapping
;
2936 idx
= vma_hugecache_offset(h
, vma
, address
);
2938 return find_lock_page(mapping
, idx
);
2942 * Return whether there is a pagecache page to back given address within VMA.
2943 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2945 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2946 struct vm_area_struct
*vma
, unsigned long address
)
2948 struct address_space
*mapping
;
2952 mapping
= vma
->vm_file
->f_mapping
;
2953 idx
= vma_hugecache_offset(h
, vma
, address
);
2955 page
= find_get_page(mapping
, idx
);
2958 return page
!= NULL
;
2961 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2962 struct address_space
*mapping
, pgoff_t idx
,
2963 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2965 struct hstate
*h
= hstate_vma(vma
);
2966 int ret
= VM_FAULT_SIGBUS
;
2974 * Currently, we are forced to kill the process in the event the
2975 * original mapper has unmapped pages from the child due to a failed
2976 * COW. Warn that such a situation has occurred as it may not be obvious
2978 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2979 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2985 * Use page lock to guard against racing truncation
2986 * before we get page_table_lock.
2989 page
= find_lock_page(mapping
, idx
);
2991 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2994 page
= alloc_huge_page(vma
, address
, 0);
2996 ret
= PTR_ERR(page
);
3000 ret
= VM_FAULT_SIGBUS
;
3003 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3004 __SetPageUptodate(page
);
3006 if (vma
->vm_flags
& VM_MAYSHARE
) {
3008 struct inode
*inode
= mapping
->host
;
3010 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3017 ClearPagePrivate(page
);
3019 spin_lock(&inode
->i_lock
);
3020 inode
->i_blocks
+= blocks_per_huge_page(h
);
3021 spin_unlock(&inode
->i_lock
);
3024 if (unlikely(anon_vma_prepare(vma
))) {
3026 goto backout_unlocked
;
3032 * If memory error occurs between mmap() and fault, some process
3033 * don't have hwpoisoned swap entry for errored virtual address.
3034 * So we need to block hugepage fault by PG_hwpoison bit check.
3036 if (unlikely(PageHWPoison(page
))) {
3037 ret
= VM_FAULT_HWPOISON
|
3038 VM_FAULT_SET_HINDEX(hstate_index(h
));
3039 goto backout_unlocked
;
3044 * If we are going to COW a private mapping later, we examine the
3045 * pending reservations for this page now. This will ensure that
3046 * any allocations necessary to record that reservation occur outside
3049 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
3050 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3052 goto backout_unlocked
;
3055 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3057 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3062 if (!huge_pte_none(huge_ptep_get(ptep
)))
3066 ClearPagePrivate(page
);
3067 hugepage_add_new_anon_rmap(page
, vma
, address
);
3069 page_dup_rmap(page
);
3070 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3071 && (vma
->vm_flags
& VM_SHARED
)));
3072 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3074 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3075 /* Optimization, do the COW without a second fault */
3076 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
3093 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3094 struct vm_area_struct
*vma
,
3095 struct address_space
*mapping
,
3096 pgoff_t idx
, unsigned long address
)
3098 unsigned long key
[2];
3101 if (vma
->vm_flags
& VM_SHARED
) {
3102 key
[0] = (unsigned long) mapping
;
3105 key
[0] = (unsigned long) mm
;
3106 key
[1] = address
>> huge_page_shift(h
);
3109 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3111 return hash
& (num_fault_mutexes
- 1);
3115 * For uniprocesor systems we always use a single mutex, so just
3116 * return 0 and avoid the hashing overhead.
