2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
26 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/hugetlb_cgroup.h>
32 #include <linux/node.h>
35 const unsigned long hugetlb_zero
= 0, hugetlb_infinity
= ~0UL;
36 static gfp_t htlb_alloc_mask
= GFP_HIGHUSER
;
37 unsigned long hugepages_treat_as_movable
;
39 int hugetlb_max_hstate __read_mostly
;
40 unsigned int default_hstate_idx
;
41 struct hstate hstates
[HUGE_MAX_HSTATE
];
43 __initdata
LIST_HEAD(huge_boot_pages
);
45 /* for command line parsing */
46 static struct hstate
* __initdata parsed_hstate
;
47 static unsigned long __initdata default_hstate_max_huge_pages
;
48 static unsigned long __initdata default_hstate_size
;
51 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
52 * free_huge_pages, and surplus_huge_pages.
54 DEFINE_SPINLOCK(hugetlb_lock
);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
58 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
60 spin_unlock(&spool
->lock
);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool
*hugepage_new_subpool(long nr_blocks
)
70 struct hugepage_subpool
*spool
;
72 spool
= kmalloc(sizeof(*spool
), GFP_KERNEL
);
76 spin_lock_init(&spool
->lock
);
78 spool
->max_hpages
= nr_blocks
;
79 spool
->used_hpages
= 0;
84 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
86 spin_lock(&spool
->lock
);
87 BUG_ON(!spool
->count
);
89 unlock_or_release_subpool(spool
);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
100 spin_lock(&spool
->lock
);
101 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
) {
102 spool
->used_hpages
+= delta
;
106 spin_unlock(&spool
->lock
);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
117 spin_lock(&spool
->lock
);
118 spool
->used_hpages
-= delta
;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool
);
124 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
126 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
129 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
131 return subpool_inode(file_inode(vma
->vm_file
));
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation_mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link
;
154 static long region_add(struct list_head
*head
, long f
, long t
)
156 struct file_region
*rg
, *nrg
, *trg
;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg
, head
, link
)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
170 if (&rg
->link
== head
)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head
*head
, long f
, long t
)
192 struct file_region
*rg
, *nrg
;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg
, head
, link
)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg
->link
== head
|| t
< rg
->from
) {
204 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
209 INIT_LIST_HEAD(&nrg
->link
);
210 list_add(&nrg
->link
, rg
->link
.prev
);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
222 if (&rg
->link
== head
)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg
-= rg
->to
- rg
->from
;
239 static long region_truncate(struct list_head
*head
, long end
)
241 struct file_region
*rg
, *trg
;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg
, head
, link
)
248 if (&rg
->link
== head
)
251 /* If we are in the middle of a region then adjust it. */
252 if (end
> rg
->from
) {
255 rg
= list_entry(rg
->link
.next
, typeof(*rg
), link
);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
260 if (&rg
->link
== head
)
262 chg
+= rg
->to
- rg
->from
;
269 static long region_count(struct list_head
*head
, long f
, long t
)
271 struct file_region
*rg
;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg
, head
, link
) {
284 seg_from
= max(rg
->from
, f
);
285 seg_to
= min(rg
->to
, t
);
287 chg
+= seg_to
- seg_from
;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
298 struct vm_area_struct
*vma
, unsigned long address
)
300 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
301 (vma
->vm_pgoff
>> huge_page_order(h
));
304 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
305 unsigned long address
)
307 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
316 struct hstate
*hstate
;
318 if (!is_vm_hugetlb_page(vma
))
321 hstate
= hstate_vma(vma
);
323 return 1UL << huge_page_shift(hstate
);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
336 return vma_kernel_pagesize(vma
);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
370 return (unsigned long)vma
->vm_private_data
;
373 static void set_vma_private_data(struct vm_area_struct
*vma
,
376 vma
->vm_private_data
= (void *)value
;
381 struct list_head regions
;
384 static struct resv_map
*resv_map_alloc(void)
386 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
390 kref_init(&resv_map
->refs
);
391 INIT_LIST_HEAD(&resv_map
->regions
);
396 static void resv_map_release(struct kref
*ref
)
398 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map
->regions
, 0);
405 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
408 if (!(vma
->vm_flags
& VM_MAYSHARE
))
409 return (struct resv_map
*)(get_vma_private_data(vma
) &
414 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
417 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
419 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
420 HPAGE_RESV_MASK
) | (unsigned long)map
);
423 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
426 VM_BUG_ON(vma
->vm_flags
& VM_MAYSHARE
);
428 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
431 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
435 return (get_vma_private_data(vma
) & flag
) != 0;
438 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
439 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
441 VM_BUG_ON(!is_vm_hugetlb_page(vma
));
442 if (!(vma
->vm_flags
& VM_MAYSHARE
))
443 vma
->vm_private_data
= (void *)0;
446 /* Returns true if the VMA has associated reserve pages */
447 static int vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
449 if (vma
->vm_flags
& VM_NORESERVE
) {
451 * This address is already reserved by other process(chg == 0),
452 * so, we should decrement reserved count. Without decrementing,
453 * reserve count remains after releasing inode, because this
454 * allocated page will go into page cache and is regarded as
455 * coming from reserved pool in releasing step. Currently, we
456 * don't have any other solution to deal with this situation
457 * properly, so add work-around here.
459 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
465 /* Shared mappings always use reserves */
466 if (vma
->vm_flags
& VM_MAYSHARE
)
470 * Only the process that called mmap() has reserves for
473 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
479 static void copy_gigantic_page(struct page
*dst
, struct page
*src
)
482 struct hstate
*h
= page_hstate(src
);
483 struct page
*dst_base
= dst
;
484 struct page
*src_base
= src
;
486 for (i
= 0; i
< pages_per_huge_page(h
); ) {
488 copy_highpage(dst
, src
);
491 dst
= mem_map_next(dst
, dst_base
, i
);
492 src
= mem_map_next(src
, src_base
, i
);
496 void copy_huge_page(struct page
*dst
, struct page
*src
)
499 struct hstate
*h
= page_hstate(src
);
501 if (unlikely(pages_per_huge_page(h
) > MAX_ORDER_NR_PAGES
)) {
502 copy_gigantic_page(dst
, src
);
507 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
509 copy_highpage(dst
+ i
, src
+ i
);
513 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
515 int nid
= page_to_nid(page
);
516 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
517 h
->free_huge_pages
++;
518 h
->free_huge_pages_node
[nid
]++;
521 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
525 if (list_empty(&h
->hugepage_freelists
[nid
]))
527 page
= list_entry(h
->hugepage_freelists
[nid
].next
, struct page
, lru
);
528 list_move(&page
->lru
, &h
->hugepage_activelist
);
529 set_page_refcounted(page
);
530 h
->free_huge_pages
--;
531 h
->free_huge_pages_node
[nid
]--;
535 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
536 struct vm_area_struct
*vma
,
537 unsigned long address
, int avoid_reserve
,
540 struct page
*page
= NULL
;
541 struct mempolicy
*mpol
;
542 nodemask_t
*nodemask
;
543 struct zonelist
*zonelist
;
546 unsigned int cpuset_mems_cookie
;
549 * A child process with MAP_PRIVATE mappings created by their parent
550 * have no page reserves. This check ensures that reservations are
551 * not "stolen". The child may still get SIGKILLed
553 if (!vma_has_reserves(vma
, chg
) &&
554 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
557 /* If reserves cannot be used, ensure enough pages are in the pool */
558 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
562 cpuset_mems_cookie
= get_mems_allowed();
563 zonelist
= huge_zonelist(vma
, address
,
564 htlb_alloc_mask
, &mpol
, &nodemask
);
566 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
567 MAX_NR_ZONES
- 1, nodemask
) {
568 if (cpuset_zone_allowed_softwall(zone
, htlb_alloc_mask
)) {
569 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
573 if (!vma_has_reserves(vma
, chg
))
576 SetPagePrivate(page
);
577 h
->resv_huge_pages
--;
584 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
592 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
596 VM_BUG_ON(h
->order
>= MAX_ORDER
);
599 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
600 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
601 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
602 1 << PG_referenced
| 1 << PG_dirty
|
603 1 << PG_active
| 1 << PG_reserved
|
604 1 << PG_private
| 1 << PG_writeback
);
606 VM_BUG_ON(hugetlb_cgroup_from_page(page
));
607 set_compound_page_dtor(page
, NULL
);
608 set_page_refcounted(page
);
609 arch_release_hugepage(page
);
610 __free_pages(page
, huge_page_order(h
));
613 struct hstate
*size_to_hstate(unsigned long size
)
618 if (huge_page_size(h
) == size
)
624 static void free_huge_page(struct page
*page
)
627 * Can't pass hstate in here because it is called from the
628 * compound page destructor.
