1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/mmdebug.h>
23 #include <linux/sched/signal.h>
24 #include <linux/rmap.h>
25 #include <linux/string_helpers.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/jhash.h>
29 #include <linux/numa.h>
30 #include <linux/llist.h>
33 #include <asm/pgtable.h>
37 #include <linux/hugetlb.h>
38 #include <linux/hugetlb_cgroup.h>
39 #include <linux/node.h>
40 #include <linux/userfaultfd_k.h>
41 #include <linux/page_owner.h>
44 int hugetlb_max_hstate __read_mostly
;
45 unsigned int default_hstate_idx
;
46 struct hstate hstates
[HUGE_MAX_HSTATE
];
48 * Minimum page order among possible hugepage sizes, set to a proper value
51 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
53 __initdata
LIST_HEAD(huge_boot_pages
);
55 /* for command line parsing */
56 static struct hstate
* __initdata parsed_hstate
;
57 static unsigned long __initdata default_hstate_max_huge_pages
;
58 static unsigned long __initdata default_hstate_size
;
59 static bool __initdata parsed_valid_hugepagesz
= true;
62 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
63 * free_huge_pages, and surplus_huge_pages.
65 DEFINE_SPINLOCK(hugetlb_lock
);
68 * Serializes faults on the same logical page. This is used to
69 * prevent spurious OOMs when the hugepage pool is fully utilized.
71 static int num_fault_mutexes
;
72 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
74 static inline bool PageHugeFreed(struct page
*head
)
76 return page_private(head
+ 4) == -1UL;
79 static inline void SetPageHugeFreed(struct page
*head
)
81 set_page_private(head
+ 4, -1UL);
84 static inline void ClearPageHugeFreed(struct page
*head
)
86 set_page_private(head
+ 4, 0);
89 /* Forward declaration */
90 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
92 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
94 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
96 spin_unlock(&spool
->lock
);
98 /* If no pages are used, and no other handles to the subpool
99 * remain, give up any reservations mased on minimum size and
100 * free the subpool */
102 if (spool
->min_hpages
!= -1)
103 hugetlb_acct_memory(spool
->hstate
,
109 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
112 struct hugepage_subpool
*spool
;
114 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
118 spin_lock_init(&spool
->lock
);
120 spool
->max_hpages
= max_hpages
;
122 spool
->min_hpages
= min_hpages
;
124 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
128 spool
->rsv_hpages
= min_hpages
;
133 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
135 spin_lock(&spool
->lock
);
136 BUG_ON(!spool
->count
);
138 unlock_or_release_subpool(spool
);
142 * Subpool accounting for allocating and reserving pages.
143 * Return -ENOMEM if there are not enough resources to satisfy the
144 * the request. Otherwise, return the number of pages by which the
145 * global pools must be adjusted (upward). The returned value may
146 * only be different than the passed value (delta) in the case where
147 * a subpool minimum size must be manitained.
149 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
157 spin_lock(&spool
->lock
);
159 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
160 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
161 spool
->used_hpages
+= delta
;
168 /* minimum size accounting */
169 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
170 if (delta
> spool
->rsv_hpages
) {
172 * Asking for more reserves than those already taken on
173 * behalf of subpool. Return difference.
175 ret
= delta
- spool
->rsv_hpages
;
176 spool
->rsv_hpages
= 0;
178 ret
= 0; /* reserves already accounted for */
179 spool
->rsv_hpages
-= delta
;
184 spin_unlock(&spool
->lock
);
189 * Subpool accounting for freeing and unreserving pages.
190 * Return the number of global page reservations that must be dropped.
191 * The return value may only be different than the passed value (delta)
192 * in the case where a subpool minimum size must be maintained.
194 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
202 spin_lock(&spool
->lock
);
204 if (spool
->max_hpages
!= -1) /* maximum size accounting */
205 spool
->used_hpages
-= delta
;
207 /* minimum size accounting */
208 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
209 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
212 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
214 spool
->rsv_hpages
+= delta
;
215 if (spool
->rsv_hpages
> spool
->min_hpages
)
216 spool
->rsv_hpages
= spool
->min_hpages
;
220 * If hugetlbfs_put_super couldn't free spool due to an outstanding
221 * quota reference, free it now.
223 unlock_or_release_subpool(spool
);
228 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
230 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
233 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
235 return subpool_inode(file_inode(vma
->vm_file
));
239 * Region tracking -- allows tracking of reservations and instantiated pages
240 * across the pages in a mapping.
242 * The region data structures are embedded into a resv_map and protected
243 * by a resv_map's lock. The set of regions within the resv_map represent
244 * reservations for huge pages, or huge pages that have already been
245 * instantiated within the map. The from and to elements are huge page
246 * indicies into the associated mapping. from indicates the starting index
247 * of the region. to represents the first index past the end of the region.
249 * For example, a file region structure with from == 0 and to == 4 represents
250 * four huge pages in a mapping. It is important to note that the to element
251 * represents the first element past the end of the region. This is used in
252 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
254 * Interval notation of the form [from, to) will be used to indicate that
255 * the endpoint from is inclusive and to is exclusive.
258 struct list_head link
;
264 * Add the huge page range represented by [f, t) to the reserve
265 * map. In the normal case, existing regions will be expanded
266 * to accommodate the specified range. Sufficient regions should
267 * exist for expansion due to the previous call to region_chg
268 * with the same range. However, it is possible that region_del
269 * could have been called after region_chg and modifed the map
270 * in such a way that no region exists to be expanded. In this
271 * case, pull a region descriptor from the cache associated with
272 * the map and use that for the new range.
274 * Return the number of new huge pages added to the map. This
275 * number is greater than or equal to zero.
277 static long region_add(struct resv_map
*resv
, long f
, long t
)
279 struct list_head
*head
= &resv
->regions
;
280 struct file_region
*rg
, *nrg
, *trg
;
283 spin_lock(&resv
->lock
);
284 /* Locate the region we are either in or before. */
285 list_for_each_entry(rg
, head
, link
)
290 * If no region exists which can be expanded to include the
291 * specified range, the list must have been modified by an
292 * interleving call to region_del(). Pull a region descriptor
293 * from the cache and use it for this range.
295 if (&rg
->link
== head
|| t
< rg
->from
) {
296 VM_BUG_ON(resv
->region_cache_count
<= 0);
298 resv
->region_cache_count
--;
299 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
,
301 list_del(&nrg
->link
);
305 list_add(&nrg
->link
, rg
->link
.prev
);
311 /* Round our left edge to the current segment if it encloses us. */
315 /* Check for and consume any regions we now overlap with. */
317 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
318 if (&rg
->link
== head
)
323 /* If this area reaches higher then extend our area to
324 * include it completely. If this is not the first area
325 * which we intend to reuse, free it. */
329 /* Decrement return value by the deleted range.
330 * Another range will span this area so that by
331 * end of routine add will be >= zero
333 add
-= (rg
->to
- rg
->from
);
339 add
+= (nrg
->from
- f
); /* Added to beginning of region */
341 add
+= t
- nrg
->to
; /* Added to end of region */
345 resv
->adds_in_progress
--;
346 spin_unlock(&resv
->lock
);
352 * Examine the existing reserve map and determine how many
353 * huge pages in the specified range [f, t) are NOT currently
354 * represented. This routine is called before a subsequent
355 * call to region_add that will actually modify the reserve
356 * map to add the specified range [f, t). region_chg does
357 * not change the number of huge pages represented by the
358 * map. However, if the existing regions in the map can not
359 * be expanded to represent the new range, a new file_region
360 * structure is added to the map as a placeholder. This is
361 * so that the subsequent region_add call will have all the
362 * regions it needs and will not fail.
364 * Upon entry, region_chg will also examine the cache of region descriptors
365 * associated with the map. If there are not enough descriptors cached, one
366 * will be allocated for the in progress add operation.
368 * Returns the number of huge pages that need to be added to the existing
369 * reservation map for the range [f, t). This number is greater or equal to
370 * zero. -ENOMEM is returned if a new file_region structure or cache entry
371 * is needed and can not be allocated.
373 static long region_chg(struct resv_map
*resv
, long f
, long t
)
375 struct list_head
*head
= &resv
->regions
;
376 struct file_region
*rg
, *nrg
= NULL
;
380 spin_lock(&resv
->lock
);
382 resv
->adds_in_progress
++;
385 * Check for sufficient descriptors in the cache to accommodate
386 * the number of in progress add operations.
388 if (resv
->adds_in_progress
> resv
->region_cache_count
) {
389 struct file_region
*trg
;
391 VM_BUG_ON(resv
->adds_in_progress
- resv
->region_cache_count
> 1);
392 /* Must drop lock to allocate a new descriptor. */
393 resv
->adds_in_progress
--;
394 spin_unlock(&resv
->lock
);
396 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
402 spin_lock(&resv
->lock
);
403 list_add(&trg
->link
, &resv
->region_cache
);
404 resv
->region_cache_count
++;
408 /* Locate the region we are before or in. */
409 list_for_each_entry(rg
, head
, link
)
413 /* If we are below the current region then a new region is required.
414 * Subtle, allocate a new region at the position but make it zero
415 * size such that we can guarantee to record the reservation. */
416 if (&rg
->link
== head
|| t
< rg
->from
) {
418 resv
->adds_in_progress
--;
419 spin_unlock(&resv
->lock
);
420 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
426 INIT_LIST_HEAD(&nrg
->link
);
430 list_add(&nrg
->link
, rg
->link
.prev
);
435 /* Round our left edge to the current segment if it encloses us. */
440 /* Check for and consume any regions we now overlap with. */
441 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
442 if (&rg
->link
== head
)
447 /* We overlap with this area, if it extends further than
448 * us then we must extend ourselves. Account for its
449 * existing reservation. */
454 chg
-= rg
->to
- rg
->from
;
458 spin_unlock(&resv
->lock
);
459 /* We already know we raced and no longer need the new region */
463 spin_unlock(&resv
->lock
);
468 * Abort the in progress add operation. The adds_in_progress field
469 * of the resv_map keeps track of the operations in progress between
470 * calls to region_chg and region_add. Operations are sometimes
471 * aborted after the call to region_chg. In such cases, region_abort
472 * is called to decrement the adds_in_progress counter.
474 * NOTE: The range arguments [f, t) are not needed or used in this
475 * routine. They are kept to make reading the calling code easier as
476 * arguments will match the associated region_chg call.
478 static void region_abort(struct resv_map
*resv
, long f
, long t
)
480 spin_lock(&resv
->lock
);
481 VM_BUG_ON(!resv
->region_cache_count
);
482 resv
->adds_in_progress
--;
483 spin_unlock(&resv
->lock
);
487 * Delete the specified range [f, t) from the reserve map. If the
488 * t parameter is LONG_MAX, this indicates that ALL regions after f
489 * should be deleted. Locate the regions which intersect [f, t)
490 * and either trim, delete or split the existing regions.
492 * Returns the number of huge pages deleted from the reserve map.
493 * In the normal case, the return value is zero or more. In the
494 * case where a region must be split, a new region descriptor must
495 * be allocated. If the allocation fails, -ENOMEM will be returned.
496 * NOTE: If the parameter t == LONG_MAX, then we will never split
497 * a region and possibly return -ENOMEM. Callers specifying
498 * t == LONG_MAX do not need to check for -ENOMEM error.
500 static long region_del(struct resv_map
*resv
, long f
, long t
)
502 struct list_head
*head
= &resv
->regions
;
503 struct file_region
*rg
, *trg
;
504 struct file_region
*nrg
= NULL
;
508 spin_lock(&resv
->lock
);
509 list_for_each_entry_safe(rg
, trg
, head
, link
) {
511 * Skip regions before the range to be deleted. file_region
512 * ranges are normally of the form [from, to). However, there
513 * may be a "placeholder" entry in the map which is of the form
514 * (from, to) with from == to. Check for placeholder entries
515 * at the beginning of the range to be deleted.
517 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
523 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
525 * Check for an entry in the cache before dropping
526 * lock and attempting allocation.
529 resv
->region_cache_count
> resv
->adds_in_progress
) {
530 nrg
= list_first_entry(&resv
->region_cache
,
533 list_del(&nrg
->link
);
534 resv
->region_cache_count
--;
538 spin_unlock(&resv
->lock
);
539 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
547 /* New entry for end of split region */
550 INIT_LIST_HEAD(&nrg
->link
);
552 /* Original entry is trimmed */
555 list_add(&nrg
->link
, &rg
->link
);
560 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
561 del
+= rg
->to
- rg
->from
;
567 if (f
<= rg
->from
) { /* Trim beginning of region */
570 } else { /* Trim end of region */
576 spin_unlock(&resv
->lock
);
582 * A rare out of memory error was encountered which prevented removal of
583 * the reserve map region for a page. The huge page itself was free'ed
584 * and removed from the page cache. This routine will adjust the subpool
585 * usage count, and the global reserve count if needed. By incrementing
586 * these counts, the reserve map entry which could not be deleted will
587 * appear as a "reserved" entry instead of simply dangling with incorrect
590 void hugetlb_fix_reserve_counts(struct inode
*inode
)
592 struct hugepage_subpool
*spool
= subpool_inode(inode
);
595 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
597 struct hstate
*h
= hstate_inode(inode
);
599 hugetlb_acct_memory(h
, 1);
604 * Count and return the number of huge pages in the reserve map
605 * that intersect with the range [f, t).
607 static long region_count(struct resv_map
*resv
, long f
, long t
)
609 struct list_head
*head
= &resv
->regions
;
610 struct file_region
*rg
;
613 spin_lock(&resv
->lock
);
614 /* Locate each segment we overlap with, and count that overlap. */
615 list_for_each_entry(rg
, head
, link
) {
624 seg_from
= max(rg
->from
, f
);
625 seg_to
= min(rg
->to
, t
);
627 chg
+= seg_to
- seg_from
;
629 spin_unlock(&resv
->lock
);
635 * Convert the address within this vma to the page offset within
636 * the mapping, in pagecache page units; huge pages here.
638 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
639 struct vm_area_struct
*vma
, unsigned long address
)
641 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
642 (vma
->vm_pgoff
>> huge_page_order(h
));
645 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
646 unsigned long address
)
648 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
650 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
653 * Return the size of the pages allocated when backing a VMA. In the majority
654 * cases this will be same size as used by the page table entries.
656 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
658 if (vma
->vm_ops
&& vma
->vm_ops
->pagesize
)
659 return vma
->vm_ops
->pagesize(vma
);
662 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
665 * Return the page size being used by the MMU to back a VMA. In the majority
666 * of cases, the page size used by the kernel matches the MMU size. On
667 * architectures where it differs, an architecture-specific 'strong'
668 * version of this symbol is required.
670 __weak
unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
672 return vma_kernel_pagesize(vma
);
676 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
677 * bits of the reservation map pointer, which are always clear due to
680 #define HPAGE_RESV_OWNER (1UL << 0)
681 #define HPAGE_RESV_UNMAPPED (1UL << 1)
682 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
685 * These helpers are used to track how many pages are reserved for
686 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
687 * is guaranteed to have their future faults succeed.
689 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
690 * the reserve counters are updated with the hugetlb_lock held. It is safe
691 * to reset the VMA at fork() time as it is not in use yet and there is no
692 * chance of the global counters getting corrupted as a result of the values.
694 * The private mapping reservation is represented in a subtly different
695 * manner to a shared mapping. A shared mapping has a region map associated
696 * with the underlying file, this region map represents the backing file
697 * pages which have ever had a reservation assigned which this persists even
698 * after the page is instantiated. A private mapping has a region map
699 * associated with the original mmap which is attached to all VMAs which
700 * reference it, this region map represents those offsets which have consumed
701 * reservation ie. where pages have been instantiated.
