2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.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/sched/signal.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/jhash.h>
28 #include <asm/pgtable.h>
32 #include <linux/hugetlb.h>
33 #include <linux/hugetlb_cgroup.h>
34 #include <linux/node.h>
35 #include <linux/userfaultfd_k.h>
38 int hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
49 __initdata
LIST_HEAD(huge_boot_pages
);
51 /* for command line parsing */
52 static struct hstate
* __initdata parsed_hstate
;
53 static unsigned long __initdata default_hstate_max_huge_pages
;
54 static unsigned long __initdata default_hstate_size
;
55 static bool __initdata parsed_valid_hugepagesz
= true;
58 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
59 * free_huge_pages, and surplus_huge_pages.
61 DEFINE_SPINLOCK(hugetlb_lock
);
64 * Serializes faults on the same logical page. This is used to
65 * prevent spurious OOMs when the hugepage pool is fully utilized.
67 static int num_fault_mutexes
;
68 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
70 /* Forward declaration */
71 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
73 static char * __init
memfmt(char *buf
, unsigned long n
)
76 sprintf(buf
, "%lu GB", n
>> 30);
77 else if (n
>= (1UL << 20))
78 sprintf(buf
, "%lu MB", n
>> 20);
80 sprintf(buf
, "%lu KB", n
>> 10);
84 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
86 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
88 spin_unlock(&spool
->lock
);
90 /* If no pages are used, and no other handles to the subpool
91 * remain, give up any reservations mased on minimum size and
94 if (spool
->min_hpages
!= -1)
95 hugetlb_acct_memory(spool
->hstate
,
101 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
104 struct hugepage_subpool
*spool
;
106 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
110 spin_lock_init(&spool
->lock
);
112 spool
->max_hpages
= max_hpages
;
114 spool
->min_hpages
= min_hpages
;
116 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
120 spool
->rsv_hpages
= min_hpages
;
125 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
127 spin_lock(&spool
->lock
);
128 BUG_ON(!spool
->count
);
130 unlock_or_release_subpool(spool
);
134 * Subpool accounting for allocating and reserving pages.
135 * Return -ENOMEM if there are not enough resources to satisfy the
136 * the request. Otherwise, return the number of pages by which the
137 * global pools must be adjusted (upward). The returned value may
138 * only be different than the passed value (delta) in the case where
139 * a subpool minimum size must be manitained.
141 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
149 spin_lock(&spool
->lock
);
151 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
152 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
153 spool
->used_hpages
+= delta
;
160 /* minimum size accounting */
161 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
162 if (delta
> spool
->rsv_hpages
) {
164 * Asking for more reserves than those already taken on
165 * behalf of subpool. Return difference.
167 ret
= delta
- spool
->rsv_hpages
;
168 spool
->rsv_hpages
= 0;
170 ret
= 0; /* reserves already accounted for */
171 spool
->rsv_hpages
-= delta
;
176 spin_unlock(&spool
->lock
);
181 * Subpool accounting for freeing and unreserving pages.
182 * Return the number of global page reservations that must be dropped.
183 * The return value may only be different than the passed value (delta)
184 * in the case where a subpool minimum size must be maintained.
186 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
194 spin_lock(&spool
->lock
);
196 if (spool
->max_hpages
!= -1) /* maximum size accounting */
197 spool
->used_hpages
-= delta
;
199 /* minimum size accounting */
200 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
201 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
204 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
206 spool
->rsv_hpages
+= delta
;
207 if (spool
->rsv_hpages
> spool
->min_hpages
)
208 spool
->rsv_hpages
= spool
->min_hpages
;
212 * If hugetlbfs_put_super couldn't free spool due to an outstanding
213 * quota reference, free it now.
215 unlock_or_release_subpool(spool
);
220 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
222 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
225 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
227 return subpool_inode(file_inode(vma
->vm_file
));
231 * Region tracking -- allows tracking of reservations and instantiated pages
232 * across the pages in a mapping.
234 * The region data structures are embedded into a resv_map and protected
235 * by a resv_map's lock. The set of regions within the resv_map represent
236 * reservations for huge pages, or huge pages that have already been
237 * instantiated within the map. The from and to elements are huge page
238 * indicies into the associated mapping. from indicates the starting index
239 * of the region. to represents the first index past the end of the region.
241 * For example, a file region structure with from == 0 and to == 4 represents
242 * four huge pages in a mapping. It is important to note that the to element
243 * represents the first element past the end of the region. This is used in
244 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
246 * Interval notation of the form [from, to) will be used to indicate that
247 * the endpoint from is inclusive and to is exclusive.
250 struct list_head link
;
256 * Add the huge page range represented by [f, t) to the reserve
257 * map. In the normal case, existing regions will be expanded
258 * to accommodate the specified range. Sufficient regions should
259 * exist for expansion due to the previous call to region_chg
260 * with the same range. However, it is possible that region_del
261 * could have been called after region_chg and modifed the map
262 * in such a way that no region exists to be expanded. In this
263 * case, pull a region descriptor from the cache associated with
264 * the map and use that for the new range.
266 * Return the number of new huge pages added to the map. This
267 * number is greater than or equal to zero.
269 static long region_add(struct resv_map
*resv
, long f
, long t
)
271 struct list_head
*head
= &resv
->regions
;
272 struct file_region
*rg
, *nrg
, *trg
;
275 spin_lock(&resv
->lock
);
276 /* Locate the region we are either in or before. */
277 list_for_each_entry(rg
, head
, link
)
282 * If no region exists which can be expanded to include the
283 * specified range, the list must have been modified by an
284 * interleving call to region_del(). Pull a region descriptor
285 * from the cache and use it for this range.
287 if (&rg
->link
== head
|| t
< rg
->from
) {
288 VM_BUG_ON(resv
->region_cache_count
<= 0);
290 resv
->region_cache_count
--;
291 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
,
293 list_del(&nrg
->link
);
297 list_add(&nrg
->link
, rg
->link
.prev
);
303 /* Round our left edge to the current segment if it encloses us. */
307 /* Check for and consume any regions we now overlap with. */
309 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
310 if (&rg
->link
== head
)
315 /* If this area reaches higher then extend our area to
316 * include it completely. If this is not the first area
317 * which we intend to reuse, free it. */
321 /* Decrement return value by the deleted range.
322 * Another range will span this area so that by
323 * end of routine add will be >= zero
325 add
-= (rg
->to
- rg
->from
);
331 add
+= (nrg
->from
- f
); /* Added to beginning of region */
333 add
+= t
- nrg
->to
; /* Added to end of region */
337 resv
->adds_in_progress
--;
338 spin_unlock(&resv
->lock
);
344 * Examine the existing reserve map and determine how many
345 * huge pages in the specified range [f, t) are NOT currently
346 * represented. This routine is called before a subsequent
347 * call to region_add that will actually modify the reserve
348 * map to add the specified range [f, t). region_chg does
349 * not change the number of huge pages represented by the
350 * map. However, if the existing regions in the map can not
351 * be expanded to represent the new range, a new file_region
352 * structure is added to the map as a placeholder. This is
353 * so that the subsequent region_add call will have all the
354 * regions it needs and will not fail.
356 * Upon entry, region_chg will also examine the cache of region descriptors
357 * associated with the map. If there are not enough descriptors cached, one
358 * will be allocated for the in progress add operation.
360 * Returns the number of huge pages that need to be added to the existing
361 * reservation map for the range [f, t). This number is greater or equal to
362 * zero. -ENOMEM is returned if a new file_region structure or cache entry
363 * is needed and can not be allocated.
365 static long region_chg(struct resv_map
*resv
, long f
, long t
)
367 struct list_head
*head
= &resv
->regions
;
368 struct file_region
*rg
, *nrg
= NULL
;
372 spin_lock(&resv
->lock
);
374 resv
->adds_in_progress
++;
377 * Check for sufficient descriptors in the cache to accommodate
378 * the number of in progress add operations.
380 if (resv
->adds_in_progress
> resv
->region_cache_count
) {
381 struct file_region
*trg
;
383 VM_BUG_ON(resv
->adds_in_progress
- resv
->region_cache_count
> 1);
384 /* Must drop lock to allocate a new descriptor. */
385 resv
->adds_in_progress
--;
386 spin_unlock(&resv
->lock
);
388 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
394 spin_lock(&resv
->lock
);
395 list_add(&trg
->link
, &resv
->region_cache
);
396 resv
->region_cache_count
++;
400 /* Locate the region we are before or in. */
401 list_for_each_entry(rg
, head
, link
)
405 /* If we are below the current region then a new region is required.
406 * Subtle, allocate a new region at the position but make it zero
407 * size such that we can guarantee to record the reservation. */
408 if (&rg
->link
== head
|| t
< rg
->from
) {
410 resv
->adds_in_progress
--;
411 spin_unlock(&resv
->lock
);
412 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
418 INIT_LIST_HEAD(&nrg
->link
);
422 list_add(&nrg
->link
, rg
->link
.prev
);
427 /* Round our left edge to the current segment if it encloses us. */
432 /* Check for and consume any regions we now overlap with. */
433 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
434 if (&rg
->link
== head
)
439 /* We overlap with this area, if it extends further than
440 * us then we must extend ourselves. Account for its
441 * existing reservation. */
446 chg
-= rg
->to
- rg
->from
;
450 spin_unlock(&resv
->lock
);
451 /* We already know we raced and no longer need the new region */
455 spin_unlock(&resv
->lock
);
460 * Abort the in progress add operation. The adds_in_progress field
461 * of the resv_map keeps track of the operations in progress between
462 * calls to region_chg and region_add. Operations are sometimes
463 * aborted after the call to region_chg. In such cases, region_abort
464 * is called to decrement the adds_in_progress counter.
466 * NOTE: The range arguments [f, t) are not needed or used in this
467 * routine. They are kept to make reading the calling code easier as
468 * arguments will match the associated region_chg call.
470 static void region_abort(struct resv_map
*resv
, long f
, long t
)
472 spin_lock(&resv
->lock
);
473 VM_BUG_ON(!resv
->region_cache_count
);
474 resv
->adds_in_progress
--;
475 spin_unlock(&resv
->lock
);
479 * Delete the specified range [f, t) from the reserve map. If the
480 * t parameter is LONG_MAX, this indicates that ALL regions after f
481 * should be deleted. Locate the regions which intersect [f, t)
482 * and either trim, delete or split the existing regions.
484 * Returns the number of huge pages deleted from the reserve map.
485 * In the normal case, the return value is zero or more. In the
486 * case where a region must be split, a new region descriptor must
487 * be allocated. If the allocation fails, -ENOMEM will be returned.
488 * NOTE: If the parameter t == LONG_MAX, then we will never split
489 * a region and possibly return -ENOMEM. Callers specifying
490 * t == LONG_MAX do not need to check for -ENOMEM error.
492 static long region_del(struct resv_map
*resv
, long f
, long t
)
494 struct list_head
*head
= &resv
->regions
;
495 struct file_region
*rg
, *trg
;
496 struct file_region
*nrg
= NULL
;
500 spin_lock(&resv
->lock
);
501 list_for_each_entry_safe(rg
, trg
, head
, link
) {
503 * Skip regions before the range to be deleted. file_region
504 * ranges are normally of the form [from, to). However, there
505 * may be a "placeholder" entry in the map which is of the form
506 * (from, to) with from == to. Check for placeholder entries
507 * at the beginning of the range to be deleted.
509 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
515 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
517 * Check for an entry in the cache before dropping
518 * lock and attempting allocation.
521 resv
->region_cache_count
> resv
->adds_in_progress
) {
522 nrg
= list_first_entry(&resv
->region_cache
,
525 list_del(&nrg
->link
);
526 resv
->region_cache_count
--;
530 spin_unlock(&resv
->lock
);
531 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
539 /* New entry for end of split region */
542 INIT_LIST_HEAD(&nrg
->link
);
544 /* Original entry is trimmed */
547 list_add(&nrg
->link
, &rg
->link
);
552 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
553 del
+= rg
->to
- rg
->from
;
559 if (f
<= rg
->from
) { /* Trim beginning of region */
562 } else { /* Trim end of region */
568 spin_unlock(&resv
->lock
);
574 * A rare out of memory error was encountered which prevented removal of
575 * the reserve map region for a page. The huge page itself was free'ed
576 * and removed from the page cache. This routine will adjust the subpool
577 * usage count, and the global reserve count if needed. By incrementing
578 * these counts, the reserve map entry which could not be deleted will
579 * appear as a "reserved" entry instead of simply dangling with incorrect
582 void hugetlb_fix_reserve_counts(struct inode
*inode
)
584 struct hugepage_subpool
*spool
= subpool_inode(inode
);
587 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
589 struct hstate
*h
= hstate_inode(inode
);
591 hugetlb_acct_memory(h
, 1);
596 * Count and return the number of huge pages in the reserve map
597 * that intersect with the range [f, t).
599 static long region_count(struct resv_map
*resv
, long f
, long t
)
601 struct list_head
*head
= &resv
->regions
;
602 struct file_region
*rg
;
605 spin_lock(&resv
->lock
);
606 /* Locate each segment we overlap with, and count that overlap. */
607 list_for_each_entry(rg
, head
, link
) {
616 seg_from
= max(rg
->from
, f
);
617 seg_to
= min(rg
->to
, t
);
619 chg
+= seg_to
- seg_from
;
621 spin_unlock(&resv
->lock
);
627 * Convert the address within this vma to the page offset within
628 * the mapping, in pagecache page units; huge pages here.
