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 unsigned long hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 __initdata
LIST_HEAD(huge_boot_pages
);
46 /* for command line parsing */
47 static struct hstate
* __initdata parsed_hstate
;
48 static unsigned long __initdata default_hstate_max_huge_pages
;
49 static unsigned long __initdata default_hstate_size
;
52 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
55 DEFINE_SPINLOCK(hugetlb_lock
);
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
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
84 spin_lock_init(&spool
->lock
);
86 spool
->max_hpages
= nr_blocks
;
87 spool
->used_hpages
= 0;
92 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
94 spin_lock(&spool
->lock
);
95 BUG_ON(!spool
->count
);
97 unlock_or_release_subpool(spool
);
100 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
108 spin_lock(&spool
->lock
);
109 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
110 spool
->used_hpages
+= delta
;
114 spin_unlock(&spool
->lock
);
119 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
125 spin_lock(&spool
->lock
);
126 spool
->used_hpages
-= delta
;
127 /* If hugetlbfs_put_super couldn't free spool due to
128 * an outstanding quota reference, free it now. */
129 unlock_or_release_subpool(spool
);
132 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
134 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
137 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
139 return subpool_inode(file_inode(vma
->vm_file
));
143 * Region tracking -- allows tracking of reservations and instantiated pages
144 * across the pages in a mapping.
146 * The region data structures are embedded into a resv_map and
147 * protected by a resv_map's lock
150 struct list_head link
;
155 static long region_add(struct resv_map
*resv
, long f
, long t
)
157 struct list_head
*head
= &resv
->regions
;
158 struct file_region
*rg
, *nrg
, *trg
;
160 spin_lock(&resv
->lock
);
161 /* Locate the region we are either in or before. */
162 list_for_each_entry(rg
, head
, link
)
166 /* Round our left edge to the current segment if it encloses us. */
170 /* Check for and consume any regions we now overlap with. */
172 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
173 if (&rg
->link
== head
)
178 /* If this area reaches higher then extend our area to
179 * include it completely. If this is not the first area
180 * which we intend to reuse, free it. */
190 spin_unlock(&resv
->lock
);
194 static long region_chg(struct resv_map
*resv
, long f
, long t
)
196 struct list_head
*head
= &resv
->regions
;
197 struct file_region
*rg
, *nrg
= NULL
;
201 spin_lock(&resv
->lock
);
202 /* Locate the region we are before or in. */
203 list_for_each_entry(rg
, head
, link
)
207 /* If we are below the current region then a new region is required.
208 * Subtle, allocate a new region at the position but make it zero
209 * size such that we can guarantee to record the reservation. */
210 if (&rg
->link
== head
|| t
< rg
->from
) {
212 spin_unlock(&resv
->lock
);
213 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
219 INIT_LIST_HEAD(&nrg
->link
);
223 list_add(&nrg
->link
, rg
->link
.prev
);
228 /* Round our left edge to the current segment if it encloses us. */
233 /* Check for and consume any regions we now overlap with. */
234 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
235 if (&rg
->link
== head
)
240 /* We overlap with this area, if it extends further than
241 * us then we must extend ourselves. Account for its
242 * existing reservation. */
247 chg
-= rg
->to
- rg
->from
;
251 spin_unlock(&resv
->lock
);
252 /* We already know we raced and no longer need the new region */
256 spin_unlock(&resv
->lock
);
260 static long region_truncate(struct resv_map
*resv
, long end
)
262 struct list_head
*head
= &resv
->regions
;
263 struct file_region
*rg
, *trg
;
266 spin_lock(&resv
->lock
);
267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg
, head
, link
)
271 if (&rg
->link
== head
)
274 /* If we are in the middle of a region then adjust it. */
275 if (end
> rg
->from
) {
278 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
281 /* Drop any remaining regions. */
282 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
283 if (&rg
->link
== head
)
285 chg
+= rg
->to
- rg
->from
;
291 spin_unlock(&resv
->lock
);
295 static long region_count(struct resv_map
*resv
, long f
, long t
)
297 struct list_head
*head
= &resv
->regions
;
298 struct file_region
*rg
;
301 spin_lock(&resv
->lock
);
302 /* Locate each segment we overlap with, and count that overlap. */
303 list_for_each_entry(rg
, head
, link
) {
312 seg_from
= max(rg
->from
, f
);
313 seg_to
= min(rg
->to
, t
);
315 chg
+= seg_to
- seg_from
;
317 spin_unlock(&resv
->lock
);
323 * Convert the address within this vma to the page offset within
324 * the mapping, in pagecache page units; huge pages here.
326 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
327 struct vm_area_struct
*vma
, unsigned long address
)
329 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
330 (vma
->vm_pgoff
>> huge_page_order(h
));
333 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
334 unsigned long address
)
336 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
340 * Return the size of the pages allocated when backing a VMA. In the majority
341 * cases this will be same size as used by the page table entries.
343 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
345 struct hstate
*hstate
;
347 if (!is_vm_hugetlb_page(vma
))
350 hstate
= hstate_vma(vma
);
352 return 1UL << huge_page_shift(hstate
);
354 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
357 * Return the page size being used by the MMU to back a VMA. In the majority
358 * of cases, the page size used by the kernel matches the MMU size. On
359 * architectures where it differs, an architecture-specific version of this
360 * function is required.
362 #ifndef vma_mmu_pagesize
363 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
365 return vma_kernel_pagesize(vma
);
370 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
371 * bits of the reservation map pointer, which are always clear due to
374 #define HPAGE_RESV_OWNER (1UL << 0)
375 #define HPAGE_RESV_UNMAPPED (1UL << 1)
376 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
379 * These helpers are used to track how many pages are reserved for
380 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
381 * is guaranteed to have their future faults succeed.
383 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
384 * the reserve counters are updated with the hugetlb_lock held. It is safe
385 * to reset the VMA at fork() time as it is not in use yet and there is no
386 * chance of the global counters getting corrupted as a result of the values.
388 * The private mapping reservation is represented in a subtly different
389 * manner to a shared mapping. A shared mapping has a region map associated
390 * with the underlying file, this region map represents the backing file
391 * pages which have ever had a reservation assigned which this persists even
392 * after the page is instantiated. A private mapping has a region map
393 * associated with the original mmap which is attached to all VMAs which
394 * reference it, this region map represents those offsets which have consumed
395 * reservation ie. where pages have been instantiated.
397 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
399 return (unsigned long)vma
->vm_private_data
;
402 static void set_vma_private_data(struct vm_area_struct
*vma
,
405 vma
->vm_private_data
= (void *)value
;
408 struct resv_map
*resv_map_alloc(void)
410 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
414 kref_init(&resv_map
->refs
);
415 spin_lock_init(&resv_map
->lock
);
416 INIT_LIST_HEAD(&resv_map
->regions
);
421 void resv_map_release(struct kref
*ref
)
423 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
425 /* Clear out any active regions before we release the map. */
426 region_truncate(resv_map
, 0);
430 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
432 return inode
->i_mapping
->private_data
;
435 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
437 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
438 if (vma
->vm_flags
& VM_MAYSHARE
) {
439 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
440 struct inode
*inode
= mapping
->host
;
442 return inode_resv_map(inode
);
445 return (struct resv_map
*)(get_vma_private_data(vma
) &
450 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
452 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
453 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
455 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
456 HPAGE_RESV_MASK
) | (unsigned long)map
);
459 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
461 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
462 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
464 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
467 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
469 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
471 return (get_vma_private_data(vma
) & flag
) != 0;
474 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
475 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
477 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
478 if (!(vma
->vm_flags
& VM_MAYSHARE
))
479 vma
->vm_private_data
= (void *)0;
482 /* Returns true if the VMA has associated reserve pages */
483 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
485 if (vma
->vm_flags
& VM_NORESERVE
) {
487 * This address is already reserved by other process(chg == 0),
488 * so, we should decrement reserved count. Without decrementing,
489 * reserve count remains after releasing inode, because this
490 * allocated page will go into page cache and is regarded as
491 * coming from reserved pool in releasing step. Currently, we
492 * don't have any other solution to deal with this situation
493 * properly, so add work-around here.
495 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
501 /* Shared mappings always use reserves */
502 if (vma
->vm_flags
& VM_MAYSHARE
)
506 * Only the process that called mmap() has reserves for
509 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
515 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
517 int nid
= page_to_nid(page
);
518 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
519 h
->free_huge_pages
++;
520 h
->free_huge_pages_node
[nid
]++;
523 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
527 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
528 if (!is_migrate_isolate_page(page
))
531 * if 'non-isolated free hugepage' not found on the list,
532 * the allocation fails.
534 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
536 list_move(&page
->lru
, &h
->hugepage_activelist
);
537 set_page_refcounted(page
);
538 h
->free_huge_pages
--;
539 h
->free_huge_pages_node
[nid
]--;
543 /* Movability of hugepages depends on migration support. */
544 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
546 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
547 return GFP_HIGHUSER_MOVABLE
;
552 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
553 struct vm_area_struct
*vma
,
554 unsigned long address
, int avoid_reserve
,
557 struct page
*page
= NULL
;
558 struct mempolicy
*mpol
;
559 nodemask_t
*nodemask
;
560 struct zonelist
*zonelist
;
563 unsigned int cpuset_mems_cookie
;
566 * A child process with MAP_PRIVATE mappings created by their parent
567 * have no page reserves. This check ensures that reservations are
568 * not "stolen". The child may still get SIGKILLed
570 if (!vma_has_reserves(vma
, chg
) &&
571 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
574 /* If reserves cannot be used, ensure enough pages are in the pool */
575 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
579 cpuset_mems_cookie
= read_mems_allowed_begin();
580 zonelist
= huge_zonelist(vma
, address
,
581 htlb_alloc_mask(h
), &mpol
, &nodemask
);
583 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
584 MAX_NR_ZONES
- 1, nodemask
) {
585 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask(h
))) {
586 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
590 if (!vma_has_reserves(vma
, chg
))
593 SetPagePrivate(page
);
594 h
->resv_huge_pages
--;
601 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
610 * common helper functions for hstate_next_node_to_{alloc|free}.
