2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
38 int hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 __initdata
LIST_HEAD(huge_boot_pages
);
46 /* for command line parsing */
47 static struct hstate
* __initdata parsed_hstate
;
48 static unsigned long __initdata default_hstate_max_huge_pages
;
49 static unsigned long __initdata default_hstate_size
;
52 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
55 DEFINE_SPINLOCK(hugetlb_lock
);
58 * Serializes faults on the same logical page. This is used to
59 * prevent spurious OOMs when the hugepage pool is fully utilized.
61 static int num_fault_mutexes
;
62 static struct mutex
*htlb_fault_mutex_table ____cacheline_aligned_in_smp
;
64 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
66 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
68 spin_unlock(&spool
->lock
);
70 /* If no pages are used, and no other handles to the subpool
71 * remain, free the subpool the subpool remain */
76 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
78 struct hugepage_subpool
*spool
;
80 spool
= 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_VMA(!is_vm_hugetlb_page(vma
), 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_VMA(!is_vm_hugetlb_page(vma
), vma
);
453 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
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_VMA(!is_vm_hugetlb_page(vma
), vma
);
462 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
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_VMA(!is_vm_hugetlb_page(vma
), 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_VMA(!is_vm_hugetlb_page(vma
), 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(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 if (!hugepages_supported())
1094 /* Set scan step to minimum hugepage size */
1096 if (order
> huge_page_order(h
))
1097 order
= huge_page_order(h
);
1098 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
1099 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
1100 dissolve_free_huge_page(pfn_to_page(pfn
));
1103 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
1108 if (hstate_is_gigantic(h
))
1112 * Assume we will successfully allocate the surplus page to
1113 * prevent racing processes from causing the surplus to exceed
1116 * This however introduces a different race, where a process B
1117 * tries to grow the static hugepage pool while alloc_pages() is
1118 * called by process A. B will only examine the per-node
1119 * counters in determining if surplus huge pages can be
1120 * converted to normal huge pages in adjust_pool_surplus(). A
1121 * won't be able to increment the per-node counter, until the
1122 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1123 * no more huge pages can be converted from surplus to normal
1124 * state (and doesn't try to convert again). Thus, we have a
1125 * case where a surplus huge page exists, the pool is grown, and
1126 * the surplus huge page still exists after, even though it
1127 * should just have been converted to a normal huge page. This
1128 * does not leak memory, though, as the hugepage will be freed
1129 * once it is out of use. It also does not allow the counters to
1130 * go out of whack in adjust_pool_surplus() as we don't modify
1131 * the node values until we've gotten the hugepage and only the
1132 * per-node value is checked there.
1134 spin_lock(&hugetlb_lock
);
1135 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1136 spin_unlock(&hugetlb_lock
);
1140 h
->surplus_huge_pages
++;
1142 spin_unlock(&hugetlb_lock
);
1144 if (nid
== NUMA_NO_NODE
)
1145 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
1146 __GFP_REPEAT
|__GFP_NOWARN
,
1147 huge_page_order(h
));
1149 page
= alloc_pages_exact_node(nid
,
1150 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1151 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
1153 if (page
&& arch_prepare_hugepage(page
)) {
1154 __free_pages(page
, huge_page_order(h
));
1158 spin_lock(&hugetlb_lock
);
1160 INIT_LIST_HEAD(&page
->lru
);
1161 r_nid
= page_to_nid(page
);
1162 set_compound_page_dtor(page
, free_huge_page
);
1163 set_hugetlb_cgroup(page
, NULL
);
1165 * We incremented the global counters already
1167 h
->nr_huge_pages_node
[r_nid
]++;
1168 h
->surplus_huge_pages_node
[r_nid
]++;
1169 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1172 h
->surplus_huge_pages
--;
1173 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1175 spin_unlock(&hugetlb_lock
);
1181 * This allocation function is useful in the context where vma is irrelevant.
1182 * E.g. soft-offlining uses this function because it only cares physical
1183 * address of error page.
1185 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1187 struct page
*page
= NULL
;
1189 spin_lock(&hugetlb_lock
);
1190 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1191 page
= dequeue_huge_page_node(h
, nid
);
1192 spin_unlock(&hugetlb_lock
);
1195 page
= alloc_buddy_huge_page(h
, nid
);
1201 * Increase the hugetlb pool such that it can accommodate a reservation
1204 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1206 struct list_head surplus_list
;
1207 struct page
*page
, *tmp
;
1209 int needed
, allocated
;
1210 bool alloc_ok
= true;
1212 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1214 h
->resv_huge_pages
+= delta
;
1219 INIT_LIST_HEAD(&surplus_list
);
1223 spin_unlock(&hugetlb_lock
);
1224 for (i
= 0; i
< needed
; i
++) {
1225 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1230 list_add(&page
->lru
, &surplus_list
);
1235 * After retaking hugetlb_lock, we need to recalculate 'needed'
1236 * because either resv_huge_pages or free_huge_pages may have changed.
1238 spin_lock(&hugetlb_lock
);
1239 needed
= (h
->resv_huge_pages
+ delta
) -
1240 (h
->free_huge_pages
+ allocated
);
1245 * We were not able to allocate enough pages to
1246 * satisfy the entire reservation so we free what
1247 * we've allocated so far.
1252 * The surplus_list now contains _at_least_ the number of extra pages
1253 * needed to accommodate the reservation. Add the appropriate number
1254 * of pages to the hugetlb pool and free the extras back to the buddy
1255 * allocator. Commit the entire reservation here to prevent another
1256 * process from stealing the pages as they are added to the pool but
1257 * before they are reserved.
1259 needed
+= allocated
;
1260 h
->resv_huge_pages
+= delta
;
1263 /* Free the needed pages to the hugetlb pool */
1264 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1268 * This page is now managed by the hugetlb allocator and has
1269 * no users -- drop the buddy allocator's reference.
1271 put_page_testzero(page
);
1272 VM_BUG_ON_PAGE(page_count(page
), page
);
1273 enqueue_huge_page(h
, page
);
1276 spin_unlock(&hugetlb_lock
);
1278 /* Free unnecessary surplus pages to the buddy allocator */
1279 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1281 spin_lock(&hugetlb_lock
);
1287 * When releasing a hugetlb pool reservation, any surplus pages that were
1288 * allocated to satisfy the reservation must be explicitly freed if they were
1290 * Called with hugetlb_lock held.
1292 static void return_unused_surplus_pages(struct hstate
*h
,
1293 unsigned long unused_resv_pages
)
1295 unsigned long nr_pages
;
1297 /* Uncommit the reservation */
1298 h
->resv_huge_pages
-= unused_resv_pages
;
1300 /* Cannot return gigantic pages currently */
1301 if (hstate_is_gigantic(h
))
1304 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1307 * We want to release as many surplus pages as possible, spread
1308 * evenly across all nodes with memory. Iterate across these nodes
1309 * until we can no longer free unreserved surplus pages. This occurs
1310 * when the nodes with surplus pages have no free pages.
1311 * free_pool_huge_page() will balance the the freed pages across the
1312 * on-line nodes with memory and will handle the hstate accounting.
