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 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
39 unsigned long hugepages_treat_as_movable
;
41 int hugetlb_max_hstate __read_mostly
;
42 unsigned int default_hstate_idx
;
43 struct hstate hstates
[HUGE_MAX_HSTATE
];
45 __initdata
LIST_HEAD(huge_boot_pages
);
47 /* for command line parsing */
48 static struct hstate
* __initdata parsed_hstate
;
49 static unsigned long __initdata default_hstate_max_huge_pages
;
50 static unsigned long __initdata default_hstate_size
;
53 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
54 * free_huge_pages, and surplus_huge_pages.
56 DEFINE_SPINLOCK(hugetlb_lock
);
59 * Serializes faults on the same logical page. This is used to
60 * prevent spurious OOMs when the hugepage pool is fully utilized.
62 static int num_fault_mutexes
;
63 static struct mutex
*htlb_fault_mutex_table ____cacheline_aligned_in_smp
;
65 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
67 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
69 spin_unlock(&spool
->lock
);
71 /* If no pages are used, and no other handles to the subpool
72 * remain, free the subpool the subpool remain */
77 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
79 struct hugepage_subpool
*spool
;
81 spool
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
85 spin_lock_init(&spool
->lock
);
87 spool
->max_hpages
= nr_blocks
;
88 spool
->used_hpages
= 0;
93 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
95 spin_lock(&spool
->lock
);
96 BUG_ON(!spool
->count
);
98 unlock_or_release_subpool(spool
);
101 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
109 spin_lock(&spool
->lock
);
110 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
111 spool
->used_hpages
+= delta
;
115 spin_unlock(&spool
->lock
);
120 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
126 spin_lock(&spool
->lock
);
127 spool
->used_hpages
-= delta
;
128 /* If hugetlbfs_put_super couldn't free spool due to
129 * an outstanding quota reference, free it now. */
130 unlock_or_release_subpool(spool
);
133 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
135 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
138 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
140 return subpool_inode(file_inode(vma
->vm_file
));
144 * Region tracking -- allows tracking of reservations and instantiated pages
145 * across the pages in a mapping.
147 * The region data structures are embedded into a resv_map and
148 * protected by a resv_map's lock
151 struct list_head link
;
156 static long region_add(struct resv_map
*resv
, long f
, long t
)
158 struct list_head
*head
= &resv
->regions
;
159 struct file_region
*rg
, *nrg
, *trg
;
161 spin_lock(&resv
->lock
);
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg
, head
, link
)
167 /* Round our left edge to the current segment if it encloses us. */
171 /* Check for and consume any regions we now overlap with. */
173 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
174 if (&rg
->link
== head
)
179 /* If this area reaches higher then extend our area to
180 * include it completely. If this is not the first area
181 * which we intend to reuse, free it. */
191 spin_unlock(&resv
->lock
);
195 static long region_chg(struct resv_map
*resv
, long f
, long t
)
197 struct list_head
*head
= &resv
->regions
;
198 struct file_region
*rg
, *nrg
= NULL
;
202 spin_lock(&resv
->lock
);
203 /* Locate the region we are before or in. */
204 list_for_each_entry(rg
, head
, link
)
208 /* If we are below the current region then a new region is required.
209 * Subtle, allocate a new region at the position but make it zero
210 * size such that we can guarantee to record the reservation. */
211 if (&rg
->link
== head
|| t
< rg
->from
) {
213 spin_unlock(&resv
->lock
);
214 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
220 INIT_LIST_HEAD(&nrg
->link
);
224 list_add(&nrg
->link
, rg
->link
.prev
);
229 /* Round our left edge to the current segment if it encloses us. */
234 /* Check for and consume any regions we now overlap with. */
235 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
236 if (&rg
->link
== head
)
241 /* We overlap with this area, if it extends further than
242 * us then we must extend ourselves. Account for its
243 * existing reservation. */
248 chg
-= rg
->to
- rg
->from
;
252 spin_unlock(&resv
->lock
);
253 /* We already know we raced and no longer need the new region */
257 spin_unlock(&resv
->lock
);
261 static long region_truncate(struct resv_map
*resv
, long end
)
263 struct list_head
*head
= &resv
->regions
;
264 struct file_region
*rg
, *trg
;
267 spin_lock(&resv
->lock
);
268 /* Locate the region we are either in or before. */
269 list_for_each_entry(rg
, head
, link
)
272 if (&rg
->link
== head
)
275 /* If we are in the middle of a region then adjust it. */
276 if (end
> rg
->from
) {
279 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
282 /* Drop any remaining regions. */
283 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
284 if (&rg
->link
== head
)
286 chg
+= rg
->to
- rg
->from
;
292 spin_unlock(&resv
->lock
);
296 static long region_count(struct resv_map
*resv
, long f
, long t
)
298 struct list_head
*head
= &resv
->regions
;
299 struct file_region
*rg
;
302 spin_lock(&resv
->lock
);
303 /* Locate each segment we overlap with, and count that overlap. */
304 list_for_each_entry(rg
, head
, link
) {
313 seg_from
= max(rg
->from
, f
);
314 seg_to
= min(rg
->to
, t
);
316 chg
+= seg_to
- seg_from
;
318 spin_unlock(&resv
->lock
);
324 * Convert the address within this vma to the page offset within
325 * the mapping, in pagecache page units; huge pages here.
327 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
328 struct vm_area_struct
*vma
, unsigned long address
)
330 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
331 (vma
->vm_pgoff
>> huge_page_order(h
));
334 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
335 unsigned long address
)
337 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
341 * Return the size of the pages allocated when backing a VMA. In the majority
342 * cases this will be same size as used by the page table entries.
344 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
346 struct hstate
*hstate
;
348 if (!is_vm_hugetlb_page(vma
))
351 hstate
= hstate_vma(vma
);
353 return 1UL << huge_page_shift(hstate
);
355 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
358 * Return the page size being used by the MMU to back a VMA. In the majority
359 * of cases, the page size used by the kernel matches the MMU size. On
360 * architectures where it differs, an architecture-specific version of this
361 * function is required.
363 #ifndef vma_mmu_pagesize
364 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
366 return vma_kernel_pagesize(vma
);
371 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
372 * bits of the reservation map pointer, which are always clear due to
375 #define HPAGE_RESV_OWNER (1UL << 0)
376 #define HPAGE_RESV_UNMAPPED (1UL << 1)
377 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
380 * These helpers are used to track how many pages are reserved for
381 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
382 * is guaranteed to have their future faults succeed.
384 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
385 * the reserve counters are updated with the hugetlb_lock held. It is safe
386 * to reset the VMA at fork() time as it is not in use yet and there is no
387 * chance of the global counters getting corrupted as a result of the values.
389 * The private mapping reservation is represented in a subtly different
390 * manner to a shared mapping. A shared mapping has a region map associated
391 * with the underlying file, this region map represents the backing file
392 * pages which have ever had a reservation assigned which this persists even
393 * after the page is instantiated. A private mapping has a region map
394 * associated with the original mmap which is attached to all VMAs which
395 * reference it, this region map represents those offsets which have consumed
396 * reservation ie. where pages have been instantiated.
398 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
400 return (unsigned long)vma
->vm_private_data
;
403 static void set_vma_private_data(struct vm_area_struct
*vma
,
406 vma
->vm_private_data
= (void *)value
;
409 struct resv_map
*resv_map_alloc(void)
411 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
415 kref_init(&resv_map
->refs
);
416 spin_lock_init(&resv_map
->lock
);
417 INIT_LIST_HEAD(&resv_map
->regions
);
422 void resv_map_release(struct kref
*ref
)
424 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
426 /* Clear out any active regions before we release the map. */
427 region_truncate(resv_map
, 0);
431 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
433 return inode
->i_mapping
->private_data
;
436 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
438 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
439 if (vma
->vm_flags
& VM_MAYSHARE
) {
440 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
441 struct inode
*inode
= mapping
->host
;
443 return inode_resv_map(inode
);
446 return (struct resv_map
*)(get_vma_private_data(vma
) &
451 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
453 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
454 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
456 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
457 HPAGE_RESV_MASK
) | (unsigned long)map
);
460 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
462 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
463 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
465 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
468 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
470 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
472 return (get_vma_private_data(vma
) & flag
) != 0;
475 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
476 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
478 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
479 if (!(vma
->vm_flags
& VM_MAYSHARE
))
480 vma
->vm_private_data
= (void *)0;
483 /* Returns true if the VMA has associated reserve pages */
484 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
486 if (vma
->vm_flags
& VM_NORESERVE
) {
488 * This address is already reserved by other process(chg == 0),
489 * so, we should decrement reserved count. Without decrementing,
490 * reserve count remains after releasing inode, because this
491 * allocated page will go into page cache and is regarded as
492 * coming from reserved pool in releasing step. Currently, we
493 * don't have any other solution to deal with this situation
494 * properly, so add work-around here.
496 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
502 /* Shared mappings always use reserves */
503 if (vma
->vm_flags
& VM_MAYSHARE
)
507 * Only the process that called mmap() has reserves for
510 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
516 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
518 int nid
= page_to_nid(page
);
519 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
520 h
->free_huge_pages
++;
521 h
->free_huge_pages_node
[nid
]++;
524 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
528 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
529 if (!is_migrate_isolate_page(page
))
532 * if 'non-isolated free hugepage' not found on the list,
533 * the allocation fails.