630 struct hstate
*h
= page_hstate(page
);
631 int nid
= page_to_nid(page
);
632 struct hugepage_subpool
*spool
=
633 (struct hugepage_subpool
*)page_private(page
);
634 bool restore_reserve
;
636 set_page_private(page
, 0);
637 page
->mapping
= NULL
;
638 BUG_ON(page_count(page
));
639 BUG_ON(page_mapcount(page
));
640 restore_reserve
= PagePrivate(page
);
642 spin_lock(&hugetlb_lock
);
643 hugetlb_cgroup_uncharge_page(hstate_index(h
),
644 pages_per_huge_page(h
), page
);
646 h
->resv_huge_pages
++;
648 if (h
->surplus_huge_pages_node
[nid
] && huge_page_order(h
) < MAX_ORDER
) {
649 /* remove the page from active list */
650 list_del(&page
->lru
);
651 update_and_free_page(h
, page
);
652 h
->surplus_huge_pages
--;
653 h
->surplus_huge_pages_node
[nid
]--;
655 arch_clear_hugepage_flags(page
);
656 enqueue_huge_page(h
, page
);
658 spin_unlock(&hugetlb_lock
);
659 hugepage_subpool_put_pages(spool
, 1);
662 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
664 INIT_LIST_HEAD(&page
->lru
);
665 set_compound_page_dtor(page
, free_huge_page
);
666 spin_lock(&hugetlb_lock
);
667 set_hugetlb_cgroup(page
, NULL
);
669 h
->nr_huge_pages_node
[nid
]++;
670 spin_unlock(&hugetlb_lock
);
671 put_page(page
); /* free it into the hugepage allocator */
674 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
677 int nr_pages
= 1 << order
;
678 struct page
*p
= page
+ 1;
680 /* we rely on prep_new_huge_page to set the destructor */
681 set_compound_order(page
, order
);
683 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
685 set_page_count(p
, 0);
686 p
->first_page
= page
;
691 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
692 * transparent huge pages. See the PageTransHuge() documentation for more
695 int PageHuge(struct page
*page
)
697 compound_page_dtor
*dtor
;
699 if (!PageCompound(page
))
702 page
= compound_head(page
);
703 dtor
= get_compound_page_dtor(page
);
705 return dtor
== free_huge_page
;
707 EXPORT_SYMBOL_GPL(PageHuge
);
709 pgoff_t
__basepage_index(struct page
*page
)
711 struct page
*page_head
= compound_head(page
);
712 pgoff_t index
= page_index(page_head
);
713 unsigned long compound_idx
;
715 if (!PageHuge(page_head
))
716 return page_index(page
);
718 if (compound_order(page_head
) >= MAX_ORDER
)
719 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
721 compound_idx
= page
- page_head
;
723 return (index
<< compound_order(page_head
)) + compound_idx
;
726 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
730 if (h
->order
>= MAX_ORDER
)
733 page
= alloc_pages_exact_node(nid
,
734 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
735 __GFP_REPEAT
|__GFP_NOWARN
,
738 if (arch_prepare_hugepage(page
)) {
739 __free_pages(page
, huge_page_order(h
));
742 prep_new_huge_page(h
, page
, nid
);
749 * common helper functions for hstate_next_node_to_{alloc|free}.
750 * We may have allocated or freed a huge page based on a different
751 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
752 * be outside of *nodes_allowed. Ensure that we use an allowed
753 * node for alloc or free.
755 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
757 nid
= next_node(nid
, *nodes_allowed
);
758 if (nid
== MAX_NUMNODES
)
759 nid
= first_node(*nodes_allowed
);
760 VM_BUG_ON(nid
>= MAX_NUMNODES
);
765 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
767 if (!node_isset(nid
, *nodes_allowed
))
768 nid
= next_node_allowed(nid
, nodes_allowed
);
773 * returns the previously saved node ["this node"] from which to
774 * allocate a persistent huge page for the pool and advance the
775 * next node from which to allocate, handling wrap at end of node
778 static int hstate_next_node_to_alloc(struct hstate
*h
,
779 nodemask_t
*nodes_allowed
)
783 VM_BUG_ON(!nodes_allowed
);
785 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
786 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
792 * helper for free_pool_huge_page() - return the previously saved
793 * node ["this node"] from which to free a huge page. Advance the
794 * next node id whether or not we find a free huge page to free so
795 * that the next attempt to free addresses the next node.
797 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
801 VM_BUG_ON(!nodes_allowed
);
803 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
804 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
809 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
810 for (nr_nodes = nodes_weight(*mask); \
812 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
815 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
816 for (nr_nodes = nodes_weight(*mask); \
818 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
821 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
827 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
828 page
= alloc_fresh_huge_page_node(h
, node
);
836 count_vm_event(HTLB_BUDDY_PGALLOC
);
838 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
844 * Free huge page from pool from next node to free.
845 * Attempt to keep persistent huge pages more or less
846 * balanced over allowed nodes.
847 * Called with hugetlb_lock locked.
849 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
855 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
857 * If we're returning unused surplus pages, only examine
858 * nodes with surplus pages.
860 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
861 !list_empty(&h
->hugepage_freelists
[node
])) {
863 list_entry(h
->hugepage_freelists
[node
].next
,
865 list_del(&page
->lru
);
866 h
->free_huge_pages
--;
867 h
->free_huge_pages_node
[node
]--;
869 h
->surplus_huge_pages
--;
870 h
->surplus_huge_pages_node
[node
]--;
872 update_and_free_page(h
, page
);
881 static struct page
*alloc_buddy_huge_page(struct hstate
*h
, int nid
)
886 if (h
->order
>= MAX_ORDER
)
890 * Assume we will successfully allocate the surplus page to
891 * prevent racing processes from causing the surplus to exceed
894 * This however introduces a different race, where a process B
895 * tries to grow the static hugepage pool while alloc_pages() is
896 * called by process A. B will only examine the per-node
897 * counters in determining if surplus huge pages can be
898 * converted to normal huge pages in adjust_pool_surplus(). A
899 * won't be able to increment the per-node counter, until the
900 * lock is dropped by B, but B doesn't drop hugetlb_lock until
901 * no more huge pages can be converted from surplus to normal
902 * state (and doesn't try to convert again). Thus, we have a
903 * case where a surplus huge page exists, the pool is grown, and
904 * the surplus huge page still exists after, even though it
905 * should just have been converted to a normal huge page. This
906 * does not leak memory, though, as the hugepage will be freed
907 * once it is out of use. It also does not allow the counters to
908 * go out of whack in adjust_pool_surplus() as we don't modify
909 * the node values until we've gotten the hugepage and only the
910 * per-node value is checked there.
912 spin_lock(&hugetlb_lock
);
913 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
914 spin_unlock(&hugetlb_lock
);
918 h
->surplus_huge_pages
++;
920 spin_unlock(&hugetlb_lock
);
922 if (nid
== NUMA_NO_NODE
)
923 page
= alloc_pages(htlb_alloc_mask
|__GFP_COMP
|
924 __GFP_REPEAT
|__GFP_NOWARN
,
927 page
= alloc_pages_exact_node(nid
,
928 htlb_alloc_mask
|__GFP_COMP
|__GFP_THISNODE
|
929 __GFP_REPEAT
|__GFP_NOWARN
, huge_page_order(h
));
931 if (page
&& arch_prepare_hugepage(page
)) {
932 __free_pages(page
, huge_page_order(h
));
936 spin_lock(&hugetlb_lock
);
938 INIT_LIST_HEAD(&page
->lru
);
939 r_nid
= page_to_nid(page
);
940 set_compound_page_dtor(page
, free_huge_page
);
941 set_hugetlb_cgroup(page
, NULL
);
943 * We incremented the global counters already
945 h
->nr_huge_pages_node
[r_nid
]++;
946 h
->surplus_huge_pages_node
[r_nid
]++;
947 __count_vm_event(HTLB_BUDDY_PGALLOC
);
950 h
->surplus_huge_pages
--;
951 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
953 spin_unlock(&hugetlb_lock
);
959 * This allocation function is useful in the context where vma is irrelevant.
960 * E.g. soft-offlining uses this function because it only cares physical
961 * address of error page.
963 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
965 struct page
*page
= NULL
;
967 spin_lock(&hugetlb_lock
);
968 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
969 page
= dequeue_huge_page_node(h
, nid
);
970 spin_unlock(&hugetlb_lock
);
973 page
= alloc_buddy_huge_page(h
, nid
);
979 * Increase the hugetlb pool such that it can accommodate a reservation
982 static int gather_surplus_pages(struct hstate
*h
, int delta
)
984 struct list_head surplus_list
;
985 struct page
*page
, *tmp
;
987 int needed
, allocated
;
988 bool alloc_ok
= true;
990 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
992 h
->resv_huge_pages
+= delta
;
997 INIT_LIST_HEAD(&surplus_list
);
1001 spin_unlock(&hugetlb_lock
);
1002 for (i
= 0; i
< needed
; i
++) {
1003 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1008 list_add(&page
->lru
, &surplus_list
);
1013 * After retaking hugetlb_lock, we need to recalculate 'needed'
1014 * because either resv_huge_pages or free_huge_pages may have changed.
1016 spin_lock(&hugetlb_lock
);
1017 needed
= (h
->resv_huge_pages
+ delta
) -
1018 (h
->free_huge_pages
+ allocated
);
1023 * We were not able to allocate enough pages to
1024 * satisfy the entire reservation so we free what
1025 * we've allocated so far.
1030 * The surplus_list now contains _at_least_ the number of extra pages
1031 * needed to accommodate the reservation. Add the appropriate number
1032 * of pages to the hugetlb pool and free the extras back to the buddy
1033 * allocator. Commit the entire reservation here to prevent another
1034 * process from stealing the pages as they are added to the pool but
1035 * before they are reserved.
1037 needed
+= allocated
;
1038 h
->resv_huge_pages
+= delta
;
1041 /* Free the needed pages to the hugetlb pool */
1042 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1046 * This page is now managed by the hugetlb allocator and has
1047 * no users -- drop the buddy allocator's reference.
1049 put_page_testzero(page
);
1050 VM_BUG_ON(page_count(page
));
1051 enqueue_huge_page(h
, page
);
1054 spin_unlock(&hugetlb_lock
);
1056 /* Free unnecessary surplus pages to the buddy allocator */
1057 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1059 spin_lock(&hugetlb_lock
);
1065 * When releasing a hugetlb pool reservation, any surplus pages that were
1066 * allocated to satisfy the reservation must be explicitly freed if they were
1068 * Called with hugetlb_lock held.