1314 while (nr_pages
--) {
1315 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1317 cond_resched_lock(&hugetlb_lock
);
1322 * Determine if the huge page at addr within the vma has an associated
1323 * reservation. Where it does not we will need to logically increase
1324 * reservation and actually increase subpool usage before an allocation
1325 * can occur. Where any new reservation would be required the
1326 * reservation change is prepared, but not committed. Once the page
1327 * has been allocated from the subpool and instantiated the change should
1328 * be committed via vma_commit_reservation. No action is required on
1331 static long vma_needs_reservation(struct hstate
*h
,
1332 struct vm_area_struct
*vma
, unsigned long addr
)
1334 struct resv_map
*resv
;
1338 resv
= vma_resv_map(vma
);
1342 idx
= vma_hugecache_offset(h
, vma
, addr
);
1343 chg
= region_chg(resv
, idx
, idx
+ 1);
1345 if (vma
->vm_flags
& VM_MAYSHARE
)
1348 return chg
< 0 ? chg
: 0;
1350 static void vma_commit_reservation(struct hstate
*h
,
1351 struct vm_area_struct
*vma
, unsigned long addr
)
1353 struct resv_map
*resv
;
1356 resv
= vma_resv_map(vma
);
1360 idx
= vma_hugecache_offset(h
, vma
, addr
);
1361 region_add(resv
, idx
, idx
+ 1);
1364 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1365 unsigned long addr
, int avoid_reserve
)
1367 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1368 struct hstate
*h
= hstate_vma(vma
);
1372 struct hugetlb_cgroup
*h_cg
;
1374 idx
= hstate_index(h
);
1376 * Processes that did not create the mapping will have no
1377 * reserves and will not have accounted against subpool
1378 * limit. Check that the subpool limit can be made before
1379 * satisfying the allocation MAP_NORESERVE mappings may also
1380 * need pages and subpool limit allocated allocated if no reserve
1383 chg
= vma_needs_reservation(h
, vma
, addr
);
1385 return ERR_PTR(-ENOMEM
);
1386 if (chg
|| avoid_reserve
)
1387 if (hugepage_subpool_get_pages(spool
, 1))
1388 return ERR_PTR(-ENOSPC
);
1390 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1392 goto out_subpool_put
;
1394 spin_lock(&hugetlb_lock
);
1395 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1397 spin_unlock(&hugetlb_lock
);
1398 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1400 goto out_uncharge_cgroup
;
1402 spin_lock(&hugetlb_lock
);
1403 list_move(&page
->lru
, &h
->hugepage_activelist
);
1406 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1407 spin_unlock(&hugetlb_lock
);
1409 set_page_private(page
, (unsigned long)spool
);
1411 vma_commit_reservation(h
, vma
, addr
);
1414 out_uncharge_cgroup
:
1415 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
1417 if (chg
|| avoid_reserve
)
1418 hugepage_subpool_put_pages(spool
, 1);
1419 return ERR_PTR(-ENOSPC
);
1423 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1424 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1425 * where no ERR_VALUE is expected to be returned.
1427 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1428 unsigned long addr
, int avoid_reserve
)
1430 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1436 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1438 struct huge_bootmem_page
*m
;
1441 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1444 addr
= memblock_virt_alloc_try_nid_nopanic(
1445 huge_page_size(h
), huge_page_size(h
),
1446 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1449 * Use the beginning of the huge page to store the
1450 * huge_bootmem_page struct (until gather_bootmem
1451 * puts them into the mem_map).
1460 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
1461 /* Put them into a private list first because mem_map is not up yet */
1462 list_add(&m
->list
, &huge_boot_pages
);
1467 static void __init
prep_compound_huge_page(struct page
*page
, int order
)
1469 if (unlikely(order
> (MAX_ORDER
- 1)))
1470 prep_compound_gigantic_page(page
, order
);
1472 prep_compound_page(page
, order
);
1475 /* Put bootmem huge pages into the standard lists after mem_map is up */
1476 static void __init
gather_bootmem_prealloc(void)
1478 struct huge_bootmem_page
*m
;
1480 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1481 struct hstate
*h
= m
->hstate
;
1484 #ifdef CONFIG_HIGHMEM
1485 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1486 memblock_free_late(__pa(m
),
1487 sizeof(struct huge_bootmem_page
));
1489 page
= virt_to_page(m
);
1491 WARN_ON(page_count(page
) != 1);
1492 prep_compound_huge_page(page
, h
->order
);
1493 WARN_ON(PageReserved(page
));
1494 prep_new_huge_page(h
, page
, page_to_nid(page
));
1496 * If we had gigantic hugepages allocated at boot time, we need
1497 * to restore the 'stolen' pages to totalram_pages in order to
1498 * fix confusing memory reports from free(1) and another
1499 * side-effects, like CommitLimit going negative.
1501 if (hstate_is_gigantic(h
))
1502 adjust_managed_page_count(page
, 1 << h
->order
);
1506 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1510 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1511 if (hstate_is_gigantic(h
)) {
1512 if (!alloc_bootmem_huge_page(h
))
1514 } else if (!alloc_fresh_huge_page(h
,
1515 &node_states
[N_MEMORY
]))
1518 h
->max_huge_pages
= i
;
1521 static void __init
hugetlb_init_hstates(void)
1525 for_each_hstate(h
) {
1526 /* oversize hugepages were init'ed in early boot */
1527 if (!hstate_is_gigantic(h
))
1528 hugetlb_hstate_alloc_pages(h
);
1532 static char * __init
memfmt(char *buf
, unsigned long n
)
1534 if (n
>= (1UL << 30))
1535 sprintf(buf
, "%lu GB", n
>> 30);
1536 else if (n
>= (1UL << 20))
1537 sprintf(buf
, "%lu MB", n
>> 20);
1539 sprintf(buf
, "%lu KB", n
>> 10);
1543 static void __init
report_hugepages(void)
1547 for_each_hstate(h
) {
1549 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1550 memfmt(buf
, huge_page_size(h
)),
1551 h
->free_huge_pages
);
1555 #ifdef CONFIG_HIGHMEM
1556 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1557 nodemask_t
*nodes_allowed
)
1561 if (hstate_is_gigantic(h
))
1564 for_each_node_mask(i
, *nodes_allowed
) {
1565 struct page
*page
, *next
;
1566 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1567 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1568 if (count
>= h
->nr_huge_pages
)
1570 if (PageHighMem(page
))
1572 list_del(&page
->lru
);
1573 update_and_free_page(h
, page
);
1574 h
->free_huge_pages
--;
1575 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1580 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1581 nodemask_t
*nodes_allowed
)
1587 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1588 * balanced by operating on them in a round-robin fashion.
1589 * Returns 1 if an adjustment was made.
1591 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1596 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1599 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1600 if (h
->surplus_huge_pages_node
[node
])
1604 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1605 if (h
->surplus_huge_pages_node
[node
] <
1606 h
->nr_huge_pages_node
[node
])
1613 h
->surplus_huge_pages
+= delta
;
1614 h
->surplus_huge_pages_node
[node
] += delta
;
1618 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1619 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1620 nodemask_t
*nodes_allowed
)
1622 unsigned long min_count
, ret
;
1624 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1625 return h
->max_huge_pages
;
1628 * Increase the pool size
1629 * First take pages out of surplus state. Then make up the
1630 * remaining difference by allocating fresh huge pages.
1632 * We might race with alloc_buddy_huge_page() here and be unable
1633 * to convert a surplus huge page to a normal huge page. That is
1634 * not critical, though, it just means the overall size of the
1635 * pool might be one hugepage larger than it needs to be, but
1636 * within all the constraints specified by the sysctls.
1638 spin_lock(&hugetlb_lock
);
1639 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1640 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1644 while (count
> persistent_huge_pages(h
)) {
1646 * If this allocation races such that we no longer need the
1647 * page, free_huge_page will handle it by freeing the page
1648 * and reducing the surplus.
1650 spin_unlock(&hugetlb_lock
);
1651 if (hstate_is_gigantic(h
))
1652 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
1654 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1655 spin_lock(&hugetlb_lock
);
1659 /* Bail for signals. Probably ctrl-c from user */
1660 if (signal_pending(current
))
1665 * Decrease the pool size
1666 * First return free pages to the buddy allocator (being careful
1667 * to keep enough around to satisfy reservations). Then place
1668 * pages into surplus state as needed so the pool will shrink
1669 * to the desired size as pages become free.
1671 * By placing pages into the surplus state independent of the
1672 * overcommit value, we are allowing the surplus pool size to
1673 * exceed overcommit. There are few sane options here. Since
1674 * alloc_buddy_huge_page() is checking the global counter,
1675 * though, we'll note that we're not allowed to exceed surplus
1676 * and won't grow the pool anywhere else. Not until one of the
1677 * sysctls are changed, or the surplus pages go out of use.