535 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
537 list_move(&page
->lru
, &h
->hugepage_activelist
);
538 set_page_refcounted(page
);
539 h
->free_huge_pages
--;
540 h
->free_huge_pages_node
[nid
]--;
544 /* Movability of hugepages depends on migration support. */
545 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
547 if (hugepages_treat_as_movable
|| hugepage_migration_support(h
))
548 return GFP_HIGHUSER_MOVABLE
;
553 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
554 struct vm_area_struct
*vma
,
555 unsigned long address
, int avoid_reserve
,
558 struct page
*page
= NULL
;
559 struct mempolicy
*mpol
;
560 nodemask_t
*nodemask
;
561 struct zonelist
*zonelist
;
564 unsigned int cpuset_mems_cookie
;
567 * A child process with MAP_PRIVATE mappings created by their parent
568 * have no page reserves. This check ensures that reservations are
569 * not "stolen". The child may still get SIGKILLed
571 if (!vma_has_reserves(vma
, chg
) &&
572 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
575 /* If reserves cannot be used, ensure enough pages are in the pool */
576 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
580 cpuset_mems_cookie
= read_mems_allowed_begin();
581 zonelist
= huge_zonelist(vma
, address
,
582 htlb_alloc_mask(h
), &mpol
, &nodemask
);
584 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
585 MAX_NR_ZONES
- 1, nodemask
) {
586 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask(h
))) {
587 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
591 if (!vma_has_reserves(vma
, chg
))
594 SetPagePrivate(page
);
595 h
->resv_huge_pages
--;
602 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
610 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
614 VM_BUG_ON(h
->order
>= MAX_ORDER
);
617 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
618 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
619 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
620 1 << PG_referenced
| 1 << PG_dirty
|
621 1 << PG_active
| 1 << PG_reserved
|
622 1 << PG_private
| 1 << PG_writeback
);
624 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
625 set_compound_page_dtor(page
, NULL
);
626 set_page_refcounted(page
);
627 arch_release_hugepage(page
);
628 __free_pages(page
, huge_page_order(h
));
631 struct hstate
*size_to_hstate(unsigned long size
)
636 if (huge_page_size(h
) == size
)
642 static void free_huge_page(struct page
*page
)
645 * Can't pass hstate in here because it is called from the
646 * compound page destructor.
648 struct hstate
*h
= page_hstate(page
);
649 int nid
= page_to_nid(page
);
650 struct hugepage_subpool
*spool
=
651 (struct hugepage_subpool
*)page_private(page
);
652 bool restore_reserve
;
654 set_page_private(page
, 0);
655 page
->mapping
= NULL
;
656 BUG_ON(page_count(page
));
657 BUG_ON(page_mapcount(page
));
658 restore_reserve
= PagePrivate(page
);
659 ClearPagePrivate(page
);
661 spin_lock(&hugetlb_lock
);
662 hugetlb_cgroup_uncharge_page(hstate_index(h
),
663 pages_per_huge_page(h
), page
);
665 h
->resv_huge_pages
++;
667 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
668 /* remove the page from active list */
669 list_del(&page
->lru
);
670 update_and_free_page(h
, page
);
671 h
->surplus_huge_pages
--;
672 h
->surplus_huge_pages_node
[nid
]--;
674 arch_clear_hugepage_flags(page
);
675 enqueue_huge_page(h
, page
);
677 spin_unlock(&hugetlb_lock
);
678 hugepage_subpool_put_pages(spool
, 1);
681 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
683 INIT_LIST_HEAD(&page
->lru
);
684 set_compound_page_dtor(page
, free_huge_page
);
685 spin_lock(&hugetlb_lock
);
686 set_hugetlb_cgroup(page
, NULL
);
688 h
->nr_huge_pages_node
[nid
]++;
689 spin_unlock(&hugetlb_lock
);
690 put_page(page
); /* free it into the hugepage allocator */
693 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
696 int nr_pages
= 1 << order
;
697 struct page
*p
= page
+ 1;
699 /* we rely on prep_new_huge_page to set the destructor */
700 set_compound_order(page
, order
);
702 __ClearPageReserved(page
);
703 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
706 * For gigantic hugepages allocated through bootmem at
707 * boot, it's safer to be consistent with the not-gigantic
708 * hugepages and clear the PG_reserved bit from all tail pages
709 * too. Otherwse drivers using get_user_pages() to access tail
710 * pages may get the reference counting wrong if they see
711 * PG_reserved set on a tail page (despite the head page not
712 * having PG_reserved set). Enforcing this consistency between
713 * head and tail pages allows drivers to optimize away a check
714 * on the head page when they need know if put_page() is needed
715 * after get_user_pages().
717 __ClearPageReserved(p
);
718 set_page_count(p
, 0);
719 p
->first_page
= page
;
724 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
725 * transparent huge pages. See the PageTransHuge() documentation for more
728 int PageHuge(struct page
*page
)
730 if (!PageCompound(page
))
733 page
= compound_head(page
);
734 return get_compound_page_dtor(page
) == free_huge_page
;
736 EXPORT_SYMBOL_GPL(PageHuge
);
739 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
740 * normal or transparent huge pages.
742 int PageHeadHuge(struct page
*page_head
)
744 if (!PageHead(page_head
))
747 return get_compound_page_dtor(page_head
) == free_huge_page
;
750 pgoff_t
__basepage_index(struct page
*page
)
752 struct page
*page_head
= compound_head(page
);
753 pgoff_t index
= page_index(page_head
);
754 unsigned long compound_idx
;
756 if (!PageHuge(page_head
))
757 return page_index(page
);
759 if (compound_order(page_head
) >= MAX_ORDER
)
760 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
762 compound_idx
= page
- page_head
;
764 return (index
<< compound_order(page_head
)) + compound_idx
;
767 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
771 if (h
->order
>= MAX_ORDER
)
774 page
= alloc_pages_exact_node(nid
,
775 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
776 __GFP_REPEAT
|__GFP_NOWARN
,
779 if (arch_prepare_hugepage(page
)) {
780 __free_pages(page
, huge_page_order(h
));
783 prep_new_huge_page(h
, page
, nid
);
790 * common helper functions for hstate_next_node_to_{alloc|free}.
791 * We may have allocated or freed a huge page based on a different
792 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
793 * be outside of *nodes_allowed. Ensure that we use an allowed
794 * node for alloc or free.
796 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
798 nid
= next_node(nid
, *nodes_allowed
);
799 if (nid
== MAX_NUMNODES
)
800 nid
= first_node(*nodes_allowed
);
801 VM_BUG_ON(nid
>= MAX_NUMNODES
);
806 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
808 if (!node_isset(nid
, *nodes_allowed
))
809 nid
= next_node_allowed(nid
, nodes_allowed
);
814 * returns the previously saved node ["this node"] from which to
815 * allocate a persistent huge page for the pool and advance the
816 * next node from which to allocate, handling wrap at end of node
819 static int hstate_next_node_to_alloc(struct hstate
*h
,
820 nodemask_t
*nodes_allowed
)
824 VM_BUG_ON(!nodes_allowed
);
826 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
827 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
833 * helper for free_pool_huge_page() - return the previously saved
834 * node ["this node"] from which to free a huge page. Advance the
835 * next node id whether or not we find a free huge page to free so
836 * that the next attempt to free addresses the next node.
838 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
842 VM_BUG_ON(!nodes_allowed
);
844 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
845 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
850 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
851 for (nr_nodes = nodes_weight(*mask); \
853 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
856 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
857 for (nr_nodes = nodes_weight(*mask); \
859 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
862 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
868 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
869 page
= alloc_fresh_huge_page_node(h
, node
);
877 count_vm_event(HTLB_BUDDY_PGALLOC
);
879 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
885 * Free huge page from pool from next node to free.
886 * Attempt to keep persistent huge pages more or less
887 * balanced over allowed nodes.
888 * Called with hugetlb_lock locked.
890 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
896 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
898 * If we're returning unused surplus pages, only examine
899 * nodes with surplus pages.
901 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
902 !list_empty(&h
->hugepage_freelists
[node
])) {
904 list_entry(h
->hugepage_freelists
[node
].next
,
906 list_del(&page
->lru
);
907 h
->free_huge_pages
--;
908 h
->free_huge_pages_node
[node
]--;
910 h
->surplus_huge_pages
--;
911 h
->surplus_huge_pages_node
[node
]--;
913 update_and_free_page(h
, page
);
923 * Dissolve a given free hugepage into free buddy pages. This function does
924 * nothing for in-use (including surplus) hugepages.
926 static void dissolve_free_huge_page(struct page
*page
)
928 spin_lock(&hugetlb_lock
);
929 if (PageHuge(page
) && !page_count(page
)) {
930 struct hstate
*h
= page_hstate(page
);
931 int nid
= page_to_nid(page
);
932 list_del(&page
->lru
);
933 h
->free_huge_pages
--;
934 h
->free_huge_pages_node
[nid
]--;
935 update_and_free_page(h
, page
);
937 spin_unlock(&hugetlb_lock
);
941 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
942 * make specified memory blocks removable from the system.
943 * Note that start_pfn should aligned with (minimum) hugepage size.
945 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
947 unsigned int order
= 8 * sizeof(void *);
951 /* Set scan step to minimum hugepage size */
953 if (order
> huge_page_order(h
))
954 order
= huge_page_order(h
);
955 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << order
));
956 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << order
)
957 dissolve_free_huge_page(pfn_to_page(pfn
));
960 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
965 if (h
->order
>= MAX_ORDER
)
969 * Assume we will successfully allocate the surplus page to
970 * prevent racing processes from causing the surplus to exceed
973 * This however introduces a different race, where a process B
974 * tries to grow the static hugepage pool while alloc_pages() is
975 * called by process A. B will only examine the per-node
976 * counters in determining if surplus huge pages can be
977 * converted to normal huge pages in adjust_pool_surplus(). A
978 * won't be able to increment the per-node counter, until the
979 * lock is dropped by B, but B doesn't drop hugetlb_lock until
980 * no more huge pages can be converted from surplus to normal
981 * state (and doesn't try to convert again). Thus, we have a
982 * case where a surplus huge page exists, the pool is grown, and
983 * the surplus huge page still exists after, even though it
984 * should just have been converted to a normal huge page. This
985 * does not leak memory, though, as the hugepage will be freed
986 * once it is out of use. It also does not allow the counters to
987 * go out of whack in adjust_pool_surplus() as we don't modify
988 * the node values until we've gotten the hugepage and only the
989 * per-node value is checked there.
