2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
38 int hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
49 __initdata
LIST_HEAD(huge_boot_pages
);
51 /* for command line parsing */
52 static struct hstate
* __initdata parsed_hstate
;
53 static unsigned long __initdata default_hstate_max_huge_pages
;
54 static unsigned long __initdata default_hstate_size
;
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
60 DEFINE_SPINLOCK(hugetlb_lock
);
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 static int num_fault_mutexes
;
67 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
72 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
74 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
76 spin_unlock(&spool
->lock
);
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
82 if (spool
->min_hpages
!= -1)
83 hugetlb_acct_memory(spool
->hstate
,
89 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
92 struct hugepage_subpool
*spool
;
94 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
98 spin_lock_init(&spool
->lock
);
100 spool
->max_hpages
= max_hpages
;
102 spool
->min_hpages
= min_hpages
;
104 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
108 spool
->rsv_hpages
= min_hpages
;
113 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
115 spin_lock(&spool
->lock
);
116 BUG_ON(!spool
->count
);
118 unlock_or_release_subpool(spool
);
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
129 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
137 spin_lock(&spool
->lock
);
139 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
140 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
141 spool
->used_hpages
+= delta
;
148 if (spool
->min_hpages
!= -1) { /* minimum size accounting */
149 if (delta
> spool
->rsv_hpages
) {
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
154 ret
= delta
- spool
->rsv_hpages
;
155 spool
->rsv_hpages
= 0;
157 ret
= 0; /* reserves already accounted for */
158 spool
->rsv_hpages
-= delta
;
163 spin_unlock(&spool
->lock
);
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
173 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
181 spin_lock(&spool
->lock
);
183 if (spool
->max_hpages
!= -1) /* maximum size accounting */
184 spool
->used_hpages
-= delta
;
186 if (spool
->min_hpages
!= -1) { /* minimum size accounting */
187 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
190 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
192 spool
->rsv_hpages
+= delta
;
193 if (spool
->rsv_hpages
> spool
->min_hpages
)
194 spool
->rsv_hpages
= spool
->min_hpages
;
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
201 unlock_or_release_subpool(spool
);
206 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
208 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
211 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
213 return subpool_inode(file_inode(vma
->vm_file
));
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
236 struct list_head link
;
242 * Add the huge page range represented by [f, t) to the reserve
243 * map. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
255 static long region_add(struct resv_map
*resv
, long f
, long t
)
257 struct list_head
*head
= &resv
->regions
;
258 struct file_region
*rg
, *nrg
, *trg
;
261 spin_lock(&resv
->lock
);
262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg
, head
, link
)
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
273 if (&rg
->link
== head
|| t
< rg
->from
) {
274 VM_BUG_ON(resv
->region_cache_count
<= 0);
276 resv
->region_cache_count
--;
277 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
,
279 list_del(&nrg
->link
);
283 list_add(&nrg
->link
, rg
->link
.prev
);
289 /* Round our left edge to the current segment if it encloses us. */
293 /* Check for and consume any regions we now overlap with. */
295 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
296 if (&rg
->link
== head
)
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
311 add
-= (rg
->to
- rg
->from
);
317 add
+= (nrg
->from
- f
); /* Added to beginning of region */
319 add
+= t
- nrg
->to
; /* Added to end of region */
323 resv
->adds_in_progress
--;
324 spin_unlock(&resv
->lock
);
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
351 static long region_chg(struct resv_map
*resv
, long f
, long t
)
353 struct list_head
*head
= &resv
->regions
;
354 struct file_region
*rg
, *nrg
= NULL
;
358 spin_lock(&resv
->lock
);
360 resv
->adds_in_progress
++;
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
366 if (resv
->adds_in_progress
> resv
->region_cache_count
) {
367 struct file_region
*trg
;
369 VM_BUG_ON(resv
->adds_in_progress
- resv
->region_cache_count
> 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv
->adds_in_progress
--;
372 spin_unlock(&resv
->lock
);
374 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
378 spin_lock(&resv
->lock
);
379 list_add(&trg
->link
, &resv
->region_cache
);
380 resv
->region_cache_count
++;
384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg
, head
, link
)
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg
->link
== head
|| t
< rg
->from
) {
394 resv
->adds_in_progress
--;
395 spin_unlock(&resv
->lock
);
396 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
402 INIT_LIST_HEAD(&nrg
->link
);
406 list_add(&nrg
->link
, rg
->link
.prev
);
411 /* Round our left edge to the current segment if it encloses us. */
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
418 if (&rg
->link
== head
)
423 /* We overlap with this area, if it extends further than
424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
430 chg
-= rg
->to
- rg
->from
;
434 spin_unlock(&resv
->lock
);
435 /* We already know we raced and no longer need the new region */
439 spin_unlock(&resv
->lock
);
444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
454 static void region_abort(struct resv_map
*resv
, long f
, long t
)
456 spin_lock(&resv
->lock
);
457 VM_BUG_ON(!resv
->region_cache_count
);
458 resv
->adds_in_progress
--;
459 spin_unlock(&resv
->lock
);
463 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
476 static long region_del(struct resv_map
*resv
, long f
, long t
)
478 struct list_head
*head
= &resv
->regions
;
479 struct file_region
*rg
, *trg
;
480 struct file_region
*nrg
= NULL
;
484 spin_lock(&resv
->lock
);
485 list_for_each_entry_safe(rg
, trg
, head
, link
) {
491 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
497 resv
->region_cache_count
> resv
->adds_in_progress
) {
498 nrg
= list_first_entry(&resv
->region_cache
,
501 list_del(&nrg
->link
);
502 resv
->region_cache_count
--;
506 spin_unlock(&resv
->lock
);
507 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
515 /* New entry for end of split region */
518 INIT_LIST_HEAD(&nrg
->link
);
520 /* Original entry is trimmed */
523 list_add(&nrg
->link
, &rg
->link
);
528 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
529 del
+= rg
->to
- rg
->from
;
535 if (f
<= rg
->from
) { /* Trim beginning of region */
538 } else { /* Trim end of region */
544 spin_unlock(&resv
->lock
);
550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
558 void hugetlb_fix_reserve_counts(struct inode
*inode
, bool restore_reserve
)
560 struct hugepage_subpool
*spool
= subpool_inode(inode
);
563 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
564 if (restore_reserve
&& rsv_adjust
) {
565 struct hstate
*h
= hstate_inode(inode
);
567 hugetlb_acct_memory(h
, 1);
572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
575 static long region_count(struct resv_map
*resv
, long f
, long t
)
577 struct list_head
*head
= &resv
->regions
;
578 struct file_region
*rg
;
581 spin_lock(&resv
->lock
);
582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg
, head
, link
) {
592 seg_from
= max(rg
->from
, f
);
593 seg_to
= min(rg
->to
, t
);
595 chg
+= seg_to
- seg_from
;
597 spin_unlock(&resv
->lock
);
603 * Convert the address within this vma to the page offset within
604 * the mapping, in pagecache page units; huge pages here.
606 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
607 struct vm_area_struct
*vma
, unsigned long address
)
609 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
610 (vma
->vm_pgoff
>> huge_page_order(h
));
613 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
614 unsigned long address
)
616 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
623 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
625 struct hstate
*hstate
;
627 if (!is_vm_hugetlb_page(vma
))
630 hstate
= hstate_vma(vma
);
632 return 1UL << huge_page_shift(hstate
);
634 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
642 #ifndef vma_mmu_pagesize
643 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
645 return vma_kernel_pagesize(vma
);
650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
654 #define HPAGE_RESV_OWNER (1UL << 0)
655 #define HPAGE_RESV_UNMAPPED (1UL << 1)
656 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
677 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
679 return (unsigned long)vma
->vm_private_data
;
682 static void set_vma_private_data(struct vm_area_struct
*vma
,
685 vma
->vm_private_data
= (void *)value
;
688 struct resv_map
*resv_map_alloc(void)
690 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
691 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
693 if (!resv_map
|| !rg
) {
699 kref_init(&resv_map
->refs
);
700 spin_lock_init(&resv_map
->lock
);
701 INIT_LIST_HEAD(&resv_map
->regions
);
703 resv_map
->adds_in_progress
= 0;
705 INIT_LIST_HEAD(&resv_map
->region_cache
);
706 list_add(&rg
->link
, &resv_map
->region_cache
);
707 resv_map
->region_cache_count
= 1;
712 void resv_map_release(struct kref
*ref
)
714 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
715 struct list_head
*head
= &resv_map
->region_cache
;
716 struct file_region
*rg
, *trg
;
718 /* Clear out any active regions before we release the map. */
719 region_del(resv_map
, 0, LONG_MAX
);
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg
, trg
, head
, link
) {
727 VM_BUG_ON(resv_map
->adds_in_progress
);
732 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
734 return inode
->i_mapping
->private_data
;
737 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
740 if (vma
->vm_flags
& VM_MAYSHARE
) {
741 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
742 struct inode
*inode
= mapping
->host
;
744 return inode_resv_map(inode
);
747 return (struct resv_map
*)(get_vma_private_data(vma
) &
752 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
755 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
757 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
758 HPAGE_RESV_MASK
) | (unsigned long)map
);
761 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
764 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
766 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
769 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
773 return (get_vma_private_data(vma
) & flag
) != 0;
776 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
777 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
780 if (!(vma
->vm_flags
& VM_MAYSHARE
))
781 vma
->vm_private_data
= (void *)0;
784 /* Returns true if the VMA has associated reserve pages */
785 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
787 if (vma
->vm_flags
& VM_NORESERVE
) {
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
797 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
803 /* Shared mappings always use reserves */
804 if (vma
->vm_flags
& VM_MAYSHARE
) {
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
819 * Only the process that called mmap() has reserves for
822 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
828 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
830 int nid
= page_to_nid(page
);
831 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
832 h
->free_huge_pages
++;
833 h
->free_huge_pages_node
[nid
]++;
836 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
840 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
841 if (!is_migrate_isolate_page(page
))
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
847 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
849 list_move(&page
->lru
, &h
->hugepage_activelist
);
850 set_page_refcounted(page
);
851 h
->free_huge_pages
--;
852 h
->free_huge_pages_node
[nid
]--;
856 /* Movability of hugepages depends on migration support. */
857 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
859 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
860 return GFP_HIGHUSER_MOVABLE
;
865 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
866 struct vm_area_struct
*vma
,
867 unsigned long address
, int avoid_reserve
,
870 struct page
*page
= NULL
;
871 struct mempolicy
*mpol
;
872 nodemask_t
*nodemask
;
873 struct zonelist
*zonelist
;
876 unsigned int cpuset_mems_cookie
;
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
883 if (!vma_has_reserves(vma
, chg
) &&
884 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
887 /* If reserves cannot be used, ensure enough pages are in the pool */
888 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
892 cpuset_mems_cookie
= read_mems_allowed_begin();
893 zonelist
= huge_zonelist(vma
, address
,
894 htlb_alloc_mask(h
), &mpol
, &nodemask
);
896 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
897 MAX_NR_ZONES
- 1, nodemask
) {
898 if (cpuset_zone_allowed(zone
, htlb_alloc_mask(h
))) {
899 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
903 if (!vma_has_reserves(vma
, chg
))
906 SetPagePrivate(page
);
907 h
->resv_huge_pages
--;
914 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
929 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
931 nid
= next_node(nid
, *nodes_allowed
);
932 if (nid
== MAX_NUMNODES
)
933 nid
= first_node(*nodes_allowed
);
934 VM_BUG_ON(nid
>= MAX_NUMNODES
);
939 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
941 if (!node_isset(nid
, *nodes_allowed
))
942 nid
= next_node_allowed(nid
, nodes_allowed
);
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
952 static int hstate_next_node_to_alloc(struct hstate
*h
,
953 nodemask_t
*nodes_allowed
)
957 VM_BUG_ON(!nodes_allowed
);
959 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
960 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
971 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
975 VM_BUG_ON(!nodes_allowed
);
977 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
978 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
983 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
989 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
995 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996 static void destroy_compound_gigantic_page(struct page
*page
,
1000 int nr_pages
= 1 << order
;
1001 struct page
*p
= page
+ 1;
1003 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1004 clear_compound_head(p
);
1005 set_page_refcounted(p
);
1008 set_compound_order(page
, 0);
1009 __ClearPageHead(page
);
1012 static void free_gigantic_page(struct page
*page
, unsigned order
)
1014 free_contig_range(page_to_pfn(page
), 1 << order
);
1017 static int __alloc_gigantic_page(unsigned long start_pfn
,
1018 unsigned long nr_pages
)
1020 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1021 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
);
1024 static bool pfn_range_valid_gigantic(unsigned long start_pfn
,
1025 unsigned long nr_pages
)
1027 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1030 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1034 page
= pfn_to_page(i
);
1036 if (PageReserved(page
))
1039 if (page_count(page
) > 0)
1049 static bool zone_spans_last_pfn(const struct zone
*zone
,
1050 unsigned long start_pfn
, unsigned long nr_pages
)
1052 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1053 return zone_spans_pfn(zone
, last_pfn
);
1056 static struct page
*alloc_gigantic_page(int nid
, unsigned order
)
1058 unsigned long nr_pages
= 1 << order
;
1059 unsigned long ret
, pfn
, flags
;
1062 z
= NODE_DATA(nid
)->node_zones
;
1063 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
1064 spin_lock_irqsave(&z
->lock
, flags
);
1066 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
1067 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
1068 if (pfn_range_valid_gigantic(pfn
, nr_pages
)) {
1070 * We release the zone lock here because
1071 * alloc_contig_range() will also lock the zone
1072 * at some point. If there's an allocation
1073 * spinning on this lock, it may win the race
1074 * and cause alloc_contig_range() to fail...