3118 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3119 struct vm_area_struct
*vma
,
3120 struct address_space
*mapping
,
3121 pgoff_t idx
, unsigned long address
)
3127 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3128 unsigned long address
, unsigned int flags
)
3135 struct page
*page
= NULL
;
3136 struct page
*pagecache_page
= NULL
;
3137 struct hstate
*h
= hstate_vma(vma
);
3138 struct address_space
*mapping
;
3139 int need_wait_lock
= 0;
3141 address
&= huge_page_mask(h
);
3143 ptep
= huge_pte_offset(mm
, address
);
3145 entry
= huge_ptep_get(ptep
);
3146 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3147 migration_entry_wait_huge(vma
, mm
, ptep
);
3149 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3150 return VM_FAULT_HWPOISON_LARGE
|
3151 VM_FAULT_SET_HINDEX(hstate_index(h
));
3154 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3156 return VM_FAULT_OOM
;
3158 mapping
= vma
->vm_file
->f_mapping
;
3159 idx
= vma_hugecache_offset(h
, vma
, address
);
3162 * Serialize hugepage allocation and instantiation, so that we don't
3163 * get spurious allocation failures if two CPUs race to instantiate
3164 * the same page in the page cache.
3166 hash
= fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3167 mutex_lock(&htlb_fault_mutex_table
[hash
]);
3169 entry
= huge_ptep_get(ptep
);
3170 if (huge_pte_none(entry
)) {
3171 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3178 * entry could be a migration/hwpoison entry at this point, so this
3179 * check prevents the kernel from going below assuming that we have
3180 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3181 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3184 if (!pte_present(entry
))
3188 * If we are going to COW the mapping later, we examine the pending
3189 * reservations for this page now. This will ensure that any
3190 * allocations necessary to record that reservation occur outside the
3191 * spinlock. For private mappings, we also lookup the pagecache
3192 * page now as it is used to determine if a reservation has been
3195 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3196 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3201 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3202 pagecache_page
= hugetlbfs_pagecache_page(h
,
3206 ptl
= huge_pte_lock(h
, mm
, ptep
);
3208 /* Check for a racing update before calling hugetlb_cow */
3209 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3213 * hugetlb_cow() requires page locks of pte_page(entry) and
3214 * pagecache_page, so here we need take the former one
3215 * when page != pagecache_page or !pagecache_page.
3217 page
= pte_page(entry
);
3218 if (page
!= pagecache_page
)
3219 if (!trylock_page(page
)) {
3226 if (flags
& FAULT_FLAG_WRITE
) {
3227 if (!huge_pte_write(entry
)) {
3228 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3229 pagecache_page
, ptl
);
3232 entry
= huge_pte_mkdirty(entry
);
3234 entry
= pte_mkyoung(entry
);
3235 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3236 flags
& FAULT_FLAG_WRITE
))
3237 update_mmu_cache(vma
, address
, ptep
);
3239 if (page
!= pagecache_page
)
3245 if (pagecache_page
) {
3246 unlock_page(pagecache_page
);
3247 put_page(pagecache_page
);
3250 mutex_unlock(&htlb_fault_mutex_table
[hash
]);
3252 * Generally it's safe to hold refcount during waiting page lock. But
3253 * here we just wait to defer the next page fault to avoid busy loop and
3254 * the page is not used after unlocked before returning from the current
3255 * page fault. So we are safe from accessing freed page, even if we wait
3256 * here without taking refcount.
3259 wait_on_page_locked(page
);
3263 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3264 struct page
**pages
, struct vm_area_struct
**vmas
,
3265 unsigned long *position
, unsigned long *nr_pages
,
3266 long i
, unsigned int flags
)
3268 unsigned long pfn_offset
;
3269 unsigned long vaddr
= *position
;
3270 unsigned long remainder
= *nr_pages
;
3271 struct hstate
*h
= hstate_vma(vma
);
3273 while (vaddr
< vma
->vm_end
&& remainder
) {
3275 spinlock_t
*ptl
= NULL
;
3280 * If we have a pending SIGKILL, don't keep faulting pages and
3281 * potentially allocating memory.
3283 if (unlikely(fatal_signal_pending(current
))) {
3289 * Some archs (sparc64, sh*) have multiple pte_ts to
3290 * each hugepage. We have to make sure we get the
3291 * first, for the page indexing below to work.
3293 * Note that page table lock is not held when pte is null.