611 * We may have allocated or freed a huge page based on a different
612 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
613 * be outside of *nodes_allowed. Ensure that we use an allowed
614 * node for alloc or free.
616 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
618 nid
= next_node(nid
, *nodes_allowed
);
619 if (nid
== MAX_NUMNODES
)
620 nid
= first_node(*nodes_allowed
);
621 VM_BUG_ON(nid
>= MAX_NUMNODES
);
626 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
628 if (!node_isset(nid
, *nodes_allowed
))
629 nid
= next_node_allowed(nid
, nodes_allowed
);
634 * returns the previously saved node ["this node"] from which to
635 * allocate a persistent huge page for the pool and advance the
636 * next node from which to allocate, handling wrap at end of node
639 static int hstate_next_node_to_alloc(struct hstate
*h
,
640 nodemask_t
*nodes_allowed
)
644 VM_BUG_ON(!nodes_allowed
);
646 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
647 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
653 * helper for free_pool_huge_page() - return the previously saved
654 * node ["this node"] from which to free a huge page. Advance the
655 * next node id whether or not we find a free huge page to free so
656 * that the next attempt to free addresses the next node.
658 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
662 VM_BUG_ON(!nodes_allowed
);
664 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
665 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
670 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
671 for (nr_nodes = nodes_weight(*mask); \
673 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
676 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
677 for (nr_nodes = nodes_weight(*mask); \
679 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
682 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
683 static void destroy_compound_gigantic_page(struct page
*page
,
687 int nr_pages
= 1 << order
;
688 struct page
*p
= page
+ 1;
690 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
692 set_page_refcounted(p
);
693 p
->first_page
= NULL
;
696 set_compound_order(page
, 0);
697 __ClearPageHead(page
);
700 static void free_gigantic_page(struct page
*page
, unsigned order
)
702 free_contig_range(page_to_pfn(page
), 1 << order
);
705 static int __alloc_gigantic_page(unsigned long start_pfn
,
706 unsigned long nr_pages
)
708 unsigned long end_pfn
= start_pfn
+ nr_pages
;
709 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
);
712 static bool pfn_range_valid_gigantic(unsigned long start_pfn
,
713 unsigned long nr_pages
)
715 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
718 for (i
= start_pfn
; i
< end_pfn
; i
++) {
722 page
= pfn_to_page(i
);
724 if (PageReserved(page
))
727 if (page_count(page
) > 0)
737 static bool zone_spans_last_pfn(const struct zone
*zone
,
738 unsigned long start_pfn
, unsigned long nr_pages
)
740 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
741 return zone_spans_pfn(zone
, last_pfn
);
744 static struct page
*alloc_gigantic_page(int nid
, unsigned order
)
746 unsigned long nr_pages
= 1 << order
;
747 unsigned long ret
, pfn
, flags
;
750 z
= NODE_DATA(nid
)->node_zones
;
751 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
752 spin_lock_irqsave(&z
->lock
, flags
);
754 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
755 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
756 if (pfn_range_valid_gigantic(pfn
, nr_pages
)) {
758 * We release the zone lock here because
759 * alloc_contig_range() will also lock the zone
760 * at some point. If there's an allocation
761 * spinning on this lock, it may win the race
762 * and cause alloc_contig_range() to fail...
764 spin_unlock_irqrestore(&z
->lock
, flags
);
765 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
767 return pfn_to_page(pfn
);
768 spin_lock_irqsave(&z
->lock
, flags
);
773 spin_unlock_irqrestore(&z
->lock
, flags
);
779 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
780 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
);
782 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
786 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
788 prep_compound_gigantic_page(page
, huge_page_order(h
));
789 prep_new_huge_page(h
, page
, nid
);
795 static int alloc_fresh_gigantic_page(struct hstate
*h
,
796 nodemask_t
*nodes_allowed
)
798 struct page
*page
= NULL
;
801 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
802 page
= alloc_fresh_gigantic_page_node(h
, node
);
810 static inline bool gigantic_page_supported(void) { return true; }
812 static inline bool gigantic_page_supported(void) { return false; }
813 static inline void free_gigantic_page(struct page
*page
, unsigned order
) { }
814 static inline void destroy_compound_gigantic_page(struct page
*page
,
815 unsigned long order
) { }
816 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
817 nodemask_t
*nodes_allowed
) { return 0; }
820 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
824 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
828 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
829 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
830 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
831 1 << PG_referenced
| 1 << PG_dirty
|
832 1 << PG_active
| 1 << PG_private
|
835 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
836 set_compound_page_dtor(page
, NULL
);
837 set_page_refcounted(page
);
838 if (hstate_is_gigantic(h
)) {
839 destroy_compound_gigantic_page(page
, huge_page_order(h
));
840 free_gigantic_page(page
, huge_page_order(h
));
842 arch_release_hugepage(page
);
843 __free_pages(page
, huge_page_order(h
));
847 struct hstate
*size_to_hstate(unsigned long size
)
852 if (huge_page_size(h
) == size
)
858 void free_huge_page(struct page
*page
)
861 * Can't pass hstate in here because it is called from the
862 * compound page destructor.
864 struct hstate
*h
= page_hstate(page
);
865 int nid
= page_to_nid(page
);
866 struct hugepage_subpool
*spool
=
867 (struct hugepage_subpool
*)page_private(page
);
868 bool restore_reserve
;
870 set_page_private(page
, 0);
871 page
->mapping
= NULL
;
872 BUG_ON(page_count(page
));
873 BUG_ON(page_mapcount(page
));
874 restore_reserve
= PagePrivate(page
);
875 ClearPagePrivate(page
);
877 spin_lock(&hugetlb_lock
);
878 hugetlb_cgroup_uncharge_page(hstate_index(h
),
879 pages_per_huge_page(h
), page
);
881 h
->resv_huge_pages
++;
883 if (h
->surplus_huge_pages_node
[nid
]) {
884 /* remove the page from active list */
885 list_del(&page
->lru
);
886 update_and_free_page(h
, page
);
887 h
->surplus_huge_pages
--;
888 h
->surplus_huge_pages_node
[nid
]--;
890 arch_clear_hugepage_flags(page
);
891 enqueue_huge_page(h
, page
);
893 spin_unlock(&hugetlb_lock
);
894 hugepage_subpool_put_pages(spool
, 1);
897 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
899 INIT_LIST_HEAD(&page
->lru
);
900 set_compound_page_dtor(page
, free_huge_page
);
901 spin_lock(&hugetlb_lock
);
902 set_hugetlb_cgroup(page
, NULL
);
904 h
->nr_huge_pages_node
[nid
]++;
905 spin_unlock(&hugetlb_lock
);
906 put_page(page
); /* free it into the hugepage allocator */
909 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
912 int nr_pages
= 1 << order
;
913 struct page
*p
= page
+ 1;
915 /* we rely on prep_new_huge_page to set the destructor */
916 set_compound_order(page
, order
);
918 __ClearPageReserved(page
);
919 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
922 * For gigantic hugepages allocated through bootmem at
923 * boot, it's safer to be consistent with the not-gigantic
924 * hugepages and clear the PG_reserved bit from all tail pages
925 * too. Otherwse drivers using get_user_pages() to access tail
926 * pages may get the reference counting wrong if they see
927 * PG_reserved set on a tail page (despite the head page not
928 * having PG_reserved set). Enforcing this consistency between
929 * head and tail pages allows drivers to optimize away a check
930 * on the head page when they need know if put_page() is needed
931 * after get_user_pages().
933 __ClearPageReserved(p
);
934 set_page_count(p
, 0);
935 p
->first_page
= page
;
940 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
941 * transparent huge pages. See the PageTransHuge() documentation for more
944 int PageHuge(struct page
*page
)
946 if (!PageCompound(page
))
949 page
= compound_head(page
);
950 return get_compound_page_dtor(page
) == free_huge_page
;
952 EXPORT_SYMBOL_GPL(PageHuge
);
955 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
956 * normal or transparent huge pages.
958 int PageHeadHuge(struct page
*page_head
)
960 if (!PageHead(page_head
))
963 return get_compound_page_dtor(page_head
) == free_huge_page
;
966 pgoff_t
__basepage_index(struct page
*page
)
968 struct page
*page_head
= compound_head(page
);
969 pgoff_t index
= page_index(page_head
);
970 unsigned long compound_idx
;
972 if (!PageHuge(page_head
))
973 return page_index(page
);
975 if (compound_order(page_head
) >= MAX_ORDER
)
976 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
978 compound_idx
= page
- page_head
;
980 return (index
<< compound_order(page_head
)) + compound_idx
;
983 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
987 page
= alloc_pages_exact_node(nid
,
988 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
989 __GFP_REPEAT
|__GFP_NOWARN
,
992 if (arch_prepare_hugepage(page
)) {
993 __free_pages(page
, huge_page_order(h
));
996 prep_new_huge_page(h
, page
, nid
);
1002 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1008 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1009 page
= alloc_fresh_huge_page_node(h
, node
);
1017 count_vm_event(HTLB_BUDDY_PGALLOC
);
1019 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1025 * Free huge page from pool from next node to free.
1026 * Attempt to keep persistent huge pages more or less
1027 * balanced over allowed nodes.
1028 * Called with hugetlb_lock locked.
1030 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1036 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1038 * If we're returning unused surplus pages, only examine
1039 * nodes with surplus pages.
1041 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1042 !list_empty(&h
->hugepage_freelists
[node
])) {
1044 list_entry(h
->hugepage_freelists
[node
].next
,
1046 list_del(&page
->lru
);
1047 h
->free_huge_pages
--;
1048 h
->free_huge_pages_node
[node
]--;
1050 h
->surplus_huge_pages
--;
1051 h
->surplus_huge_pages_node
[node
]--;
1053 update_and_free_page(h
, page
);
1063 * Dissolve a given free hugepage into free buddy pages. This function does
1064 * nothing for in-use (including surplus) hugepages.