1076 spin_unlock_irqrestore(&z
->lock
, flags
);
1077 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
1079 return pfn_to_page(pfn
);
1080 spin_lock_irqsave(&z
->lock
, flags
);
1085 spin_unlock_irqrestore(&z
->lock
, flags
);
1091 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1092 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
);
1094 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
1098 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
1100 prep_compound_gigantic_page(page
, huge_page_order(h
));
1101 prep_new_huge_page(h
, page
, nid
);
1107 static int alloc_fresh_gigantic_page(struct hstate
*h
,
1108 nodemask_t
*nodes_allowed
)
1110 struct page
*page
= NULL
;
1113 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1114 page
= alloc_fresh_gigantic_page_node(h
, node
);
1122 static inline bool gigantic_page_supported(void) { return true; }
1124 static inline bool gigantic_page_supported(void) { return false; }
1125 static inline void free_gigantic_page(struct page
*page
, unsigned order
) { }
1126 static inline void destroy_compound_gigantic_page(struct page
*page
,
1127 unsigned long order
) { }
1128 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
1129 nodemask_t
*nodes_allowed
) { return 0; }
1132 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1136 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1140 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1141 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1142 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1143 1 << PG_referenced
| 1 << PG_dirty
|
1144 1 << PG_active
| 1 << PG_private
|
1147 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1148 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1149 set_page_refcounted(page
);
1150 if (hstate_is_gigantic(h
)) {
1151 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1152 free_gigantic_page(page
, huge_page_order(h
));
1154 __free_pages(page
, huge_page_order(h
));
1158 struct hstate
*size_to_hstate(unsigned long size
)
1162 for_each_hstate(h
) {
1163 if (huge_page_size(h
) == size
)
1170 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1171 * to hstate->hugepage_activelist.)
1173 * This function can be called for tail pages, but never returns true for them.
1175 bool page_huge_active(struct page
*page
)
1177 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1178 return PageHead(page
) && PagePrivate(&page
[1]);
1181 /* never called for tail page */
1182 static void set_page_huge_active(struct page
*page
)
1184 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1185 SetPagePrivate(&page
[1]);
1188 static void clear_page_huge_active(struct page
*page
)
1190 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1191 ClearPagePrivate(&page
[1]);
1194 void free_huge_page(struct page
*page
)
1197 * Can't pass hstate in here because it is called from the
1198 * compound page destructor.
1200 struct hstate
*h
= page_hstate(page
);
1201 int nid
= page_to_nid(page
);
1202 struct hugepage_subpool
*spool
=
1203 (struct hugepage_subpool
*)page_private(page
);
1204 bool restore_reserve
;
1206 set_page_private(page
, 0);
1207 page
->mapping
= NULL
;
1208 BUG_ON(page_count(page
));
1209 BUG_ON(page_mapcount(page
));
1210 restore_reserve
= PagePrivate(page
);
1211 ClearPagePrivate(page
);
1214 * A return code of zero implies that the subpool will be under its
1215 * minimum size if the reservation is not restored after page is free.
1216 * Therefore, force restore_reserve operation.
1218 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1219 restore_reserve
= true;
1221 spin_lock(&hugetlb_lock
);
1222 clear_page_huge_active(page
);
1223 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1224 pages_per_huge_page(h
), page
);
1225 if (restore_reserve
)
1226 h
->resv_huge_pages
++;
1228 if (h
->surplus_huge_pages_node
[nid
]) {
1229 /* remove the page from active list */
1230 list_del(&page
->lru
);
1231 update_and_free_page(h
, page
);
1232 h
->surplus_huge_pages
--;
1233 h
->surplus_huge_pages_node
[nid
]--;
1235 arch_clear_hugepage_flags(page
);
1236 enqueue_huge_page(h
, page
);
1238 spin_unlock(&hugetlb_lock
);
1241 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1243 INIT_LIST_HEAD(&page
->lru
);
1244 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1245 spin_lock(&hugetlb_lock
);
1246 set_hugetlb_cgroup(page
, NULL
);
1248 h
->nr_huge_pages_node
[nid
]++;
1249 spin_unlock(&hugetlb_lock
);
1250 put_page(page
); /* free it into the hugepage allocator */
1253 static void prep_compound_gigantic_page(struct page
*page
, unsigned long order
)
1256 int nr_pages
= 1 << order
;
1257 struct page
*p
= page
+ 1;
1259 /* we rely on prep_new_huge_page to set the destructor */
1260 set_compound_order(page
, order
);
1261 __SetPageHead(page
);
1262 __ClearPageReserved(page
);
1263 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1265 * For gigantic hugepages allocated through bootmem at
1266 * boot, it's safer to be consistent with the not-gigantic
1267 * hugepages and clear the PG_reserved bit from all tail pages
1268 * too. Otherwse drivers using get_user_pages() to access tail
1269 * pages may get the reference counting wrong if they see
1270 * PG_reserved set on a tail page (despite the head page not
1271 * having PG_reserved set). Enforcing this consistency between
1272 * head and tail pages allows drivers to optimize away a check
1273 * on the head page when they need know if put_page() is needed
1274 * after get_user_pages().
1276 __ClearPageReserved(p
);
1277 set_page_count(p
, 0);
1278 set_compound_head(p
, page
);
1283 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1284 * transparent huge pages. See the PageTransHuge() documentation for more
1287 int PageHuge(struct page
*page
)
1289 if (!PageCompound(page
))
1292 page
= compound_head(page
);
1293 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1295 EXPORT_SYMBOL_GPL(PageHuge
);
1298 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1299 * normal or transparent huge pages.
1301 int PageHeadHuge(struct page
*page_head
)
1303 if (!PageHead(page_head
))
1306 return get_compound_page_dtor(page_head
) == free_huge_page
;
1309 pgoff_t
__basepage_index(struct page
*page
)
1311 struct page
*page_head
= compound_head(page
);
1312 pgoff_t index
= page_index(page_head
);
1313 unsigned long compound_idx
;
1315 if (!PageHuge(page_head
))
1316 return page_index(page
);
1318 if (compound_order(page_head
) >= MAX_ORDER
)
1319 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1321 compound_idx
= page
- page_head
;
1323 return (index
<< compound_order(page_head
)) + compound_idx
;
1326 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
1330 page
= __alloc_pages_node(nid
,
1331 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1332 __GFP_REPEAT
|__GFP_NOWARN
,
1333 huge_page_order(h
));
1335 prep_new_huge_page(h
, page
, nid
);
1341 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1347 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1348 page
= alloc_fresh_huge_page_node(h
, node
);
1356 count_vm_event(HTLB_BUDDY_PGALLOC
);
1358 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1364 * Free huge page from pool from next node to free.
1365 * Attempt to keep persistent huge pages more or less
1366 * balanced over allowed nodes.
1367 * Called with hugetlb_lock locked.
1369 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1375 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1377 * If we're returning unused surplus pages, only examine
1378 * nodes with surplus pages.
1380 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1381 !list_empty(&h
->hugepage_freelists
[node
])) {
1383 list_entry(h
->hugepage_freelists
[node
].next
,
1385 list_del(&page
->lru
);
1386 h
->free_huge_pages
--;
1387 h
->free_huge_pages_node
[node
]--;
1389 h
->surplus_huge_pages
--;
1390 h
->surplus_huge_pages_node
[node
]--;
1392 update_and_free_page(h
, page
);
1402 * Dissolve a given free hugepage into free buddy pages. This function does
1403 * nothing for in-use (including surplus) hugepages.
1405 static void dissolve_free_huge_page(struct page
*page
)
1407 spin_lock(&hugetlb_lock
);
1408 if (PageHuge(page
) && !page_count(page
)) {
1409 struct hstate
*h
= page_hstate(page
);
1410 int nid
= page_to_nid(page
);
1411 list_del(&page
->lru
);
1412 h
->free_huge_pages
--;
1413 h
->free_huge_pages_node
[nid
]--;
1414 update_and_free_page(h
, page
);
1416 spin_unlock(&hugetlb_lock
);
1420 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1421 * make specified memory blocks removable from the system.
1422 * Note that start_pfn should aligned with (minimum) hugepage size.
1424 void dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1428 if (!hugepages_supported())
1431 VM_BUG_ON(!IS_ALIGNED(start_pfn
, 1 << minimum_order
));
1432 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
)
1433 dissolve_free_huge_page(pfn_to_page(pfn
));
1437 * There are 3 ways this can get called:
1438 * 1. With vma+addr: we use the VMA's memory policy
1439 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1440 * page from any node, and let the buddy allocator itself figure
1442 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1443 * strictly from 'nid'
1445 static struct page
*__hugetlb_alloc_buddy_huge_page(struct hstate
*h
,
1446 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1448 int order
= huge_page_order(h
);
1449 gfp_t gfp
= htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_REPEAT
|__GFP_NOWARN
;
1450 unsigned int cpuset_mems_cookie
;
1453 * We need a VMA to get a memory policy. If we do not
1454 * have one, we use the 'nid' argument.