1679 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1680 min_count
= max(count
, min_count
);
1681 try_to_free_low(h
, min_count
, nodes_allowed
);
1682 while (min_count
< persistent_huge_pages(h
)) {
1683 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1685 cond_resched_lock(&hugetlb_lock
);
1687 while (count
< persistent_huge_pages(h
)) {
1688 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1692 ret
= persistent_huge_pages(h
);
1693 spin_unlock(&hugetlb_lock
);
1697 #define HSTATE_ATTR_RO(_name) \
1698 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1700 #define HSTATE_ATTR(_name) \
1701 static struct kobj_attribute _name##_attr = \
1702 __ATTR(_name, 0644, _name##_show, _name##_store)
1704 static struct kobject
*hugepages_kobj
;
1705 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1707 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1709 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1713 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1714 if (hstate_kobjs
[i
] == kobj
) {
1716 *nidp
= NUMA_NO_NODE
;
1720 return kobj_to_node_hstate(kobj
, nidp
);
1723 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1724 struct kobj_attribute
*attr
, char *buf
)
1727 unsigned long nr_huge_pages
;
1730 h
= kobj_to_hstate(kobj
, &nid
);
1731 if (nid
== NUMA_NO_NODE
)
1732 nr_huge_pages
= h
->nr_huge_pages
;
1734 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1736 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1739 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
1740 struct hstate
*h
, int nid
,
1741 unsigned long count
, size_t len
)
1744 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1746 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
1751 if (nid
== NUMA_NO_NODE
) {
1753 * global hstate attribute
1755 if (!(obey_mempolicy
&&
1756 init_nodemask_of_mempolicy(nodes_allowed
))) {
1757 NODEMASK_FREE(nodes_allowed
);
1758 nodes_allowed
= &node_states
[N_MEMORY
];
1760 } else if (nodes_allowed
) {
1762 * per node hstate attribute: adjust count to global,
1763 * but restrict alloc/free to the specified node.
1765 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1766 init_nodemask_of_node(nodes_allowed
, nid
);
1768 nodes_allowed
= &node_states
[N_MEMORY
];
1770 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1772 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1773 NODEMASK_FREE(nodes_allowed
);
1777 NODEMASK_FREE(nodes_allowed
);
1781 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1782 struct kobject
*kobj
, const char *buf
,
1786 unsigned long count
;
1790 err
= kstrtoul(buf
, 10, &count
);
1794 h
= kobj_to_hstate(kobj
, &nid
);
1795 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
1798 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1799 struct kobj_attribute
*attr
, char *buf
)
1801 return nr_hugepages_show_common(kobj
, attr
, buf
);
1804 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1805 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1807 return nr_hugepages_store_common(false, kobj
, buf
, len
);
1809 HSTATE_ATTR(nr_hugepages
);
1814 * hstate attribute for optionally mempolicy-based constraint on persistent
1815 * huge page alloc/free.
1817 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1818 struct kobj_attribute
*attr
, char *buf
)
1820 return nr_hugepages_show_common(kobj
, attr
, buf
);
1823 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1824 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1826 return nr_hugepages_store_common(true, kobj
, buf
, len
);
1828 HSTATE_ATTR(nr_hugepages_mempolicy
);
1832 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1833 struct kobj_attribute
*attr
, char *buf
)
1835 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1836 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1839 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1840 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1843 unsigned long input
;
1844 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1846 if (hstate_is_gigantic(h
))
1849 err
= kstrtoul(buf
, 10, &input
);
1853 spin_lock(&hugetlb_lock
);
1854 h
->nr_overcommit_huge_pages
= input
;
1855 spin_unlock(&hugetlb_lock
);
1859 HSTATE_ATTR(nr_overcommit_hugepages
);
1861 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1862 struct kobj_attribute
*attr
, char *buf
)
1865 unsigned long free_huge_pages
;
1868 h
= kobj_to_hstate(kobj
, &nid
);
1869 if (nid
== NUMA_NO_NODE
)
1870 free_huge_pages
= h
->free_huge_pages
;
1872 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1874 return sprintf(buf
, "%lu\n", free_huge_pages
);
1876 HSTATE_ATTR_RO(free_hugepages
);
1878 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1879 struct kobj_attribute
*attr
, char *buf
)
1881 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1882 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1884 HSTATE_ATTR_RO(resv_hugepages
);
1886 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1887 struct kobj_attribute
*attr
, char *buf
)
1890 unsigned long surplus_huge_pages
;
1893 h
= kobj_to_hstate(kobj
, &nid
);
1894 if (nid
== NUMA_NO_NODE
)
1895 surplus_huge_pages
= h
->surplus_huge_pages
;
1897 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1899 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1901 HSTATE_ATTR_RO(surplus_hugepages
);
1903 static struct attribute
*hstate_attrs
[] = {
1904 &nr_hugepages_attr
.attr
,
1905 &nr_overcommit_hugepages_attr
.attr
,
1906 &free_hugepages_attr
.attr
,
1907 &resv_hugepages_attr
.attr
,
1908 &surplus_hugepages_attr
.attr
,
1910 &nr_hugepages_mempolicy_attr
.attr
,
1915 static struct attribute_group hstate_attr_group
= {
1916 .attrs
= hstate_attrs
,
1919 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1920 struct kobject
**hstate_kobjs
,
1921 struct attribute_group
*hstate_attr_group
)
1924 int hi
= hstate_index(h
);
1926 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1927 if (!hstate_kobjs
[hi
])
1930 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1932 kobject_put(hstate_kobjs
[hi
]);
1937 static void __init
hugetlb_sysfs_init(void)
1942 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1943 if (!hugepages_kobj
)
1946 for_each_hstate(h
) {
1947 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1948 hstate_kobjs
, &hstate_attr_group
);
1950 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1957 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1958 * with node devices in node_devices[] using a parallel array. The array
1959 * index of a node device or _hstate == node id.
1960 * This is here to avoid any static dependency of the node device driver, in
1961 * the base kernel, on the hugetlb module.
1963 struct node_hstate
{
1964 struct kobject
*hugepages_kobj
;
1965 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1967 struct node_hstate node_hstates
[MAX_NUMNODES
];
1970 * A subset of global hstate attributes for node devices
1972 static struct attribute
*per_node_hstate_attrs
[] = {
1973 &nr_hugepages_attr
.attr
,
1974 &free_hugepages_attr
.attr
,
1975 &surplus_hugepages_attr
.attr
,
1979 static struct attribute_group per_node_hstate_attr_group
= {
1980 .attrs
= per_node_hstate_attrs
,
1984 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1985 * Returns node id via non-NULL nidp.
1987 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1991 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1992 struct node_hstate
*nhs
= &node_hstates
[nid
];
1994 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1995 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2007 * Unregister hstate attributes from a single node device.
2008 * No-op if no hstate attributes attached.
2010 static void hugetlb_unregister_node(struct node
*node
)
2013 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2015 if (!nhs
->hugepages_kobj
)
2016 return; /* no hstate attributes */
2018 for_each_hstate(h
) {
2019 int idx
= hstate_index(h
);
2020 if (nhs
->hstate_kobjs
[idx
]) {
2021 kobject_put(nhs
->hstate_kobjs
[idx
]);
2022 nhs
->hstate_kobjs
[idx
] = NULL
;
2026 kobject_put(nhs
->hugepages_kobj
);
2027 nhs
->hugepages_kobj
= NULL
;
2031 * hugetlb module exit: unregister hstate attributes from node devices
2034 static void hugetlb_unregister_all_nodes(void)
2039 * disable node device registrations.
2041 register_hugetlbfs_with_node(NULL
, NULL
);
2044 * remove hstate attributes from any nodes that have them.
2046 for (nid
= 0; nid
< nr_node_ids
; nid
++)
2047 hugetlb_unregister_node(node_devices
[nid
]);
2051 * Register hstate attributes for a single node device.