703 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
705 return (unsigned long)vma
->vm_private_data
;
708 static void set_vma_private_data(struct vm_area_struct
*vma
,
711 vma
->vm_private_data
= (void *)value
;
714 struct resv_map
*resv_map_alloc(void)
716 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
717 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
719 if (!resv_map
|| !rg
) {
725 kref_init(&resv_map
->refs
);
726 spin_lock_init(&resv_map
->lock
);
727 INIT_LIST_HEAD(&resv_map
->regions
);
729 resv_map
->adds_in_progress
= 0;
731 INIT_LIST_HEAD(&resv_map
->region_cache
);
732 list_add(&rg
->link
, &resv_map
->region_cache
);
733 resv_map
->region_cache_count
= 1;
738 void resv_map_release(struct kref
*ref
)
740 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
741 struct list_head
*head
= &resv_map
->region_cache
;
742 struct file_region
*rg
, *trg
;
744 /* Clear out any active regions before we release the map. */
745 region_del(resv_map
, 0, LONG_MAX
);
747 /* ... and any entries left in the cache */
748 list_for_each_entry_safe(rg
, trg
, head
, link
) {
753 VM_BUG_ON(resv_map
->adds_in_progress
);
758 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
761 * At inode evict time, i_mapping may not point to the original
762 * address space within the inode. This original address space
763 * contains the pointer to the resv_map. So, always use the
764 * address space embedded within the inode.
765 * The VERY common case is inode->mapping == &inode->i_data but,
766 * this may not be true for device special inodes.
768 return (struct resv_map
*)(&inode
->i_data
)->private_data
;
771 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
773 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
774 if (vma
->vm_flags
& VM_MAYSHARE
) {
775 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
776 struct inode
*inode
= mapping
->host
;
778 return inode_resv_map(inode
);
781 return (struct resv_map
*)(get_vma_private_data(vma
) &
786 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
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
) &
792 HPAGE_RESV_MASK
) | (unsigned long)map
);
795 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
797 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
798 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
800 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
803 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
805 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
807 return (get_vma_private_data(vma
) & flag
) != 0;
810 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
811 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
813 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
814 if (!(vma
->vm_flags
& VM_MAYSHARE
))
815 vma
->vm_private_data
= (void *)0;
818 /* Returns true if the VMA has associated reserve pages */
819 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
821 if (vma
->vm_flags
& VM_NORESERVE
) {
823 * This address is already reserved by other process(chg == 0),
824 * so, we should decrement reserved count. Without decrementing,
825 * reserve count remains after releasing inode, because this
826 * allocated page will go into page cache and is regarded as
827 * coming from reserved pool in releasing step. Currently, we
828 * don't have any other solution to deal with this situation
829 * properly, so add work-around here.
831 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
837 /* Shared mappings always use reserves */
838 if (vma
->vm_flags
& VM_MAYSHARE
) {
840 * We know VM_NORESERVE is not set. Therefore, there SHOULD
841 * be a region map for all pages. The only situation where
842 * there is no region map is if a hole was punched via
843 * fallocate. In this case, there really are no reverves to
844 * use. This situation is indicated if chg != 0.
853 * Only the process that called mmap() has reserves for
856 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
858 * Like the shared case above, a hole punch or truncate
859 * could have been performed on the private mapping.
860 * Examine the value of chg to determine if reserves
861 * actually exist or were previously consumed.
862 * Very Subtle - The value of chg comes from a previous
863 * call to vma_needs_reserves(). The reserve map for
864 * private mappings has different (opposite) semantics
865 * than that of shared mappings. vma_needs_reserves()
866 * has already taken this difference in semantics into
867 * account. Therefore, the meaning of chg is the same
868 * as in the shared case above. Code could easily be
869 * combined, but keeping it separate draws attention to
870 * subtle differences.
881 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
883 int nid
= page_to_nid(page
);
884 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
885 h
->free_huge_pages
++;
886 h
->free_huge_pages_node
[nid
]++;
887 SetPageHugeFreed(page
);
890 static struct page
*dequeue_huge_page_node_exact(struct hstate
*h
, int nid
)
894 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
895 if (!PageHWPoison(page
))
898 * if 'non-isolated free hugepage' not found on the list,
899 * the allocation fails.
901 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
903 list_move(&page
->lru
, &h
->hugepage_activelist
);
904 set_page_refcounted(page
);
905 ClearPageHugeFreed(page
);
906 h
->free_huge_pages
--;
907 h
->free_huge_pages_node
[nid
]--;
911 static struct page
*dequeue_huge_page_nodemask(struct hstate
*h
, gfp_t gfp_mask
, int nid
,
914 unsigned int cpuset_mems_cookie
;
915 struct zonelist
*zonelist
;
918 int node
= NUMA_NO_NODE
;
920 zonelist
= node_zonelist(nid
, gfp_mask
);
923 cpuset_mems_cookie
= read_mems_allowed_begin();
924 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nmask
) {
927 if (!cpuset_zone_allowed(zone
, gfp_mask
))
930 * no need to ask again on the same node. Pool is node rather than
933 if (zone_to_nid(zone
) == node
)
935 node
= zone_to_nid(zone
);
937 page
= dequeue_huge_page_node_exact(h
, node
);
941 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie
)))
947 /* Movability of hugepages depends on migration support. */
948 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
950 if (hugepage_movable_supported(h
))
951 return GFP_HIGHUSER_MOVABLE
;
956 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
957 struct vm_area_struct
*vma
,
958 unsigned long address
, int avoid_reserve
,
962 struct mempolicy
*mpol
;
964 nodemask_t
*nodemask
;
968 * A child process with MAP_PRIVATE mappings created by their parent
969 * have no page reserves. This check ensures that reservations are
970 * not "stolen". The child may still get SIGKILLed
972 if (!vma_has_reserves(vma
, chg
) &&
973 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
976 /* If reserves cannot be used, ensure enough pages are in the pool */
977 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
980 gfp_mask
= htlb_alloc_mask(h
);
981 nid
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
982 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, nodemask
);
983 if (page
&& !avoid_reserve
&& vma_has_reserves(vma
, chg
)) {
984 SetPagePrivate(page
);
985 h
->resv_huge_pages
--;
996 * common helper functions for hstate_next_node_to_{alloc|free}.
997 * We may have allocated or freed a huge page based on a different
998 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
999 * be outside of *nodes_allowed. Ensure that we use an allowed
1000 * node for alloc or free.
1002 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1004 nid
= next_node_in(nid
, *nodes_allowed
);
1005 VM_BUG_ON(nid
>= MAX_NUMNODES
);
1010 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
1012 if (!node_isset(nid
, *nodes_allowed
))
1013 nid
= next_node_allowed(nid
, nodes_allowed
);
1018 * returns the previously saved node ["this node"] from which to
1019 * allocate a persistent huge page for the pool and advance the
1020 * next node from which to allocate, handling wrap at end of node
1023 static int hstate_next_node_to_alloc(struct hstate
*h
,
1024 nodemask_t
*nodes_allowed
)
1028 VM_BUG_ON(!nodes_allowed
);
1030 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
1031 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
1037 * helper for free_pool_huge_page() - return the previously saved
1038 * node ["this node"] from which to free a huge page. Advance the
1039 * next node id whether or not we find a free huge page to free so
1040 * that the next attempt to free addresses the next node.
1042 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1046 VM_BUG_ON(!nodes_allowed
);
1048 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1049 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1054 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1055 for (nr_nodes = nodes_weight(*mask); \
1057 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1060 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1061 for (nr_nodes = nodes_weight(*mask); \
1063 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1066 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1067 static void destroy_compound_gigantic_page(struct page
*page
,
1071 int nr_pages
= 1 << order
;
1072 struct page
*p
= page
+ 1;
1074 atomic_set(compound_mapcount_ptr(page
), 0);
1075 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1076 clear_compound_head(p
);
1077 set_page_refcounted(p
);
1080 set_compound_order(page
, 0);
1081 __ClearPageHead(page
);
1084 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1086 free_contig_range(page_to_pfn(page
), 1 << order
);
1089 #ifdef CONFIG_CONTIG_ALLOC
1090 static int __alloc_gigantic_page(unsigned long start_pfn
,
1091 unsigned long nr_pages
, gfp_t gfp_mask
)
1093 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1094 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
1098 static bool pfn_range_valid_gigantic(struct zone
*z
,
1099 unsigned long start_pfn
, unsigned long nr_pages
)
1101 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1104 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1105 page
= pfn_to_online_page(i
);
1109 if (page_zone(page
) != z
)
1112 if (PageReserved(page
))
1115 if (page_count(page
) > 0)
1125 static bool zone_spans_last_pfn(const struct zone
*zone
,
1126 unsigned long start_pfn
, unsigned long nr_pages
)
1128 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1129 return zone_spans_pfn(zone
, last_pfn
);
1132 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1133 int nid
, nodemask_t
*nodemask
)
1135 unsigned int order
= huge_page_order(h
);
1136 unsigned long nr_pages
= 1 << order
;
1137 unsigned long ret
, pfn
, flags
;
1138 struct zonelist
*zonelist
;
1142 zonelist
= node_zonelist(nid
, gfp_mask
);
1143 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, gfp_zone(gfp_mask
), nodemask
) {
1144 spin_lock_irqsave(&zone
->lock
, flags
);
1146 pfn
= ALIGN(zone
->zone_start_pfn
, nr_pages
);
1147 while (zone_spans_last_pfn(zone
, pfn
, nr_pages
)) {
1148 if (pfn_range_valid_gigantic(zone
, pfn
, nr_pages
)) {
1150 * We release the zone lock here because
1151 * alloc_contig_range() will also lock the zone
1152 * at some point. If there's an allocation
1153 * spinning on this lock, it may win the race
1154 * and cause alloc_contig_range() to fail...
1156 spin_unlock_irqrestore(&zone
->lock
, flags
);
1157 ret
= __alloc_gigantic_page(pfn
, nr_pages
, gfp_mask
);
1159 return pfn_to_page(pfn
);
1160 spin_lock_irqsave(&zone
->lock
, flags
);
1165 spin_unlock_irqrestore(&zone
->lock
, flags
);
1171 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1172 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1173 #else /* !CONFIG_CONTIG_ALLOC */
1174 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1175 int nid
, nodemask_t
*nodemask
)
1179 #endif /* CONFIG_CONTIG_ALLOC */
1181 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1182 static struct page
*alloc_gigantic_page(struct hstate
*h
, gfp_t gfp_mask
,
1183 int nid
, nodemask_t
*nodemask
)
1187 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1188 static inline void destroy_compound_gigantic_page(struct page
*page
,
1189 unsigned int order
) { }
1192 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1195 struct page
*subpage
= page
;
1197 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
1201 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1202 for (i
= 0; i
< pages_per_huge_page(h
);
1203 i
++, subpage
= mem_map_next(subpage
, page
, i
)) {
1204 subpage
->flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1205 1 << PG_referenced
| 1 << PG_dirty
|
1206 1 << PG_active
| 1 << PG_private
|
1209 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1210 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1211 set_page_refcounted(page
);
1212 if (hstate_is_gigantic(h
)) {
1213 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1214 free_gigantic_page(page
, huge_page_order(h
));
1216 __free_pages(page
, huge_page_order(h
));
1220 struct hstate
*size_to_hstate(unsigned long size
)
1224 for_each_hstate(h
) {
1225 if (huge_page_size(h
) == size
)
1232 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1233 * to hstate->hugepage_activelist.)
1235 * This function can be called for tail pages, but never returns true for them.
1237 bool page_huge_active(struct page
*page
)
1239 return PageHeadHuge(page
) && PagePrivate(&page
[1]);
1242 /* never called for tail page */
1243 void set_page_huge_active(struct page
*page
)
1245 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1246 SetPagePrivate(&page
[1]);
1249 static void clear_page_huge_active(struct page
*page
)
1251 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1252 ClearPagePrivate(&page
[1]);
1256 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1259 static inline bool PageHugeTemporary(struct page
*page
)
1261 if (!PageHuge(page
))
1264 return (unsigned long)page
[2].mapping
== -1U;
1267 static inline void SetPageHugeTemporary(struct page
*page
)
1269 page
[2].mapping
= (void *)-1U;
1272 static inline void ClearPageHugeTemporary(struct page
*page
)
1274 page
[2].mapping
= NULL
;
1277 static void __free_huge_page(struct page
*page
)
1280 * Can't pass hstate in here because it is called from the
1281 * compound page destructor.
1283 struct hstate
*h
= page_hstate(page
);
1284 int nid
= page_to_nid(page
);
1285 struct hugepage_subpool
*spool
=
1286 (struct hugepage_subpool
*)page_private(page
);
1287 bool restore_reserve
;
1289 VM_BUG_ON_PAGE(page_count(page
), page
);
1290 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1292 set_page_private(page
, 0);
1293 page
->mapping
= NULL
;
1294 restore_reserve
= PagePrivate(page
);
1295 ClearPagePrivate(page
);
1298 * If PagePrivate() was set on page, page allocation consumed a
1299 * reservation. If the page was associated with a subpool, there
1300 * would have been a page reserved in the subpool before allocation
1301 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1302 * reservtion, do not call hugepage_subpool_put_pages() as this will
1303 * remove the reserved page from the subpool.
1305 if (!restore_reserve
) {
1307 * A return code of zero implies that the subpool will be
1308 * under its minimum size if the reservation is not restored
1309 * after page is free. Therefore, force restore_reserve
1312 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1313 restore_reserve
= true;
1316 spin_lock(&hugetlb_lock
);
1317 clear_page_huge_active(page
);
1318 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1319 pages_per_huge_page(h
), page
);
1320 if (restore_reserve
)
1321 h
->resv_huge_pages
++;
1323 if (PageHugeTemporary(page
)) {
1324 list_del(&page
->lru
);
1325 ClearPageHugeTemporary(page
);
1326 update_and_free_page(h
, page
);
1327 } else if (h
->surplus_huge_pages_node
[nid
]) {
1328 /* remove the page from active list */
1329 list_del(&page
->lru
);
1330 update_and_free_page(h
, page
);
1331 h
->surplus_huge_pages
--;
1332 h
->surplus_huge_pages_node
[nid
]--;
1334 arch_clear_hugepage_flags(page
);
1335 enqueue_huge_page(h
, page
);
1337 spin_unlock(&hugetlb_lock
);
1341 * As free_huge_page() can be called from a non-task context, we have
1342 * to defer the actual freeing in a workqueue to prevent potential
1343 * hugetlb_lock deadlock.
1345 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1346 * be freed and frees them one-by-one. As the page->mapping pointer is
1347 * going to be cleared in __free_huge_page() anyway, it is reused as the
1348 * llist_node structure of a lockless linked list of huge pages to be freed.
1350 static LLIST_HEAD(hpage_freelist
);
1352 static void free_hpage_workfn(struct work_struct
*work
)
1354 struct llist_node
*node
;
1357 node
= llist_del_all(&hpage_freelist
);
1360 page
= container_of((struct address_space
**)node
,
1361 struct page
, mapping
);
1363 __free_huge_page(page
);
1366 static DECLARE_WORK(free_hpage_work
, free_hpage_workfn
);
1368 void free_huge_page(struct page
*page
)
1371 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1375 * Only call schedule_work() if hpage_freelist is previously
1376 * empty. Otherwise, schedule_work() had been called but the
1377 * workfn hasn't retrieved the list yet.