1070 static void return_unused_surplus_pages(struct hstate
*h
,
1071 unsigned long unused_resv_pages
)
1073 unsigned long nr_pages
;
1075 /* Uncommit the reservation */
1076 h
->resv_huge_pages
-= unused_resv_pages
;
1078 /* Cannot return gigantic pages currently */
1079 if (h
->order
>= MAX_ORDER
)
1082 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1085 * We want to release as many surplus pages as possible, spread
1086 * evenly across all nodes with memory. Iterate across these nodes
1087 * until we can no longer free unreserved surplus pages. This occurs
1088 * when the nodes with surplus pages have no free pages.
1089 * free_pool_huge_page() will balance the the freed pages across the
1090 * on-line nodes with memory and will handle the hstate accounting.
1092 while (nr_pages
--) {
1093 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1099 * Determine if the huge page at addr within the vma has an associated
1100 * reservation. Where it does not we will need to logically increase
1101 * reservation and actually increase subpool usage before an allocation
1102 * can occur. Where any new reservation would be required the
1103 * reservation change is prepared, but not committed. Once the page
1104 * has been allocated from the subpool and instantiated the change should
1105 * be committed via vma_commit_reservation. No action is required on
1108 static long vma_needs_reservation(struct hstate
*h
,
1109 struct vm_area_struct
*vma
, unsigned long addr
)
1111 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1112 struct inode
*inode
= mapping
->host
;
1114 if (vma
->vm_flags
& VM_MAYSHARE
) {
1115 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1116 return region_chg(&inode
->i_mapping
->private_list
,
1119 } else if (!is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1124 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1125 struct resv_map
*resv
= vma_resv_map(vma
);
1127 err
= region_chg(&resv
->regions
, idx
, idx
+ 1);
1133 static void vma_commit_reservation(struct hstate
*h
,
1134 struct vm_area_struct
*vma
, unsigned long addr
)
1136 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
1137 struct inode
*inode
= mapping
->host
;
1139 if (vma
->vm_flags
& VM_MAYSHARE
) {
1140 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1141 region_add(&inode
->i_mapping
->private_list
, idx
, idx
+ 1);
1143 } else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
1144 pgoff_t idx
= vma_hugecache_offset(h
, vma
, addr
);
1145 struct resv_map
*resv
= vma_resv_map(vma
);
1147 /* Mark this page used in the map. */
1148 region_add(&resv
->regions
, idx
, idx
+ 1);
1152 static struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1153 unsigned long addr
, int avoid_reserve
)
1155 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1156 struct hstate
*h
= hstate_vma(vma
);
1160 struct hugetlb_cgroup
*h_cg
;
1162 idx
= hstate_index(h
);
1164 * Processes that did not create the mapping will have no
1165 * reserves and will not have accounted against subpool
1166 * limit. Check that the subpool limit can be made before
1167 * satisfying the allocation MAP_NORESERVE mappings may also
1168 * need pages and subpool limit allocated allocated if no reserve
1171 chg
= vma_needs_reservation(h
, vma
, addr
);
1173 return ERR_PTR(-ENOMEM
);
1174 if (chg
|| avoid_reserve
)
1175 if (hugepage_subpool_get_pages(spool
, 1))
1176 return ERR_PTR(-ENOSPC
);
1178 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1180 if (chg
|| avoid_reserve
)
1181 hugepage_subpool_put_pages(spool
, 1);
1182 return ERR_PTR(-ENOSPC
);
1184 spin_lock(&hugetlb_lock
);
1185 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, chg
);
1187 spin_unlock(&hugetlb_lock
);
1188 page
= alloc_buddy_huge_page(h
, NUMA_NO_NODE
);
1190 hugetlb_cgroup_uncharge_cgroup(idx
,
1191 pages_per_huge_page(h
),
1193 if (chg
|| avoid_reserve
)
1194 hugepage_subpool_put_pages(spool
, 1);
1195 return ERR_PTR(-ENOSPC
);
1197 spin_lock(&hugetlb_lock
);
1198 list_move(&page
->lru
, &h
->hugepage_activelist
);
1201 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1202 spin_unlock(&hugetlb_lock
);
1204 set_page_private(page
, (unsigned long)spool
);
1206 vma_commit_reservation(h
, vma
, addr
);
1211 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1212 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1213 * where no ERR_VALUE is expected to be returned.
1215 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1216 unsigned long addr
, int avoid_reserve
)
1218 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1224 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1226 struct huge_bootmem_page
*m
;
1229 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1232 addr
= __alloc_bootmem_node_nopanic(NODE_DATA(node
),
1233 huge_page_size(h
), huge_page_size(h
), 0);
1237 * Use the beginning of the huge page to store the
1238 * huge_bootmem_page struct (until gather_bootmem
1239 * puts them into the mem_map).
1248 BUG_ON((unsigned long)virt_to_phys(m
) & (huge_page_size(h
) - 1));
1249 /* Put them into a private list first because mem_map is not up yet */
1250 list_add(&m
->list
, &huge_boot_pages
);
1255 static void prep_compound_huge_page(struct page
*page
, int order
)
1257 if (unlikely(order
> (MAX_ORDER
- 1)))
1258 prep_compound_gigantic_page(page
, order
);
1260 prep_compound_page(page
, order
);
1263 /* Put bootmem huge pages into the standard lists after mem_map is up */
1264 static void __init
gather_bootmem_prealloc(void)
1266 struct huge_bootmem_page
*m
;
1268 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1269 struct hstate
*h
= m
->hstate
;
1272 #ifdef CONFIG_HIGHMEM
1273 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1274 free_bootmem_late((unsigned long)m
,
1275 sizeof(struct huge_bootmem_page
));
1277 page
= virt_to_page(m
);
1279 __ClearPageReserved(page
);
1280 WARN_ON(page_count(page
) != 1);
1281 prep_compound_huge_page(page
, h
->order
);
1282 prep_new_huge_page(h
, page
, page_to_nid(page
));
1284 * If we had gigantic hugepages allocated at boot time, we need
1285 * to restore the 'stolen' pages to totalram_pages in order to
1286 * fix confusing memory reports from free(1) and another
1287 * side-effects, like CommitLimit going negative.
1289 if (h
->order
> (MAX_ORDER
- 1))
1290 adjust_managed_page_count(page
, 1 << h
->order
);
1294 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
1298 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
1299 if (h
->order
>= MAX_ORDER
) {
1300 if (!alloc_bootmem_huge_page(h
))
1302 } else if (!alloc_fresh_huge_page(h
,
1303 &node_states
[N_MEMORY
]))
1306 h
->max_huge_pages
= i
;
1309 static void __init
hugetlb_init_hstates(void)
1313 for_each_hstate(h
) {
1314 /* oversize hugepages were init'ed in early boot */
1315 if (h
->order
< MAX_ORDER
)
1316 hugetlb_hstate_alloc_pages(h
);
1320 static char * __init
memfmt(char *buf
, unsigned long n
)
1322 if (n
>= (1UL << 30))
1323 sprintf(buf
, "%lu GB", n
>> 30);
1324 else if (n
>= (1UL << 20))
1325 sprintf(buf
, "%lu MB", n
>> 20);
1327 sprintf(buf
, "%lu KB", n
>> 10);
1331 static void __init
report_hugepages(void)
1335 for_each_hstate(h
) {
1337 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1338 memfmt(buf
, huge_page_size(h
)),
1339 h
->free_huge_pages
);
1343 #ifdef CONFIG_HIGHMEM
1344 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
1345 nodemask_t
*nodes_allowed
)
1349 if (h
->order
>= MAX_ORDER
)
1352 for_each_node_mask(i
, *nodes_allowed
) {
1353 struct page
*page
, *next
;
1354 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
1355 list_for_each_entry_safe(page
, next
, freel
, lru
) {
1356 if (count
>= h
->nr_huge_pages
)
1358 if (PageHighMem(page
))
1360 list_del(&page
->lru
);
1361 update_and_free_page(h
, page
);
1362 h
->free_huge_pages
--;
1363 h
->free_huge_pages_node
[page_to_nid(page
)]--;
1368 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
1369 nodemask_t
*nodes_allowed
)
1375 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1376 * balanced by operating on them in a round-robin fashion.
1377 * Returns 1 if an adjustment was made.
1379 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1384 VM_BUG_ON(delta
!= -1 && delta
!= 1);
1387 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1388 if (h
->surplus_huge_pages_node
[node
])
1392 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1393 if (h
->surplus_huge_pages_node
[node
] <
1394 h
->nr_huge_pages_node
[node
])
1401 h
->surplus_huge_pages
+= delta
;
1402 h
->surplus_huge_pages_node
[node
] += delta
;
1406 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1407 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
1408 nodemask_t
*nodes_allowed
)
1410 unsigned long min_count
, ret
;
1412 if (h
->order
>= MAX_ORDER
)
1413 return h
->max_huge_pages
;
1416 * Increase the pool size
1417 * First take pages out of surplus state. Then make up the
1418 * remaining difference by allocating fresh huge pages.
1420 * We might race with alloc_buddy_huge_page() here and be unable
1421 * to convert a surplus huge page to a normal huge page. That is
1422 * not critical, though, it just means the overall size of the
1423 * pool might be one hugepage larger than it needs to be, but
1424 * within all the constraints specified by the sysctls.
1426 spin_lock(&hugetlb_lock
);
1427 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
1428 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
1432 while (count
> persistent_huge_pages(h
)) {
1434 * If this allocation races such that we no longer need the
1435 * page, free_huge_page will handle it by freeing the page
1436 * and reducing the surplus.