630 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
631 struct vm_area_struct
*vma
, unsigned long address
)
633 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
634 (vma
->vm_pgoff
>> huge_page_order(h
));
637 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
638 unsigned long address
)
640 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
642 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
645 * Return the size of the pages allocated when backing a VMA. In the majority
646 * cases this will be same size as used by the page table entries.
648 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
650 struct hstate
*hstate
;
652 if (!is_vm_hugetlb_page(vma
))
655 hstate
= hstate_vma(vma
);
657 return 1UL << huge_page_shift(hstate
);
659 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
662 * Return the page size being used by the MMU to back a VMA. In the majority
663 * of cases, the page size used by the kernel matches the MMU size. On
664 * architectures where it differs, an architecture-specific version of this
665 * function is required.
667 #ifndef vma_mmu_pagesize
668 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
670 return vma_kernel_pagesize(vma
);
675 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
676 * bits of the reservation map pointer, which are always clear due to
679 #define HPAGE_RESV_OWNER (1UL << 0)
680 #define HPAGE_RESV_UNMAPPED (1UL << 1)
681 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
684 * These helpers are used to track how many pages are reserved for
685 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
686 * is guaranteed to have their future faults succeed.
688 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
689 * the reserve counters are updated with the hugetlb_lock held. It is safe
690 * to reset the VMA at fork() time as it is not in use yet and there is no
691 * chance of the global counters getting corrupted as a result of the values.
693 * The private mapping reservation is represented in a subtly different
694 * manner to a shared mapping. A shared mapping has a region map associated
695 * with the underlying file, this region map represents the backing file
696 * pages which have ever had a reservation assigned which this persists even
697 * after the page is instantiated. A private mapping has a region map
698 * associated with the original mmap which is attached to all VMAs which
699 * reference it, this region map represents those offsets which have consumed
700 * reservation ie. where pages have been instantiated.
702 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
704 return (unsigned long)vma
->vm_private_data
;
707 static void set_vma_private_data(struct vm_area_struct
*vma
,
710 vma
->vm_private_data
= (void *)value
;
713 struct resv_map
*resv_map_alloc(void)
715 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
716 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
718 if (!resv_map
|| !rg
) {
724 kref_init(&resv_map
->refs
);
725 spin_lock_init(&resv_map
->lock
);
726 INIT_LIST_HEAD(&resv_map
->regions
);
728 resv_map
->adds_in_progress
= 0;
730 INIT_LIST_HEAD(&resv_map
->region_cache
);
731 list_add(&rg
->link
, &resv_map
->region_cache
);
732 resv_map
->region_cache_count
= 1;
737 void resv_map_release(struct kref
*ref
)
739 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
740 struct list_head
*head
= &resv_map
->region_cache
;
741 struct file_region
*rg
, *trg
;
743 /* Clear out any active regions before we release the map. */
744 region_del(resv_map
, 0, LONG_MAX
);
746 /* ... and any entries left in the cache */
747 list_for_each_entry_safe(rg
, trg
, head
, link
) {
752 VM_BUG_ON(resv_map
->adds_in_progress
);
757 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
759 return inode
->i_mapping
->private_data
;
762 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
764 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
765 if (vma
->vm_flags
& VM_MAYSHARE
) {
766 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
767 struct inode
*inode
= mapping
->host
;
769 return inode_resv_map(inode
);
772 return (struct resv_map
*)(get_vma_private_data(vma
) &
777 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
780 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
782 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
783 HPAGE_RESV_MASK
) | (unsigned long)map
);
786 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
789 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
791 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
794 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
796 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
798 return (get_vma_private_data(vma
) & flag
) != 0;
801 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
802 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
804 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
805 if (!(vma
->vm_flags
& VM_MAYSHARE
))
806 vma
->vm_private_data
= (void *)0;
809 /* Returns true if the VMA has associated reserve pages */
810 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
812 if (vma
->vm_flags
& VM_NORESERVE
) {
814 * This address is already reserved by other process(chg == 0),
815 * so, we should decrement reserved count. Without decrementing,
816 * reserve count remains after releasing inode, because this
817 * allocated page will go into page cache and is regarded as
818 * coming from reserved pool in releasing step. Currently, we
819 * don't have any other solution to deal with this situation
820 * properly, so add work-around here.
822 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
828 /* Shared mappings always use reserves */
829 if (vma
->vm_flags
& VM_MAYSHARE
) {
831 * We know VM_NORESERVE is not set. Therefore, there SHOULD
832 * be a region map for all pages. The only situation where
833 * there is no region map is if a hole was punched via
834 * fallocate. In this case, there really are no reverves to
835 * use. This situation is indicated if chg != 0.
844 * Only the process that called mmap() has reserves for
847 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
849 * Like the shared case above, a hole punch or truncate
850 * could have been performed on the private mapping.
851 * Examine the value of chg to determine if reserves
852 * actually exist or were previously consumed.
853 * Very Subtle - The value of chg comes from a previous
854 * call to vma_needs_reserves(). The reserve map for
855 * private mappings has different (opposite) semantics
856 * than that of shared mappings. vma_needs_reserves()
857 * has already taken this difference in semantics into
858 * account. Therefore, the meaning of chg is the same
859 * as in the shared case above. Code could easily be
860 * combined, but keeping it separate draws attention to
861 * subtle differences.
872 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
874 int nid
= page_to_nid(page
);
875 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
876 h
->free_huge_pages
++;
877 h
->free_huge_pages_node
[nid
]++;
880 static struct page
*dequeue_huge_page_node_exact(struct hstate
*h
, int nid
)
884 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
885 if (!PageHWPoison(page
))
888 * if 'non-isolated free hugepage' not found on the list,
889 * the allocation fails.
891 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
893 list_move(&page
->lru
, &h
->hugepage_activelist
);
894 set_page_refcounted(page
);
895 h
->free_huge_pages
--;
896 h
->free_huge_pages_node
[nid
]--;
900 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
905 if (nid
!= NUMA_NO_NODE
)
906 return dequeue_huge_page_node_exact(h
, nid
);
908 for_each_online_node(node
) {
909 page
= dequeue_huge_page_node_exact(h
, node
);
916 /* Movability of hugepages depends on migration support. */
917 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
919 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
920 return GFP_HIGHUSER_MOVABLE
;
925 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
926 struct vm_area_struct
*vma
,
927 unsigned long address
, int avoid_reserve
,
930 struct page
*page
= NULL
;
931 struct mempolicy
*mpol
;
932 nodemask_t
*nodemask
;
935 struct zonelist
*zonelist
;
938 unsigned int cpuset_mems_cookie
;
941 * A child process with MAP_PRIVATE mappings created by their parent
942 * have no page reserves. This check ensures that reservations are
943 * not "stolen". The child may still get SIGKILLed
945 if (!vma_has_reserves(vma
, chg
) &&
946 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
949 /* If reserves cannot be used, ensure enough pages are in the pool */
950 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
954 cpuset_mems_cookie
= read_mems_allowed_begin();
955 gfp_mask
= htlb_alloc_mask(h
);
956 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
957 zonelist
= node_zonelist(nid
, gfp_mask
);
959 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
960 MAX_NR_ZONES
- 1, nodemask
) {
961 if (cpuset_zone_allowed(zone
, gfp_mask
)) {
962 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
966 if (!vma_has_reserves(vma
, chg
))
969 SetPagePrivate(page
);
970 h
->resv_huge_pages
--;
977 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
986 * common helper functions for hstate_next_node_to_{alloc|free}.
987 * We may have allocated or freed a huge page based on a different
988 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
989 * be outside of *nodes_allowed. Ensure that we use an allowed
990 * node for alloc or free.
992 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
994 nid
= next_node_in(nid
, *nodes_allowed
);
995 VM_BUG_ON(nid
>= MAX_NUMNODES
);
1000 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1002 if (!node_isset(nid
, *nodes_allowed
))
1003 nid
= next_node_allowed(nid
, nodes_allowed
);
1008 * returns the previously saved node ["this node"] from which to
1009 * allocate a persistent huge page for the pool and advance the
1010 * next node from which to allocate, handling wrap at end of node
1013 static int hstate_next_node_to_alloc(struct hstate
*h
,
1014 nodemask_t
*nodes_allowed
)
1018 VM_BUG_ON(!nodes_allowed
);
1020 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1021 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1027 * helper for free_pool_huge_page() - return the previously saved
1028 * node ["this node"] from which to free a huge page. Advance the
1029 * next node id whether or not we find a free huge page to free so
1030 * that the next attempt to free addresses the next node.
1032 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1036 VM_BUG_ON(!nodes_allowed
);
1038 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1039 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1044 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1045 for (nr_nodes = nodes_weight(*mask); \
1047 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1050 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1051 for (nr_nodes = nodes_weight(*mask); \
1053 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1056 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1057 static void destroy_compound_gigantic_page(struct page
*page
,
1061 int nr_pages
= 1 << order
;
1062 struct page
*p
= page
+ 1;
1064 atomic_set(compound_mapcount_ptr(page
), 0);
1065 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1066 clear_compound_head(p
);
1067 set_page_refcounted(p
);
1070 set_compound_order(page
, 0);
1071 __ClearPageHead(page
);
1074 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1076 free_contig_range(page_to_pfn(page
), 1 << order
);
1079 static int __alloc_gigantic_page(unsigned long start_pfn
,
1080 unsigned long nr_pages
)
1082 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1083 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
1087 static bool pfn_range_valid_gigantic(struct zone
*z
,
1088 unsigned long start_pfn
, unsigned long nr_pages
)
1090 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1093 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1097 page
= pfn_to_page(i
);
1099 if (page_zone(page
) != z
)
1102 if (PageReserved(page
))
1105 if (page_count(page
) > 0)
1115 static bool zone_spans_last_pfn(const struct zone
*zone
,
1116 unsigned long start_pfn
, unsigned long nr_pages
)
1118 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1119 return zone_spans_pfn(zone
, last_pfn
);
1122 static struct page
*alloc_gigantic_page(int nid
, unsigned int order
)
1124 unsigned long nr_pages
= 1 << order
;
1125 unsigned long ret
, pfn
, flags
;
1128 z
= NODE_DATA(nid
)->node_zones
;
1129 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
1130 spin_lock_irqsave(&z
->lock
, flags
);
1132 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
1133 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
1134 if (pfn_range_valid_gigantic(z
, pfn
, nr_pages
)) {
1136 * We release the zone lock here because
1137 * alloc_contig_range() will also lock the zone
1138 * at some point. If there's an allocation
1139 * spinning on this lock, it may win the race
1140 * and cause alloc_contig_range() to fail...
1142 spin_unlock_irqrestore(&z
->lock
, flags
);
1143 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
1145 return pfn_to_page(pfn
);
1146 spin_lock_irqsave(&z
->lock
, flags
);
1151 spin_unlock_irqrestore(&z
->lock
, flags
);
1157 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1158 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1160 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
1164 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
1166 prep_compound_gigantic_page(page
, huge_page_order(h
));
1167 prep_new_huge_page(h
, page
, nid
);
1173 static int alloc_fresh_gigantic_page(struct hstate
*h
,
1174 nodemask_t
*nodes_allowed
)
1176 struct page
*page
= NULL
;
1179 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1180 page
= alloc_fresh_gigantic_page_node(h
, node
);
1188 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1189 static inline bool gigantic_page_supported(void) { return false; }
1190 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1191 static inline void destroy_compound_gigantic_page(struct page
*page
,
1192 unsigned int order
) { }
1193 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
1194 nodemask_t
*nodes_allowed
) { return 0; }
1197 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1201 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1205 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1206 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1207 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1208 1 << PG_referenced
| 1 << PG_dirty
|
1209 1 << PG_active
| 1 << PG_private
|
1212 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1213 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1214 set_page_refcounted(page
);
1215 if (hstate_is_gigantic(h
)) {
1216 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1217 free_gigantic_page(page
, huge_page_order(h
));
1219 __free_pages(page
, huge_page_order(h
));
1223 struct hstate
*size_to_hstate(unsigned long size
)
1227 for_each_hstate(h
) {
1228 if (huge_page_size(h
) == size
)
1235 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1236 * to hstate->hugepage_activelist.)
1238 * This function can be called for tail pages, but never returns true for them.
1240 bool page_huge_active(struct page
*page
)
1242 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1243 return PageHead(page
) && PagePrivate(&page
[1]);
1246 /* never called for tail page */
1247 static void set_page_huge_active(struct page
*page
)
1249 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1250 SetPagePrivate(&page
[1]);
1253 static void clear_page_huge_active(struct page
*page
)
1255 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1256 ClearPagePrivate(&page
[1]);
1259 void free_huge_page(struct page
*page
)
1262 * Can't pass hstate in here because it is called from the
1263 * compound page destructor.
1265 struct hstate
*h
= page_hstate(page
);
1266 int nid
= page_to_nid(page
);
1267 struct hugepage_subpool
*spool
=
1268 (struct hugepage_subpool
*)page_private(page
);
1269 bool restore_reserve
;
1271 set_page_private(page
, 0);
1272 page
->mapping
= NULL
;
1273 VM_BUG_ON_PAGE(page_count(page
), page
);
1274 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1275 restore_reserve
= PagePrivate(page
);
1276 ClearPagePrivate(page
);
1279 * A return code of zero implies that the subpool will be under its
1280 * minimum size if the reservation is not restored after page is free.