1379 if (llist_add((struct llist_node
*)&page
->mapping
,
1381 schedule_work(&free_hpage_work
);
1385 __free_huge_page(page
);
1388 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1390 INIT_LIST_HEAD(&page
->lru
);
1391 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1392 spin_lock(&hugetlb_lock
);
1393 set_hugetlb_cgroup(page
, NULL
);
1395 h
->nr_huge_pages_node
[nid
]++;
1396 ClearPageHugeFreed(page
);
1397 spin_unlock(&hugetlb_lock
);
1400 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1403 int nr_pages
= 1 << order
;
1404 struct page
*p
= page
+ 1;
1406 /* we rely on prep_new_huge_page to set the destructor */
1407 set_compound_order(page
, order
);
1408 __ClearPageReserved(page
);
1409 __SetPageHead(page
);
1410 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1412 * For gigantic hugepages allocated through bootmem at
1413 * boot, it's safer to be consistent with the not-gigantic
1414 * hugepages and clear the PG_reserved bit from all tail pages
1415 * too. Otherwse drivers using get_user_pages() to access tail
1416 * pages may get the reference counting wrong if they see
1417 * PG_reserved set on a tail page (despite the head page not
1418 * having PG_reserved set). Enforcing this consistency between
1419 * head and tail pages allows drivers to optimize away a check
1420 * on the head page when they need know if put_page() is needed
1421 * after get_user_pages().
1423 __ClearPageReserved(p
);
1424 set_page_count(p
, 0);
1425 set_compound_head(p
, page
);
1427 atomic_set(compound_mapcount_ptr(page
), -1);
1431 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1432 * transparent huge pages. See the PageTransHuge() documentation for more
1435 int PageHuge(struct page
*page
)
1437 if (!PageCompound(page
))
1440 page
= compound_head(page
);
1441 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1443 EXPORT_SYMBOL_GPL(PageHuge
);
1446 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1447 * normal or transparent huge pages.
1449 int PageHeadHuge(struct page
*page_head
)
1451 if (!PageHead(page_head
))
1454 return get_compound_page_dtor(page_head
) == free_huge_page
;
1457 pgoff_t
__basepage_index(struct page
*page
)
1459 struct page
*page_head
= compound_head(page
);
1460 pgoff_t index
= page_index(page_head
);
1461 unsigned long compound_idx
;
1463 if (!PageHuge(page_head
))
1464 return page_index(page
);
1466 if (compound_order(page_head
) >= MAX_ORDER
)
1467 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1469 compound_idx
= page
- page_head
;
1471 return (index
<< compound_order(page_head
)) + compound_idx
;
1474 static struct page
*alloc_buddy_huge_page(struct hstate
*h
,
1475 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1476 nodemask_t
*node_alloc_noretry
)
1478 int order
= huge_page_order(h
);
1480 bool alloc_try_hard
= true;
1483 * By default we always try hard to allocate the page with
1484 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1485 * a loop (to adjust global huge page counts) and previous allocation
1486 * failed, do not continue to try hard on the same node. Use the
1487 * node_alloc_noretry bitmap to manage this state information.
1489 if (node_alloc_noretry
&& node_isset(nid
, *node_alloc_noretry
))
1490 alloc_try_hard
= false;
1491 gfp_mask
|= __GFP_COMP
|__GFP_NOWARN
;
1493 gfp_mask
|= __GFP_RETRY_MAYFAIL
;
1494 if (nid
== NUMA_NO_NODE
)
1495 nid
= numa_mem_id();
1496 page
= __alloc_pages_nodemask(gfp_mask
, order
, nid
, nmask
);
1498 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1500 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1503 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1504 * indicates an overall state change. Clear bit so that we resume
1505 * normal 'try hard' allocations.
1507 if (node_alloc_noretry
&& page
&& !alloc_try_hard
)
1508 node_clear(nid
, *node_alloc_noretry
);
1511 * If we tried hard to get a page but failed, set bit so that
1512 * subsequent attempts will not try as hard until there is an
1513 * overall state change.
1515 if (node_alloc_noretry
&& !page
&& alloc_try_hard
)
1516 node_set(nid
, *node_alloc_noretry
);
1522 * Common helper to allocate a fresh hugetlb page. All specific allocators
1523 * should use this function to get new hugetlb pages
1525 static struct page
*alloc_fresh_huge_page(struct hstate
*h
,
1526 gfp_t gfp_mask
, int nid
, nodemask_t
*nmask
,
1527 nodemask_t
*node_alloc_noretry
)
1531 if (hstate_is_gigantic(h
))
1532 page
= alloc_gigantic_page(h
, gfp_mask
, nid
, nmask
);
1534 page
= alloc_buddy_huge_page(h
, gfp_mask
,
1535 nid
, nmask
, node_alloc_noretry
);
1539 if (hstate_is_gigantic(h
))
1540 prep_compound_gigantic_page(page
, huge_page_order(h
));
1541 prep_new_huge_page(h
, page
, page_to_nid(page
));
1547 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1550 static int alloc_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1551 nodemask_t
*node_alloc_noretry
)
1555 gfp_t gfp_mask
= htlb_alloc_mask(h
) | __GFP_THISNODE
;
1557 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1558 page
= alloc_fresh_huge_page(h
, gfp_mask
, node
, nodes_allowed
,
1559 node_alloc_noretry
);
1567 put_page(page
); /* free it into the hugepage allocator */
1573 * Free huge page from pool from next node to free.
1574 * Attempt to keep persistent huge pages more or less
1575 * balanced over allowed nodes.
1576 * Called with hugetlb_lock locked.
1578 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1584 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1586 * If we're returning unused surplus pages, only examine
1587 * nodes with surplus pages.
1589 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1590 !list_empty(&h
->hugepage_freelists
[node
])) {
1592 list_entry(h
->hugepage_freelists
[node
].next
,
1594 list_del(&page
->lru
);
1595 h
->free_huge_pages
--;
1596 h
->free_huge_pages_node
[node
]--;
1598 h
->surplus_huge_pages
--;
1599 h
->surplus_huge_pages_node
[node
]--;
1601 update_and_free_page(h
, page
);
1611 * Dissolve a given free hugepage into free buddy pages. This function does
1612 * nothing for in-use hugepages and non-hugepages.
1613 * This function returns values like below:
1615 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1616 * (allocated or reserved.)
1617 * 0: successfully dissolved free hugepages or the page is not a
1618 * hugepage (considered as already dissolved)
1620 int dissolve_free_huge_page(struct page
*page
)
1625 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1626 if (!PageHuge(page
))
1629 spin_lock(&hugetlb_lock
);
1630 if (!PageHuge(page
)) {
1635 if (!page_count(page
)) {
1636 struct page
*head
= compound_head(page
);
1637 struct hstate
*h
= page_hstate(head
);
1638 int nid
= page_to_nid(head
);
1639 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0)
1643 * We should make sure that the page is already on the free list
1644 * when it is dissolved.
1646 if (unlikely(!PageHugeFreed(head
))) {
1647 spin_unlock(&hugetlb_lock
);
1651 * Theoretically, we should return -EBUSY when we
1652 * encounter this race. In fact, we have a chance
1653 * to successfully dissolve the page if we do a
1654 * retry. Because the race window is quite small.
1655 * If we seize this opportunity, it is an optimization
1656 * for increasing the success rate of dissolving page.
1662 * Move PageHWPoison flag from head page to the raw error page,
1663 * which makes any subpages rather than the error page reusable.
1665 if (PageHWPoison(head
) && page
!= head
) {
1666 SetPageHWPoison(page
);
1667 ClearPageHWPoison(head
);
1669 list_del(&head
->lru
);
1670 h
->free_huge_pages
--;
1671 h
->free_huge_pages_node
[nid
]--;
1672 h
->max_huge_pages
--;
1673 update_and_free_page(h
, head
);
1677 spin_unlock(&hugetlb_lock
);
1682 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1683 * make specified memory blocks removable from the system.
1684 * Note that this will dissolve a free gigantic hugepage completely, if any
1685 * part of it lies within the given range.
1686 * Also note that if dissolve_free_huge_page() returns with an error, all
1687 * free hugepages that were dissolved before that error are lost.
1689 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1695 if (!hugepages_supported())
1698 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1699 page
= pfn_to_page(pfn
);
1700 rc
= dissolve_free_huge_page(page
);
1709 * Allocates a fresh surplus page from the page allocator.
1711 static struct page
*alloc_surplus_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1712 int nid
, nodemask_t
*nmask
)
1714 struct page
*page
= NULL
;
1716 if (hstate_is_gigantic(h
))
1719 spin_lock(&hugetlb_lock
);
1720 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
)
1722 spin_unlock(&hugetlb_lock
);
1724 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1728 spin_lock(&hugetlb_lock
);
1730 * We could have raced with the pool size change.
1731 * Double check that and simply deallocate the new page
1732 * if we would end up overcommiting the surpluses. Abuse
1733 * temporary page to workaround the nasty free_huge_page
1736 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1737 SetPageHugeTemporary(page
);
1738 spin_unlock(&hugetlb_lock
);
1742 h
->surplus_huge_pages
++;
1743 h
->surplus_huge_pages_node
[page_to_nid(page
)]++;
1747 spin_unlock(&hugetlb_lock
);
1752 struct page
*alloc_migrate_huge_page(struct hstate
*h
, gfp_t gfp_mask
,
1753 int nid
, nodemask_t
*nmask
)
1757 if (hstate_is_gigantic(h
))
1760 page
= alloc_fresh_huge_page(h
, gfp_mask
, nid
, nmask
, NULL
);
1765 * We do not account these pages as surplus because they are only
1766 * temporary and will be released properly on the last reference
1768 SetPageHugeTemporary(page
);
1774 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1777 struct page
*alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1778 struct vm_area_struct
*vma
, unsigned long addr
)
1781 struct mempolicy
*mpol
;
1782 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1784 nodemask_t
*nodemask
;
1786 nid
= huge_node(vma
, addr
, gfp_mask
, &mpol
, &nodemask
);
1787 page
= alloc_surplus_huge_page(h
, gfp_mask
, nid
, nodemask
);
1788 mpol_cond_put(mpol
);
1793 /* page migration callback function */
1794 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1796 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1797 struct page
*page
= NULL
;
1799 if (nid
!= NUMA_NO_NODE
)
1800 gfp_mask
|= __GFP_THISNODE
;
1802 spin_lock(&hugetlb_lock
);
1803 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1804 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, nid
, NULL
);
1805 spin_unlock(&hugetlb_lock
);
1808 page
= alloc_migrate_huge_page(h
, gfp_mask
, nid
, NULL
);
1813 /* page migration callback function */
1814 struct page
*alloc_huge_page_nodemask(struct hstate
*h
, int preferred_nid
,
1817 gfp_t gfp_mask
= htlb_alloc_mask(h
);
1819 spin_lock(&hugetlb_lock
);
1820 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0) {
1823 page
= dequeue_huge_page_nodemask(h
, gfp_mask
, preferred_nid
, nmask
);
1825 spin_unlock(&hugetlb_lock
);
1829 spin_unlock(&hugetlb_lock
);
1831 return alloc_migrate_huge_page(h
, gfp_mask
, preferred_nid
, nmask
);
1834 /* mempolicy aware migration callback */
1835 struct page
*alloc_huge_page_vma(struct hstate
*h
, struct vm_area_struct
*vma
,
1836 unsigned long address
)
1838 struct mempolicy
*mpol
;
1839 nodemask_t
*nodemask
;
1844 gfp_mask
= htlb_alloc_mask(h
);
1845 node
= huge_node(vma
, address
, gfp_mask
, &mpol
, &nodemask
);
1846 page
= alloc_huge_page_nodemask(h
, node
, nodemask
);
1847 mpol_cond_put(mpol
);
1853 * Increase the hugetlb pool such that it can accommodate a reservation
1856 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1858 struct list_head surplus_list
;
1859 struct page
*page
, *tmp
;
1861 int needed
, allocated
;
1862 bool alloc_ok
= true;
1864 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1866 h
->resv_huge_pages
+= delta
;
1871 INIT_LIST_HEAD(&surplus_list
);
1875 spin_unlock(&hugetlb_lock
);
1876 for (i
= 0; i
< needed
; i
++) {
1877 page
= alloc_surplus_huge_page(h
, htlb_alloc_mask(h
),
1878 NUMA_NO_NODE
, NULL
);
1883 list_add(&page
->lru
, &surplus_list
);
1889 * After retaking hugetlb_lock, we need to recalculate 'needed'
1890 * because either resv_huge_pages or free_huge_pages may have changed.
1892 spin_lock(&hugetlb_lock
);
1893 needed
= (h
->resv_huge_pages
+ delta
) -
1894 (h
->free_huge_pages
+ allocated
);
1899 * We were not able to allocate enough pages to
1900 * satisfy the entire reservation so we free what
1901 * we've allocated so far.
1906 * The surplus_list now contains _at_least_ the number of extra pages
1907 * needed to accommodate the reservation. Add the appropriate number
1908 * of pages to the hugetlb pool and free the extras back to the buddy
1909 * allocator. Commit the entire reservation here to prevent another
1910 * process from stealing the pages as they are added to the pool but
1911 * before they are reserved.
1913 needed
+= allocated
;
1914 h
->resv_huge_pages
+= delta
;
1917 /* Free the needed pages to the hugetlb pool */
1918 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1922 * This page is now managed by the hugetlb allocator and has
1923 * no users -- drop the buddy allocator's reference.
1925 put_page_testzero(page
);
1926 VM_BUG_ON_PAGE(page_count(page
), page
);
1927 enqueue_huge_page(h
, page
);
1930 spin_unlock(&hugetlb_lock
);
1932 /* Free unnecessary surplus pages to the buddy allocator */
1933 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1935 spin_lock(&hugetlb_lock
);
1941 * This routine has two main purposes:
1942 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1943 * in unused_resv_pages. This corresponds to the prior adjustments made
1944 * to the associated reservation map.
1945 * 2) Free any unused surplus pages that may have been allocated to satisfy
1946 * the reservation. As many as unused_resv_pages may be freed.
1948 * Called with hugetlb_lock held. However, the lock could be dropped (and
1949 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1950 * we must make sure nobody else can claim pages we are in the process of
1951 * freeing. Do this by ensuring resv_huge_page always is greater than the
1952 * number of huge pages we plan to free when dropping the lock.
1954 static void return_unused_surplus_pages(struct hstate
*h
,
1955 unsigned long unused_resv_pages
)
1957 unsigned long nr_pages
;
1959 /* Cannot return gigantic pages currently */
1960 if (hstate_is_gigantic(h
))
1964 * Part (or even all) of the reservation could have been backed
1965 * by pre-allocated pages. Only free surplus pages.
1967 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1970 * We want to release as many surplus pages as possible, spread
1971 * evenly across all nodes with memory. Iterate across these nodes
1972 * until we can no longer free unreserved surplus pages. This occurs
1973 * when the nodes with surplus pages have no free pages.
1974 * free_pool_huge_page() will balance the the freed pages across the
1975 * on-line nodes with memory and will handle the hstate accounting.
1977 * Note that we decrement resv_huge_pages as we free the pages. If
1978 * we drop the lock, resv_huge_pages will still be sufficiently large
1979 * to cover subsequent pages we may free.
1981 while (nr_pages
--) {
1982 h
->resv_huge_pages
--;
1983 unused_resv_pages
--;
1984 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1986 cond_resched_lock(&hugetlb_lock
);
1990 /* Fully uncommit the reservation */
1991 h
->resv_huge_pages
-= unused_resv_pages
;
1996 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1997 * are used by the huge page allocation routines to manage reservations.