2052 * No-op if attributes already registered.
2054 static void hugetlb_register_node(struct node
*node
)
2057 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2060 if (nhs
->hugepages_kobj
)
2061 return; /* already allocated */
2063 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2065 if (!nhs
->hugepages_kobj
)
2068 for_each_hstate(h
) {
2069 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2071 &per_node_hstate_attr_group
);
2073 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2074 h
->name
, node
->dev
.id
);
2075 hugetlb_unregister_node(node
);
2082 * hugetlb init time: register hstate attributes for all registered node
2083 * devices of nodes that have memory. All on-line nodes should have
2084 * registered their associated device by this time.
2086 static void __init
hugetlb_register_all_nodes(void)
2090 for_each_node_state(nid
, N_MEMORY
) {
2091 struct node
*node
= node_devices
[nid
];
2092 if (node
->dev
.id
== nid
)
2093 hugetlb_register_node(node
);
2097 * Let the node device driver know we're here so it can
2098 * [un]register hstate attributes on node hotplug.
2100 register_hugetlbfs_with_node(hugetlb_register_node
,
2101 hugetlb_unregister_node
);
2103 #else /* !CONFIG_NUMA */
2105 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2113 static void hugetlb_unregister_all_nodes(void) { }
2115 static void hugetlb_register_all_nodes(void) { }
2119 static void __exit
hugetlb_exit(void)
2123 hugetlb_unregister_all_nodes();
2125 for_each_hstate(h
) {
2126 kobject_put(hstate_kobjs
[hstate_index(h
)]);
2129 kobject_put(hugepages_kobj
);
2130 kfree(htlb_fault_mutex_table
);
2132 module_exit(hugetlb_exit
);
2134 static int __init
hugetlb_init(void)
2138 if (!hugepages_supported())
2141 if (!size_to_hstate(default_hstate_size
)) {
2142 default_hstate_size
= HPAGE_SIZE
;
2143 if (!size_to_hstate(default_hstate_size
))
2144 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2146 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2147 if (default_hstate_max_huge_pages
)
2148 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2150 hugetlb_init_hstates();
2151 gather_bootmem_prealloc();
2154 hugetlb_sysfs_init();
2155 hugetlb_register_all_nodes();
2156 hugetlb_cgroup_file_init();
2159 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2161 num_fault_mutexes
= 1;
2163 htlb_fault_mutex_table
=
2164 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2165 BUG_ON(!htlb_fault_mutex_table
);
2167 for (i
= 0; i
< num_fault_mutexes
; i
++)
2168 mutex_init(&htlb_fault_mutex_table
[i
]);
2171 module_init(hugetlb_init
);
2173 /* Should be called on processing a hugepagesz=... option */
2174 void __init
hugetlb_add_hstate(unsigned order
)
2179 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2180 pr_warning("hugepagesz= specified twice, ignoring\n");
2183 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2185 h
= &hstates
[hugetlb_max_hstate
++];
2187 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2188 h
->nr_huge_pages
= 0;
2189 h
->free_huge_pages
= 0;
2190 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2191 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2192 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2193 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2194 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2195 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2196 huge_page_size(h
)/1024);
2201 static int __init
hugetlb_nrpages_setup(char *s
)
2204 static unsigned long *last_mhp
;
2207 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2208 * so this hugepages= parameter goes to the "default hstate".
2210 if (!hugetlb_max_hstate
)
2211 mhp
= &default_hstate_max_huge_pages
;
2213 mhp
= &parsed_hstate
->max_huge_pages
;
2215 if (mhp
== last_mhp
) {
2216 pr_warning("hugepages= specified twice without "
2217 "interleaving hugepagesz=, ignoring\n");
2221 if (sscanf(s
, "%lu", mhp
) <= 0)
2225 * Global state is always initialized later in hugetlb_init.
2226 * But we need to allocate >= MAX_ORDER hstates here early to still
2227 * use the bootmem allocator.
2229 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2230 hugetlb_hstate_alloc_pages(parsed_hstate
);
2236 __setup("hugepages=", hugetlb_nrpages_setup
);
2238 static int __init
hugetlb_default_setup(char *s
)
2240 default_hstate_size
= memparse(s
, &s
);
2243 __setup("default_hugepagesz=", hugetlb_default_setup
);
2245 static unsigned int cpuset_mems_nr(unsigned int *array
)
2248 unsigned int nr
= 0;
2250 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2256 #ifdef CONFIG_SYSCTL
2257 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2258 struct ctl_table
*table
, int write
,
2259 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2261 struct hstate
*h
= &default_hstate
;
2262 unsigned long tmp
= h
->max_huge_pages
;
2265 if (!hugepages_supported())
2269 table
->maxlen
= sizeof(unsigned long);
2270 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2275 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2276 NUMA_NO_NODE
, tmp
, *length
);
2281 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2282 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2285 return hugetlb_sysctl_handler_common(false, table
, write
,
2286 buffer
, length
, ppos
);
2290 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2291 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2293 return hugetlb_sysctl_handler_common(true, table
, write
,
2294 buffer
, length
, ppos
);
2296 #endif /* CONFIG_NUMA */
2298 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2299 void __user
*buffer
,
2300 size_t *length
, loff_t
*ppos
)
2302 struct hstate
*h
= &default_hstate
;
2306 if (!hugepages_supported())
2309 tmp
= h
->nr_overcommit_huge_pages
;
2311 if (write
&& hstate_is_gigantic(h
))
2315 table
->maxlen
= sizeof(unsigned long);
2316 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2321 spin_lock(&hugetlb_lock
);
2322 h
->nr_overcommit_huge_pages
= tmp
;
2323 spin_unlock(&hugetlb_lock
);
2329 #endif /* CONFIG_SYSCTL */
2331 void hugetlb_report_meminfo(struct seq_file
*m
)
2333 struct hstate
*h
= &default_hstate
;
2334 if (!hugepages_supported())
2337 "HugePages_Total: %5lu\n"
2338 "HugePages_Free: %5lu\n"
2339 "HugePages_Rsvd: %5lu\n"
2340 "HugePages_Surp: %5lu\n"
2341 "Hugepagesize: %8lu kB\n",
2345 h
->surplus_huge_pages
,
2346 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2349 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2351 struct hstate
*h
= &default_hstate
;
2352 if (!hugepages_supported())
2355 "Node %d HugePages_Total: %5u\n"
2356 "Node %d HugePages_Free: %5u\n"
2357 "Node %d HugePages_Surp: %5u\n",
2358 nid
, h
->nr_huge_pages_node
[nid
],
2359 nid
, h
->free_huge_pages_node
[nid
],
2360 nid
, h
->surplus_huge_pages_node
[nid
]);
2363 void hugetlb_show_meminfo(void)
2368 if (!hugepages_supported())
2371 for_each_node_state(nid
, N_MEMORY
)
2373 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2375 h
->nr_huge_pages_node
[nid
],
2376 h
->free_huge_pages_node
[nid
],
2377 h
->surplus_huge_pages_node
[nid
],
2378 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2381 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2382 unsigned long hugetlb_total_pages(void)
2385 unsigned long nr_total_pages
= 0;
2388 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2389 return nr_total_pages
;
2392 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2396 spin_lock(&hugetlb_lock
);
2398 * When cpuset is configured, it breaks the strict hugetlb page
2399 * reservation as the accounting is done on a global variable. Such
2400 * reservation is completely rubbish in the presence of cpuset because
2401 * the reservation is not checked against page availability for the
2402 * current cpuset. Application can still potentially OOM'ed by kernel
2403 * with lack of free htlb page in cpuset that the task is in.
2404 * Attempt to enforce strict accounting with cpuset is almost
2405 * impossible (or too ugly) because cpuset is too fluid that
2406 * task or memory node can be dynamically moved between cpusets.
2408 * The change of semantics for shared hugetlb mapping with cpuset is
2409 * undesirable. However, in order to preserve some of the semantics,
2410 * we fall back to check against current free page availability as
2411 * a best attempt and hopefully to minimize the impact of changing
2412 * semantics that cpuset has.