1456 * The mempolicy stuff below has some non-inlined bits
1457 * and calls ->vm_ops. That makes it hard to optimize at
1458 * compile-time, even when NUMA is off and it does
1459 * nothing. This helps the compiler optimize it out.
1461 if (!IS_ENABLED(CONFIG_NUMA
) || !vma
) {
1463 * If a specific node is requested, make sure to
1464 * get memory from there, but only when a node
1465 * is explicitly specified.
1467 if (nid
!= NUMA_NO_NODE
)
1468 gfp
|= __GFP_THISNODE
;
1470 * Make sure to call something that can handle
1473 return alloc_pages_node(nid
, gfp
, order
);
1477 * OK, so we have a VMA. Fetch the mempolicy and try to
1478 * allocate a huge page with it. We will only reach this
1479 * when CONFIG_NUMA=y.
1483 struct mempolicy
*mpol
;
1484 struct zonelist
*zl
;
1485 nodemask_t
*nodemask
;
1487 cpuset_mems_cookie
= read_mems_allowed_begin();
1488 zl
= huge_zonelist(vma
, addr
, gfp
, &mpol
, &nodemask
);
1489 mpol_cond_put(mpol
);
1490 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nodemask
);
1493 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1499 * There are two ways to allocate a huge page:
1500 * 1. When you have a VMA and an address (like a fault)
1501 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1503 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1504 * this case which signifies that the allocation should be done with
1505 * respect for the VMA's memory policy.
1507 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1508 * implies that memory policies will not be taken in to account.
1510 static struct page
*__alloc_buddy_huge_page(struct hstate
*h
,
1511 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1516 if (hstate_is_gigantic(h
))
1520 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1521 * This makes sure the caller is picking _one_ of the modes with which
1522 * we can call this function, not both.
1524 if (vma
|| (addr
!= -1)) {
1525 VM_WARN_ON_ONCE(addr
== -1);
1526 VM_WARN_ON_ONCE(nid
!= NUMA_NO_NODE
);
1529 * Assume we will successfully allocate the surplus page to
1530 * prevent racing processes from causing the surplus to exceed
1533 * This however introduces a different race, where a process B
1534 * tries to grow the static hugepage pool while alloc_pages() is
1535 * called by process A. B will only examine the per-node
1536 * counters in determining if surplus huge pages can be
1537 * converted to normal huge pages in adjust_pool_surplus(). A
1538 * won't be able to increment the per-node counter, until the
1539 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1540 * no more huge pages can be converted from surplus to normal
1541 * state (and doesn't try to convert again). Thus, we have a
1542 * case where a surplus huge page exists, the pool is grown, and
1543 * the surplus huge page still exists after, even though it
1544 * should just have been converted to a normal huge page. This
1545 * does not leak memory, though, as the hugepage will be freed
1546 * once it is out of use. It also does not allow the counters to
1547 * go out of whack in adjust_pool_surplus() as we don't modify
1548 * the node values until we've gotten the hugepage and only the
1549 * per-node value is checked there.
1551 spin_lock(&hugetlb_lock
);
1552 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1553 spin_unlock(&hugetlb_lock
);
1557 h
->surplus_huge_pages
++;
1559 spin_unlock(&hugetlb_lock
);
1561 page
= __hugetlb_alloc_buddy_huge_page(h
, vma
, addr
, nid
);
1563 spin_lock(&hugetlb_lock
);
1565 INIT_LIST_HEAD(&page
->lru
);
1566 r_nid
= page_to_nid(page
);
1567 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1568 set_hugetlb_cgroup(page
, NULL
);
1570 * We incremented the global counters already
1572 h
->nr_huge_pages_node
[r_nid
]++;
1573 h
->surplus_huge_pages_node
[r_nid
]++;
1574 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1577 h
->surplus_huge_pages
--;
1578 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1580 spin_unlock(&hugetlb_lock
);
1586 * Allocate a huge page from 'nid'. Note, 'nid' may be
1587 * NUMA_NO_NODE, which means that it may be allocated
1591 struct page
*__alloc_buddy_huge_page_no_mpol(struct hstate
*h
, int nid
)
1593 unsigned long addr
= -1;
1595 return __alloc_buddy_huge_page(h
, NULL
, addr
, nid
);
1599 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1602 struct page
*__alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1603 struct vm_area_struct
*vma
, unsigned long addr
)
1605 return __alloc_buddy_huge_page(h
, vma
, addr
, NUMA_NO_NODE
);
1609 * This allocation function is useful in the context where vma is irrelevant.
1610 * E.g. soft-offlining uses this function because it only cares physical
1611 * address of error page.
1613 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1615 struct page
*page
= NULL
;
1617 spin_lock(&hugetlb_lock
);
1618 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1619 page
= dequeue_huge_page_node(h
, nid
);
1620 spin_unlock(&hugetlb_lock
);
1623 page
= __alloc_buddy_huge_page_no_mpol(h
, nid
);
1629 * Increase the hugetlb pool such that it can accommodate a reservation
1632 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1634 struct list_head surplus_list
;
1635 struct page
*page
, *tmp
;
1637 int needed
, allocated
;
1638 bool alloc_ok
= true;
1640 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1642 h
->resv_huge_pages
+= delta
;
1647 INIT_LIST_HEAD(&surplus_list
);
1651 spin_unlock(&hugetlb_lock
);
1652 for (i
= 0; i
< needed
; i
++) {
1653 page
= __alloc_buddy_huge_page_no_mpol(h
, NUMA_NO_NODE
);
1658 list_add(&page
->lru
, &surplus_list
);
1663 * After retaking hugetlb_lock, we need to recalculate 'needed'
1664 * because either resv_huge_pages or free_huge_pages may have changed.
1666 spin_lock(&hugetlb_lock
);
1667 needed
= (h
->resv_huge_pages
+ delta
) -
1668 (h
->free_huge_pages
+ allocated
);
1673 * We were not able to allocate enough pages to
1674 * satisfy the entire reservation so we free what
1675 * we've allocated so far.
1680 * The surplus_list now contains _at_least_ the number of extra pages
1681 * needed to accommodate the reservation. Add the appropriate number
1682 * of pages to the hugetlb pool and free the extras back to the buddy
1683 * allocator. Commit the entire reservation here to prevent another
1684 * process from stealing the pages as they are added to the pool but
1685 * before they are reserved.
1687 needed
+= allocated
;
1688 h
->resv_huge_pages
+= delta
;
1691 /* Free the needed pages to the hugetlb pool */
1692 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1696 * This page is now managed by the hugetlb allocator and has
1697 * no users -- drop the buddy allocator's reference.
1699 put_page_testzero(page
);
1700 VM_BUG_ON_PAGE(page_count(page
), page
);
1701 enqueue_huge_page(h
, page
);
1704 spin_unlock(&hugetlb_lock
);
1706 /* Free unnecessary surplus pages to the buddy allocator */
1707 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1709 spin_lock(&hugetlb_lock
);
1715 * When releasing a hugetlb pool reservation, any surplus pages that were
1716 * allocated to satisfy the reservation must be explicitly freed if they were
1718 * Called with hugetlb_lock held.
1720 static void return_unused_surplus_pages(struct hstate
*h
,
1721 unsigned long unused_resv_pages
)
1723 unsigned long nr_pages
;
1725 /* Uncommit the reservation */
1726 h
->resv_huge_pages
-= unused_resv_pages
;
1728 /* Cannot return gigantic pages currently */
1729 if (hstate_is_gigantic(h
))
1732 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1735 * We want to release as many surplus pages as possible, spread
1736 * evenly across all nodes with memory. Iterate across these nodes
1737 * until we can no longer free unreserved surplus pages. This occurs
1738 * when the nodes with surplus pages have no free pages.
1739 * free_pool_huge_page() will balance the the freed pages across the
1740 * on-line nodes with memory and will handle the hstate accounting.
1742 while (nr_pages
--) {
1743 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1745 cond_resched_lock(&hugetlb_lock
);
1751 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1752 * are used by the huge page allocation routines to manage reservations.
1754 * vma_needs_reservation is called to determine if the huge page at addr
1755 * within the vma has an associated reservation. If a reservation is
1756 * needed, the value 1 is returned. The caller is then responsible for
1757 * managing the global reservation and subpool usage counts. After
1758 * the huge page has been allocated, vma_commit_reservation is called
1759 * to add the page to the reservation map. If the page allocation fails,
1760 * the reservation must be ended instead of committed. vma_end_reservation
1761 * is called in such cases.
1763 * In the normal case, vma_commit_reservation returns the same value
1764 * as the preceding vma_needs_reservation call. The only time this
1765 * is not the case is if a reserve map was changed between calls. It
1766 * is the responsibility of the caller to notice the difference and
1767 * take appropriate action.
1769 enum vma_resv_mode
{
1774 static long __vma_reservation_common(struct hstate
*h
,
1775 struct vm_area_struct
*vma
, unsigned long addr
,
1776 enum vma_resv_mode mode
)
1778 struct resv_map
*resv
;
1782 resv
= vma_resv_map(vma
);
1786 idx
= vma_hugecache_offset(h
, vma
, addr
);
1788 case VMA_NEEDS_RESV
:
1789 ret
= region_chg(resv
, idx
, idx
+ 1);
1791 case VMA_COMMIT_RESV
:
1792 ret
= region_add(resv
, idx
, idx
+ 1);
1795 region_abort(resv
, idx
, idx
+ 1);
1802 if (vma
->vm_flags
& VM_MAYSHARE
)
1805 return ret
< 0 ? ret
: 0;
1808 static long vma_needs_reservation(struct hstate
*h
,
1809 struct vm_area_struct
*vma
, unsigned long addr
)
1811 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
1814 static long vma_commit_reservation(struct hstate
*h
,
1815 struct vm_area_struct
*vma
, unsigned long addr
)
1817 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
1820 static void vma_end_reservation(struct hstate
*h
,
1821 struct vm_area_struct
*vma
, unsigned long addr
)
1823 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
1826 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1827 unsigned long addr
, int avoid_reserve
)
1829 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1830 struct hstate
*h
= hstate_vma(vma
);
1832 long map_chg
, map_commit
;
1835 struct hugetlb_cgroup
*h_cg
;
1837 idx
= hstate_index(h
);
1839 * Examine the region/reserve map to determine if the process
1840 * has a reservation for the page to be allocated. A return
1841 * code of zero indicates a reservation exists (no change).
1843 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
1845 return ERR_PTR(-ENOMEM
);
1848 * Processes that did not create the mapping will have no
1849 * reserves as indicated by the region/reserve map. Check
1850 * that the allocation will not exceed the subpool limit.
1851 * Allocations for MAP_NORESERVE mappings also need to be
1852 * checked against any subpool limit.
1854 if (map_chg
|| avoid_reserve
) {
1855 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
1857 vma_end_reservation(h
, vma
, addr
);
1858 return ERR_PTR(-ENOSPC
);
1862 * Even though there was no reservation in the region/reserve
1863 * map, there could be reservations associated with the
1864 * subpool that can be used. This would be indicated if the
1865 * return value of hugepage_subpool_get_pages() is zero.
1866 * However, if avoid_reserve is specified we still avoid even
1867 * the subpool reservations.