1438 spin_unlock(&hugetlb_lock
);
1439 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
1440 spin_lock(&hugetlb_lock
);
1444 /* Bail for signals. Probably ctrl-c from user */
1445 if (signal_pending(current
))
1450 * Decrease the pool size
1451 * First return free pages to the buddy allocator (being careful
1452 * to keep enough around to satisfy reservations). Then place
1453 * pages into surplus state as needed so the pool will shrink
1454 * to the desired size as pages become free.
1456 * By placing pages into the surplus state independent of the
1457 * overcommit value, we are allowing the surplus pool size to
1458 * exceed overcommit. There are few sane options here. Since
1459 * alloc_buddy_huge_page() is checking the global counter,
1460 * though, we'll note that we're not allowed to exceed surplus
1461 * and won't grow the pool anywhere else. Not until one of the
1462 * sysctls are changed, or the surplus pages go out of use.
1464 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
1465 min_count
= max(count
, min_count
);
1466 try_to_free_low(h
, min_count
, nodes_allowed
);
1467 while (min_count
< persistent_huge_pages(h
)) {
1468 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
1471 while (count
< persistent_huge_pages(h
)) {
1472 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
1476 ret
= persistent_huge_pages(h
);
1477 spin_unlock(&hugetlb_lock
);
1481 #define HSTATE_ATTR_RO(_name) \
1482 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1484 #define HSTATE_ATTR(_name) \
1485 static struct kobj_attribute _name##_attr = \
1486 __ATTR(_name, 0644, _name##_show, _name##_store)
1488 static struct kobject
*hugepages_kobj
;
1489 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1491 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
1493 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
1497 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1498 if (hstate_kobjs
[i
] == kobj
) {
1500 *nidp
= NUMA_NO_NODE
;
1504 return kobj_to_node_hstate(kobj
, nidp
);
1507 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
1508 struct kobj_attribute
*attr
, char *buf
)
1511 unsigned long nr_huge_pages
;
1514 h
= kobj_to_hstate(kobj
, &nid
);
1515 if (nid
== NUMA_NO_NODE
)
1516 nr_huge_pages
= h
->nr_huge_pages
;
1518 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
1520 return sprintf(buf
, "%lu\n", nr_huge_pages
);
1523 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
1524 struct kobject
*kobj
, struct kobj_attribute
*attr
,
1525 const char *buf
, size_t len
)
1529 unsigned long count
;
1531 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
1533 err
= kstrtoul(buf
, 10, &count
);
1537 h
= kobj_to_hstate(kobj
, &nid
);
1538 if (h
->order
>= MAX_ORDER
) {
1543 if (nid
== NUMA_NO_NODE
) {
1545 * global hstate attribute
1547 if (!(obey_mempolicy
&&
1548 init_nodemask_of_mempolicy(nodes_allowed
))) {
1549 NODEMASK_FREE(nodes_allowed
);
1550 nodes_allowed
= &node_states
[N_MEMORY
];
1552 } else if (nodes_allowed
) {
1554 * per node hstate attribute: adjust count to global,
1555 * but restrict alloc/free to the specified node.
1557 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
1558 init_nodemask_of_node(nodes_allowed
, nid
);
1560 nodes_allowed
= &node_states
[N_MEMORY
];
1562 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
1564 if (nodes_allowed
!= &node_states
[N_MEMORY
])
1565 NODEMASK_FREE(nodes_allowed
);
1569 NODEMASK_FREE(nodes_allowed
);
1573 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
1574 struct kobj_attribute
*attr
, char *buf
)
1576 return nr_hugepages_show_common(kobj
, attr
, buf
);
1579 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
1580 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1582 return nr_hugepages_store_common(false, kobj
, attr
, buf
, len
);
1584 HSTATE_ATTR(nr_hugepages
);
1589 * hstate attribute for optionally mempolicy-based constraint on persistent
1590 * huge page alloc/free.
1592 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
1593 struct kobj_attribute
*attr
, char *buf
)
1595 return nr_hugepages_show_common(kobj
, attr
, buf
);
1598 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
1599 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
1601 return nr_hugepages_store_common(true, kobj
, attr
, buf
, len
);
1603 HSTATE_ATTR(nr_hugepages_mempolicy
);
1607 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
1608 struct kobj_attribute
*attr
, char *buf
)
1610 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1611 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
1614 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
1615 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
1618 unsigned long input
;
1619 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1621 if (h
->order
>= MAX_ORDER
)
1624 err
= kstrtoul(buf
, 10, &input
);
1628 spin_lock(&hugetlb_lock
);
1629 h
->nr_overcommit_huge_pages
= input
;
1630 spin_unlock(&hugetlb_lock
);
1634 HSTATE_ATTR(nr_overcommit_hugepages
);
1636 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
1637 struct kobj_attribute
*attr
, char *buf
)
1640 unsigned long free_huge_pages
;
1643 h
= kobj_to_hstate(kobj
, &nid
);
1644 if (nid
== NUMA_NO_NODE
)
1645 free_huge_pages
= h
->free_huge_pages
;
1647 free_huge_pages
= h
->free_huge_pages_node
[nid
];
1649 return sprintf(buf
, "%lu\n", free_huge_pages
);
1651 HSTATE_ATTR_RO(free_hugepages
);
1653 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
1654 struct kobj_attribute
*attr
, char *buf
)
1656 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
1657 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
1659 HSTATE_ATTR_RO(resv_hugepages
);
1661 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
1662 struct kobj_attribute
*attr
, char *buf
)
1665 unsigned long surplus_huge_pages
;
1668 h
= kobj_to_hstate(kobj
, &nid
);
1669 if (nid
== NUMA_NO_NODE
)
1670 surplus_huge_pages
= h
->surplus_huge_pages
;
1672 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
1674 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
1676 HSTATE_ATTR_RO(surplus_hugepages
);
1678 static struct attribute
*hstate_attrs
[] = {
1679 &nr_hugepages_attr
.attr
,
1680 &nr_overcommit_hugepages_attr
.attr
,
1681 &free_hugepages_attr
.attr
,
1682 &resv_hugepages_attr
.attr
,
1683 &surplus_hugepages_attr
.attr
,
1685 &nr_hugepages_mempolicy_attr
.attr
,
1690 static struct attribute_group hstate_attr_group
= {
1691 .attrs
= hstate_attrs
,
1694 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
1695 struct kobject
**hstate_kobjs
,
1696 struct attribute_group
*hstate_attr_group
)
1699 int hi
= hstate_index(h
);
1701 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
1702 if (!hstate_kobjs
[hi
])
1705 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
1707 kobject_put(hstate_kobjs
[hi
]);
1712 static void __init
hugetlb_sysfs_init(void)
1717 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
1718 if (!hugepages_kobj
)
1721 for_each_hstate(h
) {
1722 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
1723 hstate_kobjs
, &hstate_attr_group
);
1725 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
1732 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1733 * with node devices in node_devices[] using a parallel array. The array
1734 * index of a node device or _hstate == node id.
1735 * This is here to avoid any static dependency of the node device driver, in
1736 * the base kernel, on the hugetlb module.
1738 struct node_hstate
{
1739 struct kobject
*hugepages_kobj
;
1740 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
1742 struct node_hstate node_hstates
[MAX_NUMNODES
];
1745 * A subset of global hstate attributes for node devices
1747 static struct attribute
*per_node_hstate_attrs
[] = {
1748 &nr_hugepages_attr
.attr
,
1749 &free_hugepages_attr
.attr
,
1750 &surplus_hugepages_attr
.attr
,
1754 static struct attribute_group per_node_hstate_attr_group
= {
1755 .attrs
= per_node_hstate_attrs
,
1759 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1760 * Returns node id via non-NULL nidp.
1762 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1766 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
1767 struct node_hstate
*nhs
= &node_hstates
[nid
];
1769 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
1770 if (nhs
->hstate_kobjs
[i
] == kobj
) {
1782 * Unregister hstate attributes from a single node device.
1783 * No-op if no hstate attributes attached.
1785 static void hugetlb_unregister_node(struct node
*node
)
1788 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1790 if (!nhs
->hugepages_kobj
)
1791 return; /* no hstate attributes */
1793 for_each_hstate(h
) {
1794 int idx
= hstate_index(h
);
1795 if (nhs
->hstate_kobjs
[idx
]) {
1796 kobject_put(nhs
->hstate_kobjs
[idx
]);
1797 nhs
->hstate_kobjs
[idx
] = NULL
;
1801 kobject_put(nhs
->hugepages_kobj
);
1802 nhs
->hugepages_kobj
= NULL
;
1806 * hugetlb module exit: unregister hstate attributes from node devices
1809 static void hugetlb_unregister_all_nodes(void)
1814 * disable node device registrations.
1816 register_hugetlbfs_with_node(NULL
, NULL
);
1819 * remove hstate attributes from any nodes that have them.
1821 for (nid
= 0; nid
< nr_node_ids
; nid
++)
1822 hugetlb_unregister_node(node_devices
[nid
]);
1826 * Register hstate attributes for a single node device.
1827 * No-op if attributes already registered.
1829 static void hugetlb_register_node(struct node
*node
)
1832 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
1835 if (nhs
->hugepages_kobj
)
1836 return; /* already allocated */
1838 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
1840 if (!nhs
->hugepages_kobj
)
1843 for_each_hstate(h
) {
1844 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
1846 &per_node_hstate_attr_group
);
1848 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1849 h
->name
, node
->dev
.id
);
1850 hugetlb_unregister_node(node
);
1857 * hugetlb init time: register hstate attributes for all registered node
1858 * devices of nodes that have memory. All on-line nodes should have
1859 * registered their associated device by this time.