2415 if (gather_surplus_pages(h
, delta
) < 0)
2418 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2419 return_unused_surplus_pages(h
, delta
);
2426 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2429 spin_unlock(&hugetlb_lock
);
2433 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2435 struct resv_map
*resv
= vma_resv_map(vma
);
2438 * This new VMA should share its siblings reservation map if present.
2439 * The VMA will only ever have a valid reservation map pointer where
2440 * it is being copied for another still existing VMA. As that VMA
2441 * has a reference to the reservation map it cannot disappear until
2442 * after this open call completes. It is therefore safe to take a
2443 * new reference here without additional locking.
2445 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2446 kref_get(&resv
->refs
);
2449 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2451 struct hstate
*h
= hstate_vma(vma
);
2452 struct resv_map
*resv
= vma_resv_map(vma
);
2453 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2454 unsigned long reserve
, start
, end
;
2456 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2459 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2460 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2462 reserve
= (end
- start
) - region_count(resv
, start
, end
);
2464 kref_put(&resv
->refs
, resv_map_release
);
2467 hugetlb_acct_memory(h
, -reserve
);
2468 hugepage_subpool_put_pages(spool
, reserve
);
2473 * We cannot handle pagefaults against hugetlb pages at all. They cause
2474 * handle_mm_fault() to try to instantiate regular-sized pages in the
2475 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2478 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2484 const struct vm_operations_struct hugetlb_vm_ops
= {
2485 .fault
= hugetlb_vm_op_fault
,
2486 .open
= hugetlb_vm_op_open
,
2487 .close
= hugetlb_vm_op_close
,
2490 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2496 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2497 vma
->vm_page_prot
)));
2499 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2500 vma
->vm_page_prot
));
2502 entry
= pte_mkyoung(entry
);
2503 entry
= pte_mkhuge(entry
);
2504 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2509 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2510 unsigned long address
, pte_t
*ptep
)
2514 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2515 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2516 update_mmu_cache(vma
, address
, ptep
);
2519 static int is_hugetlb_entry_migration(pte_t pte
)
2523 if (huge_pte_none(pte
) || pte_present(pte
))
2525 swp
= pte_to_swp_entry(pte
);
2526 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2532 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2536 if (huge_pte_none(pte
) || pte_present(pte
))
2538 swp
= pte_to_swp_entry(pte
);
2539 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2545 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2546 struct vm_area_struct
*vma
)
2548 pte_t
*src_pte
, *dst_pte
, entry
;
2549 struct page
*ptepage
;
2552 struct hstate
*h
= hstate_vma(vma
);
2553 unsigned long sz
= huge_page_size(h
);
2554 unsigned long mmun_start
; /* For mmu_notifiers */
2555 unsigned long mmun_end
; /* For mmu_notifiers */
2558 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2560 mmun_start
= vma
->vm_start
;
2561 mmun_end
= vma
->vm_end
;
2563 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
2565 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2566 spinlock_t
*src_ptl
, *dst_ptl
;
2567 src_pte
= huge_pte_offset(src
, addr
);
2570 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2576 /* If the pagetables are shared don't copy or take references */
2577 if (dst_pte
== src_pte
)
2580 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
2581 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
2582 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
2583 entry
= huge_ptep_get(src_pte
);
2584 if (huge_pte_none(entry
)) { /* skip none entry */
2586 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
2587 is_hugetlb_entry_hwpoisoned(entry
))) {
2588 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
2590 if (is_write_migration_entry(swp_entry
) && cow
) {
2592 * COW mappings require pages in both
2593 * parent and child to be set to read.
2595 make_migration_entry_read(&swp_entry
);
2596 entry
= swp_entry_to_pte(swp_entry
);
2597 set_huge_pte_at(src
, addr
, src_pte
, entry
);
2599 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2602 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2603 mmu_notifier_invalidate_range(src
, mmun_start
,
2606 entry
= huge_ptep_get(src_pte
);
2607 ptepage
= pte_page(entry
);
2609 page_dup_rmap(ptepage
);
2610 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2612 spin_unlock(src_ptl
);
2613 spin_unlock(dst_ptl
);
2617 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
2622 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2623 unsigned long start
, unsigned long end
,
2624 struct page
*ref_page
)
2626 int force_flush
= 0;
2627 struct mm_struct
*mm
= vma
->vm_mm
;
2628 unsigned long address
;
2633 struct hstate
*h
= hstate_vma(vma
);
2634 unsigned long sz
= huge_page_size(h
);
2635 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2636 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2638 WARN_ON(!is_vm_hugetlb_page(vma
));
2639 BUG_ON(start
& ~huge_page_mask(h
));
2640 BUG_ON(end
& ~huge_page_mask(h
));
2642 tlb_start_vma(tlb
, vma
);
2643 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2646 for (; address
< end
; address
+= sz
) {
2647 ptep
= huge_pte_offset(mm
, address
);
2651 ptl
= huge_pte_lock(h
, mm
, ptep
);
2652 if (huge_pmd_unshare(mm
, &address
, ptep
))
2655 pte
= huge_ptep_get(ptep
);
2656 if (huge_pte_none(pte
))
2660 * Migrating hugepage or HWPoisoned hugepage is already
2661 * unmapped and its refcount is dropped, so just clear pte here.
2663 if (unlikely(!pte_present(pte
))) {
2664 huge_pte_clear(mm
, address
, ptep
);
2668 page
= pte_page(pte
);
2670 * If a reference page is supplied, it is because a specific
2671 * page is being unmapped, not a range. Ensure the page we
2672 * are about to unmap is the actual page of interest.
2675 if (page
!= ref_page
)
2679 * Mark the VMA as having unmapped its page so that
2680 * future faults in this VMA will fail rather than
2681 * looking like data was lost
2683 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2686 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2687 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2688 if (huge_pte_dirty(pte
))
2689 set_page_dirty(page
);
2691 page_remove_rmap(page
);
2692 force_flush
= !__tlb_remove_page(tlb
, page
);
2698 /* Bail out after unmapping reference page if supplied */
2707 * mmu_gather ran out of room to batch pages, we break out of
2708 * the PTE lock to avoid doing the potential expensive TLB invalidate
2709 * and page-free while holding it.
2714 if (address
< end
&& !ref_page
)
2717 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2718 tlb_end_vma(tlb
, vma
);
2721 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2722 struct vm_area_struct
*vma
, unsigned long start
,
2723 unsigned long end
, struct page
*ref_page
)
2725 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2728 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2729 * test will fail on a vma being torn down, and not grab a page table
2730 * on its way out. We're lucky that the flag has such an appropriate
2731 * name, and can in fact be safely cleared here. We could clear it
2732 * before the __unmap_hugepage_range above, but all that's necessary
2733 * is to clear it before releasing the i_mmap_rwsem. This works
2734 * because in the context this is called, the VMA is about to be
2735 * destroyed and the i_mmap_rwsem is held.
2737 vma
->vm_flags
&= ~VM_MAYSHARE
;
2740 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2741 unsigned long end
, struct page
*ref_page
)
2743 struct mm_struct
*mm
;
2744 struct mmu_gather tlb
;
2748 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2749 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2750 tlb_finish_mmu(&tlb
, start
, end
);
2754 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2755 * mappping it owns the reserve page for. The intention is to unmap the page
2756 * from other VMAs and let the children be SIGKILLed if they are faulting the
2759 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2760 struct page
*page
, unsigned long address
)
2762 struct hstate
*h
= hstate_vma(vma
);
2763 struct vm_area_struct
*iter_vma
;
2764 struct address_space
*mapping
;
2768 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2769 * from page cache lookup which is in HPAGE_SIZE units.