991 spin_lock(&hugetlb_lock
);
992 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
993 spin_unlock(&hugetlb_lock
);
997 h
->surplus_huge_pages
++;
999 spin_unlock(&hugetlb_lock
);
1001 if (nid
== NUMA_NO_NODE
)
1002 page
= alloc_pages(htlb_alloc_mask(h
)|__GFP_COMP
|
1003 __GFP_REPEAT
|__GFP_NOWARN
,
1004 huge_page_order(h
));
1006 page
= alloc_pages_exact_node(nid
,
1007 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1008 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
1010 if (page
&& arch_prepare_hugepage(page
)) {
1011 __free_pages(page
, huge_page_order(h
));
1015 spin_lock(&hugetlb_lock
);
1017 INIT_LIST_HEAD(&page
->lru
);
1018 r_nid
= page_to_nid(page
);
1019 set_compound_page_dtor(page
, free_huge_page
);
1020 set_hugetlb_cgroup(page
, NULL
);
1022 * We incremented the global counters already
1024 h
->nr_huge_pages_node
[r_nid
]++;
1025 h
->surplus_huge_pages_node
[r_nid
]++;
1026 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1029 h
->surplus_huge_pages
--;
1030 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1032 spin_unlock(&hugetlb_lock
);
1038 * This allocation function is useful in the context where vma is irrelevant.
1039 * E.g. soft-offlining uses this function because it only cares physical
1040 * address of error page.
1042 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1044 struct page
*page
= NULL
;
1046 spin_lock(&hugetlb_lock
);
1047 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1048 page
= dequeue_huge_page_node(h
, nid
);
1049 spin_unlock(&hugetlb_lock
);
1052 page
= alloc_buddy_huge_page(h
, nid
);
1058 * Increase the hugetlb pool such that it can accommodate a reservation
1061 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1063 struct list_head surplus_list
;
1064 struct page
*page
, *tmp
;
1066 int needed
, allocated
;
1067 bool alloc_ok
= true;
1069 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1071 h
->resv_huge_pages
+= delta
;
1076 INIT_LIST_HEAD(&surplus_list
);
1080 spin_unlock(&hugetlb_lock
);
1081 for (i
= 0; i
< needed
; i
++) {
1082 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1087 list_add(&page
->lru
, &surplus_list
);
1092 * After retaking hugetlb_lock, we need to recalculate 'needed'
1093 * because either resv_huge_pages or free_huge_pages may have changed.
1095 spin_lock(&hugetlb_lock
);
1096 needed
= (h
->resv_huge_pages
+ delta
) -
1097 (h
->free_huge_pages
+ allocated
);
1102 * We were not able to allocate enough pages to
1103 * satisfy the entire reservation so we free what
1104 * we've allocated so far.
1109 * The surplus_list now contains _at_least_ the number of extra pages
1110 * needed to accommodate the reservation. Add the appropriate number
1111 * of pages to the hugetlb pool and free the extras back to the buddy
1112 * allocator. Commit the entire reservation here to prevent another
1113 * process from stealing the pages as they are added to the pool but
1114 * before they are reserved.
1116 needed
+= allocated
;
1117 h
->resv_huge_pages
+= delta
;
1120 /* Free the needed pages to the hugetlb pool */
1121 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1125 * This page is now managed by the hugetlb allocator and has
1126 * no users -- drop the buddy allocator's reference.
1128 put_page_testzero(page
);
1129 VM_BUG_ON_PAGE(page_count(page
), page
);
1130 enqueue_huge_page(h
, page
);
1133 spin_unlock(&hugetlb_lock
);
1135 /* Free unnecessary surplus pages to the buddy allocator */
1136 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1138 spin_lock(&hugetlb_lock
);
1144 * When releasing a hugetlb pool reservation, any surplus pages that were
1145 * allocated to satisfy the reservation must be explicitly freed if they were
1147 * Called with hugetlb_lock held.
1149 static void return_unused_surplus_pages(struct hstate
*h
,
1150 unsigned long unused_resv_pages
)
1152 unsigned long nr_pages
;
1154 /* Uncommit the reservation */
1155 h
->resv_huge_pages
-= unused_resv_pages
;
1157 /* Cannot return gigantic pages currently */
1158 if (h
->order
>= MAX_ORDER
)
1161 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1164 * We want to release as many surplus pages as possible, spread
1165 * evenly across all nodes with memory. Iterate across these nodes
1166 * until we can no longer free unreserved surplus pages. This occurs
1167 * when the nodes with surplus pages have no free pages.
1168 * free_pool_huge_page() will balance the the freed pages across the
1169 * on-line nodes with memory and will handle the hstate accounting.
1171 while (nr_pages
--) {
1172 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1174 cond_resched_lock(&hugetlb_lock
);
1179 * Determine if the huge page at addr within the vma has an associated
1180 * reservation. Where it does not we will need to logically increase
1181 * reservation and actually increase subpool usage before an allocation
1182 * can occur. Where any new reservation would be required the
1183 * reservation change is prepared, but not committed. Once the page
1184 * has been allocated from the subpool and instantiated the change should
1185 * be committed via vma_commit_reservation. No action is required on
1188 static long vma_needs_reservation(struct hstate
*h
,
1189 struct vm_area_struct
*vma
, unsigned long addr
)
1191 struct resv_map
*resv
;
1195 resv
= vma_resv_map(vma
);
1199 idx
= vma_hugecache_offset(h
, vma
, addr
);
1200 chg
= region_chg(resv
, idx
, idx
+ 1);
1202 if (vma
->vm_flags
& VM_MAYSHARE
)
1205 return chg
< 0 ? chg
: 0;
1207 static void vma_commit_reservation(struct hstate
*h
,
1208 struct vm_area_struct
*vma
, unsigned long addr
)
1210 struct resv_map
*resv
;
1213 resv
= vma_resv_map(vma
);
1217 idx
= vma_hugecache_offset(h
, vma
, addr
);
1218 region_add(resv
, idx
, idx
+ 1);
1221 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1222 unsigned long addr
, int avoid_reserve
)
1224 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1225 struct hstate
*h
= hstate_vma(vma
);
1229 struct hugetlb_cgroup
*h_cg
;
1231 idx
= hstate_index(h
);
1233 * Processes that did not create the mapping will have no
1234 * reserves and will not have accounted against subpool
1235 * limit. Check that the subpool limit can be made before
1236 * satisfying the allocation MAP_NORESERVE mappings may also
1237 * need pages and subpool limit allocated allocated if no reserve
1240 chg
= vma_needs_reservation(h
, vma
, addr
);
1242 return ERR_PTR(-ENOMEM
);
1243 if (chg
|| avoid_reserve
)
1244 if (hugepage_subpool_get_pages(spool
, 1))
1245 return ERR_PTR(-ENOSPC
);
1247 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1249 if (chg
|| avoid_reserve
)
1250 hugepage_subpool_put_pages(spool
, 1);
1251 return ERR_PTR(-ENOSPC
);
1253 spin_lock(&hugetlb_lock
);
1254 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1256 spin_unlock(&hugetlb_lock
);
1257 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1259 hugetlb_cgroup_uncharge_cgroup(idx
,
1260 pages_per_huge_page(h
),
1262 if (chg
|| avoid_reserve
)
1263 hugepage_subpool_put_pages(spool
, 1);
1264 return ERR_PTR(-ENOSPC
);
1266 spin_lock(&hugetlb_lock
);
1267 list_move(&page
->lru
, &h
->hugepage_activelist
);
1270 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1271 spin_unlock(&hugetlb_lock
);
1273 set_page_private(page
, (unsigned long)spool
);
1275 vma_commit_reservation(h
, vma
, addr
);
1280 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1281 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1282 * where no ERR_VALUE is expected to be returned.
1284 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1285 unsigned long addr
, int avoid_reserve
)
1287 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1293 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1295 struct huge_bootmem_page
*m
;
1298 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1301 addr
= memblock_virt_alloc_try_nid_nopanic(
1302 huge_page_size(h
), huge_page_size(h
),
1303 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1306 * Use the beginning of the huge page to store the
1307 * huge_bootmem_page struct (until gather_bootmem
1308 * puts them into the mem_map).
1317 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1318 /* Put them into a private list first because mem_map is not up yet */
1319 list_add(&m
->list
, &huge_boot_pages
);
1324 static void __init
prep_compound_huge_page(struct page
*page
, int order
)
1326 if (unlikely(order
> (MAX_ORDER
- 1)))
1327 prep_compound_gigantic_page(page
, order
);
1329 prep_compound_page(page
, order
);
1332 /* Put bootmem huge pages into the standard lists after mem_map is up */
1333 static void __init
gather_bootmem_prealloc(void)
1335 struct huge_bootmem_page
*m
;
1337 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1338 struct hstate
*h
= m
->hstate
;
1341 #ifdef CONFIG_HIGHMEM
1342 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1343 memblock_free_late(__pa(m
),
1344 sizeof(struct huge_bootmem_page
));
1346 page
= virt_to_page(m
);
1348 WARN_ON(page_count(page
) != 1);
1349 prep_compound_huge_page(page
, h
->order
);
1350 WARN_ON(PageReserved(page
));
1351 prep_new_huge_page(h
, page
, page_to_nid(page
));
1353 * If we had gigantic hugepages allocated at boot time, we need
1354 * to restore the 'stolen' pages to totalram_pages in order to
1355 * fix confusing memory reports from free(1) and another
1356 * side-effects, like CommitLimit going negative.