3295 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3297 ptl
= huge_pte_lock(h
, mm
, pte
);
3298 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3301 * When coredumping, it suits get_dump_page if we just return
3302 * an error where there's an empty slot with no huge pagecache
3303 * to back it. This way, we avoid allocating a hugepage, and
3304 * the sparse dumpfile avoids allocating disk blocks, but its
3305 * huge holes still show up with zeroes where they need to be.
3307 if (absent
&& (flags
& FOLL_DUMP
) &&
3308 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3316 * We need call hugetlb_fault for both hugepages under migration
3317 * (in which case hugetlb_fault waits for the migration,) and
3318 * hwpoisoned hugepages (in which case we need to prevent the
3319 * caller from accessing to them.) In order to do this, we use
3320 * here is_swap_pte instead of is_hugetlb_entry_migration and
3321 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3322 * both cases, and because we can't follow correct pages
3323 * directly from any kind of swap entries.
3325 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3326 ((flags
& FOLL_WRITE
) &&
3327 !huge_pte_write(huge_ptep_get(pte
)))) {
3332 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3333 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3334 if (!(ret
& VM_FAULT_ERROR
))
3341 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3342 page
= pte_page(huge_ptep_get(pte
));
3345 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3346 get_page_foll(pages
[i
]);
3356 if (vaddr
< vma
->vm_end
&& remainder
&&
3357 pfn_offset
< pages_per_huge_page(h
)) {
3359 * We use pfn_offset to avoid touching the pageframes
3360 * of this compound page.
3366 *nr_pages
= remainder
;
3369 return i
? i
: -EFAULT
;
3372 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3373 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3375 struct mm_struct
*mm
= vma
->vm_mm
;
3376 unsigned long start
= address
;
3379 struct hstate
*h
= hstate_vma(vma
);
3380 unsigned long pages
= 0;
3382 BUG_ON(address
>= end
);
3383 flush_cache_range(vma
, address
, end
);
3385 mmu_notifier_invalidate_range_start(mm
, start
, end
);
3386 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
3387 for (; address
< end
; address
+= huge_page_size(h
)) {
3389 ptep
= huge_pte_offset(mm
, address
);
3392 ptl
= huge_pte_lock(h
, mm
, ptep
);
3393 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3398 pte
= huge_ptep_get(ptep
);
3399 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
3403 if (unlikely(is_hugetlb_entry_migration(pte
))) {
3404 swp_entry_t entry
= pte_to_swp_entry(pte
);
3406 if (is_write_migration_entry(entry
)) {
3409 make_migration_entry_read(&entry
);
3410 newpte
= swp_entry_to_pte(entry
);
3411 set_huge_pte_at(mm
, address
, ptep
, newpte
);
3417 if (!huge_pte_none(pte
)) {
3418 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3419 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3420 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3421 set_huge_pte_at(mm
, address
, ptep
, pte
);
3427 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3428 * may have cleared our pud entry and done put_page on the page table:
3429 * once we release i_mmap_rwsem, another task can do the final put_page
3430 * and that page table be reused and filled with junk.
3432 flush_tlb_range(vma
, start
, end
);
3433 mmu_notifier_invalidate_range(mm
, start
, end
);
3434 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
3435 mmu_notifier_invalidate_range_end(mm
, start
, end
);
3437 return pages
<< h
->order
;
3440 int hugetlb_reserve_pages(struct inode
*inode
,
3442 struct vm_area_struct
*vma
,
3443 vm_flags_t vm_flags
)
3446 struct hstate
*h
= hstate_inode(inode
);
3447 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3448 struct resv_map
*resv_map
;
3451 * Only apply hugepage reservation if asked. At fault time, an
3452 * attempt will be made for VM_NORESERVE to allocate a page
3453 * without using reserves
3455 if (vm_flags
& VM_NORESERVE
)
3459 * Shared mappings base their reservation on the number of pages that
3460 * are already allocated on behalf of the file. Private mappings need
3461 * to reserve the full area even if read-only as mprotect() may be
3462 * called to make the mapping read-write. Assume !vma is a shm mapping
3464 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
3465 resv_map
= inode_resv_map(inode
);
3467 chg
= region_chg(resv_map
, from
, to
);
3470 resv_map
= resv_map_alloc();
3476 set_vma_resv_map(vma
, resv_map
);
3477 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3485 /* There must be enough pages in the subpool for the mapping */
3486 if (hugepage_subpool_get_pages(spool
, chg
)) {
3492 * Check enough hugepages are available for the reservation.