1281 * Therefore, force restore_reserve operation.
1283 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1284 restore_reserve
= true;
1286 spin_lock(&hugetlb_lock
);
1287 clear_page_huge_active(page
);
1288 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1289 pages_per_huge_page(h
), page
);
1290 if (restore_reserve
)
1291 h
->resv_huge_pages
++;
1293 if (h
->surplus_huge_pages_node
[nid
]) {
1294 /* remove the page from active list */
1295 list_del(&page
->lru
);
1296 update_and_free_page(h
, page
);
1297 h
->surplus_huge_pages
--;
1298 h
->surplus_huge_pages_node
[nid
]--;
1300 arch_clear_hugepage_flags(page
);
1301 enqueue_huge_page(h
, page
);
1303 spin_unlock(&hugetlb_lock
);
1306 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1308 INIT_LIST_HEAD(&page
->lru
);
1309 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1310 spin_lock(&hugetlb_lock
);
1311 set_hugetlb_cgroup(page
, NULL
);
1313 h
->nr_huge_pages_node
[nid
]++;
1314 spin_unlock(&hugetlb_lock
);
1315 put_page(page
); /* free it into the hugepage allocator */
1318 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1321 int nr_pages
= 1 << order
;
1322 struct page
*p
= page
+ 1;
1324 /* we rely on prep_new_huge_page to set the destructor */
1325 set_compound_order(page
, order
);
1326 __ClearPageReserved(page
);
1327 __SetPageHead(page
);
1328 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1330 * For gigantic hugepages allocated through bootmem at
1331 * boot, it's safer to be consistent with the not-gigantic
1332 * hugepages and clear the PG_reserved bit from all tail pages
1333 * too. Otherwse drivers using get_user_pages() to access tail
1334 * pages may get the reference counting wrong if they see
1335 * PG_reserved set on a tail page (despite the head page not
1336 * having PG_reserved set). Enforcing this consistency between
1337 * head and tail pages allows drivers to optimize away a check
1338 * on the head page when they need know if put_page() is needed
1339 * after get_user_pages().
1341 __ClearPageReserved(p
);
1342 set_page_count(p
, 0);
1343 set_compound_head(p
, page
);
1345 atomic_set(compound_mapcount_ptr(page
), -1);
1349 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1350 * transparent huge pages. See the PageTransHuge() documentation for more
1353 int PageHuge(struct page
*page
)
1355 if (!PageCompound(page
))
1358 page
= compound_head(page
);
1359 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1361 EXPORT_SYMBOL_GPL(PageHuge
);
1364 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1365 * normal or transparent huge pages.
1367 int PageHeadHuge(struct page
*page_head
)
1369 if (!PageHead(page_head
))
1372 return get_compound_page_dtor(page_head
) == free_huge_page
;
1375 pgoff_t
__basepage_index(struct page
*page
)
1377 struct page
*page_head
= compound_head(page
);
1378 pgoff_t index
= page_index(page_head
);
1379 unsigned long compound_idx
;
1381 if (!PageHuge(page_head
))
1382 return page_index(page
);
1384 if (compound_order(page_head
) >= MAX_ORDER
)
1385 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1387 compound_idx
= page
- page_head
;
1389 return (index
<< compound_order(page_head
)) + compound_idx
;
1392 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
1396 page
= __alloc_pages_node(nid
,
1397 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1398 __GFP_REPEAT
|__GFP_NOWARN
,
1399 huge_page_order(h
));
1401 prep_new_huge_page(h
, page
, nid
);
1407 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1413 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1414 page
= alloc_fresh_huge_page_node(h
, node
);
1422 count_vm_event(HTLB_BUDDY_PGALLOC
);
1424 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1430 * Free huge page from pool from next node to free.
1431 * Attempt to keep persistent huge pages more or less
1432 * balanced over allowed nodes.
1433 * Called with hugetlb_lock locked.
1435 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1441 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1443 * If we're returning unused surplus pages, only examine
1444 * nodes with surplus pages.
1446 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1447 !list_empty(&h
->hugepage_freelists
[node
])) {
1449 list_entry(h
->hugepage_freelists
[node
].next
,
1451 list_del(&page
->lru
);
1452 h
->free_huge_pages
--;
1453 h
->free_huge_pages_node
[node
]--;
1455 h
->surplus_huge_pages
--;
1456 h
->surplus_huge_pages_node
[node
]--;
1458 update_and_free_page(h
, page
);
1468 * Dissolve a given free hugepage into free buddy pages. This function does
1469 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1470 * number of free hugepages would be reduced below the number of reserved
1473 int dissolve_free_huge_page(struct page
*page
)
1477 spin_lock(&hugetlb_lock
);
1478 if (PageHuge(page
) && !page_count(page
)) {
1479 struct page
*head
= compound_head(page
);
1480 struct hstate
*h
= page_hstate(head
);
1481 int nid
= page_to_nid(head
);
1482 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0) {
1487 * Move PageHWPoison flag from head page to the raw error page,
1488 * which makes any subpages rather than the error page reusable.
1490 if (PageHWPoison(head
) && page
!= head
) {
1491 SetPageHWPoison(page
);
1492 ClearPageHWPoison(head
);
1494 list_del(&head
->lru
);
1495 h
->free_huge_pages
--;
1496 h
->free_huge_pages_node
[nid
]--;
1497 h
->max_huge_pages
--;
1498 update_and_free_page(h
, head
);
1501 spin_unlock(&hugetlb_lock
);
1506 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1507 * make specified memory blocks removable from the system.
1508 * Note that this will dissolve a free gigantic hugepage completely, if any
1509 * part of it lies within the given range.
1510 * Also note that if dissolve_free_huge_page() returns with an error, all
1511 * free hugepages that were dissolved before that error are lost.
1513 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1519 if (!hugepages_supported())
1522 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1523 page
= pfn_to_page(pfn
);
1524 if (PageHuge(page
) && !page_count(page
)) {
1525 rc
= dissolve_free_huge_page(page
);
1535 * There are 3 ways this can get called:
1536 * 1. With vma+addr: we use the VMA's memory policy
1537 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1538 * page from any node, and let the buddy allocator itself figure
1540 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1541 * strictly from 'nid'
1543 static struct page
*__hugetlb_alloc_buddy_huge_page(struct hstate
*h
,
1544 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1546 int order
= huge_page_order(h
);
1547 gfp_t gfp
= htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_REPEAT
|__GFP_NOWARN
;
1548 unsigned int cpuset_mems_cookie
;
1551 * We need a VMA to get a memory policy. If we do not
1552 * have one, we use the 'nid' argument.
1554 * The mempolicy stuff below has some non-inlined bits
1555 * and calls ->vm_ops. That makes it hard to optimize at
1556 * compile-time, even when NUMA is off and it does
1557 * nothing. This helps the compiler optimize it out.
1559 if (!IS_ENABLED(CONFIG_NUMA
) || !vma
) {
1561 * If a specific node is requested, make sure to
1562 * get memory from there, but only when a node
1563 * is explicitly specified.
1565 if (nid
!= NUMA_NO_NODE
)
1566 gfp
|= __GFP_THISNODE
;
1568 * Make sure to call something that can handle
1571 return alloc_pages_node(nid
, gfp
, order
);
1575 * OK, so we have a VMA. Fetch the mempolicy and try to
1576 * allocate a huge page with it. We will only reach this
1577 * when CONFIG_NUMA=y.
1581 struct mempolicy
*mpol
;
1583 nodemask_t
*nodemask
;
1585 cpuset_mems_cookie
= read_mems_allowed_begin();
1586 nid
= huge_node(vma
, addr
, gfp
, &mpol
, &nodemask
);
1587 mpol_cond_put(mpol
);
1588 page
= __alloc_pages_nodemask(gfp
, order
, nid
, nodemask
);
1591 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1597 * There are two ways to allocate a huge page:
1598 * 1. When you have a VMA and an address (like a fault)
1599 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1601 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1602 * this case which signifies that the allocation should be done with
1603 * respect for the VMA's memory policy.
1605 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1606 * implies that memory policies will not be taken in to account.
1608 static struct page
*__alloc_buddy_huge_page(struct hstate
*h
,
1609 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1614 if (hstate_is_gigantic(h
))
1618 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1619 * This makes sure the caller is picking _one_ of the modes with which
1620 * we can call this function, not both.
1622 if (vma
|| (addr
!= -1)) {
1623 VM_WARN_ON_ONCE(addr
== -1);
1624 VM_WARN_ON_ONCE(nid
!= NUMA_NO_NODE
);
1627 * Assume we will successfully allocate the surplus page to
1628 * prevent racing processes from causing the surplus to exceed
1631 * This however introduces a different race, where a process B
1632 * tries to grow the static hugepage pool while alloc_pages() is
1633 * called by process A. B will only examine the per-node
1634 * counters in determining if surplus huge pages can be
1635 * converted to normal huge pages in adjust_pool_surplus(). A
1636 * won't be able to increment the per-node counter, until the
1637 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1638 * no more huge pages can be converted from surplus to normal
1639 * state (and doesn't try to convert again). Thus, we have a
1640 * case where a surplus huge page exists, the pool is grown, and
1641 * the surplus huge page still exists after, even though it
1642 * should just have been converted to a normal huge page. This
1643 * does not leak memory, though, as the hugepage will be freed
1644 * once it is out of use. It also does not allow the counters to
1645 * go out of whack in adjust_pool_surplus() as we don't modify
1646 * the node values until we've gotten the hugepage and only the
1647 * per-node value is checked there.
1649 spin_lock(&hugetlb_lock
);
1650 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1651 spin_unlock(&hugetlb_lock
);
1655 h
->surplus_huge_pages
++;
1657 spin_unlock(&hugetlb_lock
);
1659 page
= __hugetlb_alloc_buddy_huge_page(h
, vma
, addr
, nid
);
1661 spin_lock(&hugetlb_lock
);
1663 INIT_LIST_HEAD(&page
->lru
);
1664 r_nid
= page_to_nid(page
);
1665 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1666 set_hugetlb_cgroup(page
, NULL
);
1668 * We incremented the global counters already
1670 h
->nr_huge_pages_node
[r_nid
]++;
1671 h
->surplus_huge_pages_node
[r_nid
]++;
1672 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1675 h
->surplus_huge_pages
--;
1676 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1678 spin_unlock(&hugetlb_lock
);
1684 * Allocate a huge page from 'nid'. Note, 'nid' may be
1685 * NUMA_NO_NODE, which means that it may be allocated
1689 struct page
*__alloc_buddy_huge_page_no_mpol(struct hstate
*h
, int nid
)
1691 unsigned long addr
= -1;
1693 return __alloc_buddy_huge_page(h
, NULL
, addr
, nid
);
1697 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1700 struct page
*__alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1701 struct vm_area_struct
*vma
, unsigned long addr
)
1703 return __alloc_buddy_huge_page(h
, vma
, addr
, NUMA_NO_NODE
);
1707 * This allocation function is useful in the context where vma is irrelevant.
1708 * E.g. soft-offlining uses this function because it only cares physical
1709 * address of error page.
1711 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1713 struct page
*page
= NULL
;
1715 spin_lock(&hugetlb_lock
);
1716 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1717 page
= dequeue_huge_page_node(h
, nid
);
1718 spin_unlock(&hugetlb_lock
);
1721 page
= __alloc_buddy_huge_page_no_mpol(h
, nid
);
1726 struct page
*alloc_huge_page_nodemask(struct hstate
*h
, const nodemask_t
*nmask
)
1728 struct page
*page
= NULL
;
1731 spin_lock(&hugetlb_lock
);
1732 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0) {
1733 for_each_node_mask(node
, *nmask
) {
1734 page
= dequeue_huge_page_node_exact(h
, node
);
1739 spin_unlock(&hugetlb_lock
);
1743 /* No reservations, try to overcommit */
1744 for_each_node_mask(node
, *nmask
) {
1745 page
= __alloc_buddy_huge_page_no_mpol(h
, node
);
1754 * Increase the hugetlb pool such that it can accommodate a reservation
1757 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1759 struct list_head surplus_list
;
1760 struct page
*page
, *tmp
;
1762 int needed
, allocated
;
1763 bool alloc_ok
= true;
1765 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1767 h
->resv_huge_pages
+= delta
;
1772 INIT_LIST_HEAD(&surplus_list
);
1776 spin_unlock(&hugetlb_lock
);
1777 for (i
= 0; i
< needed
; i
++) {
1778 page
= __alloc_buddy_huge_page_no_mpol(h
, NUMA_NO_NODE
);
1783 list_add(&page
->lru
, &surplus_list
);
1788 * After retaking hugetlb_lock, we need to recalculate 'needed'
1789 * because either resv_huge_pages or free_huge_pages may have changed.
1791 spin_lock(&hugetlb_lock
);
1792 needed
= (h
->resv_huge_pages
+ delta
) -
1793 (h
->free_huge_pages
+ allocated
);
1798 * We were not able to allocate enough pages to
1799 * satisfy the entire reservation so we free what
1800 * we've allocated so far.