1358 if (h
->order
> (MAX_ORDER
- 1))
1359 adjust_managed_page_count(page
, 1 << h
->order
);
1363 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1367 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1368 if (h
->order
>= MAX_ORDER
) {
1369 if (!alloc_bootmem_huge_page(h
))
1371 } else if (!alloc_fresh_huge_page(h
,
1372 &node_states
[N_MEMORY
]))
1375 h
->max_huge_pages
= i
;
1378 static void __init
hugetlb_init_hstates(void)
1382 for_each_hstate(h
) {
1383 /* oversize hugepages were init'ed in early boot */
1384 if (h
->order
< MAX_ORDER
)
1385 hugetlb_hstate_alloc_pages(h
);
1389 static char * __init
memfmt(char *buf
, unsigned long n
)
1391 if (n
>= (1UL << 30))
1392 sprintf(buf
, "%lu GB", n
>> 30);
1393 else if (n
>= (1UL << 20))
1394 sprintf(buf
, "%lu MB", n
>> 20);
1396 sprintf(buf
, "%lu KB", n
>> 10);
1400 static void __init
report_hugepages(void)
1404 for_each_hstate(h
) {
1406 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1407 memfmt(buf
, huge_page_size(h
)),
1408 h
->free_huge_pages
);
1412 #ifdef CONFIG_HIGHMEM
1413 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1414 nodemask_t
*nodes_allowed
)
1418 if (h
->order
>= MAX_ORDER
)
1421 for_each_node_mask(i
, *nodes_allowed
) {
1422 struct page
*page
, *next
;
1423 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1424 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1425 if (count
>= h
->nr_huge_pages
)
1427 if (PageHighMem(page
))
1429 list_del(&page
->lru
);
1430 update_and_free_page(h
, page
);
1431 h
->free_huge_pages
--;
1432 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1437 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1438 nodemask_t
*nodes_allowed
)
1444 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1445 * balanced by operating on them in a round-robin fashion.
1446 * Returns 1 if an adjustment was made.
1448 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1453 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1456 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1457 if (h
->surplus_huge_pages_node
[node
])
1461 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1462 if (h
->surplus_huge_pages_node
[node
] <
1463 h
->nr_huge_pages_node
[node
])
1470 h
->surplus_huge_pages
+= delta
;
1471 h
->surplus_huge_pages_node
[node
] += delta
;
1475 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1476 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1477 nodemask_t
*nodes_allowed
)
1479 unsigned long min_count
, ret
;
1481 if (h
->order
>= MAX_ORDER
)
1482 return h
->max_huge_pages
;
1485 * Increase the pool size
1486 * First take pages out of surplus state. Then make up the
1487 * remaining difference by allocating fresh huge pages.
1489 * We might race with alloc_buddy_huge_page() here and be unable
1490 * to convert a surplus huge page to a normal huge page. That is
1491 * not critical, though, it just means the overall size of the
1492 * pool might be one hugepage larger than it needs to be, but
1493 * within all the constraints specified by the sysctls.
1495 spin_lock(&hugetlb_lock
);
1496 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1497 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1501 while (count
> persistent_huge_pages(h
)) {
1503 * If this allocation races such that we no longer need the
1504 * page, free_huge_page will handle it by freeing the page
1505 * and reducing the surplus.
1507 spin_unlock(&hugetlb_lock
);
1508 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1509 spin_lock(&hugetlb_lock
);
1513 /* Bail for signals. Probably ctrl-c from user */
1514 if (signal_pending(current
))
1519 * Decrease the pool size
1520 * First return free pages to the buddy allocator (being careful
1521 * to keep enough around to satisfy reservations). Then place
1522 * pages into surplus state as needed so the pool will shrink
1523 * to the desired size as pages become free.
1525 * By placing pages into the surplus state independent of the
1526 * overcommit value, we are allowing the surplus pool size to
1527 * exceed overcommit. There are few sane options here. Since
1528 * alloc_buddy_huge_page() is checking the global counter,
1529 * though, we'll note that we're not allowed to exceed surplus
1530 * and won't grow the pool anywhere else. Not until one of the
1531 * sysctls are changed, or the surplus pages go out of use.
1533 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1534 min_count
= max(count
, min_count
);
1535 try_to_free_low(h
, min_count
, nodes_allowed
);
1536 while (min_count
< persistent_huge_pages(h
)) {
1537 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1539 cond_resched_lock(&hugetlb_lock
);
1541 while (count
< persistent_huge_pages(h
)) {
1542 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1546 ret
= persistent_huge_pages(h
);
1547 spin_unlock(&hugetlb_lock
);
1551 #define HSTATE_ATTR_RO(_name) \
1552 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1554 #define HSTATE_ATTR(_name) \
1555 static struct kobj_attribute _name##_attr = \
1556 __ATTR(_name, 0644, _name##_show, _name##_store)
1558 static struct kobject
*hugepages_kobj
;
1559 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1561 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1563 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1567 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1568 if (hstate_kobjs
[i
] == kobj
) {
1570 *nidp
= NUMA_NO_NODE
;
1574 return kobj_to_node_hstate(kobj
, nidp
);
1577 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1578 struct kobj_attribute
*attr
, char *buf
)
1581 unsigned long nr_huge_pages
;
1584 h
= kobj_to_hstate(kobj
, &nid
);
1585 if (nid
== NUMA_NO_NODE
)
1586 nr_huge_pages
= h
->nr_huge_pages
;
1588 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1590 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1593 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1594 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1595 const char *buf
, size_t len
)
1599 unsigned long count
;
1601 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1603 err
= kstrtoul(buf
, 10, &count
);
1607 h
= kobj_to_hstate(kobj
, &nid
);
1608 if (h
->order
>= MAX_ORDER
) {
1613 if (nid
== NUMA_NO_NODE
) {
1615 * global hstate attribute
1617 if (!(obey_mempolicy
&&
1618 init_nodemask_of_mempolicy(nodes_allowed
))) {
1619 NODEMASK_FREE(nodes_allowed
);
1620 nodes_allowed
= &node_states
[N_MEMORY
];
1622 } else if (nodes_allowed
) {
1624 * per node hstate attribute: adjust count to global,
1625 * but restrict alloc/free to the specified node.
1627 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1628 init_nodemask_of_node(nodes_allowed
, nid
);
1630 nodes_allowed
= &node_states
[N_MEMORY
];
1632 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1634 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1635 NODEMASK_FREE(nodes_allowed
);
1639 NODEMASK_FREE(nodes_allowed
);
1643 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1644 struct kobj_attribute
*attr
, char *buf
)
1646 return nr_hugepages_show_common(kobj
, attr
, buf
);
1649 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1650 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1652 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1654 HSTATE_ATTR(nr_hugepages
);
1659 * hstate attribute for optionally mempolicy-based constraint on persistent
1660 * huge page alloc/free.
1662 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1663 struct kobj_attribute
*attr
, char *buf
)
1665 return nr_hugepages_show_common(kobj
, attr
, buf
);
1668 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1669 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1671 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1673 HSTATE_ATTR(nr_hugepages_mempolicy
);
1677 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1678 struct kobj_attribute
*attr
, char *buf
)
1680 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1681 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1684 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1685 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1688 unsigned long input
;
1689 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1691 if (h
->order
>= MAX_ORDER
)
1694 err
= kstrtoul(buf
, 10, &input
);
1698 spin_lock(&hugetlb_lock
);
1699 h
->nr_overcommit_huge_pages
= input
;
1700 spin_unlock(&hugetlb_lock
);
1704 HSTATE_ATTR(nr_overcommit_hugepages
);
1706 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1707 struct kobj_attribute
*attr
, char *buf
)
1710 unsigned long free_huge_pages
;
1713 h
= kobj_to_hstate(kobj
, &nid
);
1714 if (nid
== NUMA_NO_NODE
)
1715 free_huge_pages
= h
->free_huge_pages
;
1717 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1719 return sprintf(buf
, "%lu\n", free_huge_pages
);
1721 HSTATE_ATTR_RO(free_hugepages
);
1723 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1724 struct kobj_attribute
*attr
, char *buf
)
1726 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1727 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1729 HSTATE_ATTR_RO(resv_hugepages
);
1731 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1732 struct kobj_attribute
*attr
, char *buf
)
1735 unsigned long surplus_huge_pages
;
1738 h
= kobj_to_hstate(kobj
, &nid
);
1739 if (nid
== NUMA_NO_NODE
)
1740 surplus_huge_pages
= h
->surplus_huge_pages
;
1742 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1744 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1746 HSTATE_ATTR_RO(surplus_hugepages
);
1748 static struct attribute
*hstate_attrs
[] = {
1749 &nr_hugepages_attr
.attr
,
1750 &nr_overcommit_hugepages_attr
.attr
,
1751 &free_hugepages_attr
.attr
,
1752 &resv_hugepages_attr
.attr
,
1753 &surplus_hugepages_attr
.attr
,
1755 &nr_hugepages_mempolicy_attr
.attr
,
1760 static struct attribute_group hstate_attr_group
= {
1761 .attrs
= hstate_attrs
,
1764 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1765 struct kobject
**hstate_kobjs
,
1766 struct attribute_group
*hstate_attr_group
)
1769 int hi
= hstate_index(h
);
1771 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1772 if (!hstate_kobjs
[hi
])
1775 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1777 kobject_put(hstate_kobjs
[hi
]);
1782 static void __init
hugetlb_sysfs_init(void)
1787 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1788 if (!hugepages_kobj
)
1791 for_each_hstate(h
) {
1792 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1793 hstate_kobjs
, &hstate_attr_group
);
1795 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1802 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1803 * with node devices in node_devices[] using a parallel array. The array
1804 * index of a node device or _hstate == node id.
1805 * This is here to avoid any static dependency of the node device driver, in
1806 * the base kernel, on the hugetlb module.