3493 * Hand the pages back to the subpool if there are not
3495 ret
= hugetlb_acct_memory(h
, chg
);
3497 hugepage_subpool_put_pages(spool
, chg
);
3502 * Account for the reservations made. Shared mappings record regions
3503 * that have reservations as they are shared by multiple VMAs.
3504 * When the last VMA disappears, the region map says how much
3505 * the reservation was and the page cache tells how much of
3506 * the reservation was consumed. Private mappings are per-VMA and
3507 * only the consumed reservations are tracked. When the VMA
3508 * disappears, the original reservation is the VMA size and the
3509 * consumed reservations are stored in the map. Hence, nothing
3510 * else has to be done for private mappings here
3512 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3513 region_add(resv_map
, from
, to
);
3516 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3517 kref_put(&resv_map
->refs
, resv_map_release
);
3521 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3523 struct hstate
*h
= hstate_inode(inode
);
3524 struct resv_map
*resv_map
= inode_resv_map(inode
);
3526 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3529 chg
= region_truncate(resv_map
, offset
);
3530 spin_lock(&inode
->i_lock
);
3531 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3532 spin_unlock(&inode
->i_lock
);
3534 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3535 hugetlb_acct_memory(h
, -(chg
- freed
));
3538 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3539 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3540 struct vm_area_struct
*vma
,
3541 unsigned long addr
, pgoff_t idx
)
3543 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3545 unsigned long sbase
= saddr
& PUD_MASK
;
3546 unsigned long s_end
= sbase
+ PUD_SIZE
;
3548 /* Allow segments to share if only one is marked locked */
3549 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3550 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3553 * match the virtual addresses, permission and the alignment of the
3556 if (pmd_index(addr
) != pmd_index(saddr
) ||
3557 vm_flags
!= svm_flags
||
3558 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3564 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3566 unsigned long base
= addr
& PUD_MASK
;
3567 unsigned long end
= base
+ PUD_SIZE
;
3570 * check on proper vm_flags and page table alignment
3572 if (vma
->vm_flags
& VM_MAYSHARE
&&
3573 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3579 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3580 * and returns the corresponding pte. While this is not necessary for the
3581 * !shared pmd case because we can allocate the pmd later as well, it makes the
3582 * code much cleaner. pmd allocation is essential for the shared case because
3583 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3584 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3585 * bad pmd for sharing.
3587 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3589 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3590 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3591 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3593 struct vm_area_struct
*svma
;
3594 unsigned long saddr
;
3599 if (!vma_shareable(vma
, addr
))
3600 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3602 i_mmap_lock_write(mapping
);
3603 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3607 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3609 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3612 get_page(virt_to_page(spte
));
3621 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
3623 if (pud_none(*pud
)) {
3624 pud_populate(mm
, pud
,
3625 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3627 put_page(virt_to_page(spte
));
3632 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3633 i_mmap_unlock_write(mapping
);
3638 * unmap huge page backed by shared pte.
3640 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3641 * indicated by page_count > 1, unmap is achieved by clearing pud and
3642 * decrementing the ref count. If count == 1, the pte page is not shared.
3644 * called with page table lock held.