1805 * The surplus_list now contains _at_least_ the number of extra pages
1806 * needed to accommodate the reservation. Add the appropriate number
1807 * of pages to the hugetlb pool and free the extras back to the buddy
1808 * allocator. Commit the entire reservation here to prevent another
1809 * process from stealing the pages as they are added to the pool but
1810 * before they are reserved.
1812 needed
+= allocated
;
1813 h
->resv_huge_pages
+= delta
;
1816 /* Free the needed pages to the hugetlb pool */
1817 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1821 * This page is now managed by the hugetlb allocator and has
1822 * no users -- drop the buddy allocator's reference.
1824 put_page_testzero(page
);
1825 VM_BUG_ON_PAGE(page_count(page
), page
);
1826 enqueue_huge_page(h
, page
);
1829 spin_unlock(&hugetlb_lock
);
1831 /* Free unnecessary surplus pages to the buddy allocator */
1832 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1834 spin_lock(&hugetlb_lock
);
1840 * This routine has two main purposes:
1841 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1842 * in unused_resv_pages. This corresponds to the prior adjustments made
1843 * to the associated reservation map.
1844 * 2) Free any unused surplus pages that may have been allocated to satisfy
1845 * the reservation. As many as unused_resv_pages may be freed.
1847 * Called with hugetlb_lock held. However, the lock could be dropped (and
1848 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1849 * we must make sure nobody else can claim pages we are in the process of
1850 * freeing. Do this by ensuring resv_huge_page always is greater than the
1851 * number of huge pages we plan to free when dropping the lock.
1853 static void return_unused_surplus_pages(struct hstate
*h
,
1854 unsigned long unused_resv_pages
)
1856 unsigned long nr_pages
;
1858 /* Cannot return gigantic pages currently */
1859 if (hstate_is_gigantic(h
))
1863 * Part (or even all) of the reservation could have been backed
1864 * by pre-allocated pages. Only free surplus pages.
1866 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1869 * We want to release as many surplus pages as possible, spread
1870 * evenly across all nodes with memory. Iterate across these nodes
1871 * until we can no longer free unreserved surplus pages. This occurs
1872 * when the nodes with surplus pages have no free pages.
1873 * free_pool_huge_page() will balance the the freed pages across the
1874 * on-line nodes with memory and will handle the hstate accounting.
1876 * Note that we decrement resv_huge_pages as we free the pages. If
1877 * we drop the lock, resv_huge_pages will still be sufficiently large
1878 * to cover subsequent pages we may free.
1880 while (nr_pages
--) {
1881 h
->resv_huge_pages
--;
1882 unused_resv_pages
--;
1883 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1885 cond_resched_lock(&hugetlb_lock
);
1889 /* Fully uncommit the reservation */
1890 h
->resv_huge_pages
-= unused_resv_pages
;
1895 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1896 * are used by the huge page allocation routines to manage reservations.
1898 * vma_needs_reservation is called to determine if the huge page at addr
1899 * within the vma has an associated reservation. If a reservation is
1900 * needed, the value 1 is returned. The caller is then responsible for
1901 * managing the global reservation and subpool usage counts. After
1902 * the huge page has been allocated, vma_commit_reservation is called
1903 * to add the page to the reservation map. If the page allocation fails,
1904 * the reservation must be ended instead of committed. vma_end_reservation
1905 * is called in such cases.
1907 * In the normal case, vma_commit_reservation returns the same value
1908 * as the preceding vma_needs_reservation call. The only time this
1909 * is not the case is if a reserve map was changed between calls. It
1910 * is the responsibility of the caller to notice the difference and
1911 * take appropriate action.
1913 * vma_add_reservation is used in error paths where a reservation must
1914 * be restored when a newly allocated huge page must be freed. It is
1915 * to be called after calling vma_needs_reservation to determine if a
1916 * reservation exists.
1918 enum vma_resv_mode
{
1924 static long __vma_reservation_common(struct hstate
*h
,
1925 struct vm_area_struct
*vma
, unsigned long addr
,
1926 enum vma_resv_mode mode
)
1928 struct resv_map
*resv
;
1932 resv
= vma_resv_map(vma
);
1936 idx
= vma_hugecache_offset(h
, vma
, addr
);
1938 case VMA_NEEDS_RESV
:
1939 ret
= region_chg(resv
, idx
, idx
+ 1);
1941 case VMA_COMMIT_RESV
:
1942 ret
= region_add(resv
, idx
, idx
+ 1);
1945 region_abort(resv
, idx
, idx
+ 1);
1949 if (vma
->vm_flags
& VM_MAYSHARE
)
1950 ret
= region_add(resv
, idx
, idx
+ 1);
1952 region_abort(resv
, idx
, idx
+ 1);
1953 ret
= region_del(resv
, idx
, idx
+ 1);
1960 if (vma
->vm_flags
& VM_MAYSHARE
)
1962 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
1964 * In most cases, reserves always exist for private mappings.
1965 * However, a file associated with mapping could have been
1966 * hole punched or truncated after reserves were consumed.
1967 * As subsequent fault on such a range will not use reserves.
1968 * Subtle - The reserve map for private mappings has the
1969 * opposite meaning than that of shared mappings. If NO
1970 * entry is in the reserve map, it means a reservation exists.
1971 * If an entry exists in the reserve map, it means the
1972 * reservation has already been consumed. As a result, the
1973 * return value of this routine is the opposite of the
1974 * value returned from reserve map manipulation routines above.
1982 return ret
< 0 ? ret
: 0;
1985 static long vma_needs_reservation(struct hstate
*h
,
1986 struct vm_area_struct
*vma
, unsigned long addr
)
1988 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
1991 static long vma_commit_reservation(struct hstate
*h
,
1992 struct vm_area_struct
*vma
, unsigned long addr
)
1994 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
1997 static void vma_end_reservation(struct hstate
*h
,
1998 struct vm_area_struct
*vma
, unsigned long addr
)
2000 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
2003 static long vma_add_reservation(struct hstate
*h
,
2004 struct vm_area_struct
*vma
, unsigned long addr
)
2006 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
2010 * This routine is called to restore a reservation on error paths. In the
2011 * specific error paths, a huge page was allocated (via alloc_huge_page)
2012 * and is about to be freed. If a reservation for the page existed,
2013 * alloc_huge_page would have consumed the reservation and set PagePrivate
2014 * in the newly allocated page. When the page is freed via free_huge_page,
2015 * the global reservation count will be incremented if PagePrivate is set.
2016 * However, free_huge_page can not adjust the reserve map. Adjust the
2017 * reserve map here to be consistent with global reserve count adjustments
2018 * to be made by free_huge_page.
2020 static void restore_reserve_on_error(struct hstate
*h
,
2021 struct vm_area_struct
*vma
, unsigned long address
,
2024 if (unlikely(PagePrivate(page
))) {
2025 long rc
= vma_needs_reservation(h
, vma
, address
);
2027 if (unlikely(rc
< 0)) {
2029 * Rare out of memory condition in reserve map
2030 * manipulation. Clear PagePrivate so that
2031 * global reserve count will not be incremented
2032 * by free_huge_page. This will make it appear
2033 * as though the reservation for this page was
2034 * consumed. This may prevent the task from
2035 * faulting in the page at a later time. This
2036 * is better than inconsistent global huge page
2037 * accounting of reserve counts.
2039 ClearPagePrivate(page
);
2041 rc
= vma_add_reservation(h
, vma
, address
);
2042 if (unlikely(rc
< 0))
2044 * See above comment about rare out of
2047 ClearPagePrivate(page
);
2049 vma_end_reservation(h
, vma
, address
);
2053 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2054 unsigned long addr
, int avoid_reserve
)
2056 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2057 struct hstate
*h
= hstate_vma(vma
);
2059 long map_chg
, map_commit
;
2062 struct hugetlb_cgroup
*h_cg
;
2064 idx
= hstate_index(h
);
2066 * Examine the region/reserve map to determine if the process
2067 * has a reservation for the page to be allocated. A return
2068 * code of zero indicates a reservation exists (no change).
2070 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2072 return ERR_PTR(-ENOMEM
);
2075 * Processes that did not create the mapping will have no
2076 * reserves as indicated by the region/reserve map. Check
2077 * that the allocation will not exceed the subpool limit.
2078 * Allocations for MAP_NORESERVE mappings also need to be
2079 * checked against any subpool limit.
2081 if (map_chg
|| avoid_reserve
) {
2082 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2084 vma_end_reservation(h
, vma
, addr
);
2085 return ERR_PTR(-ENOSPC
);
2089 * Even though there was no reservation in the region/reserve
2090 * map, there could be reservations associated with the
2091 * subpool that can be used. This would be indicated if the
2092 * return value of hugepage_subpool_get_pages() is zero.
2093 * However, if avoid_reserve is specified we still avoid even
2094 * the subpool reservations.
2100 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2102 goto out_subpool_put
;
2104 spin_lock(&hugetlb_lock
);
2106 * glb_chg is passed to indicate whether or not a page must be taken
2107 * from the global free pool (global change). gbl_chg == 0 indicates
2108 * a reservation exists for the allocation.
2110 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2112 spin_unlock(&hugetlb_lock
);
2113 page
= __alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2115 goto out_uncharge_cgroup
;
2116 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2117 SetPagePrivate(page
);
2118 h
->resv_huge_pages
--;
2120 spin_lock(&hugetlb_lock
);
2121 list_move(&page
->lru
, &h
->hugepage_activelist
);
2124 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2125 spin_unlock(&hugetlb_lock
);
2127 set_page_private(page
, (unsigned long)spool
);
2129 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2130 if (unlikely(map_chg
> map_commit
)) {
2132 * The page was added to the reservation map between
2133 * vma_needs_reservation and vma_commit_reservation.
2134 * This indicates a race with hugetlb_reserve_pages.
2135 * Adjust for the subpool count incremented above AND
2136 * in hugetlb_reserve_pages for the same page. Also,
2137 * the reservation count added in hugetlb_reserve_pages
2138 * no longer applies.
2142 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2143 hugetlb_acct_memory(h
, -rsv_adjust
);
2147 out_uncharge_cgroup
:
2148 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2150 if (map_chg
|| avoid_reserve
)
2151 hugepage_subpool_put_pages(spool
, 1);
2152 vma_end_reservation(h
, vma
, addr
);
2153 return ERR_PTR(-ENOSPC
);
2157 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2158 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2159 * where no ERR_VALUE is expected to be returned.
2161 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
2162 unsigned long addr
, int avoid_reserve
)
2164 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
2170 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
2172 struct huge_bootmem_page
*m
;
2175 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2178 addr
= memblock_virt_alloc_try_nid_nopanic(
2179 huge_page_size(h
), huge_page_size(h
),
2180 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
2183 * Use the beginning of the huge page to store the
2184 * huge_bootmem_page struct (until gather_bootmem
2185 * puts them into the mem_map).
2194 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2195 /* Put them into a private list first because mem_map is not up yet */
2196 list_add(&m
->list
, &huge_boot_pages
);
2201 static void __init
prep_compound_huge_page(struct page
*page
,
2204 if (unlikely(order
> (MAX_ORDER
- 1)))
2205 prep_compound_gigantic_page(page
, order
);
2207 prep_compound_page(page
, order
);
2210 /* Put bootmem huge pages into the standard lists after mem_map is up */
2211 static void __init
gather_bootmem_prealloc(void)
2213 struct huge_bootmem_page
*m
;
2215 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2216 struct hstate
*h
= m
->hstate
;
2219 #ifdef CONFIG_HIGHMEM
2220 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
2221 memblock_free_late(__pa(m
),
2222 sizeof(struct huge_bootmem_page
));
2224 page
= virt_to_page(m
);
2226 WARN_ON(page_count(page
) != 1);
2227 prep_compound_huge_page(page
, h
->order
);
2228 WARN_ON(PageReserved(page
));
2229 prep_new_huge_page(h
, page
, page_to_nid(page
));
2231 * If we had gigantic hugepages allocated at boot time, we need
2232 * to restore the 'stolen' pages to totalram_pages in order to
2233 * fix confusing memory reports from free(1) and another
2234 * side-effects, like CommitLimit going negative.
2236 if (hstate_is_gigantic(h
))
2237 adjust_managed_page_count(page
, 1 << h
->order
);
2241 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2245 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2246 if (hstate_is_gigantic(h
)) {
2247 if (!alloc_bootmem_huge_page(h
))
2249 } else if (!alloc_fresh_huge_page(h
,
2250 &node_states
[N_MEMORY
]))
2253 if (i
< h
->max_huge_pages
) {
2256 memfmt(buf
, huge_page_size(h
)),
2257 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2258 h
->max_huge_pages
, buf
, i
);
2259 h
->max_huge_pages
= i
;
2263 static void __init
hugetlb_init_hstates(void)
2267 for_each_hstate(h
) {
2268 if (minimum_order
> huge_page_order(h
))
2269 minimum_order
= huge_page_order(h
);
2271 /* oversize hugepages were init'ed in early boot */
2272 if (!hstate_is_gigantic(h
))
2273 hugetlb_hstate_alloc_pages(h
);
2275 VM_BUG_ON(minimum_order
== UINT_MAX
);
2278 static void __init
report_hugepages(void)
2282 for_each_hstate(h
) {
2284 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2285 memfmt(buf
, huge_page_size(h
)),
2286 h
->free_huge_pages
);
2290 #ifdef CONFIG_HIGHMEM
2291 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2292 nodemask_t
*nodes_allowed
)
2296 if (hstate_is_gigantic(h
))
2299 for_each_node_mask(i
, *nodes_allowed
) {
2300 struct page
*page
, *next
;
2301 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2302 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2303 if (count
>= h
->nr_huge_pages
)
2305 if (PageHighMem(page
))
2307 list_del(&page
->lru
);
2308 update_and_free_page(h
, page
);
2309 h
->free_huge_pages
--;
2310 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2315 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2316 nodemask_t
*nodes_allowed
)
2322 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2323 * balanced by operating on them in a round-robin fashion.