1999 * vma_needs_reservation is called to determine if the huge page at addr
2000 * within the vma has an associated reservation. If a reservation is
2001 * needed, the value 1 is returned. The caller is then responsible for
2002 * managing the global reservation and subpool usage counts. After
2003 * the huge page has been allocated, vma_commit_reservation is called
2004 * to add the page to the reservation map. If the page allocation fails,
2005 * the reservation must be ended instead of committed. vma_end_reservation
2006 * is called in such cases.
2008 * In the normal case, vma_commit_reservation returns the same value
2009 * as the preceding vma_needs_reservation call. The only time this
2010 * is not the case is if a reserve map was changed between calls. It
2011 * is the responsibility of the caller to notice the difference and
2012 * take appropriate action.
2014 * vma_add_reservation is used in error paths where a reservation must
2015 * be restored when a newly allocated huge page must be freed. It is
2016 * to be called after calling vma_needs_reservation to determine if a
2017 * reservation exists.
2019 enum vma_resv_mode
{
2025 static long __vma_reservation_common(struct hstate
*h
,
2026 struct vm_area_struct
*vma
, unsigned long addr
,
2027 enum vma_resv_mode mode
)
2029 struct resv_map
*resv
;
2033 resv
= vma_resv_map(vma
);
2037 idx
= vma_hugecache_offset(h
, vma
, addr
);
2039 case VMA_NEEDS_RESV
:
2040 ret
= region_chg(resv
, idx
, idx
+ 1);
2042 case VMA_COMMIT_RESV
:
2043 ret
= region_add(resv
, idx
, idx
+ 1);
2046 region_abort(resv
, idx
, idx
+ 1);
2050 if (vma
->vm_flags
& VM_MAYSHARE
)
2051 ret
= region_add(resv
, idx
, idx
+ 1);
2053 region_abort(resv
, idx
, idx
+ 1);
2054 ret
= region_del(resv
, idx
, idx
+ 1);
2061 if (vma
->vm_flags
& VM_MAYSHARE
)
2063 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
2065 * In most cases, reserves always exist for private mappings.
2066 * However, a file associated with mapping could have been
2067 * hole punched or truncated after reserves were consumed.
2068 * As subsequent fault on such a range will not use reserves.
2069 * Subtle - The reserve map for private mappings has the
2070 * opposite meaning than that of shared mappings. If NO
2071 * entry is in the reserve map, it means a reservation exists.
2072 * If an entry exists in the reserve map, it means the
2073 * reservation has already been consumed. As a result, the
2074 * return value of this routine is the opposite of the
2075 * value returned from reserve map manipulation routines above.
2083 return ret
< 0 ? ret
: 0;
2086 static long vma_needs_reservation(struct hstate
*h
,
2087 struct vm_area_struct
*vma
, unsigned long addr
)
2089 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
2092 static long vma_commit_reservation(struct hstate
*h
,
2093 struct vm_area_struct
*vma
, unsigned long addr
)
2095 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
2098 static void vma_end_reservation(struct hstate
*h
,
2099 struct vm_area_struct
*vma
, unsigned long addr
)
2101 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
2104 static long vma_add_reservation(struct hstate
*h
,
2105 struct vm_area_struct
*vma
, unsigned long addr
)
2107 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
2111 * This routine is called to restore a reservation on error paths. In the
2112 * specific error paths, a huge page was allocated (via alloc_huge_page)
2113 * and is about to be freed. If a reservation for the page existed,
2114 * alloc_huge_page would have consumed the reservation and set PagePrivate
2115 * in the newly allocated page. When the page is freed via free_huge_page,
2116 * the global reservation count will be incremented if PagePrivate is set.
2117 * However, free_huge_page can not adjust the reserve map. Adjust the
2118 * reserve map here to be consistent with global reserve count adjustments
2119 * to be made by free_huge_page.
2121 static void restore_reserve_on_error(struct hstate
*h
,
2122 struct vm_area_struct
*vma
, unsigned long address
,
2125 if (unlikely(PagePrivate(page
))) {
2126 long rc
= vma_needs_reservation(h
, vma
, address
);
2128 if (unlikely(rc
< 0)) {
2130 * Rare out of memory condition in reserve map
2131 * manipulation. Clear PagePrivate so that
2132 * global reserve count will not be incremented
2133 * by free_huge_page. This will make it appear
2134 * as though the reservation for this page was
2135 * consumed. This may prevent the task from
2136 * faulting in the page at a later time. This
2137 * is better than inconsistent global huge page
2138 * accounting of reserve counts.
2140 ClearPagePrivate(page
);
2142 rc
= vma_add_reservation(h
, vma
, address
);
2143 if (unlikely(rc
< 0))
2145 * See above comment about rare out of
2148 ClearPagePrivate(page
);
2150 vma_end_reservation(h
, vma
, address
);
2154 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
2155 unsigned long addr
, int avoid_reserve
)
2157 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2158 struct hstate
*h
= hstate_vma(vma
);
2160 long map_chg
, map_commit
;
2163 struct hugetlb_cgroup
*h_cg
;
2165 idx
= hstate_index(h
);
2167 * Examine the region/reserve map to determine if the process
2168 * has a reservation for the page to be allocated. A return
2169 * code of zero indicates a reservation exists (no change).
2171 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2173 return ERR_PTR(-ENOMEM
);
2176 * Processes that did not create the mapping will have no
2177 * reserves as indicated by the region/reserve map. Check
2178 * that the allocation will not exceed the subpool limit.
2179 * Allocations for MAP_NORESERVE mappings also need to be
2180 * checked against any subpool limit.
2182 if (map_chg
|| avoid_reserve
) {
2183 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2185 vma_end_reservation(h
, vma
, addr
);
2186 return ERR_PTR(-ENOSPC
);
2190 * Even though there was no reservation in the region/reserve
2191 * map, there could be reservations associated with the
2192 * subpool that can be used. This would be indicated if the
2193 * return value of hugepage_subpool_get_pages() is zero.
2194 * However, if avoid_reserve is specified we still avoid even
2195 * the subpool reservations.
2201 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2203 goto out_subpool_put
;
2205 spin_lock(&hugetlb_lock
);
2207 * glb_chg is passed to indicate whether or not a page must be taken
2208 * from the global free pool (global change). gbl_chg == 0 indicates
2209 * a reservation exists for the allocation.
2211 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2213 spin_unlock(&hugetlb_lock
);
2214 page
= alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2216 goto out_uncharge_cgroup
;
2217 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2218 SetPagePrivate(page
);
2219 h
->resv_huge_pages
--;
2221 spin_lock(&hugetlb_lock
);
2222 list_move(&page
->lru
, &h
->hugepage_activelist
);
2225 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2226 spin_unlock(&hugetlb_lock
);
2228 set_page_private(page
, (unsigned long)spool
);
2230 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2231 if (unlikely(map_chg
> map_commit
)) {
2233 * The page was added to the reservation map between
2234 * vma_needs_reservation and vma_commit_reservation.
2235 * This indicates a race with hugetlb_reserve_pages.
2236 * Adjust for the subpool count incremented above AND
2237 * in hugetlb_reserve_pages for the same page. Also,
2238 * the reservation count added in hugetlb_reserve_pages
2239 * no longer applies.
2243 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2244 hugetlb_acct_memory(h
, -rsv_adjust
);
2248 out_uncharge_cgroup
:
2249 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2251 if (map_chg
|| avoid_reserve
)
2252 hugepage_subpool_put_pages(spool
, 1);
2253 vma_end_reservation(h
, vma
, addr
);
2254 return ERR_PTR(-ENOSPC
);
2257 int alloc_bootmem_huge_page(struct hstate
*h
)
2258 __attribute__ ((weak
, alias("__alloc_bootmem_huge_page")));
2259 int __alloc_bootmem_huge_page(struct hstate
*h
)
2261 struct huge_bootmem_page
*m
;
2264 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2267 addr
= memblock_alloc_try_nid_raw(
2268 huge_page_size(h
), huge_page_size(h
),
2269 0, MEMBLOCK_ALLOC_ACCESSIBLE
, node
);
2272 * Use the beginning of the huge page to store the
2273 * huge_bootmem_page struct (until gather_bootmem
2274 * puts them into the mem_map).
2283 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2284 /* Put them into a private list first because mem_map is not up yet */
2285 INIT_LIST_HEAD(&m
->list
);
2286 list_add(&m
->list
, &huge_boot_pages
);
2291 static void __init
prep_compound_huge_page(struct page
*page
,
2294 if (unlikely(order
> (MAX_ORDER
- 1)))
2295 prep_compound_gigantic_page(page
, order
);
2297 prep_compound_page(page
, order
);
2300 /* Put bootmem huge pages into the standard lists after mem_map is up */
2301 static void __init
gather_bootmem_prealloc(void)
2303 struct huge_bootmem_page
*m
;
2305 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2306 struct page
*page
= virt_to_page(m
);
2307 struct hstate
*h
= m
->hstate
;
2309 WARN_ON(page_count(page
) != 1);
2310 prep_compound_huge_page(page
, h
->order
);
2311 WARN_ON(PageReserved(page
));
2312 prep_new_huge_page(h
, page
, page_to_nid(page
));
2313 put_page(page
); /* free it into the hugepage allocator */
2316 * If we had gigantic hugepages allocated at boot time, we need
2317 * to restore the 'stolen' pages to totalram_pages in order to
2318 * fix confusing memory reports from free(1) and another
2319 * side-effects, like CommitLimit going negative.
2321 if (hstate_is_gigantic(h
))
2322 adjust_managed_page_count(page
, 1 << h
->order
);
2327 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2330 nodemask_t
*node_alloc_noretry
;
2332 if (!hstate_is_gigantic(h
)) {
2334 * Bit mask controlling how hard we retry per-node allocations.
2335 * Ignore errors as lower level routines can deal with
2336 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2337 * time, we are likely in bigger trouble.
2339 node_alloc_noretry
= kmalloc(sizeof(*node_alloc_noretry
),
2342 /* allocations done at boot time */
2343 node_alloc_noretry
= NULL
;
2346 /* bit mask controlling how hard we retry per-node allocations */
2347 if (node_alloc_noretry
)
2348 nodes_clear(*node_alloc_noretry
);
2350 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2351 if (hstate_is_gigantic(h
)) {
2352 if (!alloc_bootmem_huge_page(h
))
2354 } else if (!alloc_pool_huge_page(h
,
2355 &node_states
[N_MEMORY
],
2356 node_alloc_noretry
))
2360 if (i
< h
->max_huge_pages
) {
2363 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2364 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2365 h
->max_huge_pages
, buf
, i
);
2366 h
->max_huge_pages
= i
;
2369 kfree(node_alloc_noretry
);
2372 static void __init
hugetlb_init_hstates(void)
2376 for_each_hstate(h
) {
2377 if (minimum_order
> huge_page_order(h
))
2378 minimum_order
= huge_page_order(h
);
2380 /* oversize hugepages were init'ed in early boot */
2381 if (!hstate_is_gigantic(h
))
2382 hugetlb_hstate_alloc_pages(h
);
2384 VM_BUG_ON(minimum_order
== UINT_MAX
);
2387 static void __init
report_hugepages(void)
2391 for_each_hstate(h
) {
2394 string_get_size(huge_page_size(h
), 1, STRING_UNITS_2
, buf
, 32);
2395 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2396 buf
, h
->free_huge_pages
);
2400 #ifdef CONFIG_HIGHMEM
2401 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2402 nodemask_t
*nodes_allowed
)
2406 if (hstate_is_gigantic(h
))
2409 for_each_node_mask(i
, *nodes_allowed
) {
2410 struct page
*page
, *next
;
2411 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2412 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2413 if (count
>= h
->nr_huge_pages
)
2415 if (PageHighMem(page
))
2417 list_del(&page
->lru
);
2418 update_and_free_page(h
, page
);
2419 h
->free_huge_pages
--;
2420 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2425 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2426 nodemask_t
*nodes_allowed
)
2432 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2433 * balanced by operating on them in a round-robin fashion.
2434 * Returns 1 if an adjustment was made.
2436 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2441 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2444 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2445 if (h
->surplus_huge_pages_node
[node
])
2449 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2450 if (h
->surplus_huge_pages_node
[node
] <
2451 h
->nr_huge_pages_node
[node
])
2458 h
->surplus_huge_pages
+= delta
;
2459 h
->surplus_huge_pages_node
[node
] += delta
;
2463 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2464 static int set_max_huge_pages(struct hstate
*h
, unsigned long count
, int nid
,
2465 nodemask_t
*nodes_allowed
)
2467 unsigned long min_count
, ret
;
2468 NODEMASK_ALLOC(nodemask_t
, node_alloc_noretry
, GFP_KERNEL
);
2471 * Bit mask controlling how hard we retry per-node allocations.
2472 * If we can not allocate the bit mask, do not attempt to allocate
2473 * the requested huge pages.
2475 if (node_alloc_noretry
)
2476 nodes_clear(*node_alloc_noretry
);
2480 spin_lock(&hugetlb_lock
);
2483 * Check for a node specific request.
2484 * Changing node specific huge page count may require a corresponding
2485 * change to the global count. In any case, the passed node mask
2486 * (nodes_allowed) will restrict alloc/free to the specified node.
2488 if (nid
!= NUMA_NO_NODE
) {
2489 unsigned long old_count
= count
;
2491 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2493 * User may have specified a large count value which caused the
2494 * above calculation to overflow. In this case, they wanted
2495 * to allocate as many huge pages as possible. Set count to
2496 * largest possible value to align with their intention.
2498 if (count
< old_count
)
2503 * Gigantic pages runtime allocation depend on the capability for large
2504 * page range allocation.
2505 * If the system does not provide this feature, return an error when
2506 * the user tries to allocate gigantic pages but let the user free the
2507 * boottime allocated gigantic pages.
2509 if (hstate_is_gigantic(h
) && !IS_ENABLED(CONFIG_CONTIG_ALLOC
)) {
2510 if (count
> persistent_huge_pages(h
)) {
2511 spin_unlock(&hugetlb_lock
);
2512 NODEMASK_FREE(node_alloc_noretry
);
2515 /* Fall through to decrease pool */
2519 * Increase the pool size
2520 * First take pages out of surplus state. Then make up the
2521 * remaining difference by allocating fresh huge pages.
2523 * We might race with alloc_surplus_huge_page() here and be unable
2524 * to convert a surplus huge page to a normal huge page. That is
2525 * not critical, though, it just means the overall size of the
2526 * pool might be one hugepage larger than it needs to be, but
2527 * within all the constraints specified by the sysctls.
2529 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2530 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2534 while (count
> persistent_huge_pages(h
)) {
2536 * If this allocation races such that we no longer need the
2537 * page, free_huge_page will handle it by freeing the page
2538 * and reducing the surplus.
2540 spin_unlock(&hugetlb_lock
);
2542 /* yield cpu to avoid soft lockup */
2545 ret
= alloc_pool_huge_page(h
, nodes_allowed
,
2546 node_alloc_noretry
);
2547 spin_lock(&hugetlb_lock
);
2551 /* Bail for signals. Probably ctrl-c from user */
2552 if (signal_pending(current
))
2557 * Decrease the pool size
2558 * First return free pages to the buddy allocator (being careful
2559 * to keep enough around to satisfy reservations). Then place
2560 * pages into surplus state as needed so the pool will shrink
2561 * to the desired size as pages become free.