1808 struct node_hstate
{
1809 struct kobject
*hugepages_kobj
;
1810 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1812 struct node_hstate node_hstates
[MAX_NUMNODES
];
1815 * A subset of global hstate attributes for node devices
1817 static struct attribute
*per_node_hstate_attrs
[] = {
1818 &nr_hugepages_attr
.attr
,
1819 &free_hugepages_attr
.attr
,
1820 &surplus_hugepages_attr
.attr
,
1824 static struct attribute_group per_node_hstate_attr_group
= {
1825 .attrs
= per_node_hstate_attrs
,
1829 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1830 * Returns node id via non-NULL nidp.
1832 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1836 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1837 struct node_hstate
*nhs
= &node_hstates
[nid
];
1839 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1840 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1852 * Unregister hstate attributes from a single node device.
1853 * No-op if no hstate attributes attached.
1855 static void hugetlb_unregister_node(struct node
*node
)
1858 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1860 if (!nhs
->hugepages_kobj
)
1861 return; /* no hstate attributes */
1863 for_each_hstate(h
) {
1864 int idx
= hstate_index(h
);
1865 if (nhs
->hstate_kobjs
[idx
]) {
1866 kobject_put(nhs
->hstate_kobjs
[idx
]);
1867 nhs
->hstate_kobjs
[idx
] = NULL
;
1871 kobject_put(nhs
->hugepages_kobj
);
1872 nhs
->hugepages_kobj
= NULL
;
1876 * hugetlb module exit: unregister hstate attributes from node devices
1879 static void hugetlb_unregister_all_nodes(void)
1884 * disable node device registrations.
1886 register_hugetlbfs_with_node(NULL
, NULL
);
1889 * remove hstate attributes from any nodes that have them.
1891 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1892 hugetlb_unregister_node(node_devices
[nid
]);
1896 * Register hstate attributes for a single node device.
1897 * No-op if attributes already registered.
1899 static void hugetlb_register_node(struct node
*node
)
1902 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1905 if (nhs
->hugepages_kobj
)
1906 return; /* already allocated */
1908 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1910 if (!nhs
->hugepages_kobj
)
1913 for_each_hstate(h
) {
1914 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1916 &per_node_hstate_attr_group
);
1918 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1919 h
->name
, node
->dev
.id
);
1920 hugetlb_unregister_node(node
);
1927 * hugetlb init time: register hstate attributes for all registered node
1928 * devices of nodes that have memory. All on-line nodes should have
1929 * registered their associated device by this time.
1931 static void hugetlb_register_all_nodes(void)
1935 for_each_node_state(nid
, N_MEMORY
) {
1936 struct node
*node
= node_devices
[nid
];
1937 if (node
->dev
.id
== nid
)
1938 hugetlb_register_node(node
);
1942 * Let the node device driver know we're here so it can
1943 * [un]register hstate attributes on node hotplug.
1945 register_hugetlbfs_with_node(hugetlb_register_node
,
1946 hugetlb_unregister_node
);
1948 #else /* !CONFIG_NUMA */
1950 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1958 static void hugetlb_unregister_all_nodes(void) { }
1960 static void hugetlb_register_all_nodes(void) { }
1964 static void __exit
hugetlb_exit(void)
1968 hugetlb_unregister_all_nodes();
1970 for_each_hstate(h
) {
1971 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1974 kobject_put(hugepages_kobj
);
1975 kfree(htlb_fault_mutex_table
);
1977 module_exit(hugetlb_exit
);
1979 static int __init
hugetlb_init(void)
1983 if (!hugepages_supported())
1986 if (!size_to_hstate(default_hstate_size
)) {
1987 default_hstate_size
= HPAGE_SIZE
;
1988 if (!size_to_hstate(default_hstate_size
))
1989 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1991 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1992 if (default_hstate_max_huge_pages
)
1993 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1995 hugetlb_init_hstates();
1996 gather_bootmem_prealloc();
1999 hugetlb_sysfs_init();
2000 hugetlb_register_all_nodes();
2001 hugetlb_cgroup_file_init();
2004 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2006 num_fault_mutexes
= 1;
2008 htlb_fault_mutex_table
=
2009 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2010 BUG_ON(!htlb_fault_mutex_table
);
2012 for (i
= 0; i
< num_fault_mutexes
; i
++)
2013 mutex_init(&htlb_fault_mutex_table
[i
]);
2016 module_init(hugetlb_init
);
2018 /* Should be called on processing a hugepagesz=... option */
2019 void __init
hugetlb_add_hstate(unsigned order
)
2024 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2025 pr_warning("hugepagesz= specified twice, ignoring\n");
2028 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2030 h
= &hstates
[hugetlb_max_hstate
++];
2032 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2033 h
->nr_huge_pages
= 0;
2034 h
->free_huge_pages
= 0;
2035 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2036 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2037 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2038 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2039 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2040 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2041 huge_page_size(h
)/1024);
2046 static int __init
hugetlb_nrpages_setup(char *s
)
2049 static unsigned long *last_mhp
;
2052 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2053 * so this hugepages= parameter goes to the "default hstate".
2055 if (!hugetlb_max_hstate
)
2056 mhp
= &default_hstate_max_huge_pages
;
2058 mhp
= &parsed_hstate
->max_huge_pages
;
2060 if (mhp
== last_mhp
) {
2061 pr_warning("hugepages= specified twice without "
2062 "interleaving hugepagesz=, ignoring\n");
2066 if (sscanf(s
, "%lu", mhp
) <= 0)
2070 * Global state is always initialized later in hugetlb_init.
2071 * But we need to allocate >= MAX_ORDER hstates here early to still
2072 * use the bootmem allocator.
2074 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2075 hugetlb_hstate_alloc_pages(parsed_hstate
);
2081 __setup("hugepages=", hugetlb_nrpages_setup
);
2083 static int __init
hugetlb_default_setup(char *s
)
2085 default_hstate_size
= memparse(s
, &s
);
2088 __setup("default_hugepagesz=", hugetlb_default_setup
);
2090 static unsigned int cpuset_mems_nr(unsigned int *array
)
2093 unsigned int nr
= 0;
2095 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2101 #ifdef CONFIG_SYSCTL
2102 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2103 struct ctl_table
*table
, int write
,
2104 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2106 struct hstate
*h
= &default_hstate
;
2110 if (!hugepages_supported())
2113 tmp
= h
->max_huge_pages
;
2115 if (write
&& h
->order
>= MAX_ORDER
)
2119 table
->maxlen
= sizeof(unsigned long);
2120 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2125 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2126 GFP_KERNEL
| __GFP_NORETRY
);
2127 if (!(obey_mempolicy
&&
2128 init_nodemask_of_mempolicy(nodes_allowed
))) {
2129 NODEMASK_FREE(nodes_allowed
);
2130 nodes_allowed
= &node_states
[N_MEMORY
];
2132 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2134 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2135 NODEMASK_FREE(nodes_allowed
);
2141 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2142 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2145 return hugetlb_sysctl_handler_common(false, table
, write
,
2146 buffer
, length
, ppos
);
2150 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2151 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2153 return hugetlb_sysctl_handler_common(true, table
, write
,
2154 buffer
, length
, ppos
);
2156 #endif /* CONFIG_NUMA */
2158 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2159 void __user
*buffer
,
2160 size_t *length
, loff_t
*ppos
)
2162 struct hstate
*h
= &default_hstate
;
2166 if (!hugepages_supported())
2169 tmp
= h
->nr_overcommit_huge_pages
;
2171 if (write
&& h
->order
>= MAX_ORDER
)
2175 table
->maxlen
= sizeof(unsigned long);
2176 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2181 spin_lock(&hugetlb_lock
);
2182 h
->nr_overcommit_huge_pages
= tmp
;
2183 spin_unlock(&hugetlb_lock
);
2189 #endif /* CONFIG_SYSCTL */
2191 void hugetlb_report_meminfo(struct seq_file
*m
)
2193 struct hstate
*h
= &default_hstate
;
2194 if (!hugepages_supported())
2197 "HugePages_Total: %5lu\n"
2198 "HugePages_Free: %5lu\n"
2199 "HugePages_Rsvd: %5lu\n"
2200 "HugePages_Surp: %5lu\n"
2201 "Hugepagesize: %8lu kB\n",
2205 h
->surplus_huge_pages
,
2206 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2209 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2211 struct hstate
*h
= &default_hstate
;
2212 if (!hugepages_supported())
2215 "Node %d HugePages_Total: %5u\n"
2216 "Node %d HugePages_Free: %5u\n"
2217 "Node %d HugePages_Surp: %5u\n",
2218 nid
, h
->nr_huge_pages_node
[nid
],
2219 nid
, h
->free_huge_pages_node
[nid
],
2220 nid
, h
->surplus_huge_pages_node
[nid
]);
2223 void hugetlb_show_meminfo(void)
2228 if (!hugepages_supported())
2231 for_each_node_state(nid
, N_MEMORY
)
2233 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2235 h
->nr_huge_pages_node
[nid
],
2236 h
->free_huge_pages_node
[nid
],
2237 h
->surplus_huge_pages_node
[nid
],
2238 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2241 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2242 unsigned long hugetlb_total_pages(void)
2245 unsigned long nr_total_pages
= 0;
2248 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2249 return nr_total_pages
;
2252 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2256 spin_lock(&hugetlb_lock
);
2258 * When cpuset is configured, it breaks the strict hugetlb page
2259 * reservation as the accounting is done on a global variable. Such
2260 * reservation is completely rubbish in the presence of cpuset because
2261 * the reservation is not checked against page availability for the
2262 * current cpuset. Application can still potentially OOM'ed by kernel
2263 * with lack of free htlb page in cpuset that the task is in.
2264 * Attempt to enforce strict accounting with cpuset is almost
2265 * impossible (or too ugly) because cpuset is too fluid that
2266 * task or memory node can be dynamically moved between cpusets.