2324 * Returns 1 if an adjustment was made.
2326 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2331 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2334 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2335 if (h
->surplus_huge_pages_node
[node
])
2339 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2340 if (h
->surplus_huge_pages_node
[node
] <
2341 h
->nr_huge_pages_node
[node
])
2348 h
->surplus_huge_pages
+= delta
;
2349 h
->surplus_huge_pages_node
[node
] += delta
;
2353 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2354 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
2355 nodemask_t
*nodes_allowed
)
2357 unsigned long min_count
, ret
;
2359 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
2360 return h
->max_huge_pages
;
2363 * Increase the pool size
2364 * First take pages out of surplus state. Then make up the
2365 * remaining difference by allocating fresh huge pages.
2367 * We might race with __alloc_buddy_huge_page() here and be unable
2368 * to convert a surplus huge page to a normal huge page. That is
2369 * not critical, though, it just means the overall size of the
2370 * pool might be one hugepage larger than it needs to be, but
2371 * within all the constraints specified by the sysctls.
2373 spin_lock(&hugetlb_lock
);
2374 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2375 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2379 while (count
> persistent_huge_pages(h
)) {
2381 * If this allocation races such that we no longer need the
2382 * page, free_huge_page will handle it by freeing the page
2383 * and reducing the surplus.
2385 spin_unlock(&hugetlb_lock
);
2387 /* yield cpu to avoid soft lockup */
2390 if (hstate_is_gigantic(h
))
2391 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
2393 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
2394 spin_lock(&hugetlb_lock
);
2398 /* Bail for signals. Probably ctrl-c from user */
2399 if (signal_pending(current
))
2404 * Decrease the pool size
2405 * First return free pages to the buddy allocator (being careful
2406 * to keep enough around to satisfy reservations). Then place
2407 * pages into surplus state as needed so the pool will shrink
2408 * to the desired size as pages become free.
2410 * By placing pages into the surplus state independent of the
2411 * overcommit value, we are allowing the surplus pool size to
2412 * exceed overcommit. There are few sane options here. Since
2413 * __alloc_buddy_huge_page() is checking the global counter,
2414 * though, we'll note that we're not allowed to exceed surplus
2415 * and won't grow the pool anywhere else. Not until one of the
2416 * sysctls are changed, or the surplus pages go out of use.
2418 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2419 min_count
= max(count
, min_count
);
2420 try_to_free_low(h
, min_count
, nodes_allowed
);
2421 while (min_count
< persistent_huge_pages(h
)) {
2422 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2424 cond_resched_lock(&hugetlb_lock
);
2426 while (count
< persistent_huge_pages(h
)) {
2427 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2431 ret
= persistent_huge_pages(h
);
2432 spin_unlock(&hugetlb_lock
);
2436 #define HSTATE_ATTR_RO(_name) \
2437 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2439 #define HSTATE_ATTR(_name) \
2440 static struct kobj_attribute _name##_attr = \
2441 __ATTR(_name, 0644, _name##_show, _name##_store)
2443 static struct kobject
*hugepages_kobj
;
2444 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2446 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2448 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2452 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2453 if (hstate_kobjs
[i
] == kobj
) {
2455 *nidp
= NUMA_NO_NODE
;
2459 return kobj_to_node_hstate(kobj
, nidp
);
2462 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2463 struct kobj_attribute
*attr
, char *buf
)
2466 unsigned long nr_huge_pages
;
2469 h
= kobj_to_hstate(kobj
, &nid
);
2470 if (nid
== NUMA_NO_NODE
)
2471 nr_huge_pages
= h
->nr_huge_pages
;
2473 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2475 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2478 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2479 struct hstate
*h
, int nid
,
2480 unsigned long count
, size_t len
)
2483 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
2485 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
2490 if (nid
== NUMA_NO_NODE
) {
2492 * global hstate attribute
2494 if (!(obey_mempolicy
&&
2495 init_nodemask_of_mempolicy(nodes_allowed
))) {
2496 NODEMASK_FREE(nodes_allowed
);
2497 nodes_allowed
= &node_states
[N_MEMORY
];
2499 } else if (nodes_allowed
) {
2501 * per node hstate attribute: adjust count to global,
2502 * but restrict alloc/free to the specified node.
2504 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2505 init_nodemask_of_node(nodes_allowed
, nid
);
2507 nodes_allowed
= &node_states
[N_MEMORY
];
2509 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
2511 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2512 NODEMASK_FREE(nodes_allowed
);
2516 NODEMASK_FREE(nodes_allowed
);
2520 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2521 struct kobject
*kobj
, const char *buf
,
2525 unsigned long count
;
2529 err
= kstrtoul(buf
, 10, &count
);
2533 h
= kobj_to_hstate(kobj
, &nid
);
2534 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2537 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2538 struct kobj_attribute
*attr
, char *buf
)
2540 return nr_hugepages_show_common(kobj
, attr
, buf
);
2543 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2544 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2546 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2548 HSTATE_ATTR(nr_hugepages
);
2553 * hstate attribute for optionally mempolicy-based constraint on persistent
2554 * huge page alloc/free.
2556 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2557 struct kobj_attribute
*attr
, char *buf
)
2559 return nr_hugepages_show_common(kobj
, attr
, buf
);
2562 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2563 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2565 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2567 HSTATE_ATTR(nr_hugepages_mempolicy
);
2571 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2572 struct kobj_attribute
*attr
, char *buf
)
2574 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2575 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2578 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2579 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2582 unsigned long input
;
2583 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2585 if (hstate_is_gigantic(h
))
2588 err
= kstrtoul(buf
, 10, &input
);
2592 spin_lock(&hugetlb_lock
);
2593 h
->nr_overcommit_huge_pages
= input
;
2594 spin_unlock(&hugetlb_lock
);
2598 HSTATE_ATTR(nr_overcommit_hugepages
);
2600 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2601 struct kobj_attribute
*attr
, char *buf
)
2604 unsigned long free_huge_pages
;
2607 h
= kobj_to_hstate(kobj
, &nid
);
2608 if (nid
== NUMA_NO_NODE
)
2609 free_huge_pages
= h
->free_huge_pages
;
2611 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2613 return sprintf(buf
, "%lu\n", free_huge_pages
);
2615 HSTATE_ATTR_RO(free_hugepages
);
2617 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2618 struct kobj_attribute
*attr
, char *buf
)
2620 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2621 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2623 HSTATE_ATTR_RO(resv_hugepages
);
2625 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2626 struct kobj_attribute
*attr
, char *buf
)
2629 unsigned long surplus_huge_pages
;
2632 h
= kobj_to_hstate(kobj
, &nid
);
2633 if (nid
== NUMA_NO_NODE
)
2634 surplus_huge_pages
= h
->surplus_huge_pages
;
2636 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2638 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2640 HSTATE_ATTR_RO(surplus_hugepages
);
2642 static struct attribute
*hstate_attrs
[] = {
2643 &nr_hugepages_attr
.attr
,
2644 &nr_overcommit_hugepages_attr
.attr
,
2645 &free_hugepages_attr
.attr
,
2646 &resv_hugepages_attr
.attr
,
2647 &surplus_hugepages_attr
.attr
,
2649 &nr_hugepages_mempolicy_attr
.attr
,
2654 static struct attribute_group hstate_attr_group
= {
2655 .attrs
= hstate_attrs
,
2658 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2659 struct kobject
**hstate_kobjs
,
2660 struct attribute_group
*hstate_attr_group
)
2663 int hi
= hstate_index(h
);
2665 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2666 if (!hstate_kobjs
[hi
])
2669 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2671 kobject_put(hstate_kobjs
[hi
]);
2676 static void __init
hugetlb_sysfs_init(void)
2681 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2682 if (!hugepages_kobj
)
2685 for_each_hstate(h
) {
2686 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2687 hstate_kobjs
, &hstate_attr_group
);
2689 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2696 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2697 * with node devices in node_devices[] using a parallel array. The array
2698 * index of a node device or _hstate == node id.
2699 * This is here to avoid any static dependency of the node device driver, in
2700 * the base kernel, on the hugetlb module.
2702 struct node_hstate
{
2703 struct kobject
*hugepages_kobj
;
2704 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2706 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2709 * A subset of global hstate attributes for node devices
2711 static struct attribute
*per_node_hstate_attrs
[] = {
2712 &nr_hugepages_attr
.attr
,
2713 &free_hugepages_attr
.attr
,
2714 &surplus_hugepages_attr
.attr
,
2718 static struct attribute_group per_node_hstate_attr_group
= {
2719 .attrs
= per_node_hstate_attrs
,
2723 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2724 * Returns node id via non-NULL nidp.
2726 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2730 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2731 struct node_hstate
*nhs
= &node_hstates
[nid
];
2733 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2734 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2746 * Unregister hstate attributes from a single node device.
2747 * No-op if no hstate attributes attached.
2749 static void hugetlb_unregister_node(struct node
*node
)
2752 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2754 if (!nhs
->hugepages_kobj
)
2755 return; /* no hstate attributes */
2757 for_each_hstate(h
) {
2758 int idx
= hstate_index(h
);
2759 if (nhs
->hstate_kobjs
[idx
]) {
2760 kobject_put(nhs
->hstate_kobjs
[idx
]);
2761 nhs
->hstate_kobjs
[idx
] = NULL
;
2765 kobject_put(nhs
->hugepages_kobj
);
2766 nhs
->hugepages_kobj
= NULL
;
2771 * Register hstate attributes for a single node device.
2772 * No-op if attributes already registered.
2774 static void hugetlb_register_node(struct node
*node
)
2777 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2780 if (nhs
->hugepages_kobj
)
2781 return; /* already allocated */
2783 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2785 if (!nhs
->hugepages_kobj
)
2788 for_each_hstate(h
) {
2789 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2791 &per_node_hstate_attr_group
);
2793 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2794 h
->name
, node
->dev
.id
);
2795 hugetlb_unregister_node(node
);
2802 * hugetlb init time: register hstate attributes for all registered node
2803 * devices of nodes that have memory. All on-line nodes should have
2804 * registered their associated device by this time.
2806 static void __init
hugetlb_register_all_nodes(void)
2810 for_each_node_state(nid
, N_MEMORY
) {
2811 struct node
*node
= node_devices
[nid
];
2812 if (node
->dev
.id
== nid
)
2813 hugetlb_register_node(node
);
2817 * Let the node device driver know we're here so it can
2818 * [un]register hstate attributes on node hotplug.
2820 register_hugetlbfs_with_node(hugetlb_register_node
,
2821 hugetlb_unregister_node
);
2823 #else /* !CONFIG_NUMA */
2825 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2833 static void hugetlb_register_all_nodes(void) { }
2837 static int __init
hugetlb_init(void)
2841 if (!hugepages_supported())
2844 if (!size_to_hstate(default_hstate_size
)) {
2845 if (default_hstate_size
!= 0) {
2846 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2847 default_hstate_size
, HPAGE_SIZE
);
2850 default_hstate_size
= HPAGE_SIZE
;
2851 if (!size_to_hstate(default_hstate_size
))
2852 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2854 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2855 if (default_hstate_max_huge_pages
) {
2856 if (!default_hstate
.max_huge_pages
)
2857 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2860 hugetlb_init_hstates();
2861 gather_bootmem_prealloc();
2864 hugetlb_sysfs_init();
2865 hugetlb_register_all_nodes();
2866 hugetlb_cgroup_file_init();
2869 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2871 num_fault_mutexes
= 1;
2873 hugetlb_fault_mutex_table
=
2874 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2875 BUG_ON(!hugetlb_fault_mutex_table
);
2877 for (i
= 0; i
< num_fault_mutexes
; i
++)
2878 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2881 subsys_initcall(hugetlb_init
);
2883 /* Should be called on processing a hugepagesz=... option */
2884 void __init
hugetlb_bad_size(void)
2886 parsed_valid_hugepagesz
= false;
2889 void __init
hugetlb_add_hstate(unsigned int order
)
2894 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2895 pr_warn("hugepagesz= specified twice, ignoring\n");
2898 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2900 h
= &hstates
[hugetlb_max_hstate
++];
2902 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2903 h
->nr_huge_pages
= 0;
2904 h
->free_huge_pages
= 0;
2905 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2906 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2907 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2908 h
->next_nid_to_alloc
= first_memory_node
;
2909 h
->next_nid_to_free
= first_memory_node
;
2910 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2911 huge_page_size(h
)/1024);
2916 static int __init
hugetlb_nrpages_setup(char *s
)
2919 static unsigned long *last_mhp
;
2921 if (!parsed_valid_hugepagesz
) {
2922 pr_warn("hugepages = %s preceded by "
2923 "an unsupported hugepagesz, ignoring\n", s
);
2924 parsed_valid_hugepagesz
= true;
2928 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2929 * so this hugepages= parameter goes to the "default hstate".