1861 static void hugetlb_register_all_nodes(void)
1865 for_each_node_state(nid
, N_MEMORY
) {
1866 struct node
*node
= node_devices
[nid
];
1867 if (node
->dev
.id
== nid
)
1868 hugetlb_register_node(node
);
1872 * Let the node device driver know we're here so it can
1873 * [un]register hstate attributes on node hotplug.
1875 register_hugetlbfs_with_node(hugetlb_register_node
,
1876 hugetlb_unregister_node
);
1878 #else /* !CONFIG_NUMA */
1880 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
1888 static void hugetlb_unregister_all_nodes(void) { }
1890 static void hugetlb_register_all_nodes(void) { }
1894 static void __exit
hugetlb_exit(void)
1898 hugetlb_unregister_all_nodes();
1900 for_each_hstate(h
) {
1901 kobject_put(hstate_kobjs
[hstate_index(h
)]);
1904 kobject_put(hugepages_kobj
);
1906 module_exit(hugetlb_exit
);
1908 static int __init
hugetlb_init(void)
1910 /* Some platform decide whether they support huge pages at boot
1911 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1912 * there is no such support
1914 if (HPAGE_SHIFT
== 0)
1917 if (!size_to_hstate(default_hstate_size
)) {
1918 default_hstate_size
= HPAGE_SIZE
;
1919 if (!size_to_hstate(default_hstate_size
))
1920 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
1922 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
1923 if (default_hstate_max_huge_pages
)
1924 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
1926 hugetlb_init_hstates();
1927 gather_bootmem_prealloc();
1930 hugetlb_sysfs_init();
1931 hugetlb_register_all_nodes();
1932 hugetlb_cgroup_file_init();
1936 module_init(hugetlb_init
);
1938 /* Should be called on processing a hugepagesz=... option */
1939 void __init
hugetlb_add_hstate(unsigned order
)
1944 if (size_to_hstate(PAGE_SIZE
<< order
)) {
1945 pr_warning("hugepagesz= specified twice, ignoring\n");
1948 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
1950 h
= &hstates
[hugetlb_max_hstate
++];
1952 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
1953 h
->nr_huge_pages
= 0;
1954 h
->free_huge_pages
= 0;
1955 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
1956 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
1957 INIT_LIST_HEAD(&h
->hugepage_activelist
);
1958 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
1959 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
1960 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
1961 huge_page_size(h
)/1024);
1966 static int __init
hugetlb_nrpages_setup(char *s
)
1969 static unsigned long *last_mhp
;
1972 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1973 * so this hugepages= parameter goes to the "default hstate".
1975 if (!hugetlb_max_hstate
)
1976 mhp
= &default_hstate_max_huge_pages
;
1978 mhp
= &parsed_hstate
->max_huge_pages
;
1980 if (mhp
== last_mhp
) {
1981 pr_warning("hugepages= specified twice without "
1982 "interleaving hugepagesz=, ignoring\n");
1986 if (sscanf(s
, "%lu", mhp
) <= 0)
1990 * Global state is always initialized later in hugetlb_init.
1991 * But we need to allocate >= MAX_ORDER hstates here early to still
1992 * use the bootmem allocator.
1994 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
1995 hugetlb_hstate_alloc_pages(parsed_hstate
);
2001 __setup("hugepages=", hugetlb_nrpages_setup
);
2003 static int __init
hugetlb_default_setup(char *s
)
2005 default_hstate_size
= memparse(s
, &s
);
2008 __setup("default_hugepagesz=", hugetlb_default_setup
);
2010 static unsigned int cpuset_mems_nr(unsigned int *array
)
2013 unsigned int nr
= 0;
2015 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2021 #ifdef CONFIG_SYSCTL
2022 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2023 struct ctl_table
*table
, int write
,
2024 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2026 struct hstate
*h
= &default_hstate
;
2030 tmp
= h
->max_huge_pages
;
2032 if (write
&& h
->order
>= MAX_ORDER
)
2036 table
->maxlen
= sizeof(unsigned long);
2037 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2042 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
,
2043 GFP_KERNEL
| __GFP_NORETRY
);
2044 if (!(obey_mempolicy
&&
2045 init_nodemask_of_mempolicy(nodes_allowed
))) {
2046 NODEMASK_FREE(nodes_allowed
);
2047 nodes_allowed
= &node_states
[N_MEMORY
];
2049 h
->max_huge_pages
= set_max_huge_pages(h
, tmp
, nodes_allowed
);
2051 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2052 NODEMASK_FREE(nodes_allowed
);
2058 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2059 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2062 return hugetlb_sysctl_handler_common(false, table
, write
,
2063 buffer
, length
, ppos
);
2067 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2068 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2070 return hugetlb_sysctl_handler_common(true, table
, write
,
2071 buffer
, length
, ppos
);
2073 #endif /* CONFIG_NUMA */
2075 int hugetlb_treat_movable_handler(struct ctl_table
*table
, int write
,
2076 void __user
*buffer
,
2077 size_t *length
, loff_t
*ppos
)
2079 proc_dointvec(table
, write
, buffer
, length
, ppos
);
2080 if (hugepages_treat_as_movable
)
2081 htlb_alloc_mask
= GFP_HIGHUSER_MOVABLE
;
2083 htlb_alloc_mask
= GFP_HIGHUSER
;
2087 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2088 void __user
*buffer
,
2089 size_t *length
, loff_t
*ppos
)
2091 struct hstate
*h
= &default_hstate
;
2095 tmp
= h
->nr_overcommit_huge_pages
;
2097 if (write
&& h
->order
>= MAX_ORDER
)
2101 table
->maxlen
= sizeof(unsigned long);
2102 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2107 spin_lock(&hugetlb_lock
);
2108 h
->nr_overcommit_huge_pages
= tmp
;
2109 spin_unlock(&hugetlb_lock
);
2115 #endif /* CONFIG_SYSCTL */
2117 void hugetlb_report_meminfo(struct seq_file
*m
)
2119 struct hstate
*h
= &default_hstate
;
2121 "HugePages_Total: %5lu\n"
2122 "HugePages_Free: %5lu\n"
2123 "HugePages_Rsvd: %5lu\n"
2124 "HugePages_Surp: %5lu\n"
2125 "Hugepagesize: %8lu kB\n",
2129 h
->surplus_huge_pages
,
2130 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2133 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2135 struct hstate
*h
= &default_hstate
;
2137 "Node %d HugePages_Total: %5u\n"
2138 "Node %d HugePages_Free: %5u\n"
2139 "Node %d HugePages_Surp: %5u\n",
2140 nid
, h
->nr_huge_pages_node
[nid
],
2141 nid
, h
->free_huge_pages_node
[nid
],
2142 nid
, h
->surplus_huge_pages_node
[nid
]);
2145 void hugetlb_show_meminfo(void)
2150 for_each_node_state(nid
, N_MEMORY
)
2152 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2154 h
->nr_huge_pages_node
[nid
],
2155 h
->free_huge_pages_node
[nid
],
2156 h
->surplus_huge_pages_node
[nid
],
2157 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2160 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2161 unsigned long hugetlb_total_pages(void)
2164 unsigned long nr_total_pages
= 0;
2167 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2168 return nr_total_pages
;
2171 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2175 spin_lock(&hugetlb_lock
);
2177 * When cpuset is configured, it breaks the strict hugetlb page
2178 * reservation as the accounting is done on a global variable. Such
2179 * reservation is completely rubbish in the presence of cpuset because
2180 * the reservation is not checked against page availability for the
2181 * current cpuset. Application can still potentially OOM'ed by kernel
2182 * with lack of free htlb page in cpuset that the task is in.
2183 * Attempt to enforce strict accounting with cpuset is almost
2184 * impossible (or too ugly) because cpuset is too fluid that
2185 * task or memory node can be dynamically moved between cpusets.
2187 * The change of semantics for shared hugetlb mapping with cpuset is
2188 * undesirable. However, in order to preserve some of the semantics,
2189 * we fall back to check against current free page availability as
2190 * a best attempt and hopefully to minimize the impact of changing
2191 * semantics that cpuset has.
2194 if (gather_surplus_pages(h
, delta
) < 0)
2197 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2198 return_unused_surplus_pages(h
, delta
);
2205 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2208 spin_unlock(&hugetlb_lock
);
2212 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2214 struct resv_map
*resv
= vma_resv_map(vma
);
2217 * This new VMA should share its siblings reservation map if present.
2218 * The VMA will only ever have a valid reservation map pointer where
2219 * it is being copied for another still existing VMA. As that VMA
2220 * has a reference to the reservation map it cannot disappear until
2221 * after this open call completes. It is therefore safe to take a
2222 * new reference here without additional locking.