2268 * The change of semantics for shared hugetlb mapping with cpuset is
2269 * undesirable. However, in order to preserve some of the semantics,
2270 * we fall back to check against current free page availability as
2271 * a best attempt and hopefully to minimize the impact of changing
2272 * semantics that cpuset has.
2275 if (gather_surplus_pages(h
, delta
) < 0)
2278 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2279 return_unused_surplus_pages(h
, delta
);
2286 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2289 spin_unlock(&hugetlb_lock
);
2293 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2295 struct resv_map
*resv
= vma_resv_map(vma
);
2298 * This new VMA should share its siblings reservation map if present.
2299 * The VMA will only ever have a valid reservation map pointer where
2300 * it is being copied for another still existing VMA. As that VMA
2301 * has a reference to the reservation map it cannot disappear until
2302 * after this open call completes. It is therefore safe to take a
2303 * new reference here without additional locking.
2305 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2306 kref_get(&resv
->refs
);
2309 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2311 struct hstate
*h
= hstate_vma(vma
);
2312 struct resv_map
*resv
= vma_resv_map(vma
);
2313 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2314 unsigned long reserve
, start
, end
;
2316 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2319 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2320 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2322 reserve
= (end
- start
) - region_count(resv
, start
, end
);
2324 kref_put(&resv
->refs
, resv_map_release
);
2327 hugetlb_acct_memory(h
, -reserve
);
2328 hugepage_subpool_put_pages(spool
, reserve
);
2333 * We cannot handle pagefaults against hugetlb pages at all. They cause
2334 * handle_mm_fault() to try to instantiate regular-sized pages in the
2335 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2338 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2344 const struct vm_operations_struct hugetlb_vm_ops
= {
2345 .fault
= hugetlb_vm_op_fault
,
2346 .open
= hugetlb_vm_op_open
,
2347 .close
= hugetlb_vm_op_close
,
2350 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2356 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2357 vma
->vm_page_prot
)));
2359 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2360 vma
->vm_page_prot
));
2362 entry
= pte_mkyoung(entry
);
2363 entry
= pte_mkhuge(entry
);
2364 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2369 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2370 unsigned long address
, pte_t
*ptep
)
2374 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2375 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2376 update_mmu_cache(vma
, address
, ptep
);
2380 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2381 struct vm_area_struct
*vma
)
2383 pte_t
*src_pte
, *dst_pte
, entry
;
2384 struct page
*ptepage
;
2387 struct hstate
*h
= hstate_vma(vma
);
2388 unsigned long sz
= huge_page_size(h
);
2389 unsigned long mmun_start
; /* For mmu_notifiers */
2390 unsigned long mmun_end
; /* For mmu_notifiers */
2393 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2395 mmun_start
= vma
->vm_start
;
2396 mmun_end
= vma
->vm_end
;
2398 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
2400 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2401 spinlock_t
*src_ptl
, *dst_ptl
;
2402 src_pte
= huge_pte_offset(src
, addr
);
2405 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2411 /* If the pagetables are shared don't copy or take references */
2412 if (dst_pte
== src_pte
)
2415 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
2416 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
2417 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
2418 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2420 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2421 entry
= huge_ptep_get(src_pte
);
2422 ptepage
= pte_page(entry
);
2424 page_dup_rmap(ptepage
);
2425 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2427 spin_unlock(src_ptl
);
2428 spin_unlock(dst_ptl
);
2432 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
2437 static int is_hugetlb_entry_migration(pte_t pte
)
2441 if (huge_pte_none(pte
) || pte_present(pte
))
2443 swp
= pte_to_swp_entry(pte
);
2444 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2450 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2454 if (huge_pte_none(pte
) || pte_present(pte
))
2456 swp
= pte_to_swp_entry(pte
);
2457 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2463 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2464 unsigned long start
, unsigned long end
,
2465 struct page
*ref_page
)
2467 int force_flush
= 0;
2468 struct mm_struct
*mm
= vma
->vm_mm
;
2469 unsigned long address
;
2474 struct hstate
*h
= hstate_vma(vma
);
2475 unsigned long sz
= huge_page_size(h
);
2476 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2477 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2479 WARN_ON(!is_vm_hugetlb_page(vma
));
2480 BUG_ON(start
& ~huge_page_mask(h
));
2481 BUG_ON(end
& ~huge_page_mask(h
));
2483 tlb_start_vma(tlb
, vma
);
2484 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2486 for (address
= start
; address
< end
; address
+= sz
) {
2487 ptep
= huge_pte_offset(mm
, address
);
2491 ptl
= huge_pte_lock(h
, mm
, ptep
);
2492 if (huge_pmd_unshare(mm
, &address
, ptep
))
2495 pte
= huge_ptep_get(ptep
);
2496 if (huge_pte_none(pte
))
2500 * HWPoisoned hugepage is already unmapped and dropped reference
2502 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2503 huge_pte_clear(mm
, address
, ptep
);
2507 page
= pte_page(pte
);
2509 * If a reference page is supplied, it is because a specific
2510 * page is being unmapped, not a range. Ensure the page we
2511 * are about to unmap is the actual page of interest.
2514 if (page
!= ref_page
)
2518 * Mark the VMA as having unmapped its page so that
2519 * future faults in this VMA will fail rather than
2520 * looking like data was lost
2522 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2525 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2526 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2527 if (huge_pte_dirty(pte
))
2528 set_page_dirty(page
);
2530 page_remove_rmap(page
);
2531 force_flush
= !__tlb_remove_page(tlb
, page
);
2536 /* Bail out after unmapping reference page if supplied */
2545 * mmu_gather ran out of room to batch pages, we break out of
2546 * the PTE lock to avoid doing the potential expensive TLB invalidate
2547 * and page-free while holding it.
2552 if (address
< end
&& !ref_page
)
2555 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2556 tlb_end_vma(tlb
, vma
);
2559 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2560 struct vm_area_struct
*vma
, unsigned long start
,
2561 unsigned long end
, struct page
*ref_page
)
2563 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2566 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2567 * test will fail on a vma being torn down, and not grab a page table
2568 * on its way out. We're lucky that the flag has such an appropriate
2569 * name, and can in fact be safely cleared here. We could clear it
2570 * before the __unmap_hugepage_range above, but all that's necessary
2571 * is to clear it before releasing the i_mmap_mutex. This works
2572 * because in the context this is called, the VMA is about to be
2573 * destroyed and the i_mmap_mutex is held.
2575 vma
->vm_flags
&= ~VM_MAYSHARE
;
2578 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2579 unsigned long end
, struct page
*ref_page
)
2581 struct mm_struct
*mm
;
2582 struct mmu_gather tlb
;
2586 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2587 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2588 tlb_finish_mmu(&tlb
, start
, end
);
2592 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2593 * mappping it owns the reserve page for. The intention is to unmap the page
2594 * from other VMAs and let the children be SIGKILLed if they are faulting the
2597 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2598 struct page
*page
, unsigned long address
)
2600 struct hstate
*h
= hstate_vma(vma
);
2601 struct vm_area_struct
*iter_vma
;
2602 struct address_space
*mapping
;
2606 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2607 * from page cache lookup which is in HPAGE_SIZE units.
2609 address
= address
& huge_page_mask(h
);
2610 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2612 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2615 * Take the mapping lock for the duration of the table walk. As
2616 * this mapping should be shared between all the VMAs,
2617 * __unmap_hugepage_range() is called as the lock is already held
2619 mutex_lock(&mapping
->i_mmap_mutex
);
2620 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2621 /* Do not unmap the current VMA */
2622 if (iter_vma
== vma
)
2626 * Unmap the page from other VMAs without their own reserves.
2627 * They get marked to be SIGKILLed if they fault in these
2628 * areas. This is because a future no-page fault on this VMA
2629 * could insert a zeroed page instead of the data existing
2630 * from the time of fork. This would look like data corruption
2632 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2633 unmap_hugepage_range(iter_vma
, address
,
2634 address
+ huge_page_size(h
), page
);
2636 mutex_unlock(&mapping
->i_mmap_mutex
);
2642 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2643 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2644 * cannot race with other handlers or page migration.
2645 * Keep the pte_same checks anyway to make transition from the mutex easier.
2647 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2648 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2649 struct page
*pagecache_page
, spinlock_t
*ptl
)
2651 struct hstate
*h
= hstate_vma(vma
);
2652 struct page
*old_page
, *new_page
;
2653 int outside_reserve
= 0;
2654 unsigned long mmun_start
; /* For mmu_notifiers */
2655 unsigned long mmun_end
; /* For mmu_notifiers */
2657 old_page
= pte_page(pte
);
2660 /* If no-one else is actually using this page, avoid the copy
2661 * and just make the page writable */
2662 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2663 page_move_anon_rmap(old_page
, vma
, address
);
2664 set_huge_ptep_writable(vma
, address
, ptep
);
2669 * If the process that created a MAP_PRIVATE mapping is about to
2670 * perform a COW due to a shared page count, attempt to satisfy
2671 * the allocation without using the existing reserves. The pagecache
2672 * page is used to determine if the reserve at this address was
2673 * consumed or not. If reserves were used, a partial faulted mapping
2674 * at the time of fork() could consume its reserves on COW instead
2675 * of the full address range.
2677 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2678 old_page
!= pagecache_page
)
2679 outside_reserve
= 1;
2681 page_cache_get(old_page
);
2683 /* Drop page table lock as buddy allocator may be called */
2685 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2687 if (IS_ERR(new_page
)) {
2688 long err
= PTR_ERR(new_page
);
2689 page_cache_release(old_page
);
2692 * If a process owning a MAP_PRIVATE mapping fails to COW,
2693 * it is due to references held by a child and an insufficient
2694 * huge page pool. To guarantee the original mappers
2695 * reliability, unmap the page from child processes. The child
2696 * may get SIGKILLed if it later faults.