1066 static void dissolve_free_huge_page(struct page
*page
)
1068 spin_lock(&hugetlb_lock
);
1069 if (PageHuge(page
) && !page_count(page
)) {
1070 struct hstate
*h
= page_hstate(page
);
1071 int nid
= page_to_nid(page
);
1072 list_del(&page
->lru
);
1073 h
->free_huge_pages
--;
1074 h
->free_huge_pages_node
[nid
]--;
1075 update_and_free_page(h
, page
);
1077 spin_unlock(&hugetlb_lock
);
1081 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1082 * make specified memory blocks removable from the system.
1083 * Note that start_pfn should aligned with (minimum) hugepage size.
1085 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1087 unsigned int order
= 8 * sizeof(void *);
1091 /* Set scan step to minimum hugepage size */
1093 if (order
> huge_page_order(h
))
1094 order
= huge_page_order(h
);
1095 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
1096 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
1097 dissolve_free_huge_page(pfn_to_page(pfn
));
1100 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
1105 if (hstate_is_gigantic(h
))
1109 * Assume we will successfully allocate the surplus page to
1110 * prevent racing processes from causing the surplus to exceed
1113 * This however introduces a different race, where a process B
1114 * tries to grow the static hugepage pool while alloc_pages() is
1115 * called by process A. B will only examine the per-node
1116 * counters in determining if surplus huge pages can be
1117 * converted to normal huge pages in adjust_pool_surplus(). A
1118 * won't be able to increment the per-node counter, until the
1119 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1120 * no more huge pages can be converted from surplus to normal
1121 * state (and doesn't try to convert again). Thus, we have a
1122 * case where a surplus huge page exists, the pool is grown, and
1123 * the surplus huge page still exists after, even though it
1124 * should just have been converted to a normal huge page. This
1125 * does not leak memory, though, as the hugepage will be freed
1126 * once it is out of use. It also does not allow the counters to
1127 * go out of whack in adjust_pool_surplus() as we don't modify
1128 * the node values until we've gotten the hugepage and only the
1129 * per-node value is checked there.
1131 spin_lock(&hugetlb_lock
);
1132 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1133 spin_unlock(&hugetlb_lock
);
1137 h
->surplus_huge_pages
++;
1139 spin_unlock(&hugetlb_lock
);
1141 if (nid
== NUMA_NO_NODE
)
1142 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
1143 __GFP_REPEAT
|__GFP_NOWARN
,
1144 huge_page_order(h
));
1146 page
= alloc_pages_exact_node(nid
,
1147 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1148 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
1150 if (page
&& arch_prepare_hugepage(page
)) {
1151 __free_pages(page
, huge_page_order(h
));
1155 spin_lock(&hugetlb_lock
);
1157 INIT_LIST_HEAD(&page
->lru
);
1158 r_nid
= page_to_nid(page
);
1159 set_compound_page_dtor(page
, free_huge_page
);
1160 set_hugetlb_cgroup(page
, NULL
);
1162 * We incremented the global counters already
1164 h
->nr_huge_pages_node
[r_nid
]++;
1165 h
->surplus_huge_pages_node
[r_nid
]++;
1166 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1169 h
->surplus_huge_pages
--;
1170 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1172 spin_unlock(&hugetlb_lock
);
1178 * This allocation function is useful in the context where vma is irrelevant.
1179 * E.g. soft-offlining uses this function because it only cares physical
1180 * address of error page.
1182 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1184 struct page
*page
= NULL
;
1186 spin_lock(&hugetlb_lock
);
1187 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1188 page
= dequeue_huge_page_node(h
, nid
);
1189 spin_unlock(&hugetlb_lock
);
1192 page
= alloc_buddy_huge_page(h
, nid
);
1198 * Increase the hugetlb pool such that it can accommodate a reservation
1201 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1203 struct list_head surplus_list
;
1204 struct page
*page
, *tmp
;
1206 int needed
, allocated
;
1207 bool alloc_ok
= true;
1209 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1211 h
->resv_huge_pages
+= delta
;
1216 INIT_LIST_HEAD(&surplus_list
);
1220 spin_unlock(&hugetlb_lock
);
1221 for (i
= 0; i
< needed
; i
++) {
1222 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1227 list_add(&page
->lru
, &surplus_list
);
1232 * After retaking hugetlb_lock, we need to recalculate 'needed'
1233 * because either resv_huge_pages or free_huge_pages may have changed.
1235 spin_lock(&hugetlb_lock
);
1236 needed
= (h
->resv_huge_pages
+ delta
) -
1237 (h
->free_huge_pages
+ allocated
);
1242 * We were not able to allocate enough pages to
1243 * satisfy the entire reservation so we free what
1244 * we've allocated so far.
1249 * The surplus_list now contains _at_least_ the number of extra pages
1250 * needed to accommodate the reservation. Add the appropriate number
1251 * of pages to the hugetlb pool and free the extras back to the buddy
1252 * allocator. Commit the entire reservation here to prevent another
1253 * process from stealing the pages as they are added to the pool but
1254 * before they are reserved.
1256 needed
+= allocated
;
1257 h
->resv_huge_pages
+= delta
;
1260 /* Free the needed pages to the hugetlb pool */
1261 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1265 * This page is now managed by the hugetlb allocator and has
1266 * no users -- drop the buddy allocator's reference.
1268 put_page_testzero(page
);
1269 VM_BUG_ON_PAGE(page_count(page
), page
);
1270 enqueue_huge_page(h
, page
);
1273 spin_unlock(&hugetlb_lock
);
1275 /* Free unnecessary surplus pages to the buddy allocator */
1276 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1278 spin_lock(&hugetlb_lock
);
1284 * When releasing a hugetlb pool reservation, any surplus pages that were
1285 * allocated to satisfy the reservation must be explicitly freed if they were
1287 * Called with hugetlb_lock held.
1289 static void return_unused_surplus_pages(struct hstate
*h
,
1290 unsigned long unused_resv_pages
)
1292 unsigned long nr_pages
;
1294 /* Uncommit the reservation */
1295 h
->resv_huge_pages
-= unused_resv_pages
;
1297 /* Cannot return gigantic pages currently */
1298 if (hstate_is_gigantic(h
))
1301 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1304 * We want to release as many surplus pages as possible, spread
1305 * evenly across all nodes with memory. Iterate across these nodes
1306 * until we can no longer free unreserved surplus pages. This occurs
1307 * when the nodes with surplus pages have no free pages.
1308 * free_pool_huge_page() will balance the the freed pages across the
1309 * on-line nodes with memory and will handle the hstate accounting.
1311 while (nr_pages
--) {
1312 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1314 cond_resched_lock(&hugetlb_lock
);
1319 * Determine if the huge page at addr within the vma has an associated
1320 * reservation. Where it does not we will need to logically increase
1321 * reservation and actually increase subpool usage before an allocation
1322 * can occur. Where any new reservation would be required the
1323 * reservation change is prepared, but not committed. Once the page
1324 * has been allocated from the subpool and instantiated the change should
1325 * be committed via vma_commit_reservation. No action is required on
1328 static long vma_needs_reservation(struct hstate
*h
,
1329 struct vm_area_struct
*vma
, unsigned long addr
)
1331 struct resv_map
*resv
;
1335 resv
= vma_resv_map(vma
);
1339 idx
= vma_hugecache_offset(h
, vma
, addr
);
1340 chg
= region_chg(resv
, idx
, idx
+ 1);
1342 if (vma
->vm_flags
& VM_MAYSHARE
)
1345 return chg
< 0 ? chg
: 0;
1347 static void vma_commit_reservation(struct hstate
*h
,
1348 struct vm_area_struct
*vma
, unsigned long addr
)
1350 struct resv_map
*resv
;
1353 resv
= vma_resv_map(vma
);
1357 idx
= vma_hugecache_offset(h
, vma
, addr
);
1358 region_add(resv
, idx
, idx
+ 1);
1361 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1362 unsigned long addr
, int avoid_reserve
)
1364 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1365 struct hstate
*h
= hstate_vma(vma
);
1369 struct hugetlb_cgroup
*h_cg
;
1371 idx
= hstate_index(h
);
1373 * Processes that did not create the mapping will have no
1374 * reserves and will not have accounted against subpool
1375 * limit. Check that the subpool limit can be made before
1376 * satisfying the allocation MAP_NORESERVE mappings may also
1377 * need pages and subpool limit allocated allocated if no reserve
1380 chg
= vma_needs_reservation(h
, vma
, addr
);
1382 return ERR_PTR(-ENOMEM
);
1383 if (chg
|| avoid_reserve
)
1384 if (hugepage_subpool_get_pages(spool
, 1))
1385 return ERR_PTR(-ENOSPC
);
1387 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1389 goto out_subpool_put
;
1391 spin_lock(&hugetlb_lock
);
1392 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1394 spin_unlock(&hugetlb_lock
);
1395 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1397 goto out_uncharge_cgroup
;
1399 spin_lock(&hugetlb_lock
);
1400 list_move(&page
->lru
, &h
->hugepage_activelist
);
1403 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1404 spin_unlock(&hugetlb_lock
);
1406 set_page_private(page
, (unsigned long)spool
);
1408 vma_commit_reservation(h
, vma
, addr
);
1411 out_uncharge_cgroup
:
1412 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
1414 if (chg
|| avoid_reserve
)
1415 hugepage_subpool_put_pages(spool
, 1);
1416 return ERR_PTR(-ENOSPC
);
1420 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1421 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1422 * where no ERR_VALUE is expected to be returned.
1424 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1425 unsigned long addr
, int avoid_reserve
)
1427 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1433 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1435 struct huge_bootmem_page
*m
;
1438 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1441 addr
= memblock_virt_alloc_try_nid_nopanic(
1442 huge_page_size(h
), huge_page_size(h
),
1443 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1446 * Use the beginning of the huge page to store the
1447 * huge_bootmem_page struct (until gather_bootmem
1448 * puts them into the mem_map).