2225 kref_get(&resv
->refs
);
2228 static void resv_map_put(struct vm_area_struct
*vma
)
2230 struct resv_map
*resv
= vma_resv_map(vma
);
2234 kref_put(&resv
->refs
, resv_map_release
);
2237 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2239 struct hstate
*h
= hstate_vma(vma
);
2240 struct resv_map
*resv
= vma_resv_map(vma
);
2241 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2242 unsigned long reserve
;
2243 unsigned long start
;
2247 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2248 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2250 reserve
= (end
- start
) -
2251 region_count(&resv
->regions
, start
, end
);
2256 hugetlb_acct_memory(h
, -reserve
);
2257 hugepage_subpool_put_pages(spool
, reserve
);
2263 * We cannot handle pagefaults against hugetlb pages at all. They cause
2264 * handle_mm_fault() to try to instantiate regular-sized pages in the
2265 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2268 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
2274 const struct vm_operations_struct hugetlb_vm_ops
= {
2275 .fault
= hugetlb_vm_op_fault
,
2276 .open
= hugetlb_vm_op_open
,
2277 .close
= hugetlb_vm_op_close
,
2280 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
2286 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
2287 vma
->vm_page_prot
)));
2289 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
2290 vma
->vm_page_prot
));
2292 entry
= pte_mkyoung(entry
);
2293 entry
= pte_mkhuge(entry
);
2294 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
2299 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
2300 unsigned long address
, pte_t
*ptep
)
2304 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
2305 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
2306 update_mmu_cache(vma
, address
, ptep
);
2310 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
2311 struct vm_area_struct
*vma
)
2313 pte_t
*src_pte
, *dst_pte
, entry
;
2314 struct page
*ptepage
;
2317 struct hstate
*h
= hstate_vma(vma
);
2318 unsigned long sz
= huge_page_size(h
);
2320 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
2322 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
2323 src_pte
= huge_pte_offset(src
, addr
);
2326 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
2330 /* If the pagetables are shared don't copy or take references */
2331 if (dst_pte
== src_pte
)
2334 spin_lock(&dst
->page_table_lock
);
2335 spin_lock_nested(&src
->page_table_lock
, SINGLE_DEPTH_NESTING
);
2336 if (!huge_pte_none(huge_ptep_get(src_pte
))) {
2338 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
2339 entry
= huge_ptep_get(src_pte
);
2340 ptepage
= pte_page(entry
);
2342 page_dup_rmap(ptepage
);
2343 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
2345 spin_unlock(&src
->page_table_lock
);
2346 spin_unlock(&dst
->page_table_lock
);
2354 static int is_hugetlb_entry_migration(pte_t pte
)
2358 if (huge_pte_none(pte
) || pte_present(pte
))
2360 swp
= pte_to_swp_entry(pte
);
2361 if (non_swap_entry(swp
) && is_migration_entry(swp
))
2367 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
2371 if (huge_pte_none(pte
) || pte_present(pte
))
2373 swp
= pte_to_swp_entry(pte
);
2374 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
2380 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
2381 unsigned long start
, unsigned long end
,
2382 struct page
*ref_page
)
2384 int force_flush
= 0;
2385 struct mm_struct
*mm
= vma
->vm_mm
;
2386 unsigned long address
;
2390 struct hstate
*h
= hstate_vma(vma
);
2391 unsigned long sz
= huge_page_size(h
);
2392 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
2393 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
2395 WARN_ON(!is_vm_hugetlb_page(vma
));
2396 BUG_ON(start
& ~huge_page_mask(h
));
2397 BUG_ON(end
& ~huge_page_mask(h
));
2399 tlb_start_vma(tlb
, vma
);
2400 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2402 spin_lock(&mm
->page_table_lock
);
2403 for (address
= start
; address
< end
; address
+= sz
) {
2404 ptep
= huge_pte_offset(mm
, address
);
2408 if (huge_pmd_unshare(mm
, &address
, ptep
))
2411 pte
= huge_ptep_get(ptep
);
2412 if (huge_pte_none(pte
))
2416 * HWPoisoned hugepage is already unmapped and dropped reference
2418 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
2419 huge_pte_clear(mm
, address
, ptep
);
2423 page
= pte_page(pte
);
2425 * If a reference page is supplied, it is because a specific
2426 * page is being unmapped, not a range. Ensure the page we
2427 * are about to unmap is the actual page of interest.
2430 if (page
!= ref_page
)
2434 * Mark the VMA as having unmapped its page so that
2435 * future faults in this VMA will fail rather than
2436 * looking like data was lost
2438 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
2441 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
2442 tlb_remove_tlb_entry(tlb
, ptep
, address
);
2443 if (huge_pte_dirty(pte
))
2444 set_page_dirty(page
);
2446 page_remove_rmap(page
);
2447 force_flush
= !__tlb_remove_page(tlb
, page
);
2450 /* Bail out after unmapping reference page if supplied */
2454 spin_unlock(&mm
->page_table_lock
);
2456 * mmu_gather ran out of room to batch pages, we break out of
2457 * the PTE lock to avoid doing the potential expensive TLB invalidate
2458 * and page-free while holding it.
2463 if (address
< end
&& !ref_page
)
2466 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2467 tlb_end_vma(tlb
, vma
);
2470 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
2471 struct vm_area_struct
*vma
, unsigned long start
,
2472 unsigned long end
, struct page
*ref_page
)
2474 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
2477 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2478 * test will fail on a vma being torn down, and not grab a page table
2479 * on its way out. We're lucky that the flag has such an appropriate
2480 * name, and can in fact be safely cleared here. We could clear it
2481 * before the __unmap_hugepage_range above, but all that's necessary
2482 * is to clear it before releasing the i_mmap_mutex. This works
2483 * because in the context this is called, the VMA is about to be
2484 * destroyed and the i_mmap_mutex is held.
2486 vma
->vm_flags
&= ~VM_MAYSHARE
;
2489 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
2490 unsigned long end
, struct page
*ref_page
)
2492 struct mm_struct
*mm
;
2493 struct mmu_gather tlb
;
2497 tlb_gather_mmu(&tlb
, mm
, start
, end
);
2498 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
2499 tlb_finish_mmu(&tlb
, start
, end
);
2503 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2504 * mappping it owns the reserve page for. The intention is to unmap the page
2505 * from other VMAs and let the children be SIGKILLed if they are faulting the
2508 static int unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2509 struct page
*page
, unsigned long address
)
2511 struct hstate
*h
= hstate_vma(vma
);
2512 struct vm_area_struct
*iter_vma
;
2513 struct address_space
*mapping
;
2517 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2518 * from page cache lookup which is in HPAGE_SIZE units.
2520 address
= address
& huge_page_mask(h
);
2521 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
2523 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
2526 * Take the mapping lock for the duration of the table walk. As
2527 * this mapping should be shared between all the VMAs,
2528 * __unmap_hugepage_range() is called as the lock is already held
2530 mutex_lock(&mapping
->i_mmap_mutex
);
2531 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
2532 /* Do not unmap the current VMA */
2533 if (iter_vma
== vma
)
2537 * Unmap the page from other VMAs without their own reserves.
2538 * They get marked to be SIGKILLed if they fault in these
2539 * areas. This is because a future no-page fault on this VMA
2540 * could insert a zeroed page instead of the data existing
2541 * from the time of fork. This would look like data corruption
2543 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
2544 unmap_hugepage_range(iter_vma
, address
,
2545 address
+ huge_page_size(h
), page
);
2547 mutex_unlock(&mapping
->i_mmap_mutex
);
2553 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2554 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2555 * cannot race with other handlers or page migration.
2556 * Keep the pte_same checks anyway to make transition from the mutex easier.
2558 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2559 unsigned long address
, pte_t
*ptep
, pte_t pte
,
2560 struct page
*pagecache_page
)
2562 struct hstate
*h
= hstate_vma(vma
);
2563 struct page
*old_page
, *new_page
;
2564 int outside_reserve
= 0;
2565 unsigned long mmun_start
; /* For mmu_notifiers */
2566 unsigned long mmun_end
; /* For mmu_notifiers */
2568 old_page
= pte_page(pte
);
2571 /* If no-one else is actually using this page, avoid the copy
2572 * and just make the page writable */
2573 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
2574 page_move_anon_rmap(old_page
, vma
, address
);
2575 set_huge_ptep_writable(vma
, address
, ptep
);
2580 * If the process that created a MAP_PRIVATE mapping is about to
2581 * perform a COW due to a shared page count, attempt to satisfy
2582 * the allocation without using the existing reserves. The pagecache
2583 * page is used to determine if the reserve at this address was
2584 * consumed or not. If reserves were used, a partial faulted mapping
2585 * at the time of fork() could consume its reserves on COW instead
2586 * of the full address range.
2588 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
2589 old_page
!= pagecache_page
)
2590 outside_reserve
= 1;
2592 page_cache_get(old_page
);
2594 /* Drop page_table_lock as buddy allocator may be called */
2595 spin_unlock(&mm
->page_table_lock
);
2596 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
2598 if (IS_ERR(new_page
)) {
2599 long err
= PTR_ERR(new_page
);
2600 page_cache_release(old_page
);
2603 * If a process owning a MAP_PRIVATE mapping fails to COW,
2604 * it is due to references held by a child and an insufficient
2605 * huge page pool. To guarantee the original mappers
2606 * reliability, unmap the page from child processes. The child
2607 * may get SIGKILLed if it later faults.
2609 if (outside_reserve
) {
2610 BUG_ON(huge_pte_none(pte
));
2611 if (unmap_ref_private(mm
, vma
, old_page
, address
)) {
2612 BUG_ON(huge_pte_none(pte
));
2613 spin_lock(&mm
->page_table_lock
);
2614 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2615 if (likely(pte_same(huge_ptep_get(ptep
), pte
)))
2616 goto retry_avoidcopy
;
2618 * race occurs while re-acquiring page_table_lock, and
2626 /* Caller expects lock to be held */
2627 spin_lock(&mm
->page_table_lock
);
2629 return VM_FAULT_OOM
;
2631 return VM_FAULT_SIGBUS
;
2635 * When the original hugepage is shared one, it does not have
2636 * anon_vma prepared.