2931 else if (!hugetlb_max_hstate
)
2932 mhp
= &default_hstate_max_huge_pages
;
2934 mhp
= &parsed_hstate
->max_huge_pages
;
2936 if (mhp
== last_mhp
) {
2937 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2941 if (sscanf(s
, "%lu", mhp
) <= 0)
2945 * Global state is always initialized later in hugetlb_init.
2946 * But we need to allocate >= MAX_ORDER hstates here early to still
2947 * use the bootmem allocator.
2949 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2950 hugetlb_hstate_alloc_pages(parsed_hstate
);
2956 __setup("hugepages=", hugetlb_nrpages_setup
);
2958 static int __init
hugetlb_default_setup(char *s
)
2960 default_hstate_size
= memparse(s
, &s
);
2963 __setup("default_hugepagesz=", hugetlb_default_setup
);
2965 static unsigned int cpuset_mems_nr(unsigned int *array
)
2968 unsigned int nr
= 0;
2970 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2976 #ifdef CONFIG_SYSCTL
2977 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2978 struct ctl_table
*table
, int write
,
2979 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2981 struct hstate
*h
= &default_hstate
;
2982 unsigned long tmp
= h
->max_huge_pages
;
2985 if (!hugepages_supported())
2989 table
->maxlen
= sizeof(unsigned long);
2990 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2995 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2996 NUMA_NO_NODE
, tmp
, *length
);
3001 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
3002 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3005 return hugetlb_sysctl_handler_common(false, table
, write
,
3006 buffer
, length
, ppos
);
3010 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
3011 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3013 return hugetlb_sysctl_handler_common(true, table
, write
,
3014 buffer
, length
, ppos
);
3016 #endif /* CONFIG_NUMA */
3018 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
3019 void __user
*buffer
,
3020 size_t *length
, loff_t
*ppos
)
3022 struct hstate
*h
= &default_hstate
;
3026 if (!hugepages_supported())
3029 tmp
= h
->nr_overcommit_huge_pages
;
3031 if (write
&& hstate_is_gigantic(h
))
3035 table
->maxlen
= sizeof(unsigned long);
3036 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3041 spin_lock(&hugetlb_lock
);
3042 h
->nr_overcommit_huge_pages
= tmp
;
3043 spin_unlock(&hugetlb_lock
);
3049 #endif /* CONFIG_SYSCTL */
3051 void hugetlb_report_meminfo(struct seq_file
*m
)
3053 struct hstate
*h
= &default_hstate
;
3054 if (!hugepages_supported())
3057 "HugePages_Total: %5lu\n"
3058 "HugePages_Free: %5lu\n"
3059 "HugePages_Rsvd: %5lu\n"
3060 "HugePages_Surp: %5lu\n"
3061 "Hugepagesize: %8lu kB\n",
3065 h
->surplus_huge_pages
,
3066 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3069 int hugetlb_report_node_meminfo(int nid
, char *buf
)
3071 struct hstate
*h
= &default_hstate
;
3072 if (!hugepages_supported())
3075 "Node %d HugePages_Total: %5u\n"
3076 "Node %d HugePages_Free: %5u\n"
3077 "Node %d HugePages_Surp: %5u\n",
3078 nid
, h
->nr_huge_pages_node
[nid
],
3079 nid
, h
->free_huge_pages_node
[nid
],
3080 nid
, h
->surplus_huge_pages_node
[nid
]);
3083 void hugetlb_show_meminfo(void)
3088 if (!hugepages_supported())
3091 for_each_node_state(nid
, N_MEMORY
)
3093 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3095 h
->nr_huge_pages_node
[nid
],
3096 h
->free_huge_pages_node
[nid
],
3097 h
->surplus_huge_pages_node
[nid
],
3098 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3101 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3103 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3104 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3107 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3108 unsigned long hugetlb_total_pages(void)
3111 unsigned long nr_total_pages
= 0;
3114 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3115 return nr_total_pages
;
3118 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3122 spin_lock(&hugetlb_lock
);
3124 * When cpuset is configured, it breaks the strict hugetlb page
3125 * reservation as the accounting is done on a global variable. Such
3126 * reservation is completely rubbish in the presence of cpuset because
3127 * the reservation is not checked against page availability for the
3128 * current cpuset. Application can still potentially OOM'ed by kernel
3129 * with lack of free htlb page in cpuset that the task is in.
3130 * Attempt to enforce strict accounting with cpuset is almost
3131 * impossible (or too ugly) because cpuset is too fluid that
3132 * task or memory node can be dynamically moved between cpusets.
3134 * The change of semantics for shared hugetlb mapping with cpuset is
3135 * undesirable. However, in order to preserve some of the semantics,
3136 * we fall back to check against current free page availability as
3137 * a best attempt and hopefully to minimize the impact of changing
3138 * semantics that cpuset has.
3141 if (gather_surplus_pages(h
, delta
) < 0)
3144 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
3145 return_unused_surplus_pages(h
, delta
);
3152 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3155 spin_unlock(&hugetlb_lock
);
3159 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3161 struct resv_map
*resv
= vma_resv_map(vma
);
3164 * This new VMA should share its siblings reservation map if present.
3165 * The VMA will only ever have a valid reservation map pointer where
3166 * it is being copied for another still existing VMA. As that VMA
3167 * has a reference to the reservation map it cannot disappear until
3168 * after this open call completes. It is therefore safe to take a
3169 * new reference here without additional locking.
3171 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3172 kref_get(&resv
->refs
);
3175 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3177 struct hstate
*h
= hstate_vma(vma
);
3178 struct resv_map
*resv
= vma_resv_map(vma
);
3179 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3180 unsigned long reserve
, start
, end
;
3183 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3186 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3187 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3189 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3191 kref_put(&resv
->refs
, resv_map_release
);
3195 * Decrement reserve counts. The global reserve count may be
3196 * adjusted if the subpool has a minimum size.
3198 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3199 hugetlb_acct_memory(h
, -gbl_reserve
);
3204 * We cannot handle pagefaults against hugetlb pages at all. They cause
3205 * handle_mm_fault() to try to instantiate regular-sized pages in the
3206 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3209 static int hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3215 const struct vm_operations_struct hugetlb_vm_ops
= {
3216 .fault
= hugetlb_vm_op_fault
,
3217 .open
= hugetlb_vm_op_open
,
3218 .close
= hugetlb_vm_op_close
,
3221 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3227 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3228 vma
->vm_page_prot
)));
3230 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3231 vma
->vm_page_prot
));
3233 entry
= pte_mkyoung(entry
);
3234 entry
= pte_mkhuge(entry
);
3235 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3240 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3241 unsigned long address
, pte_t
*ptep
)
3245 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3246 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3247 update_mmu_cache(vma
, address
, ptep
);
3250 bool is_hugetlb_entry_migration(pte_t pte
)
3254 if (huge_pte_none(pte
) || pte_present(pte
))
3256 swp
= pte_to_swp_entry(pte
);
3257 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3263 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3267 if (huge_pte_none(pte
) || pte_present(pte
))
3269 swp
= pte_to_swp_entry(pte
);
3270 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3276 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3277 struct vm_area_struct
*vma
)
3279 pte_t
*src_pte
, *dst_pte
, entry
;
3280 struct page
*ptepage
;
3283 struct hstate
*h
= hstate_vma(vma
);
3284 unsigned long sz
= huge_page_size(h
);
3285 unsigned long mmun_start
; /* For mmu_notifiers */
3286 unsigned long mmun_end
; /* For mmu_notifiers */
3289 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3291 mmun_start
= vma
->vm_start
;
3292 mmun_end
= vma
->vm_end
;
3294 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
3296 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3297 spinlock_t
*src_ptl
, *dst_ptl
;
3298 src_pte
= huge_pte_offset(src
, addr
, sz
);
3301 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3307 /* If the pagetables are shared don't copy or take references */
3308 if (dst_pte
== src_pte
)
3311 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3312 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3313 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3314 entry
= huge_ptep_get(src_pte
);
3315 if (huge_pte_none(entry
)) { /* skip none entry */
3317 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3318 is_hugetlb_entry_hwpoisoned(entry
))) {
3319 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3321 if (is_write_migration_entry(swp_entry
) && cow
) {
3323 * COW mappings require pages in both
3324 * parent and child to be set to read.
3326 make_migration_entry_read(&swp_entry
);
3327 entry
= swp_entry_to_pte(swp_entry
);
3328 set_huge_swap_pte_at(src
, addr
, src_pte
,
3331 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
3334 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3335 mmu_notifier_invalidate_range(src
, mmun_start
,
3338 entry
= huge_ptep_get(src_pte
);
3339 ptepage
= pte_page(entry
);
3341 page_dup_rmap(ptepage
, true);
3342 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3343 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3345 spin_unlock(src_ptl
);
3346 spin_unlock(dst_ptl
);
3350 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
3355 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3356 unsigned long start
, unsigned long end
,
3357 struct page
*ref_page
)
3359 struct mm_struct
*mm
= vma
->vm_mm
;
3360 unsigned long address
;
3365 struct hstate
*h
= hstate_vma(vma
);
3366 unsigned long sz
= huge_page_size(h
);
3367 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
3368 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
3370 WARN_ON(!is_vm_hugetlb_page(vma
));
3371 BUG_ON(start
& ~huge_page_mask(h
));
3372 BUG_ON(end
& ~huge_page_mask(h
));
3375 * This is a hugetlb vma, all the pte entries should point
3378 tlb_remove_check_page_size_change(tlb
, sz
);
3379 tlb_start_vma(tlb
, vma
);
3380 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3382 for (; address
< end
; address
+= sz
) {
3383 ptep
= huge_pte_offset(mm
, address
, sz
);
3387 ptl
= huge_pte_lock(h
, mm
, ptep
);
3388 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3393 pte
= huge_ptep_get(ptep
);
3394 if (huge_pte_none(pte
)) {
3400 * Migrating hugepage or HWPoisoned hugepage is already
3401 * unmapped and its refcount is dropped, so just clear pte here.
3403 if (unlikely(!pte_present(pte
))) {
3404 huge_pte_clear(mm
, address
, ptep
, sz
);
3409 page
= pte_page(pte
);
3411 * If a reference page is supplied, it is because a specific
3412 * page is being unmapped, not a range. Ensure the page we
3413 * are about to unmap is the actual page of interest.
3416 if (page
!= ref_page
) {
3421 * Mark the VMA as having unmapped its page so that
3422 * future faults in this VMA will fail rather than
3423 * looking like data was lost
3425 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3428 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3429 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3430 if (huge_pte_dirty(pte
))
3431 set_page_dirty(page
);
3433 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3434 page_remove_rmap(page
, true);
3437 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3439 * Bail out after unmapping reference page if supplied
3444 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3445 tlb_end_vma(tlb
, vma
);
3448 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3449 struct vm_area_struct
*vma
, unsigned long start
,
3450 unsigned long end
, struct page
*ref_page
)
3452 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3455 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3456 * test will fail on a vma being torn down, and not grab a page table
3457 * on its way out. We're lucky that the flag has such an appropriate
3458 * name, and can in fact be safely cleared here. We could clear it
3459 * before the __unmap_hugepage_range above, but all that's necessary
3460 * is to clear it before releasing the i_mmap_rwsem. This works
3461 * because in the context this is called, the VMA is about to be
3462 * destroyed and the i_mmap_rwsem is held.
3464 vma
->vm_flags
&= ~VM_MAYSHARE
;
3467 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3468 unsigned long end
, struct page
*ref_page
)
3470 struct mm_struct
*mm
;
3471 struct mmu_gather tlb
;
3475 tlb_gather_mmu(&tlb
, mm
, start
, end
);
3476 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3477 tlb_finish_mmu(&tlb
, start
, end
);
3481 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3482 * mappping it owns the reserve page for. The intention is to unmap the page
3483 * from other VMAs and let the children be SIGKILLed if they are faulting the
3486 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3487 struct page
*page
, unsigned long address
)
3489 struct hstate
*h
= hstate_vma(vma
);
3490 struct vm_area_struct
*iter_vma
;
3491 struct address_space
*mapping
;
3495 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3496 * from page cache lookup which is in HPAGE_SIZE units.
3498 address
= address
& huge_page_mask(h
);
3499 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3501 mapping
= vma
->vm_file
->f_mapping
;
3504 * Take the mapping lock for the duration of the table walk. As
3505 * this mapping should be shared between all the VMAs,
3506 * __unmap_hugepage_range() is called as the lock is already held
3508 i_mmap_lock_write(mapping
);
3509 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3510 /* Do not unmap the current VMA */
3511 if (iter_vma
== vma
)
3515 * Shared VMAs have their own reserves and do not affect
3516 * MAP_PRIVATE accounting but it is possible that a shared
3517 * VMA is using the same page so check and skip such VMAs.
3519 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3523 * Unmap the page from other VMAs without their own reserves.