1457 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1458 /* Put them into a private list first because mem_map is not up yet */
1459 list_add(&m
->list
, &huge_boot_pages
);
1464 static void __init
prep_compound_huge_page(struct page
*page
, int order
)
1466 if (unlikely(order
> (MAX_ORDER
- 1)))
1467 prep_compound_gigantic_page(page
, order
);
1469 prep_compound_page(page
, order
);
1472 /* Put bootmem huge pages into the standard lists after mem_map is up */
1473 static void __init
gather_bootmem_prealloc(void)
1475 struct huge_bootmem_page
*m
;
1477 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1478 struct hstate
*h
= m
->hstate
;
1481 #ifdef CONFIG_HIGHMEM
1482 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1483 memblock_free_late(__pa(m
),
1484 sizeof(struct huge_bootmem_page
));
1486 page
= virt_to_page(m
);
1488 WARN_ON(page_count(page
) != 1);
1489 prep_compound_huge_page(page
, h
->order
);
1490 WARN_ON(PageReserved(page
));
1491 prep_new_huge_page(h
, page
, page_to_nid(page
));
1493 * If we had gigantic hugepages allocated at boot time, we need
1494 * to restore the 'stolen' pages to totalram_pages in order to
1495 * fix confusing memory reports from free(1) and another
1496 * side-effects, like CommitLimit going negative.
1498 if (hstate_is_gigantic(h
))
1499 adjust_managed_page_count(page
, 1 << h
->order
);
1503 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1507 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1508 if (hstate_is_gigantic(h
)) {
1509 if (!alloc_bootmem_huge_page(h
))
1511 } else if (!alloc_fresh_huge_page(h
,
1512 &node_states
[N_MEMORY
]))
1515 h
->max_huge_pages
= i
;
1518 static void __init
hugetlb_init_hstates(void)
1522 for_each_hstate(h
) {
1523 /* oversize hugepages were init'ed in early boot */
1524 if (!hstate_is_gigantic(h
))
1525 hugetlb_hstate_alloc_pages(h
);
1529 static char * __init
memfmt(char *buf
, unsigned long n
)
1531 if (n
>= (1UL << 30))
1532 sprintf(buf
, "%lu GB", n
>> 30);
1533 else if (n
>= (1UL << 20))
1534 sprintf(buf
, "%lu MB", n
>> 20);
1536 sprintf(buf
, "%lu KB", n
>> 10);
1540 static void __init
report_hugepages(void)
1544 for_each_hstate(h
) {
1546 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1547 memfmt(buf
, huge_page_size(h
)),
1548 h
->free_huge_pages
);
1552 #ifdef CONFIG_HIGHMEM
1553 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1554 nodemask_t
*nodes_allowed
)
1558 if (hstate_is_gigantic(h
))
1561 for_each_node_mask(i
, *nodes_allowed
) {
1562 struct page
*page
, *next
;
1563 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1564 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1565 if (count
>= h
->nr_huge_pages
)
1567 if (PageHighMem(page
))
1569 list_del(&page
->lru
);
1570 update_and_free_page(h
, page
);
1571 h
->free_huge_pages
--;
1572 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1577 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1578 nodemask_t
*nodes_allowed
)
1584 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1585 * balanced by operating on them in a round-robin fashion.
1586 * Returns 1 if an adjustment was made.
1588 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1593 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1596 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1597 if (h
->surplus_huge_pages_node
[node
])
1601 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1602 if (h
->surplus_huge_pages_node
[node
] <
1603 h
->nr_huge_pages_node
[node
])
1610 h
->surplus_huge_pages
+= delta
;
1611 h
->surplus_huge_pages_node
[node
] += delta
;
1615 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1616 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1617 nodemask_t
*nodes_allowed
)
1619 unsigned long min_count
, ret
;
1621 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1622 return h
->max_huge_pages
;
1625 * Increase the pool size
1626 * First take pages out of surplus state. Then make up the
1627 * remaining difference by allocating fresh huge pages.
1629 * We might race with alloc_buddy_huge_page() here and be unable
1630 * to convert a surplus huge page to a normal huge page. That is
1631 * not critical, though, it just means the overall size of the
1632 * pool might be one hugepage larger than it needs to be, but
1633 * within all the constraints specified by the sysctls.
1635 spin_lock(&hugetlb_lock
);
1636 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1637 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1641 while (count
> persistent_huge_pages(h
)) {
1643 * If this allocation races such that we no longer need the
1644 * page, free_huge_page will handle it by freeing the page
1645 * and reducing the surplus.
1647 spin_unlock(&hugetlb_lock
);
1648 if (hstate_is_gigantic(h
))
1649 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
1651 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1652 spin_lock(&hugetlb_lock
);
1656 /* Bail for signals. Probably ctrl-c from user */
1657 if (signal_pending(current
))
1662 * Decrease the pool size
1663 * First return free pages to the buddy allocator (being careful
1664 * to keep enough around to satisfy reservations). Then place
1665 * pages into surplus state as needed so the pool will shrink
1666 * to the desired size as pages become free.
1668 * By placing pages into the surplus state independent of the
1669 * overcommit value, we are allowing the surplus pool size to
1670 * exceed overcommit. There are few sane options here. Since
1671 * alloc_buddy_huge_page() is checking the global counter,
1672 * though, we'll note that we're not allowed to exceed surplus
1673 * and won't grow the pool anywhere else. Not until one of the
1674 * sysctls are changed, or the surplus pages go out of use.
1676 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1677 min_count
= max(count
, min_count
);
1678 try_to_free_low(h
, min_count
, nodes_allowed
);
1679 while (min_count
< persistent_huge_pages(h
)) {
1680 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1682 cond_resched_lock(&hugetlb_lock
);
1684 while (count
< persistent_huge_pages(h
)) {
1685 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1689 ret
= persistent_huge_pages(h
);
1690 spin_unlock(&hugetlb_lock
);
1694 #define HSTATE_ATTR_RO(_name) \
1695 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1697 #define HSTATE_ATTR(_name) \
1698 static struct kobj_attribute _name##_attr = \
1699 __ATTR(_name, 0644, _name##_show, _name##_store)
1701 static struct kobject
*hugepages_kobj
;
1702 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1704 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1706 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1710 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1711 if (hstate_kobjs
[i
] == kobj
) {
1713 *nidp
= NUMA_NO_NODE
;
1717 return kobj_to_node_hstate(kobj
, nidp
);
1720 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1721 struct kobj_attribute
*attr
, char *buf
)
1724 unsigned long nr_huge_pages
;
1727 h
= kobj_to_hstate(kobj
, &nid
);
1728 if (nid
== NUMA_NO_NODE
)
1729 nr_huge_pages
= h
->nr_huge_pages
;
1731 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1733 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1736 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
1737 struct hstate
*h
, int nid
,
1738 unsigned long count
, size_t len
)
1741 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1743 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
1748 if (nid
== NUMA_NO_NODE
) {
1750 * global hstate attribute
1752 if (!(obey_mempolicy
&&
1753 init_nodemask_of_mempolicy(nodes_allowed
))) {
1754 NODEMASK_FREE(nodes_allowed
);
1755 nodes_allowed
= &node_states
[N_MEMORY
];
1757 } else if (nodes_allowed
) {
1759 * per node hstate attribute: adjust count to global,
1760 * but restrict alloc/free to the specified node.
1762 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1763 init_nodemask_of_node(nodes_allowed
, nid
);
1765 nodes_allowed
= &node_states
[N_MEMORY
];
1767 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1769 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1770 NODEMASK_FREE(nodes_allowed
);
1774 NODEMASK_FREE(nodes_allowed
);
1778 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1779 struct kobject
*kobj
, const char *buf
,
1783 unsigned long count
;
1787 err
= kstrtoul(buf
, 10, &count
);
1791 h
= kobj_to_hstate(kobj
, &nid
);
1792 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
1795 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1796 struct kobj_attribute
*attr
, char *buf
)
1798 return nr_hugepages_show_common(kobj
, attr
, buf
);
1801 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1802 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1804 return nr_hugepages_store_common(false, kobj
, buf
, len
);
1806 HSTATE_ATTR(nr_hugepages
);
1811 * hstate attribute for optionally mempolicy-based constraint on persistent
1812 * huge page alloc/free.
1814 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1815 struct kobj_attribute
*attr
, char *buf
)
1817 return nr_hugepages_show_common(kobj
, attr
, buf
);
1820 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1821 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1823 return nr_hugepages_store_common(true, kobj
, buf
, len
);
1825 HSTATE_ATTR(nr_hugepages_mempolicy
);
1829 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1830 struct kobj_attribute
*attr
, char *buf
)
1832 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1833 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1836 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1837 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1840 unsigned long input
;
1841 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1843 if (hstate_is_gigantic(h
))
1846 err
= kstrtoul(buf
, 10, &input
);
1850 spin_lock(&hugetlb_lock
);
1851 h
->nr_overcommit_huge_pages
= input
;
1852 spin_unlock(&hugetlb_lock
);
1856 HSTATE_ATTR(nr_overcommit_hugepages
);
1858 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1859 struct kobj_attribute
*attr
, char *buf
)
1862 unsigned long free_huge_pages
;
1865 h
= kobj_to_hstate(kobj
, &nid
);
1866 if (nid
== NUMA_NO_NODE
)
1867 free_huge_pages
= h
->free_huge_pages
;
1869 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1871 return sprintf(buf
, "%lu\n", free_huge_pages
);
1873 HSTATE_ATTR_RO(free_hugepages
);
1875 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1876 struct kobj_attribute
*attr
, char *buf
)
1878 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1879 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1881 HSTATE_ATTR_RO(resv_hugepages
);
1883 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1884 struct kobj_attribute
*attr
, char *buf
)
1887 unsigned long surplus_huge_pages
;
1890 h
= kobj_to_hstate(kobj
, &nid
);
1891 if (nid
== NUMA_NO_NODE
)
1892 surplus_huge_pages
= h
->surplus_huge_pages
;
1894 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1896 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1898 HSTATE_ATTR_RO(surplus_hugepages
);
1900 static struct attribute
*hstate_attrs
[] = {
1901 &nr_hugepages_attr
.attr
,
1902 &nr_overcommit_hugepages_attr
.attr
,
1903 &free_hugepages_attr
.attr
,
1904 &resv_hugepages_attr
.attr
,
1905 &surplus_hugepages_attr
.attr
,
1907 &nr_hugepages_mempolicy_attr
.attr
,
1912 static struct attribute_group hstate_attr_group
= {
1913 .attrs
= hstate_attrs
,
1916 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1917 struct kobject
**hstate_kobjs
,
1918 struct attribute_group
*hstate_attr_group
)
1921 int hi
= hstate_index(h
);
1923 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1924 if (!hstate_kobjs
[hi
])
1927 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1929 kobject_put(hstate_kobjs
[hi
]);
1934 static void __init
hugetlb_sysfs_init(void)
1939 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1940 if (!hugepages_kobj
)
1943 for_each_hstate(h
) {
1944 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1945 hstate_kobjs
, &hstate_attr_group
);
1947 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1954 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1955 * with node devices in node_devices[] using a parallel array. The array
1956 * index of a node device or _hstate == node id.