2638 if (unlikely(anon_vma_prepare(vma
))) {
2639 page_cache_release(new_page
);
2640 page_cache_release(old_page
);
2641 /* Caller expects lock to be held */
2642 spin_lock(&mm
->page_table_lock
);
2643 return VM_FAULT_OOM
;
2646 copy_user_huge_page(new_page
, old_page
, address
, vma
,
2647 pages_per_huge_page(h
));
2648 __SetPageUptodate(new_page
);
2650 mmun_start
= address
& huge_page_mask(h
);
2651 mmun_end
= mmun_start
+ huge_page_size(h
);
2652 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2654 * Retake the page_table_lock to check for racing updates
2655 * before the page tables are altered
2657 spin_lock(&mm
->page_table_lock
);
2658 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
2659 if (likely(pte_same(huge_ptep_get(ptep
), pte
))) {
2660 ClearPagePrivate(new_page
);
2663 huge_ptep_clear_flush(vma
, address
, ptep
);
2664 set_huge_pte_at(mm
, address
, ptep
,
2665 make_huge_pte(vma
, new_page
, 1));
2666 page_remove_rmap(old_page
);
2667 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
2668 /* Make the old page be freed below */
2669 new_page
= old_page
;
2671 spin_unlock(&mm
->page_table_lock
);
2672 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2673 page_cache_release(new_page
);
2674 page_cache_release(old_page
);
2676 /* Caller expects lock to be held */
2677 spin_lock(&mm
->page_table_lock
);
2681 /* Return the pagecache page at a given address within a VMA */
2682 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
2683 struct vm_area_struct
*vma
, unsigned long address
)
2685 struct address_space
*mapping
;
2688 mapping
= vma
->vm_file
->f_mapping
;
2689 idx
= vma_hugecache_offset(h
, vma
, address
);
2691 return find_lock_page(mapping
, idx
);
2695 * Return whether there is a pagecache page to back given address within VMA.
2696 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2698 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
2699 struct vm_area_struct
*vma
, unsigned long address
)
2701 struct address_space
*mapping
;
2705 mapping
= vma
->vm_file
->f_mapping
;
2706 idx
= vma_hugecache_offset(h
, vma
, address
);
2708 page
= find_get_page(mapping
, idx
);
2711 return page
!= NULL
;
2714 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2715 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
2717 struct hstate
*h
= hstate_vma(vma
);
2718 int ret
= VM_FAULT_SIGBUS
;
2723 struct address_space
*mapping
;
2727 * Currently, we are forced to kill the process in the event the
2728 * original mapper has unmapped pages from the child due to a failed
2729 * COW. Warn that such a situation has occurred as it may not be obvious
2731 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
2732 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2737 mapping
= vma
->vm_file
->f_mapping
;
2738 idx
= vma_hugecache_offset(h
, vma
, address
);
2741 * Use page lock to guard against racing truncation
2742 * before we get page_table_lock.
2745 page
= find_lock_page(mapping
, idx
);
2747 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2750 page
= alloc_huge_page(vma
, address
, 0);
2752 ret
= PTR_ERR(page
);
2756 ret
= VM_FAULT_SIGBUS
;
2759 clear_huge_page(page
, address
, pages_per_huge_page(h
));
2760 __SetPageUptodate(page
);
2762 if (vma
->vm_flags
& VM_MAYSHARE
) {
2764 struct inode
*inode
= mapping
->host
;
2766 err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
2773 ClearPagePrivate(page
);
2775 spin_lock(&inode
->i_lock
);
2776 inode
->i_blocks
+= blocks_per_huge_page(h
);
2777 spin_unlock(&inode
->i_lock
);
2780 if (unlikely(anon_vma_prepare(vma
))) {
2782 goto backout_unlocked
;
2788 * If memory error occurs between mmap() and fault, some process
2789 * don't have hwpoisoned swap entry for errored virtual address.
2790 * So we need to block hugepage fault by PG_hwpoison bit check.
2792 if (unlikely(PageHWPoison(page
))) {
2793 ret
= VM_FAULT_HWPOISON
|
2794 VM_FAULT_SET_HINDEX(hstate_index(h
));
2795 goto backout_unlocked
;
2800 * If we are going to COW a private mapping later, we examine the
2801 * pending reservations for this page now. This will ensure that
2802 * any allocations necessary to record that reservation occur outside
2805 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
))
2806 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2808 goto backout_unlocked
;
2811 spin_lock(&mm
->page_table_lock
);
2812 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
2817 if (!huge_pte_none(huge_ptep_get(ptep
)))
2821 ClearPagePrivate(page
);
2822 hugepage_add_new_anon_rmap(page
, vma
, address
);
2825 page_dup_rmap(page
);
2826 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
2827 && (vma
->vm_flags
& VM_SHARED
)));
2828 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
2830 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
2831 /* Optimization, do the COW without a second fault */
2832 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
);
2835 spin_unlock(&mm
->page_table_lock
);
2841 spin_unlock(&mm
->page_table_lock
);
2848 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2849 unsigned long address
, unsigned int flags
)
2854 struct page
*page
= NULL
;
2855 struct page
*pagecache_page
= NULL
;
2856 static DEFINE_MUTEX(hugetlb_instantiation_mutex
);
2857 struct hstate
*h
= hstate_vma(vma
);
2859 address
&= huge_page_mask(h
);
2861 ptep
= huge_pte_offset(mm
, address
);
2863 entry
= huge_ptep_get(ptep
);
2864 if (unlikely(is_hugetlb_entry_migration(entry
))) {
2865 migration_entry_wait_huge(mm
, ptep
);
2867 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
2868 return VM_FAULT_HWPOISON_LARGE
|
2869 VM_FAULT_SET_HINDEX(hstate_index(h
));
2872 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
2874 return VM_FAULT_OOM
;
2877 * Serialize hugepage allocation and instantiation, so that we don't
2878 * get spurious allocation failures if two CPUs race to instantiate
2879 * the same page in the page cache.
2881 mutex_lock(&hugetlb_instantiation_mutex
);
2882 entry
= huge_ptep_get(ptep
);
2883 if (huge_pte_none(entry
)) {
2884 ret
= hugetlb_no_page(mm
, vma
, address
, ptep
, flags
);
2891 * If we are going to COW the mapping later, we examine the pending
2892 * reservations for this page now. This will ensure that any
2893 * allocations necessary to record that reservation occur outside the
2894 * spinlock. For private mappings, we also lookup the pagecache
2895 * page now as it is used to determine if a reservation has been
2898 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
2899 if (vma_needs_reservation(h
, vma
, address
) < 0) {
2904 if (!(vma
->vm_flags
& VM_MAYSHARE
))
2905 pagecache_page
= hugetlbfs_pagecache_page(h
,
2910 * hugetlb_cow() requires page locks of pte_page(entry) and
2911 * pagecache_page, so here we need take the former one
2912 * when page != pagecache_page or !pagecache_page.
2913 * Note that locking order is always pagecache_page -> page,
2914 * so no worry about deadlock.
2916 page
= pte_page(entry
);
2918 if (page
!= pagecache_page
)
2921 spin_lock(&mm
->page_table_lock
);
2922 /* Check for a racing update before calling hugetlb_cow */
2923 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
2924 goto out_page_table_lock
;
2927 if (flags
& FAULT_FLAG_WRITE
) {
2928 if (!huge_pte_write(entry
)) {
2929 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
2931 goto out_page_table_lock
;
2933 entry
= huge_pte_mkdirty(entry
);
2935 entry
= pte_mkyoung(entry
);
2936 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
2937 flags
& FAULT_FLAG_WRITE
))
2938 update_mmu_cache(vma
, address
, ptep
);
2940 out_page_table_lock
:
2941 spin_unlock(&mm
->page_table_lock
);
2943 if (pagecache_page
) {
2944 unlock_page(pagecache_page
);
2945 put_page(pagecache_page
);
2947 if (page
!= pagecache_page
)
2952 mutex_unlock(&hugetlb_instantiation_mutex
);
2957 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
2958 struct page
**pages
, struct vm_area_struct
**vmas
,
2959 unsigned long *position
, unsigned long *nr_pages
,
2960 long i
, unsigned int flags
)
2962 unsigned long pfn_offset
;
2963 unsigned long vaddr
= *position
;
2964 unsigned long remainder
= *nr_pages
;
2965 struct hstate
*h
= hstate_vma(vma
);
2967 spin_lock(&mm
->page_table_lock
);
2968 while (vaddr
< vma
->vm_end
&& remainder
) {
2974 * Some archs (sparc64, sh*) have multiple pte_ts to
2975 * each hugepage. We have to make sure we get the
2976 * first, for the page indexing below to work.
2978 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
2979 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
2982 * When coredumping, it suits get_dump_page if we just return
2983 * an error where there's an empty slot with no huge pagecache
2984 * to back it. This way, we avoid allocating a hugepage, and
2985 * the sparse dumpfile avoids allocating disk blocks, but its
2986 * huge holes still show up with zeroes where they need to be.
2988 if (absent
&& (flags
& FOLL_DUMP
) &&
2989 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
2995 * We need call hugetlb_fault for both hugepages under migration
2996 * (in which case hugetlb_fault waits for the migration,) and
2997 * hwpoisoned hugepages (in which case we need to prevent the
2998 * caller from accessing to them.) In order to do this, we use
2999 * here is_swap_pte instead of is_hugetlb_entry_migration and
3000 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3001 * both cases, and because we can't follow correct pages
3002 * directly from any kind of swap entries.
3004 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3005 ((flags
& FOLL_WRITE
) &&
3006 !huge_pte_write(huge_ptep_get(pte
)))) {
3009 spin_unlock(&mm
->page_table_lock
);
3010 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3011 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3012 spin_lock(&mm
->page_table_lock
);
3013 if (!(ret
& VM_FAULT_ERROR
))
3020 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3021 page
= pte_page(huge_ptep_get(pte
));
3024 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3035 if (vaddr
< vma
->vm_end
&& remainder
&&
3036 pfn_offset
< pages_per_huge_page(h
)) {
3038 * We use pfn_offset to avoid touching the pageframes
3039 * of this compound page.