3524 * They get marked to be SIGKILLed if they fault in these
3525 * areas. This is because a future no-page fault on this VMA
3526 * could insert a zeroed page instead of the data existing
3527 * from the time of fork. This would look like data corruption
3529 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3530 unmap_hugepage_range(iter_vma
, address
,
3531 address
+ huge_page_size(h
), page
);
3533 i_mmap_unlock_write(mapping
);
3537 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3538 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3539 * cannot race with other handlers or page migration.
3540 * Keep the pte_same checks anyway to make transition from the mutex easier.
3542 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3543 unsigned long address
, pte_t
*ptep
,
3544 struct page
*pagecache_page
, spinlock_t
*ptl
)
3547 struct hstate
*h
= hstate_vma(vma
);
3548 struct page
*old_page
, *new_page
;
3549 int ret
= 0, outside_reserve
= 0;
3550 unsigned long mmun_start
; /* For mmu_notifiers */
3551 unsigned long mmun_end
; /* For mmu_notifiers */
3553 pte
= huge_ptep_get(ptep
);
3554 old_page
= pte_page(pte
);
3557 /* If no-one else is actually using this page, avoid the copy
3558 * and just make the page writable */
3559 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3560 page_move_anon_rmap(old_page
, vma
);
3561 set_huge_ptep_writable(vma
, address
, ptep
);
3566 * If the process that created a MAP_PRIVATE mapping is about to
3567 * perform a COW due to a shared page count, attempt to satisfy
3568 * the allocation without using the existing reserves. The pagecache
3569 * page is used to determine if the reserve at this address was
3570 * consumed or not. If reserves were used, a partial faulted mapping
3571 * at the time of fork() could consume its reserves on COW instead
3572 * of the full address range.
3574 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3575 old_page
!= pagecache_page
)
3576 outside_reserve
= 1;
3581 * Drop page table lock as buddy allocator may be called. It will
3582 * be acquired again before returning to the caller, as expected.
3585 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
3587 if (IS_ERR(new_page
)) {
3589 * If a process owning a MAP_PRIVATE mapping fails to COW,
3590 * it is due to references held by a child and an insufficient
3591 * huge page pool. To guarantee the original mappers
3592 * reliability, unmap the page from child processes. The child
3593 * may get SIGKILLed if it later faults.
3595 if (outside_reserve
) {
3597 BUG_ON(huge_pte_none(pte
));
3598 unmap_ref_private(mm
, vma
, old_page
, address
);
3599 BUG_ON(huge_pte_none(pte
));
3601 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
),
3604 pte_same(huge_ptep_get(ptep
), pte
)))
3605 goto retry_avoidcopy
;
3607 * race occurs while re-acquiring page table
3608 * lock, and our job is done.
3613 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
3614 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
3615 goto out_release_old
;
3619 * When the original hugepage is shared one, it does not have
3620 * anon_vma prepared.
3622 if (unlikely(anon_vma_prepare(vma
))) {
3624 goto out_release_all
;
3627 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3628 pages_per_huge_page(h
));
3629 __SetPageUptodate(new_page
);
3630 set_page_huge_active(new_page
);
3632 mmun_start
= address
& huge_page_mask(h
);
3633 mmun_end
= mmun_start
+ huge_page_size(h
);
3634 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3637 * Retake the page table lock to check for racing updates
3638 * before the page tables are altered
3641 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
),
3643 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3644 ClearPagePrivate(new_page
);
3647 huge_ptep_clear_flush(vma
, address
, ptep
);
3648 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
3649 set_huge_pte_at(mm
, address
, ptep
,
3650 make_huge_pte(vma
, new_page
, 1));
3651 page_remove_rmap(old_page
, true);
3652 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
3653 /* Make the old page be freed below */
3654 new_page
= old_page
;
3657 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3659 restore_reserve_on_error(h
, vma
, address
, new_page
);
3664 spin_lock(ptl
); /* Caller expects lock to be held */
3668 /* Return the pagecache page at a given address within a VMA */
3669 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3670 struct vm_area_struct
*vma
, unsigned long address
)
3672 struct address_space
*mapping
;
3675 mapping
= vma
->vm_file
->f_mapping
;
3676 idx
= vma_hugecache_offset(h
, vma
, address
);
3678 return find_lock_page(mapping
, idx
);
3682 * Return whether there is a pagecache page to back given address within VMA.
3683 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3685 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3686 struct vm_area_struct
*vma
, unsigned long address
)
3688 struct address_space
*mapping
;
3692 mapping
= vma
->vm_file
->f_mapping
;
3693 idx
= vma_hugecache_offset(h
, vma
, address
);
3695 page
= find_get_page(mapping
, idx
);
3698 return page
!= NULL
;
3701 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3704 struct inode
*inode
= mapping
->host
;
3705 struct hstate
*h
= hstate_inode(inode
);
3706 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3710 ClearPagePrivate(page
);
3712 spin_lock(&inode
->i_lock
);
3713 inode
->i_blocks
+= blocks_per_huge_page(h
);
3714 spin_unlock(&inode
->i_lock
);
3718 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3719 struct address_space
*mapping
, pgoff_t idx
,
3720 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3722 struct hstate
*h
= hstate_vma(vma
);
3723 int ret
= VM_FAULT_SIGBUS
;
3731 * Currently, we are forced to kill the process in the event the
3732 * original mapper has unmapped pages from the child due to a failed
3733 * COW. Warn that such a situation has occurred as it may not be obvious
3735 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3736 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3742 * Use page lock to guard against racing truncation
3743 * before we get page_table_lock.
3746 page
= find_lock_page(mapping
, idx
);
3748 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3753 * Check for page in userfault range
3755 if (userfaultfd_missing(vma
)) {
3757 struct vm_fault vmf
= {
3762 * Hard to debug if it ends up being
3763 * used by a callee that assumes
3764 * something about the other
3765 * uninitialized fields... same as in
3771 * hugetlb_fault_mutex must be dropped before
3772 * handling userfault. Reacquire after handling
3773 * fault to make calling code simpler.
3775 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
,
3777 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3778 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
3779 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3783 page
= alloc_huge_page(vma
, address
, 0);
3785 ret
= PTR_ERR(page
);
3789 ret
= VM_FAULT_SIGBUS
;
3792 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3793 __SetPageUptodate(page
);
3794 set_page_huge_active(page
);
3796 if (vma
->vm_flags
& VM_MAYSHARE
) {
3797 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3806 if (unlikely(anon_vma_prepare(vma
))) {
3808 goto backout_unlocked
;
3814 * If memory error occurs between mmap() and fault, some process
3815 * don't have hwpoisoned swap entry for errored virtual address.
3816 * So we need to block hugepage fault by PG_hwpoison bit check.
3818 if (unlikely(PageHWPoison(page
))) {
3819 ret
= VM_FAULT_HWPOISON
|
3820 VM_FAULT_SET_HINDEX(hstate_index(h
));
3821 goto backout_unlocked
;
3826 * If we are going to COW a private mapping later, we examine the
3827 * pending reservations for this page now. This will ensure that
3828 * any allocations necessary to record that reservation occur outside
3831 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3832 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3834 goto backout_unlocked
;
3836 /* Just decrements count, does not deallocate */
3837 vma_end_reservation(h
, vma
, address
);
3840 ptl
= huge_pte_lock(h
, mm
, ptep
);
3841 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3846 if (!huge_pte_none(huge_ptep_get(ptep
)))
3850 ClearPagePrivate(page
);
3851 hugepage_add_new_anon_rmap(page
, vma
, address
);
3853 page_dup_rmap(page
, true);
3854 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3855 && (vma
->vm_flags
& VM_SHARED
)));
3856 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3858 hugetlb_count_add(pages_per_huge_page(h
), mm
);
3859 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3860 /* Optimization, do the COW without a second fault */
3861 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
3873 restore_reserve_on_error(h
, vma
, address
, page
);
3879 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3880 struct vm_area_struct
*vma
,
3881 struct address_space
*mapping
,
3882 pgoff_t idx
, unsigned long address
)
3884 unsigned long key
[2];
3887 if (vma
->vm_flags
& VM_SHARED
) {
3888 key
[0] = (unsigned long) mapping
;
3891 key
[0] = (unsigned long) mm
;
3892 key
[1] = address
>> huge_page_shift(h
);
3895 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3897 return hash
& (num_fault_mutexes
- 1);
3901 * For uniprocesor systems we always use a single mutex, so just
3902 * return 0 and avoid the hashing overhead.
3904 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3905 struct vm_area_struct
*vma
,
3906 struct address_space
*mapping
,
3907 pgoff_t idx
, unsigned long address
)
3913 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3914 unsigned long address
, unsigned int flags
)
3921 struct page
*page
= NULL
;
3922 struct page
*pagecache_page
= NULL
;
3923 struct hstate
*h
= hstate_vma(vma
);
3924 struct address_space
*mapping
;
3925 int need_wait_lock
= 0;
3927 address
&= huge_page_mask(h
);
3929 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
3931 entry
= huge_ptep_get(ptep
);
3932 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3933 migration_entry_wait_huge(vma
, mm
, ptep
);
3935 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3936 return VM_FAULT_HWPOISON_LARGE
|
3937 VM_FAULT_SET_HINDEX(hstate_index(h
));
3939 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3941 return VM_FAULT_OOM
;
3944 mapping
= vma
->vm_file
->f_mapping
;
3945 idx
= vma_hugecache_offset(h
, vma
, address
);
3948 * Serialize hugepage allocation and instantiation, so that we don't
3949 * get spurious allocation failures if two CPUs race to instantiate
3950 * the same page in the page cache.
3952 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3953 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3955 entry
= huge_ptep_get(ptep
);
3956 if (huge_pte_none(entry
)) {
3957 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3964 * entry could be a migration/hwpoison entry at this point, so this
3965 * check prevents the kernel from going below assuming that we have
3966 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3967 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3970 if (!pte_present(entry
))
3974 * If we are going to COW the mapping later, we examine the pending
3975 * reservations for this page now. This will ensure that any
3976 * allocations necessary to record that reservation occur outside the
3977 * spinlock. For private mappings, we also lookup the pagecache
3978 * page now as it is used to determine if a reservation has been
3981 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3982 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3986 /* Just decrements count, does not deallocate */
3987 vma_end_reservation(h
, vma
, address
);
3989 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3990 pagecache_page
= hugetlbfs_pagecache_page(h
,
3994 ptl
= huge_pte_lock(h
, mm
, ptep
);
3996 /* Check for a racing update before calling hugetlb_cow */
3997 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
4001 * hugetlb_cow() requires page locks of pte_page(entry) and
4002 * pagecache_page, so here we need take the former one
4003 * when page != pagecache_page or !pagecache_page.
4005 page
= pte_page(entry
);
4006 if (page
!= pagecache_page
)
4007 if (!trylock_page(page
)) {
4014 if (flags
& FAULT_FLAG_WRITE
) {
4015 if (!huge_pte_write(entry
)) {
4016 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
4017 pagecache_page
, ptl
);
4020 entry
= huge_pte_mkdirty(entry
);
4022 entry
= pte_mkyoung(entry
);
4023 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
4024 flags
& FAULT_FLAG_WRITE
))
4025 update_mmu_cache(vma
, address
, ptep
);
4027 if (page
!= pagecache_page
)
4033 if (pagecache_page
) {
4034 unlock_page(pagecache_page
);
4035 put_page(pagecache_page
);
4038 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4040 * Generally it's safe to hold refcount during waiting page lock. But
4041 * here we just wait to defer the next page fault to avoid busy loop and
4042 * the page is not used after unlocked before returning from the current
4043 * page fault. So we are safe from accessing freed page, even if we wait
4044 * here without taking refcount.
4047 wait_on_page_locked(page
);
4052 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4053 * modifications for huge pages.
4055 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
4057 struct vm_area_struct
*dst_vma
,
4058 unsigned long dst_addr
,
4059 unsigned long src_addr
,
4060 struct page
**pagep
)
4062 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
4063 struct hstate
*h
= hstate_vma(dst_vma
);
4071 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4075 ret
= copy_huge_page_from_user(page
,
4076 (const void __user
*) src_addr
,
4077 pages_per_huge_page(h
), false);
4079 /* fallback to copy_from_user outside mmap_sem */
4080 if (unlikely(ret
)) {
4083 /* don't free the page */
4092 * The memory barrier inside __SetPageUptodate makes sure that
4093 * preceding stores to the page contents become visible before
4094 * the set_pte_at() write.
4096 __SetPageUptodate(page
);
4097 set_page_huge_active(page
);
4100 * If shared, add to page cache
4103 struct address_space
*mapping
= dst_vma
->vm_file
->f_mapping
;
4104 pgoff_t idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4106 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4108 goto out_release_nounlock
;
4111 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4115 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4116 goto out_release_unlock
;
4119 page_dup_rmap(page
, true);
4121 ClearPagePrivate(page
);
4122 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4125 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4126 if (dst_vma
->vm_flags
& VM_WRITE
)
4127 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4128 _dst_pte
= pte_mkyoung(_dst_pte
);
4130 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4132 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4133 dst_vma
->vm_flags
& VM_WRITE
);
4134 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4136 /* No need to invalidate - it was non-present before */
4137 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4147 out_release_nounlock
:
4154 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4155 struct page
**pages
, struct vm_area_struct
**vmas
,
4156 unsigned long *position
, unsigned long *nr_pages
,
4157 long i
, unsigned int flags
, int *nonblocking
)
4159 unsigned long pfn_offset
;
4160 unsigned long vaddr
= *position
;
4161 unsigned long remainder
= *nr_pages
;
4162 struct hstate
*h
= hstate_vma(vma
);
4164 while (vaddr
< vma
->vm_end
&& remainder
) {
4166 spinlock_t
*ptl
= NULL
;
4171 * If we have a pending SIGKILL, don't keep faulting pages and
4172 * potentially allocating memory.