1873 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
1875 goto out_subpool_put
;
1877 spin_lock(&hugetlb_lock
);
1879 * glb_chg is passed to indicate whether or not a page must be taken
1880 * from the global free pool (global change). gbl_chg == 0 indicates
1881 * a reservation exists for the allocation.
1883 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
1885 spin_unlock(&hugetlb_lock
);
1886 page
= __alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
1888 goto out_uncharge_cgroup
;
1890 spin_lock(&hugetlb_lock
);
1891 list_move(&page
->lru
, &h
->hugepage_activelist
);
1894 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
1895 spin_unlock(&hugetlb_lock
);
1897 set_page_private(page
, (unsigned long)spool
);
1899 map_commit
= vma_commit_reservation(h
, vma
, addr
);
1900 if (unlikely(map_chg
> map_commit
)) {
1902 * The page was added to the reservation map between
1903 * vma_needs_reservation and vma_commit_reservation.
1904 * This indicates a race with hugetlb_reserve_pages.
1905 * Adjust for the subpool count incremented above AND
1906 * in hugetlb_reserve_pages for the same page. Also,
1907 * the reservation count added in hugetlb_reserve_pages
1908 * no longer applies.
1912 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
1913 hugetlb_acct_memory(h
, -rsv_adjust
);
1917 out_uncharge_cgroup
:
1918 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
1920 if (map_chg
|| avoid_reserve
)
1921 hugepage_subpool_put_pages(spool
, 1);
1922 vma_end_reservation(h
, vma
, addr
);
1923 return ERR_PTR(-ENOSPC
);
1927 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1928 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1929 * where no ERR_VALUE is expected to be returned.
1931 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
1932 unsigned long addr
, int avoid_reserve
)
1934 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
1940 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
1942 struct huge_bootmem_page
*m
;
1945 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
1948 addr
= memblock_virt_alloc_try_nid_nopanic(
1949 huge_page_size(h
), huge_page_size(h
),
1950 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
1953 * Use the beginning of the huge page to store the
1954 * huge_bootmem_page struct (until gather_bootmem
1955 * puts them into the mem_map).
1964 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
1965 /* Put them into a private list first because mem_map is not up yet */
1966 list_add(&m
->list
, &huge_boot_pages
);
1971 static void __init
prep_compound_huge_page(struct page
*page
, int order
)
1973 if (unlikely(order
> (MAX_ORDER
- 1)))
1974 prep_compound_gigantic_page(page
, order
);
1976 prep_compound_page(page
, order
);
1979 /* Put bootmem huge pages into the standard lists after mem_map is up */
1980 static void __init
gather_bootmem_prealloc(void)
1982 struct huge_bootmem_page
*m
;
1984 list_for_each_entry(m
, &huge_boot_pages
, list
) {
1985 struct hstate
*h
= m
->hstate
;
1988 #ifdef CONFIG_HIGHMEM
1989 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
1990 memblock_free_late(__pa(m
),
1991 sizeof(struct huge_bootmem_page
));
1993 page
= virt_to_page(m
);
1995 WARN_ON(page_count(page
) != 1);
1996 prep_compound_huge_page(page
, h
->order
);
1997 WARN_ON(PageReserved(page
));
1998 prep_new_huge_page(h
, page
, page_to_nid(page
));
2000 * If we had gigantic hugepages allocated at boot time, we need
2001 * to restore the 'stolen' pages to totalram_pages in order to
2002 * fix confusing memory reports from free(1) and another
2003 * side-effects, like CommitLimit going negative.
2005 if (hstate_is_gigantic(h
))
2006 adjust_managed_page_count(page
, 1 << h
->order
);
2010 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2014 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2015 if (hstate_is_gigantic(h
)) {
2016 if (!alloc_bootmem_huge_page(h
))
2018 } else if (!alloc_fresh_huge_page(h
,
2019 &node_states
[N_MEMORY
]))
2022 h
->max_huge_pages
= i
;
2025 static void __init
hugetlb_init_hstates(void)
2029 for_each_hstate(h
) {
2030 if (minimum_order
> huge_page_order(h
))
2031 minimum_order
= huge_page_order(h
);
2033 /* oversize hugepages were init'ed in early boot */
2034 if (!hstate_is_gigantic(h
))
2035 hugetlb_hstate_alloc_pages(h
);
2037 VM_BUG_ON(minimum_order
== UINT_MAX
);
2040 static char * __init
memfmt(char *buf
, unsigned long n
)
2042 if (n
>= (1UL << 30))
2043 sprintf(buf
, "%lu GB", n
>> 30);
2044 else if (n
>= (1UL << 20))
2045 sprintf(buf
, "%lu MB", n
>> 20);
2047 sprintf(buf
, "%lu KB", n
>> 10);
2051 static void __init
report_hugepages(void)
2055 for_each_hstate(h
) {
2057 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2058 memfmt(buf
, huge_page_size(h
)),
2059 h
->free_huge_pages
);
2063 #ifdef CONFIG_HIGHMEM
2064 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2065 nodemask_t
*nodes_allowed
)
2069 if (hstate_is_gigantic(h
))
2072 for_each_node_mask(i
, *nodes_allowed
) {
2073 struct page
*page
, *next
;
2074 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2075 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2076 if (count
>= h
->nr_huge_pages
)
2078 if (PageHighMem(page
))
2080 list_del(&page
->lru
);
2081 update_and_free_page(h
, page
);
2082 h
->free_huge_pages
--;
2083 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2088 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2089 nodemask_t
*nodes_allowed
)
2095 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2096 * balanced by operating on them in a round-robin fashion.
2097 * Returns 1 if an adjustment was made.
2099 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2104 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2107 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2108 if (h
->surplus_huge_pages_node
[node
])
2112 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2113 if (h
->surplus_huge_pages_node
[node
] <
2114 h
->nr_huge_pages_node
[node
])
2121 h
->surplus_huge_pages
+= delta
;
2122 h
->surplus_huge_pages_node
[node
] += delta
;
2126 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2127 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
2128 nodemask_t
*nodes_allowed
)
2130 unsigned long min_count
, ret
;
2132 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
2133 return h
->max_huge_pages
;
2136 * Increase the pool size
2137 * First take pages out of surplus state. Then make up the
2138 * remaining difference by allocating fresh huge pages.
2140 * We might race with alloc_buddy_huge_page() here and be unable
2141 * to convert a surplus huge page to a normal huge page. That is
2142 * not critical, though, it just means the overall size of the
2143 * pool might be one hugepage larger than it needs to be, but
2144 * within all the constraints specified by the sysctls.
2146 spin_lock(&hugetlb_lock
);
2147 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2148 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2152 while (count
> persistent_huge_pages(h
)) {
2154 * If this allocation races such that we no longer need the
2155 * page, free_huge_page will handle it by freeing the page
2156 * and reducing the surplus.
2158 spin_unlock(&hugetlb_lock
);
2159 if (hstate_is_gigantic(h
))
2160 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
2162 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
2163 spin_lock(&hugetlb_lock
);
2167 /* Bail for signals. Probably ctrl-c from user */
2168 if (signal_pending(current
))
2173 * Decrease the pool size
2174 * First return free pages to the buddy allocator (being careful
2175 * to keep enough around to satisfy reservations). Then place
2176 * pages into surplus state as needed so the pool will shrink
2177 * to the desired size as pages become free.
2179 * By placing pages into the surplus state independent of the
2180 * overcommit value, we are allowing the surplus pool size to
2181 * exceed overcommit. There are few sane options here. Since
2182 * alloc_buddy_huge_page() is checking the global counter,
2183 * though, we'll note that we're not allowed to exceed surplus
2184 * and won't grow the pool anywhere else. Not until one of the
2185 * sysctls are changed, or the surplus pages go out of use.
2187 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2188 min_count
= max(count
, min_count
);
2189 try_to_free_low(h
, min_count
, nodes_allowed
);
2190 while (min_count
< persistent_huge_pages(h
)) {
2191 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2193 cond_resched_lock(&hugetlb_lock
);
2195 while (count
< persistent_huge_pages(h
)) {
2196 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2200 ret
= persistent_huge_pages(h
);
2201 spin_unlock(&hugetlb_lock
);
2205 #define HSTATE_ATTR_RO(_name) \
2206 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2208 #define HSTATE_ATTR(_name) \
2209 static struct kobj_attribute _name##_attr = \
2210 __ATTR(_name, 0644, _name##_show, _name##_store)
2212 static struct kobject
*hugepages_kobj
;
2213 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2215 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2217 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2221 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2222 if (hstate_kobjs
[i
] == kobj
) {
2224 *nidp
= NUMA_NO_NODE
;
2228 return kobj_to_node_hstate(kobj
, nidp
);
2231 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2232 struct kobj_attribute
*attr
, char *buf
)
2235 unsigned long nr_huge_pages
;
2238 h
= kobj_to_hstate(kobj
, &nid
);
2239 if (nid
== NUMA_NO_NODE
)
2240 nr_huge_pages
= h
->nr_huge_pages
;
2242 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2244 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2247 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2248 struct hstate
*h
, int nid
,
2249 unsigned long count
, size_t len
)
2252 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
2254 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
2259 if (nid
== NUMA_NO_NODE
) {
2261 * global hstate attribute
2263 if (!(obey_mempolicy
&&
2264 init_nodemask_of_mempolicy(nodes_allowed
))) {
2265 NODEMASK_FREE(nodes_allowed
);
2266 nodes_allowed
= &node_states
[N_MEMORY
];
2268 } else if (nodes_allowed
) {
2270 * per node hstate attribute: adjust count to global,
2271 * but restrict alloc/free to the specified node.
2273 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2274 init_nodemask_of_node(nodes_allowed
, nid
);
2276 nodes_allowed
= &node_states
[N_MEMORY
];
2278 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
2280 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2281 NODEMASK_FREE(nodes_allowed
);
2285 NODEMASK_FREE(nodes_allowed
);
2289 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2290 struct kobject
*kobj
, const char *buf
,
2294 unsigned long count
;
2298 err
= kstrtoul(buf
, 10, &count
);
2302 h
= kobj_to_hstate(kobj
, &nid
);
2303 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2306 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2307 struct kobj_attribute
*attr
, char *buf
)
2309 return nr_hugepages_show_common(kobj
, attr
, buf
);
2312 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2313 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2315 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2317 HSTATE_ATTR(nr_hugepages
);
2322 * hstate attribute for optionally mempolicy-based constraint on persistent
2323 * huge page alloc/free.