1957 * This is here to avoid any static dependency of the node device driver, in
1958 * the base kernel, on the hugetlb module.
1960 struct node_hstate
{
1961 struct kobject
*hugepages_kobj
;
1962 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1964 struct node_hstate node_hstates
[MAX_NUMNODES
];
1967 * A subset of global hstate attributes for node devices
1969 static struct attribute
*per_node_hstate_attrs
[] = {
1970 &nr_hugepages_attr
.attr
,
1971 &free_hugepages_attr
.attr
,
1972 &surplus_hugepages_attr
.attr
,
1976 static struct attribute_group per_node_hstate_attr_group
= {
1977 .attrs
= per_node_hstate_attrs
,
1981 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1982 * Returns node id via non-NULL nidp.
1984 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1988 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1989 struct node_hstate
*nhs
= &node_hstates
[nid
];
1991 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1992 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2004 * Unregister hstate attributes from a single node device.
2005 * No-op if no hstate attributes attached.
2007 static void hugetlb_unregister_node(struct node
*node
)
2010 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2012 if (!nhs
->hugepages_kobj
)
2013 return; /* no hstate attributes */
2015 for_each_hstate(h
) {
2016 int idx
= hstate_index(h
);
2017 if (nhs
->hstate_kobjs
[idx
]) {
2018 kobject_put(nhs
->hstate_kobjs
[idx
]);
2019 nhs
->hstate_kobjs
[idx
] = NULL
;
2023 kobject_put(nhs
->hugepages_kobj
);
2024 nhs
->hugepages_kobj
= NULL
;
2028 * hugetlb module exit: unregister hstate attributes from node devices
2031 static void hugetlb_unregister_all_nodes(void)
2036 * disable node device registrations.
2038 register_hugetlbfs_with_node(NULL
, NULL
);
2041 * remove hstate attributes from any nodes that have them.
2043 for (nid
= 0; nid
< nr_node_ids
; nid
++)
2044 hugetlb_unregister_node(node_devices
[nid
]);
2048 * Register hstate attributes for a single node device.
2049 * No-op if attributes already registered.
2051 static void hugetlb_register_node(struct node
*node
)
2054 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2057 if (nhs
->hugepages_kobj
)
2058 return; /* already allocated */
2060 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2062 if (!nhs
->hugepages_kobj
)
2065 for_each_hstate(h
) {
2066 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2068 &per_node_hstate_attr_group
);
2070 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2071 h
->name
, node
->dev
.id
);
2072 hugetlb_unregister_node(node
);
2079 * hugetlb init time: register hstate attributes for all registered node
2080 * devices of nodes that have memory. All on-line nodes should have
2081 * registered their associated device by this time.
2083 static void hugetlb_register_all_nodes(void)
2087 for_each_node_state(nid
, N_MEMORY
) {
2088 struct node
*node
= node_devices
[nid
];
2089 if (node
->dev
.id
== nid
)
2090 hugetlb_register_node(node
);
2094 * Let the node device driver know we're here so it can
2095 * [un]register hstate attributes on node hotplug.
2097 register_hugetlbfs_with_node(hugetlb_register_node
,
2098 hugetlb_unregister_node
);
2100 #else /* !CONFIG_NUMA */
2102 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2110 static void hugetlb_unregister_all_nodes(void) { }
2112 static void hugetlb_register_all_nodes(void) { }
2116 static void __exit
hugetlb_exit(void)
2120 hugetlb_unregister_all_nodes();
2122 for_each_hstate(h
) {
2123 kobject_put(hstate_kobjs
[hstate_index(h
)]);
2126 kobject_put(hugepages_kobj
);
2127 kfree(htlb_fault_mutex_table
);
2129 module_exit(hugetlb_exit
);
2131 static int __init
hugetlb_init(void)
2135 if (!hugepages_supported())
2138 if (!size_to_hstate(default_hstate_size
)) {
2139 default_hstate_size
= HPAGE_SIZE
;
2140 if (!size_to_hstate(default_hstate_size
))
2141 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2143 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2144 if (default_hstate_max_huge_pages
)
2145 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2147 hugetlb_init_hstates();
2148 gather_bootmem_prealloc();
2151 hugetlb_sysfs_init();
2152 hugetlb_register_all_nodes();
2153 hugetlb_cgroup_file_init();
2156 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2158 num_fault_mutexes
= 1;
2160 htlb_fault_mutex_table
=
2161 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2162 BUG_ON(!htlb_fault_mutex_table
);
2164 for (i
= 0; i
< num_fault_mutexes
; i
++)
2165 mutex_init(&htlb_fault_mutex_table
[i
]);
2168 module_init(hugetlb_init
);
2170 /* Should be called on processing a hugepagesz=... option */
2171 void __init
hugetlb_add_hstate(unsigned order
)
2176 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2177 pr_warning("hugepagesz= specified twice, ignoring\n");
2180 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2182 h
= &hstates
[hugetlb_max_hstate
++];
2184 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2185 h
->nr_huge_pages
= 0;
2186 h
->free_huge_pages
= 0;
2187 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2188 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2189 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2190 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2191 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2192 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2193 huge_page_size(h
)/1024);
2198 static int __init
hugetlb_nrpages_setup(char *s
)
2201 static unsigned long *last_mhp
;
2204 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2205 * so this hugepages= parameter goes to the "default hstate".
2207 if (!hugetlb_max_hstate
)
2208 mhp
= &default_hstate_max_huge_pages
;
2210 mhp
= &parsed_hstate
->max_huge_pages
;
2212 if (mhp
== last_mhp
) {
2213 pr_warning("hugepages= specified twice without "
2214 "interleaving hugepagesz=, ignoring\n");
2218 if (sscanf(s
, "%lu", mhp
) <= 0)
2222 * Global state is always initialized later in hugetlb_init.
2223 * But we need to allocate >= MAX_ORDER hstates here early to still
2224 * use the bootmem allocator.
2226 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2227 hugetlb_hstate_alloc_pages(parsed_hstate
);
2233 __setup("hugepages=", hugetlb_nrpages_setup
);
2235 static int __init
hugetlb_default_setup(char *s
)
2237 default_hstate_size
= memparse(s
, &s
);
2240 __setup("default_hugepagesz=", hugetlb_default_setup
);
2242 static unsigned int cpuset_mems_nr(unsigned int *array
)
2245 unsigned int nr
= 0;
2247 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2253 #ifdef CONFIG_SYSCTL
2254 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2255 struct ctl_table
*table
, int write
,
2256 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2258 struct hstate
*h
= &default_hstate
;
2259 unsigned long tmp
= h
->max_huge_pages
;
2262 if (!hugepages_supported())
2266 table
->maxlen
= sizeof(unsigned long);
2267 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2272 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2273 NUMA_NO_NODE
, tmp
, *length
);
2278 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2279 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2282 return hugetlb_sysctl_handler_common(false, table
, write
,
2283 buffer
, length
, ppos
);
2287 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2288 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2290 return hugetlb_sysctl_handler_common(true, table
, write
,
2291 buffer
, length
, ppos
);
2293 #endif /* CONFIG_NUMA */
2295 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2296 void __user
*buffer
,
2297 size_t *length
, loff_t
*ppos
)
2299 struct hstate
*h
= &default_hstate
;
2303 if (!hugepages_supported())
2306 tmp
= h
->nr_overcommit_huge_pages
;
2308 if (write
&& hstate_is_gigantic(h
))
2312 table
->maxlen
= sizeof(unsigned long);
2313 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2318 spin_lock(&hugetlb_lock
);
2319 h
->nr_overcommit_huge_pages
= tmp
;
2320 spin_unlock(&hugetlb_lock
);
2326 #endif /* CONFIG_SYSCTL */
2328 void hugetlb_report_meminfo(struct seq_file
*m
)
2330 struct hstate
*h
= &default_hstate
;
2331 if (!hugepages_supported())
2334 "HugePages_Total: %5lu\n"
2335 "HugePages_Free: %5lu\n"
2336 "HugePages_Rsvd: %5lu\n"
2337 "HugePages_Surp: %5lu\n"
2338 "Hugepagesize: %8lu kB\n",
2342 h
->surplus_huge_pages
,
2343 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2346 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2348 struct hstate
*h
= &default_hstate
;
2349 if (!hugepages_supported())
2352 "Node %d HugePages_Total: %5u\n"
2353 "Node %d HugePages_Free: %5u\n"
2354 "Node %d HugePages_Surp: %5u\n",
2355 nid
, h
->nr_huge_pages_node
[nid
],
2356 nid
, h
->free_huge_pages_node
[nid
],
2357 nid
, h
->surplus_huge_pages_node
[nid
]);
2360 void hugetlb_show_meminfo(void)
2365 if (!hugepages_supported())
2368 for_each_node_state(nid
, N_MEMORY
)
2370 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2372 h
->nr_huge_pages_node
[nid
],
2373 h
->free_huge_pages_node
[nid
],
2374 h
->surplus_huge_pages_node
[nid
],
2375 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2378 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2379 unsigned long hugetlb_total_pages(void)
2382 unsigned long nr_total_pages
= 0;
2385 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2386 return nr_total_pages
;
2389 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2393 spin_lock(&hugetlb_lock
);
2395 * When cpuset is configured, it breaks the strict hugetlb page
2396 * reservation as the accounting is done on a global variable. Such
2397 * reservation is completely rubbish in the presence of cpuset because
2398 * the reservation is not checked against page availability for the
2399 * current cpuset. Application can still potentially OOM'ed by kernel
2400 * with lack of free htlb page in cpuset that the task is in.