3044 spin_unlock(&mm
->page_table_lock
);
3045 *nr_pages
= remainder
;
3048 return i
? i
: -EFAULT
;
3051 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3052 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3054 struct mm_struct
*mm
= vma
->vm_mm
;
3055 unsigned long start
= address
;
3058 struct hstate
*h
= hstate_vma(vma
);
3059 unsigned long pages
= 0;
3061 BUG_ON(address
>= end
);
3062 flush_cache_range(vma
, address
, end
);
3064 mutex_lock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3065 spin_lock(&mm
->page_table_lock
);
3066 for (; address
< end
; address
+= huge_page_size(h
)) {
3067 ptep
= huge_pte_offset(mm
, address
);
3070 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3074 if (!huge_pte_none(huge_ptep_get(ptep
))) {
3075 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3076 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3077 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3078 set_huge_pte_at(mm
, address
, ptep
, pte
);
3082 spin_unlock(&mm
->page_table_lock
);
3084 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3085 * may have cleared our pud entry and done put_page on the page table:
3086 * once we release i_mmap_mutex, another task can do the final put_page
3087 * and that page table be reused and filled with junk.
3089 flush_tlb_range(vma
, start
, end
);
3090 mutex_unlock(&vma
->vm_file
->f_mapping
->i_mmap_mutex
);
3092 return pages
<< h
->order
;
3095 int hugetlb_reserve_pages(struct inode
*inode
,
3097 struct vm_area_struct
*vma
,
3098 vm_flags_t vm_flags
)
3101 struct hstate
*h
= hstate_inode(inode
);
3102 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3105 * Only apply hugepage reservation if asked. At fault time, an
3106 * attempt will be made for VM_NORESERVE to allocate a page
3107 * without using reserves
3109 if (vm_flags
& VM_NORESERVE
)
3113 * Shared mappings base their reservation on the number of pages that
3114 * are already allocated on behalf of the file. Private mappings need
3115 * to reserve the full area even if read-only as mprotect() may be
3116 * called to make the mapping read-write. Assume !vma is a shm mapping
3118 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3119 chg
= region_chg(&inode
->i_mapping
->private_list
, from
, to
);
3121 struct resv_map
*resv_map
= resv_map_alloc();
3127 set_vma_resv_map(vma
, resv_map
);
3128 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
3136 /* There must be enough pages in the subpool for the mapping */
3137 if (hugepage_subpool_get_pages(spool
, chg
)) {
3143 * Check enough hugepages are available for the reservation.
3144 * Hand the pages back to the subpool if there are not
3146 ret
= hugetlb_acct_memory(h
, chg
);
3148 hugepage_subpool_put_pages(spool
, chg
);
3153 * Account for the reservations made. Shared mappings record regions
3154 * that have reservations as they are shared by multiple VMAs.
3155 * When the last VMA disappears, the region map says how much
3156 * the reservation was and the page cache tells how much of
3157 * the reservation was consumed. Private mappings are per-VMA and
3158 * only the consumed reservations are tracked. When the VMA
3159 * disappears, the original reservation is the VMA size and the
3160 * consumed reservations are stored in the map. Hence, nothing
3161 * else has to be done for private mappings here
3163 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
3164 region_add(&inode
->i_mapping
->private_list
, from
, to
);
3172 void hugetlb_unreserve_pages(struct inode
*inode
, long offset
, long freed
)
3174 struct hstate
*h
= hstate_inode(inode
);
3175 long chg
= region_truncate(&inode
->i_mapping
->private_list
, offset
);
3176 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3178 spin_lock(&inode
->i_lock
);
3179 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
3180 spin_unlock(&inode
->i_lock
);
3182 hugepage_subpool_put_pages(spool
, (chg
- freed
));
3183 hugetlb_acct_memory(h
, -(chg
- freed
));
3186 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3187 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
3188 struct vm_area_struct
*vma
,
3189 unsigned long addr
, pgoff_t idx
)
3191 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
3193 unsigned long sbase
= saddr
& PUD_MASK
;
3194 unsigned long s_end
= sbase
+ PUD_SIZE
;
3196 /* Allow segments to share if only one is marked locked */
3197 unsigned long vm_flags
= vma
->vm_flags
& ~VM_LOCKED
;
3198 unsigned long svm_flags
= svma
->vm_flags
& ~VM_LOCKED
;
3201 * match the virtual addresses, permission and the alignment of the
3204 if (pmd_index(addr
) != pmd_index(saddr
) ||
3205 vm_flags
!= svm_flags
||
3206 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
3212 static int vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
3214 unsigned long base
= addr
& PUD_MASK
;
3215 unsigned long end
= base
+ PUD_SIZE
;
3218 * check on proper vm_flags and page table alignment
3220 if (vma
->vm_flags
& VM_MAYSHARE
&&
3221 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
3227 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3228 * and returns the corresponding pte. While this is not necessary for the
3229 * !shared pmd case because we can allocate the pmd later as well, it makes the
3230 * code much cleaner. pmd allocation is essential for the shared case because
3231 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3232 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3233 * bad pmd for sharing.
3235 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3237 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
3238 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
3239 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
3241 struct vm_area_struct
*svma
;
3242 unsigned long saddr
;
3246 if (!vma_shareable(vma
, addr
))
3247 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3249 mutex_lock(&mapping
->i_mmap_mutex
);
3250 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
3254 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
3256 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
3258 get_page(virt_to_page(spte
));
3267 spin_lock(&mm
->page_table_lock
);
3269 pud_populate(mm
, pud
,
3270 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
3272 put_page(virt_to_page(spte
));
3273 spin_unlock(&mm
->page_table_lock
);
3275 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3276 mutex_unlock(&mapping
->i_mmap_mutex
);
3281 * unmap huge page backed by shared pte.
3283 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3284 * indicated by page_count > 1, unmap is achieved by clearing pud and
3285 * decrementing the ref count. If count == 1, the pte page is not shared.
3287 * called with vma->vm_mm->page_table_lock held.
3289 * returns: 1 successfully unmapped a shared pte page
3290 * 0 the underlying pte page is not shared, or it is the last user
3292 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
3294 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
3295 pud_t
*pud
= pud_offset(pgd
, *addr
);
3297 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
3298 if (page_count(virt_to_page(ptep
)) == 1)
3302 put_page(virt_to_page(ptep
));
3303 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
3306 #define want_pmd_share() (1)
3307 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3308 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
3312 #define want_pmd_share() (0)
3313 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3315 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3316 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
3317 unsigned long addr
, unsigned long sz
)
3323 pgd
= pgd_offset(mm
, addr
);
3324 pud
= pud_alloc(mm
, pgd
, addr
);
3326 if (sz
== PUD_SIZE
) {
3329 BUG_ON(sz
!= PMD_SIZE
);
3330 if (want_pmd_share() && pud_none(*pud
))
3331 pte
= huge_pmd_share(mm
, addr
, pud
);
3333 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
3336 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
3341 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
3347 pgd
= pgd_offset(mm
, addr
);
3348 if (pgd_present(*pgd
)) {
3349 pud
= pud_offset(pgd
, addr
);
3350 if (pud_present(*pud
)) {
3352 return (pte_t
*)pud
;
3353 pmd
= pmd_offset(pud
, addr
);
3356 return (pte_t
*) pmd
;
3360 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
3361 pmd_t
*pmd
, int write
)
3365 page
= pte_page(*(pte_t
*)pmd
);
3367 page
+= ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
3372 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3373 pud_t
*pud
, int write
)
3377 page
= pte_page(*(pte_t
*)pud
);
3379 page
+= ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
3383 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3385 /* Can be overriden by architectures */
3386 __attribute__((weak
)) struct page
*
3387 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
3388 pud_t
*pud
, int write
)
3394 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3396 #ifdef CONFIG_MEMORY_FAILURE
3398 /* Should be called in hugetlb_lock */
3399 static int is_hugepage_on_freelist(struct page
*hpage
)
3403 struct hstate
*h
= page_hstate(hpage
);
3404 int nid
= page_to_nid(hpage
);
3406 list_for_each_entry_safe(page
, tmp
, &h
->hugepage_freelists
[nid
], lru
)
3413 * This function is called from memory failure code.
3414 * Assume the caller holds page lock of the head page.
3416 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
3418 struct hstate
*h
= page_hstate(hpage
);
3419 int nid
= page_to_nid(hpage
);
3422 spin_lock(&hugetlb_lock
);
3423 if (is_hugepage_on_freelist(hpage
)) {
3425 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3426 * but dangling hpage->lru can trigger list-debug warnings
3427 * (this happens when we call unpoison_memory() on it),
3428 * so let it point to itself with list_del_init().
3430 list_del_init(&hpage
->lru
);
3431 set_page_refcounted(hpage
);
3432 h
->free_huge_pages
--;
3433 h
->free_huge_pages_node
[nid
]--;
3436 spin_unlock(&hugetlb_lock
);
3441 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
3443 VM_BUG_ON(!PageHead(page
));
3444 if (!get_page_unless_zero(page
))
3446 spin_lock(&hugetlb_lock
);
3447 list_move_tail(&page
->lru
, list
);
3448 spin_unlock(&hugetlb_lock
);
3452 void putback_active_hugepage(struct page
*page
)
3454 VM_BUG_ON(!PageHead(page
));
3455 spin_lock(&hugetlb_lock
);
3456 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
3457 spin_unlock(&hugetlb_lock
);