2325 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2326 struct kobj_attribute
*attr
, char *buf
)
2328 return nr_hugepages_show_common(kobj
, attr
, buf
);
2331 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2332 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2334 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2336 HSTATE_ATTR(nr_hugepages_mempolicy
);
2340 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2341 struct kobj_attribute
*attr
, char *buf
)
2343 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2344 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2347 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2348 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2351 unsigned long input
;
2352 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2354 if (hstate_is_gigantic(h
))
2357 err
= kstrtoul(buf
, 10, &input
);
2361 spin_lock(&hugetlb_lock
);
2362 h
->nr_overcommit_huge_pages
= input
;
2363 spin_unlock(&hugetlb_lock
);
2367 HSTATE_ATTR(nr_overcommit_hugepages
);
2369 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2370 struct kobj_attribute
*attr
, char *buf
)
2373 unsigned long free_huge_pages
;
2376 h
= kobj_to_hstate(kobj
, &nid
);
2377 if (nid
== NUMA_NO_NODE
)
2378 free_huge_pages
= h
->free_huge_pages
;
2380 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2382 return sprintf(buf
, "%lu\n", free_huge_pages
);
2384 HSTATE_ATTR_RO(free_hugepages
);
2386 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2387 struct kobj_attribute
*attr
, char *buf
)
2389 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2390 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2392 HSTATE_ATTR_RO(resv_hugepages
);
2394 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2395 struct kobj_attribute
*attr
, char *buf
)
2398 unsigned long surplus_huge_pages
;
2401 h
= kobj_to_hstate(kobj
, &nid
);
2402 if (nid
== NUMA_NO_NODE
)
2403 surplus_huge_pages
= h
->surplus_huge_pages
;
2405 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2407 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2409 HSTATE_ATTR_RO(surplus_hugepages
);
2411 static struct attribute
*hstate_attrs
[] = {
2412 &nr_hugepages_attr
.attr
,
2413 &nr_overcommit_hugepages_attr
.attr
,
2414 &free_hugepages_attr
.attr
,
2415 &resv_hugepages_attr
.attr
,
2416 &surplus_hugepages_attr
.attr
,
2418 &nr_hugepages_mempolicy_attr
.attr
,
2423 static struct attribute_group hstate_attr_group
= {
2424 .attrs
= hstate_attrs
,
2427 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2428 struct kobject
**hstate_kobjs
,
2429 struct attribute_group
*hstate_attr_group
)
2432 int hi
= hstate_index(h
);
2434 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2435 if (!hstate_kobjs
[hi
])
2438 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2440 kobject_put(hstate_kobjs
[hi
]);
2445 static void __init
hugetlb_sysfs_init(void)
2450 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2451 if (!hugepages_kobj
)
2454 for_each_hstate(h
) {
2455 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2456 hstate_kobjs
, &hstate_attr_group
);
2458 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2465 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2466 * with node devices in node_devices[] using a parallel array. The array
2467 * index of a node device or _hstate == node id.
2468 * This is here to avoid any static dependency of the node device driver, in
2469 * the base kernel, on the hugetlb module.
2471 struct node_hstate
{
2472 struct kobject
*hugepages_kobj
;
2473 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2475 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2478 * A subset of global hstate attributes for node devices
2480 static struct attribute
*per_node_hstate_attrs
[] = {
2481 &nr_hugepages_attr
.attr
,
2482 &free_hugepages_attr
.attr
,
2483 &surplus_hugepages_attr
.attr
,
2487 static struct attribute_group per_node_hstate_attr_group
= {
2488 .attrs
= per_node_hstate_attrs
,
2492 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2493 * Returns node id via non-NULL nidp.
2495 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2499 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2500 struct node_hstate
*nhs
= &node_hstates
[nid
];
2502 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2503 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2515 * Unregister hstate attributes from a single node device.
2516 * No-op if no hstate attributes attached.
2518 static void hugetlb_unregister_node(struct node
*node
)
2521 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2523 if (!nhs
->hugepages_kobj
)
2524 return; /* no hstate attributes */
2526 for_each_hstate(h
) {
2527 int idx
= hstate_index(h
);
2528 if (nhs
->hstate_kobjs
[idx
]) {
2529 kobject_put(nhs
->hstate_kobjs
[idx
]);
2530 nhs
->hstate_kobjs
[idx
] = NULL
;
2534 kobject_put(nhs
->hugepages_kobj
);
2535 nhs
->hugepages_kobj
= NULL
;
2539 * hugetlb module exit: unregister hstate attributes from node devices
2542 static void hugetlb_unregister_all_nodes(void)
2547 * disable node device registrations.
2549 register_hugetlbfs_with_node(NULL
, NULL
);
2552 * remove hstate attributes from any nodes that have them.
2554 for (nid
= 0; nid
< nr_node_ids
; nid
++)
2555 hugetlb_unregister_node(node_devices
[nid
]);
2559 * Register hstate attributes for a single node device.
2560 * No-op if attributes already registered.
2562 static void hugetlb_register_node(struct node
*node
)
2565 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2568 if (nhs
->hugepages_kobj
)
2569 return; /* already allocated */
2571 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2573 if (!nhs
->hugepages_kobj
)
2576 for_each_hstate(h
) {
2577 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2579 &per_node_hstate_attr_group
);
2581 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2582 h
->name
, node
->dev
.id
);
2583 hugetlb_unregister_node(node
);
2590 * hugetlb init time: register hstate attributes for all registered node
2591 * devices of nodes that have memory. All on-line nodes should have
2592 * registered their associated device by this time.
2594 static void __init
hugetlb_register_all_nodes(void)
2598 for_each_node_state(nid
, N_MEMORY
) {
2599 struct node
*node
= node_devices
[nid
];
2600 if (node
->dev
.id
== nid
)
2601 hugetlb_register_node(node
);
2605 * Let the node device driver know we're here so it can
2606 * [un]register hstate attributes on node hotplug.
2608 register_hugetlbfs_with_node(hugetlb_register_node
,
2609 hugetlb_unregister_node
);
2611 #else /* !CONFIG_NUMA */
2613 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2621 static void hugetlb_unregister_all_nodes(void) { }
2623 static void hugetlb_register_all_nodes(void) { }
2627 static void __exit
hugetlb_exit(void)
2631 hugetlb_unregister_all_nodes();
2633 for_each_hstate(h
) {
2634 kobject_put(hstate_kobjs
[hstate_index(h
)]);
2637 kobject_put(hugepages_kobj
);
2638 kfree(hugetlb_fault_mutex_table
);
2640 module_exit(hugetlb_exit
);
2642 static int __init
hugetlb_init(void)
2646 if (!hugepages_supported())
2649 if (!size_to_hstate(default_hstate_size
)) {
2650 default_hstate_size
= HPAGE_SIZE
;
2651 if (!size_to_hstate(default_hstate_size
))
2652 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2654 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2655 if (default_hstate_max_huge_pages
)
2656 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2658 hugetlb_init_hstates();
2659 gather_bootmem_prealloc();
2662 hugetlb_sysfs_init();
2663 hugetlb_register_all_nodes();
2664 hugetlb_cgroup_file_init();
2667 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2669 num_fault_mutexes
= 1;
2671 hugetlb_fault_mutex_table
=
2672 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2673 BUG_ON(!hugetlb_fault_mutex_table
);
2675 for (i
= 0; i
< num_fault_mutexes
; i
++)
2676 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2679 module_init(hugetlb_init
);
2681 /* Should be called on processing a hugepagesz=... option */
2682 void __init
hugetlb_add_hstate(unsigned order
)
2687 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2688 pr_warning("hugepagesz= specified twice, ignoring\n");
2691 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2693 h
= &hstates
[hugetlb_max_hstate
++];
2695 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2696 h
->nr_huge_pages
= 0;
2697 h
->free_huge_pages
= 0;
2698 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2699 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2700 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2701 h
->next_nid_to_alloc
= first_node(node_states
[N_MEMORY
]);
2702 h
->next_nid_to_free
= first_node(node_states
[N_MEMORY
]);
2703 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2704 huge_page_size(h
)/1024);
2709 static int __init
hugetlb_nrpages_setup(char *s
)
2712 static unsigned long *last_mhp
;
2715 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2716 * so this hugepages= parameter goes to the "default hstate".
2718 if (!hugetlb_max_hstate
)
2719 mhp
= &default_hstate_max_huge_pages
;
2721 mhp
= &parsed_hstate
->max_huge_pages
;
2723 if (mhp
== last_mhp
) {
2724 pr_warning("hugepages= specified twice without "
2725 "interleaving hugepagesz=, ignoring\n");
2729 if (sscanf(s
, "%lu", mhp
) <= 0)
2733 * Global state is always initialized later in hugetlb_init.
2734 * But we need to allocate >= MAX_ORDER hstates here early to still
2735 * use the bootmem allocator.
2737 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2738 hugetlb_hstate_alloc_pages(parsed_hstate
);
2744 __setup("hugepages=", hugetlb_nrpages_setup
);
2746 static int __init
hugetlb_default_setup(char *s
)
2748 default_hstate_size
= memparse(s
, &s
);
2751 __setup("default_hugepagesz=", hugetlb_default_setup
);
2753 static unsigned int cpuset_mems_nr(unsigned int *array
)
2756 unsigned int nr
= 0;
2758 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2764 #ifdef CONFIG_SYSCTL
2765 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2766 struct ctl_table
*table
, int write
,
2767 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2769 struct hstate
*h
= &default_hstate
;
2770 unsigned long tmp
= h
->max_huge_pages
;
2773 if (!hugepages_supported())
2777 table
->maxlen
= sizeof(unsigned long);
2778 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2783 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2784 NUMA_NO_NODE
, tmp
, *length
);
2789 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2790 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2793 return hugetlb_sysctl_handler_common(false, table
, write
,
2794 buffer
, length
, ppos
);
2798 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2799 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2801 return hugetlb_sysctl_handler_common(true, table
, write
,
2802 buffer
, length
, ppos
);
2804 #endif /* CONFIG_NUMA */
2806 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2807 void __user
*buffer
,
2808 size_t *length
, loff_t
*ppos
)
2810 struct hstate
*h
= &default_hstate
;
2814 if (!hugepages_supported())
2817 tmp
= h
->nr_overcommit_huge_pages
;
2819 if (write
&& hstate_is_gigantic(h
))
2823 table
->maxlen
= sizeof(unsigned long);
2824 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2829 spin_lock(&hugetlb_lock
);
2830 h
->nr_overcommit_huge_pages
= tmp
;
2831 spin_unlock(&hugetlb_lock
);
2837 #endif /* CONFIG_SYSCTL */
2839 void hugetlb_report_meminfo(struct seq_file
*m
)
2841 struct hstate
*h
= &default_hstate
;
2842 if (!hugepages_supported())
2845 "HugePages_Total: %5lu\n"
2846 "HugePages_Free: %5lu\n"
2847 "HugePages_Rsvd: %5lu\n"
2848 "HugePages_Surp: %5lu\n"
2849 "Hugepagesize: %8lu kB\n",
2853 h
->surplus_huge_pages
,
2854 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2857 int hugetlb_report_node_meminfo(int nid
, char *buf
)
2859 struct hstate
*h
= &default_hstate
;
2860 if (!hugepages_supported())
2863 "Node %d HugePages_Total: %5u\n"
2864 "Node %d HugePages_Free: %5u\n"
2865 "Node %d HugePages_Surp: %5u\n",
2866 nid
, h
->nr_huge_pages_node
[nid
],
2867 nid
, h
->free_huge_pages_node
[nid
],
2868 nid
, h
->surplus_huge_pages_node
[nid
]);
2871 void hugetlb_show_meminfo(void)
2876 if (!hugepages_supported())
2879 for_each_node_state(nid
, N_MEMORY
)
2881 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2883 h
->nr_huge_pages_node
[nid
],
2884 h
->free_huge_pages_node
[nid
],
2885 h
->surplus_huge_pages_node
[nid
],
2886 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
2889 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
2891 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
2892 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
2895 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2896 unsigned long hugetlb_total_pages(void)
2899 unsigned long nr_total_pages
= 0;
2902 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
2903 return nr_total_pages
;
2906 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
2910 spin_lock(&hugetlb_lock
);
2912 * When cpuset is configured, it breaks the strict hugetlb page
2913 * reservation as the accounting is done on a global variable. Such
2914 * reservation is completely rubbish in the presence of cpuset because
2915 * the reservation is not checked against page availability for the
2916 * current cpuset. Application can still potentially OOM'ed by kernel
2917 * with lack of free htlb page in cpuset that the task is in.