2401 * Attempt to enforce strict accounting with cpuset is almost
2402 * impossible (or too ugly) because cpuset is too fluid that
2403 * task or memory node can be dynamically moved between cpusets.
2405 * The change of semantics for shared hugetlb mapping with cpuset is
2406 * undesirable. However, in order to preserve some of the semantics,
2407 * we fall back to check against current free page availability as
2408 * a best attempt and hopefully to minimize the impact of changing
2409 * semantics that cpuset has.
2412 if (gather_surplus_pages(h
, delta
) < 0)
2415 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2416 return_unused_surplus_pages(h
, delta
);
2423 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2426 spin_unlock(&hugetlb_lock
);
2430 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2432 struct resv_map
*resv
= vma_resv_map(vma
);
2435 * This new VMA should share its siblings reservation map if present.
2436 * The VMA will only ever have a valid reservation map pointer where
2437 * it is being copied for another still existing VMA. As that VMA
2438 * has a reference to the reservation map it cannot disappear until
2439 * after this open call completes. It is therefore safe to take a
2440 * new reference here without additional locking.
2442 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2443 kref_get(&resv
->refs
);
2446 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2448 struct hstate
*h
= hstate_vma(vma
);
2449 struct resv_map
*resv
= vma_resv_map(vma
);
2450 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2451 unsigned long reserve
, start
, end
;
2453 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2456 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2457 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2459 reserve
= (end
- start
) - region_count(resv
, start
, end
);
2461 kref_put(&resv
->refs
, resv_map_release
);
2464 hugetlb_acct_memory(h
, -reserve
);
2465 hugepage_subpool_put_pages(spool
, reserve
);
2470 * We cannot handle pagefaults against hugetlb pages at all. They cause
2471 * handle_mm_fault() to try to instantiate regular-sized pages in the
2472 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2475 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2481 const struct vm_operations_struct hugetlb_vm_ops
= {
2482 .fault
= hugetlb_vm_op_fault
,
2483 .open
= hugetlb_vm_op_open
,
2484 .close
= hugetlb_vm_op_close
,
2487 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2493 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2494 vma
->vm_page_prot
)));
2496 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2497 vma
->vm_page_prot
));
2499 entry
= pte_mkyoung(entry
);
2500 entry
= pte_mkhuge(entry
);
2501 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2506 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2507 unsigned long address
, pte_t
*ptep
)
2511 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2512 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2513 update_mmu_cache(vma
, address
, ptep
);
2516 static int is_hugetlb_entry_migration(pte_t pte
)
2520 if (huge_pte_none(pte
) || pte_present(pte
))
2522 swp
= pte_to_swp_entry(pte
);
2523 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2529 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2533 if (huge_pte_none(pte
) || pte_present(pte
))
2535 swp
= pte_to_swp_entry(pte
);
2536 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2542 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2543 struct vm_area_struct
*vma
)
2545 pte_t
*src_pte
, *dst_pte
, entry
;
2546 struct page
*ptepage
;
2549 struct hstate
*h
= hstate_vma(vma
);
2550 unsigned long sz
= huge_page_size(h
);
2551 unsigned long mmun_start
; /* For mmu_notifiers */
2552 unsigned long mmun_end
; /* For mmu_notifiers */
2555 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2557 mmun_start
= vma
->vm_start
;
2558 mmun_end
= vma
->vm_end
;
2560 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
2562 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2563 spinlock_t
*src_ptl
, *dst_ptl
;
2564 src_pte
= huge_pte_offset(src
, addr
);
2567 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2573 /* If the pagetables are shared don't copy or take references */
2574 if (dst_pte
== src_pte
)
2577 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
2578 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
2579 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
2580 entry
= huge_ptep_get(src_pte
);
2581 if (huge_pte_none(entry
)) { /* skip none entry */
2583 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
2584 is_hugetlb_entry_hwpoisoned(entry
))) {
2585 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
2587 if (is_write_migration_entry(swp_entry
) && cow
) {
2589 * COW mappings require pages in both
2590 * parent and child to be set to read.
2592 make_migration_entry_read(&swp_entry
);
2593 entry
= swp_entry_to_pte(swp_entry
);
2594 set_huge_pte_at(src
, addr
, src_pte
, entry
);
2596 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2599 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2600 entry
= huge_ptep_get(src_pte
);
2601 ptepage
= pte_page(entry
);
2603 page_dup_rmap(ptepage
);
2604 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2606 spin_unlock(src_ptl
);
2607 spin_unlock(dst_ptl
);
2611 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
2616 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2617 unsigned long start
, unsigned long end
,
2618 struct page
*ref_page
)
2620 int force_flush
= 0;
2621 struct mm_struct
*mm
= vma
->vm_mm
;
2622 unsigned long address
;
2627 struct hstate
*h
= hstate_vma(vma
);
2628 unsigned long sz
= huge_page_size(h
);
2629 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2630 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2632 WARN_ON(!is_vm_hugetlb_page(vma
));
2633 BUG_ON(start
& ~huge_page_mask(h
));
2634 BUG_ON(end
& ~huge_page_mask(h
));
2636 tlb_start_vma(tlb
, vma
);
2637 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2639 for (address
= start
; address
< end
; address
+= sz
) {
2640 ptep
= huge_pte_offset(mm
, address
);
2644 ptl
= huge_pte_lock(h
, mm
, ptep
);
2645 if (huge_pmd_unshare(mm
, &address
, ptep
))
2648 pte
= huge_ptep_get(ptep
);
2649 if (huge_pte_none(pte
))
2653 * HWPoisoned hugepage is already unmapped and dropped reference
2655 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2656 huge_pte_clear(mm
, address
, ptep
);
2660 page
= pte_page(pte
);
2662 * If a reference page is supplied, it is because a specific
2663 * page is being unmapped, not a range. Ensure the page we
2664 * are about to unmap is the actual page of interest.
2667 if (page
!= ref_page
)
2671 * Mark the VMA as having unmapped its page so that
2672 * future faults in this VMA will fail rather than
2673 * looking like data was lost
2675 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2678 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2679 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2680 if (huge_pte_dirty(pte
))
2681 set_page_dirty(page
);
2683 page_remove_rmap(page
);
2684 force_flush
= !__tlb_remove_page(tlb
, page
);
2689 /* Bail out after unmapping reference page if supplied */
2698 * mmu_gather ran out of room to batch pages, we break out of
2699 * the PTE lock to avoid doing the potential expensive TLB invalidate
2700 * and page-free while holding it.
2705 if (address
< end
&& !ref_page
)
2708 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2709 tlb_end_vma(tlb
, vma
);
2712 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2713 struct vm_area_struct
*vma
, unsigned long start
,
2714 unsigned long end
, struct page
*ref_page
)
2716 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2719 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2720 * test will fail on a vma being torn down, and not grab a page table
2721 * on its way out. We're lucky that the flag has such an appropriate
2722 * name, and can in fact be safely cleared here. We could clear it
2723 * before the __unmap_hugepage_range above, but all that's necessary
2724 * is to clear it before releasing the i_mmap_mutex. This works
2725 * because in the context this is called, the VMA is about to be
2726 * destroyed and the i_mmap_mutex is held.
2728 vma
->vm_flags
&= ~VM_MAYSHARE
;
2731 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2732 unsigned long end
, struct page
*ref_page
)
2734 struct mm_struct
*mm
;
2735 struct mmu_gather tlb
;
2739 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2740 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2741 tlb_finish_mmu(&tlb
, start
, end
);
2745 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2746 * mappping it owns the reserve page for. The intention is to unmap the page
2747 * from other VMAs and let the children be SIGKILLed if they are faulting the
2750 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2751 struct page
*page
, unsigned long address
)
2753 struct hstate
*h
= hstate_vma(vma
);
2754 struct vm_area_struct
*iter_vma
;
2755 struct address_space
*mapping
;
2759 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2760 * from page cache lookup which is in HPAGE_SIZE units.
2762 address
= address
& huge_page_mask(h
);
2763 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2765 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2768 * Take the mapping lock for the duration of the table walk. As
2769 * this mapping should be shared between all the VMAs,
2770 * __unmap_hugepage_range() is called as the lock is already held
2772 mutex_lock(&mapping
->i_mmap_mutex
);
2773 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2774 /* Do not unmap the current VMA */
2775 if (iter_vma
== vma
)
2779 * Unmap the page from other VMAs without their own reserves.
2780 * They get marked to be SIGKILLed if they fault in these
2781 * areas. This is because a future no-page fault on this VMA
2782 * could insert a zeroed page instead of the data existing
2783 * from the time of fork. This would look like data corruption
2785 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2786 unmap_hugepage_range(iter_vma
, address
,
2787 address
+ huge_page_size(h
), page
);
2789 mutex_unlock(&mapping
->i_mmap_mutex
);
2793 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2794 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2795 * cannot race with other handlers or page migration.
2796 * Keep the pte_same checks anyway to make transition from the mutex easier.
2798 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2799 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2800 struct page
*pagecache_page
, spinlock_t
*ptl
)
2802 struct hstate
*h
= hstate_vma(vma
);
2803 struct page
*old_page
, *new_page
;
2804 int ret
= 0, outside_reserve
= 0;
2805 unsigned long mmun_start
; /* For mmu_notifiers */
2806 unsigned long mmun_end
; /* For mmu_notifiers */
2808 old_page
= pte_page(pte
);
2811 /* If no-one else is actually using this page, avoid the copy
2812 * and just make the page writable */
2813 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2814 page_move_anon_rmap(old_page
, vma
, address
);
2815 set_huge_ptep_writable(vma
, address
, ptep
);
2820 * If the process that created a MAP_PRIVATE mapping is about to
2821 * perform a COW due to a shared page count, attempt to satisfy
2822 * the allocation without using the existing reserves. The pagecache
2823 * page is used to determine if the reserve at this address was
2824 * consumed or not. If reserves were used, a partial faulted mapping
2825 * at the time of fork() could consume its reserves on COW instead
2826 * of the full address range.