2563 * By placing pages into the surplus state independent of the
2564 * overcommit value, we are allowing the surplus pool size to
2565 * exceed overcommit. There are few sane options here. Since
2566 * alloc_surplus_huge_page() is checking the global counter,
2567 * though, we'll note that we're not allowed to exceed surplus
2568 * and won't grow the pool anywhere else. Not until one of the
2569 * sysctls are changed, or the surplus pages go out of use.
2571 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2572 min_count
= max(count
, min_count
);
2573 try_to_free_low(h
, min_count
, nodes_allowed
);
2574 while (min_count
< persistent_huge_pages(h
)) {
2575 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2577 cond_resched_lock(&hugetlb_lock
);
2579 while (count
< persistent_huge_pages(h
)) {
2580 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2584 h
->max_huge_pages
= persistent_huge_pages(h
);
2585 spin_unlock(&hugetlb_lock
);
2587 NODEMASK_FREE(node_alloc_noretry
);
2592 #define HSTATE_ATTR_RO(_name) \
2593 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2595 #define HSTATE_ATTR(_name) \
2596 static struct kobj_attribute _name##_attr = \
2597 __ATTR(_name, 0644, _name##_show, _name##_store)
2599 static struct kobject
*hugepages_kobj
;
2600 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2602 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2604 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2608 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2609 if (hstate_kobjs
[i
] == kobj
) {
2611 *nidp
= NUMA_NO_NODE
;
2615 return kobj_to_node_hstate(kobj
, nidp
);
2618 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2619 struct kobj_attribute
*attr
, char *buf
)
2622 unsigned long nr_huge_pages
;
2625 h
= kobj_to_hstate(kobj
, &nid
);
2626 if (nid
== NUMA_NO_NODE
)
2627 nr_huge_pages
= h
->nr_huge_pages
;
2629 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2631 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2634 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2635 struct hstate
*h
, int nid
,
2636 unsigned long count
, size_t len
)
2639 nodemask_t nodes_allowed
, *n_mask
;
2641 if (hstate_is_gigantic(h
) && !gigantic_page_runtime_supported())
2644 if (nid
== NUMA_NO_NODE
) {
2646 * global hstate attribute
2648 if (!(obey_mempolicy
&&
2649 init_nodemask_of_mempolicy(&nodes_allowed
)))
2650 n_mask
= &node_states
[N_MEMORY
];
2652 n_mask
= &nodes_allowed
;
2655 * Node specific request. count adjustment happens in
2656 * set_max_huge_pages() after acquiring hugetlb_lock.
2658 init_nodemask_of_node(&nodes_allowed
, nid
);
2659 n_mask
= &nodes_allowed
;
2662 err
= set_max_huge_pages(h
, count
, nid
, n_mask
);
2664 return err
? err
: len
;
2667 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2668 struct kobject
*kobj
, const char *buf
,
2672 unsigned long count
;
2676 err
= kstrtoul(buf
, 10, &count
);
2680 h
= kobj_to_hstate(kobj
, &nid
);
2681 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2684 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2685 struct kobj_attribute
*attr
, char *buf
)
2687 return nr_hugepages_show_common(kobj
, attr
, buf
);
2690 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2691 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2693 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2695 HSTATE_ATTR(nr_hugepages
);
2700 * hstate attribute for optionally mempolicy-based constraint on persistent
2701 * huge page alloc/free.
2703 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2704 struct kobj_attribute
*attr
, char *buf
)
2706 return nr_hugepages_show_common(kobj
, attr
, buf
);
2709 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2710 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2712 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2714 HSTATE_ATTR(nr_hugepages_mempolicy
);
2718 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2719 struct kobj_attribute
*attr
, char *buf
)
2721 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2722 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2725 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2726 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2729 unsigned long input
;
2730 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2732 if (hstate_is_gigantic(h
))
2735 err
= kstrtoul(buf
, 10, &input
);
2739 spin_lock(&hugetlb_lock
);
2740 h
->nr_overcommit_huge_pages
= input
;
2741 spin_unlock(&hugetlb_lock
);
2745 HSTATE_ATTR(nr_overcommit_hugepages
);
2747 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2748 struct kobj_attribute
*attr
, char *buf
)
2751 unsigned long free_huge_pages
;
2754 h
= kobj_to_hstate(kobj
, &nid
);
2755 if (nid
== NUMA_NO_NODE
)
2756 free_huge_pages
= h
->free_huge_pages
;
2758 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2760 return sprintf(buf
, "%lu\n", free_huge_pages
);
2762 HSTATE_ATTR_RO(free_hugepages
);
2764 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2765 struct kobj_attribute
*attr
, char *buf
)
2767 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2768 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2770 HSTATE_ATTR_RO(resv_hugepages
);
2772 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2773 struct kobj_attribute
*attr
, char *buf
)
2776 unsigned long surplus_huge_pages
;
2779 h
= kobj_to_hstate(kobj
, &nid
);
2780 if (nid
== NUMA_NO_NODE
)
2781 surplus_huge_pages
= h
->surplus_huge_pages
;
2783 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2785 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2787 HSTATE_ATTR_RO(surplus_hugepages
);
2789 static struct attribute
*hstate_attrs
[] = {
2790 &nr_hugepages_attr
.attr
,
2791 &nr_overcommit_hugepages_attr
.attr
,
2792 &free_hugepages_attr
.attr
,
2793 &resv_hugepages_attr
.attr
,
2794 &surplus_hugepages_attr
.attr
,
2796 &nr_hugepages_mempolicy_attr
.attr
,
2801 static const struct attribute_group hstate_attr_group
= {
2802 .attrs
= hstate_attrs
,
2805 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2806 struct kobject
**hstate_kobjs
,
2807 const struct attribute_group
*hstate_attr_group
)
2810 int hi
= hstate_index(h
);
2812 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2813 if (!hstate_kobjs
[hi
])
2816 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2818 kobject_put(hstate_kobjs
[hi
]);
2819 hstate_kobjs
[hi
] = NULL
;
2825 static void __init
hugetlb_sysfs_init(void)
2830 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2831 if (!hugepages_kobj
)
2834 for_each_hstate(h
) {
2835 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2836 hstate_kobjs
, &hstate_attr_group
);
2838 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2845 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2846 * with node devices in node_devices[] using a parallel array. The array
2847 * index of a node device or _hstate == node id.
2848 * This is here to avoid any static dependency of the node device driver, in
2849 * the base kernel, on the hugetlb module.
2851 struct node_hstate
{
2852 struct kobject
*hugepages_kobj
;
2853 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2855 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2858 * A subset of global hstate attributes for node devices
2860 static struct attribute
*per_node_hstate_attrs
[] = {
2861 &nr_hugepages_attr
.attr
,
2862 &free_hugepages_attr
.attr
,
2863 &surplus_hugepages_attr
.attr
,
2867 static const struct attribute_group per_node_hstate_attr_group
= {
2868 .attrs
= per_node_hstate_attrs
,
2872 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2873 * Returns node id via non-NULL nidp.
2875 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2879 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2880 struct node_hstate
*nhs
= &node_hstates
[nid
];
2882 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2883 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2895 * Unregister hstate attributes from a single node device.
2896 * No-op if no hstate attributes attached.
2898 static void hugetlb_unregister_node(struct node
*node
)
2901 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2903 if (!nhs
->hugepages_kobj
)
2904 return; /* no hstate attributes */
2906 for_each_hstate(h
) {
2907 int idx
= hstate_index(h
);
2908 if (nhs
->hstate_kobjs
[idx
]) {
2909 kobject_put(nhs
->hstate_kobjs
[idx
]);
2910 nhs
->hstate_kobjs
[idx
] = NULL
;
2914 kobject_put(nhs
->hugepages_kobj
);
2915 nhs
->hugepages_kobj
= NULL
;
2920 * Register hstate attributes for a single node device.
2921 * No-op if attributes already registered.
2923 static void hugetlb_register_node(struct node
*node
)
2926 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2929 if (nhs
->hugepages_kobj
)
2930 return; /* already allocated */
2932 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2934 if (!nhs
->hugepages_kobj
)
2937 for_each_hstate(h
) {
2938 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2940 &per_node_hstate_attr_group
);
2942 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2943 h
->name
, node
->dev
.id
);
2944 hugetlb_unregister_node(node
);
2951 * hugetlb init time: register hstate attributes for all registered node
2952 * devices of nodes that have memory. All on-line nodes should have
2953 * registered their associated device by this time.
2955 static void __init
hugetlb_register_all_nodes(void)
2959 for_each_node_state(nid
, N_MEMORY
) {
2960 struct node
*node
= node_devices
[nid
];
2961 if (node
->dev
.id
== nid
)
2962 hugetlb_register_node(node
);
2966 * Let the node device driver know we're here so it can
2967 * [un]register hstate attributes on node hotplug.
2969 register_hugetlbfs_with_node(hugetlb_register_node
,
2970 hugetlb_unregister_node
);
2972 #else /* !CONFIG_NUMA */
2974 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2982 static void hugetlb_register_all_nodes(void) { }
2986 static int __init
hugetlb_init(void)
2990 if (!hugepages_supported())
2993 if (!size_to_hstate(default_hstate_size
)) {
2994 if (default_hstate_size
!= 0) {
2995 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2996 default_hstate_size
, HPAGE_SIZE
);
2999 default_hstate_size
= HPAGE_SIZE
;
3000 if (!size_to_hstate(default_hstate_size
))
3001 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
3003 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
3004 if (default_hstate_max_huge_pages
) {
3005 if (!default_hstate
.max_huge_pages
)
3006 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
3009 hugetlb_init_hstates();
3010 gather_bootmem_prealloc();
3013 hugetlb_sysfs_init();
3014 hugetlb_register_all_nodes();
3015 hugetlb_cgroup_file_init();
3018 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
3020 num_fault_mutexes
= 1;
3022 hugetlb_fault_mutex_table
=
3023 kmalloc_array(num_fault_mutexes
, sizeof(struct mutex
),
3025 BUG_ON(!hugetlb_fault_mutex_table
);
3027 for (i
= 0; i
< num_fault_mutexes
; i
++)
3028 mutex_init(&hugetlb_fault_mutex_table
[i
]);
3031 subsys_initcall(hugetlb_init
);
3033 /* Should be called on processing a hugepagesz=... option */
3034 void __init
hugetlb_bad_size(void)
3036 parsed_valid_hugepagesz
= false;
3039 void __init
hugetlb_add_hstate(unsigned int order
)
3044 if (size_to_hstate(PAGE_SIZE
<< order
)) {
3045 pr_warn("hugepagesz= specified twice, ignoring\n");
3048 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
3050 h
= &hstates
[hugetlb_max_hstate
++];
3052 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
3053 h
->nr_huge_pages
= 0;
3054 h
->free_huge_pages
= 0;
3055 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
3056 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
3057 INIT_LIST_HEAD(&h
->hugepage_activelist
);
3058 h
->next_nid_to_alloc
= first_memory_node
;
3059 h
->next_nid_to_free
= first_memory_node
;
3060 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
3061 huge_page_size(h
)/1024);
3066 static int __init
hugetlb_nrpages_setup(char *s
)
3069 static unsigned long *last_mhp
;
3071 if (!parsed_valid_hugepagesz
) {
3072 pr_warn("hugepages = %s preceded by "
3073 "an unsupported hugepagesz, ignoring\n", s
);
3074 parsed_valid_hugepagesz
= true;
3078 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
3079 * so this hugepages= parameter goes to the "default hstate".
3081 else if (!hugetlb_max_hstate
)
3082 mhp
= &default_hstate_max_huge_pages
;
3084 mhp
= &parsed_hstate
->max_huge_pages
;
3086 if (mhp
== last_mhp
) {
3087 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
3091 if (sscanf(s
, "%lu", mhp
) <= 0)
3095 * Global state is always initialized later in hugetlb_init.
3096 * But we need to allocate >= MAX_ORDER hstates here early to still
3097 * use the bootmem allocator.
3099 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
3100 hugetlb_hstate_alloc_pages(parsed_hstate
);
3106 __setup("hugepages=", hugetlb_nrpages_setup
);
3108 static int __init
hugetlb_default_setup(char *s
)
3110 default_hstate_size
= memparse(s
, &s
);
3113 __setup("default_hugepagesz=", hugetlb_default_setup
);
3115 static unsigned int cpuset_mems_nr(unsigned int *array
)
3118 unsigned int nr
= 0;
3120 for_each_node_mask(node
, cpuset_current_mems_allowed
)
3126 #ifdef CONFIG_SYSCTL
3127 static int proc_hugetlb_doulongvec_minmax(struct ctl_table
*table
, int write
,
3128 void *buffer
, size_t *length
,
3129 loff_t
*ppos
, unsigned long *out
)
3131 struct ctl_table dup_table
;
3134 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3135 * can duplicate the @table and alter the duplicate of it.
3138 dup_table
.data
= out
;
3140 return proc_doulongvec_minmax(&dup_table
, write
, buffer
, length
, ppos
);
3143 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
3144 struct ctl_table
*table
, int write
,
3145 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3147 struct hstate
*h
= &default_hstate
;
3148 unsigned long tmp
= h
->max_huge_pages
;
3151 if (!hugepages_supported())
3154 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3160 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
3161 NUMA_NO_NODE
, tmp
, *length
);
3166 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
3167 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3170 return hugetlb_sysctl_handler_common(false, table
, write
,
3171 buffer
, length
, ppos
);
3175 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
3176 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
3178 return hugetlb_sysctl_handler_common(true, table
, write
,
3179 buffer
, length
, ppos
);
3181 #endif /* CONFIG_NUMA */
3183 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
3184 void __user
*buffer
,
3185 size_t *length
, loff_t
*ppos
)
3187 struct hstate
*h
= &default_hstate
;
3191 if (!hugepages_supported())
3194 tmp
= h
->nr_overcommit_huge_pages
;
3196 if (write
&& hstate_is_gigantic(h
))
3199 ret
= proc_hugetlb_doulongvec_minmax(table
, write
, buffer
, length
, ppos
,
3205 spin_lock(&hugetlb_lock
);
3206 h
->nr_overcommit_huge_pages
= tmp
;
3207 spin_unlock(&hugetlb_lock
);
3213 #endif /* CONFIG_SYSCTL */
3215 void hugetlb_report_meminfo(struct seq_file
*m
)
3218 unsigned long total
= 0;
3220 if (!hugepages_supported())
3223 for_each_hstate(h
) {
3224 unsigned long count
= h
->nr_huge_pages
;
3226 total
+= (PAGE_SIZE
<< huge_page_order(h
)) * count
;
3228 if (h
== &default_hstate
)
3230 "HugePages_Total: %5lu\n"
3231 "HugePages_Free: %5lu\n"
3232 "HugePages_Rsvd: %5lu\n"
3233 "HugePages_Surp: %5lu\n"
3234 "Hugepagesize: %8lu kB\n",
3238 h
->surplus_huge_pages
,
3239 (PAGE_SIZE
<< huge_page_order(h
)) / 1024);
3242 seq_printf(m
, "Hugetlb: %8lu kB\n", total
/ 1024);
3245 int hugetlb_report_node_meminfo(int nid
, char *buf
)
3247 struct hstate
*h
= &default_hstate
;
3248 if (!hugepages_supported())
3251 "Node %d HugePages_Total: %5u\n"
3252 "Node %d HugePages_Free: %5u\n"
3253 "Node %d HugePages_Surp: %5u\n",
3254 nid
, h
->nr_huge_pages_node
[nid
],
3255 nid
, h
->free_huge_pages_node
[nid
],
3256 nid
, h
->surplus_huge_pages_node
[nid
]);
3259 void hugetlb_show_meminfo(void)
3264 if (!hugepages_supported())
3267 for_each_node_state(nid
, N_MEMORY
)
3269 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3271 h
->nr_huge_pages_node
[nid
],
3272 h
->free_huge_pages_node
[nid
],
3273 h
->surplus_huge_pages_node
[nid
],
3274 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3277 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3279 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3280 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3283 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3284 unsigned long hugetlb_total_pages(void)
3287 unsigned long nr_total_pages
= 0;
3290 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3291 return nr_total_pages
;
3294 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3298 spin_lock(&hugetlb_lock
);
3300 * When cpuset is configured, it breaks the strict hugetlb page
3301 * reservation as the accounting is done on a global variable. Such
3302 * reservation is completely rubbish in the presence of cpuset because
3303 * the reservation is not checked against page availability for the
3304 * current cpuset. Application can still potentially OOM'ed by kernel
3305 * with lack of free htlb page in cpuset that the task is in.