2771 address
= address
& huge_page_mask(h
);
2772 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2774 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2777 * Take the mapping lock for the duration of the table walk. As
2778 * this mapping should be shared between all the VMAs,
2779 * __unmap_hugepage_range() is called as the lock is already held
2781 i_mmap_lock_write(mapping
);
2782 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2783 /* Do not unmap the current VMA */
2784 if (iter_vma
== vma
)
2788 * Unmap the page from other VMAs without their own reserves.
2789 * They get marked to be SIGKILLed if they fault in these
2790 * areas. This is because a future no-page fault on this VMA
2791 * could insert a zeroed page instead of the data existing
2792 * from the time of fork. This would look like data corruption
2794 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2795 unmap_hugepage_range(iter_vma
, address
,
2796 address
+ huge_page_size(h
), page
);
2798 i_mmap_unlock_write(mapping
);
2802 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2803 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2804 * cannot race with other handlers or page migration.
2805 * Keep the pte_same checks anyway to make transition from the mutex easier.
2807 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2808 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2809 struct page
*pagecache_page
, spinlock_t
*ptl
)
2811 struct hstate
*h
= hstate_vma(vma
);
2812 struct page
*old_page
, *new_page
;
2813 int ret
= 0, outside_reserve
= 0;
2814 unsigned long mmun_start
; /* For mmu_notifiers */
2815 unsigned long mmun_end
; /* For mmu_notifiers */
2817 old_page
= pte_page(pte
);
2820 /* If no-one else is actually using this page, avoid the copy
2821 * and just make the page writable */
2822 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2823 page_move_anon_rmap(old_page
, vma
, address
);
2824 set_huge_ptep_writable(vma
, address
, ptep
);
2829 * If the process that created a MAP_PRIVATE mapping is about to
2830 * perform a COW due to a shared page count, attempt to satisfy
2831 * the allocation without using the existing reserves. The pagecache
2832 * page is used to determine if the reserve at this address was
2833 * consumed or not. If reserves were used, a partial faulted mapping
2834 * at the time of fork() could consume its reserves on COW instead
2835 * of the full address range.
2837 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2838 old_page
!= pagecache_page
)
2839 outside_reserve
= 1;
2841 page_cache_get(old_page
);
2844 * Drop page table lock as buddy allocator may be called. It will
2845 * be acquired again before returning to the caller, as expected.
2848 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2850 if (IS_ERR(new_page
)) {
2852 * If a process owning a MAP_PRIVATE mapping fails to COW,
2853 * it is due to references held by a child and an insufficient
2854 * huge page pool. To guarantee the original mappers
2855 * reliability, unmap the page from child processes. The child
2856 * may get SIGKILLed if it later faults.
2858 if (outside_reserve
) {
2859 page_cache_release(old_page
);
2860 BUG_ON(huge_pte_none(pte
));
2861 unmap_ref_private(mm
, vma
, old_page
, address
);
2862 BUG_ON(huge_pte_none(pte
));
2864 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2866 pte_same(huge_ptep_get(ptep
), pte
)))
2867 goto retry_avoidcopy
;
2869 * race occurs while re-acquiring page table
2870 * lock, and our job is done.
2875 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
2876 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
2877 goto out_release_old
;
2881 * When the original hugepage is shared one, it does not have
2882 * anon_vma prepared.
2884 if (unlikely(anon_vma_prepare(vma
))) {
2886 goto out_release_all
;
2889 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2890 pages_per_huge_page(h
));
2891 __SetPageUptodate(new_page
);
2893 mmun_start
= address
& huge_page_mask(h
);
2894 mmun_end
= mmun_start
+ huge_page_size(h
);
2895 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2898 * Retake the page table lock to check for racing updates
2899 * before the page tables are altered
2902 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2903 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
2904 ClearPagePrivate(new_page
);
2907 huge_ptep_clear_flush(vma
, address
, ptep
);
2908 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
2909 set_huge_pte_at(mm
, address
, ptep
,
2910 make_huge_pte(vma
, new_page
, 1));
2911 page_remove_rmap(old_page
);
2912 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2913 /* Make the old page be freed below */
2914 new_page
= old_page
;
2917 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2919 page_cache_release(new_page
);
2921 page_cache_release(old_page
);
2923 spin_lock(ptl
); /* Caller expects lock to be held */
2927 /* Return the pagecache page at a given address within a VMA */
2928 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2929 struct vm_area_struct
*vma
, unsigned long address
)
2931 struct address_space
*mapping
;
2934 mapping
= vma
->vm_file
->f_mapping
;
2935 idx
= vma_hugecache_offset(h
, vma
, address
);
2937 return find_lock_page(mapping
, idx
);
2941 * Return whether there is a pagecache page to back given address within VMA.
2942 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2944 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2945 struct vm_area_struct
*vma
, unsigned long address
)
2947 struct address_space
*mapping
;
2951 mapping
= vma
->vm_file
->f_mapping
;
2952 idx
= vma_hugecache_offset(h
, vma
, address
);
2954 page
= find_get_page(mapping
, idx
);
2957 return page
!= NULL
;
2960 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2961 struct address_space
*mapping
, pgoff_t idx
,
2962 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2964 struct hstate
*h
= hstate_vma(vma
);
2965 int ret
= VM_FAULT_SIGBUS
;
2973 * Currently, we are forced to kill the process in the event the
2974 * original mapper has unmapped pages from the child due to a failed
2975 * COW. Warn that such a situation has occurred as it may not be obvious
2977 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2978 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2984 * Use page lock to guard against racing truncation
2985 * before we get page_table_lock.
2988 page
= find_lock_page(mapping
, idx
);
2990 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2993 page
= alloc_huge_page(vma
, address
, 0);
2995 ret
= PTR_ERR(page
);
2999 ret
= VM_FAULT_SIGBUS
;
3002 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3003 __SetPageUptodate(page
);
3005 if (vma
->vm_flags
& VM_MAYSHARE
) {
3007 struct inode
*inode
= mapping
->host
;
3009 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3016 ClearPagePrivate(page
);
3018 spin_lock(&inode
->i_lock
);
3019 inode
->i_blocks
+= blocks_per_huge_page(h
);
3020 spin_unlock(&inode
->i_lock
);
3023 if (unlikely(anon_vma_prepare(vma
))) {
3025 goto backout_unlocked
;
3031 * If memory error occurs between mmap() and fault, some process
3032 * don't have hwpoisoned swap entry for errored virtual address.
3033 * So we need to block hugepage fault by PG_hwpoison bit check.
3035 if (unlikely(PageHWPoison(page
))) {
3036 ret
= VM_FAULT_HWPOISON
|
3037 VM_FAULT_SET_HINDEX(hstate_index(h
));
3038 goto backout_unlocked
;
3043 * If we are going to COW a private mapping later, we examine the
3044 * pending reservations for this page now. This will ensure that
3045 * any allocations necessary to record that reservation occur outside
3048 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
3049 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3051 goto backout_unlocked
;
3054 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3056 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3061 if (!huge_pte_none(huge_ptep_get(ptep
)))
3065 ClearPagePrivate(page
);
3066 hugepage_add_new_anon_rmap(page
, vma
, address
);
3068 page_dup_rmap(page
);
3069 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3070 && (vma
->vm_flags
& VM_SHARED
)));
3071 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3073 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3074 /* Optimization, do the COW without a second fault */
3075 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
3092 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3093 struct vm_area_struct
*vma
,
3094 struct address_space
*mapping
,
3095 pgoff_t idx
, unsigned long address
)
3097 unsigned long key
[2];
3100 if (vma
->vm_flags
& VM_SHARED
) {
3101 key
[0] = (unsigned long) mapping
;
3104 key
[0] = (unsigned long) mm
;
3105 key
[1] = address
>> huge_page_shift(h
);
3108 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3110 return hash
& (num_fault_mutexes
- 1);
3114 * For uniprocesor systems we always use a single mutex, so just
3115 * return 0 and avoid the hashing overhead.