2828 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2829 old_page
!= pagecache_page
)
2830 outside_reserve
= 1;
2832 page_cache_get(old_page
);
2835 * Drop page table lock as buddy allocator may be called. It will
2836 * be acquired again before returning to the caller, as expected.
2839 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2841 if (IS_ERR(new_page
)) {
2843 * If a process owning a MAP_PRIVATE mapping fails to COW,
2844 * it is due to references held by a child and an insufficient
2845 * huge page pool. To guarantee the original mappers
2846 * reliability, unmap the page from child processes. The child
2847 * may get SIGKILLed if it later faults.
2849 if (outside_reserve
) {
2850 page_cache_release(old_page
);
2851 BUG_ON(huge_pte_none(pte
));
2852 unmap_ref_private(mm
, vma
, old_page
, address
);
2853 BUG_ON(huge_pte_none(pte
));
2855 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2857 pte_same(huge_ptep_get(ptep
), pte
)))
2858 goto retry_avoidcopy
;
2860 * race occurs while re-acquiring page table
2861 * lock, and our job is done.
2866 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
2867 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
2868 goto out_release_old
;
2872 * When the original hugepage is shared one, it does not have
2873 * anon_vma prepared.
2875 if (unlikely(anon_vma_prepare(vma
))) {
2877 goto out_release_all
;
2880 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2881 pages_per_huge_page(h
));
2882 __SetPageUptodate(new_page
);
2884 mmun_start
= address
& huge_page_mask(h
);
2885 mmun_end
= mmun_start
+ huge_page_size(h
);
2886 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2889 * Retake the page table lock to check for racing updates
2890 * before the page tables are altered
2893 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2894 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
2895 ClearPagePrivate(new_page
);
2898 huge_ptep_clear_flush(vma
, address
, ptep
);
2899 set_huge_pte_at(mm
, address
, ptep
,
2900 make_huge_pte(vma
, new_page
, 1));
2901 page_remove_rmap(old_page
);
2902 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2903 /* Make the old page be freed below */
2904 new_page
= old_page
;
2907 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2909 page_cache_release(new_page
);
2911 page_cache_release(old_page
);
2913 spin_lock(ptl
); /* Caller expects lock to be held */
2917 /* Return the pagecache page at a given address within a VMA */
2918 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2919 struct vm_area_struct
*vma
, unsigned long address
)
2921 struct address_space
*mapping
;
2924 mapping
= vma
->vm_file
->f_mapping
;
2925 idx
= vma_hugecache_offset(h
, vma
, address
);
2927 return find_lock_page(mapping
, idx
);
2931 * Return whether there is a pagecache page to back given address within VMA.
2932 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2934 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2935 struct vm_area_struct
*vma
, unsigned long address
)
2937 struct address_space
*mapping
;
2941 mapping
= vma
->vm_file
->f_mapping
;
2942 idx
= vma_hugecache_offset(h
, vma
, address
);
2944 page
= find_get_page(mapping
, idx
);
2947 return page
!= NULL
;
2950 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2951 struct address_space
*mapping
, pgoff_t idx
,
2952 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2954 struct hstate
*h
= hstate_vma(vma
);
2955 int ret
= VM_FAULT_SIGBUS
;
2963 * Currently, we are forced to kill the process in the event the
2964 * original mapper has unmapped pages from the child due to a failed
2965 * COW. Warn that such a situation has occurred as it may not be obvious
2967 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2968 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2974 * Use page lock to guard against racing truncation
2975 * before we get page_table_lock.
2978 page
= find_lock_page(mapping
, idx
);
2980 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2983 page
= alloc_huge_page(vma
, address
, 0);
2985 ret
= PTR_ERR(page
);
2989 ret
= VM_FAULT_SIGBUS
;
2992 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2993 __SetPageUptodate(page
);
2995 if (vma
->vm_flags
& VM_MAYSHARE
) {
2997 struct inode
*inode
= mapping
->host
;
2999 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3006 ClearPagePrivate(page
);
3008 spin_lock(&inode
->i_lock
);
3009 inode
->i_blocks
+= blocks_per_huge_page(h
);
3010 spin_unlock(&inode
->i_lock
);
3013 if (unlikely(anon_vma_prepare(vma
))) {
3015 goto backout_unlocked
;
3021 * If memory error occurs between mmap() and fault, some process
3022 * don't have hwpoisoned swap entry for errored virtual address.
3023 * So we need to block hugepage fault by PG_hwpoison bit check.
3025 if (unlikely(PageHWPoison(page
))) {
3026 ret
= VM_FAULT_HWPOISON
|
3027 VM_FAULT_SET_HINDEX(hstate_index(h
));
3028 goto backout_unlocked
;
3033 * If we are going to COW a private mapping later, we examine the
3034 * pending reservations for this page now. This will ensure that
3035 * any allocations necessary to record that reservation occur outside
3038 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
3039 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3041 goto backout_unlocked
;
3044 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3046 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3051 if (!huge_pte_none(huge_ptep_get(ptep
)))
3055 ClearPagePrivate(page
);
3056 hugepage_add_new_anon_rmap(page
, vma
, address
);
3058 page_dup_rmap(page
);
3059 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3060 && (vma
->vm_flags
& VM_SHARED
)));
3061 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3063 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3064 /* Optimization, do the COW without a second fault */
3065 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
3082 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3083 struct vm_area_struct
*vma
,
3084 struct address_space
*mapping
,
3085 pgoff_t idx
, unsigned long address
)
3087 unsigned long key
[2];
3090 if (vma
->vm_flags
& VM_SHARED
) {
3091 key
[0] = (unsigned long) mapping
;
3094 key
[0] = (unsigned long) mm
;
3095 key
[1] = address
>> huge_page_shift(h
);
3098 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3100 return hash
& (num_fault_mutexes
- 1);
3104 * For uniprocesor systems we always use a single mutex, so just
3105 * return 0 and avoid the hashing overhead.
3107 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3108 struct vm_area_struct
*vma
,
3109 struct address_space
*mapping
,
3110 pgoff_t idx
, unsigned long address
)
3116 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3117 unsigned long address
, unsigned int flags
)
3124 struct page
*page
= NULL
;
3125 struct page
*pagecache_page
= NULL
;
3126 struct hstate
*h
= hstate_vma(vma
);
3127 struct address_space
*mapping
;
3129 address
&= huge_page_mask(h
);
3131 ptep
= huge_pte_offset(mm
, address
);
3133 entry
= huge_ptep_get(ptep
);
3134 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3135 migration_entry_wait_huge(vma
, mm
, ptep
);
3137 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3138 return VM_FAULT_HWPOISON_LARGE
|
3139 VM_FAULT_SET_HINDEX(hstate_index(h
));
3142 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3144 return VM_FAULT_OOM
;
3146 mapping
= vma
->vm_file
->f_mapping
;
3147 idx
= vma_hugecache_offset(h
, vma
, address
);
3150 * Serialize hugepage allocation and instantiation, so that we don't
3151 * get spurious allocation failures if two CPUs race to instantiate
3152 * the same page in the page cache.
3154 hash
= fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3155 mutex_lock(&htlb_fault_mutex_table
[hash
]);
3157 entry
= huge_ptep_get(ptep
);
3158 if (huge_pte_none(entry
)) {
3159 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3166 * If we are going to COW the mapping later, we examine the pending
3167 * reservations for this page now. This will ensure that any
3168 * allocations necessary to record that reservation occur outside the
3169 * spinlock. For private mappings, we also lookup the pagecache
3170 * page now as it is used to determine if a reservation has been
3173 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3174 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3179 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3180 pagecache_page
= hugetlbfs_pagecache_page(h
,
3185 * hugetlb_cow() requires page locks of pte_page(entry) and
3186 * pagecache_page, so here we need take the former one
3187 * when page != pagecache_page or !pagecache_page.
3188 * Note that locking order is always pagecache_page -> page,
3189 * so no worry about deadlock.
3191 page
= pte_page(entry
);
3193 if (page
!= pagecache_page
)
3196 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3198 /* Check for a racing update before calling hugetlb_cow */
3199 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3203 if (flags
& FAULT_FLAG_WRITE
) {
3204 if (!huge_pte_write(entry
)) {
3205 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3206 pagecache_page
, ptl
);
3209 entry
= huge_pte_mkdirty(entry
);
3211 entry
= pte_mkyoung(entry
);
3212 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3213 flags
& FAULT_FLAG_WRITE
))
3214 update_mmu_cache(vma
, address
, ptep
);
3219 if (pagecache_page
) {
3220 unlock_page(pagecache_page
);
3221 put_page(pagecache_page
);
3223 if (page
!= pagecache_page
)
3228 mutex_unlock(&htlb_fault_mutex_table
[hash
]);
3232 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3233 struct page
**pages
, struct vm_area_struct
**vmas
,
3234 unsigned long *position
, unsigned long *nr_pages
,
3235 long i
, unsigned int flags
)
3237 unsigned long pfn_offset
;
3238 unsigned long vaddr
= *position
;
3239 unsigned long remainder
= *nr_pages
;
3240 struct hstate
*h
= hstate_vma(vma
);
3242 while (vaddr
< vma
->vm_end
&& remainder
) {
3244 spinlock_t
*ptl
= NULL
;
3249 * Some archs (sparc64, sh*) have multiple pte_ts to
3250 * each hugepage. We have to make sure we get the
3251 * first, for the page indexing below to work.
3253 * Note that page table lock is not held when pte is null.
3255 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3257 ptl
= huge_pte_lock(h
, mm
, pte
);
3258 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3261 * When coredumping, it suits get_dump_page if we just return
3262 * an error where there's an empty slot with no huge pagecache
3263 * to back it. This way, we avoid allocating a hugepage, and
3264 * the sparse dumpfile avoids allocating disk blocks, but its
3265 * huge holes still show up with zeroes where they need to be.