2918 * Attempt to enforce strict accounting with cpuset is almost
2919 * impossible (or too ugly) because cpuset is too fluid that
2920 * task or memory node can be dynamically moved between cpusets.
2922 * The change of semantics for shared hugetlb mapping with cpuset is
2923 * undesirable. However, in order to preserve some of the semantics,
2924 * we fall back to check against current free page availability as
2925 * a best attempt and hopefully to minimize the impact of changing
2926 * semantics that cpuset has.
2929 if (gather_surplus_pages(h
, delta
) < 0)
2932 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
2933 return_unused_surplus_pages(h
, delta
);
2940 return_unused_surplus_pages(h
, (unsigned long) -delta
);
2943 spin_unlock(&hugetlb_lock
);
2947 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
2949 struct resv_map
*resv
= vma_resv_map(vma
);
2952 * This new VMA should share its siblings reservation map if present.
2953 * The VMA will only ever have a valid reservation map pointer where
2954 * it is being copied for another still existing VMA. As that VMA
2955 * has a reference to the reservation map it cannot disappear until
2956 * after this open call completes. It is therefore safe to take a
2957 * new reference here without additional locking.
2959 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2960 kref_get(&resv
->refs
);
2963 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
2965 struct hstate
*h
= hstate_vma(vma
);
2966 struct resv_map
*resv
= vma_resv_map(vma
);
2967 struct hugepage_subpool
*spool
= subpool_vma(vma
);
2968 unsigned long reserve
, start
, end
;
2971 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
2974 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
2975 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
2977 reserve
= (end
- start
) - region_count(resv
, start
, end
);
2979 kref_put(&resv
->refs
, resv_map_release
);
2983 * Decrement reserve counts. The global reserve count may be
2984 * adjusted if the subpool has a minimum size.
2986 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
2987 hugetlb_acct_memory(h
, -gbl_reserve
);
2992 * We cannot handle pagefaults against hugetlb pages at all. They cause
2993 * handle_mm_fault() to try to instantiate regular-sized pages in the
2994 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2997 static int hugetlb_vm_op_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
3003 const struct vm_operations_struct hugetlb_vm_ops
= {
3004 .fault
= hugetlb_vm_op_fault
,
3005 .open
= hugetlb_vm_op_open
,
3006 .close
= hugetlb_vm_op_close
,
3009 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3015 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3016 vma
->vm_page_prot
)));
3018 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3019 vma
->vm_page_prot
));
3021 entry
= pte_mkyoung(entry
);
3022 entry
= pte_mkhuge(entry
);
3023 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3028 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3029 unsigned long address
, pte_t
*ptep
)
3033 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3034 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3035 update_mmu_cache(vma
, address
, ptep
);
3038 static int is_hugetlb_entry_migration(pte_t pte
)
3042 if (huge_pte_none(pte
) || pte_present(pte
))
3044 swp
= pte_to_swp_entry(pte
);
3045 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3051 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3055 if (huge_pte_none(pte
) || pte_present(pte
))
3057 swp
= pte_to_swp_entry(pte
);
3058 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3064 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3065 struct vm_area_struct
*vma
)
3067 pte_t
*src_pte
, *dst_pte
, entry
;
3068 struct page
*ptepage
;
3071 struct hstate
*h
= hstate_vma(vma
);
3072 unsigned long sz
= huge_page_size(h
);
3073 unsigned long mmun_start
; /* For mmu_notifiers */
3074 unsigned long mmun_end
; /* For mmu_notifiers */
3077 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3079 mmun_start
= vma
->vm_start
;
3080 mmun_end
= vma
->vm_end
;
3082 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
3084 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3085 spinlock_t
*src_ptl
, *dst_ptl
;
3086 src_pte
= huge_pte_offset(src
, addr
);
3089 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3095 /* If the pagetables are shared don't copy or take references */
3096 if (dst_pte
== src_pte
)
3099 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3100 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3101 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3102 entry
= huge_ptep_get(src_pte
);
3103 if (huge_pte_none(entry
)) { /* skip none entry */
3105 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3106 is_hugetlb_entry_hwpoisoned(entry
))) {
3107 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3109 if (is_write_migration_entry(swp_entry
) && cow
) {
3111 * COW mappings require pages in both
3112 * parent and child to be set to read.
3114 make_migration_entry_read(&swp_entry
);
3115 entry
= swp_entry_to_pte(swp_entry
);
3116 set_huge_pte_at(src
, addr
, src_pte
, entry
);
3118 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3121 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3122 mmu_notifier_invalidate_range(src
, mmun_start
,
3125 entry
= huge_ptep_get(src_pte
);
3126 ptepage
= pte_page(entry
);
3128 page_dup_rmap(ptepage
);
3129 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3130 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3132 spin_unlock(src_ptl
);
3133 spin_unlock(dst_ptl
);
3137 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
3142 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3143 unsigned long start
, unsigned long end
,
3144 struct page
*ref_page
)
3146 int force_flush
= 0;
3147 struct mm_struct
*mm
= vma
->vm_mm
;
3148 unsigned long address
;
3153 struct hstate
*h
= hstate_vma(vma
);
3154 unsigned long sz
= huge_page_size(h
);
3155 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
3156 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
3158 WARN_ON(!is_vm_hugetlb_page(vma
));
3159 BUG_ON(start
& ~huge_page_mask(h
));
3160 BUG_ON(end
& ~huge_page_mask(h
));
3162 tlb_start_vma(tlb
, vma
);
3163 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3166 for (; address
< end
; address
+= sz
) {
3167 ptep
= huge_pte_offset(mm
, address
);
3171 ptl
= huge_pte_lock(h
, mm
, ptep
);
3172 if (huge_pmd_unshare(mm
, &address
, ptep
))
3175 pte
= huge_ptep_get(ptep
);
3176 if (huge_pte_none(pte
))
3180 * Migrating hugepage or HWPoisoned hugepage is already
3181 * unmapped and its refcount is dropped, so just clear pte here.
3183 if (unlikely(!pte_present(pte
))) {
3184 huge_pte_clear(mm
, address
, ptep
);
3188 page
= pte_page(pte
);
3190 * If a reference page is supplied, it is because a specific
3191 * page is being unmapped, not a range. Ensure the page we
3192 * are about to unmap is the actual page of interest.
3195 if (page
!= ref_page
)
3199 * Mark the VMA as having unmapped its page so that
3200 * future faults in this VMA will fail rather than
3201 * looking like data was lost
3203 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3206 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3207 tlb_remove_tlb_entry(tlb
, ptep
, address
);
3208 if (huge_pte_dirty(pte
))
3209 set_page_dirty(page
);
3211 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3212 page_remove_rmap(page
);
3213 force_flush
= !__tlb_remove_page(tlb
, page
);
3219 /* Bail out after unmapping reference page if supplied */
3228 * mmu_gather ran out of room to batch pages, we break out of
3229 * the PTE lock to avoid doing the potential expensive TLB invalidate
3230 * and page-free while holding it.
3235 if (address
< end
&& !ref_page
)
3238 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3239 tlb_end_vma(tlb
, vma
);
3242 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3243 struct vm_area_struct
*vma
, unsigned long start
,
3244 unsigned long end
, struct page
*ref_page
)
3246 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3249 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3250 * test will fail on a vma being torn down, and not grab a page table
3251 * on its way out. We're lucky that the flag has such an appropriate
3252 * name, and can in fact be safely cleared here. We could clear it
3253 * before the __unmap_hugepage_range above, but all that's necessary
3254 * is to clear it before releasing the i_mmap_rwsem. This works
3255 * because in the context this is called, the VMA is about to be
3256 * destroyed and the i_mmap_rwsem is held.
3258 vma
->vm_flags
&= ~VM_MAYSHARE
;
3261 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3262 unsigned long end
, struct page
*ref_page
)
3264 struct mm_struct
*mm
;
3265 struct mmu_gather tlb
;
3269 tlb_gather_mmu(&tlb
, mm
, start
, end
);
3270 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3271 tlb_finish_mmu(&tlb
, start
, end
);
3275 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3276 * mappping it owns the reserve page for. The intention is to unmap the page
3277 * from other VMAs and let the children be SIGKILLed if they are faulting the
3280 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3281 struct page
*page
, unsigned long address
)
3283 struct hstate
*h
= hstate_vma(vma
);
3284 struct vm_area_struct
*iter_vma
;
3285 struct address_space
*mapping
;
3289 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3290 * from page cache lookup which is in HPAGE_SIZE units.
3292 address
= address
& huge_page_mask(h
);
3293 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3295 mapping
= file_inode(vma
->vm_file
)->i_mapping
;
3298 * Take the mapping lock for the duration of the table walk. As
3299 * this mapping should be shared between all the VMAs,
3300 * __unmap_hugepage_range() is called as the lock is already held
3302 i_mmap_lock_write(mapping
);
3303 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3304 /* Do not unmap the current VMA */
3305 if (iter_vma
== vma
)
3309 * Shared VMAs have their own reserves and do not affect
3310 * MAP_PRIVATE accounting but it is possible that a shared
3311 * VMA is using the same page so check and skip such VMAs.
3313 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3317 * Unmap the page from other VMAs without their own reserves.
3318 * They get marked to be SIGKILLed if they fault in these
3319 * areas. This is because a future no-page fault on this VMA
3320 * could insert a zeroed page instead of the data existing
3321 * from the time of fork. This would look like data corruption
3323 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3324 unmap_hugepage_range(iter_vma
, address
,
3325 address
+ huge_page_size(h
), page
);
3327 i_mmap_unlock_write(mapping
);
3331 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3332 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3333 * cannot race with other handlers or page migration.
3334 * Keep the pte_same checks anyway to make transition from the mutex easier.
3336 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3337 unsigned long address
, pte_t
*ptep
, pte_t pte
,
3338 struct page
*pagecache_page
, spinlock_t
*ptl
)
3340 struct hstate
*h
= hstate_vma(vma
);
3341 struct page
*old_page
, *new_page
;
3342 int ret
= 0, outside_reserve
= 0;
3343 unsigned long mmun_start
; /* For mmu_notifiers */
3344 unsigned long mmun_end
; /* For mmu_notifiers */
3346 old_page
= pte_page(pte
);
3349 /* If no-one else is actually using this page, avoid the copy
3350 * and just make the page writable */
3351 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3352 page_move_anon_rmap(old_page
, vma
, address
);
3353 set_huge_ptep_writable(vma
, address
, ptep
);
3358 * If the process that created a MAP_PRIVATE mapping is about to
3359 * perform a COW due to a shared page count, attempt to satisfy
3360 * the allocation without using the existing reserves. The pagecache
3361 * page is used to determine if the reserve at this address was
3362 * consumed or not. If reserves were used, a partial faulted mapping
3363 * at the time of fork() could consume its reserves on COW instead
3364 * of the full address range.
3366 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3367 old_page
!= pagecache_page
)
3368 outside_reserve
= 1;
3370 page_cache_get(old_page
);
3373 * Drop page table lock as buddy allocator may be called. It will
3374 * be acquired again before returning to the caller, as expected.