2698 if (outside_reserve
) {
2699 BUG_ON(huge_pte_none(pte
));
2700 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2701 BUG_ON(huge_pte_none(pte
));
2703 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2705 pte_same(huge_ptep_get(ptep
), pte
)))
2706 goto retry_avoidcopy
;
2708 * race occurs while re-acquiring page table
2709 * lock, and our job is done.
2716 /* Caller expects lock to be held */
2719 return VM_FAULT_OOM
;
2721 return VM_FAULT_SIGBUS
;
2725 * When the original hugepage is shared one, it does not have
2726 * anon_vma prepared.
2728 if (unlikely(anon_vma_prepare(vma
))) {
2729 page_cache_release(new_page
);
2730 page_cache_release(old_page
);
2731 /* Caller expects lock to be held */
2733 return VM_FAULT_OOM
;
2736 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2737 pages_per_huge_page(h
));
2738 __SetPageUptodate(new_page
);
2740 mmun_start
= address
& huge_page_mask(h
);
2741 mmun_end
= mmun_start
+ huge_page_size(h
);
2742 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2744 * Retake the page table lock to check for racing updates
2745 * before the page tables are altered
2748 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2749 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
2750 ClearPagePrivate(new_page
);
2753 huge_ptep_clear_flush(vma
, address
, ptep
);
2754 set_huge_pte_at(mm
, address
, ptep
,
2755 make_huge_pte(vma
, new_page
, 1));
2756 page_remove_rmap(old_page
);
2757 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2758 /* Make the old page be freed below */
2759 new_page
= old_page
;
2762 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2763 page_cache_release(new_page
);
2764 page_cache_release(old_page
);
2766 /* Caller expects lock to be held */
2771 /* Return the pagecache page at a given address within a VMA */
2772 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2773 struct vm_area_struct
*vma
, unsigned long address
)
2775 struct address_space
*mapping
;
2778 mapping
= vma
->vm_file
->f_mapping
;
2779 idx
= vma_hugecache_offset(h
, vma
, address
);
2781 return find_lock_page(mapping
, idx
);
2785 * Return whether there is a pagecache page to back given address within VMA.
2786 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2788 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2789 struct vm_area_struct
*vma
, unsigned long address
)
2791 struct address_space
*mapping
;
2795 mapping
= vma
->vm_file
->f_mapping
;
2796 idx
= vma_hugecache_offset(h
, vma
, address
);
2798 page
= find_get_page(mapping
, idx
);
2801 return page
!= NULL
;
2804 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2805 struct address_space
*mapping
, pgoff_t idx
,
2806 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2808 struct hstate
*h
= hstate_vma(vma
);
2809 int ret
= VM_FAULT_SIGBUS
;
2817 * Currently, we are forced to kill the process in the event the
2818 * original mapper has unmapped pages from the child due to a failed
2819 * COW. Warn that such a situation has occurred as it may not be obvious
2821 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2822 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2828 * Use page lock to guard against racing truncation
2829 * before we get page_table_lock.
2832 page
= find_lock_page(mapping
, idx
);
2834 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2837 page
= alloc_huge_page(vma
, address
, 0);
2839 ret
= PTR_ERR(page
);
2843 ret
= VM_FAULT_SIGBUS
;
2846 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2847 __SetPageUptodate(page
);
2849 if (vma
->vm_flags
& VM_MAYSHARE
) {
2851 struct inode
*inode
= mapping
->host
;
2853 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2860 ClearPagePrivate(page
);
2862 spin_lock(&inode
->i_lock
);
2863 inode
->i_blocks
+= blocks_per_huge_page(h
);
2864 spin_unlock(&inode
->i_lock
);
2867 if (unlikely(anon_vma_prepare(vma
))) {
2869 goto backout_unlocked
;
2875 * If memory error occurs between mmap() and fault, some process
2876 * don't have hwpoisoned swap entry for errored virtual address.
2877 * So we need to block hugepage fault by PG_hwpoison bit check.
2879 if (unlikely(PageHWPoison(page
))) {
2880 ret
= VM_FAULT_HWPOISON
|
2881 VM_FAULT_SET_HINDEX(hstate_index(h
));
2882 goto backout_unlocked
;
2887 * If we are going to COW a private mapping later, we examine the
2888 * pending reservations for this page now. This will ensure that
2889 * any allocations necessary to record that reservation occur outside
2892 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2893 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2895 goto backout_unlocked
;
2898 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
2900 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2905 if (!huge_pte_none(huge_ptep_get(ptep
)))
2909 ClearPagePrivate(page
);
2910 hugepage_add_new_anon_rmap(page
, vma
, address
);
2912 page_dup_rmap(page
);
2913 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2914 && (vma
->vm_flags
& VM_SHARED
)));
2915 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2917 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2918 /* Optimization, do the COW without a second fault */
2919 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
2936 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
2937 struct vm_area_struct
*vma
,
2938 struct address_space
*mapping
,
2939 pgoff_t idx
, unsigned long address
)
2941 unsigned long key
[2];
2944 if (vma
->vm_flags
& VM_SHARED
) {
2945 key
[0] = (unsigned long) mapping
;
2948 key
[0] = (unsigned long) mm
;
2949 key
[1] = address
>> huge_page_shift(h
);
2952 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
2954 return hash
& (num_fault_mutexes
- 1);
2958 * For uniprocesor systems we always use a single mutex, so just
2959 * return 0 and avoid the hashing overhead.
2961 static u32
fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
2962 struct vm_area_struct
*vma
,
2963 struct address_space
*mapping
,
2964 pgoff_t idx
, unsigned long address
)
2970 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2971 unsigned long address
, unsigned int flags
)
2978 struct page
*page
= NULL
;
2979 struct page
*pagecache_page
= NULL
;
2980 struct hstate
*h
= hstate_vma(vma
);
2981 struct address_space
*mapping
;
2983 address
&= huge_page_mask(h
);
2985 ptep
= huge_pte_offset(mm
, address
);
2987 entry
= huge_ptep_get(ptep
);
2988 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2989 migration_entry_wait_huge(vma
, mm
, ptep
);
2991 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2992 return VM_FAULT_HWPOISON_LARGE
|
2993 VM_FAULT_SET_HINDEX(hstate_index(h
));
2996 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2998 return VM_FAULT_OOM
;
3000 mapping
= vma
->vm_file
->f_mapping
;
3001 idx
= vma_hugecache_offset(h
, vma
, address
);
3004 * Serialize hugepage allocation and instantiation, so that we don't
3005 * get spurious allocation failures if two CPUs race to instantiate
3006 * the same page in the page cache.
3008 hash
= fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3009 mutex_lock(&htlb_fault_mutex_table
[hash
]);
3011 entry
= huge_ptep_get(ptep
);
3012 if (huge_pte_none(entry
)) {
3013 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3020 * If we are going to COW the mapping later, we examine the pending
3021 * reservations for this page now. This will ensure that any
3022 * allocations necessary to record that reservation occur outside the
3023 * spinlock. For private mappings, we also lookup the pagecache
3024 * page now as it is used to determine if a reservation has been
3027 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3028 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3033 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3034 pagecache_page
= hugetlbfs_pagecache_page(h
,
3039 * hugetlb_cow() requires page locks of pte_page(entry) and
3040 * pagecache_page, so here we need take the former one
3041 * when page != pagecache_page or !pagecache_page.
3042 * Note that locking order is always pagecache_page -> page,
3043 * so no worry about deadlock.
3045 page
= pte_page(entry
);
3047 if (page
!= pagecache_page
)
3050 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3052 /* Check for a racing update before calling hugetlb_cow */
3053 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3057 if (flags
& FAULT_FLAG_WRITE
) {
3058 if (!huge_pte_write(entry
)) {
3059 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3060 pagecache_page
, ptl
);
3063 entry
= huge_pte_mkdirty(entry
);
3065 entry
= pte_mkyoung(entry
);
3066 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3067 flags
& FAULT_FLAG_WRITE
))
3068 update_mmu_cache(vma
, address
, ptep
);
3073 if (pagecache_page
) {
3074 unlock_page(pagecache_page
);
3075 put_page(pagecache_page
);
3077 if (page
!= pagecache_page
)
3082 mutex_unlock(&htlb_fault_mutex_table
[hash
]);
3086 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3087 struct page
**pages
, struct vm_area_struct
**vmas
,
3088 unsigned long *position
, unsigned long *nr_pages
,
3089 long i
, unsigned int flags
)
3091 unsigned long pfn_offset
;
3092 unsigned long vaddr
= *position
;
3093 unsigned long remainder
= *nr_pages
;
3094 struct hstate
*h
= hstate_vma(vma
);
3096 while (vaddr
< vma
->vm_end
&& remainder
) {
3098 spinlock_t
*ptl
= NULL
;
3103 * Some archs (sparc64, sh*) have multiple pte_ts to
3104 * each hugepage. We have to make sure we get the
3105 * first, for the page indexing below to work.
3107 * Note that page table lock is not held when pte is null.
3109 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3111 ptl
= huge_pte_lock(h
, mm
, pte
);
3112 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3115 * When coredumping, it suits get_dump_page if we just return
3116 * an error where there's an empty slot with no huge pagecache
3117 * to back it. This way, we avoid allocating a hugepage, and
3118 * the sparse dumpfile avoids allocating disk blocks, but its
3119 * huge holes still show up with zeroes where they need to be.
3121 if (absent
&& (flags
& FOLL_DUMP
) &&
3122 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3130 * We need call hugetlb_fault for both hugepages under migration
3131 * (in which case hugetlb_fault waits for the migration,) and
3132 * hwpoisoned hugepages (in which case we need to prevent the
3133 * caller from accessing to them.) In order to do this, we use
3134 * here is_swap_pte instead of is_hugetlb_entry_migration and
3135 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3136 * both cases, and because we can't follow correct pages
3137 * directly from any kind of swap entries.