3267 if (absent
&& (flags
& FOLL_DUMP
) &&
3268 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3276 * We need call hugetlb_fault for both hugepages under migration
3277 * (in which case hugetlb_fault waits for the migration,) and
3278 * hwpoisoned hugepages (in which case we need to prevent the
3279 * caller from accessing to them.) In order to do this, we use
3280 * here is_swap_pte instead of is_hugetlb_entry_migration and
3281 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3282 * both cases, and because we can't follow correct pages
3283 * directly from any kind of swap entries.
3285 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3286 ((flags
& FOLL_WRITE
) &&
3287 !huge_pte_write(huge_ptep_get(pte
)))) {
3292 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3293 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3294 if (!(ret
& VM_FAULT_ERROR
))
3301 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3302 page
= pte_page(huge_ptep_get(pte
));
3305 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3306 get_page_foll(pages
[i
]);
3316 if (vaddr
< vma
->vm_end
&& remainder
&&
3317 pfn_offset
< pages_per_huge_page(h
)) {
3319 * We use pfn_offset to avoid touching the pageframes
3320 * of this compound page.
3326 *nr_pages
= remainder
;
3329 return i
? i
: -EFAULT
;
3332 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3333 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3335 struct mm_struct
*mm
= vma
->vm_mm
;
3336 unsigned long start
= address
;
3339 struct hstate
*h
= hstate_vma(vma
);
3340 unsigned long pages
= 0;
3342 BUG_ON(address
>= end
);
3343 flush_cache_range(vma
, address
, end
);
3345 mmu_notifier_invalidate_range_start(mm
, start
, end
);
3346 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3347 for (; address
< end
; address
+= huge_page_size(h
)) {
3349 ptep
= huge_pte_offset(mm
, address
);
3352 ptl
= huge_pte_lock(h
, mm
, ptep
);
3353 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3358 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3359 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3360 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3361 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3362 set_huge_pte_at(mm
, address
, ptep
, pte
);
3368 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3369 * may have cleared our pud entry and done put_page on the page table:
3370 * once we release i_mmap_mutex, another task can do the final put_page
3371 * and that page table be reused and filled with junk.
3373 flush_tlb_range(vma
, start
, end
);
3374 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3375 mmu_notifier_invalidate_range_end(mm
, start
, end
);
3377 return pages
<< h
->order
;
3380 int hugetlb_reserve_pages(struct inode
*inode
,
3382 struct vm_area_struct
*vma
,
3383 vm_flags_t vm_flags
)
3386 struct hstate
*h
= hstate_inode(inode
);
3387 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3388 struct resv_map
*resv_map
;
3391 * Only apply hugepage reservation if asked. At fault time, an
3392 * attempt will be made for VM_NORESERVE to allocate a page
3393 * without using reserves
3395 if (vm_flags
& VM_NORESERVE
)
3399 * Shared mappings base their reservation on the number of pages that
3400 * are already allocated on behalf of the file. Private mappings need
3401 * to reserve the full area even if read-only as mprotect() may be
3402 * called to make the mapping read-write. Assume !vma is a shm mapping
3404 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
3405 resv_map
= inode_resv_map(inode
);
3407 chg
= region_chg(resv_map
, from
, to
);
3410 resv_map
= resv_map_alloc();
3416 set_vma_resv_map(vma
, resv_map
);
3417 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3425 /* There must be enough pages in the subpool for the mapping */
3426 if (hugepage_subpool_get_pages(spool
, chg
)) {
3432 * Check enough hugepages are available for the reservation.
3433 * Hand the pages back to the subpool if there are not
3435 ret
= hugetlb_acct_memory(h
, chg
);
3437 hugepage_subpool_put_pages(spool
, chg
);
3442 * Account for the reservations made. Shared mappings record regions
3443 * that have reservations as they are shared by multiple VMAs.
3444 * When the last VMA disappears, the region map says how much
3445 * the reservation was and the page cache tells how much of
3446 * the reservation was consumed. Private mappings are per-VMA and
3447 * only the consumed reservations are tracked. When the VMA
3448 * disappears, the original reservation is the VMA size and the
3449 * consumed reservations are stored in the map. Hence, nothing
3450 * else has to be done for private mappings here
3452 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3453 region_add(resv_map
, from
, to
);
3456 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3457 kref_put(&resv_map
->refs
, resv_map_release
);
3461 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3463 struct hstate
*h
= hstate_inode(inode
);
3464 struct resv_map
*resv_map
= inode_resv_map(inode
);
3466 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3469 chg
= region_truncate(resv_map
, offset
);
3470 spin_lock(&inode
->i_lock
);
3471 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3472 spin_unlock(&inode
->i_lock
);
3474 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3475 hugetlb_acct_memory(h
, -(chg
- freed
));
3478 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3479 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3480 struct vm_area_struct
*vma
,
3481 unsigned long addr
, pgoff_t idx
)
3483 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3485 unsigned long sbase
= saddr
& PUD_MASK
;
3486 unsigned long s_end
= sbase
+ PUD_SIZE
;
3488 /* Allow segments to share if only one is marked locked */
3489 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3490 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3493 * match the virtual addresses, permission and the alignment of the
3496 if (pmd_index(addr
) != pmd_index(saddr
) ||
3497 vm_flags
!= svm_flags
||
3498 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3504 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3506 unsigned long base
= addr
& PUD_MASK
;
3507 unsigned long end
= base
+ PUD_SIZE
;
3510 * check on proper vm_flags and page table alignment
3512 if (vma
->vm_flags
& VM_MAYSHARE
&&
3513 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3519 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3520 * and returns the corresponding pte. While this is not necessary for the
3521 * !shared pmd case because we can allocate the pmd later as well, it makes the
3522 * code much cleaner. pmd allocation is essential for the shared case because
3523 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3524 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3525 * bad pmd for sharing.
3527 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3529 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3530 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3531 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3533 struct vm_area_struct
*svma
;
3534 unsigned long saddr
;
3539 if (!vma_shareable(vma
, addr
))
3540 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3542 mutex_lock(&mapping
->i_mmap_mutex
);
3543 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3547 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3549 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3551 get_page(virt_to_page(spte
));
3560 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
3563 pud_populate(mm
, pud
,
3564 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3566 put_page(virt_to_page(spte
));
3569 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3570 mutex_unlock(&mapping
->i_mmap_mutex
);
3575 * unmap huge page backed by shared pte.
3577 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3578 * indicated by page_count > 1, unmap is achieved by clearing pud and
3579 * decrementing the ref count. If count == 1, the pte page is not shared.
3581 * called with page table lock held.
3583 * returns: 1 successfully unmapped a shared pte page
3584 * 0 the underlying pte page is not shared, or it is the last user
3586 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3588 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3589 pud_t
*pud
= pud_offset(pgd
, *addr
);
3591 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3592 if (page_count(virt_to_page(ptep
)) == 1)
3596 put_page(virt_to_page(ptep
));
3597 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3600 #define want_pmd_share() (1)
3601 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3602 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3606 #define want_pmd_share() (0)
3607 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3609 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3610 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3611 unsigned long addr
, unsigned long sz
)
3617 pgd
= pgd_offset(mm
, addr
);
3618 pud
= pud_alloc(mm
, pgd
, addr
);
3620 if (sz
== PUD_SIZE
) {
3623 BUG_ON(sz
!= PMD_SIZE
);
3624 if (want_pmd_share() && pud_none(*pud
))
3625 pte
= huge_pmd_share(mm
, addr
, pud
);
3627 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3630 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3635 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3641 pgd
= pgd_offset(mm
, addr
);
3642 if (pgd_present(*pgd
)) {
3643 pud
= pud_offset(pgd
, addr
);
3644 if (pud_present(*pud
)) {
3646 return (pte_t
*)pud
;
3647 pmd
= pmd_offset(pud
, addr
);
3650 return (pte_t
*) pmd
;
3654 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3655 pmd_t
*pmd
, int write
)
3659 page
= pte_page(*(pte_t
*)pmd
);
3661 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3666 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3667 pud_t
*pud
, int write
)
3671 page
= pte_page(*(pte_t
*)pud
);
3673 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3677 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3679 /* Can be overriden by architectures */
3680 struct page
* __weak
3681 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3682 pud_t
*pud
, int write
)
3688 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3690 #ifdef CONFIG_MEMORY_FAILURE
3692 /* Should be called in hugetlb_lock */
3693 static int is_hugepage_on_freelist(struct page
*hpage
)
3697 struct hstate
*h
= page_hstate(hpage
);
3698 int nid
= page_to_nid(hpage
);
3700 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3707 * This function is called from memory failure code.
3708 * Assume the caller holds page lock of the head page.
3710 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3712 struct hstate
*h
= page_hstate(hpage
);
3713 int nid
= page_to_nid(hpage
);
3716 spin_lock(&hugetlb_lock
);
3717 if (is_hugepage_on_freelist(hpage
)) {
3719 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3720 * but dangling hpage->lru can trigger list-debug warnings
3721 * (this happens when we call unpoison_memory() on it),
3722 * so let it point to itself with list_del_init().
3724 list_del_init(&hpage
->lru
);
3725 set_page_refcounted(hpage
);
3726 h
->free_huge_pages
--;
3727 h
->free_huge_pages_node
[nid
]--;
3730 spin_unlock(&hugetlb_lock
);
3735 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3737 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3738 if (!get_page_unless_zero(page
))
3740 spin_lock(&hugetlb_lock
);
3741 list_move_tail(&page
->lru
, list
);
3742 spin_unlock(&hugetlb_lock
);
3746 void putback_active_hugepage(struct page
*page
)
3748 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3749 spin_lock(&hugetlb_lock
);
3750 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3751 spin_unlock(&hugetlb_lock
);
3755 bool is_hugepage_active(struct page
*page
)
3757 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
3759 * This function can be called for a tail page because the caller,
3760 * scan_movable_pages, scans through a given pfn-range which typically
3761 * covers one memory block. In systems using gigantic hugepage (1GB
3762 * for x86_64,) a hugepage is larger than a memory block, and we don't
3763 * support migrating such large hugepages for now, so return false
3764 * when called for tail pages.
3769 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3770 * so we should return false for them.
3772 if (unlikely(PageHWPoison(page
)))
3774 return page_count(page
) > 0;