3377 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
3379 if (IS_ERR(new_page
)) {
3381 * If a process owning a MAP_PRIVATE mapping fails to COW,
3382 * it is due to references held by a child and an insufficient
3383 * huge page pool. To guarantee the original mappers
3384 * reliability, unmap the page from child processes. The child
3385 * may get SIGKILLed if it later faults.
3387 if (outside_reserve
) {
3388 page_cache_release(old_page
);
3389 BUG_ON(huge_pte_none(pte
));
3390 unmap_ref_private(mm
, vma
, old_page
, address
);
3391 BUG_ON(huge_pte_none(pte
));
3393 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3395 pte_same(huge_ptep_get(ptep
), pte
)))
3396 goto retry_avoidcopy
;
3398 * race occurs while re-acquiring page table
3399 * lock, and our job is done.
3404 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
3405 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
3406 goto out_release_old
;
3410 * When the original hugepage is shared one, it does not have
3411 * anon_vma prepared.
3413 if (unlikely(anon_vma_prepare(vma
))) {
3415 goto out_release_all
;
3418 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3419 pages_per_huge_page(h
));
3420 __SetPageUptodate(new_page
);
3421 set_page_huge_active(new_page
);
3423 mmun_start
= address
& huge_page_mask(h
);
3424 mmun_end
= mmun_start
+ huge_page_size(h
);
3425 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3428 * Retake the page table lock to check for racing updates
3429 * before the page tables are altered
3432 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3433 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3434 ClearPagePrivate(new_page
);
3437 huge_ptep_clear_flush(vma
, address
, ptep
);
3438 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
3439 set_huge_pte_at(mm
, address
, ptep
,
3440 make_huge_pte(vma
, new_page
, 1));
3441 page_remove_rmap(old_page
);
3442 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
3443 /* Make the old page be freed below */
3444 new_page
= old_page
;
3447 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3449 page_cache_release(new_page
);
3451 page_cache_release(old_page
);
3453 spin_lock(ptl
); /* Caller expects lock to be held */
3457 /* Return the pagecache page at a given address within a VMA */
3458 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3459 struct vm_area_struct
*vma
, unsigned long address
)
3461 struct address_space
*mapping
;
3464 mapping
= vma
->vm_file
->f_mapping
;
3465 idx
= vma_hugecache_offset(h
, vma
, address
);
3467 return find_lock_page(mapping
, idx
);
3471 * Return whether there is a pagecache page to back given address within VMA.
3472 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3474 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3475 struct vm_area_struct
*vma
, unsigned long address
)
3477 struct address_space
*mapping
;
3481 mapping
= vma
->vm_file
->f_mapping
;
3482 idx
= vma_hugecache_offset(h
, vma
, address
);
3484 page
= find_get_page(mapping
, idx
);
3487 return page
!= NULL
;
3490 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3493 struct inode
*inode
= mapping
->host
;
3494 struct hstate
*h
= hstate_inode(inode
);
3495 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3499 ClearPagePrivate(page
);
3501 spin_lock(&inode
->i_lock
);
3502 inode
->i_blocks
+= blocks_per_huge_page(h
);
3503 spin_unlock(&inode
->i_lock
);
3507 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3508 struct address_space
*mapping
, pgoff_t idx
,
3509 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3511 struct hstate
*h
= hstate_vma(vma
);
3512 int ret
= VM_FAULT_SIGBUS
;
3520 * Currently, we are forced to kill the process in the event the
3521 * original mapper has unmapped pages from the child due to a failed
3522 * COW. Warn that such a situation has occurred as it may not be obvious
3524 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3525 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3531 * Use page lock to guard against racing truncation
3532 * before we get page_table_lock.
3535 page
= find_lock_page(mapping
, idx
);
3537 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3540 page
= alloc_huge_page(vma
, address
, 0);
3542 ret
= PTR_ERR(page
);
3546 ret
= VM_FAULT_SIGBUS
;
3549 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3550 __SetPageUptodate(page
);
3551 set_page_huge_active(page
);
3553 if (vma
->vm_flags
& VM_MAYSHARE
) {
3554 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3563 if (unlikely(anon_vma_prepare(vma
))) {
3565 goto backout_unlocked
;
3571 * If memory error occurs between mmap() and fault, some process
3572 * don't have hwpoisoned swap entry for errored virtual address.
3573 * So we need to block hugepage fault by PG_hwpoison bit check.
3575 if (unlikely(PageHWPoison(page
))) {
3576 ret
= VM_FAULT_HWPOISON
|
3577 VM_FAULT_SET_HINDEX(hstate_index(h
));
3578 goto backout_unlocked
;
3583 * If we are going to COW a private mapping later, we examine the
3584 * pending reservations for this page now. This will ensure that
3585 * any allocations necessary to record that reservation occur outside
3588 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3589 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3591 goto backout_unlocked
;
3593 /* Just decrements count, does not deallocate */
3594 vma_end_reservation(h
, vma
, address
);
3597 ptl
= huge_pte_lockptr(h
, mm
, ptep
);
3599 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3604 if (!huge_pte_none(huge_ptep_get(ptep
)))
3608 ClearPagePrivate(page
);
3609 hugepage_add_new_anon_rmap(page
, vma
, address
);
3611 page_dup_rmap(page
);
3612 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3613 && (vma
->vm_flags
& VM_SHARED
)));
3614 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3616 hugetlb_count_add(pages_per_huge_page(h
), mm
);
3617 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3618 /* Optimization, do the COW without a second fault */
3619 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, new_pte
, page
, ptl
);
3636 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3637 struct vm_area_struct
*vma
,
3638 struct address_space
*mapping
,
3639 pgoff_t idx
, unsigned long address
)
3641 unsigned long key
[2];
3644 if (vma
->vm_flags
& VM_SHARED
) {
3645 key
[0] = (unsigned long) mapping
;
3648 key
[0] = (unsigned long) mm
;
3649 key
[1] = address
>> huge_page_shift(h
);
3652 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3654 return hash
& (num_fault_mutexes
- 1);
3658 * For uniprocesor systems we always use a single mutex, so just
3659 * return 0 and avoid the hashing overhead.
3661 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3662 struct vm_area_struct
*vma
,
3663 struct address_space
*mapping
,
3664 pgoff_t idx
, unsigned long address
)
3670 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3671 unsigned long address
, unsigned int flags
)
3678 struct page
*page
= NULL
;
3679 struct page
*pagecache_page
= NULL
;
3680 struct hstate
*h
= hstate_vma(vma
);
3681 struct address_space
*mapping
;
3682 int need_wait_lock
= 0;
3684 address
&= huge_page_mask(h
);
3686 ptep
= huge_pte_offset(mm
, address
);
3688 entry
= huge_ptep_get(ptep
);
3689 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3690 migration_entry_wait_huge(vma
, mm
, ptep
);
3692 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3693 return VM_FAULT_HWPOISON_LARGE
|
3694 VM_FAULT_SET_HINDEX(hstate_index(h
));
3697 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3699 return VM_FAULT_OOM
;
3701 mapping
= vma
->vm_file
->f_mapping
;
3702 idx
= vma_hugecache_offset(h
, vma
, address
);
3705 * Serialize hugepage allocation and instantiation, so that we don't
3706 * get spurious allocation failures if two CPUs race to instantiate
3707 * the same page in the page cache.
3709 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3710 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3712 entry
= huge_ptep_get(ptep
);
3713 if (huge_pte_none(entry
)) {
3714 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3721 * entry could be a migration/hwpoison entry at this point, so this
3722 * check prevents the kernel from going below assuming that we have
3723 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3724 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3727 if (!pte_present(entry
))
3731 * If we are going to COW the mapping later, we examine the pending
3732 * reservations for this page now. This will ensure that any
3733 * allocations necessary to record that reservation occur outside the
3734 * spinlock. For private mappings, we also lookup the pagecache
3735 * page now as it is used to determine if a reservation has been
3738 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3739 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3743 /* Just decrements count, does not deallocate */
3744 vma_end_reservation(h
, vma
, address
);
3746 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3747 pagecache_page
= hugetlbfs_pagecache_page(h
,
3751 ptl
= huge_pte_lock(h
, mm
, ptep
);
3753 /* Check for a racing update before calling hugetlb_cow */
3754 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3758 * hugetlb_cow() requires page locks of pte_page(entry) and
3759 * pagecache_page, so here we need take the former one
3760 * when page != pagecache_page or !pagecache_page.
3762 page
= pte_page(entry
);
3763 if (page
!= pagecache_page
)
3764 if (!trylock_page(page
)) {
3771 if (flags
& FAULT_FLAG_WRITE
) {
3772 if (!huge_pte_write(entry
)) {
3773 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, entry
,
3774 pagecache_page
, ptl
);
3777 entry
= huge_pte_mkdirty(entry
);
3779 entry
= pte_mkyoung(entry
);
3780 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3781 flags
& FAULT_FLAG_WRITE
))
3782 update_mmu_cache(vma
, address
, ptep
);
3784 if (page
!= pagecache_page
)
3790 if (pagecache_page
) {
3791 unlock_page(pagecache_page
);
3792 put_page(pagecache_page
);
3795 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3797 * Generally it's safe to hold refcount during waiting page lock. But
3798 * here we just wait to defer the next page fault to avoid busy loop and
3799 * the page is not used after unlocked before returning from the current
3800 * page fault. So we are safe from accessing freed page, even if we wait
3801 * here without taking refcount.
3804 wait_on_page_locked(page
);
3808 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3809 struct page
**pages
, struct vm_area_struct
**vmas
,
3810 unsigned long *position
, unsigned long *nr_pages
,
3811 long i
, unsigned int flags
)
3813 unsigned long pfn_offset
;
3814 unsigned long vaddr
= *position
;
3815 unsigned long remainder
= *nr_pages
;
3816 struct hstate
*h
= hstate_vma(vma
);
3818 while (vaddr
< vma
->vm_end
&& remainder
) {
3820 spinlock_t
*ptl
= NULL
;
3825 * If we have a pending SIGKILL, don't keep faulting pages and
3826 * potentially allocating memory.
3828 if (unlikely(fatal_signal_pending(current
))) {
3834 * Some archs (sparc64, sh*) have multiple pte_ts to
3835 * each hugepage. We have to make sure we get the
3836 * first, for the page indexing below to work.
3838 * Note that page table lock is not held when pte is null.
3840 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
3842 ptl
= huge_pte_lock(h
, mm
, pte
);
3843 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
3846 * When coredumping, it suits get_dump_page if we just return
3847 * an error where there's an empty slot with no huge pagecache
3848 * to back it. This way, we avoid allocating a hugepage, and
3849 * the sparse dumpfile avoids allocating disk blocks, but its
3850 * huge holes still show up with zeroes where they need to be.
3852 if (absent
&& (flags
& FOLL_DUMP
) &&
3853 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
3861 * We need call hugetlb_fault for both hugepages under migration
3862 * (in which case hugetlb_fault waits for the migration,) and
3863 * hwpoisoned hugepages (in which case we need to prevent the
3864 * caller from accessing to them.) In order to do this, we use
3865 * here is_swap_pte instead of is_hugetlb_entry_migration and
3866 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3867 * both cases, and because we can't follow correct pages
3868 * directly from any kind of swap entries.