3306 * Attempt to enforce strict accounting with cpuset is almost
3307 * impossible (or too ugly) because cpuset is too fluid that
3308 * task or memory node can be dynamically moved between cpusets.
3310 * The change of semantics for shared hugetlb mapping with cpuset is
3311 * undesirable. However, in order to preserve some of the semantics,
3312 * we fall back to check against current free page availability as
3313 * a best attempt and hopefully to minimize the impact of changing
3314 * semantics that cpuset has.
3317 if (gather_surplus_pages(h
, delta
) < 0)
3320 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
3321 return_unused_surplus_pages(h
, delta
);
3328 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3331 spin_unlock(&hugetlb_lock
);
3335 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3337 struct resv_map
*resv
= vma_resv_map(vma
);
3340 * This new VMA should share its siblings reservation map if present.
3341 * The VMA will only ever have a valid reservation map pointer where
3342 * it is being copied for another still existing VMA. As that VMA
3343 * has a reference to the reservation map it cannot disappear until
3344 * after this open call completes. It is therefore safe to take a
3345 * new reference here without additional locking.
3347 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3348 kref_get(&resv
->refs
);
3351 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3353 struct hstate
*h
= hstate_vma(vma
);
3354 struct resv_map
*resv
= vma_resv_map(vma
);
3355 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3356 unsigned long reserve
, start
, end
;
3359 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3362 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3363 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3365 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3367 kref_put(&resv
->refs
, resv_map_release
);
3371 * Decrement reserve counts. The global reserve count may be
3372 * adjusted if the subpool has a minimum size.
3374 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3375 hugetlb_acct_memory(h
, -gbl_reserve
);
3379 static int hugetlb_vm_op_split(struct vm_area_struct
*vma
, unsigned long addr
)
3381 if (addr
& ~(huge_page_mask(hstate_vma(vma
))))
3386 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct
*vma
)
3388 struct hstate
*hstate
= hstate_vma(vma
);
3390 return 1UL << huge_page_shift(hstate
);
3394 * We cannot handle pagefaults against hugetlb pages at all. They cause
3395 * handle_mm_fault() to try to instantiate regular-sized pages in the
3396 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3399 static vm_fault_t
hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3406 * When a new function is introduced to vm_operations_struct and added
3407 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3408 * This is because under System V memory model, mappings created via
3409 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3410 * their original vm_ops are overwritten with shm_vm_ops.
3412 const struct vm_operations_struct hugetlb_vm_ops
= {
3413 .fault
= hugetlb_vm_op_fault
,
3414 .open
= hugetlb_vm_op_open
,
3415 .close
= hugetlb_vm_op_close
,
3416 .split
= hugetlb_vm_op_split
,
3417 .pagesize
= hugetlb_vm_op_pagesize
,
3420 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3426 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3427 vma
->vm_page_prot
)));
3429 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3430 vma
->vm_page_prot
));
3432 entry
= pte_mkyoung(entry
);
3433 entry
= pte_mkhuge(entry
);
3434 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3439 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3440 unsigned long address
, pte_t
*ptep
)
3444 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3445 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3446 update_mmu_cache(vma
, address
, ptep
);
3449 bool is_hugetlb_entry_migration(pte_t pte
)
3453 if (huge_pte_none(pte
) || pte_present(pte
))
3455 swp
= pte_to_swp_entry(pte
);
3456 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3462 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3466 if (huge_pte_none(pte
) || pte_present(pte
))
3468 swp
= pte_to_swp_entry(pte
);
3469 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3475 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3476 struct vm_area_struct
*vma
)
3478 pte_t
*src_pte
, *dst_pte
, entry
, dst_entry
;
3479 struct page
*ptepage
;
3482 struct hstate
*h
= hstate_vma(vma
);
3483 unsigned long sz
= huge_page_size(h
);
3484 struct mmu_notifier_range range
;
3487 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3490 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, src
,
3493 mmu_notifier_invalidate_range_start(&range
);
3496 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3497 spinlock_t
*src_ptl
, *dst_ptl
;
3498 src_pte
= huge_pte_offset(src
, addr
, sz
);
3501 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3508 * If the pagetables are shared don't copy or take references.
3509 * dst_pte == src_pte is the common case of src/dest sharing.
3511 * However, src could have 'unshared' and dst shares with
3512 * another vma. If dst_pte !none, this implies sharing.
3513 * Check here before taking page table lock, and once again
3514 * after taking the lock below.
3516 dst_entry
= huge_ptep_get(dst_pte
);
3517 if ((dst_pte
== src_pte
) || !huge_pte_none(dst_entry
))
3520 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3521 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3522 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3523 entry
= huge_ptep_get(src_pte
);
3524 dst_entry
= huge_ptep_get(dst_pte
);
3525 if (huge_pte_none(entry
) || !huge_pte_none(dst_entry
)) {
3527 * Skip if src entry none. Also, skip in the
3528 * unlikely case dst entry !none as this implies
3529 * sharing with another vma.
3532 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3533 is_hugetlb_entry_hwpoisoned(entry
))) {
3534 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3536 if (is_write_migration_entry(swp_entry
) && cow
) {
3538 * COW mappings require pages in both
3539 * parent and child to be set to read.
3541 make_migration_entry_read(&swp_entry
);
3542 entry
= swp_entry_to_pte(swp_entry
);
3543 set_huge_swap_pte_at(src
, addr
, src_pte
,
3546 set_huge_swap_pte_at(dst
, addr
, dst_pte
, entry
, sz
);
3550 * No need to notify as we are downgrading page
3551 * table protection not changing it to point
3554 * See Documentation/vm/mmu_notifier.rst
3556 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3558 entry
= huge_ptep_get(src_pte
);
3559 ptepage
= pte_page(entry
);
3561 page_dup_rmap(ptepage
, true);
3562 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3563 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3565 spin_unlock(src_ptl
);
3566 spin_unlock(dst_ptl
);
3570 mmu_notifier_invalidate_range_end(&range
);
3575 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3576 unsigned long start
, unsigned long end
,
3577 struct page
*ref_page
)
3579 struct mm_struct
*mm
= vma
->vm_mm
;
3580 unsigned long address
;
3585 struct hstate
*h
= hstate_vma(vma
);
3586 unsigned long sz
= huge_page_size(h
);
3587 struct mmu_notifier_range range
;
3589 WARN_ON(!is_vm_hugetlb_page(vma
));
3590 BUG_ON(start
& ~huge_page_mask(h
));
3591 BUG_ON(end
& ~huge_page_mask(h
));
3594 * This is a hugetlb vma, all the pte entries should point
3597 tlb_change_page_size(tlb
, sz
);
3598 tlb_start_vma(tlb
, vma
);
3601 * If sharing possible, alert mmu notifiers of worst case.
3603 mmu_notifier_range_init(&range
, MMU_NOTIFY_UNMAP
, 0, vma
, mm
, start
,
3605 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
3606 mmu_notifier_invalidate_range_start(&range
);
3608 for (; address
< end
; address
+= sz
) {
3609 ptep
= huge_pte_offset(mm
, address
, sz
);
3613 ptl
= huge_pte_lock(h
, mm
, ptep
);
3614 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3617 * We just unmapped a page of PMDs by clearing a PUD.
3618 * The caller's TLB flush range should cover this area.
3623 pte
= huge_ptep_get(ptep
);
3624 if (huge_pte_none(pte
)) {
3630 * Migrating hugepage or HWPoisoned hugepage is already
3631 * unmapped and its refcount is dropped, so just clear pte here.
3633 if (unlikely(!pte_present(pte
))) {
3634 huge_pte_clear(mm
, address
, ptep
, sz
);
3639 page
= pte_page(pte
);
3641 * If a reference page is supplied, it is because a specific
3642 * page is being unmapped, not a range. Ensure the page we
3643 * are about to unmap is the actual page of interest.
3646 if (page
!= ref_page
) {
3651 * Mark the VMA as having unmapped its page so that
3652 * future faults in this VMA will fail rather than
3653 * looking like data was lost
3655 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3658 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3659 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3660 if (huge_pte_dirty(pte
))
3661 set_page_dirty(page
);
3663 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3664 page_remove_rmap(page
, true);
3667 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3669 * Bail out after unmapping reference page if supplied
3674 mmu_notifier_invalidate_range_end(&range
);
3675 tlb_end_vma(tlb
, vma
);
3678 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3679 struct vm_area_struct
*vma
, unsigned long start
,
3680 unsigned long end
, struct page
*ref_page
)
3682 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3685 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3686 * test will fail on a vma being torn down, and not grab a page table
3687 * on its way out. We're lucky that the flag has such an appropriate
3688 * name, and can in fact be safely cleared here. We could clear it
3689 * before the __unmap_hugepage_range above, but all that's necessary
3690 * is to clear it before releasing the i_mmap_rwsem. This works
3691 * because in the context this is called, the VMA is about to be
3692 * destroyed and the i_mmap_rwsem is held.
3694 vma
->vm_flags
&= ~VM_MAYSHARE
;
3697 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3698 unsigned long end
, struct page
*ref_page
)
3700 struct mm_struct
*mm
;
3701 struct mmu_gather tlb
;
3702 unsigned long tlb_start
= start
;
3703 unsigned long tlb_end
= end
;
3706 * If shared PMDs were possibly used within this vma range, adjust
3707 * start/end for worst case tlb flushing.
3708 * Note that we can not be sure if PMDs are shared until we try to
3709 * unmap pages. However, we want to make sure TLB flushing covers
3710 * the largest possible range.
3712 adjust_range_if_pmd_sharing_possible(vma
, &tlb_start
, &tlb_end
);
3716 tlb_gather_mmu(&tlb
, mm
, tlb_start
, tlb_end
);
3717 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3718 tlb_finish_mmu(&tlb
, tlb_start
, tlb_end
);
3722 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3723 * mappping it owns the reserve page for. The intention is to unmap the page
3724 * from other VMAs and let the children be SIGKILLed if they are faulting the
3727 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3728 struct page
*page
, unsigned long address
)
3730 struct hstate
*h
= hstate_vma(vma
);
3731 struct vm_area_struct
*iter_vma
;
3732 struct address_space
*mapping
;
3736 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3737 * from page cache lookup which is in HPAGE_SIZE units.
3739 address
= address
& huge_page_mask(h
);
3740 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3742 mapping
= vma
->vm_file
->f_mapping
;
3745 * Take the mapping lock for the duration of the table walk. As
3746 * this mapping should be shared between all the VMAs,
3747 * __unmap_hugepage_range() is called as the lock is already held
3749 i_mmap_lock_write(mapping
);
3750 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3751 /* Do not unmap the current VMA */
3752 if (iter_vma
== vma
)
3756 * Shared VMAs have their own reserves and do not affect
3757 * MAP_PRIVATE accounting but it is possible that a shared
3758 * VMA is using the same page so check and skip such VMAs.
3760 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3764 * Unmap the page from other VMAs without their own reserves.
3765 * They get marked to be SIGKILLed if they fault in these
3766 * areas. This is because a future no-page fault on this VMA
3767 * could insert a zeroed page instead of the data existing
3768 * from the time of fork. This would look like data corruption
3770 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3771 unmap_hugepage_range(iter_vma
, address
,
3772 address
+ huge_page_size(h
), page
);
3774 i_mmap_unlock_write(mapping
);
3778 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3779 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3780 * cannot race with other handlers or page migration.
3781 * Keep the pte_same checks anyway to make transition from the mutex easier.
3783 static vm_fault_t
hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3784 unsigned long address
, pte_t
*ptep
,
3785 struct page
*pagecache_page
, spinlock_t
*ptl
)
3788 struct hstate
*h
= hstate_vma(vma
);
3789 struct page
*old_page
, *new_page
;
3790 int outside_reserve
= 0;
3792 unsigned long haddr
= address
& huge_page_mask(h
);
3793 struct mmu_notifier_range range
;
3795 pte
= huge_ptep_get(ptep
);
3796 old_page
= pte_page(pte
);
3799 /* If no-one else is actually using this page, avoid the copy
3800 * and just make the page writable */
3801 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3802 page_move_anon_rmap(old_page
, vma
);
3803 set_huge_ptep_writable(vma
, haddr
, ptep
);
3808 * If the process that created a MAP_PRIVATE mapping is about to
3809 * perform a COW due to a shared page count, attempt to satisfy
3810 * the allocation without using the existing reserves. The pagecache
3811 * page is used to determine if the reserve at this address was
3812 * consumed or not. If reserves were used, a partial faulted mapping
3813 * at the time of fork() could consume its reserves on COW instead
3814 * of the full address range.
3816 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3817 old_page
!= pagecache_page
)
3818 outside_reserve
= 1;
3823 * Drop page table lock as buddy allocator may be called. It will
3824 * be acquired again before returning to the caller, as expected.
3827 new_page
= alloc_huge_page(vma
, haddr
, outside_reserve
);
3829 if (IS_ERR(new_page
)) {
3831 * If a process owning a MAP_PRIVATE mapping fails to COW,
3832 * it is due to references held by a child and an insufficient
3833 * huge page pool. To guarantee the original mappers
3834 * reliability, unmap the page from child processes. The child
3835 * may get SIGKILLed if it later faults.
3837 if (outside_reserve
) {
3839 BUG_ON(huge_pte_none(pte
));
3840 unmap_ref_private(mm
, vma
, old_page
, haddr
);
3841 BUG_ON(huge_pte_none(pte
));
3843 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
3845 pte_same(huge_ptep_get(ptep
), pte
)))
3846 goto retry_avoidcopy
;
3848 * race occurs while re-acquiring page table
3849 * lock, and our job is done.
3854 ret
= vmf_error(PTR_ERR(new_page
));
3855 goto out_release_old
;
3859 * When the original hugepage is shared one, it does not have
3860 * anon_vma prepared.
3862 if (unlikely(anon_vma_prepare(vma
))) {
3864 goto out_release_all
;
3867 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3868 pages_per_huge_page(h
));
3869 __SetPageUptodate(new_page
);
3871 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
, haddr
,
3872 haddr
+ huge_page_size(h
));
3873 mmu_notifier_invalidate_range_start(&range
);
3876 * Retake the page table lock to check for racing updates
3877 * before the page tables are altered
3880 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
3881 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3882 ClearPagePrivate(new_page
);
3885 huge_ptep_clear_flush(vma
, haddr
, ptep
);
3886 mmu_notifier_invalidate_range(mm
, range
.start
, range
.end
);
3887 set_huge_pte_at(mm
, haddr
, ptep
,
3888 make_huge_pte(vma
, new_page
, 1));
3889 page_remove_rmap(old_page
, true);
3890 hugepage_add_new_anon_rmap(new_page
, vma
, haddr
);
3891 set_page_huge_active(new_page
);
3892 /* Make the old page be freed below */
3893 new_page
= old_page
;
3896 mmu_notifier_invalidate_range_end(&range
);
3898 restore_reserve_on_error(h
, vma
, haddr
, new_page
);
3903 spin_lock(ptl
); /* Caller expects lock to be held */
3907 /* Return the pagecache page at a given address within a VMA */
3908 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3909 struct vm_area_struct
*vma
, unsigned long address
)
3911 struct address_space
*mapping
;
3914 mapping
= vma
->vm_file
->f_mapping
;
3915 idx
= vma_hugecache_offset(h
, vma
, address
);
3917 return find_lock_page(mapping
, idx
);
3921 * Return whether there is a pagecache page to back given address within VMA.