3646 * returns: 1 successfully unmapped a shared pte page
3647 * 0 the underlying pte page is not shared, or it is the last user
3649 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3651 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3652 pud_t
*pud
= pud_offset(pgd
, *addr
);
3654 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3655 if (page_count(virt_to_page(ptep
)) == 1)
3659 put_page(virt_to_page(ptep
));
3661 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3664 #define want_pmd_share() (1)
3665 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3666 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3670 #define want_pmd_share() (0)
3671 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3673 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3674 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3675 unsigned long addr
, unsigned long sz
)
3681 pgd
= pgd_offset(mm
, addr
);
3682 pud
= pud_alloc(mm
, pgd
, addr
);
3684 if (sz
== PUD_SIZE
) {
3687 BUG_ON(sz
!= PMD_SIZE
);
3688 if (want_pmd_share() && pud_none(*pud
))
3689 pte
= huge_pmd_share(mm
, addr
, pud
);
3691 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3694 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3699 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3705 pgd
= pgd_offset(mm
, addr
);
3706 if (pgd_present(*pgd
)) {
3707 pud
= pud_offset(pgd
, addr
);
3708 if (pud_present(*pud
)) {
3710 return (pte_t
*)pud
;
3711 pmd
= pmd_offset(pud
, addr
);
3714 return (pte_t
*) pmd
;
3717 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3720 * These functions are overwritable if your architecture needs its own
3723 struct page
* __weak
3724 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
3727 return ERR_PTR(-EINVAL
);
3730 struct page
* __weak
3731 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3732 pmd_t
*pmd
, int flags
)
3734 struct page
*page
= NULL
;
3737 ptl
= pmd_lockptr(mm
, pmd
);
3740 * make sure that the address range covered by this pmd is not
3741 * unmapped from other threads.
3743 if (!pmd_huge(*pmd
))
3745 if (pmd_present(*pmd
)) {
3746 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3747 if (flags
& FOLL_GET
)
3750 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t
*)pmd
))) {
3752 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
3756 * hwpoisoned entry is treated as no_page_table in
3757 * follow_page_mask().
3765 struct page
* __weak
3766 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3767 pud_t
*pud
, int flags
)
3769 if (flags
& FOLL_GET
)
3772 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3775 #ifdef CONFIG_MEMORY_FAILURE
3777 /* Should be called in hugetlb_lock */
3778 static int is_hugepage_on_freelist(struct page
*hpage
)
3782 struct hstate
*h
= page_hstate(hpage
);
3783 int nid
= page_to_nid(hpage
);
3785 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3792 * This function is called from memory failure code.
3793 * Assume the caller holds page lock of the head page.
3795 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3797 struct hstate
*h
= page_hstate(hpage
);
3798 int nid
= page_to_nid(hpage
);
3801 spin_lock(&hugetlb_lock
);
3802 if (is_hugepage_on_freelist(hpage
)) {
3804 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3805 * but dangling hpage->lru can trigger list-debug warnings
3806 * (this happens when we call unpoison_memory() on it),
3807 * so let it point to itself with list_del_init().
3809 list_del_init(&hpage
->lru
);
3810 set_page_refcounted(hpage
);
3811 h
->free_huge_pages
--;
3812 h
->free_huge_pages_node
[nid
]--;
3815 spin_unlock(&hugetlb_lock
);
3820 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3822 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3823 if (!get_page_unless_zero(page
))
3825 spin_lock(&hugetlb_lock
);
3826 list_move_tail(&page
->lru
, list
);
3827 spin_unlock(&hugetlb_lock
);
3831 void putback_active_hugepage(struct page
*page
)
3833 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3834 spin_lock(&hugetlb_lock
);
3835 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3836 spin_unlock(&hugetlb_lock
);
3840 bool is_hugepage_active(struct page
*page
)
3842 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
3844 * This function can be called for a tail page because the caller,
3845 * scan_movable_pages, scans through a given pfn-range which typically
3846 * covers one memory block. In systems using gigantic hugepage (1GB
3847 * for x86_64,) a hugepage is larger than a memory block, and we don't
3848 * support migrating such large hugepages for now, so return false
3849 * when called for tail pages.
3854 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3855 * so we should return false for them.
3857 if (unlikely(PageHWPoison(page
)))
3859 return page_count(page
) > 0;