3870 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
3871 ((flags
& FOLL_WRITE
) &&
3872 !huge_pte_write(huge_ptep_get(pte
)))) {
3877 ret
= hugetlb_fault(mm
, vma
, vaddr
,
3878 (flags
& FOLL_WRITE
) ? FAULT_FLAG_WRITE
: 0);
3879 if (!(ret
& VM_FAULT_ERROR
))
3886 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
3887 page
= pte_page(huge_ptep_get(pte
));
3890 pages
[i
] = mem_map_offset(page
, pfn_offset
);
3891 get_page_foll(pages
[i
]);
3901 if (vaddr
< vma
->vm_end
&& remainder
&&
3902 pfn_offset
< pages_per_huge_page(h
)) {
3904 * We use pfn_offset to avoid touching the pageframes
3905 * of this compound page.
3911 *nr_pages
= remainder
;
3914 return i
? i
: -EFAULT
;
3917 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
3918 unsigned long address
, unsigned long end
, pgprot_t newprot
)
3920 struct mm_struct
*mm
= vma
->vm_mm
;
3921 unsigned long start
= address
;
3924 struct hstate
*h
= hstate_vma(vma
);
3925 unsigned long pages
= 0;
3927 BUG_ON(address
>= end
);
3928 flush_cache_range(vma
, address
, end
);
3930 mmu_notifier_invalidate_range_start(mm
, start
, end
);
3931 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
3932 for (; address
< end
; address
+= huge_page_size(h
)) {
3934 ptep
= huge_pte_offset(mm
, address
);
3937 ptl
= huge_pte_lock(h
, mm
, ptep
);
3938 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3943 pte
= huge_ptep_get(ptep
);
3944 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
3948 if (unlikely(is_hugetlb_entry_migration(pte
))) {
3949 swp_entry_t entry
= pte_to_swp_entry(pte
);
3951 if (is_write_migration_entry(entry
)) {
3954 make_migration_entry_read(&entry
);
3955 newpte
= swp_entry_to_pte(entry
);
3956 set_huge_pte_at(mm
, address
, ptep
, newpte
);
3962 if (!huge_pte_none(pte
)) {
3963 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3964 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
3965 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
3966 set_huge_pte_at(mm
, address
, ptep
, pte
);
3972 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3973 * may have cleared our pud entry and done put_page on the page table:
3974 * once we release i_mmap_rwsem, another task can do the final put_page
3975 * and that page table be reused and filled with junk.
3977 flush_tlb_range(vma
, start
, end
);
3978 mmu_notifier_invalidate_range(mm
, start
, end
);
3979 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
3980 mmu_notifier_invalidate_range_end(mm
, start
, end
);
3982 return pages
<< h
->order
;
3985 int hugetlb_reserve_pages(struct inode
*inode
,
3987 struct vm_area_struct
*vma
,
3988 vm_flags_t vm_flags
)
3991 struct hstate
*h
= hstate_inode(inode
);
3992 struct hugepage_subpool
*spool
= subpool_inode(inode
);
3993 struct resv_map
*resv_map
;
3997 * Only apply hugepage reservation if asked. At fault time, an
3998 * attempt will be made for VM_NORESERVE to allocate a page
3999 * without using reserves
4001 if (vm_flags
& VM_NORESERVE
)
4005 * Shared mappings base their reservation on the number of pages that
4006 * are already allocated on behalf of the file. Private mappings need
4007 * to reserve the full area even if read-only as mprotect() may be
4008 * called to make the mapping read-write. Assume !vma is a shm mapping
4010 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4011 resv_map
= inode_resv_map(inode
);
4013 chg
= region_chg(resv_map
, from
, to
);
4016 resv_map
= resv_map_alloc();
4022 set_vma_resv_map(vma
, resv_map
);
4023 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4032 * There must be enough pages in the subpool for the mapping. If
4033 * the subpool has a minimum size, there may be some global
4034 * reservations already in place (gbl_reserve).
4036 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4037 if (gbl_reserve
< 0) {
4043 * Check enough hugepages are available for the reservation.
4044 * Hand the pages back to the subpool if there are not
4046 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4048 /* put back original number of pages, chg */
4049 (void)hugepage_subpool_put_pages(spool
, chg
);
4054 * Account for the reservations made. Shared mappings record regions
4055 * that have reservations as they are shared by multiple VMAs.
4056 * When the last VMA disappears, the region map says how much
4057 * the reservation was and the page cache tells how much of
4058 * the reservation was consumed. Private mappings are per-VMA and
4059 * only the consumed reservations are tracked. When the VMA
4060 * disappears, the original reservation is the VMA size and the
4061 * consumed reservations are stored in the map. Hence, nothing
4062 * else has to be done for private mappings here
4064 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4065 long add
= region_add(resv_map
, from
, to
);
4067 if (unlikely(chg
> add
)) {
4069 * pages in this range were added to the reserve
4070 * map between region_chg and region_add. This
4071 * indicates a race with alloc_huge_page. Adjust
4072 * the subpool and reserve counts modified above
4073 * based on the difference.
4077 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4079 hugetlb_acct_memory(h
, -rsv_adjust
);
4084 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4085 region_abort(resv_map
, from
, to
);
4086 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4087 kref_put(&resv_map
->refs
, resv_map_release
);
4091 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4094 struct hstate
*h
= hstate_inode(inode
);
4095 struct resv_map
*resv_map
= inode_resv_map(inode
);
4097 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4101 chg
= region_del(resv_map
, start
, end
);
4103 * region_del() can fail in the rare case where a region
4104 * must be split and another region descriptor can not be
4105 * allocated. If end == LONG_MAX, it will not fail.
4111 spin_lock(&inode
->i_lock
);
4112 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4113 spin_unlock(&inode
->i_lock
);
4116 * If the subpool has a minimum size, the number of global
4117 * reservations to be released may be adjusted.
4119 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4120 hugetlb_acct_memory(h
, -gbl_reserve
);
4125 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4126 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4127 struct vm_area_struct
*vma
,
4128 unsigned long addr
, pgoff_t idx
)
4130 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4132 unsigned long sbase
= saddr
& PUD_MASK
;
4133 unsigned long s_end
= sbase
+ PUD_SIZE
;
4135 /* Allow segments to share if only one is marked locked */
4136 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4137 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4140 * match the virtual addresses, permission and the alignment of the
4143 if (pmd_index(addr
) != pmd_index(saddr
) ||
4144 vm_flags
!= svm_flags
||
4145 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4151 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4153 unsigned long base
= addr
& PUD_MASK
;
4154 unsigned long end
= base
+ PUD_SIZE
;
4157 * check on proper vm_flags and page table alignment
4159 if (vma
->vm_flags
& VM_MAYSHARE
&&
4160 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
4166 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4167 * and returns the corresponding pte. While this is not necessary for the
4168 * !shared pmd case because we can allocate the pmd later as well, it makes the
4169 * code much cleaner. pmd allocation is essential for the shared case because
4170 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4171 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4172 * bad pmd for sharing.
4174 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4176 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4177 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4178 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4180 struct vm_area_struct
*svma
;
4181 unsigned long saddr
;
4186 if (!vma_shareable(vma
, addr
))
4187 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4189 i_mmap_lock_write(mapping
);
4190 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4194 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4196 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
4199 get_page(virt_to_page(spte
));
4208 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, spte
);
4210 if (pud_none(*pud
)) {
4211 pud_populate(mm
, pud
,
4212 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4214 put_page(virt_to_page(spte
));
4219 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4220 i_mmap_unlock_write(mapping
);
4225 * unmap huge page backed by shared pte.
4227 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4228 * indicated by page_count > 1, unmap is achieved by clearing pud and
4229 * decrementing the ref count. If count == 1, the pte page is not shared.
4231 * called with page table lock held.
4233 * returns: 1 successfully unmapped a shared pte page
4234 * 0 the underlying pte page is not shared, or it is the last user
4236 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4238 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4239 pud_t
*pud
= pud_offset(pgd
, *addr
);
4241 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4242 if (page_count(virt_to_page(ptep
)) == 1)
4246 put_page(virt_to_page(ptep
));
4248 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4251 #define want_pmd_share() (1)
4252 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4253 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4258 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4262 #define want_pmd_share() (0)
4263 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4265 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4266 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4267 unsigned long addr
, unsigned long sz
)
4273 pgd
= pgd_offset(mm
, addr
);
4274 pud
= pud_alloc(mm
, pgd
, addr
);
4276 if (sz
== PUD_SIZE
) {
4279 BUG_ON(sz
!= PMD_SIZE
);
4280 if (want_pmd_share() && pud_none(*pud
))
4281 pte
= huge_pmd_share(mm
, addr
, pud
);
4283 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4286 BUG_ON(pte
&& !pte_none(*pte
) && !pte_huge(*pte
));
4291 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
4297 pgd
= pgd_offset(mm
, addr
);
4298 if (pgd_present(*pgd
)) {
4299 pud
= pud_offset(pgd
, addr
);
4300 if (pud_present(*pud
)) {
4302 return (pte_t
*)pud
;
4303 pmd
= pmd_offset(pud
, addr
);
4306 return (pte_t
*) pmd
;
4309 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4312 * These functions are overwritable if your architecture needs its own
4315 struct page
* __weak
4316 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
4319 return ERR_PTR(-EINVAL
);
4322 struct page
* __weak
4323 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
4324 pmd_t
*pmd
, int flags
)
4326 struct page
*page
= NULL
;
4329 ptl
= pmd_lockptr(mm
, pmd
);
4332 * make sure that the address range covered by this pmd is not
4333 * unmapped from other threads.
4335 if (!pmd_huge(*pmd
))
4337 if (pmd_present(*pmd
)) {
4338 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
4339 if (flags
& FOLL_GET
)
4342 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t
*)pmd
))) {
4344 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
4348 * hwpoisoned entry is treated as no_page_table in
4349 * follow_page_mask().
4357 struct page
* __weak
4358 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
4359 pud_t
*pud
, int flags
)
4361 if (flags
& FOLL_GET
)
4364 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
4367 #ifdef CONFIG_MEMORY_FAILURE
4370 * This function is called from memory failure code.
4371 * Assume the caller holds page lock of the head page.
4373 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
4375 struct hstate
*h
= page_hstate(hpage
);
4376 int nid
= page_to_nid(hpage
);
4379 spin_lock(&hugetlb_lock
);
4381 * Just checking !page_huge_active is not enough, because that could be
4382 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4384 if (!page_huge_active(hpage
) && !page_count(hpage
)) {
4386 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4387 * but dangling hpage->lru can trigger list-debug warnings
4388 * (this happens when we call unpoison_memory() on it),
4389 * so let it point to itself with list_del_init().
4391 list_del_init(&hpage
->lru
);
4392 set_page_refcounted(hpage
);
4393 h
->free_huge_pages
--;
4394 h
->free_huge_pages_node
[nid
]--;
4397 spin_unlock(&hugetlb_lock
);
4402 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
4406 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4407 spin_lock(&hugetlb_lock
);
4408 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
4412 clear_page_huge_active(page
);
4413 list_move_tail(&page
->lru
, list
);
4415 spin_unlock(&hugetlb_lock
);
4419 void putback_active_hugepage(struct page
*page
)
4421 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4422 spin_lock(&hugetlb_lock
);
4423 set_page_huge_active(page
);
4424 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
4425 spin_unlock(&hugetlb_lock
);