3922 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3924 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3925 struct vm_area_struct
*vma
, unsigned long address
)
3927 struct address_space
*mapping
;
3931 mapping
= vma
->vm_file
->f_mapping
;
3932 idx
= vma_hugecache_offset(h
, vma
, address
);
3934 page
= find_get_page(mapping
, idx
);
3937 return page
!= NULL
;
3940 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3943 struct inode
*inode
= mapping
->host
;
3944 struct hstate
*h
= hstate_inode(inode
);
3945 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3949 ClearPagePrivate(page
);
3952 * set page dirty so that it will not be removed from cache/file
3953 * by non-hugetlbfs specific code paths.
3955 set_page_dirty(page
);
3957 spin_lock(&inode
->i_lock
);
3958 inode
->i_blocks
+= blocks_per_huge_page(h
);
3959 spin_unlock(&inode
->i_lock
);
3963 static vm_fault_t
hugetlb_no_page(struct mm_struct
*mm
,
3964 struct vm_area_struct
*vma
,
3965 struct address_space
*mapping
, pgoff_t idx
,
3966 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3968 struct hstate
*h
= hstate_vma(vma
);
3969 vm_fault_t ret
= VM_FAULT_SIGBUS
;
3975 unsigned long haddr
= address
& huge_page_mask(h
);
3976 bool new_page
= false;
3979 * Currently, we are forced to kill the process in the event the
3980 * original mapper has unmapped pages from the child due to a failed
3981 * COW. Warn that such a situation has occurred as it may not be obvious
3983 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3984 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3990 * Use page lock to guard against racing truncation
3991 * before we get page_table_lock.
3994 page
= find_lock_page(mapping
, idx
);
3996 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4001 * Check for page in userfault range
4003 if (userfaultfd_missing(vma
)) {
4005 struct vm_fault vmf
= {
4010 * Hard to debug if it ends up being
4011 * used by a callee that assumes
4012 * something about the other
4013 * uninitialized fields... same as in
4019 * hugetlb_fault_mutex must be dropped before
4020 * handling userfault. Reacquire after handling
4021 * fault to make calling code simpler.
4023 hash
= hugetlb_fault_mutex_hash(h
, mapping
, idx
);
4024 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4025 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
4026 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4030 page
= alloc_huge_page(vma
, haddr
, 0);
4033 * Returning error will result in faulting task being
4034 * sent SIGBUS. The hugetlb fault mutex prevents two
4035 * tasks from racing to fault in the same page which
4036 * could result in false unable to allocate errors.
4037 * Page migration does not take the fault mutex, but
4038 * does a clear then write of pte's under page table
4039 * lock. Page fault code could race with migration,
4040 * notice the clear pte and try to allocate a page
4041 * here. Before returning error, get ptl and make
4042 * sure there really is no pte entry.
4044 ptl
= huge_pte_lock(h
, mm
, ptep
);
4045 if (!huge_pte_none(huge_ptep_get(ptep
))) {
4051 ret
= vmf_error(PTR_ERR(page
));
4054 clear_huge_page(page
, address
, pages_per_huge_page(h
));
4055 __SetPageUptodate(page
);
4058 if (vma
->vm_flags
& VM_MAYSHARE
) {
4059 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
4068 if (unlikely(anon_vma_prepare(vma
))) {
4070 goto backout_unlocked
;
4076 * If memory error occurs between mmap() and fault, some process
4077 * don't have hwpoisoned swap entry for errored virtual address.
4078 * So we need to block hugepage fault by PG_hwpoison bit check.
4080 if (unlikely(PageHWPoison(page
))) {
4081 ret
= VM_FAULT_HWPOISON_LARGE
|
4082 VM_FAULT_SET_HINDEX(hstate_index(h
));
4083 goto backout_unlocked
;
4088 * If we are going to COW a private mapping later, we examine the
4089 * pending reservations for this page now. This will ensure that
4090 * any allocations necessary to record that reservation occur outside
4093 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4094 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4096 goto backout_unlocked
;
4098 /* Just decrements count, does not deallocate */
4099 vma_end_reservation(h
, vma
, haddr
);
4102 ptl
= huge_pte_lock(h
, mm
, ptep
);
4103 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4108 if (!huge_pte_none(huge_ptep_get(ptep
)))
4112 ClearPagePrivate(page
);
4113 hugepage_add_new_anon_rmap(page
, vma
, haddr
);
4115 page_dup_rmap(page
, true);
4116 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
4117 && (vma
->vm_flags
& VM_SHARED
)));
4118 set_huge_pte_at(mm
, haddr
, ptep
, new_pte
);
4120 hugetlb_count_add(pages_per_huge_page(h
), mm
);
4121 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
4122 /* Optimization, do the COW without a second fault */
4123 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
4129 * Only make newly allocated pages active. Existing pages found
4130 * in the pagecache could be !page_huge_active() if they have been
4131 * isolated for migration.
4134 set_page_huge_active(page
);
4144 restore_reserve_on_error(h
, vma
, haddr
, page
);
4150 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct address_space
*mapping
,
4153 unsigned long key
[2];
4156 key
[0] = (unsigned long) mapping
;
4159 hash
= jhash2((u32
*)&key
, sizeof(key
)/(sizeof(u32
)), 0);
4161 return hash
& (num_fault_mutexes
- 1);
4165 * For uniprocesor systems we always use a single mutex, so just
4166 * return 0 and avoid the hashing overhead.
4168 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct address_space
*mapping
,
4175 vm_fault_t
hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4176 unsigned long address
, unsigned int flags
)
4183 struct page
*page
= NULL
;
4184 struct page
*pagecache_page
= NULL
;
4185 struct hstate
*h
= hstate_vma(vma
);
4186 struct address_space
*mapping
;
4187 int need_wait_lock
= 0;
4188 unsigned long haddr
= address
& huge_page_mask(h
);
4190 ptep
= huge_pte_offset(mm
, haddr
, huge_page_size(h
));
4192 entry
= huge_ptep_get(ptep
);
4193 if (unlikely(is_hugetlb_entry_migration(entry
))) {
4194 migration_entry_wait_huge(vma
, mm
, ptep
);
4196 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
4197 return VM_FAULT_HWPOISON_LARGE
|
4198 VM_FAULT_SET_HINDEX(hstate_index(h
));
4200 ptep
= huge_pte_alloc(mm
, haddr
, huge_page_size(h
));
4202 return VM_FAULT_OOM
;
4205 mapping
= vma
->vm_file
->f_mapping
;
4206 idx
= vma_hugecache_offset(h
, vma
, haddr
);
4209 * Serialize hugepage allocation and instantiation, so that we don't
4210 * get spurious allocation failures if two CPUs race to instantiate
4211 * the same page in the page cache.
4213 hash
= hugetlb_fault_mutex_hash(h
, mapping
, idx
);
4214 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
4216 entry
= huge_ptep_get(ptep
);
4217 if (huge_pte_none(entry
)) {
4218 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
4225 * entry could be a migration/hwpoison entry at this point, so this
4226 * check prevents the kernel from going below assuming that we have
4227 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4228 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4231 if (!pte_present(entry
))
4235 * If we are going to COW the mapping later, we examine the pending
4236 * reservations for this page now. This will ensure that any
4237 * allocations necessary to record that reservation occur outside the
4238 * spinlock. For private mappings, we also lookup the pagecache
4239 * page now as it is used to determine if a reservation has been
4242 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
4243 if (vma_needs_reservation(h
, vma
, haddr
) < 0) {
4247 /* Just decrements count, does not deallocate */
4248 vma_end_reservation(h
, vma
, haddr
);
4250 if (!(vma
->vm_flags
& VM_MAYSHARE
))
4251 pagecache_page
= hugetlbfs_pagecache_page(h
,
4255 ptl
= huge_pte_lock(h
, mm
, ptep
);
4257 /* Check for a racing update before calling hugetlb_cow */
4258 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
4262 * hugetlb_cow() requires page locks of pte_page(entry) and
4263 * pagecache_page, so here we need take the former one
4264 * when page != pagecache_page or !pagecache_page.
4266 page
= pte_page(entry
);
4267 if (page
!= pagecache_page
)
4268 if (!trylock_page(page
)) {
4275 if (flags
& FAULT_FLAG_WRITE
) {
4276 if (!huge_pte_write(entry
)) {
4277 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
4278 pagecache_page
, ptl
);
4281 entry
= huge_pte_mkdirty(entry
);
4283 entry
= pte_mkyoung(entry
);
4284 if (huge_ptep_set_access_flags(vma
, haddr
, ptep
, entry
,
4285 flags
& FAULT_FLAG_WRITE
))
4286 update_mmu_cache(vma
, haddr
, ptep
);
4288 if (page
!= pagecache_page
)
4294 if (pagecache_page
) {
4295 unlock_page(pagecache_page
);
4296 put_page(pagecache_page
);
4299 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
4301 * Generally it's safe to hold refcount during waiting page lock. But
4302 * here we just wait to defer the next page fault to avoid busy loop and
4303 * the page is not used after unlocked before returning from the current
4304 * page fault. So we are safe from accessing freed page, even if we wait
4305 * here without taking refcount.
4308 wait_on_page_locked(page
);
4313 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4314 * modifications for huge pages.
4316 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
4318 struct vm_area_struct
*dst_vma
,
4319 unsigned long dst_addr
,
4320 unsigned long src_addr
,
4321 struct page
**pagep
)
4323 struct address_space
*mapping
;
4326 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
4327 struct hstate
*h
= hstate_vma(dst_vma
);
4335 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4339 ret
= copy_huge_page_from_user(page
,
4340 (const void __user
*) src_addr
,
4341 pages_per_huge_page(h
), false);
4343 /* fallback to copy_from_user outside mmap_sem */
4344 if (unlikely(ret
)) {
4347 /* don't free the page */
4356 * The memory barrier inside __SetPageUptodate makes sure that
4357 * preceding stores to the page contents become visible before
4358 * the set_pte_at() write.
4360 __SetPageUptodate(page
);
4362 mapping
= dst_vma
->vm_file
->f_mapping
;
4363 idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4366 * If shared, add to page cache
4369 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4372 goto out_release_nounlock
;
4375 * Serialization between remove_inode_hugepages() and
4376 * huge_add_to_page_cache() below happens through the
4377 * hugetlb_fault_mutex_table that here must be hold by
4380 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4382 goto out_release_nounlock
;
4385 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4389 * Recheck the i_size after holding PT lock to make sure not
4390 * to leave any page mapped (as page_mapped()) beyond the end
4391 * of the i_size (remove_inode_hugepages() is strict about
4392 * enforcing that). If we bail out here, we'll also leave a
4393 * page in the radix tree in the vm_shared case beyond the end
4394 * of the i_size, but remove_inode_hugepages() will take care
4395 * of it as soon as we drop the hugetlb_fault_mutex_table.
4397 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
4400 goto out_release_unlock
;
4403 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4404 goto out_release_unlock
;
4407 page_dup_rmap(page
, true);
4409 ClearPagePrivate(page
);
4410 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4413 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4414 if (dst_vma
->vm_flags
& VM_WRITE
)
4415 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4416 _dst_pte
= pte_mkyoung(_dst_pte
);
4418 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4420 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4421 dst_vma
->vm_flags
& VM_WRITE
);
4422 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4424 /* No need to invalidate - it was non-present before */
4425 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4428 set_page_huge_active(page
);
4438 out_release_nounlock
:
4443 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4444 struct page
**pages
, struct vm_area_struct
**vmas
,
4445 unsigned long *position
, unsigned long *nr_pages
,
4446 long i
, unsigned int flags
, int *nonblocking
)
4448 unsigned long pfn_offset
;
4449 unsigned long vaddr
= *position
;
4450 unsigned long remainder
= *nr_pages
;
4451 struct hstate
*h
= hstate_vma(vma
);
4454 while (vaddr
< vma
->vm_end
&& remainder
) {
4456 spinlock_t
*ptl
= NULL
;
4461 * If we have a pending SIGKILL, don't keep faulting pages and
4462 * potentially allocating memory.
4464 if (fatal_signal_pending(current
)) {
4470 * Some archs (sparc64, sh*) have multiple pte_ts to
4471 * each hugepage. We have to make sure we get the
4472 * first, for the page indexing below to work.
4474 * Note that page table lock is not held when pte is null.
4476 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
),
4479 ptl
= huge_pte_lock(h
, mm
, pte
);
4480 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4483 * When coredumping, it suits get_dump_page if we just return
4484 * an error where there's an empty slot with no huge pagecache
4485 * to back it. This way, we avoid allocating a hugepage, and
4486 * the sparse dumpfile avoids allocating disk blocks, but its
4487 * huge holes still show up with zeroes where they need to be.
4489 if (absent
&& (flags
& FOLL_DUMP
) &&
4490 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4498 * We need call hugetlb_fault for both hugepages under migration
4499 * (in which case hugetlb_fault waits for the migration,) and
4500 * hwpoisoned hugepages (in which case we need to prevent the
4501 * caller from accessing to them.) In order to do this, we use
4502 * here is_swap_pte instead of is_hugetlb_entry_migration and
4503 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4504 * both cases, and because we can't follow correct pages
4505 * directly from any kind of swap entries.
4507 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4508 ((flags
& FOLL_WRITE
) &&
4509 !huge_pte_write(huge_ptep_get(pte
)))) {
4511 unsigned int fault_flags
= 0;
4515 if (flags
& FOLL_WRITE
)
4516 fault_flags
|= FAULT_FLAG_WRITE
;
4518 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
4519 if (flags
& FOLL_NOWAIT
)
4520 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4521 FAULT_FLAG_RETRY_NOWAIT
;
4522 if (flags
& FOLL_TRIED
) {
4523 VM_WARN_ON_ONCE(fault_flags
&
4524 FAULT_FLAG_ALLOW_RETRY
);
4525 fault_flags
|= FAULT_FLAG_TRIED
;
4527 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4528 if (ret
& VM_FAULT_ERROR
) {
4529 err
= vm_fault_to_errno(ret
, flags
);
4533 if (ret
& VM_FAULT_RETRY
) {
4535 !(fault_flags
& FAULT_FLAG_RETRY_NOWAIT
))
4539 * VM_FAULT_RETRY must not return an
4540 * error, it will return zero
4543 * No need to update "position" as the
4544 * caller will not check it after
4545 * *nr_pages is set to 0.
4552 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4553 page
= pte_page(huge_ptep_get(pte
));
4556 * Instead of doing 'try_get_page()' below in the same_page
4557 * loop, just check the count once here.
4559 if (unlikely(page_count(page
) <= 0)) {
4569 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4580 if (vaddr
< vma
->vm_end
&& remainder
&&
4581 pfn_offset
< pages_per_huge_page(h
)) {
4583 * We use pfn_offset to avoid touching the pageframes
4584 * of this compound page.