3117 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3118 struct vm_area_struct
*vma
,
3119 struct address_space
*mapping
,
3120 pgoff_t idx
, unsigned long address
)
3126 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3127 unsigned long address
, unsigned int flags
)
3134 struct page
*page
= NULL
;
3135 struct page
*pagecache_page
= NULL
;
3136 struct hstate
*h
= hstate_vma(vma
);
3137 struct address_space
*mapping
;
3138 int need_wait_lock
= 0;
3140 address
&= huge_page_mask(h
);
3142 ptep
= huge_pte_offset(mm
, address
);
3144 entry
= huge_ptep_get(ptep
);
3145 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3146 migration_entry_wait_huge(vma
, mm
, ptep
);
3148 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3149 return VM_FAULT_HWPOISON_LARGE
|
3150 VM_FAULT_SET_HINDEX(hstate_index(h
));
3153 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3155 return VM_FAULT_OOM
;
3157 mapping
= vma
->vm_file
->f_mapping
;
3158 idx
= vma_hugecache_offset(h
, vma
, address
);
3161 * Serialize hugepage allocation and instantiation, so that we don't
3162 * get spurious allocation failures if two CPUs race to instantiate
3163 * the same page in the page cache.
3165 hash
= fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3166 mutex_lock(&htlb_fault_mutex_table
[hash
]);
3168 entry
= huge_ptep_get(ptep
);
3169 if (huge_pte_none(entry
)) {
3170 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3177 * entry could be a migration/hwpoison entry at this point, so this
3178 * check prevents the kernel from going below assuming that we have
3179 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3180 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3183 if (!pte_present(entry
))
3187 * If we are going to COW the mapping later, we examine the pending
3188 * reservations for this page now. This will ensure that any
3189 * allocations necessary to record that reservation occur outside the
3190 * spinlock. For private mappings, we also lookup the pagecache
3191 * page now as it is used to determine if a reservation has been
3194 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3195 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3200 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3201 pagecache_page
= hugetlbfs_pagecache_page(h
,
3205 ptl
= huge_pte_lock(h
, mm
, ptep
);
3207 /* Check for a racing update before calling hugetlb_cow */
3208 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3212 * hugetlb_cow() requires page locks of pte_page(entry) and
3213 * pagecache_page, so here we need take the former one
3214 * when page != pagecache_page or !pagecache_page.
3216 page
= pte_page(entry
);
3217 if (page
!= pagecache_page
)
3218 if (!trylock_page(page
)) {
3225 if (flags
& FAULT_FLAG_WRITE
) {
3226 if (!huge_pte_write(entry
)) {
3227 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3228 pagecache_page
, ptl
);
3231 entry
= huge_pte_mkdirty(entry
);
3233 entry
= pte_mkyoung(entry
);
3234 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3235 flags
& FAULT_FLAG_WRITE
))
3236 update_mmu_cache(vma
, address
, ptep
);
3238 if (page
!= pagecache_page
)
3244 if (pagecache_page
) {
3245 unlock_page(pagecache_page
);
3246 put_page(pagecache_page
);
3249 mutex_unlock(&htlb_fault_mutex_table
[hash
]);
3251 * Generally it's safe to hold refcount during waiting page lock. But
3252 * here we just wait to defer the next page fault to avoid busy loop and
3253 * the page is not used after unlocked before returning from the current
3254 * page fault. So we are safe from accessing freed page, even if we wait
3255 * here without taking refcount.
3258 wait_on_page_locked(page
);
3262 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3263 struct page
**pages
, struct vm_area_struct
**vmas
,
3264 unsigned long *position
, unsigned long *nr_pages
,
3265 long i
, unsigned int flags
)
3267 unsigned long pfn_offset
;
3268 unsigned long vaddr
= *position
;
3269 unsigned long remainder
= *nr_pages
;
3270 struct hstate
*h
= hstate_vma(vma
);
3272 while (vaddr
< vma
->vm_end
&& remainder
) {
3274 spinlock_t
*ptl
= NULL
;
3279 * Some archs (sparc64, sh*) have multiple pte_ts to
3280 * each hugepage. We have to make sure we get the
3281 * first, for the page indexing below to work.
3283 * Note that page table lock is not held when pte is null.
3285 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3287 ptl
= huge_pte_lock(h
, mm
, pte
);
3288 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3291 * When coredumping, it suits get_dump_page if we just return
3292 * an error where there's an empty slot with no huge pagecache
3293 * to back it. This way, we avoid allocating a hugepage, and
3294 * the sparse dumpfile avoids allocating disk blocks, but its
3295 * huge holes still show up with zeroes where they need to be.
3297 if (absent
&& (flags
& FOLL_DUMP
) &&
3298 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3306 * We need call hugetlb_fault for both hugepages under migration
3307 * (in which case hugetlb_fault waits for the migration,) and
3308 * hwpoisoned hugepages (in which case we need to prevent the
3309 * caller from accessing to them.) In order to do this, we use
3310 * here is_swap_pte instead of is_hugetlb_entry_migration and
3311 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3312 * both cases, and because we can't follow correct pages
3313 * directly from any kind of swap entries.
3315 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3316 ((flags
& FOLL_WRITE
) &&
3317 !huge_pte_write(huge_ptep_get(pte
)))) {
3322 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3323 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3324 if (!(ret
& VM_FAULT_ERROR
))
3331 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3332 page
= pte_page(huge_ptep_get(pte
));
3335 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3336 get_page_foll(pages
[i
]);
3346 if (vaddr
< vma
->vm_end
&& remainder
&&
3347 pfn_offset
< pages_per_huge_page(h
)) {
3349 * We use pfn_offset to avoid touching the pageframes
3350 * of this compound page.
3356 *nr_pages
= remainder
;
3359 return i
? i
: -EFAULT
;
3362 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3363 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3365 struct mm_struct
*mm
= vma
->vm_mm
;
3366 unsigned long start
= address
;
3369 struct hstate
*h
= hstate_vma(vma
);
3370 unsigned long pages
= 0;
3372 BUG_ON(address
>= end
);
3373 flush_cache_range(vma
, address
, end
);
3375 mmu_notifier_invalidate_range_start(mm
, start
, end
);
3376 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
3377 for (; address
< end
; address
+= huge_page_size(h
)) {
3379 ptep
= huge_pte_offset(mm
, address
);
3382 ptl
= huge_pte_lock(h
, mm
, ptep
);
3383 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3388 pte
= huge_ptep_get(ptep
);
3389 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
3393 if (unlikely(is_hugetlb_entry_migration(pte
))) {
3394 swp_entry_t entry
= pte_to_swp_entry(pte
);
3396 if (is_write_migration_entry(entry
)) {
3399 make_migration_entry_read(&entry
);
3400 newpte
= swp_entry_to_pte(entry
);
3401 set_huge_pte_at(mm
, address
, ptep
, newpte
);
3407 if (!huge_pte_none(pte
)) {
3408 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3409 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3410 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3411 set_huge_pte_at(mm
, address
, ptep
, pte
);
3417 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3418 * may have cleared our pud entry and done put_page on the page table:
3419 * once we release i_mmap_rwsem, another task can do the final put_page
3420 * and that page table be reused and filled with junk.
3422 flush_tlb_range(vma
, start
, end
);
3423 mmu_notifier_invalidate_range(mm
, start
, end
);
3424 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
3425 mmu_notifier_invalidate_range_end(mm
, start
, end
);
3427 return pages
<< h
->order
;
3430 int hugetlb_reserve_pages(struct inode
*inode
,
3432 struct vm_area_struct
*vma
,
3433 vm_flags_t vm_flags
)
3436 struct hstate
*h
= hstate_inode(inode
);
3437 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3438 struct resv_map
*resv_map
;
3441 * Only apply hugepage reservation if asked. At fault time, an
3442 * attempt will be made for VM_NORESERVE to allocate a page
3443 * without using reserves
3445 if (vm_flags
& VM_NORESERVE
)
3449 * Shared mappings base their reservation on the number of pages that
3450 * are already allocated on behalf of the file. Private mappings need
3451 * to reserve the full area even if read-only as mprotect() may be
3452 * called to make the mapping read-write. Assume !vma is a shm mapping
3454 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
3455 resv_map
= inode_resv_map(inode
);
3457 chg
= region_chg(resv_map
, from
, to
);
3460 resv_map
= resv_map_alloc();
3466 set_vma_resv_map(vma
, resv_map
);
3467 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3475 /* There must be enough pages in the subpool for the mapping */
3476 if (hugepage_subpool_get_pages(spool
, chg
)) {
3482 * Check enough hugepages are available for the reservation.