3139 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3140 ((flags
& FOLL_WRITE
) &&
3141 !huge_pte_write(huge_ptep_get(pte
)))) {
3146 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3147 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3148 if (!(ret
& VM_FAULT_ERROR
))
3155 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3156 page
= pte_page(huge_ptep_get(pte
));
3159 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3160 get_page_foll(pages
[i
]);
3170 if (vaddr
< vma
->vm_end
&& remainder
&&
3171 pfn_offset
< pages_per_huge_page(h
)) {
3173 * We use pfn_offset to avoid touching the pageframes
3174 * of this compound page.
3180 *nr_pages
= remainder
;
3183 return i
? i
: -EFAULT
;
3186 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3187 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3189 struct mm_struct
*mm
= vma
->vm_mm
;
3190 unsigned long start
= address
;
3193 struct hstate
*h
= hstate_vma(vma
);
3194 unsigned long pages
= 0;
3196 BUG_ON(address
>= end
);
3197 flush_cache_range(vma
, address
, end
);
3199 mmu_notifier_invalidate_range_start(mm
, start
, end
);
3200 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3201 for (; address
< end
; address
+= huge_page_size(h
)) {
3203 ptep
= huge_pte_offset(mm
, address
);
3206 ptl
= huge_pte_lock(h
, mm
, ptep
);
3207 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3212 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3213 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3214 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3215 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3216 set_huge_pte_at(mm
, address
, ptep
, pte
);
3222 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3223 * may have cleared our pud entry and done put_page on the page table:
3224 * once we release i_mmap_mutex, another task can do the final put_page
3225 * and that page table be reused and filled with junk.
3227 flush_tlb_range(vma
, start
, end
);
3228 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3229 mmu_notifier_invalidate_range_end(mm
, start
, end
);
3231 return pages
<< h
->order
;
3234 int hugetlb_reserve_pages(struct inode
*inode
,
3236 struct vm_area_struct
*vma
,
3237 vm_flags_t vm_flags
)
3240 struct hstate
*h
= hstate_inode(inode
);
3241 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3242 struct resv_map
*resv_map
;
3245 * Only apply hugepage reservation if asked. At fault time, an
3246 * attempt will be made for VM_NORESERVE to allocate a page
3247 * without using reserves
3249 if (vm_flags
& VM_NORESERVE
)
3253 * Shared mappings base their reservation on the number of pages that
3254 * are already allocated on behalf of the file. Private mappings need
3255 * to reserve the full area even if read-only as mprotect() may be
3256 * called to make the mapping read-write. Assume !vma is a shm mapping
3258 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
3259 resv_map
= inode_resv_map(inode
);
3261 chg
= region_chg(resv_map
, from
, to
);
3264 resv_map
= resv_map_alloc();
3270 set_vma_resv_map(vma
, resv_map
);
3271 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3279 /* There must be enough pages in the subpool for the mapping */
3280 if (hugepage_subpool_get_pages(spool
, chg
)) {
3286 * Check enough hugepages are available for the reservation.
3287 * Hand the pages back to the subpool if there are not
3289 ret
= hugetlb_acct_memory(h
, chg
);
3291 hugepage_subpool_put_pages(spool
, chg
);
3296 * Account for the reservations made. Shared mappings record regions
3297 * that have reservations as they are shared by multiple VMAs.
3298 * When the last VMA disappears, the region map says how much
3299 * the reservation was and the page cache tells how much of
3300 * the reservation was consumed. Private mappings are per-VMA and
3301 * only the consumed reservations are tracked. When the VMA
3302 * disappears, the original reservation is the VMA size and the
3303 * consumed reservations are stored in the map. Hence, nothing
3304 * else has to be done for private mappings here
3306 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3307 region_add(resv_map
, from
, to
);
3310 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3311 kref_put(&resv_map
->refs
, resv_map_release
);
3315 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3317 struct hstate
*h
= hstate_inode(inode
);
3318 struct resv_map
*resv_map
= inode_resv_map(inode
);
3320 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3323 chg
= region_truncate(resv_map
, offset
);
3324 spin_lock(&inode
->i_lock
);
3325 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3326 spin_unlock(&inode
->i_lock
);
3328 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3329 hugetlb_acct_memory(h
, -(chg
- freed
));
3332 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3333 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3334 struct vm_area_struct
*vma
,
3335 unsigned long addr
, pgoff_t idx
)
3337 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3339 unsigned long sbase
= saddr
& PUD_MASK
;
3340 unsigned long s_end
= sbase
+ PUD_SIZE
;
3342 /* Allow segments to share if only one is marked locked */
3343 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3344 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3347 * match the virtual addresses, permission and the alignment of the
3350 if (pmd_index(addr
) != pmd_index(saddr
) ||
3351 vm_flags
!= svm_flags
||
3352 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3358 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3360 unsigned long base
= addr
& PUD_MASK
;
3361 unsigned long end
= base
+ PUD_SIZE
;
3364 * check on proper vm_flags and page table alignment
3366 if (vma
->vm_flags
& VM_MAYSHARE
&&
3367 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3373 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3374 * and returns the corresponding pte. While this is not necessary for the
3375 * !shared pmd case because we can allocate the pmd later as well, it makes the
3376 * code much cleaner. pmd allocation is essential for the shared case because
3377 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3378 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3379 * bad pmd for sharing.
3381 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3383 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3384 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3385 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3387 struct vm_area_struct
*svma
;
3388 unsigned long saddr
;
3393 if (!vma_shareable(vma
, addr
))
3394 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3396 mutex_lock(&mapping
->i_mmap_mutex
);
3397 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3401 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3403 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3405 get_page(virt_to_page(spte
));
3414 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
3417 pud_populate(mm
, pud
,
3418 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3420 put_page(virt_to_page(spte
));
3423 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3424 mutex_unlock(&mapping
->i_mmap_mutex
);
3429 * unmap huge page backed by shared pte.
3431 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3432 * indicated by page_count > 1, unmap is achieved by clearing pud and
3433 * decrementing the ref count. If count == 1, the pte page is not shared.
3435 * called with page table lock held.
3437 * returns: 1 successfully unmapped a shared pte page
3438 * 0 the underlying pte page is not shared, or it is the last user
3440 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3442 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3443 pud_t
*pud
= pud_offset(pgd
, *addr
);
3445 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3446 if (page_count(virt_to_page(ptep
)) == 1)
3450 put_page(virt_to_page(ptep
));
3451 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3454 #define want_pmd_share() (1)
3455 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3456 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3460 #define want_pmd_share() (0)
3461 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3463 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3464 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3465 unsigned long addr
, unsigned long sz
)
3471 pgd
= pgd_offset(mm
, addr
);
3472 pud
= pud_alloc(mm
, pgd
, addr
);
3474 if (sz
== PUD_SIZE
) {
3477 BUG_ON(sz
!= PMD_SIZE
);
3478 if (want_pmd_share() && pud_none(*pud
))
3479 pte
= huge_pmd_share(mm
, addr
, pud
);
3481 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3484 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3489 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3495 pgd
= pgd_offset(mm
, addr
);
3496 if (pgd_present(*pgd
)) {
3497 pud
= pud_offset(pgd
, addr
);
3498 if (pud_present(*pud
)) {
3500 return (pte_t
*)pud
;
3501 pmd
= pmd_offset(pud
, addr
);
3504 return (pte_t
*) pmd
;
3508 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3509 pmd_t
*pmd
, int write
)
3513 page
= pte_page(*(pte_t
*)pmd
);
3515 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3520 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3521 pud_t
*pud
, int write
)
3525 page
= pte_page(*(pte_t
*)pud
);
3527 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3531 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3533 /* Can be overriden by architectures */
3534 struct page
* __weak
3535 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3536 pud_t
*pud
, int write
)
3542 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3544 #ifdef CONFIG_MEMORY_FAILURE
3546 /* Should be called in hugetlb_lock */
3547 static int is_hugepage_on_freelist(struct page
*hpage
)
3551 struct hstate
*h
= page_hstate(hpage
);
3552 int nid
= page_to_nid(hpage
);
3554 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3561 * This function is called from memory failure code.
3562 * Assume the caller holds page lock of the head page.
3564 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3566 struct hstate
*h
= page_hstate(hpage
);
3567 int nid
= page_to_nid(hpage
);
3570 spin_lock(&hugetlb_lock
);
3571 if (is_hugepage_on_freelist(hpage
)) {
3573 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3574 * but dangling hpage->lru can trigger list-debug warnings
3575 * (this happens when we call unpoison_memory() on it),
3576 * so let it point to itself with list_del_init().
3578 list_del_init(&hpage
->lru
);
3579 set_page_refcounted(hpage
);
3580 h
->free_huge_pages
--;
3581 h
->free_huge_pages_node
[nid
]--;
3584 spin_unlock(&hugetlb_lock
);
3589 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3591 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3592 if (!get_page_unless_zero(page
))
3594 spin_lock(&hugetlb_lock
);
3595 list_move_tail(&page
->lru
, list
);
3596 spin_unlock(&hugetlb_lock
);
3600 void putback_active_hugepage(struct page
*page
)
3602 VM_BUG_ON_PAGE(!PageHead(page
), page
);
3603 spin_lock(&hugetlb_lock
);
3604 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3605 spin_unlock(&hugetlb_lock
);
3609 bool is_hugepage_active(struct page
*page
)
3611 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
3613 * This function can be called for a tail page because the caller,
3614 * scan_movable_pages, scans through a given pfn-range which typically
3615 * covers one memory block. In systems using gigantic hugepage (1GB
3616 * for x86_64,) a hugepage is larger than a memory block, and we don't
3617 * support migrating such large hugepages for now, so return false
3618 * when called for tail pages.
3623 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3624 * so we should return false for them.
3626 if (unlikely(PageHWPoison(page
)))
3628 return page_count(page
) > 0;