4174 if (unlikely(fatal_signal_pending(current
))) {
4180 * Some archs (sparc64, sh*) have multiple pte_ts to
4181 * each hugepage. We have to make sure we get the
4182 * first, for the page indexing below to work.
4184 * Note that page table lock is not held when pte is null.
4186 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
4189 ptl
= huge_pte_lock(h
, mm
, pte
);
4190 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4193 * When coredumping, it suits get_dump_page if we just return
4194 * an error where there's an empty slot with no huge pagecache
4195 * to back it. This way, we avoid allocating a hugepage, and
4196 * the sparse dumpfile avoids allocating disk blocks, but its
4197 * huge holes still show up with zeroes where they need to be.
4199 if (absent
&& (flags
& FOLL_DUMP
) &&
4200 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4208 * We need call hugetlb_fault for both hugepages under migration
4209 * (in which case hugetlb_fault waits for the migration,) and
4210 * hwpoisoned hugepages (in which case we need to prevent the
4211 * caller from accessing to them.) In order to do this, we use
4212 * here is_swap_pte instead of is_hugetlb_entry_migration and
4213 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4214 * both cases, and because we can't follow correct pages
4215 * directly from any kind of swap entries.
4217 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4218 ((flags
& FOLL_WRITE
) &&
4219 !huge_pte_write(huge_ptep_get(pte
)))) {
4221 unsigned int fault_flags
= 0;
4225 if (flags
& FOLL_WRITE
)
4226 fault_flags
|= FAULT_FLAG_WRITE
;
4228 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
4229 if (flags
& FOLL_NOWAIT
)
4230 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4231 FAULT_FLAG_RETRY_NOWAIT
;
4232 if (flags
& FOLL_TRIED
) {
4233 VM_WARN_ON_ONCE(fault_flags
&
4234 FAULT_FLAG_ALLOW_RETRY
);
4235 fault_flags
|= FAULT_FLAG_TRIED
;
4237 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4238 if (ret
& VM_FAULT_ERROR
) {
4239 int err
= vm_fault_to_errno(ret
, flags
);
4247 if (ret
& VM_FAULT_RETRY
) {
4252 * VM_FAULT_RETRY must not return an
4253 * error, it will return zero
4256 * No need to update "position" as the
4257 * caller will not check it after
4258 * *nr_pages is set to 0.
4265 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4266 page
= pte_page(huge_ptep_get(pte
));
4269 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4280 if (vaddr
< vma
->vm_end
&& remainder
&&
4281 pfn_offset
< pages_per_huge_page(h
)) {
4283 * We use pfn_offset to avoid touching the pageframes
4284 * of this compound page.
4290 *nr_pages
= remainder
;
4292 * setting position is actually required only if remainder is
4293 * not zero but it's faster not to add a "if (remainder)"
4298 return i
? i
: -EFAULT
;
4301 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4303 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4306 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4309 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
4310 unsigned long address
, unsigned long end
, pgprot_t newprot
)
4312 struct mm_struct
*mm
= vma
->vm_mm
;
4313 unsigned long start
= address
;
4316 struct hstate
*h
= hstate_vma(vma
);
4317 unsigned long pages
= 0;
4319 BUG_ON(address
>= end
);
4320 flush_cache_range(vma
, address
, end
);
4322 mmu_notifier_invalidate_range_start(mm
, start
, end
);
4323 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
4324 for (; address
< end
; address
+= huge_page_size(h
)) {
4326 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
4329 ptl
= huge_pte_lock(h
, mm
, ptep
);
4330 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
4335 pte
= huge_ptep_get(ptep
);
4336 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
4340 if (unlikely(is_hugetlb_entry_migration(pte
))) {
4341 swp_entry_t entry
= pte_to_swp_entry(pte
);
4343 if (is_write_migration_entry(entry
)) {
4346 make_migration_entry_read(&entry
);
4347 newpte
= swp_entry_to_pte(entry
);
4348 set_huge_swap_pte_at(mm
, address
, ptep
,
4349 newpte
, huge_page_size(h
));
4355 if (!huge_pte_none(pte
)) {
4356 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
4357 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
4358 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
4359 set_huge_pte_at(mm
, address
, ptep
, pte
);
4365 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4366 * may have cleared our pud entry and done put_page on the page table:
4367 * once we release i_mmap_rwsem, another task can do the final put_page
4368 * and that page table be reused and filled with junk.
4370 flush_hugetlb_tlb_range(vma
, start
, end
);
4371 mmu_notifier_invalidate_range(mm
, start
, end
);
4372 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
4373 mmu_notifier_invalidate_range_end(mm
, start
, end
);
4375 return pages
<< h
->order
;
4378 int hugetlb_reserve_pages(struct inode
*inode
,
4380 struct vm_area_struct
*vma
,
4381 vm_flags_t vm_flags
)
4384 struct hstate
*h
= hstate_inode(inode
);
4385 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4386 struct resv_map
*resv_map
;
4390 * Only apply hugepage reservation if asked. At fault time, an
4391 * attempt will be made for VM_NORESERVE to allocate a page
4392 * without using reserves
4394 if (vm_flags
& VM_NORESERVE
)
4398 * Shared mappings base their reservation on the number of pages that
4399 * are already allocated on behalf of the file. Private mappings need
4400 * to reserve the full area even if read-only as mprotect() may be
4401 * called to make the mapping read-write. Assume !vma is a shm mapping
4403 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4404 resv_map
= inode_resv_map(inode
);
4406 chg
= region_chg(resv_map
, from
, to
);
4409 resv_map
= resv_map_alloc();
4415 set_vma_resv_map(vma
, resv_map
);
4416 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4425 * There must be enough pages in the subpool for the mapping. If
4426 * the subpool has a minimum size, there may be some global
4427 * reservations already in place (gbl_reserve).
4429 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4430 if (gbl_reserve
< 0) {
4436 * Check enough hugepages are available for the reservation.
4437 * Hand the pages back to the subpool if there are not
4439 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4441 /* put back original number of pages, chg */
4442 (void)hugepage_subpool_put_pages(spool
, chg
);
4447 * Account for the reservations made. Shared mappings record regions
4448 * that have reservations as they are shared by multiple VMAs.
4449 * When the last VMA disappears, the region map says how much
4450 * the reservation was and the page cache tells how much of
4451 * the reservation was consumed. Private mappings are per-VMA and
4452 * only the consumed reservations are tracked. When the VMA
4453 * disappears, the original reservation is the VMA size and the
4454 * consumed reservations are stored in the map. Hence, nothing
4455 * else has to be done for private mappings here
4457 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4458 long add
= region_add(resv_map
, from
, to
);
4460 if (unlikely(chg
> add
)) {
4462 * pages in this range were added to the reserve
4463 * map between region_chg and region_add. This
4464 * indicates a race with alloc_huge_page. Adjust
4465 * the subpool and reserve counts modified above
4466 * based on the difference.
4470 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4472 hugetlb_acct_memory(h
, -rsv_adjust
);
4477 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4478 /* Don't call region_abort if region_chg failed */
4480 region_abort(resv_map
, from
, to
);
4481 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4482 kref_put(&resv_map
->refs
, resv_map_release
);
4486 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4489 struct hstate
*h
= hstate_inode(inode
);
4490 struct resv_map
*resv_map
= inode_resv_map(inode
);
4492 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4496 chg
= region_del(resv_map
, start
, end
);
4498 * region_del() can fail in the rare case where a region
4499 * must be split and another region descriptor can not be
4500 * allocated. If end == LONG_MAX, it will not fail.
4506 spin_lock(&inode
->i_lock
);
4507 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4508 spin_unlock(&inode
->i_lock
);
4511 * If the subpool has a minimum size, the number of global
4512 * reservations to be released may be adjusted.
4514 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4515 hugetlb_acct_memory(h
, -gbl_reserve
);
4520 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4521 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4522 struct vm_area_struct
*vma
,
4523 unsigned long addr
, pgoff_t idx
)
4525 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4527 unsigned long sbase
= saddr
& PUD_MASK
;
4528 unsigned long s_end
= sbase
+ PUD_SIZE
;
4530 /* Allow segments to share if only one is marked locked */
4531 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4532 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4535 * match the virtual addresses, permission and the alignment of the
4538 if (pmd_index(addr
) != pmd_index(saddr
) ||
4539 vm_flags
!= svm_flags
||
4540 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4546 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4548 unsigned long base
= addr
& PUD_MASK
;
4549 unsigned long end
= base
+ PUD_SIZE
;
4552 * check on proper vm_flags and page table alignment
4554 if (vma
->vm_flags
& VM_MAYSHARE
&&
4555 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
4561 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4562 * and returns the corresponding pte. While this is not necessary for the
4563 * !shared pmd case because we can allocate the pmd later as well, it makes the
4564 * code much cleaner. pmd allocation is essential for the shared case because
4565 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4566 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4567 * bad pmd for sharing.
4569 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4571 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4572 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4573 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4575 struct vm_area_struct
*svma
;
4576 unsigned long saddr
;
4581 if (!vma_shareable(vma
, addr
))
4582 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4584 i_mmap_lock_write(mapping
);
4585 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4589 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4591 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
4592 vma_mmu_pagesize(svma
));
4594 get_page(virt_to_page(spte
));
4603 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
4604 if (pud_none(*pud
)) {
4605 pud_populate(mm
, pud
,
4606 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4609 put_page(virt_to_page(spte
));
4613 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4614 i_mmap_unlock_write(mapping
);
4619 * unmap huge page backed by shared pte.
4621 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4622 * indicated by page_count > 1, unmap is achieved by clearing pud and
4623 * decrementing the ref count. If count == 1, the pte page is not shared.
4625 * called with page table lock held.
4627 * returns: 1 successfully unmapped a shared pte page
4628 * 0 the underlying pte page is not shared, or it is the last user
4630 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4632 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4633 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
4634 pud_t
*pud
= pud_offset(p4d
, *addr
);
4636 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4637 if (page_count(virt_to_page(ptep
)) == 1)
4641 put_page(virt_to_page(ptep
));
4643 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4646 #define want_pmd_share() (1)
4647 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4648 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4653 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4657 #define want_pmd_share() (0)
4658 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4660 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4661 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4662 unsigned long addr
, unsigned long sz
)
4669 pgd
= pgd_offset(mm
, addr
);
4670 p4d
= p4d_offset(pgd
, addr
);
4671 pud
= pud_alloc(mm
, p4d
, addr
);
4673 if (sz
== PUD_SIZE
) {
4676 BUG_ON(sz
!= PMD_SIZE
);
4677 if (want_pmd_share() && pud_none(*pud
))
4678 pte
= huge_pmd_share(mm
, addr
, pud
);
4680 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4683 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
4688 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
4689 unsigned long addr
, unsigned long sz
)
4696 pgd
= pgd_offset(mm
, addr
);
4697 if (!pgd_present(*pgd
))
4699 p4d
= p4d_offset(pgd
, addr
);
4700 if (!p4d_present(*p4d
))
4702 pud
= pud_offset(p4d
, addr
);
4703 if (!pud_present(*pud
))
4706 return (pte_t
*)pud
;
4707 pmd
= pmd_offset(pud
, addr
);
4708 return (pte_t
*) pmd
;
4711 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4714 * These functions are overwritable if your architecture needs its own
4717 struct page
* __weak
4718 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
4721 return ERR_PTR(-EINVAL
);
4724 struct page
* __weak
4725 follow_huge_pd(struct vm_area_struct
*vma
,
4726 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
4728 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4732 struct page
* __weak
4733 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
4734 pmd_t
*pmd
, int flags
)
4736 struct page
*page
= NULL
;
4740 ptl
= pmd_lockptr(mm
, pmd
);
4743 * make sure that the address range covered by this pmd is not
4744 * unmapped from other threads.
4746 if (!pmd_huge(*pmd
))
4748 pte
= huge_ptep_get((pte_t
*)pmd
);
4749 if (pte_present(pte
)) {
4750 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
4751 if (flags
& FOLL_GET
)
4754 if (is_hugetlb_entry_migration(pte
)) {
4756 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
4760 * hwpoisoned entry is treated as no_page_table in
4761 * follow_page_mask().
4769 struct page
* __weak
4770 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
4771 pud_t
*pud
, int flags
)
4773 if (flags
& FOLL_GET
)
4776 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
4779 struct page
* __weak
4780 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
4782 if (flags
& FOLL_GET
)
4785 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
4788 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
4792 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4793 spin_lock(&hugetlb_lock
);
4794 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
4798 clear_page_huge_active(page
);
4799 list_move_tail(&page
->lru
, list
);
4801 spin_unlock(&hugetlb_lock
);
4805 void putback_active_hugepage(struct page
*page
)
4807 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4808 spin_lock(&hugetlb_lock
);
4809 set_page_huge_active(page
);
4810 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
4811 spin_unlock(&hugetlb_lock
);