3483 * Hand the pages back to the subpool if there are not
3485 ret
= hugetlb_acct_memory(h
, chg
);
3487 hugepage_subpool_put_pages(spool
, chg
);
3492 * Account for the reservations made. Shared mappings record regions
3493 * that have reservations as they are shared by multiple VMAs.
3494 * When the last VMA disappears, the region map says how much
3495 * the reservation was and the page cache tells how much of
3496 * the reservation was consumed. Private mappings are per-VMA and
3497 * only the consumed reservations are tracked. When the VMA
3498 * disappears, the original reservation is the VMA size and the
3499 * consumed reservations are stored in the map. Hence, nothing
3500 * else has to be done for private mappings here
3502 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3503 region_add(resv_map
, from
, to
);
3506 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3507 kref_put(&resv_map
->refs
, resv_map_release
);
3511 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3513 struct hstate
*h
= hstate_inode(inode
);
3514 struct resv_map
*resv_map
= inode_resv_map(inode
);
3516 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3519 chg
= region_truncate(resv_map
, offset
);
3520 spin_lock(&inode
->i_lock
);
3521 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3522 spin_unlock(&inode
->i_lock
);
3524 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3525 hugetlb_acct_memory(h
, -(chg
- freed
));
3528 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3529 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3530 struct vm_area_struct
*vma
,
3531 unsigned long addr
, pgoff_t idx
)
3533 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3535 unsigned long sbase
= saddr
& PUD_MASK
;
3536 unsigned long s_end
= sbase
+ PUD_SIZE
;
3538 /* Allow segments to share if only one is marked locked */
3539 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3540 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3543 * match the virtual addresses, permission and the alignment of the
3546 if (pmd_index(addr
) != pmd_index(saddr
) ||
3547 vm_flags
!= svm_flags
||
3548 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3554 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3556 unsigned long base
= addr
& PUD_MASK
;
3557 unsigned long end
= base
+ PUD_SIZE
;
3560 * check on proper vm_flags and page table alignment
3562 if (vma
->vm_flags
& VM_MAYSHARE
&&
3563 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3569 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3570 * and returns the corresponding pte. While this is not necessary for the
3571 * !shared pmd case because we can allocate the pmd later as well, it makes the
3572 * code much cleaner. pmd allocation is essential for the shared case because
3573 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3574 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3575 * bad pmd for sharing.
3577 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3579 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3580 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3581 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3583 struct vm_area_struct
*svma
;
3584 unsigned long saddr
;
3589 if (!vma_shareable(vma
, addr
))
3590 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3592 i_mmap_lock_write(mapping
);
3593 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3597 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3599 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3602 get_page(virt_to_page(spte
));
3611 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
3613 if (pud_none(*pud
)) {
3614 pud_populate(mm
, pud
,
3615 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3617 put_page(virt_to_page(spte
));
3622 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3623 i_mmap_unlock_write(mapping
);
3628 * unmap huge page backed by shared pte.
3630 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3631 * indicated by page_count > 1, unmap is achieved by clearing pud and
3632 * decrementing the ref count. If count == 1, the pte page is not shared.
3634 * called with page table lock held.
3636 * returns: 1 successfully unmapped a shared pte page
3637 * 0 the underlying pte page is not shared, or it is the last user
3639 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3641 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3642 pud_t
*pud
= pud_offset(pgd
, *addr
);
3644 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3645 if (page_count(virt_to_page(ptep
)) == 1)
3649 put_page(virt_to_page(ptep
));
3651 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3654 #define want_pmd_share() (1)
3655 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3656 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3660 #define want_pmd_share() (0)
3661 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3663 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3664 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3665 unsigned long addr
, unsigned long sz
)
3671 pgd
= pgd_offset(mm
, addr
);
3672 pud
= pud_alloc(mm
, pgd
, addr
);
3674 if (sz
== PUD_SIZE
) {
3677 BUG_ON(sz
!= PMD_SIZE
);
3678 if (want_pmd_share() && pud_none(*pud
))
3679 pte
= huge_pmd_share(mm
, addr
, pud
);
3681 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3684 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3689 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3695 pgd
= pgd_offset(mm
, addr
);
3696 if (pgd_present(*pgd
)) {
3697 pud
= pud_offset(pgd
, addr
);
3698 if (pud_present(*pud
)) {
3700 return (pte_t
*)pud
;
3701 pmd
= pmd_offset(pud
, addr
);
3704 return (pte_t
*) pmd
;
3707 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3710 * These functions are overwritable if your architecture needs its own
3713 struct page
* __weak
3714 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
3717 return ERR_PTR(-EINVAL
);
3720 struct page
* __weak
3721 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3722 pmd_t
*pmd
, int flags
)
3724 struct page
*page
= NULL
;
3727 ptl
= pmd_lockptr(mm
, pmd
);
3730 * make sure that the address range covered by this pmd is not
3731 * unmapped from other threads.
3733 if (!pmd_huge(*pmd
))
3735 if (pmd_present(*pmd
)) {
3736 page
= pte_page(*(pte_t
*)pmd
) +
3737 ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3738 if (flags
& FOLL_GET
)
3741 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t
*)pmd
))) {
3743 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
3747 * hwpoisoned entry is treated as no_page_table in
3748 * follow_page_mask().
3756 struct page
* __weak
3757 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3758 pud_t
*pud
, int flags
)
3760 if (flags
& FOLL_GET
)
3763 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3766 #ifdef CONFIG_MEMORY_FAILURE
3768 /* Should be called in hugetlb_lock */
3769 static int is_hugepage_on_freelist(struct page
*hpage
)
3773 struct hstate
*h
= page_hstate(hpage
);
3774 int nid
= page_to_nid(hpage
);
3776 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3783 * This function is called from memory failure code.
3784 * Assume the caller holds page lock of the head page.
3786 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3788 struct hstate
*h
= page_hstate(hpage
);
3789 int nid
= page_to_nid(hpage
);
3792 spin_lock(&hugetlb_lock
);
3793 if (is_hugepage_on_freelist(hpage
)) {
3795 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3796 * but dangling hpage->lru can trigger list-debug warnings
3797 * (this happens when we call unpoison_memory() on it),
3798 * so let it point to itself with list_del_init().
3800 list_del_init(&hpage
->lru
);
3801 set_page_refcounted(hpage
);
3802 h
->free_huge_pages
--;
3803 h
->free_huge_pages_node
[nid
]--;
3806 spin_unlock(&hugetlb_lock
);
3811 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3813 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3814 if (!get_page_unless_zero(page
))
3816 spin_lock(&hugetlb_lock
);
3817 list_move_tail(&page
->lru
, list
);
3818 spin_unlock(&hugetlb_lock
);
3822 void putback_active_hugepage(struct page
*page
)
3824 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3825 spin_lock(&hugetlb_lock
);
3826 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3827 spin_unlock(&hugetlb_lock
);
3831 bool is_hugepage_active(struct page
*page
)
3833 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
3835 * This function can be called for a tail page because the caller,
3836 * scan_movable_pages, scans through a given pfn-range which typically
3837 * covers one memory block. In systems using gigantic hugepage (1GB
3838 * for x86_64,) a hugepage is larger than a memory block, and we don't
3839 * support migrating such large hugepages for now, so return false
3840 * when called for tail pages.
3845 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3846 * so we should return false for them.
3848 if (unlikely(PageHWPoison(page
)))
3850 return page_count(page
) > 0;