4590 *nr_pages
= remainder
;
4592 * setting position is actually required only if remainder is
4593 * not zero but it's faster not to add a "if (remainder)"
4601 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4603 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4606 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4609 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
4610 unsigned long address
, unsigned long end
, pgprot_t newprot
)
4612 struct mm_struct
*mm
= vma
->vm_mm
;
4613 unsigned long start
= address
;
4616 struct hstate
*h
= hstate_vma(vma
);
4617 unsigned long pages
= 0;
4618 bool shared_pmd
= false;
4619 struct mmu_notifier_range range
;
4622 * In the case of shared PMDs, the area to flush could be beyond
4623 * start/end. Set range.start/range.end to cover the maximum possible
4624 * range if PMD sharing is possible.
4626 mmu_notifier_range_init(&range
, MMU_NOTIFY_PROTECTION_VMA
,
4627 0, vma
, mm
, start
, end
);
4628 adjust_range_if_pmd_sharing_possible(vma
, &range
.start
, &range
.end
);
4630 BUG_ON(address
>= end
);
4631 flush_cache_range(vma
, range
.start
, range
.end
);
4633 mmu_notifier_invalidate_range_start(&range
);
4634 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
4635 for (; address
< end
; address
+= huge_page_size(h
)) {
4637 ptep
= huge_pte_offset(mm
, address
, huge_page_size(h
));
4640 ptl
= huge_pte_lock(h
, mm
, ptep
);
4641 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
4647 pte
= huge_ptep_get(ptep
);
4648 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
4652 if (unlikely(is_hugetlb_entry_migration(pte
))) {
4653 swp_entry_t entry
= pte_to_swp_entry(pte
);
4655 if (is_write_migration_entry(entry
)) {
4658 make_migration_entry_read(&entry
);
4659 newpte
= swp_entry_to_pte(entry
);
4660 set_huge_swap_pte_at(mm
, address
, ptep
,
4661 newpte
, huge_page_size(h
));
4667 if (!huge_pte_none(pte
)) {
4670 old_pte
= huge_ptep_modify_prot_start(vma
, address
, ptep
);
4671 pte
= pte_mkhuge(huge_pte_modify(old_pte
, newprot
));
4672 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
4673 huge_ptep_modify_prot_commit(vma
, address
, ptep
, old_pte
, pte
);
4679 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4680 * may have cleared our pud entry and done put_page on the page table:
4681 * once we release i_mmap_rwsem, another task can do the final put_page
4682 * and that page table be reused and filled with junk. If we actually
4683 * did unshare a page of pmds, flush the range corresponding to the pud.
4686 flush_hugetlb_tlb_range(vma
, range
.start
, range
.end
);
4688 flush_hugetlb_tlb_range(vma
, start
, end
);
4690 * No need to call mmu_notifier_invalidate_range() we are downgrading
4691 * page table protection not changing it to point to a new page.
4693 * See Documentation/vm/mmu_notifier.rst
4695 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
4696 mmu_notifier_invalidate_range_end(&range
);
4698 return pages
<< h
->order
;
4701 int hugetlb_reserve_pages(struct inode
*inode
,
4703 struct vm_area_struct
*vma
,
4704 vm_flags_t vm_flags
)
4707 struct hstate
*h
= hstate_inode(inode
);
4708 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4709 struct resv_map
*resv_map
;
4712 /* This should never happen */
4714 VM_WARN(1, "%s called with a negative range\n", __func__
);
4719 * Only apply hugepage reservation if asked. At fault time, an
4720 * attempt will be made for VM_NORESERVE to allocate a page
4721 * without using reserves
4723 if (vm_flags
& VM_NORESERVE
)
4727 * Shared mappings base their reservation on the number of pages that
4728 * are already allocated on behalf of the file. Private mappings need
4729 * to reserve the full area even if read-only as mprotect() may be
4730 * called to make the mapping read-write. Assume !vma is a shm mapping
4732 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4734 * resv_map can not be NULL as hugetlb_reserve_pages is only
4735 * called for inodes for which resv_maps were created (see
4736 * hugetlbfs_get_inode).
4738 resv_map
= inode_resv_map(inode
);
4740 chg
= region_chg(resv_map
, from
, to
);
4743 resv_map
= resv_map_alloc();
4749 set_vma_resv_map(vma
, resv_map
);
4750 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4759 * There must be enough pages in the subpool for the mapping. If
4760 * the subpool has a minimum size, there may be some global
4761 * reservations already in place (gbl_reserve).
4763 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4764 if (gbl_reserve
< 0) {
4770 * Check enough hugepages are available for the reservation.
4771 * Hand the pages back to the subpool if there are not
4773 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4775 /* put back original number of pages, chg */
4776 (void)hugepage_subpool_put_pages(spool
, chg
);
4781 * Account for the reservations made. Shared mappings record regions
4782 * that have reservations as they are shared by multiple VMAs.
4783 * When the last VMA disappears, the region map says how much
4784 * the reservation was and the page cache tells how much of
4785 * the reservation was consumed. Private mappings are per-VMA and
4786 * only the consumed reservations are tracked. When the VMA
4787 * disappears, the original reservation is the VMA size and the
4788 * consumed reservations are stored in the map. Hence, nothing
4789 * else has to be done for private mappings here
4791 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4792 long add
= region_add(resv_map
, from
, to
);
4794 if (unlikely(chg
> add
)) {
4796 * pages in this range were added to the reserve
4797 * map between region_chg and region_add. This
4798 * indicates a race with alloc_huge_page. Adjust
4799 * the subpool and reserve counts modified above
4800 * based on the difference.
4804 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4806 hugetlb_acct_memory(h
, -rsv_adjust
);
4811 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4812 /* Don't call region_abort if region_chg failed */
4814 region_abort(resv_map
, from
, to
);
4815 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4816 kref_put(&resv_map
->refs
, resv_map_release
);
4820 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4823 struct hstate
*h
= hstate_inode(inode
);
4824 struct resv_map
*resv_map
= inode_resv_map(inode
);
4826 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4830 * Since this routine can be called in the evict inode path for all
4831 * hugetlbfs inodes, resv_map could be NULL.
4834 chg
= region_del(resv_map
, start
, end
);
4836 * region_del() can fail in the rare case where a region
4837 * must be split and another region descriptor can not be
4838 * allocated. If end == LONG_MAX, it will not fail.
4844 spin_lock(&inode
->i_lock
);
4845 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4846 spin_unlock(&inode
->i_lock
);
4849 * If the subpool has a minimum size, the number of global
4850 * reservations to be released may be adjusted.
4852 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4853 hugetlb_acct_memory(h
, -gbl_reserve
);
4858 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4859 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4860 struct vm_area_struct
*vma
,
4861 unsigned long addr
, pgoff_t idx
)
4863 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4865 unsigned long sbase
= saddr
& PUD_MASK
;
4866 unsigned long s_end
= sbase
+ PUD_SIZE
;
4868 /* Allow segments to share if only one is marked locked */
4869 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4870 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4873 * match the virtual addresses, permission and the alignment of the
4876 if (pmd_index(addr
) != pmd_index(saddr
) ||
4877 vm_flags
!= svm_flags
||
4878 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4884 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4886 unsigned long base
= addr
& PUD_MASK
;
4887 unsigned long end
= base
+ PUD_SIZE
;
4890 * check on proper vm_flags and page table alignment
4892 if (vma
->vm_flags
& VM_MAYSHARE
&& range_in_vma(vma
, base
, end
))
4898 * Determine if start,end range within vma could be mapped by shared pmd.
4899 * If yes, adjust start and end to cover range associated with possible
4900 * shared pmd mappings.
4902 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
4903 unsigned long *start
, unsigned long *end
)
4905 unsigned long v_start
= ALIGN(vma
->vm_start
, PUD_SIZE
),
4906 v_end
= ALIGN_DOWN(vma
->vm_end
, PUD_SIZE
);
4909 * vma need span at least one aligned PUD size and the start,end range
4910 * must at least partialy within it.
4912 if (!(vma
->vm_flags
& VM_MAYSHARE
) || !(v_end
> v_start
) ||
4913 (*end
<= v_start
) || (*start
>= v_end
))
4916 /* Extend the range to be PUD aligned for a worst case scenario */
4917 if (*start
> v_start
)
4918 *start
= ALIGN_DOWN(*start
, PUD_SIZE
);
4921 *end
= ALIGN(*end
, PUD_SIZE
);
4925 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4926 * and returns the corresponding pte. While this is not necessary for the
4927 * !shared pmd case because we can allocate the pmd later as well, it makes the
4928 * code much cleaner. pmd allocation is essential for the shared case because
4929 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4930 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4931 * bad pmd for sharing.
4933 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4935 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4936 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4937 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4939 struct vm_area_struct
*svma
;
4940 unsigned long saddr
;
4945 if (!vma_shareable(vma
, addr
))
4946 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4948 i_mmap_lock_read(mapping
);
4949 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4953 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4955 spte
= huge_pte_offset(svma
->vm_mm
, saddr
,
4956 vma_mmu_pagesize(svma
));
4958 get_page(virt_to_page(spte
));
4967 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
4968 if (pud_none(*pud
)) {
4969 pud_populate(mm
, pud
,
4970 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4973 put_page(virt_to_page(spte
));
4977 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4978 i_mmap_unlock_read(mapping
);
4983 * unmap huge page backed by shared pte.
4985 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4986 * indicated by page_count > 1, unmap is achieved by clearing pud and
4987 * decrementing the ref count. If count == 1, the pte page is not shared.
4989 * called with page table lock held.
4991 * returns: 1 successfully unmapped a shared pte page
4992 * 0 the underlying pte page is not shared, or it is the last user
4994 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4996 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4997 p4d_t
*p4d
= p4d_offset(pgd
, *addr
);
4998 pud_t
*pud
= pud_offset(p4d
, *addr
);
5000 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
5001 if (page_count(virt_to_page(ptep
)) == 1)
5005 put_page(virt_to_page(ptep
));
5007 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
5010 #define want_pmd_share() (1)
5011 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5012 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
5017 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
5022 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct
*vma
,
5023 unsigned long *start
, unsigned long *end
)
5026 #define want_pmd_share() (0)
5027 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5029 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5030 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
5031 unsigned long addr
, unsigned long sz
)
5038 pgd
= pgd_offset(mm
, addr
);
5039 p4d
= p4d_alloc(mm
, pgd
, addr
);
5042 pud
= pud_alloc(mm
, p4d
, addr
);
5044 if (sz
== PUD_SIZE
) {
5047 BUG_ON(sz
!= PMD_SIZE
);
5048 if (want_pmd_share() && pud_none(*pud
))
5049 pte
= huge_pmd_share(mm
, addr
, pud
);
5051 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
5054 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
5060 * huge_pte_offset() - Walk the page table to resolve the hugepage
5061 * entry at address @addr
5063 * Return: Pointer to page table or swap entry (PUD or PMD) for
5064 * address @addr, or NULL if a p*d_none() entry is encountered and the
5065 * size @sz doesn't match the hugepage size at this level of the page
5068 pte_t
*huge_pte_offset(struct mm_struct
*mm
,
5069 unsigned long addr
, unsigned long sz
)
5073 pud_t
*pud
, pud_entry
;
5074 pmd_t
*pmd
, pmd_entry
;
5076 pgd
= pgd_offset(mm
, addr
);
5077 if (!pgd_present(*pgd
))
5079 p4d
= p4d_offset(pgd
, addr
);
5080 if (!p4d_present(*p4d
))
5083 pud
= pud_offset(p4d
, addr
);
5084 pud_entry
= READ_ONCE(*pud
);
5085 if (sz
!= PUD_SIZE
&& pud_none(pud_entry
))
5087 /* hugepage or swap? */
5088 if (pud_huge(pud_entry
) || !pud_present(pud_entry
))
5089 return (pte_t
*)pud
;
5091 pmd
= pmd_offset(pud
, addr
);
5092 pmd_entry
= READ_ONCE(*pmd
);
5093 if (sz
!= PMD_SIZE
&& pmd_none(pmd_entry
))
5095 /* hugepage or swap? */
5096 if (pmd_huge(pmd_entry
) || !pmd_present(pmd_entry
))
5097 return (pte_t
*)pmd
;
5102 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5105 * These functions are overwritable if your architecture needs its own
5108 struct page
* __weak
5109 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
5112 return ERR_PTR(-EINVAL
);
5115 struct page
* __weak
5116 follow_huge_pd(struct vm_area_struct
*vma
,
5117 unsigned long address
, hugepd_t hpd
, int flags
, int pdshift
)
5119 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5123 struct page
* __weak
5124 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
5125 pmd_t
*pmd
, int flags
)
5127 struct page
*page
= NULL
;
5131 ptl
= pmd_lockptr(mm
, pmd
);
5134 * make sure that the address range covered by this pmd is not
5135 * unmapped from other threads.
5137 if (!pmd_huge(*pmd
))
5139 pte
= huge_ptep_get((pte_t
*)pmd
);
5140 if (pte_present(pte
)) {
5141 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
5142 if (flags
& FOLL_GET
)
5145 if (is_hugetlb_entry_migration(pte
)) {
5147 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
5151 * hwpoisoned entry is treated as no_page_table in
5152 * follow_page_mask().
5160 struct page
* __weak
5161 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
5162 pud_t
*pud
, int flags
)
5164 if (flags
& FOLL_GET
)
5167 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
5170 struct page
* __weak
5171 follow_huge_pgd(struct mm_struct
*mm
, unsigned long address
, pgd_t
*pgd
, int flags
)
5173 if (flags
& FOLL_GET
)
5176 return pte_page(*(pte_t
*)pgd
) + ((address
& ~PGDIR_MASK
) >> PAGE_SHIFT
);
5179 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
5183 spin_lock(&hugetlb_lock
);
5184 if (!PageHeadHuge(page
) || !page_huge_active(page
) ||
5185 !get_page_unless_zero(page
)) {
5189 clear_page_huge_active(page
);
5190 list_move_tail(&page
->lru
, list
);
5192 spin_unlock(&hugetlb_lock
);
5196 void putback_active_hugepage(struct page
*page
)
5198 VM_BUG_ON_PAGE(!PageHead(page
), page
);
5199 spin_lock(&hugetlb_lock
);
5200 set_page_huge_active(page
);
5201 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
5202 spin_unlock(&hugetlb_lock
);
5206 void move_hugetlb_state(struct page
*oldpage
, struct page
*newpage
, int reason
)
5208 struct hstate
*h
= page_hstate(oldpage
);
5210 hugetlb_cgroup_migrate(oldpage
, newpage
);
5211 set_page_owner_migrate_reason(newpage
, reason
);
5214 * transfer temporary state of the new huge page. This is
5215 * reverse to other transitions because the newpage is going to
5216 * be final while the old one will be freed so it takes over
5217 * the temporary status.
5219 * Also note that we have to transfer the per-node surplus state
5220 * here as well otherwise the global surplus count will not match
5223 if (PageHugeTemporary(newpage
)) {
5224 int old_nid
= page_to_nid(oldpage
);
5225 int new_nid
= page_to_nid(newpage
);
5227 SetPageHugeTemporary(oldpage
);
5228 ClearPageHugeTemporary(newpage
);
5230 spin_lock(&hugetlb_lock
);
5231 if (h
->surplus_huge_pages_node
[old_nid
]) {
5232 h
->surplus_huge_pages_node
[old_nid
]--;
5233 h
->surplus_huge_pages_node
[new_nid
]++;
5235 spin_unlock(&hugetlb_lock
);